[Fix, Test] (MG_Impl, MG_Backend, MG_Util, MG_Test): memory-audit wave 1 - sync/query teardown and GLES GPU-object lifetimes

- advertise GL_ARB_draw_indirect when supported
- gate GL_ARB_base_instance on complete non-zero firstInstance semantics
- synchronize Driver POST reporting
- add capability and extension-advertisement regression tests

CMakeLists.txt

@@ -182,6 +182,7 @@ set(ENABLE_SPVREMAPPER         OFF CACHE BOOL "Enable SPVRemapper" FORCE)
 set(ENABLE_OPT                 ON  CACHE BOOL "Enable SPIRV-Tools opt usage in glslang" FORCE)
 set(BUILD_EXTERNAL             ON  CACHE BOOL "Build external deps in External/" FORCE)
 set(ENABLE_GLSLANG_INSTALL     OFF CACHE BOOL "Install glslang targets" FORCE)
+set(SPIRV_SKIP_EXECUTABLES     ON  CACHE BOOL "Skip building SPIRV-Tools executables" FORCE)
 
 set(SPIRV_CROSS_C_API          ON  CACHE BOOL "Enable C API" FORCE)
 set(SPIRV_CROSS_ENABLE_GLSL    ON  CACHE BOOL "Enable GLSL backend" FORCE)

MobileGL/Init.cpp

@@ -14,6 +14,7 @@
 #include <MG_State/EGLState/Core.h>
 #include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
 #include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
+#include <MG_Impl/GLImpl/Query/GL_Query.h>
 #include <MG_Impl/GLImpl/Sync/GL_Sync.h>
 #include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
@@ -45,12 +46,13 @@ namespace MobileGL {
             // both of which this function is about to destroy. This is the one
             // cancellation path in the whole design that waits.
             MG_Util::Async::ShaderCompilePool::Get().StopAndDrain();
-            // GL syncs die with their contexts, and every context is gone by the
-            // time full teardown runs: drain the live-sync registry while the
-            // backend function table can still release the backend handles (and
-            // before a re-initialized library could pair them with the wrong
-            // backend's DeleteSync).
+            // GL syncs and queries die with their contexts, and every context is gone
+            // by the time full teardown runs: drain both live registries while the
+            // backend function table can still release the backend handles (and before
+            // a re-initialized library could pair them with the wrong backend's
+            // DeleteSync / DeleteBackendQuery).
             MG_Impl::GLImpl::DestroyAllSyncObjects();
+            MG_Impl::GLImpl::DestroyAllQueryObjects();
             MG_Backend::pActiveBackendObject.reset();
             MG_State::pGLContext.reset();
             MG_State::pEGLContext.reset();

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -712,9 +712,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     {
                         .TargetGLVersion = {4, 0, 0},   // GL target version
                         .TargetGLSLVersion = {4, 6, 0}, // Target Shading Language Version
-                        // Baseline advertisement (no timer queries / anisotropy yet); reconciled
-                        // once the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
-                        .Extensions = BuildAdvertisedExtensions(false, false),
+                        // Baseline advertisement (no runtime capabilities yet); reconciled once
+                        // the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
+                        .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                         .IsCompatibilityProfile = false // Is Compatibility Profile
                     },
                 .StaticBackendCapability = {.AllowVSOnlyPrograms = false} // Backend Capability
@@ -734,9 +734,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // thread can only observe the extension string after the
         // advertisement for its context has settled; rebuilding the whole
         // list keeps the re-run after a context recreation idempotent.
-        void UpdateAdvertisedCapabilityExtensions(Bool anisotropicFilteringSupported) {
-            MutableRendererInfo().RendererGLInfo.Extensions =
-                BuildAdvertisedExtensions(AreTimerQueriesSupported(), anisotropicFilteringSupported);
+        void UpdateAdvertisedCapabilityExtensions(const MG_External::GLESCapabilities& capabilities) {
+            MutableRendererInfo().RendererGLInfo.Extensions = BuildAdvertisedExtensions(
+                AreTimerQueriesSupported(), capabilities.SupportsTextureFilterAnisotropy,
+                capabilities.SupportsDrawIndirect,
+                capabilities.SupportsDrawIndirect && capabilities.SupportsBaseInstance);
         }
     } // namespace
 
@@ -779,11 +781,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
         DirectGLES::SetGLESCapabilities(m_GLESCapabilities);
-        // Now that g_GLESCapabilities knows about GL_EXT_disjoint_timer_query and
-        // GL_EXT_texture_filter_anisotropic, reconcile the advertisement (see the comment on
-        // UpdateAdvertisedCapabilityExtensions for why it cannot happen when the extension
-        // list is first built).
-        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities.SupportsTextureFilterAnisotropy);
+        // Now that g_GLESCapabilities knows the host extensions, entry points, and ES version,
+        // reconcile every runtime-gated advertisement (see the comment on
+        // UpdateAdvertisedCapabilityExtensions for why this cannot happen when the list is first
+        // built).
+        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities);
         UpdateDynamicBackendParameters();
         PopulateFormatCapabilities(m_GLESFunctions, m_GLESCapabilities, MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -924,7 +926,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return MutableRendererInfo();
     }
 
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported) {
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
@@ -955,6 +959,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Minecraft 26.3 checks this prerequisite before it even considers
+        // GL_ARB_multi_draw_indirect. ES 3.1 supplies both single-draw entry points; the loader
+        // folds the version and pointer checks into SupportsDrawIndirect.
+        if (drawIndirectSupported) {
+            extensions.push_back(E_GL_ARB_draw_indirect);
+        }
+        // ARB_base_instance also defines the last word of an indirect command. Direct calls are
+        // emulated on every Espryt device, but without host GL_EXT_base_instance a native indirect
+        // draw cannot shift divisor attributes by a GPU-authored non-zero value, so do not promise
+        // that incomplete case.
+        if (drawIndirectSupported && nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         // GL_KHR_parallel_shader_compile is MobileGL's own capability, not the host ES
         // driver's: the compiler threads are MobileGL's, and glCompileShader/glLinkProgram
         // are serviced entirely inside the frontend. Whether the device driver advertises

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.h

@@ -67,9 +67,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
     const RendererInfo& GetRendererIdentity();
 
     // The full OpenGL extension list Espryt advertises (glGetString(GL_EXTENSIONS))
-    // for a device whose timer queries / anisotropic filtering are (or are not) usable.
+    // for a device whose timer queries / anisotropic filtering / native indirect draws /
+    // non-zero indirect baseInstance semantics are (or are not) usable.
     // The MOBILEGL_DISABLE_TIMERQUERY escape hatch is applied inside.
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported);
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <OpenGL ES Renderer>, OpenGL ES <Major>.<Minor> — the exact string an
     // initialized backend returns from GetBackendAPIVersionString (and that ends up

MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp

@@ -1600,7 +1600,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 !g_hasSyncedRenderState || std::memcmp(currentBytes + kBlendSpanEnd, syncedBytes + kBlendSpanEnd,
                                                        sizeof(RenderStateParameters) - kBlendSpanEnd) != 0;
 
-            IntVec4 backendViewport = parameters.Viewport;
+            IntVec4 backendViewport = MG_State::pGLContext->GetViewport();
             if (backendViewport.z() <= 0 || backendViewport.w() <= 0) {
                 Int surfaceWidth = 0;
                 Int surfaceHeight = 0;
@@ -1614,7 +1614,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 g_syncedBackendViewport = backendViewport;
             }
 
-            // All 12 capability bools live after LogicOp in the struct, i.e. in the tail span.
+            // Every capability bool (and the scissor-test mask below) lives after LogicOp in the
+            // struct, i.e. in the tail span.
             if (tailSpanDirty) {
 #define SYNC_CAPABILITY(cap_mg, cap_gl)                                                                                \
     if (forceFullPush || parameters.cap_mg##Enabled != g_syncedRenderStateParameters.cap_mg##Enabled) {                                 \
@@ -1633,11 +1634,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 SYNC_CAPABILITY(SampleMask, GL_SAMPLE_MASK);
                 SYNC_CAPABILITY(PolygonOffsetFill, GL_POLYGON_OFFSET_FILL);
                 SYNC_CAPABILITY(RasterizerDiscard, GL_RASTERIZER_DISCARD);
-                SYNC_CAPABILITY(ScissorTest, GL_SCISSOR_TEST);
                 SYNC_CAPABILITY(StencilTest, GL_STENCIL_TEST);
                 SYNC_CAPABILITY(CullFace, GL_CULL_FACE);
 
 #undef SYNC_CAPABILITY
+
+                // GL_SCISSOR_TEST is per-viewport enable state (ARB_viewport_array), so it is a
+                // 16-bit mask and not a "<Name>Enabled" bool the macro above could key off. ES
+                // has exactly one scissor rectangle and one scissor enable, so only bit 0 - the
+                // index every ES draw rasterizes against - can be forwarded; a program that
+                // enables the test for viewport 3 alone gets viewport 0's answer here. That is
+                // the same limitation as the unemulated gl_ViewportIndex on this backend and is
+                // why the multi-viewport half of KHR-GL43.viewport_array stays red on Espryt.
+                {
+                    const Bool scissorTest = (parameters.ScissorTestEnabledMask & 1u) != 0;
+                    const Bool syncedScissorTest =
+                        (g_syncedRenderStateParameters.ScissorTestEnabledMask & 1u) != 0;
+                    if (forceFullPush || scissorTest != syncedScissorTest) {
+                        scissorTest ? g_GLESFuncs.glEnable(GL_SCISSOR_TEST) : g_GLESFuncs.glDisable(GL_SCISSOR_TEST);
+                    }
+                }
             }
 
             if (tailSpanDirty && g_GLESCapabilities.SupportsClipDistance) {
@@ -1864,8 +1880,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (forceFullPush || parameters.DepthMask != g_syncedRenderStateParameters.DepthMask) {
                     g_GLESFuncs.glDepthMask(parameters.DepthMask ? GL_TRUE : GL_FALSE);
                 }
-                if (forceFullPush || parameters.DepthRange != g_syncedRenderStateParameters.DepthRange) {
-                    g_GLESFuncs.glDepthRangef(parameters.DepthRange.x(), parameters.DepthRange.y());
+                if (forceFullPush || parameters.DepthRanges[0] != g_syncedRenderStateParameters.DepthRanges[0]) {
+                    g_GLESFuncs.glDepthRangef(parameters.DepthRanges[0].x(), parameters.DepthRanges[0].y());
                 }
             }
 
@@ -2003,7 +2019,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
               // everything drawn with GL_SCISSOR_TEST enabled before the app's first glScissor
               // is clipped away - Minecraft 26.2 keeps only its unscissored sky and hand and
               // loses the terrain and the whole GUI.
-                IntVec4 backendScissorBox = parameters.ScissorBox;
+                IntVec4 backendScissorBox = parameters.ScissorBoxes[0];
                 if (backendScissorBox.z() <= 0 || backendScissorBox.w() <= 0) {
                     Int surfaceWidth = 0;
                     Int surfaceHeight = 0;
@@ -2189,6 +2205,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // accident - and it never covered the monolithic glUseProgram path at all - so the
             // dependency is stated here instead.
             if (!twin->GetBackendProgramId() ||
+                twin->GetContextGeneration() != g_backendContextGeneration ||
                 twin->GetSyncedLinkVersion() != currentProgram->GetLinkVersion() ||
                 twin->GetSyncedImageUnitVersion() != currentProgram->GetImageUnitVersion() ||
                 twin->GetSnormFallbackClampOutputMask() != g_snormFallbackClampOutputMask ||
@@ -3038,6 +3055,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
         const auto program = GetCurrentBackendProgram();
         if (!currentProgram || program == nullptr ||
+            program->GetContextGeneration() != g_backendContextGeneration ||
             program->GetSyncedLinkVersion() != currentProgram->GetLinkVersion()) {
             return true;
         }
@@ -3045,10 +3063,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
     }
 
     static Bool SupportsNativeIndirectDraws() {
-        const auto& version = g_GLESCapabilities.GLESVersion;
-        const Bool esVersionOk = version.Major > 3 || (version.Major == 3 && version.Minor >= 1);
-        return esVersionOk && g_GLESFuncs.glDrawElementsIndirect != nullptr &&
-               g_GLESFuncs.glDrawArraysIndirect != nullptr;
+        return g_GLESCapabilities.SupportsDrawIndirect;
     }
 
     // Runs an (indexed) indirect multi-draw. When a GL_DRAW_INDIRECT_BUFFER is bound the draws
@@ -4762,7 +4777,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
     // restores the app state on exit, tracked via the render-state shadow.
     class ScopedScissorDisable {
     public:
-        ScopedScissorDisable() : m_wasEnabled(RenderStateImpl::g_syncedRenderStateParameters.ScissorTestEnabled) {
+        ScopedScissorDisable()
+            : m_wasEnabled((RenderStateImpl::g_syncedRenderStateParameters.ScissorTestEnabledMask & 1u) != 0) {
             if (m_wasEnabled) g_GLESFuncs.glDisable(GL_SCISSOR_TEST);
         }
         ~ScopedScissorDisable() {
@@ -5563,6 +5579,56 @@ namespace MobileGL::MG_Backend::DirectGLES {
         g_GLESFuncs.glMemoryBarrierByRegion(barriers);
     }
 
+    // One endpoint of a glCopyImageSubData, expressed the way the ES driver stores it.
+    //
+    // The frontend hands this backend the target the APPLICATION named, and three of the
+    // targets core GL has do not exist in ES at all. They are not missing here either - the
+    // texture managers already store a 1D texture as a height-1 2D one, a 1D array as a
+    // height-1 2D array and a rectangle texture as a plain 2D one (MapToBackendTextureTarget) -
+    // but glCopyImageSubData was the one path that never asked for that translation and passed
+    // 0x84F5 / 0x0DE0 / 0x8C18 straight through. ES rejects the enum, the copy does not happen,
+    // and with the error only asserted on (asserts are compiled out of an INFO build) the
+    // destination silently keeps whatever it held.
+    //
+    // The 1D-array case is not just a rename: GL addresses its layers with y/height while the
+    // ES 2D array that backs it addresses them with z/depth, so the two axes swap with the
+    // target.
+    struct GLESCopyImageEndpoint {
+        GLenum target = GL_TEXTURE_2D;
+        GLint x = 0;
+        GLint y = 0;
+        GLint z = 0;
+    };
+
+    static GLESCopyImageEndpoint MakeGLESCopyImageEndpoint(GLenum appTarget, GLint x, GLint y, GLint z) {
+        const TextureTarget stateTarget = MG_Util::ConvertGLEnumToTextureTarget(appTarget);
+        GLESCopyImageEndpoint endpoint{};
+        endpoint.target = TextureImpl::ConvertTextureTargetToBackendGLEnum(stateTarget);
+        if (stateTarget == TextureTarget::Texture1DArray) {
+            endpoint.x = x;
+            endpoint.y = 0;
+            endpoint.z = y;
+            return endpoint;
+        }
+        endpoint.x = x;
+        endpoint.y = y;
+        endpoint.z = z;
+        return endpoint;
+    }
+
+    // The region extent swaps the same two axes for a 1D array, and does so for whichever side
+    // of the copy is one - GL forbids a copy whose two endpoints disagree about how many layers
+    // move, so at most one of the two can be a 1D array only in the degenerate single-layer
+    // case, where the swap is the identity anyway.
+    static void ApplyGLESCopyImageExtent(GLenum appSrcTarget, GLenum appDstTarget, GLsizei& height, GLsizei& depth) {
+        const TextureTarget srcStateTarget = MG_Util::ConvertGLEnumToTextureTarget(appSrcTarget);
+        const TextureTarget dstStateTarget = MG_Util::ConvertGLEnumToTextureTarget(appDstTarget);
+        if (srcStateTarget != TextureTarget::Texture1DArray && dstStateTarget != TextureTarget::Texture1DArray) {
+            return;
+        }
+        std::swap(height, depth);
+    }
+
     void CopyImageSubData(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture,
                           GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
                           const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
@@ -5592,6 +5658,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return;
         }
 
+        const GLESCopyImageEndpoint src = MakeGLESCopyImageEndpoint(srcTarget, srcX, srcY, srcZ);
+        const GLESCopyImageEndpoint dst = MakeGLESCopyImageEndpoint(dstTarget, dstX, dstY, dstZ);
+        GLsizei copyHeight = srcHeight;
+        GLsizei copyDepth = srcDepth;
+        ApplyGLESCopyImageExtent(srcTarget, dstTarget, copyHeight, copyDepth);
+
         const Bool srcIsDepth = MG_Util::IsDepthFormatInternalFormat(srcTexture->GetFormat());
         const Bool dstIsDepth = MG_Util::IsDepthFormatInternalFormat(dstTexture->GetFormat());
         const Bool srcStencil = MG_Util::IsStencilFormatInternalFormat(srcTexture->GetFormat());
@@ -5599,12 +5671,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
         if (srcIsDepth || dstIsDepth || srcStencil || dstStencil) {
             MOBILEGL_ASSERT(srcIsDepth && dstIsDepth && !srcStencil && !dstStencil,
                             "DirectGLES CopyImageSubData only supports depth-only image copies.");
-            MOBILEGL_ASSERT(srcTarget == GL_TEXTURE_2D && dstTarget == GL_TEXTURE_2D,
+            MOBILEGL_ASSERT(src.target == GL_TEXTURE_2D && dst.target == GL_TEXTURE_2D,
                             "DirectGLES depth CopyImageSubData only supports GL_TEXTURE_2D.");
-            MOBILEGL_ASSERT(srcZ == 0 && dstZ == 0 && srcDepth == 1,
+            MOBILEGL_ASSERT(src.z == 0 && dst.z == 0 && copyDepth == 1,
                             "DirectGLES depth CopyImageSubData only supports single-layer copies.");
-            BlitDepthTexture2D(srcBackendTexture->GetBackendTextureId(), srcLevel, srcX, srcY, srcWidth, srcHeight,
-                               dstBackendTexture->GetBackendTextureId(), dstLevel, dstX, dstY, srcWidth, srcHeight);
+            BlitDepthTexture2D(srcBackendTexture->GetBackendTextureId(), srcLevel, src.x, src.y, srcWidth, copyHeight,
+                               dstBackendTexture->GetBackendTextureId(), dstLevel, dst.x, dst.y, srcWidth, copyHeight);
             return;
         }
 
@@ -5615,29 +5687,43 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // with the always-live helper so a stale flag cannot misroute a
             // succeeded native copy into the 2D-only fallback.
             ClearGLErrors();
-            g_GLESFuncs.glCopyImageSubData(srcBackendTexture->GetBackendTextureId(), srcTarget, srcLevel, srcX, srcY, srcZ,
-                                           dstBackendTexture->GetBackendTextureId(), dstTarget, dstLevel, dstX, dstY, dstZ,
-                                           srcWidth, srcHeight, srcDepth);
+            g_GLESFuncs.glCopyImageSubData(srcBackendTexture->GetBackendTextureId(), src.target, srcLevel, src.x, src.y, src.z,
+                                           dstBackendTexture->GetBackendTextureId(), dst.target, dstLevel, dst.x, dst.y, dst.z,
+                                           srcWidth, copyHeight, copyDepth);
             const GLenum copyImageError = g_GLESFuncs.glGetError();
             if (copyImageError == GL_NO_ERROR) {
                 return;
             }
             MOBILEGL_ASSERT(IsColorOnlyFormat(srcTexture->GetFormat()) && IsColorOnlyFormat(dstTexture->GetFormat()),
                             "DirectGLES CopyImageSubData only supports color-only or depth-only copies.");
-            MOBILEGL_ASSERT(srcTarget == GL_TEXTURE_2D && dstTarget == GL_TEXTURE_2D,
+            MOBILEGL_ASSERT(src.target == GL_TEXTURE_2D && dst.target == GL_TEXTURE_2D,
                             "DirectGLES color CopyImageSubData only supports GL_TEXTURE_2D.");
-            MOBILEGL_ASSERT(srcZ == 0 && dstZ == 0 && srcDepth == 1,
+            MOBILEGL_ASSERT(src.z == 0 && dst.z == 0 && copyDepth == 1,
                             "DirectGLES color CopyImageSubData only supports single-layer copies.");
-            CopyR32FTexture2D(srcBackendTexture->GetBackendTextureId(), srcLevel, srcX, srcY, srcWidth, srcHeight,
-                              dstBackendTexture->GetBackendTextureId(), dstTarget, dstLevel, dstX, dstY);
+            CopyR32FTexture2D(srcBackendTexture->GetBackendTextureId(), srcLevel, src.x, src.y, srcWidth, copyHeight,
+                              dstBackendTexture->GetBackendTextureId(), dst.target, dstLevel, dst.x, dst.y);
             return;
         }
 
         ClearGLErrors();
-        g_GLESFuncs.glCopyImageSubData(srcBackendTexture->GetBackendTextureId(), srcTarget, srcLevel, srcX, srcY, srcZ,
-                                       dstBackendTexture->GetBackendTextureId(), dstTarget, dstLevel, dstX, dstY, dstZ,
-                                       srcWidth, srcHeight, srcDepth);
-        AssertNoGLError("glCopyImageSubData");
+        g_GLESFuncs.glCopyImageSubData(srcBackendTexture->GetBackendTextureId(), src.target, srcLevel, src.x, src.y, src.z,
+                                       dstBackendTexture->GetBackendTextureId(), dst.target, dstLevel, dst.x, dst.y, dst.z,
+                                       srcWidth, copyHeight, copyDepth);
+        // Every error condition glCopyImageSubData has was already ruled out by the frontend
+        // validator, so a driver error here is an internal invariant violation, not something
+        // the application can provoke. Say so where an INFO build can still see it, then trap
+        // in the builds that trap - the previous bare assert left a release build with a
+        // destination that silently kept its old contents.
+        const GLenum copyImageError = g_GLESFuncs.glGetError();
+        if (copyImageError != GL_NO_ERROR) {
+            MGLOG_E_ONCE("glCopyImageSubData failed: %s. src target=%s (app %s), dst target=%s (app %s)",
+                         MG_Util::ConvertGLEnumToString(copyImageError).c_str(),
+                         MG_Util::ConvertGLEnumToString(src.target).c_str(),
+                         MG_Util::ConvertGLEnumToString(srcTarget).c_str(),
+                         MG_Util::ConvertGLEnumToString(dst.target).c_str(),
+                         MG_Util::ConvertGLEnumToString(dstTarget).c_str());
+            MOBILEGL_ASSERT(false, "glCopyImageSubData failed after frontend validation accepted the request.");
+        }
     }
 
     void BindImageTexture(GLuint unit, GLuint texture, GLint level, GLboolean layered, GLint layer, GLenum access,
@@ -5815,6 +5901,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // fallen behind is about to be rebuilt anyway, and its current driver interface is
         // the PREVIOUS link's - applying to it could land the binding on an unrelated block.
         if (!backendObj->GetBackendProgramId() ||
+            backendObj->GetContextGeneration() != g_backendContextGeneration ||
             backendObj->GetSyncedLinkVersion() != programObject->GetLinkVersion()) {
             return; // SyncToBackend's reseed will carry it
         }

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -1468,6 +1468,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             m_clientAttributeBufferIds.fill(0);
+            m_contextGeneration = g_backendContextGeneration;
             g_GLESFuncs.glGenVertexArrays(1, &m_backendVAOId);
             if (m_backendVAOId == 0) {
                 MGLOG_E_ONCE("Failed to generate vertex array object.");
@@ -1481,17 +1482,28 @@ namespace MobileGL::MG_Backend::DirectGLES {
             if (InProcessTeardown()) {
                 return; // see InProcessTeardown(): the driver may be unloaded already
             }
+            const Bool contextCurrent = m_contextGeneration == g_backendContextGeneration;
             if (m_backendVAOId != 0) {
+                // Scrub the binding shadow whether or not the id can still be
+                // deleted: a recycled name must never satisfy the shadow's dedup.
                 NoteVAOIdDeleted(m_backendVAOId);
-                g_GLESFuncs.glDeleteVertexArrays(1, &m_backendVAOId);
+                if (contextCurrent && g_GLESFuncs.glDeleteVertexArrays) {
+                    g_GLESFuncs.glDeleteVertexArrays(1, &m_backendVAOId);
+                }
                 m_backendVAOId = 0;
             }
             for (auto& bufferId : m_clientAttributeBufferIds) {
-                if (bufferId != 0) {
-                    BufferImpl::NoteBufferIdDeleted(bufferId);
-                    g_GLESFuncs.glDeleteBuffers(1, &bufferId);
-                    bufferId = 0;
+                if (bufferId == 0) {
+                    continue;
                 }
+                // Same discipline as the VAO id itself: a buffer id from a dead
+                // context belongs to that context and must never be deleted as a
+                // recycled name in a successor context.
+                BufferImpl::NoteBufferIdDeleted(bufferId);
+                if (contextCurrent && g_GLESFuncs.glDeleteBuffers) {
+                    g_GLESFuncs.glDeleteBuffers(1, &bufferId);
+                }
+                bufferId = 0;
             }
         }
 
@@ -1635,6 +1647,30 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // PrepareForDraw's BindCurrentVAO establishes the draw binding regardless.
             const Uint32 currentConfigVersion = stateVAOObject->GetConfigVersion();
             const Uint16 currentIndexBufferVersion = stateVAOObject->GetIndexBufferBindingSlot().GetVersion();
+
+            // The ES context was recreated since this twin last ran. Its GL names belong to
+            // the dead context and are gone; mint a fresh VAO and force every attribute /
+            // index-binding cache to re-emit. No glDelete* here: the old names are not ours
+            // to delete in the successor context.
+            if (m_contextGeneration != g_backendContextGeneration) {
+                InvalidateVAOBindingCache();
+                m_backendVAOId = 0;
+                m_contextGeneration = g_backendContextGeneration;
+                m_clientAttributeBufferIds.fill(0);
+                m_isInitialized = false;
+                m_resolvedDrawBuffers = {};
+                m_pendingAttribValueMask = {};
+                m_hasSyncedConfigVersion = false;
+                m_syncedConfigVersion = 0;
+                m_syncedIndexBufferVersion = static_cast<Uint16>(currentIndexBufferVersion + 1);
+                m_syncedAttributeVersions.fill({});
+                m_syncedFetchBaseInstance = 0;
+                g_GLESFuncs.glGenVertexArrays(1, &m_backendVAOId);
+                if (m_backendVAOId == 0) {
+                    MGLOG_E_ONCE("Failed to recreate vertex array object for a new ES context.");
+                }
+            }
+
             const Bool attributesDirty = !m_hasSyncedConfigVersion || m_syncedConfigVersion != currentConfigVersion;
             const Bool indexBufferDirty = currentIndexBufferVersion != m_syncedIndexBufferVersion;
 
@@ -1978,6 +2014,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                     TextureSwizzleParam::Alpha};
             m_cacheDepthStencilTextureMode = GL_DEPTH_COMPONENT;
             m_forceTextureParamsResync = true;
+            m_forceSamplerResync = true;
         }
 
         // Sets the backend GL unpack state to MobileGL's upload default for the scope,
@@ -2395,6 +2432,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            // The ES context was recreated since this twin last ran. Recreate the
+            // texture id before any version-based early-out below: those versions are
+            // frontend versions and do not move when only the backend context changed.
+            if (m_contextGeneration != g_backendContextGeneration) {
+                RecreateBackendTexture();
+            }
+
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
@@ -3119,14 +3163,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            if (m_contextGeneration != g_backendContextGeneration) {
+                RecreateBackendTexture();
+            }
+
             auto* samplerObject = stateTextureObject->GetSamplerObject().get();
             Uint currentSamplerVersion = samplerObject->GetVersion();
-            if (m_syncedSamplerVersion == currentSamplerVersion) {
+            if (m_syncedSamplerVersion == currentSamplerVersion && !m_forceSamplerResync) {
                 MGLOG_D("Sampler parameters have not changed for texture ID: %u, skipping sync.", m_backendTextureId);
                 return;
             }
 
             m_syncedSamplerVersion = currentSamplerVersion;
+            m_forceSamplerResync = false;
 
             MGLOG_D("Syncing texture built-in sampler with backend ID %u to backend for state ID %u",
                     m_backendTextureId, stateTextureObject->GetExternalIndex());
@@ -3229,6 +3278,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            if (m_contextGeneration != g_backendContextGeneration) {
+                RecreateBackendTexture();
+            }
+
             Uint16 currentTextureParamsVersion = stateTextureObject->GetTextureParamsVersion();
             if (m_syncedTextureParamsVersion == currentTextureParamsVersion && !m_forceTextureParamsResync) {
                 MGLOG_D("Texture parameters have not changed for texture ID: %u, skipping sync.", m_backendTextureId);
@@ -3902,6 +3955,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 MGLOG_E_ONCE("State FBO object is null, cannot sync to backend.");
                 return;
             }
+            // Recreate the driver FBO when the ES context has moved on. The old id is
+            // gone with the old context; calling glDeleteFramebuffers on its recycled
+            // numeric value could delete a new live FBO, so simply abandon it.
+            if (m_contextGeneration != g_backendContextGeneration) {
+                m_backendFBOId = 0;
+                m_contextGeneration = g_backendContextGeneration;
+                g_GLESFuncs.glGenFramebuffers(1, &m_backendFBOId);
+                if (m_backendFBOId == 0) {
+                    MGLOG_E_ONCE("Failed to recreate framebuffer object for a new ES context.");
+                }
+                InvalidateFramebufferBindingCache();
+                InvalidateSyncedState();
+            }
             MGLOG_D("Syncing FBO with backend ID %u to backend for state ID %u, as %s FBO", m_backendFBOId,
                     stateFBOObject->GetExternalIndex(), (asTarget == FramebufferTarget::Draw ? "DRAW" : "READ"));
             GLenum glFBOTarget = MG_Util::ConvertFramebufferTargetToGLEnum(asTarget);
@@ -4412,10 +4478,27 @@ namespace MobileGL::MG_Backend::DirectGLES {
         Uint g_lastUsedBackendProgramId = 0;
         StateBackendObjectRegistry<MG_State::GLState::ProgramObject, BackendProgramObjectImpl> g_backendProgramObjects;
 
+        void DeleteBackendProgramGlobalUbo(Uint& bufferId, Uint contextGeneration) {
+            if (bufferId == 0) {
+                return;
+            }
+            // Only a buffer that belongs to the LIVE context may be deleted. A stale
+            // generation means the old ES context already reclaimed it; handing its
+            // recycled numeric id to glDeleteBuffers could delete a new live buffer.
+            if (contextGeneration == g_backendContextGeneration) {
+                BufferImpl::NoteBufferIdDeleted(bufferId);
+                if (g_GLESFuncs.glDeleteBuffers) {
+                    g_GLESFuncs.glDeleteBuffers(1, &bufferId);
+                }
+            }
+            bufferId = 0;
+        }
+
         BackendProgramObjectImpl::BackendProgramObjectImpl() {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
+            m_contextGeneration = g_backendContextGeneration;
             m_backendProgramId = g_GLESFuncs.glCreateProgram();
             if (m_backendProgramId == 0) {
                 MGLOG_E_ONCE("Failed to create program object in backend.");
@@ -4433,14 +4516,23 @@ namespace MobileGL::MG_Backend::DirectGLES {
             if (InProcessTeardown()) {
                 return; // see InProcessTeardown(): the driver may be unloaded already
             }
+            DeleteBackendProgramGlobalUbo(m_backendGlobalUBOId, m_contextGeneration);
             if (m_backendProgramId != 0) {
-                MGLOG_D("Deleting backend program object with ID: %u", m_backendProgramId);
-                g_GLESFuncs.glDeleteProgram(m_backendProgramId);
+                // Same generation rule as the global UBO: a program id from a dead
+                // context is gone already and must not be deleted as a recycled name
+                // in a successor context.
+                if (m_contextGeneration == g_backendContextGeneration) {
+                    MGLOG_D("Deleting backend program object with ID: %u", m_backendProgramId);
+                    if (g_GLESFuncs.glDeleteProgram) {
+                        g_GLESFuncs.glDeleteProgram(m_backendProgramId);
+                    }
+                }
                 // The driver may recycle this GL name for a future program; a stale
                 // guard entry would then wrongly skip the glUseProgram for it.
                 if (g_lastUsedBackendProgramId == m_backendProgramId) {
                     g_lastUsedBackendProgramId = 0;
                 }
+                m_backendProgramId = 0;
             }
         }
 
@@ -4677,6 +4769,31 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            // The ES context was recreated since this twin last ran. The old program id
+            // and global UBO id belong to the dead context; drop them without GL calls
+            // and mint a fresh program before reusing any cached reflection/version data.
+            if (m_contextGeneration != g_backendContextGeneration) {
+                DeleteBackendProgramGlobalUbo(m_backendGlobalUBOId, m_contextGeneration);
+                m_backendProgramId = 0;
+                m_contextGeneration = g_backendContextGeneration;
+                m_backendProgramId = g_GLESFuncs.glCreateProgram();
+                if (m_backendProgramId == 0) {
+                    MGLOG_E_ONCE("Failed to recreate backend program object for a new ES context.");
+                }
+                m_isInitialized = false;
+                m_backendProgramUsable = false;
+                m_syncedLinkVersion = ~0u;
+                m_syncedImageUnitVersion = ~0u;
+                m_lastUploadedGlobalUboVersion = ~0u;
+                m_globalUboBackendBlockIndex = -1;
+                m_globalUboBackendBlockSize = 0;
+                m_uniformBlockBackendIndices.clear();
+                m_samplerUniformBindings.clear();
+                m_formatlessImageUnits.clear();
+                m_imageUnitFormatSignature = 0;
+                m_globalUboRingAllocation = {};
+            }
+
             MGLOG_D("Syncing program to backend. State program ID: %u, Backend ID: %u",
                     stateProgramObject->GetExternalIndex(), m_backendProgramId);
             // Every link-derived cache below (incl. m_samplerUniformBindings and its
@@ -5154,7 +5271,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 }
             }
 
-            // Create global UBO
+            // Create global UBO. Delete any previous one first: relink reuses this
+            // backend program, and without this every relink leaked the old buffer.
+            DeleteBackendProgramGlobalUbo(m_backendGlobalUBOId, m_contextGeneration);
             if (stateProgramObject->GetUBOSize() > 0) {
                 g_GLESFuncs.glGenBuffers(1, &m_backendGlobalUBOId);
                 g_GLESFuncs.glBindBuffer(GL_UNIFORM_BUFFER, m_backendGlobalUBOId);
@@ -5389,6 +5508,20 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            if (m_contextGeneration != g_backendContextGeneration) {
+                // Old sampler id died with the old context; abandon it and mint a new
+                // one before the version-based early-out below can reuse a dead name.
+                m_backendSamplerId = 0;
+                m_contextGeneration = g_backendContextGeneration;
+                g_GLESFuncs.glGenSamplers(1, &m_backendSamplerId);
+                if (m_backendSamplerId == 0) {
+                    MGLOG_E_ONCE("Failed to recreate sampler object for a new ES context.");
+                }
+                m_isInitialized = false;
+                m_cacheSamplerParameters = {};
+                g_boundSamplersCache.fill(nullptr);
+            }
+
             Uint currentSamplerVersion = stateSamplerObject->GetVersion();
             if (m_isInitialized && m_syncedSamplerVersion == currentSamplerVersion) {
                 MGLOG_D("Sampler parameters have not changed for sampler ID: %u, skipping sync.",
@@ -5533,6 +5666,23 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
+            if (m_contextGeneration != g_backendContextGeneration) {
+                // The old renderbuffer id died with the old context. Abandon it and
+                // force a fresh allocation instead of letting the parameter early-out
+                // below keep using a dead name.
+                m_backendRBOId = 0;
+                m_contextGeneration = g_backendContextGeneration;
+                g_GLESFuncs.glGenRenderbuffers(1, &m_backendRBOId);
+                if (m_backendRBOId == 0) {
+                    MGLOG_E_ONCE("Failed to recreate renderbuffer object for a new ES context.");
+                }
+                m_isInitialized = false;
+                m_cacheInternalFormat = TextureInternalFormat::Unknown;
+                m_cacheWidth = -1;
+                m_cacheHeight = -1;
+                m_cacheSamples = -1;
+            }
+
             MGLOG_D("Syncing RBO with backend ID %u to backend for state ID %u", m_backendRBOId,
                     stateRBOObject->GetExternalIndex());
 

MobileGL/MG_Backend/DirectGLES/Managers.h

@@ -406,6 +406,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void SyncClientSideAttributesForDrawArrays(
                 const SharedPtr<MG_State::GLState::VertexArrayObject>& stateVAOObject, GLint first, GLsizei count);
             Uint GetBackendVertexArrayId() const { return m_backendVAOId; }
+            Uint GetContextGeneration() const { return m_contextGeneration; }
             void Bind() const;
 
             // Draw-path memo of SyncNeccessaryBuffers' attribute walk for this VAO: the
@@ -462,6 +463,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ResolvedDrawBuffers m_resolvedDrawBuffers;
             PendingAttribValueMask m_pendingAttribValueMask;
             Uint m_backendVAOId = 0;
+            // ES context generation the VAO id and client-attribute buffer ids were
+            // created under; ids from a dead context must never be deleted against a
+            // successor context (both contexts restart GL names at 1).
+            Uint m_contextGeneration = 0;
             Array<Uint, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS> m_clientAttributeBufferIds;
             Bool m_isInitialized = false;
             Uint16 m_syncedIndexBufferVersion = 0;
@@ -711,6 +716,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // parameter already pushed onto it: the params-version early-out has to be overridden
             // once, or an unchanged version would skip the re-push forever.
             Bool m_forceTextureParamsResync = false;
+            // Same latch for the built-in sampler parameters.
+            Bool m_forceSamplerResync = false;
         };
 
         void ActivateTextureUnit(Uint unit);
@@ -1087,6 +1094,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Bool ReadsBaseVertex() const { return m_baseVertexUniformLocation >= 0; }
             Int GetIndirectParamsBinding() const { return m_indirectParamsBinding; }
             Uint GetBackendProgramId() const { return m_backendProgramId; }
+            Uint GetContextGeneration() const { return m_contextGeneration; }
             // False when the last SyncToBackend could not produce a usable program (a
             // shader failed to transpile or compile, or the link itself failed). Use()
             // must not leave the previously bound program current in that case.
@@ -1149,6 +1157,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void CacheResourceLocations(const SharedPtr<MG_State::GLState::ProgramObject>& stateProgramObject);
 
             Uint m_backendProgramId = 0;
+            // ES context generation the backend program and its global UBO were created
+            // under. A stale twin must be recreated, never deleted against a successor
+            // context (both contexts restart GL names at 1).
+            Uint m_contextGeneration = 0;
             // GL name of the frontend program this was last synced from; diagnostics only, so
             // an unusable backend program can be traced back to the glCreateProgram id the app
             // knows it by.
@@ -1196,6 +1208,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // skip redundant rebinds. Reset to 0 wherever glUseProgram(0) is issued or the
         // ES context is recreated.
         extern Uint g_lastUsedBackendProgramId;
+        // Deletes `bufferId` only while it still belongs to the live ES context. Stale
+        // generations are abandoned without a GL call: the old context already reclaimed
+        // the buffer, and its numeric id may now name a live buffer in a successor context.
+        void DeleteBackendProgramGlobalUbo(Uint& bufferId, Uint contextGeneration);
         extern StateBackendObjectRegistry<MG_State::GLState::ProgramObject, BackendProgramObjectImpl>
             g_backendProgramObjects;
 

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -497,20 +497,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .ExtraVendor = Nullopt,
             .RendererGLInfo = {.TargetGLVersion = {4, 0, 0},
                                .TargetGLSLVersion = {4, 6, 0},
-                               // Baseline advertisement (no shader subgroup, no timer queries); a
-                               // live backend reconciles its copy in UpdateAdvertisedExtensions.
-                               .Extensions = BuildAdvertisedExtensions(false, false, false),
+                               // Baseline advertisement (no runtime-gated capabilities); a live
+                               // backend reconciles its copy in UpdateAdvertisedExtensions.
+                               .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                                .IsCompatibilityProfile = false},
             .StaticBackendCapability = {.AllowVSOnlyPrograms = false}};
         return rendererInfo;
     }
 
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported) {
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_multi_draw_indirect,
+            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_draw_indirect,
+            E_GL_ARB_multi_draw_indirect,
             E_GL_ARB_indirect_parameters, E_GL_EXT_framebuffer_object, E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage,
             E_GL_ARB_texture_storage, E_GL_ARB_texture_storage_multisample, E_GL_ARB_texture_multisample,
             E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access, E_GL_ARB_shader_draw_parameters,
@@ -530,6 +532,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Vulkan's drawIndirectFirstInstance feature is optional. Direct base-instance calls work
+        // without it, but ARB_base_instance also promises non-zero firstInstance in GPU indirect
+        // commands; the renderer supplies true only when that word is legal and gl_InstanceID can
+        // be rebased to OpenGL's zero-based semantics.
+        if (nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         if (shaderSubgroupSupported && !MG_Config::Features.DisableSubgroup) {
             extensions.push_back(E_GL_KHR_shader_subgroup);
         }
@@ -690,7 +699,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // the whole list keeps re-runs idempotent.
         m_rendererInfo.RendererGLInfo.Extensions = BuildAdvertisedExtensions(
             m_vulkanCaps.SupportsShaderSubgroup, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
-            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported());
+            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported(),
+            pVulkanRenderer && pVulkanRenderer->IsNonZeroIndirectBaseInstanceSupported());
     }
 
     void BackendObject_DirectVulkan::UpdateDynamicBackendParameters() {

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.h

@@ -62,8 +62,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // POST screen shows.
 
     // Static identity of the Magma renderer (renderer/backend names, target GL/GLSL
-    // versions, ExtraVendor) with the baseline extension advertisement (no shader
-    // subgroup, no timer queries). A live backend copies this in its constructor and
+    // versions, ExtraVendor) with the baseline extension advertisement (no runtime-gated
+    // capabilities). A live backend copies this in its constructor and
     // reconciles the Extensions in UpdateAdvertisedExtensions once real capabilities
     // exist; callers that need the advertised list for a known capability set must
     // use BuildAdvertisedExtensions instead.
@@ -74,7 +74,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // MOBILEGL_DISABLE_TIMERQUERY escape hatches are applied inside, so callers pass
     // the detected device support (passing an already-gated value is harmless).
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported);
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <GPU Name>, Vulkan <Vulkan Version>, Driver <Driver Version> — the exact
     // string an initialized backend returns from GetBackendAPIVersionString (and that

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -194,53 +194,54 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     PipelineFactory::HashType PipelineFactory::ComputeHash(const PipelineCreatePayload& payload) const {
-        XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.programHash, sizeof(payload.programHash)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.vertexInputHash, sizeof(payload.vertexInputHash)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.pipelineLayout, sizeof(payload.pipelineLayout)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.renderPass, sizeof(payload.renderPass)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.colorAttachmentCount, sizeof(payload.colorAttachmentCount)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.rasterizationSamples, sizeof(payload.rasterizationSamples)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.subpass, sizeof(payload.subpass)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.topology, sizeof(payload.topology)));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), m_config.CacheVersion));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.programHash, sizeof(payload.programHash)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.vertexInputHash, sizeof(payload.vertexInputHash)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.pipelineLayout, sizeof(payload.pipelineLayout)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.renderPass, sizeof(payload.renderPass)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.colorAttachmentCount, sizeof(payload.colorAttachmentCount)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.rasterizationSamples, sizeof(payload.rasterizationSamples)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.subpass, sizeof(payload.subpass)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.topology, sizeof(payload.topology)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.primitiveRestartEnable, sizeof(payload.primitiveRestartEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.patchControlPoints, sizeof(payload.patchControlPoints)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.polygonMode, sizeof(payload.polygonMode)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.cullMode, sizeof(payload.cullMode)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontFace, sizeof(payload.frontFace)));
+            XXH64_update(m_hashState.Get(), &payload.primitiveRestartEnable, sizeof(payload.primitiveRestartEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.patchControlPoints, sizeof(payload.patchControlPoints)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.viewportCount, sizeof(payload.viewportCount)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.polygonMode, sizeof(payload.polygonMode)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.cullMode, sizeof(payload.cullMode)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.frontFace, sizeof(payload.frontFace)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.provokingVertexMode, sizeof(payload.provokingVertexMode)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.depthTestEnable, sizeof(payload.depthTestEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.depthWriteEnable, sizeof(payload.depthWriteEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.depthBiasEnable, sizeof(payload.depthBiasEnable)));
+            XXH64_update(m_hashState.Get(), &payload.provokingVertexMode, sizeof(payload.provokingVertexMode)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.depthTestEnable, sizeof(payload.depthTestEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.depthWriteEnable, sizeof(payload.depthWriteEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.depthBiasEnable, sizeof(payload.depthBiasEnable)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.rasterizerDiscardEnable, sizeof(payload.rasterizerDiscardEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.logicOpEnable, sizeof(payload.logicOpEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.stencilTestEnable, sizeof(payload.stencilTestEnable)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.depthCompareOp, sizeof(payload.depthCompareOp)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.logicOp, sizeof(payload.logicOp)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontStencilFailOp, sizeof(payload.frontStencilFailOp)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontStencilPassOp, sizeof(payload.frontStencilPassOp)));
+            XXH64_update(m_hashState.Get(), &payload.rasterizerDiscardEnable, sizeof(payload.rasterizerDiscardEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.logicOpEnable, sizeof(payload.logicOpEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.stencilTestEnable, sizeof(payload.stencilTestEnable)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.depthCompareOp, sizeof(payload.depthCompareOp)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.logicOp, sizeof(payload.logicOp)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.frontStencilFailOp, sizeof(payload.frontStencilFailOp)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.frontStencilPassOp, sizeof(payload.frontStencilPassOp)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.frontStencilDepthFailOp, sizeof(payload.frontStencilDepthFailOp)));
+            XXH64_update(m_hashState.Get(), &payload.frontStencilDepthFailOp, sizeof(payload.frontStencilDepthFailOp)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.frontStencilCompareOp, sizeof(payload.frontStencilCompareOp)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.backStencilFailOp, sizeof(payload.backStencilFailOp)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.backStencilPassOp, sizeof(payload.backStencilPassOp)));
+            XXH64_update(m_hashState.Get(), &payload.frontStencilCompareOp, sizeof(payload.frontStencilCompareOp)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.backStencilFailOp, sizeof(payload.backStencilFailOp)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &payload.backStencilPassOp, sizeof(payload.backStencilPassOp)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.backStencilDepthFailOp, sizeof(payload.backStencilDepthFailOp)));
+            XXH64_update(m_hashState.Get(), &payload.backStencilDepthFailOp, sizeof(payload.backStencilDepthFailOp)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.backStencilCompareOp, sizeof(payload.backStencilCompareOp)));
+            XXH64_update(m_hashState.Get(), &payload.backStencilCompareOp, sizeof(payload.backStencilCompareOp)));
         XXHASH_VERIFY(
-            XXH64_update(m_hashState, &payload.fragmentReplacesDepth, sizeof(payload.fragmentReplacesDepth)));
+            XXH64_update(m_hashState.Get(), &payload.fragmentReplacesDepth, sizeof(payload.fragmentReplacesDepth)));
         if (payload.colorAttachmentCount > 0) {
             XXHASH_VERIFY(XXH64_update(
-                m_hashState,
+                m_hashState.Get(),
                 payload.colorBlendAttachments.data(),
                 sizeof(payload.colorBlendAttachments[0]) * payload.colorAttachmentCount));
         }
-        return XXH64_digest(m_hashState);
+        return XXH64_digest(m_hashState.Get());
     }
 
     VkPipeline PipelineFactory::GetOrCreatePipeline(const PipelineCreatePayload& payload) {
@@ -406,8 +407,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         tessellation.patchControlPoints = payload.patchControlPoints;
 
         VkPipelineViewportStateCreateInfo vpci{VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO};
-        vpci.viewportCount = 1;
-        vpci.scissorCount = 1;
+        // Both counts move together: GL has one scissor rectangle per viewport, and Vulkan
+        // requires viewportCount == scissorCount whenever both are dynamic
+        // (VUID-VkPipelineViewportStateCreateInfo-scissorCount-04136). The caller has already
+        // clamped this to the device's multiViewport capability.
+        vpci.viewportCount = std::max<Uint32>(payload.viewportCount, 1u);
+        vpci.scissorCount = vpci.viewportCount;
 
         VkPipelineRasterizationStateCreateInfo raster{VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO};
         raster.polygonMode = payload.polygonMode;

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -12,6 +12,7 @@
 #include "../VkIncludes.h"
 #include "MG_State/GLState/FramebufferState/FramebufferObject.h"
 #include <Includes.h>
+#include <MG_Util/Types.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
     // Enough of a fingerprint to identify the exact module the driver rejected without keeping the
@@ -42,6 +43,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool primitiveRestartEnable = false;
             // GL_PATCH_VERTICES; only read for a PATCH_LIST topology.
             Uint32 patchControlPoints = 3;
+            // How many of ARB_viewport_array's viewports this pipeline rasterizes into. 1 for
+            // every program that never assigns gl_ViewportIndex, which is all of them outside the
+            // conformance suite - the wide shape costs a longer vkCmdSetViewport/Scissor per state
+            // change and can cost hardware fast paths, so it is opt-in per program. Baked into the
+            // pipeline (viewportCount is not dynamic without VK_EXT_extended_dynamic_state) and
+            // therefore hashed; the DYNAMIC viewport/scissor arrays the draw pushes must have
+            // exactly this many elements (VUID-vkCmdDraw-viewportCount-03417/-03418).
+            Uint32 viewportCount = 1;
             VkPolygonMode polygonMode = VK_POLYGON_MODE_FILL;
             VkCullModeFlags cullMode = VK_CULL_MODE_BACK_BIT;
             VkFrontFace frontFace = VK_FRONT_FACE_CLOCKWISE;
@@ -157,7 +166,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         UnorderedMap<HashType, PipelineCacheEntry> m_cache;
         // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
         Uint64 m_frameCounter = 0;
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline MobileGL::XXH64State m_hashState;
         static inline Bool s_suppressBlendedDepthWrite = false;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -1997,6 +1997,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ReflectedDeclaresInputBuiltin(reflectModule, SpvBuiltInBaseVertex);
     }
 
+    // gl_ViewportIndex on the last pre-rasterization stage. glslang emits it natively for Vulkan
+    // (BuiltIn ViewportIndex plus OpCapability MultiViewport), and nothing in the SpirvPasses
+    // chain touches it, so a plain reflection of the declared output builtins is the whole test.
+    Bool ProgramFactory::ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule) {
+        return ReflectedDeclaresOutputBuiltin(reflectModule, SpvBuiltInViewportIndex);
+    }
+
+    Bool ProgramFactory::ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule,
+                                                       SpvBuiltIn builtin) {
+        for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
+            const SpvReflectEntryPoint& entryPoint = reflectModule.entry_points[entryIndex];
+            for (Uint32 variableIndex = 0; variableIndex < entryPoint.output_variable_count; ++variableIndex) {
+                const SpvReflectInterfaceVariable* variable = entryPoint.output_variables[variableIndex];
+                if (variable != nullptr &&
+                    (variable->decoration_flags & SPV_REFLECT_DECORATION_BUILT_IN) != 0 &&
+                    variable->built_in == builtin) {
+                    return true;
+                }
+            }
+        }
+        return false;
+    }
+
     Bool ProgramFactory::ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule,
                                                        SpvBuiltIn builtin) {
         for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
@@ -2133,26 +2156,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     ProgramFactory::HashType ProgramFactory::ComputeHash(const MG_State::GLState::ProgramObject& program,
                                                          CompileOptionFlags flags) const {
-        XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), m_config.CacheVersion));
         // We expect shader stages in program object are sorted
         const auto& spirvs = program.GetGeneratedSpirv();
         for (const auto& spv : spirvs) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, spv.data(), spv.size() * sizeof(Uint)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), spv.data(), spv.size() * sizeof(Uint)));
         }
-        XXHASH_VERIFY(XXH64_update(m_hashState, &flags, sizeof(CompileOptionFlags)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &flags, sizeof(CompileOptionFlags)));
         // Only FragCoordYFlip variants bake the height in, so mixing it unconditionally would
         // re-key every program in the cache on a resize for no reason.
         if (flags & CompileOptionBit::FragCoordYFlip) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, &m_defaultFramebufferHeight,
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &m_defaultFramebufferHeight,
                                         sizeof(m_defaultFramebufferHeight)));
         }
 
         // Include UBO block bindings in hash so different binding configurations produce different entries
         const Uint32 blockCount = static_cast<Uint32>(program.GetActiveUniformBlocksCount());
-        XXHASH_VERIFY(XXH64_update(m_hashState, &blockCount, sizeof(blockCount)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &blockCount, sizeof(blockCount)));
         for (Uint32 i = 0; i < blockCount; ++i) {
             const Uint32 binding = program.GetUniformBlockBinding(i);
-            XXHASH_VERIFY(XXH64_update(m_hashState, &binding, sizeof(binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &binding, sizeof(binding)));
         }
 
         // The transform feedback capture layout is baked into the modules by
@@ -2163,18 +2186,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // hashed for a capturing compile, so nothing else changes key.
         if (flags & CompileOptionBit::XfbCapture) {
             for (const auto& varying : program.GetTransformFeedbackVaryings()) {
-                XXHASH_VERIFY(XXH64_update(m_hashState, varying.name.data(), varying.name.size()));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &varying.bufferIndex, sizeof(varying.bufferIndex)));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &varying.offsetBytes, sizeof(varying.offsetBytes)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), varying.name.data(), varying.name.size()));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &varying.bufferIndex, sizeof(varying.bufferIndex)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &varying.offsetBytes, sizeof(varying.offsetBytes)));
             }
             const SizeT bufferCount = program.GetTransformFeedbackBufferCount();
             for (SizeT i = 0; i < bufferCount; ++i) {
                 const Uint32 stride = program.GetTransformFeedbackStride(static_cast<Uint32>(i));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &stride, sizeof(stride)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &stride, sizeof(stride)));
             }
         }
 
-        HashType hash = XXH64_digest(m_hashState);
+        HashType hash = XXH64_digest(m_hashState.Get());
         return hash;
     }
 
@@ -2339,6 +2362,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    // Which pre-rasterization stage assigns gl_ViewportIndex is not fixed: GL 4.1 allows only the
+    // geometry stage, ARB_shader_viewport_layer_array/GL 4.6 also the vertex and tessellation
+    // evaluation stages. Rather than guess which one is last, every non-fragment, non-compute
+    // module is asked - one writer anywhere means this program's draws need a multi-viewport
+    // pipeline, and a false positive costs only a wider viewportCount.
+    void ProgramFactory::ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                                   const Vector<Vector<Uint>>& spirv,
+                                                   VkProgramObject& entry) const {
+        entry.writesViewportIndexBuiltin = false;
+
+        for (SizeT moduleIndex = 0; moduleIndex < shaders.size() && moduleIndex < spirv.size(); ++moduleIndex) {
+            if (!shaders[moduleIndex]) continue;
+            const ShaderStage stage = shaders[moduleIndex]->GetShaderStage();
+            if (stage == ShaderStage::Fragment || stage == ShaderStage::Compute) continue;
+
+            const auto& module = spirv[moduleIndex];
+            if (module.empty()) continue;
+
+            SpvReflectShaderModule reflectModule{};
+            const SpvReflectResult createResult =
+                spvReflectCreateShaderModule(module.size() * sizeof(Uint), module.data(), &reflectModule);
+            if (createResult != SPV_REFLECT_RESULT_SUCCESS) {
+                // Fail toward the wide pipeline. Missing a real gl_ViewportIndex writer would
+                // silently collapse every viewport onto 0 (the exact bug this reflection exists
+                // to fix); over-declaring costs one extra viewport slot on a program that never
+                // uses it.
+                MGLOG_E_ONCE("ProgramFactory::ReflectViewportIndexUsage: reflection failed (result=%d); assuming the "
+                             "program writes gl_ViewportIndex",
+                             static_cast<Int>(createResult));
+                entry.writesViewportIndexBuiltin = true;
+                continue;
+            }
+
+            if (ReflectedWritesViewportIndexBuiltin(reflectModule)) {
+                entry.writesViewportIndexBuiltin = true;
+            }
+            spvReflectDestroyShaderModule(&reflectModule);
+        }
+    }
+
     void ProgramFactory::ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                                                 const Vector<Vector<Uint>>& spirv,
                                                 VkProgramObject& entry) const {
@@ -3189,6 +3252,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ValidateRasterizationStageInterface(shaders, moduleSpirvs, entry, program.GetExternalIndex());
 #endif
         ReflectVertexInputs(shaders, moduleSpirvs, entry);
+        ReflectViewportIndexUsage(shaders, moduleSpirvs, entry);
         ReflectFragmentOutputs(shaders, moduleSpirvs, entry);
         ReflectPassthroughTessControlNeed(shaders, moduleSpirvs, entry);
         ReflectLayout(program, moduleSpirvs, entry);

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -15,6 +15,7 @@
 #include "MG_State/GLState/TextureState/TextureEnum.h"
 
 #include <Includes.h>
+#include <MG_Util/Types.h>
 #include <spirv_reflect.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -151,6 +152,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // PROGRAM rather than of the variant: the zeroed variant leaves the variable
             // declared, so both variants answer the same and the draw path can ask either.
             Bool readsBaseVertexBuiltin = false;
+            // Some pre-rasterization stage assigns gl_ViewportIndex. Its pipeline declares
+            // viewportCount = the renderer's rasterizable viewport count instead of 1, and its
+            // draws push the whole viewport/scissor array; every other program keeps the
+            // single-viewport fast path untouched. Part of the program's identity (folded into
+            // the pipeline hash through programHash), so no memo can serve the wrong shape.
+            Bool writesViewportIndexBuiltin = false;
             // This program has a tessellation EVALUATION stage and no tessellation CONTROL
             // stage. GL allows that (4.6 core 11.2.2: with no control shader the input patch
             // is passed through unmodified, the output patch size is PATCH_VERTICES, and the
@@ -218,6 +225,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
                 needsPassthroughTessControl = other.needsPassthroughTessControl;
                 passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
@@ -234,6 +242,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
                 other.needsPassthroughTessControl = false;
                 other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
@@ -276,6 +285,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
                 needsPassthroughTessControl = other.needsPassthroughTessControl;
                 passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
@@ -292,6 +302,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
                 other.needsPassthroughTessControl = false;
                 other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
@@ -400,6 +411,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Shared by the two above: does any entry point list an input variable decorated with
         // this builtin?
         static Bool ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+        // True when an entry point writes the ViewportIndex builtin (gl_ViewportIndex), i.e. when
+        // the program can route primitives to a viewport other than 0 and its pipeline therefore
+        // has to declare more than one. Asks about OUTPUT variables because that is the direction
+        // a pre-rasterization stage declares it in.
+        static Bool ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule);
+        static Bool ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
 
         // The pass-through tessellation control stage GL 4.6 core 11.2.2 describes for a
         // program that has an evaluation stage and no control stage, for an input patch of
@@ -434,6 +451,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void ReflectVertexInputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                      const Vector<Vector<Uint>>& spirv,
                      VkProgramObject& entry) const;
+        void ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                       const Vector<Vector<Uint>>& spirv,
+                                       VkProgramObject& entry) const;
         void ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                         const Vector<Vector<Uint>>& spirv,
                         VkProgramObject& entry) const;
@@ -470,6 +490,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // ever built from one keeps referencing its module. A failed build is cached as
         // VK_NULL_HANDLE so a broken generator costs one compile, not one per draw.
         UnorderedMap<Uint32, VkPipelineShaderStageCreateInfo> m_passthroughTessControlStages;
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline MobileGL::XXH64State m_hashState;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -305,7 +305,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // texture/sampler resolution, completeness probe, sync, layout handling, sampler
         // and view lookups - would recompute the identical descriptor.
         if (trustUnchangedHint && descriptorMemoUsable && binding < m_samplerResolveMemo.size() &&
-            m_samplerResolveMemo[binding].infoValid) {
+            m_samplerResolveMemo[binding].infoValid &&
+            m_samplerResolveMemo[binding].infoProgramLifetimeId == program.GetLifetimeId()) {
             outImageInfo = m_samplerResolveMemo[binding].info;
             return true;
         }
@@ -504,6 +505,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (binding < m_samplerResolveMemo.size()) {
             if (descriptorMemoUsable) {
                 m_samplerResolveMemo[binding].info = outImageInfo;
+                m_samplerResolveMemo[binding].infoProgramLifetimeId = program.GetLifetimeId();
                 m_samplerResolveMemo[binding].infoValid = true;
             } else {
                 // An arrayed binding publishes nothing here, and clears what a previous program

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -341,8 +341,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // lifetime id, so a freed-and-reallocated sampler or texture at the same heap address
         // always gets a fresh id and misses (a raw pointer would false-hit that ABA) - so a
         // stale guess can only miss and fall through to the hash, never resolve wrong. Still
-        // reset each frame alongside the descriptor-set cache. Indexed by binding.
+        // reset each frame alongside the descriptor-set cache. Indexed by binding, but the
+        // whole-descriptor entry is additionally keyed by program lifetime: Vulkan binding
+        // numbers are layout-local and unrelated programs routinely reuse binding 0/1.
         struct SamplerResolveMemo {
+            Uint64 infoProgramLifetimeId = 0;
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -13,25 +13,25 @@
 namespace MobileGL::MG_Backend::DirectVulkan {
     VertexInputStateFactory::HashType VertexInputStateFactory::ComputeHash(
         const MG_State::GLState::VertexArrayObject& vao) const {
-        XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), m_config.CacheVersion));
 
         for (Int i = 0; i < MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS; ++i) {
             const auto& attr = vao.GetAttribute(i);
 
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Enabled, sizeof(attr.Enabled)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Enabled, sizeof(attr.Enabled)));
             if (!attr.Enabled) {
                 continue;
             }
 
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Size, sizeof(attr.Size)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Type, sizeof(attr.Type)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Normalized, sizeof(attr.Normalized)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Stride, sizeof(attr.Stride)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Offset, sizeof(attr.Offset)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsInteger, sizeof(attr.IsInteger)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsLong, sizeof(attr.IsLong)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsBgra, sizeof(attr.IsBgra)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Divisor, sizeof(attr.Divisor)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Size, sizeof(attr.Size)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Type, sizeof(attr.Type)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Normalized, sizeof(attr.Normalized)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Stride, sizeof(attr.Stride)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Offset, sizeof(attr.Offset)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.IsInteger, sizeof(attr.IsInteger)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.IsLong, sizeof(attr.IsLong)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.IsBgra, sizeof(attr.IsBgra)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attr.Divisor, sizeof(attr.Divisor)));
 
             // The bound buffer's IDENTITY is a component of the key, and it has to be the
             // buffer's never-reused lifetime id - NOT its heap address, which this used to
@@ -45,10 +45,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // test's positions) instead of its own.
             // Zero for client memory (no buffer), which is a distinct identity of its own.
             const Uint64 bufferKey = attr.Buffer ? attr.Buffer->GetLifetimeId() : 0;
-            XXHASH_VERIFY(XXH64_update(m_hashState, &bufferKey, sizeof(bufferKey)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &bufferKey, sizeof(bufferKey)));
         }
 
-        return XXH64_digest(m_hashState);
+        return XXH64_digest(m_hashState.Get());
     }
 
     VertexInputStateFactory::HashType VertexInputStateFactory::GetOrComputeHash(
@@ -225,24 +225,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.attributes = builder.GetAttributes();
         // See the layoutHash declaration: hash only the resolved layout, never
         // buffer identities, so identical layouts across VAOs/buffers agree.
-        XXHASH_VERIFY(XXH64_reset(m_hashState, 0));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), 0));
         for (const auto& binding : entry.bindings) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.binding, sizeof(binding.binding)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.stride, sizeof(binding.stride)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.inputRate, sizeof(binding.inputRate)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &binding.binding, sizeof(binding.binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &binding.stride, sizeof(binding.stride)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &binding.inputRate, sizeof(binding.inputRate)));
         }
         for (const auto& attribute : entry.attributes) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.location, sizeof(attribute.location)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.binding, sizeof(attribute.binding)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.format, sizeof(attribute.format)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.offset, sizeof(attribute.offset)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attribute.location, sizeof(attribute.location)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attribute.binding, sizeof(attribute.binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attribute.format, sizeof(attribute.format)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &attribute.offset, sizeof(attribute.offset)));
         }
         for (const auto& divisor : entry.bindingDivisors) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, &divisor.binding, sizeof(divisor.binding)));
-            XXHASH_VERIFY(XXH64_update(m_hashState, &divisor.divisor, sizeof(divisor.divisor)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &divisor.binding, sizeof(divisor.binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &divisor.divisor, sizeof(divisor.divisor)));
         }
-        XXHASH_VERIFY(XXH64_update(m_hashState, &unsupportedAttribMask, sizeof(unsupportedAttribMask)));
-        entry.layoutHash = XXH64_digest(m_hashState);
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &unsupportedAttribMask, sizeof(unsupportedAttribMask)));
+        entry.layoutHash = XXH64_digest(m_hashState.Get());
         entry.attributeLocationMask = 0;
         for (const auto& attribute : entry.attributes) {
             if (attribute.location < 32u) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -12,6 +12,7 @@
 #include "VertexInputStateBuilder.h"
 #include "MG_State/GLState/VertexArrayState/VertexArrayObject.h"
 #include <Includes.h>
+#include <MG_Util/Types.h>
 #include "../VkIncludes.h"
 
 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -126,6 +127,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // matches, so an evicted entry can never be dereferenced through a
         // stale memo.
         Uint64 m_evictionEpoch = 1;
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline MobileGL::XXH64State m_hashState;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -594,27 +594,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     VkRenderPassManager::HashType VkRenderPassManager::ComputeHash(
         const MG_State::GLState::FramebufferObject& fbo, Uint32 swapchainImageIndex, Bool includePendingClear,
         Bool includeDefaultFboDepthStencil) {
-        XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), m_config.CacheVersion));
         const Bool isDefaultFbo = fbo.IsDefaultFramebuffer();
         if (isDefaultFbo) {
-            XXHASH_VERIFY(XXH64_update(m_hashState, &swapchainImageIndex, sizeof(swapchainImageIndex)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &swapchainImageIndex, sizeof(swapchainImageIndex)));
         }
         // sRGB attachments switch between their sRGB and UNORM-twin views with this
         // capability (ResolveSrgbAttachmentWriteFormat), changing the render pass formats.
         const Bool framebufferSrgbEnabled =
             MG_State::pGLContext->IsCapabilityEnabled(MobileGL::CapabilityInput::FramebufferSrgb);
-        XXHASH_VERIFY(XXH64_update(m_hashState, &framebufferSrgbEnabled, sizeof(framebufferSrgbEnabled)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &framebufferSrgbEnabled, sizeof(framebufferSrgbEnabled)));
         auto& drawBuffers = fbo.GetDrawBuffers();
-        XXHASH_VERIFY(XXH64_update(m_hashState, drawBuffers.data(), drawBuffers.size() * sizeof(drawBuffers[0])));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), drawBuffers.data(), drawBuffers.size() * sizeof(drawBuffers[0])));
         auto readBuffer = fbo.GetReadBuffer();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &readBuffer, sizeof(FramebufferAttachmentType)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &readBuffer, sizeof(FramebufferAttachmentType)));
         Int validDrawBufCount = 0;
         for (Int i = 0; i < drawBuffers.size(); ++i) {
             auto drawbuf = drawBuffers[i];
             if (drawbuf != FramebufferAttachmentType::None)
                 validDrawBufCount = std::max(validDrawBufCount, i + 1);
         }
-        XXHASH_VERIFY(XXH64_update(m_hashState, &validDrawBufCount, sizeof(validDrawBufCount)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &validDrawBufCount, sizeof(validDrawBufCount)));
 
         auto combineFramebufferAttachmentObjHash = [&](FramebufferAttachmentType attachment) {
             auto& att = fbo.GetAttachment(attachment);
@@ -623,49 +623,49 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (att.IsEmpty()) type = 0;
             else if (att.IsTexture()) type = 1;
             else if (att.IsRenderbuffer()) type = 2;
-            XXHASH_VERIFY(XXH64_update(m_hashState, &type, sizeof(type)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &type, sizeof(type)));
             void* contentPtr = nullptr;
             if (att.IsTexture())
                 contentPtr = att.GetTexture().get();
             else if (att.IsRenderbuffer())
                 contentPtr = att.GetRenderbuffer().get();
-            XXHASH_VERIFY(XXH64_update(m_hashState, &contentPtr, sizeof(contentPtr)));
+            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &contentPtr, sizeof(contentPtr)));
             if (att.IsTexture()) {
                 const Uint64 textureLifetimeId = att.GetTexture()->GetLifetimeId();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &textureLifetimeId, sizeof(textureLifetimeId)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &textureLifetimeId, sizeof(textureLifetimeId)));
                 const Int textureLevel = att.GetTextureLevel();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &textureLevel, sizeof(textureLevel)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &textureLevel, sizeof(textureLevel)));
                 const TextureUploadTarget textureUploadTarget = att.GetTextureUploadTarget();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &textureUploadTarget, sizeof(textureUploadTarget)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &textureUploadTarget, sizeof(textureUploadTarget)));
                 const Int textureLayer = att.GetTextureLayer();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &textureLayer, sizeof(textureLayer)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &textureLayer, sizeof(textureLayer)));
                 const Bool textureLayered = att.IsLayered();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &textureLayered, sizeof(textureLayered)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &textureLayered, sizeof(textureLayered)));
 
                 Uint64 imageIdentity = 0;
                 auto* texture = att.GetTexture().get();
                 auto* resource = m_textureManager.SyncTextureAndGetDescriptor(*texture);
                 if (resource != nullptr) {
                     imageIdentity = reinterpret_cast<Uint64>(resource->image);
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &resource->sampleCount, sizeof(resource->sampleCount)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &resource->sampleCount, sizeof(resource->sampleCount)));
                 } else {
                     const VkSampleCountFlagBits fallbackSampleCount = VK_SAMPLE_COUNT_1_BIT;
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &fallbackSampleCount, sizeof(fallbackSampleCount)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &fallbackSampleCount, sizeof(fallbackSampleCount)));
                 }
-                XXHASH_VERIFY(XXH64_update(m_hashState, &imageIdentity, sizeof(imageIdentity)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &imageIdentity, sizeof(imageIdentity)));
             }
 
             if (includePendingClear && att.IsTexture()) {
                 auto* texture = att.GetTexture().get();
                 const auto pendingClearKey = VkClearManager::MakePendingClearKey(att);
                 auto hasClear = m_clearManager.HasPendingClear(pendingClearKey);
-                XXHASH_VERIFY(XXH64_update(m_hashState, &hasClear, sizeof(hasClear)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &hasClear, sizeof(hasClear)));
                 if (hasClear) {
                     ClearAttachmentPayload clearPayload{};
                     Bool hasPayload = m_clearManager.GetPendingClear(pendingClearKey, clearPayload);
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &hasPayload, sizeof(hasPayload)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &hasPayload, sizeof(hasPayload)));
                     if (hasPayload) {
-                        XXHASH_VERIFY(XXH64_update(m_hashState, &clearPayload.mask, sizeof(clearPayload.mask)));
+                        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &clearPayload.mask, sizeof(clearPayload.mask)));
                     }
                 }
 
@@ -695,7 +695,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         currentLayout = textureResource->layout;
                     }
                 }
-                XXHASH_VERIFY(XXH64_update(m_hashState, &currentLayout, sizeof(currentLayout)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &currentLayout, sizeof(currentLayout)));
             }
             if (att.IsRenderbuffer() && att.GetRenderbuffer()) {
                 const auto& renderbuffer = att.GetRenderbuffer();
@@ -703,10 +703,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Int width = renderbuffer->GetWidth();
                 const Int height = renderbuffer->GetHeight();
                 const Int samples = renderbuffer->GetSamples();
-                XXHASH_VERIFY(XXH64_update(m_hashState, &internalFormat, sizeof(internalFormat)));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &width, sizeof(width)));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &height, sizeof(height)));
-                XXHASH_VERIFY(XXH64_update(m_hashState, &samples, sizeof(samples)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &internalFormat, sizeof(internalFormat)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &width, sizeof(width)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &height, sizeof(height)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &samples, sizeof(samples)));
 
                 Uint64 imageIdentity = 0;
                 VkImageLayout currentLayout = VK_IMAGE_LAYOUT_UNDEFINED;
@@ -714,25 +714,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 if (resource != nullptr) {
                     imageIdentity = reinterpret_cast<Uint64>(resource->image);
                     currentLayout = resource->layout;
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &resource->sampleCount, sizeof(resource->sampleCount)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &resource->sampleCount, sizeof(resource->sampleCount)));
                 } else {
                     const VkSampleCountFlagBits fallbackSampleCount = VK_SAMPLE_COUNT_1_BIT;
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &fallbackSampleCount, sizeof(fallbackSampleCount)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &fallbackSampleCount, sizeof(fallbackSampleCount)));
                 }
-                XXHASH_VERIFY(XXH64_update(m_hashState, &imageIdentity, sizeof(imageIdentity)));
+                XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &imageIdentity, sizeof(imageIdentity)));
 
                 if (includePendingClear) {
                     const Bool hasClear = HasPendingRenderbufferClear(att);
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &hasClear, sizeof(hasClear)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &hasClear, sizeof(hasClear)));
                     if (hasClear) {
                         ClearAttachmentPayload clearPayload{};
                         const Bool hasPayload = GetPendingRenderbufferClear(renderbuffer.get(), clearPayload);
-                        XXHASH_VERIFY(XXH64_update(m_hashState, &hasPayload, sizeof(hasPayload)));
+                        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &hasPayload, sizeof(hasPayload)));
                         if (hasPayload) {
-                            XXHASH_VERIFY(XXH64_update(m_hashState, &clearPayload.mask, sizeof(clearPayload.mask)));
+                            XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &clearPayload.mask, sizeof(clearPayload.mask)));
                         }
                     }
-                    XXHASH_VERIFY(XXH64_update(m_hashState, &currentLayout, sizeof(currentLayout)));
+                    XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &currentLayout, sizeof(currentLayout)));
                 }
             }
         };
@@ -745,13 +745,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The depth-less default-FBO flavor omits the depth/stencil attachment
         // entirely, so it must hash differently from the depth-full flavor.
         const Bool depthStencilIncluded = !isDefaultFbo || includeDefaultFboDepthStencil;
-        XXHASH_VERIFY(XXH64_update(m_hashState, &depthStencilIncluded, sizeof(depthStencilIncluded)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &depthStencilIncluded, sizeof(depthStencilIncluded)));
         if (depthStencilIncluded) {
             combineFramebufferAttachmentObjHash(FramebufferAttachmentType::Depth);
             combineFramebufferAttachmentObjHash(FramebufferAttachmentType::Stencil);
         }
 
-        return XXH64_digest(m_hashState);
+        return XXH64_digest(m_hashState.Get());
     }
 
     RenderPassEntry& VkRenderPassManager::GetOrCreateRenderPass(const MG_State::GLState::FramebufferObject& fbo,

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -16,6 +16,7 @@
 #include "MG_State/GLState/FramebufferState/FramebufferObject.h"
 
 #include <Includes.h>
+#include <MG_Util/Types.h>
 #include <unordered_map>
 #include <vk_mem_alloc.h>
 
@@ -391,7 +392,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void DeferRenderbufferBackingRelease(RenderbufferResource& resource);
         void CollectDeferredRenderbufferReleases(Bool destroyAll);
 
-        static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline MobileGL::XXH64State m_hashState;
         static inline ActiveRenderPassInfo s_activeRenderPass{};
         static inline Bool s_hasActiveRenderPass = false;
         static inline VkClearManager* s_clearManager = nullptr;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp

@@ -134,41 +134,41 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                              const MG_State::GLState::ITextureObject& texture,
                                              Bool forceNearestFiltering, Bool singleLevelView) const {
         MOBILEGL_ASSERT(m_config != nullptr, "VkSamplerManager::BuildSamplerKey: m_config is null");
-        XXHASH_VERIFY(XXH64_reset(m_hashState, m_config->CacheVersion));
+        XXHASH_VERIFY(XXH64_reset(m_hashState.Get(), m_config->CacheVersion));
 
-        XXHASH_VERIFY(XXH64_update(m_hashState, &forceNearestFiltering, sizeof(forceNearestFiltering)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &singleLevelView, sizeof(singleLevelView)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &forceNearestFiltering, sizeof(forceNearestFiltering)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &singleLevelView, sizeof(singleLevelView)));
 
         const auto minFilter = sampler.GetMinFilter();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &minFilter, sizeof(minFilter)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &minFilter, sizeof(minFilter)));
         const auto magFilter = sampler.GetMagFilter();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &magFilter, sizeof(magFilter)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &magFilter, sizeof(magFilter)));
         const auto mipmapMode = sampler.GetMipmapMode();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &mipmapMode, sizeof(mipmapMode)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &mipmapMode, sizeof(mipmapMode)));
         const auto wrapS = sampler.GetWrapS();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &wrapS, sizeof(wrapS)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &wrapS, sizeof(wrapS)));
         const auto wrapT = sampler.GetWrapT();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &wrapT, sizeof(wrapT)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &wrapT, sizeof(wrapT)));
         const auto wrapR = sampler.GetWrapR();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &wrapR, sizeof(wrapR)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &wrapR, sizeof(wrapR)));
         const auto maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         const auto minLod = ResolveEffectiveMinLod(sampler, maxLod);
-        XXHASH_VERIFY(XXH64_update(m_hashState, &minLod, sizeof(minLod)));
-        XXHASH_VERIFY(XXH64_update(m_hashState, &maxLod, sizeof(maxLod)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &minLod, sizeof(minLod)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &maxLod, sizeof(maxLod)));
         const auto lodBias = sampler.GetLodBias();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &lodBias, sizeof(lodBias)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &lodBias, sizeof(lodBias)));
         // The RESOLVED value, not the GL request: samplers that only differ in an anisotropy Vulkan
         // will not apply (NEAREST filtering, or requests past the device limit) must still share one
         // VkSampler, while two samplers that really do differ must not collide onto the first one's.
         const auto maxAnisotropy = ResolveEffectiveMaxAnisotropy(sampler, forceNearestFiltering);
-        XXHASH_VERIFY(XXH64_update(m_hashState, &maxAnisotropy, sizeof(maxAnisotropy)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &maxAnisotropy, sizeof(maxAnisotropy)));
         const auto compareMode = sampler.GetCompareMode();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &compareMode, sizeof(compareMode)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &compareMode, sizeof(compareMode)));
         const auto compareFunc = sampler.GetSamplerCompareFunc();
-        XXHASH_VERIFY(XXH64_update(m_hashState, &compareFunc, sizeof(compareFunc)));
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &compareFunc, sizeof(compareFunc)));
         const auto borderColor = ResolveVkBorderColor(sampler, texture);
-        XXHASH_VERIFY(XXH64_update(m_hashState, &borderColor, sizeof(borderColor)));
-        return XXH64_digest(m_hashState);
+        XXHASH_VERIFY(XXH64_update(m_hashState.Get(), &borderColor, sizeof(borderColor)));
+        return XXH64_digest(m_hashState.Get());
     }
 
     VkSampler VkSamplerManager::GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.h

@@ -11,6 +11,7 @@
 #include "../VkIncludes.h"
 #include "../VulkanRendererConfig.h"
 #include <Includes.h>
+#include <MG_Util/Types.h>
 #include <MG_State/GLState/SamplerState/SamplerObject.h>
 
 namespace MobileGL::MG_State::GLState {
@@ -85,6 +86,6 @@ private:
     UnorderedMap<Uint64, SamplerCacheEntry> m_samplers;
     // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
     Uint64 m_frameBoundaryCounter = 0;
-    static inline XXH64_state_t* m_hashState = XXH64_createState();
+    static inline MobileGL::XXH64State m_hashState;
 };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -300,8 +300,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels,
-            newResource.arrayLayers);
+            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare destination image");
 
         VkImageLayout srcTrackedLayout = oldResource.layout;
@@ -311,8 +310,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, oldResource.image, srcTrackedLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels,
-            oldResource.arrayLayers);
+            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare source image");
 
         Vector<VkImageCopy> copyRegions;
@@ -344,8 +342,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, oldResource.layout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask,
-            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels,
-            newResource.arrayLayers);
+            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to restore destination layout");
 
         VK_VERIFY(vkEndCommandBuffer(commandBuffer), "vkEndCommandBuffer(texture preserve)");
@@ -1191,7 +1188,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool lowerTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, lowerMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, 0, writtenMipLevel, resource.arrayLayers);
+                    resource.aspect, 0, writtenMipLevel);
                 MOBILEGL_ASSERT(lowerTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition lower mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1203,8 +1200,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool upperTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, upperMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel,
-                    resource.arrayLayers);
+                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel);
                 MOBILEGL_ASSERT(upperTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition upper mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1257,8 +1253,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const Bool ok = TransitionImageLayout(commandBuffer, resource->image, resource->layout, targetLayout, srcStageMask,
                                               s_sampledReadStages, srcAccessMask,
-                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
-                                              resource->arrayLayers);
+                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
         StampResourceRecordingUse(*resource);
@@ -1288,7 +1283,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VK_IMAGE_LAYOUT_GENERAL, srcStageMask,
                                               VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, srcAccessMask,
                                               VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
-                                              resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                                              resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                         texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
@@ -1355,8 +1350,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
                                                  VkAccessFlags srcAccessMask, VkAccessFlags dstAccessMask,
-                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel, Uint32 levelCount,
-                                                 Uint32 layerCount) {
+                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel,
+                                                 Uint32 levelCount) {
         MOBILEGL_ASSERT(image != VK_NULL_HANDLE, "TransitionImageLayout: m_image == VK_NULL_HANDLE");
         MOBILEGL_ASSERT(!((dstAccessMask & VK_ACCESS_TRANSFER_READ_BIT) != 0 &&
                           (dstStageMask & VK_PIPELINE_STAGE_TRANSFER_BIT) == 0),
@@ -1381,7 +1376,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         barrier.subresourceRange.baseMipLevel = baseMipLevel;
         barrier.subresourceRange.levelCount = levelCount;
         barrier.subresourceRange.baseArrayLayer = 0;
-        barrier.subresourceRange.layerCount = layerCount;
+        // Every layer, always - see the declaration for why layout tracking leaves no other
+        // correct answer. VK_REMAINING_ARRAY_LAYERS rather than the image's own `arrayLayers`
+        // because those are not the same number for a 3D image: MobileGL creates 3D images
+        // 2D_ARRAY_COMPATIBLE and their arrayLayers is 1, which today Vulkan reads as "all depth
+        // slices" but will read as "depth slice 0" once VK_KHR_maintenance9 is enabled. The
+        // validation layer warns about that literal 1 by name.
+        barrier.subresourceRange.layerCount = VK_REMAINING_ARRAY_LAYERS;
         vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 
         trackedLayout = newLayout;
@@ -1993,7 +1994,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Bound the idle pool: a one-off giant upload (initial atlas define)
         // must not pin its staging memory forever.
         constexpr VkDeviceSize kMaxFreeUploadStagingBytes = 32u * 1024u * 1024u;
-        if (m_allocator == nullptr || m_freeUploadStagingBytes + block.capacity > kMaxFreeUploadStagingBytes) {
+        if (m_allocator == nullptr) {
+            // The normal shutdown path destroys the free list through
+            // DestroyUploadPools while the allocator is still valid, so this is a
+            // defensive backstop only. Never pass a null allocator to VMA.
+            MGLOG_W_ONCE("VkTextureManager::RecycleUploadStagingBlock called with a null allocator");
+            return;
+        }
+        if (m_freeUploadStagingBytes + block.capacity > kMaxFreeUploadStagingBytes) {
             vmaDestroyBuffer(m_allocator, block.buffer, block.allocation);
             return;
         }
@@ -2605,7 +2613,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                         VK_PIPELINE_STAGE_TRANSFER_BIT,
                                         uploadSrcAccessMask,
                                         VK_ACCESS_TRANSFER_WRITE_BIT,
-                                        aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                        aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL failed");
 
         // Array textures keep their GL "depth" in VkImage array layers, so the
@@ -2709,7 +2717,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                    s_sampledReadStages,
                                    VK_ACCESS_TRANSFER_WRITE_BIT,
                                    VK_ACCESS_SHADER_READ_BIT,
-                                   aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                   aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to sampled read-only layout failed");
         outResource.layout = finalLayout;
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -388,12 +388,24 @@ public:
     static Bool AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
     static Bool AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
 
+    // Moves `image` to `newLayout` and writes the new layout back through `trackedLayout`.
+    //
+    // The barrier covers EVERY array layer of the image, and there is deliberately no layer
+    // parameter to say otherwise: layout here is tracked per IMAGE (one `TextureResource::layout`,
+    // or one caller-owned variable), so a barrier narrower than the image would leave the layers it
+    // skipped in the old layout while the tracker claims they moved. Every transfer against a
+    // framebuffer attachment above layer 0 - glReadPixels, glBlitFramebuffer, glCopyTexSubImage,
+    // glCopyImageSubData - then ran its copy on a layer no barrier had transitioned.
+    //
+    // The mip range IS a parameter, because mip levels really are transitioned piecewise (see
+    // UpdateTrackedImageLayoutAfterAttachmentWrite and the mipmap generation loops): those callers
+    // move the complement of the level they wrote so the whole image converges on one layout again.
+    // Nothing does, or can, do that per layer.
     static Bool TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image, VkImageLayout& trackedLayout,
                                VkImageLayout newLayout, VkPipelineStageFlags srcStageMask,
                                VkPipelineStageFlags dstStageMask, VkAccessFlags srcAccessMask,
                                VkAccessFlags dstAccessMask, VkImageAspectFlags aspectMask,
-                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1,
-                               Uint32 layerCount = 1);
+                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1);
 
     SizeT CollectGarbage();
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -238,11 +238,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // nothing else across all of gl33.
     //
     // The mapping below is derived from - and at full extent exactly reproduces - the pixel
-    // mapping RemapDefaultFboReadbackToGLOrientation has always used:
+    // mapping VulkanRenderer::RemapDefaultFramebufferReadback uses:
     //     identity : image(x, H-1-y)      -> flip Y
     //     180      : image(W-1-x, y)      -> mirror X (the rotation already flips the rows)
-    // Quarter turns swap the axes; nothing in this renderer models that (the readback declines to
-    // remap them and the viewport path only rescales), so they are left exactly as they were.
+    // Quarter turns swap the axes and are handled by MapDefaultFramebufferReadbackRect rather than
+    // this same-axis helper.
     struct DefaultFramebufferRectMapping {
         Bool flipY = false;
         Bool mirrorX = false;
@@ -335,19 +335,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Complete input inventory of ApplyDynamicDrawStateTail, one line per reader
         // (each accessor it replaces is a verified plain field read of the same
         // RenderStateParameters field - RenderState.cpp):
-        //   ApplyGLViewportState    : Viewport, DepthRange, + extent/isDefaultFbo/preTransform
+        //   ApplyGLViewportState    : Viewports[0], DepthRanges[0], + extent/isDefaultFbo/preTransform
         //   ApplyBlendConstants     : BlendColor
         //   ApplyPolygonOffsetState : PolygonOffsetUnits, PolygonOffsetFactor
         //   ApplyLineWidthState     : LineWidth (see the caveat below)
         //   ApplyStencilState       : StencilStates[0..1].{ValueMask, WriteMask, Ref}
-        //   scissor rect            : ScissorTestEnabled, ScissorBox,
+        //   scissor rect            : ScissorTestEnabledMask bit 0, ScissorBoxes[0],
         //                             + extent/isDefaultFbo/preTransform
         // Caveat, unchanged from the version-only gate: ApplyLineWidthState also clamps
         // to the ACTIVE BACKEND OBJECT's aliased line-width range. Those are device
         // limits queried once at backend init and constant for the renderer's lifetime,
         // so they are not part of the key (the version gate never covered them either).
         struct DynamicTailKey {
-            Int viewport[4] = {0, 0, 0, 0};
+            Float viewport[4] = {0.0f, 0.0f, 0.0f, 0.0f};
             Float depthRange[2] = {0.0f, 0.0f};
             Float blendColor[4] = {0.0f, 0.0f, 0.0f, 0.0f};
             Float polygonOffsetFactor = 0.0f;
@@ -437,12 +437,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         vkCmdSetScissor(commandBuffer, 0, 1, &scissor);
     }
 
-    static void ApplyGLViewportState(VkCommandBuffer commandBuffer,
-                                     const IntVec2& framebufferExtent,
-                                     VkSurfaceTransformFlagBitsKHR preTransform,
-                                     Bool isDefaultFramebuffer) {
-        const IntVec4& viewportState = MG_State::pGLContext->GetViewport();
-        const FloatVec2& depthRange = MG_State::pGLContext->GetDepthRange();
+    // One viewport of the ARB_viewport_array state, mapped into Vulkan's frame. Split out of
+    // ApplyGLViewportState so the multi-viewport path derives index i through EXACTLY the same
+    // arithmetic as index 0 - the default-framebuffer Y-flip and pre-transform rotation
+    // especially, which is the classic way a multi-viewport port comes out upside down for every
+    // index but the one that was tested.
+    static VkViewport ComputeGLViewport(Uint32 index,
+                                        const IntVec2& framebufferExtent,
+                                        VkSurfaceTransformFlagBitsKHR preTransform,
+                                        Bool isDefaultFramebuffer) {
+        // Snapped to integers. The viewport is float STATE (glViewportIndexedf may set a
+        // fractional origin, and GetFloati_v hands it back verbatim), but what rasterizes here is
+        // the rounded rectangle - a deliberate, documented infidelity rather than a spec claim:
+        // MobileGL passes the driver's VIEWPORT_SUBPIXEL_BITS through, so it does advertise
+        // subpixel viewport precision it does not deliver. Nothing in KHR-GL43.viewport_array or
+        // in Minecraft sets a fractional viewport (the conformance checks are all on the state
+        // round trip), which is why the honest-but-lossy path was kept over widening every
+        // default-framebuffer Y-flip/pre-transform helper to floats. See the KNOWN INFIDELITY
+        // note in MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp.
+        const FloatVec4& stored = MG_State::pGLContext->GetViewportIndexed(index);
+        const IntVec4 viewportState(static_cast<Int>(std::lround(stored.x())),
+                                    static_cast<Int>(std::lround(stored.y())),
+                                    static_cast<Int>(std::lround(stored.z())),
+                                    static_cast<Int>(std::lround(stored.w())));
+        const FloatVec2& depthRange = MG_State::pGLContext->GetDepthRangeIndexed(index);
         const IntVec2 logicalExtent = isDefaultFramebuffer
             ? ResolveDefaultFramebufferLogicalExtent(preTransform, framebufferExtent)
             : framebufferExtent;
@@ -477,6 +495,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         viewport.height = static_cast<float>(viewportHeight);
         viewport.minDepth = depthRange.x();
         viewport.maxDepth = depthRange.y();
+        return viewport;
+    }
+
+    static void ApplyGLViewportState(VkCommandBuffer commandBuffer,
+                                     const IntVec2& framebufferExtent,
+                                     VkSurfaceTransformFlagBitsKHR preTransform,
+                                     Bool isDefaultFramebuffer) {
+        const VkViewport viewport = ComputeGLViewport(0, framebufferExtent, preTransform, isDefaultFramebuffer);
         auto& shadow = g_dynamicStateShadow;
         if (shadow.viewportValid && shadow.viewport.x == viewport.x && shadow.viewport.y == viewport.y &&
             shadow.viewport.width == viewport.width && shadow.viewport.height == viewport.height &&
@@ -2125,47 +2151,6 @@ void main() {
             return static_cast<Uint8>(value * 255.0f + 0.5f);
         }
 
-        // Re-order the copied BLOCK - not the whole image - from the default framebuffer's stored
-        // orientation into GL's. The caller has already aimed the copy at the right place with
-        // MapDefaultFramebufferRectAxis, so what arrives here is exactly the requested
-        // rectWidth x rectHeight rect, and all that is left is the order of rows (identity) or of
-        // columns (180) WITHIN it.
-        //
-        // This used to iterate the full swapchain extent and index both sides with that stride,
-        // which is why its caller could only use it on an exact full-extent read - and why every
-        // partial glReadPixels of the default framebuffer came back in Vulkan row order. Only
-        // identity/180 share the swapchain extent with the default framebuffer; 90/270 swap
-        // extents and are still declined.
-        static Bool RemapDefaultFboReadbackToGLOrientation(const Uint8* rawPixels,
-                                                            Uint32 rectWidth,
-                                                            Uint32 rectHeight,
-                                                            VkSurfaceTransformFlagBitsKHR preTransform,
-                                                            SizeT texelSize,
-                                                            Uint8* outPixels) {
-            if (IsQuarterTurnPreTransform(preTransform)) {
-                return false;
-            }
-            if (rectWidth == 0 || rectHeight == 0 || texelSize == 0) {
-                return false;
-            }
-            const DefaultFramebufferRectMapping mapping = GetDefaultFramebufferRectMapping(preTransform);
-            const SizeT rowBytes = static_cast<SizeT>(rectWidth) * texelSize;
-            for (Uint32 outY = 0; outY < rectHeight; ++outY) {
-                const Uint32 srcY = mapping.flipY ? (rectHeight - 1 - outY) : outY;
-                const Uint8* srcRow = rawPixels + static_cast<SizeT>(srcY) * rowBytes;
-                Uint8* dstRow = outPixels + static_cast<SizeT>(outY) * rowBytes;
-                if (!mapping.mirrorX) {
-                    Memcpy(dstRow, srcRow, rowBytes);
-                    continue;
-                }
-                for (Uint32 outX = 0; outX < rectWidth; ++outX) {
-                    Memcpy(dstRow + static_cast<SizeT>(outX) * texelSize,
-                           srcRow + static_cast<SizeT>(rectWidth - 1 - outX) * texelSize, texelSize);
-                }
-            }
-            return true;
-        }
-
         static SizeT AlignPixelRow(SizeT rowBytes, Int alignment) {
             const SizeT resolvedAlignment = static_cast<SizeT>(std::max(alignment, 1));
             return (rowBytes + resolvedAlignment - 1) & ~(resolvedAlignment - 1);
@@ -2712,6 +2697,95 @@ void main() {
         return formatInfo.texel_block_size;
     }
 
+    Bool VulkanRenderer::MapDefaultFramebufferReadbackRect(
+            GLint x, GLint y, GLsizei width, GLsizei height, VkExtent2D imageExtent,
+            VkSurfaceTransformFlagBitsKHR preTransform, VkOffset2D* imageOffset,
+            VkExtent2D* imageCopyExtent) {
+        if (width <= 0 || height <= 0 || imageOffset == nullptr || imageCopyExtent == nullptr) {
+            return false;
+        }
+
+        const Int imageWidth = static_cast<Int>(imageExtent.width);
+        const Int imageHeight = static_cast<Int>(imageExtent.height);
+        Int mappedX = x;
+        Int mappedY = y;
+        Uint32 mappedWidth = static_cast<Uint32>(width);
+        Uint32 mappedHeight = static_cast<Uint32>(height);
+
+        // InsertPositionFixup first flips GL Y and then applies the surface transform. In pixel
+        // coordinates that gives these half-open rectangle mappings into the stored image:
+        //   identity: (x, H-y-h), 90: (y, x), 180: (W-x-w, y), 270: (H-y-h, W-x-w).
+        // Quarter turns also transpose the copied block's extent.
+        switch (preTransform) {
+        case VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR:
+            mappedX = y;
+            mappedY = x;
+            mappedWidth = static_cast<Uint32>(height);
+            mappedHeight = static_cast<Uint32>(width);
+            break;
+        case VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR:
+            mappedX = imageWidth - x - width;
+            mappedY = y;
+            break;
+        case VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR:
+            mappedX = imageWidth - y - height;
+            mappedY = imageHeight - x - width;
+            mappedWidth = static_cast<Uint32>(height);
+            mappedHeight = static_cast<Uint32>(width);
+            break;
+        default:
+            mappedY = imageHeight - y - height;
+            break;
+        }
+
+        if (mappedX < 0 || mappedY < 0 || mappedWidth > imageExtent.width ||
+            mappedHeight > imageExtent.height ||
+            static_cast<Uint64>(mappedX) + mappedWidth > imageExtent.width ||
+            static_cast<Uint64>(mappedY) + mappedHeight > imageExtent.height) {
+            return false;
+        }
+        *imageOffset = {mappedX, mappedY};
+        *imageCopyExtent = {mappedWidth, mappedHeight};
+        return true;
+    }
+
+    Bool VulkanRenderer::RemapDefaultFramebufferReadback(
+            const Uint8* rawPixels, Uint32 logicalWidth, Uint32 logicalHeight,
+            VkSurfaceTransformFlagBitsKHR preTransform, SizeT texelSize, Uint8* outPixels) {
+        if (rawPixels == nullptr || outPixels == nullptr || logicalWidth == 0 || logicalHeight == 0 ||
+            texelSize == 0) {
+            return false;
+        }
+
+        const Uint32 rawWidth = IsQuarterTurnPreTransform(preTransform) ? logicalHeight : logicalWidth;
+        for (Uint32 outY = 0; outY < logicalHeight; ++outY) {
+            for (Uint32 outX = 0; outX < logicalWidth; ++outX) {
+                Uint32 srcX = outX;
+                Uint32 srcY = outY;
+                switch (preTransform) {
+                case VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR:
+                    srcX = outY;
+                    srcY = outX;
+                    break;
+                case VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR:
+                    srcX = logicalWidth - 1 - outX;
+                    break;
+                case VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR:
+                    srcX = logicalHeight - 1 - outY;
+                    srcY = logicalWidth - 1 - outX;
+                    break;
+                default:
+                    srcY = logicalHeight - 1 - outY;
+                    break;
+                }
+                Memcpy(outPixels + (static_cast<SizeT>(outY) * logicalWidth + outX) * texelSize,
+                       rawPixels + (static_cast<SizeT>(srcY) * rawWidth + srcX) * texelSize,
+                       texelSize);
+            }
+        }
+        return true;
+    }
+
     Bool VulkanRenderer::ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                                GLsizei width, GLsizei height, GLenum destinationFormat,
                                                GLenum destinationType, SizeT destinationRowStride,
@@ -4879,6 +4953,7 @@ void main() {
             .topology = vkTopology,
             .primitiveRestartEnable = primitiveRestartEnabled,
             .patchControlPoints = static_cast<Uint32>(MG_State::pGLContext->GetPatchVertices()),
+            .viewportCount = ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin),
             .polygonMode = effectivePolygonMode,
             .cullMode = cullFaceEnabled
                 ? MG_Util::ConvertCullFaceModeToVkEnum(MG_State::pGLContext->GetCullFaceMode(), invertClockwise)
@@ -5321,9 +5396,71 @@ void main() {
     }
 
 
-    void VulkanRenderer::ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent,
-                                                   Bool isDefaultFbo) {
+    // The scissor rectangle Vulkan needs for ARB_viewport_array index `index`. Vulkan has no
+    // per-viewport scissor-test TOGGLE - a scissor rectangle always applies - so an index whose
+    // GL scissor test is disabled gets the whole framebuffer, which is exactly "the test always
+    // passes" (GL 4.6 core 17.3.2).
+    VkRect2D VulkanRenderer::ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                                  VkSurfaceTransformFlagBitsKHR preTransform,
+                                                  Bool isDefaultFbo) const {
+        const auto& parameters = MG_State::pGLContext->GetRenderStateParameters();
+        if ((parameters.ScissorTestEnabledMask & (1u << index)) == 0) {
+            VkRect2D full{};
+            full.offset = {0, 0};
+            full.extent = {static_cast<Uint32>(extent.x()), static_cast<Uint32>(extent.y())};
+            return full;
+        }
+        const IntVec4& scissorBox = parameters.ScissorBoxes[index];
+        return isDefaultFbo ? MakeDefaultFramebufferScissorRect(scissorBox, extent, preTransform)
+                            : MakeClampedScissorRect(scissorBox, extent);
+    }
+
+    // The wide half of ApplyDynamicDrawStateTail: a pipeline built for a gl_ViewportIndex-writing
+    // program declares viewportCount > 1, and Vulkan then requires that many viewports AND that
+    // many scissors to have been set before the draw
+    // (VUID-vkCmdDraw-viewportCount-03417/-03418). Deliberately unmemoized: only conformance
+    // shaders reach it, the single-element dynamic-state shadow cannot describe an array, and
+    // leaving that shadow invalidated is what makes the next ordinary draw re-push its own
+    // single viewport instead of believing the array's element 0 is already bound.
+    void VulkanRenderer::ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount,
+                                                        const IntVec2& extent,
+                                                        VkSurfaceTransformFlagBitsKHR preTransform,
+                                                        Bool isDefaultFbo) {
+        MOBILEGL_ASSERT(viewportCount <= RenderStateParameters::MAX_VIEWPORTS,
+                        "ApplyMultiViewportDynamicState: viewportCount=%u exceeds the indexed state width",
+                        viewportCount);
+        const Uint32 count = std::min<Uint32>(viewportCount, RenderStateParameters::MAX_VIEWPORTS);
+
+        Array<VkViewport, RenderStateParameters::MAX_VIEWPORTS> viewports{};
+        Array<VkRect2D, RenderStateParameters::MAX_VIEWPORTS> scissors{};
+        for (Uint32 i = 0; i < count; ++i) {
+            viewports[i] = ComputeGLViewport(i, extent, preTransform, isDefaultFbo);
+            scissors[i] = ComputeGLScissorRect(i, extent, preTransform, isDefaultFbo);
+        }
+        vkCmdSetViewport(commandBuffer, 0, count, viewports.data());
+        vkCmdSetScissor(commandBuffer, 0, count, scissors.data());
+
         auto& shadow = g_dynamicStateShadow;
+        shadow.viewportValid = false;
+        shadow.scissorValid = false;
+        shadow.dynamicTailValid = false;
+    }
+
+    void VulkanRenderer::ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent,
+                                                   Bool isDefaultFbo, Uint32 viewportCount) {
+        auto& shadow = g_dynamicStateShadow;
+        if (viewportCount > 1) {
+            // The other five Apply* still run: blend constants, depth bias, line width and the
+            // stencil masks are not per-viewport and a multi-viewport draw needs them just as
+            // much. Only the viewport/scissor pair takes the array shape.
+            ApplyBlendConstants(frame.commandBuffer);
+            ApplyPolygonOffsetState(frame.commandBuffer);
+            ApplyLineWidthState(frame.commandBuffer);
+            ApplyStencilState(frame.commandBuffer);
+            ApplyMultiViewportDynamicState(frame.commandBuffer, viewportCount, extent,
+                                           m_swapchainObject.GetPreTransform(), isDefaultFbo);
+            return;
+        }
         // One compare for the whole tail: see the gate's declaration in
         // DynamicStateShadow for why (version, extent, default-FBO flag) pins every
         // input the six Apply* below read.
@@ -5342,12 +5479,14 @@ void main() {
         DynamicStateShadow::DynamicTailKey key;
         {
             const RenderStateParameters& p = MG_State::pGLContext->GetRenderStateParameters();
-            key.viewport[0] = p.Viewport.x();
-            key.viewport[1] = p.Viewport.y();
-            key.viewport[2] = p.Viewport.z();
-            key.viewport[3] = p.Viewport.w();
-            key.depthRange[0] = p.DepthRange.x();
-            key.depthRange[1] = p.DepthRange.y();
+            // Viewport 0 and its depth range: ApplyGLViewportState reads exactly those two
+            // (per-index state for indices > 0 is keyed separately, see multiViewportKey below).
+            key.viewport[0] = p.Viewports[0].x();
+            key.viewport[1] = p.Viewports[0].y();
+            key.viewport[2] = p.Viewports[0].z();
+            key.viewport[3] = p.Viewports[0].w();
+            key.depthRange[0] = p.DepthRanges[0].x();
+            key.depthRange[1] = p.DepthRanges[0].y();
             key.blendColor[0] = p.BlendColor.x();
             key.blendColor[1] = p.BlendColor.y();
             key.blendColor[2] = p.BlendColor.z();
@@ -5362,11 +5501,11 @@ void main() {
                 key.stencilWriteMask[face] = p.StencilStates[face].WriteMask;
                 key.stencilRef[face] = p.StencilStates[face].Ref;
             }
-            key.scissorEnabled = p.ScissorTestEnabled;
-            key.scissorBox[0] = p.ScissorBox.x();
-            key.scissorBox[1] = p.ScissorBox.y();
-            key.scissorBox[2] = p.ScissorBox.z();
-            key.scissorBox[3] = p.ScissorBox.w();
+            key.scissorEnabled = (p.ScissorTestEnabledMask & 1u) != 0;
+            key.scissorBox[0] = p.ScissorBoxes[0].x();
+            key.scissorBox[1] = p.ScissorBoxes[0].y();
+            key.scissorBox[2] = p.ScissorBoxes[0].z();
+            key.scissorBox[3] = p.ScissorBoxes[0].w();
             key.extentX = extent.x();
             key.extentY = extent.y();
             key.preTransform = static_cast<Uint32>(preTransform);
@@ -5759,7 +5898,7 @@ void main() {
             const Bool idxUploadOk = UploadAndBindIndexBuffer(frame, vao, pIndexBufferView);
             MOBILEGL_ASSERT(idxUploadOk, "SetupDraw fast path: failed to upload index buffer");
         }
-        ApplyDynamicDrawStateTail(frame, snap.renderPassExtent, snap.drawFboIsDefault);
+        ApplyDynamicDrawStateTail(frame, snap.renderPassExtent, snap.drawFboIsDefault, snap.viewportCount);
         return true;
     }
 
@@ -6199,7 +6338,8 @@ void main() {
             MOBILEGL_ASSERT(idxUploadOk, "SetupDraw skipped: failed to upload index buffer");
         }
 
-        ApplyDynamicDrawStateTail(frame, renderPassEntry->extent, drawFbo->IsDefaultFramebuffer());
+        ApplyDynamicDrawStateTail(frame, renderPassEntry->extent, drawFbo->IsDefaultFramebuffer(),
+                                  ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin));
 
         // Snapshot the fully resolved configuration for the consecutive-draw
         // fast path (see TrySetupDrawFastPath).
@@ -6218,6 +6358,7 @@ void main() {
                 snap.drawFbo = drawFbo.get();
                 snap.fboVersion = drawFbo->GetObjectVersion();
                 snap.drawFboIsDefault = drawFboIsDefault;
+                snap.viewportCount = ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin);
                 snap.renderStateVersion = MG_State::pGLContext->GetPipelineStateVersion();
                 snap.bindGeneration = MG_State::pGLContext->GetTextureBindGeneration();
                 snap.baseTransformFlags = GetBaseTransformFlagsRaw(drawFboIsDefault);
@@ -7113,7 +7254,7 @@ void main() {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-            resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+            resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForTexture: failed to transition textureId=%d to TRANSFER_DST",
                         texture.GetExternalIndex());
@@ -7231,8 +7372,7 @@ void main() {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, clearLayout, sampledLayout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
-            VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
-            resource->arrayLayers);
+            VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForTexture: failed to transition textureId=%d to sampled layout",
                         texture.GetExternalIndex());
@@ -7270,7 +7410,7 @@ void main() {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-            resource->aspect, 0, 1, 1);
+            resource->aspect, 0, 1);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForRenderbuffer: failed to transition renderbuffer %u to TRANSFER_DST",
                         renderbuffer->GetExternalIndex());
@@ -7321,7 +7461,7 @@ void main() {
             VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
                 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
                 VK_ACCESS_TRANSFER_READ_BIT,
-            resource->aspect, 0, 1, 1);
+            resource->aspect, 0, 1);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForRenderbuffer: failed to transition renderbuffer %u to steady layout",
                         renderbuffer->GetExternalIndex());
@@ -7928,6 +8068,9 @@ void main() {
             VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
             VkAccessFlags srcAccessMask = 0;
             GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
+            // Both blit regions below name `baseArrayLayer` from their binding, and a layered depth
+            // attachment puts that above 0. These barriers carry a mip range only - their layer
+            // range is every layer (see VkTextureManager::TransitionImageLayout).
             if (readIsDefaultFbo) {
                 VkImageLayout srcTrackedLayout = srcOriginalLayout;
                 Bool ok = VkTextureManager::TransitionImageLayout(
@@ -8755,30 +8898,22 @@ void main() {
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
         VkImageLayout srcCopyLayout = srcOriginalLayout;
-        // The barrier has to name every layer the copy touches, not just layer 0 - otherwise the
-        // slice fix above lands the copy on layers the barrier never transitioned, which is the
-        // same defect one level down. TransitionImageLayout always starts its range at
-        // baseArrayLayer 0, so VK_REMAINING_ARRAY_LAYERS is the whole range and a superset of
-        // [baseSlice, baseSlice + depth).
-        //
-        // Not `arrayLayers`, which is 1 for a 3D image: MobileGL creates 3D images
-        // 2D_ARRAY_COMPATIBLE, and a literal 1 on one of those means "every depth slice" today but
-        // "depth slice 0" once VK_KHR_maintenance9 is enabled - i.e. it would silently become a
-        // single-slice barrier again on a newer driver. The validation layer says so by name.
-        static constexpr Uint32 kAllLayers = VK_REMAINING_ARRAY_LAYERS;
+        // The barriers below name a MIP range only. Their layer range is not a parameter:
+        // TransitionImageLayout always covers every layer of the image, which is a superset of the
+        // [baseSlice, baseSlice + depth) the slice mapping above hands the copy.
         if (srcOriginalLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
             Bool srcReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcResource->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                 srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
                 srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT,
-                srcResource->aspect, 0, srcResource->mipLevels, kAllLayers);
+                srcResource->aspect, 0, srcResource->mipLevels);
             MOBILEGL_ASSERT(srcReady, "%s: failed to transition undefined source image", __func__);
             srcCopyLayout = srcResource->layout;
         } else {
             Bool srcReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcCopyLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                 srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, copyAspectMask, srcMipLevel, 1, kAllLayers);
+                srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, copyAspectMask, srcMipLevel, 1);
             MOBILEGL_ASSERT(srcReady, "%s: failed to transition source image", __func__);
         }
 
@@ -8791,14 +8926,14 @@ void main() {
                 frame.commandBuffer, dstResource->image, dstResource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                 dstStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
                 dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-                dstResource->aspect, 0, dstResource->mipLevels, kAllLayers);
+                dstResource->aspect, 0, dstResource->mipLevels);
             MOBILEGL_ASSERT(dstReady, "%s: failed to transition undefined destination image", __func__);
             dstCopyLayout = dstResource->layout;
         } else {
             Bool dstReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, dstResource->image, dstCopyLayout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                 dstStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT, copyAspectMask, dstMipLevel, 1, kAllLayers);
+                dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT, copyAspectMask, dstMipLevel, 1);
             MOBILEGL_ASSERT(dstReady, "%s: failed to transition destination image", __func__);
         }
 
@@ -8840,13 +8975,13 @@ void main() {
                 frame.commandBuffer, srcResource->image, srcResource->layout, srcRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, srcRestoreStageMask,
                 VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask,
-                srcResource->aspect, 0, srcResource->mipLevels, kAllLayers);
+                srcResource->aspect, 0, srcResource->mipLevels);
             MOBILEGL_ASSERT(srcRestored, "%s: failed to restore undefined source image layout", __func__);
         } else {
             Bool srcRestored = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcCopyLayout, srcRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, srcRestoreStageMask,
-                VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask, copyAspectMask, srcMipLevel, 1, kAllLayers);
+                VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask, copyAspectMask, srcMipLevel, 1);
             MOBILEGL_ASSERT(srcRestored, "%s: failed to restore source image layout", __func__);
         }
 
@@ -8858,13 +8993,13 @@ void main() {
                 frame.commandBuffer, dstResource->image, dstResource->layout, dstRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, dstRestoreStageMask,
                 VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask,
-                dstResource->aspect, 0, dstResource->mipLevels, kAllLayers);
+                dstResource->aspect, 0, dstResource->mipLevels);
             MOBILEGL_ASSERT(dstRestored, "%s: failed to restore undefined destination image layout", __func__);
         } else {
             Bool dstRestored = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, dstResource->image, dstCopyLayout, dstRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, dstRestoreStageMask,
-                VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask, copyAspectMask, dstMipLevel, 1, kAllLayers);
+                VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask, copyAspectMask, dstMipLevel, 1);
             MOBILEGL_ASSERT(dstRestored, "%s: failed to restore destination image layout", __func__);
         }
 
@@ -9023,6 +9158,9 @@ void main() {
         VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
+        // The copy below reads `srcBinding.baseArrayLayer`, which for a glFramebufferTextureLayer
+        // attachment is any layer of the array - the barrier covers all of them (see
+        // VkTextureManager::TransitionImageLayout), so the layer being read is one it moved.
         if (readIsDefaultFbo) {
             VkImageLayout trackedLayout = srcOriginalLayout;
             Bool ok = VkTextureManager::TransitionImageLayout(
@@ -9048,19 +9186,18 @@ void main() {
         // The GL rect, aimed at the default framebuffer's stored orientation. Using the GL y
         // verbatim copied rows [y, y+h) counted from the TOP of the image, i.e. the wrong band for
         // every read that was not full-height.
-        Int32 copyOffsetX = x;
-        Int32 copyOffsetY = y;
+        VkOffset2D copyOffset{x, y};
+        VkExtent2D copyExtent{static_cast<Uint32>(width), static_cast<Uint32>(height)};
         if (readIsDefaultFbo) {
             const VkExtent2D defaultFboExtent = m_swapchainObject.GetExtent();
-            const DefaultFramebufferRectMapping mapping =
-                GetDefaultFramebufferRectMapping(m_swapchainObject.GetPreTransform());
-            copyOffsetX = MapDefaultFramebufferRectAxis(x, width, static_cast<Int>(defaultFboExtent.width),
-                                                        mapping.mirrorX);
-            copyOffsetY = MapDefaultFramebufferRectAxis(y, height, static_cast<Int>(defaultFboExtent.height),
-                                                        mapping.flipY);
+            const Bool mapped = MapDefaultFramebufferReadbackRect(
+                x, y, width, height, defaultFboExtent, m_swapchainObject.GetPreTransform(), &copyOffset,
+                &copyExtent);
+            MOBILEGL_ASSERT(mapped, "ReadPixels: default framebuffer read rectangle is out of bounds");
+            if (!mapped) return;
         }
-        copyRegion.imageOffset = {copyOffsetX, copyOffsetY, static_cast<Int32>(srcBinding.depthOffset)};
-        copyRegion.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+        copyRegion.imageOffset = {copyOffset.x, copyOffset.y, static_cast<Int32>(srcBinding.depthOffset)};
+        copyRegion.imageExtent = {copyExtent.width, copyExtent.height, 1};
         vkCmdCopyImageToBuffer(frame.commandBuffer, srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                                readback.GetHandle(), 1, &copyRegion);
 
@@ -9102,16 +9239,14 @@ void main() {
             // already aimed with the same mapping. The gate is exactly what made every partial
             // read of the default framebuffer come back in Vulkan row order.
             Vector<Uint8> remapped(static_cast<SizeT>(width) * static_cast<SizeT>(height) * sourceTexelSize);
-            if (RemapDefaultFboReadbackToGLOrientation(mapped, static_cast<Uint32>(width),
-                                                       static_cast<Uint32>(height), preTransform, sourceTexelSize,
-                                                       remapped.data())) {
+            if (RemapDefaultFramebufferReadback(mapped, static_cast<Uint32>(width),
+                                                static_cast<Uint32>(height), preTransform, sourceTexelSize,
+                                                remapped.data())) {
                 PackReadbackToClientOrPbo(remapped.data(), srcFormat, width, height, 1, format, type, pixels,
                                           /*applyPackImageParams=*/false, /*applyReadColorClamp=*/true);
                 return;
             }
-            // Only a quarter-turn pre-transform reaches this, and nothing in this renderer models
-            // one. MGLOG_I because the INFO builds are the ones that run conformance.
-            MGLOG_D("DirectVulkan::ReadPixels: default-FBO remap declined (w=%d h=%d preTransform=%d); falling back "
+            MGLOG_D("DirectVulkan::ReadPixels: default-FBO remap failed (w=%d h=%d preTransform=%d); falling back "
                     "to raw readback",
                     width, height, static_cast<Int>(preTransform));
         }
@@ -9492,16 +9627,15 @@ void main() {
         // The swapchain's depth/stencil image is stored display-side-up like its colour twin, so
         // the GL rect has to be mapped into that space before the copy and the copied rows
         // re-oriented afterwards - the same two halves the colour ReadPixels path applies.
-        Int32 copyOffsetX = x;
-        Int32 copyOffsetY = y;
+        VkOffset2D copyOffset{x, y};
+        VkExtent2D copyExtent{static_cast<Uint32>(width), static_cast<Uint32>(height)};
         if (defaultFramebufferOrientation) {
             const VkExtent2D defaultFboExtent = m_swapchainObject.GetExtent();
-            const DefaultFramebufferRectMapping mapping =
-                GetDefaultFramebufferRectMapping(m_swapchainObject.GetPreTransform());
-            copyOffsetX = MapDefaultFramebufferRectAxis(x, width, static_cast<Int>(defaultFboExtent.width),
-                                                        mapping.mirrorX);
-            copyOffsetY = MapDefaultFramebufferRectAxis(y, height, static_cast<Int>(defaultFboExtent.height),
-                                                        mapping.flipY);
+            const Bool mapped = MapDefaultFramebufferReadbackRect(
+                x, y, width, height, defaultFboExtent, m_swapchainObject.GetPreTransform(), &copyOffset,
+                &copyExtent);
+            MOBILEGL_ASSERT(mapped, "ReadDepthStencilPixels: default framebuffer read rectangle is out of bounds");
+            if (!mapped) return;
         }
 
         VkBufferImageCopy regions[2]{};
@@ -9513,8 +9647,8 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
-            region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+            region.imageOffset = {copyOffset.x, copyOffset.y, 0};
+            region.imageExtent = {copyExtent.width, copyExtent.height, 1};
         }
         if (wantStencil) {
             auto& region = regions[regionCount++];
@@ -9523,8 +9657,8 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
-            region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+            region.imageOffset = {copyOffset.x, copyOffset.y, 0};
+            region.imageExtent = {copyExtent.width, copyExtent.height, 1};
         }
         vkCmdCopyImageToBuffer(frame.commandBuffer, image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, readback.GetHandle(),
                                regionCount, regions);
@@ -9558,23 +9692,21 @@ void main() {
             Bool remapped = true;
             if (wantDepth && depthCopyBytes > 0) {
                 remappedDepth.resize(pixelCount * depthCopyBytes);
-                remapped = RemapDefaultFboReadbackToGLOrientation(depthSrc, static_cast<Uint32>(width),
-                                                                  static_cast<Uint32>(height), preTransform,
-                                                                  depthCopyBytes, remappedDepth.data());
+                remapped = RemapDefaultFramebufferReadback(depthSrc, static_cast<Uint32>(width),
+                                                           static_cast<Uint32>(height), preTransform,
+                                                           depthCopyBytes, remappedDepth.data());
             }
             if (remapped && wantStencil) {
                 remappedStencil.resize(pixelCount);
-                remapped = RemapDefaultFboReadbackToGLOrientation(stencilSrc, static_cast<Uint32>(width),
-                                                                  static_cast<Uint32>(height), preTransform, 1,
-                                                                  remappedStencil.data());
+                remapped = RemapDefaultFramebufferReadback(stencilSrc, static_cast<Uint32>(width),
+                                                           static_cast<Uint32>(height), preTransform, 1,
+                                                           remappedStencil.data());
             }
             if (remapped) {
                 if (!remappedDepth.empty()) depthSrc = remappedDepth.data();
                 if (!remappedStencil.empty()) stencilSrc = remappedStencil.data();
             } else {
-                // Only a quarter-turn pre-transform reaches this, and nothing in this renderer
-                // models one. MGLOG_I because the INFO builds are the ones that run conformance.
-                MGLOG_D("DirectVulkan::ReadDepthStencilPixels: default-FBO remap declined (w=%d h=%d "
+                MGLOG_D("DirectVulkan::ReadDepthStencilPixels: default-FBO remap failed (w=%d h=%d "
                         "preTransform=%d); falling back to raw readback",
                         width, height, static_cast<Int>(preTransform));
             }
@@ -9825,14 +9957,13 @@ void main() {
         VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(originalLayout, srcStageMask, srcAccessMask);
-        // The copy below reads EVERY layer of the level, so the barrier has to name every layer
-        // too; a layerCount of 1 left an array texture's layers 1.. in whatever layout they were
-        // last left in while the transfer read them.
+        // The copy below reads EVERY layer of the level, which is exactly the range
+        // TransitionImageLayout barriers cover.
         Bool ok = VkTextureManager::TransitionImageLayout(
             frame.commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
             srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, resource->aspect,
-            static_cast<Uint32>(level), 1, VK_REMAINING_ARRAY_LAYERS);
+            static_cast<Uint32>(level), 1);
         MOBILEGL_ASSERT(ok, "%s: failed to transition texture image", __func__);
 
         VkBufferImageCopy copyRegion{};
@@ -9852,7 +9983,7 @@ void main() {
             frame.commandBuffer, resource->image, resource->layout, originalLayout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, restoreStageMask,
             VK_ACCESS_TRANSFER_READ_BIT, restoreAccessMask, resource->aspect,
-            static_cast<Uint32>(level), 1, VK_REMAINING_ARRAY_LAYERS);
+            static_cast<Uint32>(level), 1);
         MOBILEGL_ASSERT(ok, "%s: failed to restore texture image layout", __func__);
 
         if (!SubmitReadbackCommandsAndWait(frame)) {
@@ -9960,7 +10091,7 @@ void main() {
             Bool transitioned = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, resource->image, resource->layout, finalLayout,
                 VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
-                0, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                0, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
             MOBILEGL_ASSERT(transitioned, "GenerateMipmap: failed to transition uninitialized mip chain");
             return;
         }
@@ -12218,6 +12349,28 @@ void main() {
         m_fillModeNonSolidFeatureEnabled = deviceFeatures.fillModeNonSolid == VK_TRUE;
         deviceFeatures.dualSrcBlend = supportedDeviceFeatures.dualSrcBlend;
         m_dualSrcBlendFeatureEnabled = deviceFeatures.dualSrcBlend == VK_TRUE;
+        // ARB_viewport_array rasterization. Without multiViewport a pipeline may declare exactly
+        // one viewport (VUID-VkPipelineViewportStateCreateInfo-viewportCount-01216), so a shader's
+        // gl_ViewportIndex can only ever select viewport 0 and the other fifteen rectangles are
+        // state with nowhere to go. The GL state stays 16 wide either way - GL 4.3 core requires
+        // MAX_VIEWPORTS >= 16 and that is a frontend promise, not a device one; this gate decides
+        // only whether a DRAW can rasterize into more than one of them.
+        deviceFeatures.multiViewport = supportedDeviceFeatures.multiViewport;
+        m_multiViewportFeatureEnabled = deviceFeatures.multiViewport == VK_TRUE;
+        m_maxRasterizableViewports =
+            m_multiViewportFeatureEnabled
+                ? std::min<Uint32>(RenderStateParameters::MAX_VIEWPORTS,
+                                   std::max<Uint32>(m_physicalDevice.properties.limits.maxViewports, 1u))
+                : 1u;
+        MGLOG_I("Vulkan: multiViewport %s; rasterizable viewports=%u (device limit %u, GL state width %u)",
+                m_multiViewportFeatureEnabled ? "enabled" : "UNAVAILABLE", m_maxRasterizableViewports,
+                m_physicalDevice.properties.limits.maxViewports,
+                static_cast<Uint32>(RenderStateParameters::MAX_VIEWPORTS));
+        if (!m_multiViewportFeatureEnabled) {
+            MGLOG_W("Vulkan: the device does not support the multiViewport feature; gl_ViewportIndex will always "
+                    "select viewport 0 and per-viewport scissor/depth-range state past index 0 cannot be "
+                    "rasterized (the state itself is still stored and queryable)");
+        }
         deviceFeatures.logicOp = supportedDeviceFeatures.logicOp;
         deviceFeatures.shaderClipDistance = supportedDeviceFeatures.shaderClipDistance;
         deviceFeatures.shaderCullDistance = supportedDeviceFeatures.shaderCullDistance;

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -229,6 +229,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLint dstY, GLint width, GLint height, VkImageLayout srcRestoreLayout,
                                     VkImageLayout dstRestoreLayout, Bool stencilAspect);
         static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
+        // Map a GL bottom-left-origin rectangle into the display-oriented swapchain image.
+        // Quarter-turn surface transforms swap the copy extent's axes.
+        static Bool MapDefaultFramebufferReadbackRect(GLint x, GLint y, GLsizei width, GLsizei height,
+                                                      VkExtent2D imageExtent,
+                                                      VkSurfaceTransformFlagBitsKHR preTransform,
+                                                      VkOffset2D* imageOffset, VkExtent2D* imageCopyExtent);
+        // Reorder a tightly packed block copied with MapDefaultFramebufferReadbackRect back into
+        // GL row order. The input block has swapped dimensions for 90/270 degree transforms.
+        static Bool RemapDefaultFramebufferReadback(const Uint8* rawPixels, Uint32 logicalWidth,
+                                                    Uint32 logicalHeight,
+                                                    VkSurfaceTransformFlagBitsKHR preTransform,
+                                                    SizeT texelSize, Uint8* outPixels);
         static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                           GLsizei width, GLsizei height, GLenum destinationFormat,
                                           GLenum destinationType, SizeT destinationRowStride,
@@ -298,6 +310,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The samplerAnisotropy device feature was granted, so GL_TEXTURE_MAX_ANISOTROPY_EXT is
         // honored rather than accepted-and-ignored.
         Bool IsSamplerAnisotropySupported() const { return m_samplerAnisotropyFeatureEnabled; }
+        // ARB_base_instance extends indirect command records with a non-zero firstInstance and
+        // requires gl_InstanceID to remain zero-based. Vulkan needs both features to honor that
+        // complete contract: one legalizes the command word, the other enables the shader rebase.
+        Bool IsNonZeroIndirectBaseInstanceSupported() const {
+            return m_drawIndirectFirstInstanceFeatureEnabled && m_shaderDrawParametersFeatureEnabled;
+        }
         // Ensures the frame command buffer is recording (same lazy pattern as
         // SetupDraw) and writes a bottom-of-pipe timestamp into the current
         // frame's pool. Null when unsupported or the pool is exhausted.
@@ -547,6 +565,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // needs no feature). Both cached at device creation and drive a hard-fail-at-draw when absent.
         Bool m_dualSrcBlendFeatureEnabled = false;
         Bool m_primitiveTopologyListRestartFeatureEnabled = false;
+        // multiViewport gates rasterizing into more than one of ARB_viewport_array's 16 viewports
+        // (gl_ViewportIndex). m_maxRasterizableViewports is min(MAX_VIEWPORTS, device limit), or 1
+        // when the feature is off, and is the viewportCount a gl_ViewportIndex-writing pipeline
+        // declares - it is NOT what GL_MAX_VIEWPORTS reports, which is the frontend state width.
+        Bool m_multiViewportFeatureEnabled = false;
+        Uint32 m_maxRasterizableViewports = 1;
         // Union of shader stages sampled-read barriers may name; built at device creation
         // because geometry/tessellation stage bits are invalid in a barrier when their
         // feature is off (VUID-vkCmdPipelineBarrier-srcStageMask-04090/-04091), and
@@ -830,6 +854,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // re-resolve just the pipeline against the active pass; a change that
             // flips it must fall back to the full path's pass selection.
             Bool drawUsesDepthStencil = false;
+            // The snapshotting draw's pipeline viewportCount. A pure function of the PROGRAM
+            // (writesViewportIndexBuiltin) and of a device feature fixed at renderer init, both
+            // of which the programLifetimeId/programVersion guards above already pin - carried
+            // here so the fast path does not re-fetch the program object to re-derive it.
+            Uint32 viewportCount = 1;
             IntVec2 renderPassExtent = {0, 0};
             // colorAttachmentCount of the snapshotting draw's render pass: the
             // pipeline-state hash input, so the fast path can refresh that hash and
@@ -1120,7 +1149,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The per-draw dynamic-state tail (viewport, scissor, blend constants, depth
         // bias, line width, stencil), gated behind one render-state-parameters-version
         // compare per command buffer - see the gate fields in DynamicStateShadow.
-        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo);
+        // viewportCount is the bound pipeline's declared viewport count: 1 for every program that
+        // does not write gl_ViewportIndex (the memoized fast path), otherwise the renderer's
+        // rasterizable viewport count, which takes the unmemoized array path.
+        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo,
+                                       Uint32 viewportCount = 1);
+        void ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount, const IntVec2& extent,
+                                            VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo);
+        VkRect2D ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                      VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo) const;
+        // How many viewports a draw with this program rasterizes into: 1 unless the program
+        // assigns gl_ViewportIndex AND the device enabled multiViewport. Both the pipeline's
+        // baked viewportCount and the dynamic arrays come from this one answer, so they cannot
+        // disagree.
+        Uint32 ResolveDrawViewportCount(Bool programWritesViewportIndex) const {
+            return programWritesViewportIndex && m_multiViewportFeatureEnabled ? m_maxRasterizableViewports : 1u;
+        }
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -969,14 +969,14 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL3dv, GLuint index, const GLdoub
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL4dv, GLuint index, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribL4dv, index, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribLPointer, GLuint index, GLint size, GLenum type, GLsizei stride, const void* pointer) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribLPointer, index, size, type, stride, pointer)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetVertexAttribLdv, GLuint index, GLenum pname, GLdouble* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetVertexAttribLdv, index, pname, params)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedfv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexedv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedfv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexedv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
 DECLARE_GL_FUNCTION_HEAD(void, GetFloati_v, GLenum target, GLuint index, GLfloat* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetFloati_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, GetDoublei_v, GLenum target, GLuint index, GLdouble* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetDoublei_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, DrawArraysInstancedBaseInstance, GLenum mode, GLint first, GLsizei count, GLsizei instancecount, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawArraysInstancedBaseInstance, mode, first, count, instancecount, baseinstance)

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -16,6 +16,7 @@
 #include <MG_State/GLState/ErrorState/ErrorInfo.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
+#include <MG_Util/Converters/GLToMG/RenderStateEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/FramebufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/ErrorCodeConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -27,6 +28,11 @@
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
+    // Declared rather than #included from GL_RenderState.h on purpose: that header also declares
+    // a free function named BlendEquation, which would hide the ::MobileGL::BlendEquation enum
+    // this file's blend-state queries name unqualified.
+    GLboolean IsEnabledi(GLenum target, GLuint index);
+
     namespace {
         enum class IndexedBufferQueryKind {
             Binding,
@@ -339,26 +345,70 @@ namespace MobileGL::MG_Impl::GLImpl {
             return sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
         }
 
-        // The ARB_viewport_array indexed rectangles. MobileGL keeps exactly one viewport, one
-        // scissor box and one depth range, so every in-range index answers with that single
-        // value - but it has to come from the frontend state the non-indexed getters read.
-        // The generic path at the bottom of GetIntegeri_v is a raw backend passthrough that
-        // has no case for these, so routing them through it returned zeros.
+        // The ARB_viewport_array indexed rectangles. Each of these is genuinely per-viewport
+        // frontend state (RenderStateParameters::Viewports / ScissorBoxes / DepthRanges), so the
+        // indexed getters must read the indexed storage - the generic path at the bottom of
+        // GetIntegeri_v is a raw backend passthrough that has no case for them and returned
+        // zeros, and routing them to the NON-indexed getter (what this used to do) answered every
+        // index with viewport 0's value, which is what
+        // KHR-GL43.viewport_array.{viewport,scissor,depth_range}_api caught.
         Bool IsIndexedViewportQuery(GLenum target) {
             return target == GL_VIEWPORT || target == GL_SCISSOR_BOX || target == GL_DEPTH_RANGE;
         }
 
-        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it.
+        // Component count of an indexed viewport-array query, so every width of getter writes the
+        // caller's whole buffer instead of just element 0 (GL 4.6 core 22.1).
+        GLsizei IndexedViewportQueryComponents(GLenum target) {
+            return target == GL_DEPTH_RANGE ? 2 : 4;
+        }
+
+        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it. The bound is
+        // the frontend's own state width, which is also exactly what GL_MAX_VIEWPORTS reports -
+        // taking it from the backend caps instead would let a device limit of 1 (a Vulkan device
+        // without the multiViewport feature) make index 1 illegal even though the state exists.
         Bool ValidateViewportQueryIndex(GLuint index, const char* caller) {
-            GLint maxViewports = 0;
-            GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
-            if (index < static_cast<GLuint>(std::max(maxViewports, 1))) return true;
+            if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Viewport index is out of range."));
             return false;
         }
 
+        // The indexed viewport/scissor/depth-range state as floats, which is the widest lossless
+        // shape MobileGL stores (the viewport really is float state; the scissor box is integral
+        // and well inside float's exact range, and every depth range is in [0, 1]). Every indexed
+        // getter width funnels through this so they can never disagree with each other.
+        void ReadIndexedViewportStateFloat(GLenum target, GLuint index, GLfloat* out) {
+            switch (target) {
+            case GL_VIEWPORT: {
+                const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(index);
+                out[0] = viewport.x();
+                out[1] = viewport.y();
+                out[2] = viewport.z();
+                out[3] = viewport.w();
+                return;
+            }
+            case GL_SCISSOR_BOX: {
+                const IntVec4& box = MG_State::pGLContext->GetScissorBoxIndexed(index);
+                out[0] = static_cast<GLfloat>(box.x());
+                out[1] = static_cast<GLfloat>(box.y());
+                out[2] = static_cast<GLfloat>(box.z());
+                out[3] = static_cast<GLfloat>(box.w());
+                return;
+            }
+            case GL_DEPTH_RANGE: {
+                const FloatVec2& range = MG_State::pGLContext->GetDepthRangeIndexed(index);
+                out[0] = range.x();
+                out[1] = range.y();
+                return;
+            }
+            default:
+                MOBILEGL_ASSERT(false, "ReadIndexedViewportStateFloat: unexpected target 0x%x",
+                                static_cast<Uint32>(target));
+                return;
+            }
+        }
+
         void CopyIntsToBooleans(const GLint* src, SizeT count, GLboolean* dst) {
             for (SizeT i = 0; i < count; ++i) {
                 dst[i] = src[i] ? GL_TRUE : GL_FALSE;
@@ -629,6 +679,17 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.ViewportBoundsRangeMax;
             return;
         }
+        // Viewport 0's rectangle, verbatim. Falling through to the integer width below would
+        // round the fractional rectangle a glViewportIndexedf(0, ...) is allowed to set, and
+        // glGetFloatv(GL_VIEWPORT) is a lossless query of float state.
+        case GL_VIEWPORT: {
+            const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(0);
+            params[0] = viewport.x();
+            params[1] = viewport.y();
+            params[2] = viewport.z();
+            params[3] = viewport.w();
+            return;
+        }
         case GL_MIN_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_MAX_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_FRAGMENT_INTERPOLATION_OFFSET_BITS: {
@@ -792,15 +853,32 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // GL 4.6 core 22.1: an indexed query answers EVERY indexed state, and GL_SCISSOR_TEST is
+        // indexed by viewport just like GL_BLEND is by draw buffer. Without this the integer
+        // width fell through to the backend passthrough and answered GL_INVALID_ENUM, which is
+        // the sticky error KHR-GL43.viewport_array.queries trips over at its next error check.
+        if (MG_Util::ConvertGLEnumToCapabilityInput(target) != CapabilityInput::Unknown) {
+            *data = IsEnabledi(target, index);
+            return;
+        }
+
         switch (target) {
         // ARB_viewport_array queries the indexed rectangles through glGetIntegeri_v as well
-        // (gl4cMultiBindTests and the viewport_array group both do). The frontend keeps one
-        // viewport and one scissor box, so every in-range index reports that one.
+        // (gl4cMultiBindTests and the viewport_array group both do).
         case GL_VIEWPORT:
         case GL_SCISSOR_BOX:
+        case GL_DEPTH_RANGE: {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetIntegerv(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                // Round, not truncate: glGetIntegerv on floating-point state rounds to nearest
+                // (GL 4.6 core 22.2), so a 255.875-wide viewport reads back as 256 and not 255.
+                data[i] = static_cast<GLint>(std::lround(values[i]));
+            }
             return;
+        }
         // The vertex buffer binding points of the vertex array object that is bound. Indexed by
         // binding point, not by attribute (GL 4.6 core 10.3.1).
         case GL_VERTEX_BINDING_BUFFER:
@@ -927,7 +1005,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetFloatv(target, data);
+            // Verbatim, NOT via the integer width: the viewport is float state and
+            // KHR-GL43.viewport_array.viewport_api compares the read-back with ==, so a
+            // glViewportIndexedf(i, 0.125f, ...) has to come back as 0.125f exactly.
+            ReadIndexedViewportStateFloat(target, index, data);
             return;
         }
         GLint ints[4] = {};
@@ -944,7 +1025,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetDoublev(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                data[i] = static_cast<GLdouble>(values[i]);
+            }
             return;
         }
         GLint ints[4] = {};
@@ -1020,7 +1106,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         // frontend-only value simply is not in the driver's table.
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = static_cast<GLint64>(values[0]);
+        // The viewport-array rectangles are the only multi-component indexed state here; every
+        // other pname is scalar, so widening element 0 alone would silently truncate them.
+        const GLsizei components = IsIndexedViewportQuery(target) ? IndexedViewportQueryComponents(target) : 1;
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = static_cast<GLint64>(values[i]);
+        }
     }
 
     void GetInteger64v(GLenum pname, GLint64* params) {
@@ -2192,7 +2283,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.MaxViewportHeight;
             break;
         case GL_MAX_VIEWPORTS:
-            *params = dynamicParameters.MaxViewports;
+            // The frontend's own state width, not the backend's device limit. GL 4.3 core
+            // requires MAX_VIEWPORTS >= 16 and every indexed viewport entry point validates
+            // against RenderStateParameters::MAX_VIEWPORTS, so reporting anything else would
+            // either advertise viewports the state cannot hold or reject indices it can. A
+            // Vulkan device without the multiViewport feature reports maxViewports == 1, which
+            // limits what can be RASTERIZED to more than one rectangle (see the multiViewport
+            // gate in VulkanRenderer), not what the GL state can hold; caps.MaxViewports keeps
+            // carrying that device number for exactly that decision.
+            *params = static_cast<GLint>(RenderStateParameters::MAX_VIEWPORTS);
             break;
         case GL_MINOR_VERSION:
             *params = rendererInfo.RendererGLInfo.TargetGLVersion.Minor;

MobileGL/MG_Impl/GLImpl/Query/GL_Query.cpp

@@ -344,6 +344,41 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
+    void DestroyAllQueryObjects() {
+        // Detach the registry under the lock and release it outside. Entries the app
+        // already deleted were erased by DeleteQueries, so nothing here double-frees;
+        // a DeleteQueries racing this sweep finds an empty registry and ignores the
+        // names. The active-query slots and the name allocator are reset under the
+        // same lock: query names are context-owned state, so a fresh context must
+        // start clean instead of inheriting the dead context's allocator cursor or
+        // a stale "a query is already active on this target" latch.
+        UnorderedMap<GLuint, QueryObject*> orphans;
+        {
+            const std::lock_guard<std::mutex> lock(g_queryObjectsMutex);
+            orphans.swap(g_liveQueryObjects);
+            g_nextQueryId = 1;
+            g_activeTimeElapsedQueryId = 0;
+            g_activePrimitivesWrittenQueryId = 0;
+            g_activePrimitivesGeneratedQueryId = 0;
+            g_activeSamplesPassedQueryId = 0;
+        }
+        if (orphans.empty()) {
+            return;
+        }
+        // Both backends' DeleteBackendQuery only free the heap wrapper once their GL
+        // context/renderer is gone (generation/current-thread guards), so this is
+        // safe after the backend has released its EGL resources - but not after the
+        // function table itself is cleared.
+        const auto deleteBackendQuery = MG_Backend::gBackendFunctionsTable.GL.DeleteBackendQuery;
+        for (const auto& [_, queryObject] : orphans) {
+            if (deleteBackendQuery && queryObject->backendHandle) {
+                deleteBackendQuery(queryObject->backendHandle);
+            }
+            delete queryObject;
+        }
+        MGLOG_D("DestroyAllQueryObjects: reclaimed %zu query object(s) the app left undeleted", orphans.size());
+    }
+
     GLboolean IsQuery(GLuint id) {
         if (id == 0) {
             return GL_FALSE;

MobileGL/MG_Impl/GLImpl/Query/GL_Query.h

@@ -13,6 +13,15 @@ namespace MobileGL::MG_Impl::GLImpl {
     void GenQueries(GLsizei n, GLuint* ids);
     void CreateQueries(GLenum target, GLsizei n, GLuint* ids);
     void DeleteQueries(GLsizei n, const GLuint* ids);
+    // Destroys every still-registered query object exactly as DeleteQueries would.
+    // Query objects are context-owned, and MobileGL::Destroy() tears every context
+    // down, so the process-global registry has to be drained there: without this the
+    // QueryObject and any backend timer-query wrapper leaked across every
+    // eglTerminate/eglInitialize cycle, and the active-query/name-allocator state
+    // from the dead context survived into the next one. Must run while the backend
+    // function table is still populated, and before a re-initialized library could
+    // pair the handles with the wrong backend's DeleteBackendQuery.
+    void DestroyAllQueryObjects();
     GLboolean IsQuery(GLuint id);
     void BeginQuery(GLenum target, GLuint id);
     void EndQuery(GLenum target);

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.cpp

@@ -20,28 +20,118 @@ namespace MobileGL::MG_Impl::GLImpl {
         return std::clamp(static_cast<Float>(value), 0.0f, 1.0f);
     }
 
-    static Bool ValidateIndexedBlendCapability(GLenum target, GLuint index, const char* functionName) {
-        if (target != GL_BLEND) {
+    // GL 4.6 core 17.3.2 and 22.1 give exactly two indexed capabilities: GL_BLEND, indexed by
+    // draw buffer, and GL_SCISSOR_TEST, indexed by viewport. They have DIFFERENT bounds
+    // (MAX_DRAW_BUFFERS vs MAX_VIEWPORTS), so the limit is picked per target rather than shared.
+    static Bool ValidateIndexedCapability(GLenum target, GLuint index, const char* functionName) {
+        GLuint limit = 0;
+        const char* indexName = nullptr;
+        switch (target) {
+        case GL_BLEND:
+            limit = MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS;
+            indexName = "Buffer";
+            break;
+        case GL_SCISSOR_TEST:
+            limit = RenderStateParameters::MAX_VIEWPORTS;
+            indexName = "Viewport";
+            break;
+        default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                                             "Only GL_BLEND is supported for indexed capability state."));
+                                             "Only GL_BLEND and GL_SCISSOR_TEST are supported for indexed "
+                                             "capability state."));
             return false;
         }
 
-        if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {
+        if (index >= limit) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>(
-                    "MG_Impl/GLImpl", functionName,
-                    "Buffer index " + std::to_string(index) + " is out of range. Max supported is " +
-                        std::to_string(MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS - 1) + "."));
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             String(indexName) + " index " + std::to_string(index) +
+                                                 " is out of range. Max supported is " + std::to_string(limit - 1) +
+                                                 "."));
             return false;
         }
 
         return true;
     }
 
+    // ------------------ ARB_viewport_array parameter validation ------------------
+    // All three families share the same two shapes, so they share the two checkers. GL 4.6 core
+    // 13.6.1/17.3.2: an out-of-range index is GL_INVALID_VALUE, and so is a negative width or
+    // height. `first + count == MAX_VIEWPORTS` is LEGAL - only strictly greater is an error,
+    // which KHR-GL43.viewport_array.api_errors checks explicitly in both directions.
+    static Bool ValidateViewportIndex(GLuint index, const char* functionName) {
+        if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                         "Viewport index " + std::to_string(index) +
+                                             " is out of range. Max supported is " +
+                                             std::to_string(RenderStateParameters::MAX_VIEWPORTS - 1) + "."));
+        return false;
+    }
+
+    static Bool ValidateViewportRange(GLuint first, GLsizei count, const char* functionName) {
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "count must not be negative."));
+            return false;
+        }
+        // Widened before adding: first is a GLuint and count a GLsizei, so `first + count` in
+        // 32 bits can wrap past MAX_VIEWPORTS and let an out-of-range range through.
+        const Uint64 last = static_cast<Uint64>(first) + static_cast<Uint64>(count);
+        if (last > RenderStateParameters::MAX_VIEWPORTS) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "first (" + std::to_string(first) + ") + count (" +
+                                                 std::to_string(count) + ") exceeds GL_MAX_VIEWPORTS (" +
+                                                 std::to_string(RenderStateParameters::MAX_VIEWPORTS) + ")."));
+            return false;
+        }
+        return true;
+    }
+
+    template <typename T>
+    static Bool ValidateNonNegativeExtent(T width, T height, const char* functionName) {
+        if (width >= T(0) && height >= T(0)) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "Width and height must be non-negative."));
+        return false;
+    }
+
+    // The array forms are all-or-nothing: one bad element rejects the whole call with a SINGLE
+    // GL_INVALID_VALUE and leaves every rectangle untouched. api_errors relies on both halves -
+    // it passes a full 16-element array with exactly one negative extent and then asserts the
+    // error queue holds exactly one entry.
+    template <typename T>
+    static Bool ValidateArrayExtents(GLsizei count, const T* v, const char* functionName) {
+        for (GLsizei i = 0; i < count; ++i) {
+            if (v[i * 4 + 2] >= T(0) && v[i * 4 + 3] >= T(0)) continue;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "Width and height must be non-negative (element " + std::to_string(i) +
+                                                 ")."));
+            return false;
+        }
+        return true;
+    }
+
+    static Bool ValidateNonNullArray(const void* v, const char* functionName) {
+        if (v != nullptr) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "value pointer cannot be null."));
+        return false;
+    }
+
     static Bool TryConvertBlendEquation(GLenum mode, const char* functionName,
                                         ::MobileGL::BlendEquation& outEquation) {
         outEquation = MG_Util::ConvertGLEnumToBlendEquation(mode);
@@ -93,16 +183,70 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Viewport_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Viewport_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
+        if (!ValidateNonNegativeExtent(width, height, "Viewport_State")) return;
 
         MG_State::pGLContext->SetViewport(IntVec4(x, y, width, height));
     }
 
+    // ------------------ ARB_viewport_array setters ------------------
+    void ViewportArrayv_State(GLuint first, GLsizei count, const GLfloat* v) {
+        if (!ValidateViewportRange(first, count, "ViewportArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ViewportArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ViewportArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetViewportIndexed(first + static_cast<GLuint>(i),
+                                                     FloatVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ViewportIndexedf_State(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        if (!ValidateViewportIndex(index, "ViewportIndexedf_State")) return;
+        if (!ValidateNonNegativeExtent(w, h, "ViewportIndexedf_State")) return;
+
+        MG_State::pGLContext->SetViewportIndexed(index, FloatVec4(x, y, w, h));
+    }
+
+    void ScissorArrayv_State(GLuint first, GLsizei count, const GLint* v) {
+        if (!ValidateViewportRange(first, count, "ScissorArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ScissorArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ScissorArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetScissorBoxIndexed(first + static_cast<GLuint>(i),
+                                                       IntVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ScissorIndexed_State(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        if (!ValidateViewportIndex(index, "ScissorIndexed_State")) return;
+        if (!ValidateNonNegativeExtent(width, height, "ScissorIndexed_State")) return;
+
+        MG_State::pGLContext->SetScissorBoxIndexed(index, IntVec4(left, bottom, width, height));
+    }
+
+    void DepthRangeArrayv_State(GLuint first, GLsizei count, const GLdouble* v) {
+        if (!ValidateViewportRange(first, count, "DepthRangeArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "DepthRangeArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetDepthRangeIndexed(
+                first + static_cast<GLuint>(i),
+                FloatVec2(ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 0])),
+                          ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 1]))));
+        }
+    }
+
+    void DepthRangeIndexed_State(GLuint index, GLdouble n, GLdouble f) {
+        if (!ValidateViewportIndex(index, "DepthRangeIndexed_State")) return;
+
+        MG_State::pGLContext->SetDepthRangeIndexed(
+            index, FloatVec2(ClampUnitFloat(static_cast<GLfloat>(n)), ClampUnitFloat(static_cast<GLfloat>(f))));
+    }
+
     void StencilOpSeparate_State(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         Bool applyFront = false;
         Bool applyBack = false;
@@ -175,12 +319,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Scissor_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Scissor_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
+        if (!ValidateNonNegativeExtent(width, height, "Scissor_State")) return;
 
         MG_State::pGLContext->SetScissorBox(IntVec4(x, y, width, height));
     }
@@ -336,7 +475,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     GLboolean IsEnabledi_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "IsEnabledi_State")) {
+        if (!ValidateIndexedCapability(target, index, "IsEnabledi_State")) {
             return GL_FALSE;
         }
 
@@ -392,7 +531,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = values[0] != 0 ? GL_TRUE : GL_FALSE;
+        // The ARB_viewport_array rectangles are the only multi-component indexed state that
+        // reaches here; writing element 0 alone would leave the caller's other three untouched.
+        const GLsizei components = target == GL_VIEWPORT || target == GL_SCISSOR_BOX
+            ? 4
+            : (target == GL_DEPTH_RANGE ? 2 : 1);
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = values[i] != 0 ? GL_TRUE : GL_FALSE;
+        }
     }
 
     GLboolean IsEnabled_State(GLenum cap) {
@@ -725,7 +871,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Disablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Disablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Disablei_State")) {
             return;
         }
 
@@ -743,7 +889,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Enablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Enablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Enablei_State")) {
             return;
         }
 
@@ -797,6 +943,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         Viewport_State(x, y, width, height);
     }
 
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v) {
+        ViewportArrayv_State(first, count, v);
+    }
+
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        ViewportIndexedf_State(index, x, y, w, h);
+    }
+
+    void ViewportIndexedfv(GLuint index, const GLfloat* v) {
+        // The index is validated before the pointer is touched: glViewportIndexedfv(MAX, nullptr)
+        // must be one GL_INVALID_VALUE, not a null dereference.
+        if (!ValidateViewportIndex(index, "ViewportIndexedfv")) return;
+        if (!ValidateNonNullArray(v, "ViewportIndexedfv")) return;
+        ViewportIndexedf_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v) {
+        ScissorArrayv_State(first, count, v);
+    }
+
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        ScissorIndexed_State(index, left, bottom, width, height);
+    }
+
+    void ScissorIndexedv(GLuint index, const GLint* v) {
+        if (!ValidateViewportIndex(index, "ScissorIndexedv")) return;
+        if (!ValidateNonNullArray(v, "ScissorIndexedv")) return;
+        ScissorIndexed_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v) {
+        DepthRangeArrayv_State(first, count, v);
+    }
+
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f) {
+        DepthRangeIndexed_State(index, n, f);
+    }
+
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         StencilOpSeparate_State(face, sfail, dpfail, dppass);
     }

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.h

@@ -20,6 +20,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     void Enablei(GLenum target, GLuint index);
     void BlendFunc(GLenum sfactor, GLenum dfactor);
     void Viewport(GLint x, GLint y, GLsizei width, GLsizei height);
+    // ARB_viewport_array (core since GL 4.1). Every one of these addresses the same 16-element
+    // indexed state the classic glViewport/glScissor/glDepthRange trio broadcasts to.
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v);
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h);
+    void ViewportIndexedfv(GLuint index, const GLfloat* v);
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v);
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height);
+    void ScissorIndexedv(GLuint index, const GLint* v);
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v);
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f);
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass);
     void StencilOp(GLenum fail, GLenum zfail, GLenum zpass);
     void StencilMaskSeparate(GLenum face, GLuint mask);

MobileGL/MG_Impl/GLImpl/Sync/GL_Sync.cpp

@@ -14,10 +14,29 @@ namespace MobileGL::MG_Impl::GLImpl {
         // Frontend sync object: wraps an optional backend fence handle. A null
         // backend handle (backend has no fence support, or could not create a
         // fence at call time) keeps the legacy always-signaled behavior.
+        //
+        // SharedPtr-owned, not raw: DeleteSync can remove the registry entry while
+        // another thread is inside ClientWaitSync/GetSynciv. Those callers hold a
+        // SharedPtr copy, so the object stays alive until the last reader leaves.
+        // `mutex` then serializes backend-handle reads against the one-time
+        // backend-handle release performed by DeleteSync / DestroyAllSyncObjects.
         struct SyncObject {
+            std::mutex mutex;
             MG_Backend::BackendSyncHandle backendHandle = nullptr;
             GLenum condition = GL_SYNC_GPU_COMMANDS_COMPLETE;
             GLbitfield flags = 0;
+
+            void ReleaseBackendHandle() {
+                const std::lock_guard<std::mutex> lock(mutex);
+                if (backendHandle == nullptr) {
+                    return;
+                }
+                const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
+                if (backendDeleteSync) {
+                    backendDeleteSync(backendHandle);
+                }
+                backendHandle = nullptr;
+            }
         };
 
         // Sync calls may arrive from any thread (launchers migrate the context
@@ -25,9 +44,9 @@ namespace MobileGL::MG_Impl::GLImpl {
         // Entries left at process shutdown are simply dropped; their backend
         // handles die with the backend.
         std::mutex g_syncObjectsMutex;
-        UnorderedMap<GLsync, SyncObject*> g_liveSyncObjects;
+        UnorderedMap<GLsync, SharedPtr<SyncObject>> g_liveSyncObjects;
 
-        SyncObject* FindSyncObject(GLsync sync) {
+        SharedPtr<SyncObject> FindSyncObject(GLsync sync) {
             const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
             const auto it = g_liveSyncObjects.find(sync);
             return it != g_liveSyncObjects.end() ? it->second : nullptr;
@@ -35,13 +54,13 @@ namespace MobileGL::MG_Impl::GLImpl {
     } // namespace
 
     GLsync FenceSync(GLenum condition, GLbitfield flags) {
-        auto* syncObject = new SyncObject;
+        auto syncObject = MakeShared<SyncObject>();
         syncObject->condition = condition;
         syncObject->flags = flags;
         if (const auto backendFenceSync = MG_Backend::gBackendFunctionsTable.GL.FenceSync) {
             syncObject->backendHandle = backendFenceSync();
         }
-        const GLsync handle = reinterpret_cast<GLsync>(syncObject);
+        const GLsync handle = reinterpret_cast<GLsync>(syncObject.get());
         const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
         g_liveSyncObjects[handle] = syncObject;
         return handle;
@@ -52,24 +71,31 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     GLenum ClientWaitSync(GLsync sync, GLbitfield flags, GLuint64 timeout) {
-        const auto* syncObject = FindSyncObject(sync);
+        const SharedPtr<SyncObject> syncObject = FindSyncObject(sync);
         if (!syncObject) {
             return GL_WAIT_FAILED;
         }
         const auto backendClientWaitSync = MG_Backend::gBackendFunctionsTable.GL.ClientWaitSync;
-        if (!backendClientWaitSync || !syncObject->backendHandle) {
+        // Hold the per-object lock across the backend call: a concurrent
+        // DeleteSync may already have removed this object from the registry, but
+        // it cannot free the backend handle (or the wrapper) until this reader
+        // finishes. ClientWaitSync can block for `timeout`; that blocks only this
+        // sync object, never the registry or unrelated syncs.
+        const std::lock_guard<std::mutex> lock(syncObject->mutex);
+        if (!backendClientWaitSync || syncObject->backendHandle == nullptr) {
             return GL_ALREADY_SIGNALED; // legacy always-signaled fallback
         }
         return backendClientWaitSync(syncObject->backendHandle, flags, timeout);
     }
 
     void WaitSync(GLsync sync, GLbitfield flags, GLuint64 timeout) {
-        const auto* syncObject = FindSyncObject(sync);
+        const SharedPtr<SyncObject> syncObject = FindSyncObject(sync);
         if (!syncObject) {
             return;
         }
         const auto backendWaitSync = MG_Backend::gBackendFunctionsTable.GL.WaitSync;
-        if (backendWaitSync && syncObject->backendHandle) {
+        const std::lock_guard<std::mutex> lock(syncObject->mutex);
+        if (backendWaitSync && syncObject->backendHandle != nullptr) {
             backendWaitSync(syncObject->backendHandle, flags, timeout);
         }
     }
@@ -78,7 +104,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (sync == nullptr) {
             return; // glDeleteSync(0) is silently ignored
         }
-        SyncObject* syncObject = nullptr;
+        SharedPtr<SyncObject> syncObject;
         {
             const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
             const auto it = g_liveSyncObjects.find(sync);
@@ -88,15 +114,14 @@ namespace MobileGL::MG_Impl::GLImpl {
             syncObject = it->second;
             g_liveSyncObjects.erase(it);
         }
-        const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
-        if (backendDeleteSync && syncObject->backendHandle) {
-            backendDeleteSync(syncObject->backendHandle);
-        }
-        delete syncObject;
+        // Release the backend handle under the object lock. The local SharedPtr
+        // (and any reader's SharedPtr) keeps the wrapper itself alive until every
+        // in-flight backend call has returned.
+        syncObject->ReleaseBackendHandle();
     }
 
     void GetSynciv(GLsync sync, GLenum pname, GLsizei bufSize, GLsizei* length, GLint* values) {
-        const auto* syncObject = FindSyncObject(sync);
+        const SharedPtr<SyncObject> syncObject = FindSyncObject(sync);
         if (!syncObject) {
             if (length) {
                 *length = 0;
@@ -111,7 +136,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             break;
         case GL_SYNC_STATUS: {
             const auto backendGetSyncStatus = MG_Backend::gBackendFunctionsTable.GL.GetSyncStatus;
-            const Bool signaled = !backendGetSyncStatus || !syncObject->backendHandle ||
+            const std::lock_guard<std::mutex> lock(syncObject->mutex);
+            const Bool signaled = !backendGetSyncStatus || syncObject->backendHandle == nullptr ||
                                   backendGetSyncStatus(syncObject->backendHandle);
             value = signaled ? GL_SIGNALED : GL_UNSIGNALED;
             break;
@@ -137,11 +163,10 @@ namespace MobileGL::MG_Impl::GLImpl {
     void DestroyAllSyncObjects() {
         // Detach the registry under the lock, release outside it. Entries the app
         // already deleted were erased by DeleteSync, so nothing here double-frees;
-        // a DeleteSync racing this sweep finds an empty registry and returns. A
-        // thread still blocked inside ClientWaitSync/GetSynciv during teardown
-        // holds a raw SyncObject* these deletes invalidate - the same undefined
-        // race an app-driven DeleteSync already has.
-        UnorderedMap<GLsync, SyncObject*> orphans;
+        // a DeleteSync racing this sweep finds an empty registry and returns.
+        // Readers racing this sweep keep their SharedPtr copy alive, and each
+        // object's own lock makes the backend-handle release wait for them.
+        UnorderedMap<GLsync, SharedPtr<SyncObject>> orphans;
         {
             const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
             orphans.swap(g_liveSyncObjects);
@@ -153,12 +178,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         // context/renderer is gone (generation/current-thread guards), so this is
         // safe after the backend has released its EGL resources - but not after
         // the function table itself is cleared.
-        const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
         for (const auto& [_, syncObject] : orphans) {
-            if (backendDeleteSync && syncObject->backendHandle) {
-                backendDeleteSync(syncObject->backendHandle);
+            if (syncObject) {
+                syncObject->ReleaseBackendHandle();
             }
-            delete syncObject;
         }
         MGLOG_D("DestroyAllSyncObjects: reclaimed %zu sync object(s) the app left undeleted", orphans.size());
     }

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -3382,12 +3382,84 @@ namespace MobileGL::MG_Impl::GLImpl {
                          dstY, dstZ, srcWidth, srcHeight, srcDepth);
     }
 
+    namespace {
+        // The eleven targets GL 4.6 core 18.3.2 accepts. GL_TEXTURE_BUFFER, the six cube FACE
+        // enums and every PROXY enum all convert to a TextureTarget this frontend recognises,
+        // so ValidateTextureTarget lets them through; here they are INVALID_ENUM.
+        Bool ValidateCopyImageTarget(GLenum target, const char* endpointName) {
+            switch (target) {
+            case GL_RENDERBUFFER:
+            case GL_TEXTURE_1D:
+            case GL_TEXTURE_1D_ARRAY:
+            case GL_TEXTURE_2D:
+            case GL_TEXTURE_2D_ARRAY:
+            case GL_TEXTURE_2D_MULTISAMPLE:
+            case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+            case GL_TEXTURE_3D:
+            case GL_TEXTURE_CUBE_MAP:
+            case GL_TEXTURE_CUBE_MAP_ARRAY:
+            case GL_TEXTURE_RECTANGLE:
+                return true;
+            default:
+                break;
+            }
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("{} is not a target glCopyImageSubData accepts as the {}.",
+                                MG_Util::ConvertGLEnumToString(target), endpointName)));
+            return false;
+        }
+
+        IntVec3 GetCopyImageLevelSize(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                      TextureUploadTarget uploadTarget, GLint level) {
+            const auto* mipmapTexture = MG_State::GLState::AsMipmapTexture(textureObject.get());
+            if (!mipmapTexture) return textureObject->GetBaseSize();
+            return mipmapTexture->GetMipmapTexelSize(uploadTarget, static_cast<Uint>(level));
+        }
+
+        // glCopyImageSubData names an object that must already exist, and GL 4.6 core 18.3.2
+        // spells the failure INVALID_VALUE - "if either name does not correspond to a valid
+        // object". The shared ValidateTextureObject says INVALID_OPERATION, which is right for
+        // the ~30 entry points that reach it through a BOUND object (where the name was never
+        // in question and the fault is the binding), so this is a local rule rather than a
+        // change to the helper.
+        Bool ValidateCopyImageObjectExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                           const char* endpointName) {
+            if (textureObject) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} name does not correspond to an existing image object.", endpointName)));
+            return false;
+        }
+
+        // Same split for the target/object disagreement: GL 4.6 core 18.3.2 makes a target that
+        // does not match the object INVALID_ENUM, where the shared uniformity helper records
+        // INVALID_OPERATION for the upload paths that share it.
+        Bool ValidateCopyImageTargetMatchesObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                                  TextureTarget target, const char* endpointName) {
+            if (!textureObject || textureObject->GetTarget() == target) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} target {} does not match the target the object was created with ({}).",
+                                endpointName, MG_Util::ConvertTextureTargetToString(target),
+                                MG_Util::ConvertTextureTargetToString(textureObject->GetTarget()))));
+            return false;
+        }
+    } // namespace
+
     Bool ValidateCopyImageSubData_State(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture,
-                                        GLenum srcTarget, GLint srcLevel,
+                                        GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY,
                                         const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
-                                        GLenum dstTarget, GLint dstLevel,
+                                        GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY,
                                         GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        if (!TextureImpl::ValidateTextureObject(srcTexture) || !TextureImpl::ValidateTextureObject(dstTexture)) {
+        if (!ValidateCopyImageObjectExists(srcTexture, "source") ||
+            !ValidateCopyImageObjectExists(dstTexture, "destination")) {
             return false;
         }
         const auto srcTextureTarget = MG_Util::ConvertGLEnumToTextureTarget(srcTarget);
@@ -3396,8 +3468,13 @@ namespace MobileGL::MG_Impl::GLImpl {
             !TextureImpl::ValidateTextureTarget(dstTextureTarget)) {
             return false;
         }
-        if (!TextureImpl::ValidateTextureTargetUniformity(srcTexture, srcTextureTarget) ||
-            !TextureImpl::ValidateTextureTargetUniformity(dstTexture, dstTextureTarget)) {
+        // GL_TEXTURE_BUFFER and the cube FACE enums convert to a target this frontend knows, but
+        // 18.3.2 does not accept them here - only the eleven whole-image targets do.
+        if (!ValidateCopyImageTarget(srcTarget, "source") || !ValidateCopyImageTarget(dstTarget, "destination")) {
+            return false;
+        }
+        if (!ValidateCopyImageTargetMatchesObject(srcTexture, srcTextureTarget, "source") ||
+            !ValidateCopyImageTargetMatchesObject(dstTexture, dstTextureTarget, "destination")) {
             return false;
         }
         if (!TextureImpl::ValidateTextureLevelNumber(srcLevel) ||
@@ -3425,7 +3502,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (srcWidth == 0 || srcHeight == 0 || srcDepth == 0) {
             return false;
         }
-        if (!TextureImpl::ValidateBaseInternalFormatMatch(srcTexture->GetFormat(), dstTexture->GetFormat())) {
+        // A multisample image can only be copied to one with the same sample count, and a
+        // single-sample image reports zero - so this one comparison is also what rejects
+        // copying between a multisample target and a non-multisample one.
+        if (srcTexture->GetSamples() != dstTexture->GetSamples()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("The two images have different sample counts ({} vs. {}).",
+                                srcTexture->GetSamples(), dstTexture->GetSamples())));
+            return false;
+        }
+        // 18.3.2: both images must be complete. An incomplete one has no defined texels to copy
+        // and no defined storage to copy into.
+        if (!srcTexture->IsComplete() || !dstTexture->IsComplete()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("A copied image is incomplete (source complete: {}, destination complete: {}).",
+                                srcTexture->IsComplete(), dstTexture->IsComplete())));
+            return false;
+        }
+        const auto srcUploadTarget = GetPrimaryUploadTarget(srcTexture);
+        const auto dstUploadTarget = GetPrimaryUploadTarget(dstTexture);
+        const auto srcBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            srcTexture->GetFormat(), GetCompressedLevelFormat(srcTexture, srcUploadTarget, srcLevel));
+        const auto dstBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            dstTexture->GetFormat(), GetCompressedLevelFormat(dstTexture, dstUploadTarget, dstLevel));
+        if (!TextureImpl::ValidateCopyImageFormatCompatibility(srcBlock, dstBlock)) {
+            return false;
+        }
+        const IntVec3 srcLevelSize = GetCopyImageLevelSize(srcTexture, srcUploadTarget, srcLevel);
+        const IntVec3 dstLevelSize = GetCopyImageLevelSize(dstTexture, dstUploadTarget, dstLevel);
+        if (!TextureImpl::ValidateCopyImageBlockAlignment(srcBlock, srcX, srcY, srcWidth, srcHeight,
+                                                          srcLevelSize.x(), srcLevelSize.y(), "source") ||
+            !TextureImpl::ValidateCopyImageBlockAlignment(dstBlock, dstX, dstY, srcWidth, srcHeight,
+                                                          dstLevelSize.x(), dstLevelSize.y(), "destination")) {
             return false;
         }
         return true;
@@ -5600,10 +5714,15 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CopyImageSubData(GLuint srcName, GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
                           GLuint dstName, GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ,
                           GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        auto srcTexture = GetTextureObjectByName(srcName, __func__);
-        auto dstTexture = GetTextureObjectByName(dstName, __func__);
-        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, dstTexture, dstTarget, dstLevel,
-                                            srcWidth, srcHeight, srcDepth)) {
+        // A missing name is INVALID_VALUE here, where GetTextureObjectByName's own diagnostic is
+        // INVALID_OPERATION - so resolve through the plain lookup, which answers a null
+        // SharedPtr, and let the validator record the error this entry point owes.
+        const SharedPtr<MG_State::GLState::ITextureObject> srcTexture =
+            MG_State::pGLContext->GetTextureObject(srcName);
+        const SharedPtr<MG_State::GLState::ITextureObject> dstTexture =
+            MG_State::pGLContext->GetTextureObject(dstName);
+        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, srcX, srcY, dstTexture, dstTarget,
+                                            dstLevel, dstX, dstY, srcWidth, srcHeight, srcDepth)) {
             return;
         }
         CopyImageSubData_Backend(srcTexture, srcTarget, srcLevel, srcX, srcY, srcZ, dstTexture, dstTarget, dstLevel,

MobileGL/MG_Impl/GLImpl/Texture/Validators.cpp

@@ -15,6 +15,7 @@
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToMG/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToStr/TextureEnumConverter.h>
+#include <MG_Util/Metrics/TextureMetrics.h>
 
 namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
     Bool ValidateTextureTarget(TextureTarget target) {
@@ -515,26 +516,86 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         }
     } // namespace
 
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2) {
-        const auto unsizedFormat1 = MG_Util::ConvertInternalFormatToUnsized(format1);
-        const auto unsizedFormat2 = MG_Util::ConvertInternalFormatToUnsized(format2);
-        if (unsizedFormat1 != unsizedFormat2) {
-            // The 3-argument GenericErrorInfo constructor used to be spelled as a single
-            // std::format() call whose format string was the component name, so every
-            // diagnostic collapsed to the literal "MG_Impl/GLImpl". Format the message, then
-            // hand over component/function/message separately.
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat) {
+        CopyImageTexelBlock block{};
+        if (compressedFormat != GL_NONE) {
+            const auto info = MG_Util::GetCompressedFormatInfo(compressedFormat);
+            if (info.blockByteSize != 0) {
+                block.byteSize = info.blockByteSize;
+                block.blockWidth = info.blockWidth;
+                block.blockHeight = info.blockHeight;
+                block.compressed = true;
+                return block;
+            }
+        }
+        // The size MobileGL actually stores a texel of this format in, which for every format GL
+        // gives a required size is that required size. The handful of legacy formats GL leaves
+        // implementation-defined (R3_G3_B2, RGB4/5/10/12, RGBA2/12) have no view class in table
+        // 8.22 to be compared against anyway, and this is the size that decides whether a raw
+        // copy between them would in fact preserve the bytes.
+        block.byteSize = MG_Util::GetSizedInternalFormatSizeInBytes(format);
+        return block;
+    }
+
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock) {
+        if (srcBlock.byteSize == 0 || dstBlock.byteSize == 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                                             "A copied image has no storage whose texel size is known."));
+            return false;
+        }
+        if (srcBlock.byteSize != dstBlock.byteSize) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
                 MakeUnique<GenericErrorInfo>(
-                    "MG_Impl/GLImpl", "ValidateBaseInternalFormatMatch",
-                    std::format("The base internal format of the two formats do not match ({} vs. {})",
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat1),
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat2))));
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                    std::format("The two images' texel blocks are different sizes ({} vs. {} bytes), so the "
+                                "formats are not copy-compatible.",
+                                srcBlock.byteSize, dstBlock.byteSize)));
+            return false;
+        }
+        // Two compressed images additionally have to agree on the SHAPE of the block, not only
+        // its size: an 8-byte 4x4 block and a hypothetical 8-byte 8x8 one hold different texel
+        // counts, and GL 4.6 core 18.3.2 requires both dimensions to match.
+        if (srcBlock.compressed && dstBlock.compressed &&
+            (srcBlock.blockWidth != dstBlock.blockWidth || srcBlock.blockHeight != dstBlock.blockHeight)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                    std::format("The two compressed images have different block dimensions ({}x{} vs. {}x{}).",
+                                srcBlock.blockWidth, srcBlock.blockHeight, dstBlock.blockWidth,
+                                dstBlock.blockHeight)));
             return false;
         }
         return true;
     }
 
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName) {
+        if (!block.compressed) return true;
+        const Int blockWidth = static_cast<Int>(block.blockWidth);
+        const Int blockHeight = static_cast<Int>(block.blockHeight);
+        if (blockWidth <= 1 && blockHeight <= 1) return true;
+        // The origin is unconditional; the extent gets the "or it reaches the edge of the image"
+        // exemption GL 4.6 core 18.3.2 grants, which is what lets a 16x16 BPTC image be copied
+        // whole even when the last block is partial.
+        const Bool originAligned = (x % blockWidth == 0) && (y % blockHeight == 0);
+        const Bool widthOk = (width % blockWidth == 0) || (x + width == imageWidth);
+        const Bool heightOk = (height % blockHeight == 0) || (y + height == imageHeight);
+        if (originAligned && widthOk && heightOk) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>(
+                "MG_Impl/GLImpl", "ValidateCopyImageBlockAlignment",
+                std::format("The {} region [{}, {}] + [{} x {}] is not aligned to the {}x{} compressed block "
+                            "grid of a {} x {} image.",
+                            endpointName, x, y, width, height, blockWidth, blockHeight, imageWidth, imageHeight)));
+        return false;
+    }
+
     Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat) {
         const auto unsizedDest = MG_Util::ConvertInternalFormatToUnsized(destFormat);
         const auto unsizedSrc = MG_Util::ConvertInternalFormatToUnsized(srcFormat);

MobileGL/MG_Impl/GLImpl/Texture/Validators.h

@@ -50,8 +50,32 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
                                          TextureTarget target);
     Bool ValidateTextureSubImageOffsets(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int xoffset,
                                         Int width, Int yoffset = 0, Int height = 0, Int zoffset = 0, Int depth = 0);
-    // Exact base-format equality - what glCopyImageSubData's format compatibility needs.
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2);
+    // The texel block of one glCopyImageSubData endpoint, resolved to the two things the
+    // compatibility rule actually asks about. `compressed` is not redundant with a block bigger
+    // than 1x1: it is what distinguishes "compressed, and so the region is measured in texels of
+    // a blocked image" from "uncompressed, and so it is measured in texels".
+    struct CopyImageTexelBlock {
+        SizeT byteSize = 0;
+        Uint blockWidth = 1;
+        Uint blockHeight = 1;
+        Bool compressed = false;
+    };
+    // `compressedFormat` is the GLenum a glCompressedTexImage* upload recorded for the level, or
+    // GL_NONE. It has to be asked for separately because MobileGL stores every compressed format
+    // in uncompressed storage (ConvertGLEnumToTextureInternalFormat), so the TextureInternalFormat
+    // alone can no longer tell a BPTC image from the RGBA8 backing it.
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat);
+    // GL 4.6 core 18.3.2: the two images must be COMPATIBLE, and compatible means their texel
+    // blocks are the same SIZE - not that they share a base internal format. RGBA32UI into
+    // RGBA32F is legal (both 128-bit) while RGBA8 into RGBA32F is not, and a compressed image
+    // pairs with an uncompressed one whose texel is as big as the compressed block.
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock);
+    // GL 4.6 core 18.3.2: for a compressed image the region's origin must sit on a block
+    // boundary and its size must be a whole number of blocks - unless the edge it runs to is
+    // the edge of the image.
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName);
     // GL 4.6 SS 8.6 subset rule for glCopyTexImage*: the read buffer must supply every component
     // the requested internalformat asks for, but may supply more.
     Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat);

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -63,6 +63,7 @@ add_executable(MobileGLIntegrationTest
     Scenarios/DepthStencilReadbackMatrixScenario.cpp
     Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp
     Scenarios/ClipDistanceScenario.cpp
+    Scenarios/ViewportArrayScenario.cpp
     Scenarios/SsboArrayLengthScenario.cpp
     Scenarios/DoublePrecisionScenario.cpp
     Scenarios/UniformInitializerScenario.cpp
@@ -80,6 +81,7 @@ add_executable(MobileGLIntegrationTest
     Scenarios/VertexArrayEnableDisableScenario.cpp
     Scenarios/CopyImageLevelRangeScenario.cpp
     Scenarios/CopyImageLayeredScenario.cpp
+    Scenarios/LayeredAttachmentBarrierScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE

MobileGL/MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp

@@ -199,5 +199,61 @@ namespace MGITest {
                    "derived component limits are computed in";
         }
 
+        // ARB_viewport_array's own limits. They are advertised from three different places -
+        // GL_MAX_VIEWPORTS from the frontend's indexed state width, the bounds range and the
+        // subpixel bits from the backend caps table - and each backend fills that table from a
+        // different source, so all three are checked on both lanes.
+        //
+        // GL_VIEWPORT_BOUNDS_RANGE is the one that shipped wrong: GLES has no such query, the
+        // DirectGLES loader's glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE) therefore raised
+        // GL_INVALID_ENUM and left the probe's zero-initialized array in place, and MobileGL
+        // advertised [0, 0] - a range that admits no viewport origin at all, and the check that
+        // kept KHR-GL43.viewport_array.queries red on Espryt after the indexed-state work.
+        TEST_F(AdvertisedLimitsScenario, ViewportArrayLimitsMeetTheirGL43Floors) {
+            GLint maxViewports = -1;
+            glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(maxViewports, 16) << "GL 4.3 core table 23.53 sets the MAX_VIEWPORTS minimum at 16";
+            EXPECT_LE(maxViewports, 256) << "one viewport rectangle of indexed state is allocated per advertised "
+                                            "viewport, and the CTS sizes its arrays off this number";
+
+            GLfloat boundsRange[2] = {1.0f, -1.0f};
+            glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE, boundsRange);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_LE(boundsRange[0], -32768.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+            EXPECT_GE(boundsRange[1], 32767.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+
+            // KNOWN INFIDELITY, pinned here rather than hidden. MobileGL reports the driver's own
+            // VIEWPORT_SUBPIXEL_BITS (4 on llvmpipe, i.e. 1/16-pixel viewport precision), but the
+            // float viewport rectangle glViewportIndexedf stores is snapped to integers on its
+            // way to both backends (ComputeGLViewport, DirectGLES SyncRenderState). The STATE
+            // round trip is exact - which is all KHR-GL43.viewport_array.viewport_api checks, and
+            // all this cluster set out to fix - so the gap is in rasterization only: a fractional
+            // viewport origin rasterizes as if it had been rounded. Nothing in the suite or in
+            // Minecraft sets one. Only the spec floor is asserted; tightening this to EQ(0) would
+            // mean advertising no subpixel precision at all, which is a separate decision about a
+            // limit MobileGL currently passes through from the driver.
+            GLint subpixelBits = -1;
+            glGetIntegerv(GL_VIEWPORT_SUBPIXEL_BITS, &subpixelBits);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(subpixelBits, 0) << "GL 4.6 core table 23.60: VIEWPORT_SUBPIXEL_BITS has a minimum of 0, and "
+                                          "a negative value is what a sign-flipped uint32 looks like";
+
+            GLint viewportDims[2] = {-1, -1};
+            glGetIntegerv(GL_MAX_VIEWPORT_DIMS, viewportDims);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            GLint maxRenderbufferSize = -1;
+            glGetIntegerv(GL_MAX_RENDERBUFFER_SIZE, &maxRenderbufferSize);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            // GL 4.6 core 13.6.1: MAX_VIEWPORT_DIMS must be at least as large as the largest
+            // renderable surface, or a full-size framebuffer could not be fully viewported.
+            EXPECT_GE(viewportDims[0], maxRenderbufferSize);
+            EXPECT_GE(viewportDims[1], maxRenderbufferSize);
+        }
+
     } // namespace
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.cpp

@@ -0,0 +1,378 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - A TRANSFER OFF A NON-ZERO ATTACHMENT LAYER READS THE LAYER THE BARRIER MOVED.
+//
+// Every transfer DirectVulkan performs against a framebuffer attachment is three commands: a
+// barrier that puts the image in TRANSFER_SRC/DST, the copy or blit itself, and a barrier that
+// puts it back. The copy names the attachment's layer - glFramebufferTextureLayer(.., layer) ends
+// up in `srcSubresource.baseArrayLayer` - but TransitionImageLayout used to emit `layerCount = 1`
+// from `baseArrayLayer 0`, so for every attachment on a layer above zero the barrier moved layer 0
+// and the copy read layer N. The layer the transfer touched was never transitioned: it sat in
+// COLOR_ATTACHMENT_OPTIMAL (or DEPTH_STENCIL_ATTACHMENT_OPTIMAL) while being read as TRANSFER_SRC.
+//
+// That is undefined behaviour, not a guaranteed wrong pixel: a layout is a compression/tiling
+// promise, so a driver that stores both layouts identically returns the right bytes anyway. The
+// software lanes (lavapipe) are exactly such a driver, which is why this scenario is paired with a
+// validation-layer run - the layer names the mismatch outright
+// (VUID-vkCmdCopyImageToBuffer-srcImageLayout-00189, "srcImageLayout ... doesn't match the actual
+// current layout") where the pixels here cannot. On a tiler that really does re-tile per layout,
+// these are the reads that come back as garbage.
+//
+// The four cases below are the four transfer paths that take an attachment layer from GL:
+//
+//   glReadPixels (colour)                    -> VulkanRenderer::ReadPixels
+//   glBlitFramebuffer (colour)               -> VulkanRenderer::BlitNamedFramebuffer
+//   glReadPixels (GL_DEPTH_COMPONENT)        -> VulkanRenderer::ReadDepthStencilImageToClient
+//   glBlitFramebuffer (GL_DEPTH_BUFFER_BIT)  -> VulkanRenderer::BlitNamedFramebuffer, depth leg
+//
+// Each one renders or clears INTO the non-zero layer first, so the image is genuinely sitting in
+// its attachment layout when the transfer starts - a scenario that only uploaded texels would
+// leave it in a transfer layout already and the mismatched barrier would be a no-op.
+//
+// Every case also asserts the layers it did not name still hold their own fill, so a backend that
+// "fixed" the miss by transferring the whole image passes neither half.
+//
+// DirectGLES is the control: it hands the same calls to the driver, so a failure on both backends
+// means the scenario is wrong and a failure on DirectVulkan alone means Magma is.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr int kWidth = 8;
+        constexpr int kHeight = 8;
+        // Four layers with the subject at index 2: layers on both sides of it stay untouched, so
+        // "moved the whole image" and "moved layer 0" are both distinguishable from correct.
+        constexpr int kLayers = 4;
+        constexpr int kSubjectLayer = 2;
+
+        // A value no correct read can produce, so "the backend wrote nothing" fails loudly.
+        constexpr float kDepthPoison = 0.2f;
+
+        std::string Describe(const Rgba8& color) {
+            return "(" + std::to_string(color.r) + ", " + std::to_string(color.g) + ", " + std::to_string(color.b) +
+                   ", " + std::to_string(color.a) + ")";
+        }
+
+        // Per-layer fill, uniform within a layer: the defect is about WHICH layer is addressed, and
+        // a value that also varied inside the layer would make the assertions depend on row order.
+        Rgba8 LayerFill(int layer) {
+            return {static_cast<GLubyte>(17 + layer * 30), static_cast<GLubyte>(200 - layer * 25),
+                    static_cast<GLubyte>(60 + layer * 40), 255};
+        }
+
+        // What the draw paints - matches kFS below, and is deliberately none of the LayerFill
+        // values so "the draw never landed" cannot read as a pass.
+        constexpr Rgba8 kPaintedColor{26, 51, 204, 255};
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() { gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+
+        constexpr const char* kFS = R"(#version 330 core
+out vec4 o_color;
+void main() { o_color = vec4(0.1, 0.2, 0.8, 1.0); }
+)";
+
+        void DrawFullViewportQuad(unsigned int program) {
+            static const float kQuad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0, vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(kQuad), kQuad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+        }
+
+        class LayeredAttachmentBarrierScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVS, kFS, &error);
+                ASSERT_NE(m_program, 0u) << error;
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                for (const GLuint fbo : m_fbos) {
+                    glDeleteFramebuffers(1, &fbo);
+                }
+                m_fbos.clear();
+                for (const GLuint texture : m_textures) {
+                    glDeleteTextures(1, &texture);
+                }
+                m_textures.clear();
+                if (m_program != 0) {
+                    glUseProgram(0);
+                    glDeleteProgram(m_program);
+                    m_program = 0;
+                }
+            }
+
+            // An RGBA8 2D array with a different uniform colour per layer.
+            GLuint MakeColorArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA8, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    const std::vector<Rgba8> texels(static_cast<std::size_t>(kWidth) * kHeight, LayerFill(layer));
+                    glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, layer, kWidth, kHeight, 1, GL_RGBA,
+                                    GL_UNSIGNED_BYTE, texels.data());
+                }
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // A depth 2D array. No initial upload: depth arrays are filled by clearing through an
+            // attachment, which is also the state the transfer paths have to cope with.
+            GLuint MakeDepthArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH_COMPONENT24, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // One FBO naming `layer` of the given arrays. Depth is optional (0 = colour only).
+            GLuint MakeLayerFbo(GLuint colorArray, GLuint depthArray, int layer) {
+                GLuint fbo = 0;
+                glGenFramebuffers(1, &fbo);
+                m_fbos.push_back(fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, colorArray, 0, layer);
+                if (depthArray != 0) {
+                    glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthArray, 0, layer);
+                }
+                EXPECT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                    << "layer " << layer << " is not attachable";
+                return fbo;
+            }
+
+            // glReadPixels of one whole layer, through an FBO that names it.
+            Rgba8 ReadLayer(GLuint colorArray, int layer) {
+                const GLuint fbo = MakeLayerFbo(colorArray, 0, layer);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glReadBuffer(GL_COLOR_ATTACHMENT0);
+                glPixelStorei(GL_PACK_ALIGNMENT, 1);
+                std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+                glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                // The fill is uniform within a layer, so any disagreement between texels is itself
+                // a failure - reported here rather than silently reduced to pixels[0].
+                for (std::size_t i = 1; i < pixels.size(); ++i) {
+                    EXPECT_TRUE(pixels[i] == pixels[0])
+                        << "layer " << layer << " is not uniform: texel 0 is " << Describe(pixels[0]) << ", texel "
+                        << i << " is " << Describe(pixels[i]);
+                }
+                return pixels[0];
+            }
+
+            // Every layer but `changed` still holds its own fill.
+            void ExpectOtherLayersUntouched(GLuint colorArray, int changed, const char* what) {
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    if (layer == changed) continue;
+                    const Rgba8 actual = ReadLayer(colorArray, layer);
+                    EXPECT_TRUE(actual == LayerFill(layer))
+                        << what << ": layer " << layer << " should still hold its fill but is " << Describe(actual)
+                        << ", expected " << Describe(LayerFill(layer));
+                }
+            }
+
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            std::vector<GLuint> m_textures;
+            std::vector<GLuint> m_fbos;
+            unsigned int m_program = 0;
+        };
+
+        // glReadPixels straight off a layer that was just rendered to. The image is in
+        // COLOR_ATTACHMENT_OPTIMAL when the readback barrier runs, so the barrier and the copy
+        // disagreeing about the layer is a live layout mismatch, not a bookkeeping detail.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadPixelsOffRenderedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, 0, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glPixelStorei(GL_PACK_ALIGNMENT, 1);
+            std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+            glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            for (std::size_t i = 0; i < pixels.size(); ++i) {
+                ASSERT_NEAR(pixels[i].r, kPaintedColor.r, 2)
+                    << "texel " << i << " of the rendered layer is " << Describe(pixels[i]);
+                ASSERT_NEAR(pixels[i].g, kPaintedColor.g, 2) << "texel " << i;
+                ASSERT_NEAR(pixels[i].b, kPaintedColor.b, 2) << "texel " << i;
+            }
+
+            ExpectOtherLayersUntouched(colorArray, kSubjectLayer, "readback off a rendered layer");
+        }
+
+        // glBlitFramebuffer between two non-zero layers of two different arrays. Both endpoints are
+        // above layer 0, so the source and destination barriers are each wrong on their own side.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitBetweenNonZeroColorLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceArray = MakeColorArray();
+            const GLuint destinationArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceArray, 0, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            const GLuint destinationFbo = MakeLayerFbo(destinationArray, 0, kDestinationLayer);
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            const Rgba8 blitted = ReadLayer(destinationArray, kDestinationLayer);
+            EXPECT_NEAR(blitted.r, kPaintedColor.r, 2) << "blit destination layer is " << Describe(blitted);
+            EXPECT_NEAR(blitted.g, kPaintedColor.g, 2);
+            EXPECT_NEAR(blitted.b, kPaintedColor.b, 2);
+
+            ExpectOtherLayersUntouched(destinationArray, kDestinationLayer, "colour blit destination");
+            // The source layer was rendered, not blitted into, so it is checked separately.
+            const Rgba8 source = ReadLayer(sourceArray, kSourceLayer);
+            EXPECT_NEAR(source.r, kPaintedColor.r, 2) << "blit source layer is " << Describe(source);
+            ExpectOtherLayersUntouched(sourceArray, kSourceLayer, "colour blit source");
+        }
+
+        // The depth aspect of the same readback path: the depth image sits in
+        // DEPTH_STENCIL_ATTACHMENT_OPTIMAL after the clear, and the copy names the attached layer.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadDepthOffClearedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            const GLuint depthArray = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, depthArray, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.375);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            const float centre = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(centre, 0.375f, 1.0f / 4096.0f)
+                << "glReadPixels(GL_DEPTH_COMPONENT) off layer " << kSubjectLayer << " returned " << centre
+                << (std::fabs(centre - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+        }
+
+        // The depth leg of the blit path, both endpoints above layer 0. Verified by reading the
+        // destination's depth back, which is the same readback the case above pins - so a failure
+        // here with that one passing is the blit, not the readback.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitDepthBetweenNonZeroLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceColor = MakeColorArray();
+            const GLuint sourceDepth = MakeDepthArray();
+            const GLuint destinationColor = MakeColorArray();
+            const GLuint destinationDepth = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceColor, sourceDepth, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.625);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            // A destination pre-cleared to something the blit must overwrite, so "the blit did
+            // nothing" and "the blit landed" are different answers.
+            const GLuint destinationFbo = MakeLayerFbo(destinationColor, destinationDepth, kDestinationLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.125);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            const float blitted = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(blitted, 0.625f, 1.0f / 4096.0f)
+                << "depth blitted onto layer " << kDestinationLayer << " reads back as " << blitted
+                << (std::fabs(blitted - 0.125f) < 1e-3f ? " - the destination kept its own clear" : "");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp

@@ -0,0 +1,524 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - gl_ViewportIndex ACTUALLY ROUTES, AND THE PER-INDEX STATE IT SELECTS IS REAL.
+//
+// The state half of ARB_viewport_array is asserted in MG_Test/State/RenderStateTest.cpp, which
+// is a pure set/get exercise and would pass just as green against a backend that stores all 16
+// rectangles and rasterizes only the first. This file is the other half: every case here routes
+// primitives to a viewport OTHER than 0 and then looks at where the pixels landed.
+//
+// Three claims, one per case:
+//   1. gl_ViewportIndex selects the viewport RECTANGLE - a 4x4 grid of 32x32 viewports, one
+//      geometry-shader invocation per cell, and every cell must hold its own index.
+//   2. gl_ViewportIndex selects the DEPTH RANGE - 16 one-pixel-wide viewports whose ranges are
+//      (i/16, 1 - i/16), a quad at each end of clip space, and gl_FragCoord.z read back.
+//      This is the claim that fails loudest against a single-viewport backend, because the
+//      geometry is still in the right place while every depth comes back as viewport 0's.
+//   3. The per-index SCISSOR TEST ENABLE is honoured. Vulkan has no per-viewport scissor-test
+//      toggle, so a disabled index has to be given the whole framebuffer as its rectangle; the
+//      case draws the same primitive into the same index twice, once with the test off and once
+//      with it on, and requires the two results to differ in the documented direction.
+//
+// Case 1 runs a second time against the DEFAULT framebuffer. MobileGL Y-flips (and pre-transform
+// rotates) the default framebuffer's rectangles and does not touch an FBO's, so a port that
+// applies the flip to viewport 0 and forgets the other fifteen renders a correct-looking FBO and
+// an upside-down window - the classic multi-viewport bug, and invisible to every FBO-only case.
+//
+// HONEST LIMIT OF THIS FILE. DirectGLES SKIPS every case: GLES has one viewport, one scissor
+// rectangle and no gl_ViewportIndex, so routing to index > 0 is an emulation feature that has
+// not been built (the Espryt half of KHR-GL43.viewport_array's rendering group is deliberately
+// still red). The skip is explicit rather than silent so a future emulation lands here as a
+// failing test and not as a test that was quietly never running. DirectVulkan additionally
+// skips when the device lacks the multiViewport feature - Vulkan then forbids a pipeline from
+// declaring more than one viewport at all, which is a device limit and not a MobileGL bug;
+// lavapipe (every CI lane) and both Mali/Adreno devices support it, so the cases do run where
+// it matters.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr int kViewportCount = 16;
+        constexpr int kGridSide = 4;   // 4x4 grid of viewports
+        constexpr int kCellSize = 32;  // ... each 32x32
+        constexpr int kSurfaceSide = kGridSide * kCellSize;
+        constexpr GLint kUnwritten = -1;
+
+        // A geometry shader is the only stage GL 4.1 lets write gl_ViewportIndex, and
+        // `invocations` runs it once per viewport off a single input point - the same shape
+        // KHR-GL43.viewport_array.draw_to_single_layer_with_multiple_viewports uses.
+        const char* const kVertexSource = R"(#version 410 core
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+        const char* const kGridGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 4) out;
+flat out int gsIndex;
+void main() {
+    gsIndex = gl_InvocationID;
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        // One invocation, viewport chosen by a uniform: lets a case draw the SAME primitive into
+        // the SAME index twice under two different scissor-enable states.
+        const char* const kSingleGeometrySource = R"(#version 410 core
+layout(points, invocations = 1) in;
+layout(triangle_strip, max_vertices = 4) out;
+uniform int uViewport;
+flat out int gsIndex;
+void main() {
+    gsIndex = uViewport;
+    gl_ViewportIndex = uViewport;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kIntFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out int fragColor;
+void main() { fragColor = gsIndex; }
+)";
+
+        // Two quads, one at each end of clip space, so the fragment stage can report the depth
+        // the viewport's range mapped them to. gl_FragCoord.z IS the post-range window depth, so
+        // it reads back the per-viewport minDepth/maxDepth directly.
+        const char* const kDepthGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 8) out;
+void main() {
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  0.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, -1.0, 1.0); EmitVertex();
+    EndPrimitive();
+    gl_Position = vec4(-1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 1.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kDepthFragmentSource = R"(#version 410 core
+layout(location = 0) out float fragColor;
+void main() { fragColor = gl_FragCoord.z; }
+)";
+
+        class ViewportArrayScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                if (Gl().BackendName() == "DirectGLES") {
+                    GTEST_SKIP() << "gl_ViewportIndex routing is not emulated on DirectGLES: GLES has one viewport "
+                                    "and one scissor rectangle, so every index rasterizes as index 0. The indexed "
+                                    "STATE is still asserted (MG_Test RenderStateTest); this is the deferred "
+                                    "rendering half of KHR-GL43.viewport_array.";
+                }
+
+                GLint maxViewports = 0;
+                glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+                ASSERT_GE(maxViewports, kViewportCount) << "GL 4.3 core requires GL_MAX_VIEWPORTS >= 16";
+
+                m_program = BuildProgram(kGridGeometrySource, kIntFragmentSource);
+                ASSERT_NE(m_program, 0u) << "grid program failed to build: " << m_buildLog;
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                ResetViewportArrayState();
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready() || IsSkipped()) return;
+                ResetViewportArrayState();
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            // Every case starts from the same slate: this fixture shares its context with every
+            // other scenario in the process, and a leftover per-index scissor enable is exactly
+            // the kind of state that would make a later case pass or fail for the wrong reason.
+            static void ResetViewportArrayState() {
+                for (int i = 0; i < kViewportCount; ++i) {
+                    glDisablei(GL_SCISSOR_TEST, static_cast<GLuint>(i));
+                }
+                glDisable(GL_SCISSOR_TEST);
+                glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+                glScissor(0, 0, kSurfaceSide, kSurfaceSide);
+                glDepthRange(0.0, 1.0);
+                glDisable(GL_DEPTH_TEST);
+            }
+
+            // The 4x4 grid: viewport y*4+x covers the cell whose lower-left corner is
+            // (x*cellW, y*cellH), in GL's bottom-left-origin window coordinates. Parameterized on
+            // the cell size because the default framebuffer this scenario also renders into is
+            // deliberately non-square (HeadlessGL is 128x96, so a transposing bug cannot hide).
+            static void SetupGridViewports(int cellW, int cellH) {
+                std::vector<GLfloat> data(static_cast<size_t>(kViewportCount) * 4);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                        data[base + 0] = static_cast<GLfloat>(x * cellW);
+                        data[base + 1] = static_cast<GLfloat>(y * cellH);
+                        data[base + 2] = static_cast<GLfloat>(cellW);
+                        data[base + 3] = static_cast<GLfloat>(cellH);
+                    }
+                }
+                glViewportArrayv(0, kViewportCount, data.data());
+            }
+
+            GLuint BuildProgram(const char* geometrySource, const char* fragmentSource) {
+                const GLuint vs = CompileStage(GL_VERTEX_SHADER, kVertexSource);
+                if (vs == 0) return 0;
+                const GLuint gs = CompileStage(GL_GEOMETRY_SHADER, geometrySource);
+                if (gs == 0) {
+                    glDeleteShader(vs);
+                    return 0;
+                }
+                const GLuint fs = CompileStage(GL_FRAGMENT_SHADER, fragmentSource);
+                if (fs == 0) {
+                    glDeleteShader(vs);
+                    glDeleteShader(gs);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, vs);
+                glAttachShader(program, gs);
+                glAttachShader(program, fs);
+                glLinkProgram(program);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                glDeleteShader(vs);
+                glDeleteShader(gs);
+                glDeleteShader(fs);
+                if (!linked) {
+                    GLint length = 0;
+                    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                    std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                    glGetProgramInfoLog(program, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                    m_buildLog = log.data();
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint CompileStage(GLenum stage, const char* source) {
+                const GLuint shader = glCreateShader(stage);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled) return shader;
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                m_buildLog = log.data();
+                glDeleteShader(shader);
+                return 0;
+            }
+
+            // An R32I colour target, pre-filled with kUnwritten so "nothing was drawn here" is
+            // distinguishable from "index 0 was drawn here".
+            struct IntTarget {
+                GLuint fbo = 0;
+                GLuint texture = 0;
+            };
+
+            // The "nothing drawn here" value is UPLOADED, not cleared: the CTS fills its R32I
+            // targets the same way (fillTexture), and an upload cannot be confused with a clear
+            // that a backend defers, reorders or drops - which is exactly the ambiguity a case
+            // asserting "this cell must be untouched" cannot afford.
+            static void FillIntTarget(const IntTarget& target, int width, int height) {
+                const std::vector<GLint> unwritten(static_cast<size_t>(width) * height, kUnwritten);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED_INTEGER, GL_INT, unwritten.data());
+            }
+
+            static IntTarget MakeIntTarget(int width, int height) {
+                IntTarget target;
+                glGenTextures(1, &target.texture);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexImage2D(GL_TEXTURE_2D, 0, GL_R32I, width, height, 0, GL_RED_INTEGER, GL_INT, nullptr);
+                glGenFramebuffers(1, &target.fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, target.texture, 0);
+                FillIntTarget(target, width, height);
+                return target;
+            }
+
+            static void DestroyIntTarget(IntTarget& target) {
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                if (target.fbo != 0) glDeleteFramebuffers(1, &target.fbo);
+                if (target.texture != 0) glDeleteTextures(1, &target.texture);
+            }
+
+            static std::vector<GLint> ReadInts(int width, int height) {
+                std::vector<GLint> pixels(static_cast<size_t>(width) * height, 0);
+                glReadPixels(0, 0, width, height, GL_RED_INTEGER, GL_INT, pixels.data());
+                return pixels;
+            }
+
+            // The centre of grid cell (x, y), in the bottom-left-origin coordinates glReadPixels
+            // returns. Sampling the centre rather than a corner keeps the assertion about WHICH
+            // viewport was selected rather than about edge rounding.
+            static GLint CellCentre(const std::vector<GLint>& pixels, int stride, int x, int y) {
+                const int px = x * kCellSize + kCellSize / 2;
+                const int py = y * kCellSize + kCellSize / 2;
+                return pixels[static_cast<size_t>(py) * stride + px];
+            }
+
+            std::string m_buildLog;
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+        };
+
+        // --- 1. the viewport rectangle -------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexRasterizesIntoItsOwnRectangle) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+            SetupGridViewports(kCellSize, kCellSize);
+            glUseProgram(m_program);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const GLint expected = y * kGridSide + x;
+                    EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), expected)
+                        << "cell (" << x << ", " << y << ") should hold viewport index " << expected
+                        << "; a single-viewport backend paints the whole image with 15 (the last invocation)";
+                }
+            }
+            DestroyIntTarget(target);
+        }
+
+        // The same claim against the DEFAULT framebuffer, where MobileGL applies its Y-flip and
+        // pre-transform rotation. Index 0 alone getting the mapping is the classic bug.
+        TEST_F(ViewportArrayScenario, TheDefaultFramebufferAppliesTheSameFlipToEveryViewport) {
+            const int surfaceW = Gl().Width();
+            const int surfaceH = Gl().Height();
+            ASSERT_GE(surfaceW, kGridSide);
+            ASSERT_GE(surfaceH, kGridSide);
+            const int cellW = surfaceW / kGridSide;
+            const int cellH = surfaceH / kGridSide;
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            // Paint a value no viewport index can produce, so an unwritten cell is obvious.
+            glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
+            glClear(GL_COLOR_BUFFER_BIT);
+
+            // The default framebuffer is 8-bit RGBA, so the index travels as a colour: cell i is
+            // painted with red = i * 16, which is exact in 8 bits for i in [0, 16).
+            const char* const kColorFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = vec4(float(gsIndex) * 16.0 / 255.0, 0.0, 0.0, 1.0); }
+)";
+            const GLuint colorProgram = BuildProgram(kGridGeometrySource, kColorFragmentSource);
+            ASSERT_NE(colorProgram, 0u) << "colour program failed to build: " << m_buildLog;
+
+            SetupGridViewports(cellW, cellH);
+            glUseProgram(colorProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<unsigned char> pixels(static_cast<size_t>(surfaceW) * surfaceH * 4, 0);
+            glReadPixels(0, 0, surfaceW, surfaceH, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const int px = x * cellW + cellW / 2;
+                    const int py = y * cellH + cellH / 2;
+                    const int red = pixels[(static_cast<size_t>(py) * surfaceW + px) * 4];
+                    const int expected = (y * kGridSide + x) * 16;
+                    // One LSB of slack for an 8-bit round trip; the values are 16 apart, so this
+                    // cannot confuse two neighbouring indices.
+                    EXPECT_LE(std::abs(red - expected), 1)
+                        << "default-framebuffer cell (" << x << ", " << y << ") holds red=" << red << ", expected "
+                        << expected << ". A vertically mirrored grid means the Y-flip was applied to viewport 0 "
+                        << "only";
+                }
+            }
+            glDeleteProgram(colorProgram);
+        }
+
+        // --- 2. the depth range --------------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexUsesItsOwnDepthRange) {
+            // 16 columns one pixel wide and two rows tall: row 0 gets the near-plane quad, row 1
+            // the far-plane one, so both ends of viewport i's range land in the same column.
+            constexpr int kWidth = kViewportCount;
+            constexpr int kHeight = 2;
+
+            GLuint texture = 0;
+            GLuint fbo = 0;
+            glGenTextures(1, &texture);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, kWidth, kHeight, 0, GL_RED, GL_FLOAT, nullptr);
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
+            const GLfloat clearValue[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
+            glClearBufferfv(GL_COLOR, 0, clearValue);
+
+            std::vector<GLfloat> viewports(static_cast<size_t>(kViewportCount) * 4);
+            std::vector<GLdouble> ranges(static_cast<size_t>(kViewportCount) * 2);
+            for (int i = 0; i < kViewportCount; ++i) {
+                viewports[static_cast<size_t>(i) * 4 + 0] = static_cast<GLfloat>(i);
+                viewports[static_cast<size_t>(i) * 4 + 1] = 0.0f;
+                viewports[static_cast<size_t>(i) * 4 + 2] = 1.0f;
+                viewports[static_cast<size_t>(i) * 4 + 3] = 2.0f;
+                ranges[static_cast<size_t>(i) * 2 + 0] = static_cast<GLdouble>(i) / 16.0;
+                ranges[static_cast<size_t>(i) * 2 + 1] = 1.0 - static_cast<GLdouble>(i) / 16.0;
+            }
+            glViewportArrayv(0, kViewportCount, viewports.data());
+            glDepthRangeArrayv(0, kViewportCount, ranges.data());
+
+            const GLuint depthProgram = BuildProgram(kDepthGeometrySource, kDepthFragmentSource);
+            ASSERT_NE(depthProgram, 0u) << "depth program failed to build: " << m_buildLog;
+            glUseProgram(depthProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<GLfloat> pixels(static_cast<size_t>(kWidth) * kHeight, 0.0f);
+            glReadPixels(0, 0, kWidth, kHeight, GL_RED, GL_FLOAT, pixels.data());
+            for (int i = 0; i < kViewportCount; ++i) {
+                const float near = static_cast<float>(i) / 16.0f;
+                const float far = 1.0f - static_cast<float>(i) / 16.0f;
+                // The tolerance covers depth-buffer-free rasterization of gl_FragCoord.z on a
+                // software rasterizer; the per-index values are 1/16 apart, so it cannot let a
+                // neighbouring viewport's range through, and viewport 0's range (0, 1) differs
+                // from every other index by at least 1/16.
+                EXPECT_NEAR(pixels[i], near, 1.0e-3f)
+                    << "viewport " << i << " near-plane depth; got viewport 0's range if this is 0";
+                EXPECT_NEAR(pixels[static_cast<size_t>(kWidth) + i], far, 1.0e-3f)
+                    << "viewport " << i << " far-plane depth; got viewport 0's range if this is 1";
+            }
+
+            glDeleteProgram(depthProgram);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteTextures(1, &texture);
+        }
+
+        // --- 3. the per-index scissor-test enable --------------------------------------------
+
+        TEST_F(ViewportArrayScenario, AnIndexedScissorEnableClipsOnlyThatIndex) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+
+            // One full-size viewport per index so the scissor rectangle is the ONLY thing that
+            // can shrink the quad - the same separation KHR-GL43.viewport_array.scissor uses.
+            glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+            std::vector<GLint> boxes(static_cast<size_t>(kViewportCount) * 4);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                    boxes[base + 0] = x * kCellSize;
+                    boxes[base + 1] = y * kCellSize;
+                    boxes[base + 2] = kCellSize;
+                    boxes[base + 3] = kCellSize;
+                }
+            }
+            glScissorArrayv(0, kViewportCount, boxes.data());
+
+            const GLuint singleProgram = BuildProgram(kSingleGeometrySource, kIntFragmentSource);
+            ASSERT_NE(singleProgram, 0u) << "single-viewport program failed to build: " << m_buildLog;
+            glUseProgram(singleProgram);
+            glBindVertexArray(m_vao);
+            const GLint uViewport = glGetUniformLocation(singleProgram, "uViewport");
+            ASSERT_NE(uViewport, -1);
+
+            constexpr GLint kProbeIndex = 6;  // grid cell (2, 1)
+            constexpr int kProbeX = kProbeIndex % kGridSide;
+            constexpr int kProbeY = kProbeIndex / kGridSide;
+
+            // (a) scissor test ENABLED for this index: the quad is clipped to its 32x32 box.
+            glUniform1i(uViewport, kProbeIndex);
+            glEnablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                EXPECT_EQ(CellCentre(pixels, kSurfaceSide, kProbeX, kProbeY), kProbeIndex)
+                    << "the scissored index must still paint inside its own box";
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        if (x == kProbeX && y == kProbeY) continue;
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kUnwritten)
+                            << "cell (" << x << ", " << y << ") is outside scissor rectangle " << kProbeIndex
+                            << " and must be untouched";
+                    }
+                }
+            }
+
+            // (b) scissor test DISABLED for the same index, everything else identical: with no
+            // per-viewport toggle in Vulkan this is the case that needs the disabled index to be
+            // given the full framebuffer rectangle, and it is exactly where "leave the last
+            // rectangle bound" would show up as a still-clipped quad.
+            FillIntTarget(target, kSurfaceSide, kSurfaceSide);
+            glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+            glDisablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kProbeIndex)
+                            << "with the scissor test off for index " << kProbeIndex
+                            << ", its full-viewport quad must cover cell (" << x << ", " << y << ")";
+                    }
+                }
+            }
+
+            glDeleteProgram(singleProgram);
+            DestroyIntTarget(target);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_State/GLState/Core.cpp

@@ -712,10 +712,18 @@ namespace MobileGL::MG_State {
             m_renderState.SetViewport(viewport);
         }
 
-        const IntVec4& GLContext::GetViewport() const {
+        IntVec4 GLContext::GetViewport() const {
             return m_renderState.GetViewport();
         }
 
+        void GLContext::SetViewportIndexed(Uint index, FloatVec4 viewport) {
+            m_renderState.SetViewportIndexed(index, viewport);
+        }
+
+        const FloatVec4& GLContext::GetViewportIndexed(Uint index) const {
+            return m_renderState.GetViewportIndexed(index);
+        }
+
         void GLContext::SetLineWidth(Float width) {
             m_renderState.SetLineWidth(width);
         }
@@ -953,6 +961,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetDepthRange();
         }
 
+        void GLContext::SetDepthRangeIndexed(Uint index, FloatVec2 range) {
+            m_renderState.SetDepthRangeIndexed(index, range);
+        }
+
+        const FloatVec2& GLContext::GetDepthRangeIndexed(Uint index) const {
+            return m_renderState.GetDepthRangeIndexed(index);
+        }
+
         void GLContext::SetSampleCoverage(Float value, Bool invert) {
             m_renderState.SetSampleCoverage(value, invert);
         }
@@ -1017,6 +1033,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetScissorBox();
         }
 
+        void GLContext::SetScissorBoxIndexed(Uint index, IntVec4 box) {
+            m_renderState.SetScissorBoxIndexed(index, box);
+        }
+
+        const IntVec4& GLContext::GetScissorBoxIndexed(Uint index) const {
+            return m_renderState.GetScissorBoxIndexed(index);
+        }
+
         // Framebuffer
         void GLContext::GenFramebufferNames(Uint number, Vector<Uint>& framebuffers) {
             m_framebufferState.GenerateNames(number, framebuffers);

MobileGL/MG_State/GLState/Core.h

@@ -198,8 +198,10 @@ namespace MobileGL {
                 // Only the pipeline-relevant subset - see RenderState::m_pipelineStateVersion.
                 Uint GetPipelineStateVersion() const;
                 const RenderStateParameters& GetRenderStateParameters() const;
-                void SetViewport(IntVec4 viewport); // x, y, width, height
-                const IntVec4& GetViewport() const; // x, y, width, height
+                void SetViewport(IntVec4 viewport); // x, y, width, height; writes ALL viewports
+                IntVec4 GetViewport() const;        // x, y, width, height; viewport 0, rounded
+                void SetViewportIndexed(Uint index, FloatVec4 viewport);
+                const FloatVec4& GetViewportIndexed(Uint index) const;
                 void SetLineWidth(Float width);
                 Float GetLineWidth() const;
                 void SetPointSize(Float size);
@@ -260,8 +262,10 @@ namespace MobileGL {
                 Uint32 GetClearStencil() const;
                 void SetBlendColor(FloatVec4 color);
                 const FloatVec4& GetBlendColor() const;
-                void SetDepthRange(FloatVec2 range);
+                void SetDepthRange(FloatVec2 range); // writes ALL viewports' depth ranges
                 const FloatVec2& GetDepthRange() const;
+                void SetDepthRangeIndexed(Uint index, FloatVec2 range);
+                const FloatVec2& GetDepthRangeIndexed(Uint index) const;
                 void SetSampleCoverage(Float value, Bool invert);
                 Float GetSampleCoverageValue() const;
                 Bool GetSampleCoverageInvert() const;
@@ -276,8 +280,10 @@ namespace MobileGL {
                 FrontFaceMode GetFrontFaceMode() const;
                 void SetProvokingVertexMode(ProvokingVertexMode mode);
                 ProvokingVertexMode GetProvokingVertexMode() const;
-                void SetScissorBox(IntVec4 box);      // x, y, width, height
-                const IntVec4& GetScissorBox() const; // x, y, width, height
+                void SetScissorBox(IntVec4 box);      // x, y, width, height; writes ALL rectangles
+                const IntVec4& GetScissorBox() const; // x, y, width, height; rectangle 0
+                void SetScissorBoxIndexed(Uint index, IntVec4 box);
+                const IntVec4& GetScissorBoxIndexed(Uint index) const;
 
                 // Transform feedback. The fields below are the state of the transform
                 // feedback object currently bound to GL_TRANSFORM_FEEDBACK; see the object

MobileGL/MG_State/GLState/RenderState/RenderState.cpp

@@ -25,6 +25,12 @@ namespace MobileGL {
                         return 0;
                     }
                 }
+
+                // Every viewport's scissor-test bit set, i.e. what glEnable(GL_SCISSOR_TEST) writes.
+                constexpr Uint32 kAllViewportsMask =
+                    RenderStateParameters::MAX_VIEWPORTS >= 32
+                        ? ~0u
+                        : (1u << RenderStateParameters::MAX_VIEWPORTS) - 1u;
             } // namespace
 
             RenderState::RenderState() {
@@ -32,6 +38,15 @@ namespace MobileGL {
                 for (auto& mask : m_parameters.ColorMasks) {
                     mask = BoolVec4(true, true, true, true);
                 }
+                // Every viewport's depth range starts at (0, 1) - GL 4.6 core table 23.4. The
+                // viewport and scissor rectangles legitimately start all-zero here: their spec
+                // initial value is the size of the window the context is first made current to,
+                // which the frontend does not know yet, so an all-zero rectangle means "never
+                // written" and the backends resolve it against the live surface (see
+                // DirectGLES' SyncRenderState and VulkanRenderer's ApplyGLViewportState).
+                for (auto& range : m_parameters.DepthRanges) {
+                    range = FloatVec2(0.0f, 1.0f);
+                }
             }
 
             Uint RenderState::GetVersion() const {
@@ -47,15 +62,47 @@ namespace MobileGL {
             }
 
             // -------------------- Rasterization --------------------
+            // ARB_viewport_array, "Additions to Chapter 2": Viewport(x, y, w, h) is equivalent to
+            // ViewportIndexedf(i, x, y, w, h) for every i in [0, MAX_VIEWPORTS) - it is not a
+            // synonym for "viewport 0".
             void RenderState::SetViewport(IntVec4 viewport) {
-                if (m_parameters.Viewport == viewport) return;
+                const FloatVec4 asFloat(static_cast<Float>(viewport.x()), static_cast<Float>(viewport.y()),
+                                        static_cast<Float>(viewport.z()), static_cast<Float>(viewport.w()));
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.Viewports) {
+                    if (stored == asFloat) continue;
+                    stored = asFloat;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
+            }
 
-                m_parameters.Viewport = viewport;
+            IntVec4 RenderState::GetViewport() const {
+                const FloatVec4& viewport = m_parameters.Viewports[0];
+                // Round rather than truncate: glGetIntegerv on floating-point state rounds to
+                // nearest (GL 4.6 core 22.2), and truncating a 63.5-wide viewport to 63 would
+                // also hand the backends a rectangle one pixel short of what was asked for.
+                return IntVec4(static_cast<Int>(std::lround(viewport.x())), static_cast<Int>(std::lround(viewport.y())),
+                               static_cast<Int>(std::lround(viewport.z())), static_cast<Int>(std::lround(viewport.w())));
+            }
+
+            void RenderState::SetViewportIndexed(Uint index, FloatVec4 viewport) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Viewport index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.Viewports[index] == viewport) return;
+
+                m_parameters.Viewports[index] = viewport;
                 ++m_version;
             }
 
-            const IntVec4& RenderState::GetViewport() const {
-                return m_parameters.Viewport;
+            const FloatVec4& RenderState::GetViewportIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Viewport index out of range: %u", index);
+                    return m_parameters.Viewports[0];
+                }
+                return m_parameters.Viewports[index];
             }
 
             void RenderState::SetLineWidth(Float width) {
@@ -223,7 +270,6 @@ namespace MobileGL {
                     SET_CAPABILITY(SampleAlphaToOne, enabled);
                     SET_CAPABILITY(SampleCoverage, enabled);
                     SET_CAPABILITY(SampleMask, enabled);
-                    SET_CAPABILITY(ScissorTest, enabled);
                     SET_CAPABILITY(StencilTest, enabled);
                     SET_CAPABILITY(ProgramPointSize, enabled);
                 case CapabilityInput::Blend: {
@@ -236,6 +282,17 @@ namespace MobileGL {
                     if (stateChanged) BumpVersions();
                     break;
                 }
+                // GL 4.6 core 17.3.2: the non-indexed Enable/Disable(SCISSOR_TEST) enables or
+                // disables the test for ALL viewports, exactly like glViewport writes all
+                // viewports. Anything narrower fails KHR-GL43.viewport_array.scissor_test_state_api,
+                // whose "enable all" phase reads every index back through glIsEnabledi.
+                case CapabilityInput::ScissorTest: {
+                    const Uint32 updated = enabled ? kAllViewportsMask : 0u;
+                    if (m_parameters.ScissorTestEnabledMask == updated) break;
+                    m_parameters.ScissorTestEnabledMask = updated;
+                    BumpVersions();
+                    break;
+                }
                 case CapabilityInput::ClipDistance0:
                 case CapabilityInput::ClipDistance1:
                 case CapabilityInput::ClipDistance2:
@@ -287,11 +344,14 @@ namespace MobileGL {
                     RETURN_CAPABILITY(SampleAlphaToOne);
                     RETURN_CAPABILITY(SampleCoverage);
                     RETURN_CAPABILITY(SampleMask);
-                    RETURN_CAPABILITY(ScissorTest);
                     RETURN_CAPABILITY(StencilTest);
                     RETURN_CAPABILITY(ProgramPointSize);
                 case CapabilityInput::Blend:
                     return m_parameters.BlendStates[0].Enabled;
+                // The non-indexed query of an indexed capability answers for index 0
+                // (GL 4.6 core 22.1), which is also the only bit either backend consumes today.
+                case CapabilityInput::ScissorTest:
+                    return (m_parameters.ScissorTestEnabledMask & 1u) != 0;
                 case CapabilityInput::ClipDistance0:
                 case CapabilityInput::ClipDistance1:
                 case CapabilityInput::ClipDistance2:
@@ -307,13 +367,29 @@ namespace MobileGL {
             }
 
             void RenderState::SetCapabilityIndexed(CapabilityInput cap, Uint index, Bool enabled) {
-                // Only for BlendState currently. The GL entry points (glEnablei/glDisablei) already
-                // reject every non-GL_BLEND target with GL_INVALID_ENUM before reaching here, so this
-                // is a backstop - but it must stay a backstop: THROW_UNIMPL_EXCEPTION unwinds a C++
-                // exception through the C GL ABI and terminates the process.
+                // GL_BLEND (indexed by draw buffer) and GL_SCISSOR_TEST (indexed by viewport) are
+                // the only indexed capabilities in GL 4.6 core. The GL entry points
+                // (glEnablei/glDisablei) already reject every other target with GL_INVALID_ENUM
+                // and every out-of-range index with GL_INVALID_VALUE before reaching here, so the
+                // guards below are backstops - but they must stay backstops:
+                // THROW_UNIMPL_EXCEPTION unwinds a C++ exception through the C GL ABI and
+                // terminates the process.
+                if (cap == CapabilityInput::ScissorTest) {
+                    if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                        MOBILEGL_ASSERT(false, "Scissor test capability index out of range: %u", index);
+                        return;
+                    }
+                    const Uint32 bit = 1u << index;
+                    const Uint32 updated = enabled ? (m_parameters.ScissorTestEnabledMask | bit)
+                                                   : (m_parameters.ScissorTestEnabledMask & ~bit);
+                    if (updated == m_parameters.ScissorTestEnabledMask) return;
+                    m_parameters.ScissorTestEnabledMask = updated;
+                    BumpVersions();
+                    return;
+                }
                 if (cap != CapabilityInput::Blend) {
                     MGLOG_I("RenderState::SetCapabilityIndexed: indexed capability state exists only for "
-                            "GL_BLEND (cap=%d, index=%u); ignoring",
+                            "GL_BLEND and GL_SCISSOR_TEST (cap=%d, index=%u); ignoring",
                             static_cast<int>(cap), index);
                     return;
                 }
@@ -328,9 +404,17 @@ namespace MobileGL {
             }
 
             Bool RenderState::IsCapabilityEnabledIndexed(CapabilityInput cap, Uint index) const {
-                // Only for BlendState currently - same backstop reasoning as SetCapabilityIndexed:
-                // glIsEnabledi has already answered GL_INVALID_ENUM/GL_FALSE for anything else, and a
-                // query must never be able to terminate the process.
+                // GL_BLEND and GL_SCISSOR_TEST only - same backstop reasoning as
+                // SetCapabilityIndexed: glIsEnabledi has already answered
+                // GL_INVALID_ENUM/GL_INVALID_VALUE for anything else, and a query must never be
+                // able to terminate the process.
+                if (cap == CapabilityInput::ScissorTest) {
+                    if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                        MOBILEGL_ASSERT(false, "Scissor test capability index out of range: %u", index);
+                        return false;
+                    }
+                    return (m_parameters.ScissorTestEnabledMask & (1u << index)) != 0;
+                }
                 if (cap != CapabilityInput::Blend) {
                     MGLOG_I("RenderState::IsCapabilityEnabledIndexed: indexed capability state exists only "
                             "for GL_BLEND (cap=%d, index=%u); reporting disabled",
@@ -591,15 +675,39 @@ namespace MobileGL {
                 return m_parameters.BlendColor;
             }
 
+            // Like Viewport: ARB_viewport_array makes DepthRange(n, f) the same as
+            // DepthRangeIndexed(i, n, f) for every i.
             void RenderState::SetDepthRange(FloatVec2 range) {
-                if (m_parameters.DepthRange == range) return;
-
-                m_parameters.DepthRange = range;
-                ++m_version;
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.DepthRanges) {
+                    if (stored == range) continue;
+                    stored = range;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
             }
 
             const FloatVec2& RenderState::GetDepthRange() const {
-                return m_parameters.DepthRange;
+                return m_parameters.DepthRanges[0];
+            }
+
+            void RenderState::SetDepthRangeIndexed(Uint index, FloatVec2 range) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Depth range index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.DepthRanges[index] == range) return;
+
+                m_parameters.DepthRanges[index] = range;
+                ++m_version;
+            }
+
+            const FloatVec2& RenderState::GetDepthRangeIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Depth range index out of range: %u", index);
+                    return m_parameters.DepthRanges[0];
+                }
+                return m_parameters.DepthRanges[index];
             }
 
             void RenderState::SetSampleCoverage(Float value, Bool invert) {
@@ -726,15 +834,39 @@ namespace MobileGL {
             }
 
             // --------------------- Scissor ---------------------
+            // Like Viewport: ARB_viewport_array makes Scissor(x, y, w, h) the same as
+            // ScissorIndexed(i, x, y, w, h) for every i.
             void RenderState::SetScissorBox(IntVec4 box) {
-                if (m_parameters.ScissorBox == box) return;
-
-                m_parameters.ScissorBox = box;
-                ++m_version;
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.ScissorBoxes) {
+                    if (stored == box) continue;
+                    stored = box;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
             }
 
             const IntVec4& RenderState::GetScissorBox() const {
-                return m_parameters.ScissorBox;
+                return m_parameters.ScissorBoxes[0];
+            }
+
+            void RenderState::SetScissorBoxIndexed(Uint index, IntVec4 box) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Scissor box index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.ScissorBoxes[index] == box) return;
+
+                m_parameters.ScissorBoxes[index] = box;
+                ++m_version;
+            }
+
+            const IntVec4& RenderState::GetScissorBoxIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Scissor box index out of range: %u", index);
+                    return m_parameters.ScissorBoxes[0];
+                }
+                return m_parameters.ScissorBoxes[index];
             }
         } // namespace GLState
     } // namespace MG_State

MobileGL/MG_State/GLState/RenderState/RenderState.h

@@ -220,8 +220,22 @@ namespace MobileGL {
     };
 
     struct RenderStateParameters {
+        // ARB_viewport_array / GL 4.6 core 13.6.1: the viewport, the scissor rectangle, the depth
+        // range and the scissor-test enable are all arrays indexed by gl_ViewportIndex, and the
+        // spec floor for MAX_VIEWPORTS is 16. MobileGL advertises exactly 16 on both backends, so
+        // this is also what GL_MAX_VIEWPORTS reports (see the backend loaders' caps.MaxViewports).
+        static constexpr Uint MAX_VIEWPORTS = 16;
+
         // Rasterization
-        IntVec4 Viewport = IntVec4(0, 0, 0, 0); // x, y, width, height
+        // The viewport rectangle is FLOAT state as of GL 4.1 - ViewportIndexedf writes fractional
+        // values and GetFloati_v(GL_VIEWPORT) must hand them back bit-exact
+        // (KHR-GL43.viewport_array.viewport_api compares with ==, no tolerance). glViewport's
+        // integers are simply one way to write it. Index 0 is what a program that never assigns
+        // gl_ViewportIndex rasterizes against, and what the classic glViewport /
+        // glGetIntegerv(GL_VIEWPORT) pair addresses. Both backends rasterize the rectangle
+        // rounded back to integers; the STATE stays exact, which is the half the conformance
+        // suite checks (see the KNOWN INFIDELITY note in AdvertisedLimitsScenario.cpp).
+        Array<FloatVec4, MAX_VIEWPORTS> Viewports{}; // x, y, width, height
         Float LineWidth = 1.0f;
         Float PointSize = 1.0f;
         // GL_PATCH_VERTICES: how many vertices one tessellation patch consumes.
@@ -247,7 +261,13 @@ namespace MobileGL {
         Float ClearDepth = 1.0f;
         Uint32 ClearStencil = 0;
         FloatVec4 BlendColor = FloatVec4(0.0f, 0.0f, 0.0f, 0.0f);
-        FloatVec2 DepthRange = FloatVec2(0.0f, 1.0f);
+        // Per-viewport depth range (glDepthRangeIndexed / glDepthRangeArrayv). Every entry is
+        // initialized to (0, 1) in RenderState's constructor - a default member initializer would
+        // not survive the Array<> aggregate. Kept float rather than double: DepthRangeArrayv takes
+        // GLdouble, but the value reaches the hardware as VkViewport::minDepth/maxDepth (float) on
+        // Magma and glDepthRangef on Espryt, so a double store would only widen the readback and
+        // then lose it again at the same place.
+        Array<FloatVec2, MAX_VIEWPORTS> DepthRanges{};
         Float SampleCoverageValue = 1.0f;
         Bool SampleCoverageInvert = false;
         Uint32 SampleMaskValue = 0xffffffffu;
@@ -299,10 +319,15 @@ namespace MobileGL {
         Bool SampleAlphaToOneEnabled = false;
         Bool SampleCoverageEnabled = false;
         Bool SampleMaskEnabled = false;
-        Bool ScissorTestEnabled = false;
         Bool StencilTestEnabled = false;
         Bool ProgramPointSizeEnabled = false;
-        IntVec4 ScissorBox = IntVec4(0, 0, 0, 0); // x, y, width, height
+        // glEnable(GL_SCISSOR_TEST) enables the test for EVERY viewport, glEnablei for one
+        // (GL 4.6 core 17.3.2), so this is 16 bits and not a bool. Bit 0 is what the classic
+        // glIsEnabled(GL_SCISSOR_TEST) reports and what both backends currently consume. Unlike
+        // ClipDistanceEnabledMask below it DOES bump the pipeline version, because DirectGLES
+        // turns it into a real glEnable/glDisable.
+        Uint32 ScissorTestEnabledMask = 0;
+        Array<IntVec4, MAX_VIEWPORTS> ScissorBoxes{}; // x, y, width, height
         // glEnable(GL_CLIP_DISTANCE0 + i) for i in [0, 8), one bit each. A bitmask rather than
         // eight bools because every consumer wants the set, not an individual flag, and because
         // the SYNC_CAPABILITY/SET_CAPABILITY macros key off a "<Name>Enabled" field name that
@@ -323,8 +348,14 @@ namespace MobileGL {
                 const RenderStateParameters& GetAllParameters() const;
 
                 // Rasterization
+                // ARB_viewport_array defines glViewport as ViewportIndexedf on EVERY index, so the
+                // classic setter broadcasts; GetViewport answers for index 0 (rounded to the
+                // integers glGetIntegerv(GL_VIEWPORT) and both backends want) and is BY VALUE for
+                // that reason. The indexed pair is the verbatim float state.
                 void SetViewport(IntVec4 viewport); // x, y, width, height
-                const IntVec4& GetViewport() const; // x, y, width, height
+                IntVec4 GetViewport() const;        // x, y, width, height, viewport 0, rounded
+                void SetViewportIndexed(Uint index, FloatVec4 viewport);
+                const FloatVec4& GetViewportIndexed(Uint index) const;
                 void SetLineWidth(Float width);
                 Float GetLineWidth() const;
                 void SetPointSize(Float size);
@@ -400,8 +431,12 @@ namespace MobileGL {
                 Uint32 GetClearStencil() const;
                 void SetBlendColor(FloatVec4 color);
                 const FloatVec4& GetBlendColor() const;
+                // glDepthRange(f) writes every viewport's range (ARB_viewport_array); the indexed
+                // pair is glDepthRangeIndexed / glDepthRangeArrayv. GetDepthRange answers index 0.
                 void SetDepthRange(FloatVec2 range);
                 const FloatVec2& GetDepthRange() const;
+                void SetDepthRangeIndexed(Uint index, FloatVec2 range);
+                const FloatVec2& GetDepthRangeIndexed(Uint index) const;
                 void SetSampleCoverage(Float value, Bool invert);
                 Float GetSampleCoverageValue() const;
                 Bool GetSampleCoverageInvert() const;
@@ -421,9 +456,12 @@ namespace MobileGL {
                 void SetProvokingVertexMode(ProvokingVertexMode mode);
                 ProvokingVertexMode GetProvokingVertexMode() const;
 
-                // Scissor
+                // Scissor. glScissor writes every rectangle (ARB_viewport_array); GetScissorBox
+                // answers for index 0.
                 void SetScissorBox(IntVec4 box);      // x, y, width, height
                 const IntVec4& GetScissorBox() const; // x, y, width, height
+                void SetScissorBoxIndexed(Uint index, IntVec4 box);
+                const IntVec4& GetScissorBoxIndexed(Uint index) const;
 
             private:
                 // Bump both: any state change invalidates the draw snapshot, and this one also

MobileGL/MG_Test/Backend/DirectVulkan/SanityTest.cpp

@@ -8,9 +8,28 @@
 
 #include <gtest/gtest.h>
 #include <iostream>
+#include <utility>
 #include <vector>
 
-#include <vulkan/vulkan.h>
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+
+TEST(DirectVulkanSanity, ProgramMovePreservesViewportIndexUsage) {
+    using VkProgramObject = MobileGL::MG_Backend::DirectVulkan::ProgramFactory::VkProgramObject;
+
+    VkProgramObject moveConstructedSource;
+    moveConstructedSource.writesViewportIndexBuiltin = true;
+    VkProgramObject moveConstructed(std::move(moveConstructedSource));
+    EXPECT_TRUE(moveConstructed.writesViewportIndexBuiltin);
+    EXPECT_FALSE(moveConstructedSource.writesViewportIndexBuiltin);
+
+    VkProgramObject moveAssignedSource;
+    moveAssignedSource.writesViewportIndexBuiltin = true;
+    VkProgramObject moveAssigned;
+    moveAssigned = std::move(moveAssignedSource);
+    EXPECT_TRUE(moveAssigned.writesViewportIndexBuiltin);
+    EXPECT_FALSE(moveAssignedSource.writesViewportIndexBuiltin);
+}
+
 TEST(DirectVulkanSanity, ExtensionEnumeration) {
     uint32_t extensionCount = 0;
     vkEnumerateInstanceExtensionProperties(nullptr, &extensionCount, nullptr);

MobileGL/MG_Test/BackendLoader/BackendLoaderTest.cpp

@@ -725,22 +725,57 @@ TEST(TextureAnisotropyCapabilities, ExtensionIsAdvertisedOnlyWhenTheHostDriverSu
         return std::find(extensions.begin(), extensions.end(), wanted) != extensions.end();
     };
 
-    const auto without = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false);
+    const auto without = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false);
     EXPECT_FALSE(contains(without, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_FALSE(contains(without, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 
-    const auto with = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, true);
+    const auto with = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, true, false, false);
     EXPECT_TRUE(contains(with, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_TRUE(contains(with, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 
     // Same rule on the Vulkan backend, where the gate is the samplerAnisotropy device feature.
-    const auto vkWithout = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false);
+    const auto vkWithout = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false);
     EXPECT_FALSE(contains(vkWithout, MobileGL::E_GL_EXT_texture_filter_anisotropic));
-    const auto vkWith = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, true);
+    const auto vkWith = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, true, false);
     EXPECT_TRUE(contains(vkWith, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_TRUE(contains(vkWith, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 }
 
+// Minecraft 26.3 checks ARB_draw_indirect before it considers the already-advertised
+// ARB_multi_draw_indirect, then separately requires ARB_base_instance before enabling its terrain
+// indirect path. Pin both strings and, just as importantly, the non-zero firstInstance gate.
+TEST(IndirectDrawAdvertisement, MatchesEachBackendsUsableCommandSemantics) {
+    const auto contains = [](const MobileGL::Vector<MobileGL::GLExtension>& extensions,
+                             MobileGL::GLExtension wanted) {
+        return std::find(extensions.begin(), extensions.end(), wanted) != extensions.end();
+    };
+
+    const auto esWithoutIndirect =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false);
+    EXPECT_FALSE(contains(esWithoutIndirect, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(esWithoutIndirect, MobileGL::E_GL_ARB_base_instance));
+
+    const auto esWithoutBaseInstance =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, true, false);
+    EXPECT_TRUE(contains(esWithoutBaseInstance, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(esWithoutBaseInstance, MobileGL::E_GL_ARB_base_instance));
+
+    const auto esWithBoth =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, true, true);
+    EXPECT_TRUE(contains(esWithBoth, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_TRUE(contains(esWithBoth, MobileGL::E_GL_ARB_base_instance));
+
+    const auto vkWithoutBaseInstance =
+        MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false);
+    EXPECT_TRUE(contains(vkWithoutBaseInstance, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(vkWithoutBaseInstance, MobileGL::E_GL_ARB_base_instance));
+
+    const auto vkWithBoth =
+        MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, true);
+    EXPECT_TRUE(contains(vkWithBoth, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_TRUE(contains(vkWithBoth, MobileGL::E_GL_ARB_base_instance));
+}
+
 TEST(TextureAnisotropyCapabilities, MaxAnisotropyIsQueriedOnlyWhenTheExtensionIsPresent) {
     ResetFakeDriver();
     g_fake.maxVertexSsboBlocks = 0;
@@ -836,3 +871,63 @@ TEST(MultiDrawCapabilities, ExtensionWithoutResolvedPointerIsNotSupport) {
     EXPECT_FALSE(caps.SupportsMultiDrawIndirect);
     EXPECT_FALSE(caps.SupportsMultiDrawElementsBaseVertex);
 }
+
+TEST(DrawIndirectCapabilities, RequiresEs31AndBothCoreEntryPoints) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    auto funcs = MakeFakeGLESFunctions();
+    funcs.glDrawElementsIndirect = [](GLenum, GLenum, const void*) {};
+
+    MobileGL::MG_External::GLESCapabilities supportedCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(supportedCaps, funcs));
+    EXPECT_TRUE(supportedCaps.SupportsDrawIndirect);
+
+    // The same pointers on an ES 3.0 context are not core entry points and cannot back the
+    // desktop extension contract.
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 0;
+    MobileGL::MG_External::GLESCapabilities es30Caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(es30Caps, funcs));
+    EXPECT_FALSE(es30Caps.SupportsDrawIndirect);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    const auto missingElements = MakeFakeGLESFunctions();
+    MobileGL::MG_External::GLESCapabilities missingEntryPointCaps;
+    ASSERT_TRUE(
+        MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(missingEntryPointCaps, missingElements));
+    EXPECT_FALSE(missingEntryPointCaps.SupportsDrawIndirect);
+}
+
+TEST(BaseInstanceCapabilities, RequiresTheExtensionAndAllThreeEntryPoints) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    auto funcs = MakeFakeGLESFunctions();
+    funcs.glDrawArraysInstancedBaseInstanceEXT = [](GLenum, GLint, GLsizei, GLsizei, GLuint) {};
+    funcs.glDrawElementsInstancedBaseInstanceEXT =
+        [](GLenum, GLsizei, GLenum, const void*, GLsizei, GLuint) {};
+    funcs.glDrawElementsInstancedBaseVertexBaseInstanceEXT =
+        [](GLenum, GLsizei, GLenum, const void*, GLsizei, GLint, GLuint) {};
+
+    // Resolved stubs alone must never make the capability true.
+    MobileGL::MG_External::GLESCapabilities pointersOnlyCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(pointersOnlyCaps, funcs));
+    EXPECT_FALSE(pointersOnlyCaps.SupportsBaseInstance);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_base_instance");
+    MobileGL::MG_External::GLESCapabilities supportedCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(supportedCaps, funcs));
+    EXPECT_TRUE(supportedCaps.SupportsBaseInstance);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_base_instance");
+    funcs.glDrawElementsInstancedBaseInstanceEXT = nullptr;
+    MobileGL::MG_External::GLESCapabilities missingEntryPointCaps;
+    ASSERT_TRUE(
+        MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(missingEntryPointCaps, funcs));
+    EXPECT_FALSE(missingEntryPointCaps.SupportsBaseInstance);
+}

MobileGL/MG_Test/Pipeline/CMakeLists.txt

@@ -4,6 +4,7 @@ add_executable(
         PipelineQuirkTest
         PipelineQuirkTest.cpp
         PassthroughTessControlTest.cpp
+        ViewportIndexReflectionTest.cpp
 )
 
 target_include_directories(PipelineQuirkTest PRIVATE

MobileGL/MG_Test/Pipeline/ViewportIndexReflectionTest.cpp

@@ -0,0 +1,183 @@
+// MobileGL - MobileGL/MG_Test/Pipeline/ViewportIndexReflectionTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// ProgramFactory::ReflectedWritesViewportIndexBuiltin is the switch that decides whether a
+// DirectVulkan pipeline declares one viewport or all sixteen. Getting it wrong is silent in both
+// directions and neither direction is caught by a state test:
+//
+//   - a false NEGATIVE collapses every gl_ViewportIndex onto viewport 0, which is precisely the
+//     bug the multi-viewport work exists to fix and which a set/get round trip cannot see;
+//   - a false POSITIVE widens viewportCount for an ordinary Minecraft shader, costing a longer
+//     vkCmdSetViewport per state change and, on a tiler, possibly a hardware fast path.
+//
+// So this compiles REAL GLSL through the same glslang path the renderer uses and reflects the
+// SPIR-V that comes out, rather than asserting against hand-assembled words: what has to hold is
+// that the detector agrees with what glslang actually emits for a shader that writes the builtin,
+// including the stage-by-stage question of WHERE it may be written (GL 4.1 allows the geometry
+// stage; ARB_shader_viewport_layer_array adds vertex and tessellation evaluation).
+//
+// The end-to-end claim - that a detected writer really does route pixels to its own viewport -
+// lives in MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp.
+
+#include <gtest/gtest.h>
+
+#include <string>
+#include <vector>
+
+#include "Includes.h"
+#include "Init.h"
+
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <spirv_reflect.h>
+
+using namespace MobileGL;
+using MobileGL::MG_Backend::DirectVulkan::ProgramFactory;
+using MobileGL::MG_Util::ShaderTranspiler::ShaderCompiler;
+
+namespace {
+
+    Vector<Uint32> CompileToSpirv(GLenum stage, const String& source) {
+        using namespace MG_Util::ShaderTranspiler;
+        ShaderAttrib shaderAttrib{.shaderType = stage, .sourceStr = source};
+        auto shaderResult = ShaderCompiler::CompileShader(shaderAttrib);
+        EXPECT_TRUE(shaderResult) << (shaderResult ? String{} : shaderResult.error().log);
+        if (!shaderResult) return {};
+
+        ProgramAttrib programAttrib{.shaders = {shaderResult.value()}};
+        auto programResult = ShaderCompiler::LinkProgram(programAttrib);
+        EXPECT_TRUE(programResult) << (programResult ? String{} : programResult.error().log);
+        if (!programResult) return {};
+
+        ProgramBinaryAttrib binaryAttrib{.shaderTypes = {stage}, .program = *programResult.value()};
+        auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+        EXPECT_TRUE(binaryResult) << (binaryResult ? String{} : binaryResult.error().log);
+        if (!binaryResult || binaryResult->empty()) return {};
+        return binaryResult->front();
+    }
+
+    // Owns the reflection module so a failing EXPECT cannot leak it.
+    class ReflectModule {
+    public:
+        explicit ReflectModule(const Vector<Uint32>& spirv) {
+            if (spirv.empty()) return;
+            m_created = spvReflectCreateShaderModule(spirv.size() * sizeof(Uint32), spirv.data(), &m_module) ==
+                        SPV_REFLECT_RESULT_SUCCESS;
+        }
+        ~ReflectModule() {
+            if (m_created) spvReflectDestroyShaderModule(&m_module);
+        }
+        ReflectModule(const ReflectModule&) = delete;
+        ReflectModule& operator=(const ReflectModule&) = delete;
+
+        Bool Created() const { return m_created; }
+        const SpvReflectShaderModule& Get() const { return m_module; }
+
+    private:
+        SpvReflectShaderModule m_module{};
+        Bool m_created = false;
+    };
+
+    class ViewportIndexReflectionTest : public ::testing::Test {
+    protected:
+        void SetUp() override { MobileGL::Initialize(); }
+    };
+
+    const char* const kGeometryWritesViewportIndex = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    // Same stage, same shape, writing gl_Layer INSTEAD. Layered rendering and viewport routing
+    // are different features and the detector must not confuse them: a Minecraft-style cubemap
+    // pass writes gl_Layer and must keep the one-viewport pipeline.
+    const char* const kGeometryWritesLayerOnly = R"(#version 410 core
+layout(points, invocations = 6) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_Layer = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    const char* const kPlainGeometry = R"(#version 410 core
+layout(points, invocations = 1) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    const char* const kPlainVertex = R"(#version 410 core
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+    const char* const kPlainFragment = R"(#version 410 core
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = vec4(1.0); }
+)";
+
+    TEST_F(ViewportIndexReflectionTest, TrueForAGeometryShaderThatAssignsViewportIndex) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kGeometryWritesViewportIndex));
+        ASSERT_TRUE(module.Created());
+        EXPECT_TRUE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()))
+            << "a shader that assigns gl_ViewportIndex must get a multi-viewport pipeline; missing it is what "
+               "collapses every index onto viewport 0";
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAGeometryShaderThatOnlyAssignsLayer) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kGeometryWritesLayerOnly));
+        ASSERT_TRUE(module.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()))
+            << "gl_Layer is layered rendering, not viewport routing; widening viewportCount for it costs the "
+               "single-viewport fast path for nothing";
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAPlainGeometryShader) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kPlainGeometry));
+        ASSERT_TRUE(module.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()));
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForTheOrdinaryVertexAndFragmentStages) {
+        // The shape every real application ships: neither stage may widen the pipeline.
+        const ReflectModule vertexModule(CompileToSpirv(GL_VERTEX_SHADER, kPlainVertex));
+        ASSERT_TRUE(vertexModule.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(vertexModule.Get()));
+
+        const ReflectModule fragmentModule(CompileToSpirv(GL_FRAGMENT_SHADER, kPlainFragment));
+        ASSERT_TRUE(fragmentModule.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(fragmentModule.Get()));
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAnEmptyModuleWithoutDereferencing) {
+        // A default-constructed module has no entry points. The scan runs on every link, so it
+        // must survive a reflection that never got built rather than walk a null array.
+        SpvReflectShaderModule emptyModule{};
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(emptyModule));
+    }
+
+} // namespace

MobileGL/MG_Test/Program/ParallelShaderCompileTest.cpp

@@ -516,17 +516,17 @@ TEST_F(ParallelShaderCompileTest, MaxShaderCompilerThreadsIgnoresTheCurrentBudge
 TEST_F(ParallelShaderCompileTest, BothBackendsAdvertiseTheExtensionIffAsyncIsEnabled) {
     {
         const AsyncModeScope async(true);
-        EXPECT_TRUE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false),
+        EXPECT_TRUE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false),
                                E_GL_KHR_parallel_shader_compile));
-        EXPECT_TRUE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false),
+        EXPECT_TRUE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false),
                                E_GL_KHR_parallel_shader_compile));
     }
     {
         const AsyncModeScope async(false);
-        EXPECT_FALSE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false),
+        EXPECT_FALSE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false),
                                 E_GL_KHR_parallel_shader_compile))
             << "MOBILEGL_ASYNC_SHADER_COMPILE=0 must withdraw the extension, not only the threading";
-        EXPECT_FALSE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false),
+        EXPECT_FALSE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false),
                                 E_GL_KHR_parallel_shader_compile))
             << "MOBILEGL_ASYNC_SHADER_COMPILE=0 must withdraw the extension, not only the threading";
     }

MobileGL/MG_Test/Program/ProgramTest.cpp

@@ -2692,11 +2692,13 @@ out vec4 fragColor;
 float fma
     (float a, float b, float c) { return a * b + c; }
 float sinh(float x, float y) { return x * y; }
+float length_squared(vec3 value) { return dot(value, value); }
 float round(float x) { return floor(x + 0.5); }
 float min3(float a, float b, float c) { return min(min(a, b), c); }
 
 void main() {
-    fragColor = vec4(fma(0.1, 0.2, 0.3), sinh(0.4, 2.0), round(1.25), min3(0.1, 0.2, 0.3));
+    fragColor = vec4(fma(0.1, 0.2, 0.3), sinh(0.4, 2.0), round(1.25),
+                     min3(0.1, 0.2, 0.3) + length_squared(vec3(0.1, 0.2, 0.3)));
 }
 )";
     GLuint vs = CompileShaderChecked(GL_VERTEX_SHADER, vsSource);
@@ -2707,6 +2709,7 @@ void main() {
         if (essl.find("fragColor") == String::npos) continue; // fragment module only
         EXPECT_NE(essl.find("mg_fma("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_sinh("), String::npos) << essl;
+        EXPECT_NE(essl.find("mg_length_squared("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_round("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_min3("), String::npos) << essl;
         EXPECT_EQ(essl.find("float fma("), String::npos) << essl;

MobileGL/MG_Test/Program/ProgramUtilTest.cpp

@@ -52,20 +52,24 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
                OpEntryPoint Fragment %main "main" %outColor
                OpExecutionMode %main OriginUpperLeft
                OpName %globalSampler "sampler"
+               OpName %globalNew "new"
                OpName %paramSampler "sampler"
+               OpName %paramNew "new"
                OpName %main "main"
                OpDecorate %outColor Location 0
        %void = OpTypeVoid
       %float = OpTypeFloat 32
     %v4float = OpTypeVector %float 4
      %mainFn = OpTypeFunction %void
-    %paramFn = OpTypeFunction %void %float
+    %paramFn = OpTypeFunction %void %float %float
   %outV4Ptr = OpTypePointer Output %v4float
  %privatePtr = OpTypePointer Private %float
    %outColor = OpVariable %outV4Ptr Output
 %globalSampler = OpVariable %privatePtr Private
+    %globalNew = OpVariable %privatePtr Private
      %helper = OpFunction %void None %paramFn
 %paramSampler = OpFunctionParameter %float
+    %paramNew = OpFunctionParameter %float
  %helperBody = OpLabel
                OpReturn
                OpFunctionEnd
@@ -91,6 +95,7 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
     ASSERT_TRUE(tools.Disassemble(outputBinary, &outputText));
 
     EXPECT_NE(outputText.find("\"MGL_COMPAT_sampler\""), String::npos);
+    EXPECT_NE(outputText.find("\"MGL_COMPAT_new\""), String::npos);
 
     SizeT exactSamplerNameCount = 0;
     SizeT searchOffset = 0;
@@ -99,6 +104,14 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
         searchOffset += std::strlen("\"sampler\"");
     }
     EXPECT_EQ(exactSamplerNameCount, 1u);
+
+    SizeT exactNewNameCount = 0;
+    searchOffset = 0;
+    while ((searchOffset = outputText.find("\"new\"", searchOffset)) != String::npos) {
+        ++exactNewNameCount;
+        searchOffset += std::strlen("\"new\"");
+    }
+    EXPECT_EQ(exactNewNameCount, 1u);
 }
 
 TEST_F(ProgramUtilTest, UnformattedFloatStorageImagesKeepIntegerAtomicImagesTyped) {

MobileGL/MG_Test/Query/QueryTest.cpp

@@ -448,6 +448,39 @@ TEST_F(QueryTest, BackendResultsPropagateThroughFrontend) {
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
 
+TEST_F(QueryTest, DestroyAllQueryObjectsReclaimsRegistryAndResetsContextState) {
+    const ScopedFeaturesOverride featuresGuard;
+    const ScopedBackendFunctionsOverride backendGuard;
+    InstallStubBackendTimerQueries();
+    MG_Config::Features.DisableTimerQuery = false;
+
+    GLuint id = 0;
+    MG_Impl::GLImpl::GenQueries(1, &id);
+    ASSERT_NE(id, 0u);
+    MG_Impl::GLImpl::BeginQuery(GL_TIME_ELAPSED, id);
+
+    GLint currentQuery = -1;
+    MG_Impl::GLImpl::GetQueryiv(GL_TIME_ELAPSED, GL_CURRENT_QUERY, &currentQuery);
+    EXPECT_EQ(currentQuery, static_cast<GLint>(id));
+
+    // Full teardown drains the registry through this function while the backend
+    // table is still valid. The unread backend handle must be released, the query
+    // must disappear, and a fresh context must restart with no active query and a
+    // fresh name allocator.
+    MG_Impl::GLImpl::DestroyAllQueryObjects();
+    EXPECT_EQ(g_stubDeleteCount, 1);
+    EXPECT_EQ(MG_Impl::GLImpl::IsQuery(id), GL_FALSE);
+
+    MG_Impl::GLImpl::GetQueryiv(GL_TIME_ELAPSED, GL_CURRENT_QUERY, &currentQuery);
+    EXPECT_EQ(currentQuery, 0);
+
+    GLuint freshId = 0;
+    MG_Impl::GLImpl::GenQueries(1, &freshId);
+    EXPECT_EQ(freshId, 1u);
+    MG_Impl::GLImpl::DeleteQueries(1, &freshId);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
 // Environment-agnostic property test for the env -> ConfigLoader -> Features
 // chain: whatever MOBILEGL_DISABLE_TIMERQUERY is set to in the environment of
 // this test process, MG_ConfigLoader::Init must have parsed it with the

MobileGL/MG_Test/SanityTest.cpp

@@ -18,6 +18,7 @@
 #include <MG_Backend/DirectVulkan/BackendObject_DirectVulkan.h>
 #include <MG_Backend/BackendObjects.h>
 #include <MG_Impl/GLImpl/Getter/GL_Getter.h>
+#include <MG_Impl/GLImpl/Sync/GL_Sync.h>
 #include <MG_Impl/GLImpl/RenderState/GL_RenderState.h>
 #include <MG_Impl/GLImpl/Texture/GL_Texture.h>
 #include <MG_Impl/GLImpl/VertexArray/Validators.h>
@@ -936,6 +937,45 @@ TEST(DirectVulkanSanity, ReadbackUsesTheSourceFormatTexelSize) {
     EXPECT_EQ(VulkanRenderer::GetReadbackTexelSize(VK_FORMAT_R32G32B32A32_SFLOAT), 16u);
 }
 
+TEST(DirectVulkanSanity, DefaultFramebufferQuarterTurnReadbackMapsRectAndPixels) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+    using MobileGL::Uint8;
+
+    VkOffset2D offset{};
+    VkExtent2D copyExtent{};
+    ASSERT_TRUE(VulkanRenderer::MapDefaultFramebufferReadbackRect(
+        1, 0, 2, 1, VkExtent2D{2, 3}, VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR,
+        &offset, &copyExtent));
+    EXPECT_EQ(offset.x, 0);
+    EXPECT_EQ(offset.y, 1);
+    EXPECT_EQ(copyExtent.width, 1u);
+    EXPECT_EQ(copyExtent.height, 2u);
+
+    ASSERT_TRUE(VulkanRenderer::MapDefaultFramebufferReadbackRect(
+        1, 0, 2, 1, VkExtent2D{2, 3}, VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR,
+        &offset, &copyExtent));
+    EXPECT_EQ(offset.x, 1);
+    EXPECT_EQ(offset.y, 0);
+    EXPECT_EQ(copyExtent.width, 1u);
+    EXPECT_EQ(copyExtent.height, 2u);
+
+    // Logical GL rows, bottom to top, are abc / def. The display-oriented swapchain blocks are
+    // transposed in opposite directions for 90 and 270 degrees.
+    const Uint8 raw90[] = {'a', 'd', 'b', 'e', 'c', 'f'};
+    const Uint8 raw270[] = {'f', 'c', 'e', 'b', 'd', 'a'};
+    const Uint8 expected[] = {'a', 'b', 'c', 'd', 'e', 'f'};
+    Uint8 result[sizeof(expected)]{};
+
+    ASSERT_TRUE(VulkanRenderer::RemapDefaultFramebufferReadback(
+        raw90, 3, 2, VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR, 1, result));
+    EXPECT_TRUE(std::equal(std::begin(expected), std::end(expected), std::begin(result)));
+
+    std::fill(std::begin(result), std::end(result), 0);
+    ASSERT_TRUE(VulkanRenderer::RemapDefaultFramebufferReadback(
+        raw270, 3, 2, VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR, 1, result));
+    EXPECT_TRUE(std::equal(std::begin(expected), std::end(expected), std::begin(result)));
+}
+
 TEST(DirectVulkanSanity, ReadbackConvertsRgba8AndRgba16fPixels) {
     using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
     using MobileGL::MG_Util::EncodeFloatToHalfBits;
@@ -1932,6 +1972,9 @@ namespace {
         MobileGL::Vector<GLuint> framebuffers;
         MobileGL::Vector<GLuint> renderbuffers;
         MobileGL::Vector<GLuint> samplers;
+        MobileGL::Vector<GLuint> vertexArrays;
+        MobileGL::Vector<GLuint> programs;
+        MobileGL::Vector<GLuint> buffers;
     };
 
     TwinDeletionSinks* g_twinDeletionSinks = nullptr;
@@ -1958,6 +2001,24 @@ namespace {
         if (!g_twinDeletionSinks) return;
         for (GLsizei i = 0; i < count; ++i) g_twinDeletionSinks->samplers.push_back(ids[i]);
     }
+    void TW_GenVertexArrays(GLsizei count, GLuint* ids) {
+        for (GLsizei i = 0; i < count; ++i) ids[i] = g_nextTwinDriverId++;
+    }
+    void TW_DeleteVertexArrays(GLsizei count, const GLuint* ids) {
+        if (!g_twinDeletionSinks) return;
+        for (GLsizei i = 0; i < count; ++i) g_twinDeletionSinks->vertexArrays.push_back(ids[i]);
+    }
+    GLuint TW_CreateProgram() { return g_nextTwinDriverId++; }
+    void TW_DeleteProgram(GLuint program) {
+        if (g_twinDeletionSinks) g_twinDeletionSinks->programs.push_back(program);
+    }
+    void TW_GenBuffers(GLsizei count, GLuint* ids) {
+        for (GLsizei i = 0; i < count; ++i) ids[i] = g_nextTwinDriverId++;
+    }
+    void TW_DeleteBuffers(GLsizei count, const GLuint* ids) {
+        if (!g_twinDeletionSinks) return;
+        for (GLsizei i = 0; i < count; ++i) g_twinDeletionSinks->buffers.push_back(ids[i]);
+    }
     void TW_BindFramebuffer(GLenum target, GLuint framebuffer) {
         SG_Log("BindFramebuffer:" + std::to_string(target) + ":" + std::to_string(framebuffer));
     }
@@ -1979,6 +2040,12 @@ namespace {
             functions.glGenSamplers = TW_GenSamplers;
             functions.glDeleteSamplers = TW_DeleteSamplers;
             functions.glBindSampler = TW_BindSampler;
+            functions.glGenVertexArrays = TW_GenVertexArrays;
+            functions.glDeleteVertexArrays = TW_DeleteVertexArrays;
+            functions.glCreateProgram = TW_CreateProgram;
+            functions.glDeleteProgram = TW_DeleteProgram;
+            functions.glGenBuffers = TW_GenBuffers;
+            functions.glDeleteBuffers = TW_DeleteBuffers;
             functions.glGetError = SG_NoError;
             MobileGL::MG_Backend::DirectGLES::SetGLESFuncsTable(functions);
             g_twinDeletionSinks = &sinks;
@@ -2078,6 +2145,145 @@ TEST(DirectGLESBackendSampler, DestructorDeletesIdAndScrubsUnitCache) {
     }
 }
 
+TEST(DirectGLESBackendVertexArray, DestructorDeletesIdAndHonorsContextGeneration) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedBackendTwinMocks mocks;
+
+    GLuint id = 0;
+    {
+        auto backendVao = MobileGL::MakeShared<VertexArrayImpl::BackendVertexArrayObject>();
+        id = backendVao->GetBackendVertexArrayId();
+        ASSERT_NE(id, 0u);
+    }
+    ASSERT_EQ(mocks.sinks.vertexArrays.size(), 1u);
+    EXPECT_EQ(mocks.sinks.vertexArrays[0], id);
+
+    // A twin whose context died must NOT delete a VAO name a successor context
+    // may already have recycled (both contexts restart GL names at 1).
+    {
+        auto backendVao = MobileGL::MakeShared<VertexArrayImpl::BackendVertexArrayObject>();
+        ++g_backendContextGeneration;
+        backendVao.reset();
+        --g_backendContextGeneration; // restore for later tests
+        EXPECT_EQ(mocks.sinks.vertexArrays.size(), 1u);
+    }
+}
+
+TEST(DirectGLESBackendProgram, DestructorDeletesIdAndHonorsContextGeneration) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedBackendTwinMocks mocks;
+
+    GLuint id = 0;
+    {
+        auto backendProgram = MobileGL::MakeShared<PrgramImpl::BackendProgramObjectImpl>();
+        id = backendProgram->GetBackendProgramId();
+        ASSERT_NE(id, 0u);
+    }
+    ASSERT_EQ(mocks.sinks.programs.size(), 1u);
+    EXPECT_EQ(mocks.sinks.programs[0], id);
+
+    {
+        auto backendProgram = MobileGL::MakeShared<PrgramImpl::BackendProgramObjectImpl>();
+        ++g_backendContextGeneration;
+        backendProgram.reset();
+        --g_backendContextGeneration;
+        EXPECT_EQ(mocks.sinks.programs.size(), 1u);
+    }
+}
+
+TEST(DirectGLESBackendProgram, GlobalUboDeletionHonorsContextGeneration) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedBackendTwinMocks mocks;
+
+    MobileGL::Uint id = 123;
+    PrgramImpl::DeleteBackendProgramGlobalUbo(id, g_backendContextGeneration);
+    EXPECT_EQ(id, 0u);
+    ASSERT_EQ(mocks.sinks.buffers.size(), 1u);
+    EXPECT_EQ(mocks.sinks.buffers[0], 123u);
+
+    // A buffer belonging to a dead context must be abandoned, never deleted as a
+    // recycled name in the successor context.
+    id = 124;
+    PrgramImpl::DeleteBackendProgramGlobalUbo(id, g_backendContextGeneration - 1);
+    EXPECT_EQ(id, 0u);
+    EXPECT_EQ(mocks.sinks.buffers.size(), 1u);
+}
+
+namespace {
+    struct SyncDeleteRacePayload {
+        std::atomic<MobileGL::Bool> alive{true};
+    };
+    std::atomic<MobileGL::Int> g_syncRaceDeleteCount{0};
+
+    MobileGL::MG_Backend::BackendSyncHandle SyncRaceFenceSync() {
+        return new SyncDeleteRacePayload();
+    }
+
+    GLenum SyncRaceClientWaitSync(MobileGL::MG_Backend::BackendSyncHandle handle, GLbitfield, GLuint64) {
+        auto* payload = static_cast<SyncDeleteRacePayload*>(handle);
+        // Keep the backend call in flight while the GL thread runs DeleteSync. The
+        // frontend must not release the backend handle (or the SyncObject wrapper)
+        // until this call has returned.
+        std::this_thread::sleep_for(std::chrono::milliseconds(20));
+        return payload->alive.load(std::memory_order_acquire) ? GL_ALREADY_SIGNALED : GL_WAIT_FAILED;
+    }
+
+    void SyncRaceWaitSync(MobileGL::MG_Backend::BackendSyncHandle, GLbitfield, GLuint64) {}
+
+    void SyncRaceDeleteSync(MobileGL::MG_Backend::BackendSyncHandle handle) {
+        auto* payload = static_cast<SyncDeleteRacePayload*>(handle);
+        payload->alive.store(false, std::memory_order_release);
+        delete payload;
+        g_syncRaceDeleteCount.fetch_add(1, std::memory_order_relaxed);
+    }
+
+    MobileGL::Bool SyncRaceGetSyncStatus(MobileGL::MG_Backend::BackendSyncHandle) { return true; }
+
+    struct ScopedSyncRaceBackend {
+        ScopedSyncRaceBackend(): previous(MobileGL::MG_Backend::gBackendFunctionsTable) {
+            MobileGL::MG_Backend::GlobalBackendFunctionsTable functions{};
+            functions.GL.FenceSync = SyncRaceFenceSync;
+            functions.GL.ClientWaitSync = SyncRaceClientWaitSync;
+            functions.GL.WaitSync = SyncRaceWaitSync;
+            functions.GL.DeleteSync = SyncRaceDeleteSync;
+            functions.GL.GetSyncStatus = SyncRaceGetSyncStatus;
+            MobileGL::MG_Backend::gBackendFunctionsTable = functions;
+            g_syncRaceDeleteCount.store(0, std::memory_order_relaxed);
+        }
+
+        ~ScopedSyncRaceBackend() {
+            MobileGL::MG_Impl::GLImpl::DestroyAllSyncObjects();
+            MobileGL::MG_Backend::gBackendFunctionsTable = previous;
+        }
+
+        ScopedSyncRaceBackend(const ScopedSyncRaceBackend&) = delete;
+        ScopedSyncRaceBackend& operator=(const ScopedSyncRaceBackend&) = delete;
+
+        MobileGL::MG_Backend::GlobalBackendFunctionsTable previous;
+    };
+} // namespace
+
+TEST(SyncLifetime, DeleteWaitsForInFlightClientWait) {
+    ScopedSyncRaceBackend backend;
+
+    const GLsync sync = MobileGL::MG_Impl::GLImpl::FenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+    ASSERT_NE(sync, nullptr);
+
+    GLenum clientResult = GL_WAIT_FAILED;
+    std::thread waiter([sync, &clientResult] {
+        clientResult = MobileGL::MG_Impl::GLImpl::ClientWaitSync(sync, 0, 0);
+    });
+
+    // Give the worker a head start so ClientWaitSync is already inside the stub
+    // (and therefore holds the per-object lock) when DeleteSync runs.
+    std::this_thread::sleep_for(std::chrono::milliseconds(5));
+    MobileGL::MG_Impl::GLImpl::DeleteSync(sync);
+    waiter.join();
+
+    EXPECT_EQ(clientResult, GL_ALREADY_SIGNALED);
+    EXPECT_EQ(g_syncRaceDeleteCount.load(std::memory_order_relaxed), 1);
+}
+
 TEST(DirectGLESStateGuards, DefaultFramebufferBindGoesThroughShadow) {
     using namespace MobileGL::MG_Backend::DirectGLES;
     ScopedStateGuardMocks mocks;

MobileGL/MG_Test/State/RenderStateTest.cpp

@@ -6,11 +6,20 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header
 //
-// Indexed capability state (glEnablei/glDisablei/glIsEnabledi) exists only for GL_BLEND in this
-// stack. Every other capability must come back as GL_INVALID_ENUM per GL 4.6 sec. 17.3.3 - and,
-// far more importantly, must come back at all: RenderState::SetCapabilityIndexed and
-// IsCapabilityEnabledIndexed used to answer a non-blend capability with THROW_UNIMPL_EXCEPTION,
+// Indexed capability state (glEnablei/glDisablei/glIsEnabledi) exists for exactly two
+// capabilities: GL_BLEND, indexed by draw buffer, and GL_SCISSOR_TEST, indexed by viewport
+// (ARB_viewport_array). Every other capability must come back as GL_INVALID_ENUM per GL 4.6
+// sec. 17.3.3 - and, far more importantly, must come back at all: RenderState::SetCapabilityIndexed
+// and IsCapabilityEnabledIndexed used to answer a non-blend capability with THROW_UNIMPL_EXCEPTION,
 // which unwinds a C++ exception through the C GL ABI and terminates the process.
+//
+// The second half of this file is the ARB_viewport_array indexed rectangle state. Every one of
+// glViewportArrayv/glViewportIndexedf(v)/glScissorArrayv/glScissorIndexed(v)/glDepthRangeArrayv/
+// glDepthRangeIndexed was a MGLOG_W_ONCE stub that raised no error and stored nothing, and the
+// indexed getters answered EVERY index with viewport 0's value, so a set/get round trip silently
+// reported the initial state. The assertions below are deliberately state-shaped rather than
+// render-shaped: this IS the state machine, and the rendering half (gl_ViewportIndex routing) is
+// asserted separately in MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp.
 
 #include <gtest/gtest.h>
 
@@ -21,6 +30,7 @@
 #include <MG_Impl/GLImpl/RenderState/GL_RenderState.h>
 #include <MG_State/GLState/Core.h>
 #include <MG_State/GLState/FramebufferState/FramebufferObject.h>
+#include <MG_State/GLState/RenderState/RenderState.h>
 
 using namespace MobileGL;
 
@@ -50,10 +60,11 @@ namespace {
     };
 } // namespace
 
-TEST_F(RenderStateTest, IndexedCapabilityTogglesRejectNonBlendCapabilities) {
+TEST_F(RenderStateTest, IndexedCapabilityTogglesRejectNonIndexedCapabilities) {
     // GL_CLIP_DISTANCE0 is a real capability, just not an indexed one - the shape an application or
-    // a CTS negative test would hit.
-    for (const GLenum cap : {GL_CLIP_DISTANCE0, GL_DEPTH_TEST, GL_SCISSOR_TEST}) {
+    // a CTS negative test would hit. GL_SCISSOR_TEST used to be in this list and is not any more:
+    // ARB_viewport_array makes it the second indexed capability (see the tests below).
+    for (const GLenum cap : {GL_CLIP_DISTANCE0, GL_DEPTH_TEST, GL_STENCIL_TEST}) {
         MG_Impl::GLImpl::Enablei(cap, 0);
         ExpectSingleGlError(GL_INVALID_ENUM);
 
@@ -89,3 +100,462 @@ TEST_F(RenderStateTest, IndexedBlendTogglesStillWork) {
     EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_BLEND, 1), GL_FALSE);
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
+
+// ---------------------------------------------------------------------------------------------
+// ARB_viewport_array: indexed viewport / scissor / depth-range state
+// ---------------------------------------------------------------------------------------------
+
+namespace {
+    constexpr GLuint kMaxViewports = RenderStateParameters::MAX_VIEWPORTS;
+
+    Array<Array<GLfloat, 4>, kMaxViewports> ReadAllViewports() {
+        Array<Array<GLfloat, 4>, kMaxViewports> out{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, out[i].data());
+        }
+        return out;
+    }
+
+    Array<Array<GLdouble, 2>, kMaxViewports> ReadAllDepthRanges() {
+        Array<Array<GLdouble, 2>, kMaxViewports> out{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, i, out[i].data());
+        }
+        return out;
+    }
+} // namespace
+
+TEST_F(RenderStateTest, ScissorTestIsIndexedByViewport) {
+    // The exact shape of KHR-GL43.viewport_array.scissor_test_state_api's toggle loop: one index
+    // is flipped and EVERY index is read back, so a broadcast masquerading as an indexed write
+    // cannot pass.
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint toggled = 0; toggled < kMaxViewports; ++toggled) {
+        MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, toggled);
+        EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR) << "index " << toggled;
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), i == toggled ? GL_TRUE : GL_FALSE)
+                << "enabled index " << toggled << ", read index " << i;
+        }
+        MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, toggled);
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, toggled), GL_FALSE);
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, NonIndexedScissorTestEnableWritesEveryViewport) {
+    // GL 4.6 core 17.3.2: Enable/Disable(SCISSOR_TEST) is "for all viewports". Reading only
+    // index 0 back would let a broadcast-less implementation through, so every index is checked.
+    MG_Impl::GLImpl::Enable(GL_SCISSOR_TEST);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), GL_TRUE) << "index " << i;
+    }
+    // ... and the non-indexed query answers for viewport 0 (GL 4.6 core 22.1).
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_TRUE);
+
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), GL_FALSE) << "index " << i;
+    }
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_FALSE);
+
+    // An indexed enable on a NON-zero index must not move the non-indexed answer.
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, 3);
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_FALSE);
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, 0);
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_TRUE);
+
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ScissorTestEnableRejectsAnOutOfRangeViewportIndex) {
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, kMaxViewports);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, kMaxViewports);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, kMaxViewports), GL_FALSE);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    // MAX_VIEWPORTS - 1 is the last LEGAL index and must stay silent.
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, kMaxViewports - 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, kMaxViewports - 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, MaxViewportsMatchesTheIndexedStateWidth) {
+    // The advertised limit and the width of the state arrays are the same number by
+    // construction; a divergence would make some index simultaneously legal to the CTS and
+    // out of range to the setters.
+    GLint maxViewports = 0;
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+    ExpectSingleGlError(GL_NO_ERROR);
+    EXPECT_EQ(maxViewports, static_cast<GLint>(kMaxViewports));
+    EXPECT_GE(maxViewports, 16) << "GL 4.3 core requires MAX_VIEWPORTS >= 16";
+}
+
+TEST_F(RenderStateTest, ViewportArrayvRoundTripsThroughEveryGetterWidth) {
+    Array<GLfloat, kMaxViewports * 4> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        written[i * 4 + 0] = static_cast<GLfloat>(i) + 0.125f;
+        written[i * 4 + 1] = static_cast<GLfloat>(i) + 0.25f;
+        written[i * 4 + 2] = static_cast<GLfloat>(64 + i);
+        written[i * 4 + 3] = static_cast<GLfloat>(32 + i);
+    }
+    MG_Impl::GLImpl::ViewportArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLfloat asFloat[4] = {};
+        MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, asFloat);
+        // Bit-exact: the fractional origin is the whole point of float viewport state, and the
+        // CTS compares with == (0.125 and 0.25 are exact binary fractions, so this is fair).
+        EXPECT_EQ(asFloat[0], written[i * 4 + 0]) << "index " << i << " must round-trip verbatim";
+        EXPECT_EQ(asFloat[1], written[i * 4 + 1]) << "index " << i;
+        EXPECT_EQ(asFloat[2], written[i * 4 + 2]) << "index " << i;
+        EXPECT_EQ(asFloat[3], written[i * 4 + 3]) << "index " << i;
+
+        GLdouble asDouble[4] = {};
+        MG_Impl::GLImpl::GetDoublei_v(GL_VIEWPORT, i, asDouble);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(asDouble[c], static_cast<GLdouble>(written[i * 4 + c])) << "index " << i << " component " << c;
+        }
+
+        // The integer widths round to nearest rather than truncate; the .5+ case is pinned by
+        // ViewportRoundsRatherThanTruncatesForIntegerQueries below.
+        GLint asInt[4] = {};
+        MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, i, asInt);
+        EXPECT_EQ(asInt[2], static_cast<GLint>(64 + i)) << "index " << i;
+        EXPECT_EQ(asInt[3], static_cast<GLint>(32 + i)) << "index " << i;
+
+        GLint64 asInt64[4] = {};
+        MG_Impl::GLImpl::GetInteger64i_v(GL_VIEWPORT, i, asInt64);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(asInt64[c], static_cast<GLint64>(asInt[c])) << "index " << i << " component " << c;
+        }
+
+        GLboolean asBool[4] = {};
+        MG_Impl::GLImpl::GetBooleani_v(GL_VIEWPORT, i, asBool);
+        EXPECT_EQ(asBool[2], GL_TRUE) << "index " << i << ": a non-zero width is GL_TRUE";
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ViewportRoundsRatherThanTruncatesForIntegerQueries) {
+    MG_Impl::GLImpl::ViewportIndexedf(2, 0.0f, 0.0f, 255.875f, 63.5f);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    GLint asInt[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, 2, asInt);
+    EXPECT_EQ(asInt[2], 256);
+    EXPECT_EQ(asInt[3], 64);
+
+    GLfloat asFloat[4] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, 2, asFloat);
+    EXPECT_EQ(asFloat[2], 255.875f) << "the integer query must not disturb the stored float";
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ViewportIndexedWritesTouchExactlyOneIndex) {
+    MG_Impl::GLImpl::Viewport(0, 0, 8, 8);
+    const auto before = ReadAllViewports();
+
+    for (GLuint target = 0; target < kMaxViewports; ++target) {
+        const GLfloat value[4] = {0.375f, 0.375f, 0.625f, 0.625f};
+        // Alternate the two indexed entry points so both are covered by the isolation claim.
+        if (target % 2 == 0) {
+            MG_Impl::GLImpl::ViewportIndexedf(target, value[0], value[1], value[2], value[3]);
+        } else {
+            MG_Impl::GLImpl::ViewportIndexedfv(target, value);
+        }
+        ExpectSingleGlError(GL_NO_ERROR);
+
+        const auto after = ReadAllViewports();
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            if (i == target) {
+                EXPECT_EQ(after[i][0], value[0]) << "index " << i;
+                EXPECT_EQ(after[i][2], value[2]) << "index " << i;
+            } else {
+                EXPECT_EQ(after[i], before[i]) << "write to " << target << " disturbed index " << i;
+            }
+        }
+        MG_Impl::GLImpl::ViewportIndexedf(target, before[target][0], before[target][1], before[target][2],
+                                          before[target][3]);
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ClassicViewportWritesEveryIndexAndIsVisibleThroughIndexZero) {
+    // Both directions of the aliasing. ARB_viewport_array defines glViewport as ViewportIndexedf
+    // on every index, and glGetIntegerv(GL_VIEWPORT) as viewport 0.
+    MG_Impl::GLImpl::ViewportIndexedf(5, 1.0f, 2.0f, 3.0f, 4.0f);
+    MG_Impl::GLImpl::Viewport(0, 0, 1, 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLfloat data[4] = {};
+        MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, data);
+        EXPECT_EQ(data[0], 0.0f) << "index " << i;
+        EXPECT_EQ(data[2], 1.0f) << "glViewport must overwrite index " << i;
+    }
+
+    MG_Impl::GLImpl::ViewportIndexedf(0, 4.0f, 5.0f, 6.0f, 7.0f);
+    GLint classic[4] = {};
+    MG_Impl::GLImpl::GetIntegerv(GL_VIEWPORT, classic);
+    EXPECT_EQ(classic[0], 4);
+    EXPECT_EQ(classic[2], 6);
+    GLfloat classicFloat[4] = {};
+    MG_Impl::GLImpl::GetFloatv(GL_VIEWPORT, classicFloat);
+    EXPECT_EQ(classicFloat[2], 6.0f);
+    // Index 5 keeps its own value: writing index 0 is not a broadcast.
+    GLfloat other[4] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, 5, other);
+    EXPECT_EQ(other[2], 1.0f);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ScissorBoxRoundTripsPerIndexAndAliasesIndexZero) {
+    Array<GLint, kMaxViewports * 4> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        written[i * 4 + 0] = static_cast<GLint>(i);
+        written[i * 4 + 1] = static_cast<GLint>(i * 2);
+        written[i * 4 + 2] = static_cast<GLint>(16 + i);
+        written[i * 4 + 3] = static_cast<GLint>(8 + i);
+    }
+    MG_Impl::GLImpl::ScissorArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLint readBack[4] = {};
+        MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, i, readBack);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(readBack[c], written[i * 4 + c]) << "index " << i << " component " << c;
+        }
+    }
+
+    // Indexed writes stay indexed; both spellings.
+    MG_Impl::GLImpl::ScissorIndexed(4, 4, 4, 8, 8);
+    const GLint indexedV[4] = {9, 9, 12, 12};
+    MG_Impl::GLImpl::ScissorIndexedv(7, indexedV);
+    ExpectSingleGlError(GL_NO_ERROR);
+    GLint probe[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 4, probe);
+    EXPECT_EQ(probe[2], 8);
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 7, probe);
+    EXPECT_EQ(probe[2], 12);
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 5, probe);
+    EXPECT_EQ(probe[2], static_cast<GLint>(16 + 5)) << "index 5 must be untouched";
+
+    // glScissor writes every rectangle, and glGetIntegerv(GL_SCISSOR_BOX) reports rectangle 0.
+    MG_Impl::GLImpl::Scissor(2, 3, 5, 6);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, i, probe);
+        EXPECT_EQ(probe[0], 2) << "index " << i;
+        EXPECT_EQ(probe[2], 5) << "index " << i;
+    }
+    GLint classic[4] = {};
+    MG_Impl::GLImpl::GetIntegerv(GL_SCISSOR_BOX, classic);
+    EXPECT_EQ(classic[2], 5);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, DepthRangeRoundTripsPerIndexAndAliasesIndexZero) {
+    Array<GLdouble, kMaxViewports * 2> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        // Exact binary fractions, like the CTS uses: a float-backed store round-trips them.
+        written[i * 2 + 0] = static_cast<GLdouble>(i) / 16.0;
+        written[i * 2 + 1] = 1.0 - static_cast<GLdouble>(i) / 16.0;
+    }
+    MG_Impl::GLImpl::DepthRangeArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    const auto readBack = ReadAllDepthRanges();
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(readBack[i][0], written[i * 2 + 0]) << "index " << i;
+        EXPECT_EQ(readBack[i][1], written[i * 2 + 1]) << "index " << i;
+    }
+
+    MG_Impl::GLImpl::DepthRangeIndexed(9, 0.25, 0.75);
+    ExpectSingleGlError(GL_NO_ERROR);
+    GLdouble probe[2] = {};
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, 9, probe);
+    EXPECT_EQ(probe[0], 0.25);
+    EXPECT_EQ(probe[1], 0.75);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, 8, probe);
+    EXPECT_EQ(probe[0], 8.0 / 16.0) << "index 8 must be untouched";
+
+    GLfloat asFloat[2] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_DEPTH_RANGE, 9, asFloat);
+    EXPECT_EQ(asFloat[0], 0.25f);
+    EXPECT_EQ(asFloat[1], 0.75f);
+
+    // glDepthRange writes every range; glGetDoublev(GL_DEPTH_RANGE) reports range 0.
+    MG_Impl::GLImpl::DepthRange(0.0, 1.0);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, i, probe);
+        EXPECT_EQ(probe[0], 0.0) << "index " << i;
+        EXPECT_EQ(probe[1], 1.0) << "index " << i;
+    }
+    MG_Impl::GLImpl::DepthRangeIndexed(0, 0.125, 0.875);
+    GLdouble classic[2] = {};
+    MG_Impl::GLImpl::GetDoublev(GL_DEPTH_RANGE, classic);
+    EXPECT_EQ(classic[0], 0.125);
+    EXPECT_EQ(classic[1], 0.875);
+    MG_Impl::GLImpl::DepthRange(0.0, 1.0);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, IndexedRectangleSettersRejectAnOutOfRangeIndex) {
+    const GLfloat viewport[4] = {0.0f, 0.0f, 1.0f, 1.0f};
+    const GLint scissor[4] = {0, 0, 1, 1};
+
+    for (const GLuint index : {kMaxViewports, kMaxViewports + 1}) {
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, 1.0f, 1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ViewportIndexedfv(index, viewport);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorIndexed(index, 0, 0, 1, 1);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorIndexedv(index, scissor);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::DepthRangeIndexed(index, 0.0, 1.0);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+    }
+
+    // The last legal index must stay silent - api_errors checks both sides of the boundary.
+    MG_Impl::GLImpl::ViewportIndexedf(kMaxViewports - 1, 0.0f, 0.0f, 1.0f, 1.0f);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::ScissorIndexed(kMaxViewports - 1, 0, 0, 1, 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::DepthRangeIndexed(kMaxViewports - 1, 0.0, 1.0);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ArraySettersRejectAnOutOfRangeRangeButAcceptAnExactlyFullOne) {
+    Array<GLfloat, kMaxViewports * 4> viewports{};
+    Array<GLint, kMaxViewports * 4> scissors{};
+    Array<GLdouble, kMaxViewports * 2> depths{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        viewports[i * 4 + 2] = 1.0f;
+        viewports[i * 4 + 3] = 1.0f;
+        scissors[i * 4 + 2] = 1;
+        scissors[i * 4 + 3] = 1;
+        depths[i * 2 + 1] = 1.0;
+    }
+
+    // first == MAX_VIEWPORTS, and first + count > MAX_VIEWPORTS.
+    MG_Impl::GLImpl::ViewportArrayv(kMaxViewports, 1, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ViewportArrayv(1, kMaxViewports, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(kMaxViewports, 1, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(1, kMaxViewports, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(kMaxViewports, 1, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(1, kMaxViewports, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    // first + count == MAX_VIEWPORTS is LEGAL - the off-by-one an ">=" bound would get wrong,
+    // and one KHR-GL43.viewport_array.api_errors asserts explicitly.
+    MG_Impl::GLImpl::ViewportArrayv(1, kMaxViewports - 1, viewports.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::ScissorArrayv(1, kMaxViewports - 1, scissors.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::DepthRangeArrayv(1, kMaxViewports - 1, depths.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    // A negative count is GL_INVALID_VALUE and must not be read as a huge unsigned length.
+    MG_Impl::GLImpl::ViewportArrayv(0, -1, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(0, -1, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(0, -1, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+}
+
+TEST_F(RenderStateTest, NegativeExtentsAreRejectedWithoutDisturbingState) {
+    MG_Impl::GLImpl::Viewport(0, 0, 4, 4);
+    MG_Impl::GLImpl::Scissor(0, 0, 4, 4);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    MG_Impl::GLImpl::Viewport(0, 0, -1, 1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Viewport(0, 0, 1, -1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Scissor(0, 0, -1, 1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Scissor(0, 0, 1, -1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    for (GLuint index = 0; index < kMaxViewports; ++index) {
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, -1.0f, 1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, 1.0f, -1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        const GLfloat badW[4] = {0.0f, 0.0f, -1.0f, 1.0f};
+        MG_Impl::GLImpl::ViewportIndexedfv(index, badW);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        MG_Impl::GLImpl::ScissorIndexed(index, 0, 0, -1, 1);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        const GLint badH[4] = {0, 0, 1, -1};
+        MG_Impl::GLImpl::ScissorIndexedv(index, badH);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        // The array form must reject the WHOLE call for one bad element, exactly once, and
+        // leave every rectangle alone - api_errors submits a full 16-element array with a
+        // single negative extent and then requires the error queue to hold one entry.
+        Array<GLfloat, kMaxViewports * 4> viewports{};
+        Array<GLint, kMaxViewports * 4> scissors{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            viewports[i * 4 + 2] = 1.0f;
+            viewports[i * 4 + 3] = 1.0f;
+            scissors[i * 4 + 2] = 1;
+            scissors[i * 4 + 3] = 1;
+        }
+        viewports[index * 4 + 2] = -1.0f;
+        scissors[index * 4 + 3] = -1;
+        MG_Impl::GLImpl::ViewportArrayv(0, kMaxViewports, viewports.data());
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorArrayv(0, kMaxViewports, scissors.data());
+        ExpectSingleGlError(GL_INVALID_VALUE);
+    }
+
+    // Nothing above may have landed.
+    GLint viewport[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, 0, viewport);
+    EXPECT_EQ(viewport[2], 4);
+    EXPECT_EQ(viewport[3], 4);
+    GLint scissor[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 0, scissor);
+    EXPECT_EQ(scissor[2], 4);
+    EXPECT_EQ(scissor[3], 4);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, IndexedRectangleQueriesRejectAnOutOfRangeIndex) {
+    GLint ints[4] = {};
+    GLfloat floats[4] = {};
+    GLdouble doubles[4] = {};
+
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, kMaxViewports, ints);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, kMaxViewports, floats);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, kMaxViewports, doubles);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, kMaxViewports - 1, ints);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, kMaxViewports - 1, floats);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, kMaxViewports - 1, doubles);
+    ExpectSingleGlError(GL_NO_ERROR);
+}

MobileGL/MG_Test/Texture/TextureTest.cpp

@@ -4033,3 +4033,306 @@ TEST_F(TextureTest, TexStorage2DLeavesAGenericCompressedFormatUncompressed) {
     EXPECT_EQ(compressed, GL_FALSE);
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
+
+// ===================== glCopyImageSubData validation (KHR-GL43.copy_image) =====================
+//
+// Every case below is a mechanism the conformance group caught in the field, and each one is
+// pinned here because the backend cannot: a wrongly ACCEPTED copy shows up only as wrong pixels
+// on a device, and a wrongly REJECTED one shows up only as a conformance failure.
+
+namespace {
+    struct CopyImageSubDataCall {
+        Bool Called = false;
+        GLenum SrcTarget = GL_NONE;
+        GLenum DstTarget = GL_NONE;
+        GLint SrcZ = -1;
+        GLint DstZ = -1;
+        GLsizei Depth = -1;
+    } g_copyImageSubDataCall;
+
+    void RecordCopyImageSubData(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture, GLenum srcTarget,
+                                GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
+                                const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture, GLenum dstTarget,
+                                GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ, GLsizei srcWidth,
+                                GLsizei srcHeight, GLsizei srcDepth) {
+        (void)srcTexture;
+        (void)srcLevel;
+        (void)srcX;
+        (void)srcY;
+        (void)dstTexture;
+        (void)dstLevel;
+        (void)dstX;
+        (void)dstY;
+        (void)srcWidth;
+        (void)srcHeight;
+        g_copyImageSubDataCall = {true, srcTarget, dstTarget, srcZ, dstZ, srcDepth};
+    }
+
+    // Two storage-backed 2D textures of the requested formats, so a copy between them is a legal
+    // call in every respect except the one the test is about.
+    void MakeCopyImagePair(GLenum srcFormat, GLenum dstFormat, GLuint& srcTexture, GLuint& dstTexture,
+                           GLsizei levels = 1, GLsizei extent = 8) {
+        MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D, 1, &srcTexture);
+        MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D, 1, &dstTexture);
+        MG_Impl::GLImpl::TextureStorage2D(srcTexture, levels, srcFormat, extent, extent);
+        MG_Impl::GLImpl::TextureStorage2D(dstTexture, levels, dstFormat, extent, extent);
+    }
+} // namespace
+
+// GL 4.6 core 18.3.2 compatibility is texel-block SIZE, not base internal format. RGB10_A2 and
+// R11F_G11F_B10F are both 32-bit and their bases differ (RGBA vs RGB); the old exact-base-format
+// predicate rejected the pair, which is what took down the whole cross-format half of the
+// conformance matrix on both backends.
+TEST_F(TextureTest, CopyImageSubDataAcceptsEqualTexelSizeAcrossDifferentBaseFormats) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGB10_A2, GL_R11F_G11F_B10F, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D, 0, 0, 0, 0,
+                                      4, 4, 1);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// The other half of the same rule: equal base format is not sufficient either. RGBA8 and RGBA32F
+// are both RGBA and 32 vs 128 bits, so the copy is illegal.
+TEST_F(TextureTest, CopyImageSubDataRejectsDifferentTexelSizesWithTheSameBaseFormat) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGBA8, GL_RGBA32F, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D, 0, 0, 0, 0,
+                                      4, 4, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_OPERATION);
+}
+
+// ...and the pairing that is legal purely because the sizes agree, across integer-ness too.
+TEST_F(TextureTest, CopyImageSubDataAcceptsIntegerAndFloatOfTheSameTexelSize) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGBA32UI, GL_RGBA32F, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D, 0, 0, 0, 0,
+                                      4, 4, 1);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// 18.3.2 spells a name that is not an object INVALID_VALUE. The shared texture-object validator
+// says INVALID_OPERATION, which is right for the entry points that reach an object through a
+// BINDING - hence a rule local to this entry point rather than a change to the helper.
+TEST_F(TextureTest, CopyImageSubDataNonExistentNameIsInvalidValue) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    MG_Impl::GLImpl::CopyImageSubData(4242, GL_TEXTURE_2D, 0, 0, 0, 0, 4243, GL_TEXTURE_2D, 0, 0, 0, 0, 1, 1, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+}
+
+// A target that disagrees with the object it names is INVALID_ENUM, not the INVALID_OPERATION the
+// shared target-uniformity validator records for the upload paths.
+TEST_F(TextureTest, CopyImageSubDataTargetNotMatchingTheObjectIsInvalidEnum) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGBA8, GL_RGBA8, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D_ARRAY, 0, 0,
+                                      0, 0, 1, 1, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_ENUM);
+}
+
+// The eleven whole-image targets only: a cube FACE converts to a target the frontend knows, so the
+// generic target validator lets it through, but 18.3.2 does not accept it here.
+TEST_F(TextureTest, CopyImageSubDataRejectsTargetsOutsideTheSpecList) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGBA8, GL_RGBA8, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_CUBE_MAP_POSITIVE_X, 0, 0, 0, 0, dstTexture,
+                                      GL_TEXTURE_2D, 0, 0, 0, 0, 1, 1, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_ENUM);
+}
+
+// A level the image does not have is INVALID_VALUE; a single-level texture asked for level 1 used
+// to reach the backend with whatever the storage layer answered for that level.
+TEST_F(TextureTest, CopyImageSubDataRejectsLevelTheImageDoesNotHave) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MakeCopyImagePair(GL_RGBA8, GL_RGBA8, srcTexture, dstTexture);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D, 1, 0, 0, 0,
+                                      1, 1, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+}
+
+// Sample counts must match. A single-sample image reports zero, so this same comparison is also
+// what refuses a copy between a multisample target and a non-multisample one.
+TEST_F(TextureTest, CopyImageSubDataRejectsSampleCountMismatch) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    // Two DIFFERENT counts are the whole point, so the case needs a context that can actually
+    // create multisample storage - which this unit-test binary, with no backend behind the
+    // renderable-format and sample-count queries, may not be able to. The precondition is
+    // checked on the state objects rather than assumed, so this can only ever skip or test the
+    // real rule; it can never pass vacuously.
+    GLint maxSamples = 1;
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_SAMPLES, &maxSamples);
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D_MULTISAMPLE, 1, &srcTexture);
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D_MULTISAMPLE, 1, &dstTexture);
+    MG_Impl::GLImpl::TextureStorage2DMultisample(srcTexture, 1, GL_RGBA8, 8, 8, GL_FALSE);
+    MG_Impl::GLImpl::TextureStorage2DMultisample(dstTexture, std::max(maxSamples, 2), GL_RGBA8, 8, 8, GL_FALSE);
+    DrainPendingGlErrors();
+
+    const Int srcSamples = MG_State::pGLContext->GetTextureObject(srcTexture)->GetSamples();
+    const Int dstSamples = MG_State::pGLContext->GetTextureObject(dstTexture)->GetSamples();
+    if (srcSamples == dstSamples) {
+        GTEST_SKIP() << "this context could not give the two textures different sample counts (both " << srcSamples
+                     << "); nothing for the rule to reject";
+    }
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D_MULTISAMPLE, 0, 0, 0, 0, dstTexture,
+                                      GL_TEXTURE_2D_MULTISAMPLE, 0, 0, 0, 0, 1, 1, 1);
+    EXPECT_FALSE(g_copyImageSubDataCall.Called);
+    ExpectSingleGlError(GL_INVALID_OPERATION);
+}
+
+// The layer range has to survive the frontend intact. Both backends used to drop it - DirectVulkan
+// pinned baseArrayLayer/layerCount at 0/1 - so a 12-layer copy moved one layer and said nothing;
+// this pins the frontend half of that contract.
+TEST_F(TextureTest, CopyImageSubDataForwardsTheWholeLayerRangeToTheBackend) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D_ARRAY, 1, &srcTexture);
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D_ARRAY, 1, &dstTexture);
+    MG_Impl::GLImpl::TextureStorage3D(srcTexture, 1, GL_RGBA8, 8, 8, 12);
+    MG_Impl::GLImpl::TextureStorage3D(dstTexture, 1, GL_RGBA8, 8, 8, 12);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 2, dstTexture, GL_TEXTURE_2D_ARRAY,
+                                      0, 0, 0, 5, 4, 4, 7);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(g_copyImageSubDataCall.SrcZ, 2);
+    EXPECT_EQ(g_copyImageSubDataCall.DstZ, 5);
+    EXPECT_EQ(g_copyImageSubDataCall.Depth, 7);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// The shape KHR-GL43.copy_image.invalid_object ends on once the invalid-name cases are answered
+// correctly: two ordinary glTexImage2D textures, no storage object, one texel copied from the
+// origin. Nothing about it is exotic, which is exactly why it is worth a case of its own - every
+// rule added to this validator is a new way to reject it.
+TEST_F(TextureTest, CopyImageSubDataAcceptsAPlainMutableTexImage2DPair) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &srcTexture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, srcTexture);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 16, 16, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
+    MG_Impl::GLImpl::TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
+    MG_Impl::GLImpl::GenTextures(1, &dstTexture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, dstTexture);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 16, 16, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
+    MG_Impl::GLImpl::TexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_2D, 0, 0, 0, 0, dstTexture, GL_TEXTURE_2D, 0, 0, 0, 0,
+                                      1, 1, 1);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    // ...and again after the names have been through a delete/regenerate cycle, which is what the
+    // conformance case does between its sub-cases: it deletes an object to make it invalid, then
+    // builds the next pair from names the allocator hands straight back.
+    MG_Impl::GLImpl::DeleteTextures(1, &srcTexture);
+    MG_Impl::GLImpl::DeleteTextures(1, &dstTexture);
+    DrainPendingGlErrors();
+    g_copyImageSubDataCall = {};
+
+    GLuint reusedSrc = 0;
+    GLuint reusedDst = 0;
+    MG_Impl::GLImpl::GenTextures(1, &reusedSrc);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, reusedSrc);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 16, 16, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::GenTextures(1, &reusedDst);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, reusedDst);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 16, 16, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(reusedSrc, GL_TEXTURE_2D, 0, 0, 0, 0, reusedDst, GL_TEXTURE_2D, 0, 0, 0, 0, 1,
+                                      1, 1);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// A rectangle target reaches the backend as itself. The translation to the GL_TEXTURE_2D the ES
+// driver actually stores it in belongs to DirectGLES, not here - and putting it here would break
+// DirectVulkan, which needs the real target to tell an array copy from a flat one.
+TEST_F(TextureTest, CopyImageSubDataPassesTheRectangleTargetThroughUntranslated) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyImageSubData = RecordCopyImageSubData;
+    g_copyImageSubDataCall = {};
+
+    GLuint srcTexture = 0;
+    GLuint dstTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_RECTANGLE, 1, &srcTexture);
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_RECTANGLE, 1, &dstTexture);
+    MG_Impl::GLImpl::TextureStorage2D(srcTexture, 1, GL_RGBA8, 8, 8);
+    MG_Impl::GLImpl::TextureStorage2D(dstTexture, 1, GL_RGBA8, 8, 8);
+    ASSERT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MG_Impl::GLImpl::CopyImageSubData(srcTexture, GL_TEXTURE_RECTANGLE, 0, 0, 0, 0, dstTexture, GL_TEXTURE_RECTANGLE,
+                                      0, 0, 0, 0, 4, 4, 1);
+    EXPECT_TRUE(g_copyImageSubDataCall.Called);
+    EXPECT_EQ(g_copyImageSubDataCall.SrcTarget, static_cast<GLenum>(GL_TEXTURE_RECTANGLE));
+    EXPECT_EQ(g_copyImageSubDataCall.DstTarget, static_cast<GLenum>(GL_TEXTURE_RECTANGLE));
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp

@@ -955,6 +955,8 @@ namespace MobileGL::MG_Util::BackendLoader {
                                  (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 2);
         const Bool esAtLeast31 = caps.GLESVersion.Major > 3 ||
                                  (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 1);
+        caps.SupportsDrawIndirect = esAtLeast31 && glesFuncs.glDrawArraysIndirect != nullptr &&
+                                    glesFuncs.glDrawElementsIndirect != nullptr;
         caps.SupportsDrawElementsBaseVertex = (esAtLeast32 || hasDrawElementsBaseVertexExtension) &&
                                               glesFuncs.glDrawElementsBaseVertex != nullptr;
         caps.SupportsComputeShader = esAtLeast31 && glesFuncs.glDispatchCompute != nullptr &&
@@ -976,6 +978,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         MGLOG_I("    indexed glColorMaski: %s", caps.SupportsIndexedColorMask ? "yes" : "no");
         MGLOG_I("    dual-source blend (EXT_blend_func_extended): %s",
                 caps.SupportsDualSourceBlend ? "yes" : "no");
+        MGLOG_I("    draw indirect (ES 3.1 core): %s", caps.SupportsDrawIndirect ? "yes" : "no");
         MGLOG_I("    multi-draw indirect (EXT_multi_draw_indirect): %s",
                 caps.SupportsMultiDrawIndirect ? "yes" : "no");
         MGLOG_I("    multi-draw base vertex (EXT/OES_draw_elements_base_vertex + EXT_multi_draw_arrays): %s",
@@ -1000,7 +1003,11 @@ namespace MobileGL::MG_Util::BackendLoader {
         GLfloat smoothLineWidthRange[2] = {1.0f, 1.0f};
         GLfloat smoothLineWidthGranularity = 1.0f;
         GLfloat aliasedPointSizeRange[2] = {1.0f, 1.0f};
-        GLfloat viewportBoundsRange[2] = {0.0f, 0.0f};
+        // GL 4.6 core table 23.60 sets the MINIMUM VIEWPORT_BOUNDS_RANGE at [-32768, 32767], and
+        // KHR-GL43.viewport_array.queries asserts exactly that floor. GLES has no such query, so
+        // the glGetFloatv below raises GL_INVALID_ENUM and leaves this untouched - starting it at
+        // {0, 0} advertised a range that admits no viewport origin at all.
+        GLfloat viewportBoundsRange[2] = {-32768.0f, 32767.0f};
         GLint maxViewportDims[2] = {16384, 16384};
         GLint viewportSubpixelBits = 0;
         GLint max3DTextureSize = 16384;
@@ -1289,8 +1296,12 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.MaxViewports = maxViewports;
         caps.MaxViewportWidth = maxViewportDims[0];
         caps.MaxViewportHeight = maxViewportDims[1];
-        caps.ViewportBoundsRangeMin = viewportBoundsRange[0];
-        caps.ViewportBoundsRangeMax = viewportBoundsRange[1];
+        // Only ever WIDER than the core minimum: a driver that answered the query is allowed to
+        // exceed the floor but never to sit inside it, and a driver that rejected the query left
+        // the floor in place. Written as a clamp rather than a plain assignment so a partial
+        // write (one component answered, the other not) cannot narrow the range either.
+        caps.ViewportBoundsRangeMin = std::min(viewportBoundsRange[0], -32768.0f);
+        caps.ViewportBoundsRangeMax = std::max(viewportBoundsRange[1], 32767.0f);
         caps.ViewportSubpixelBits = viewportSubpixelBits;
         caps.MinFragmentInterpolationOffset =
             std::isfinite(minFragmentInterpolationOffset) && minFragmentInterpolationOffset <= -0.5f

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.h

@@ -1149,6 +1149,10 @@ namespace MobileGL {
             // GLES 3.2 core or GL_OES_shader_multisample_interpolation exposes
             // interpolateAtOffset and the three fragment-offset limit queries.
             Bool SupportsShaderMultisampleInterpolation = false;
+            // ES 3.1+ exposes glDrawArraysIndirect / glDrawElementsIndirect in core. Keep the
+            // version and both entry-point checks together so extension advertisement and the
+            // DirectGLES dispatch path cannot disagree on whether native indirect draws exist.
+            Bool SupportsDrawIndirect = false;
             // GL_EXT_multi_draw_indirect is present AND glMultiDrawArraysIndirectEXT /
             // glMultiDrawElementsIndirectEXT both resolved. Multi-draw is not core in any ES
             // version, and eglGetProcAddress may return a live-looking stub on drivers without

MobileGL/MG_Util/SelfTest/DriverPost.cpp

@@ -1168,7 +1168,9 @@ namespace MobileGL::MG_Util::SelfTest {
             backendApiVersionString = MG_Backend::DirectGLES::FormatBackendAPIVersionString(
                 summary.caps.GLESRendererString, summary.caps.GLESVersion.Major, summary.caps.GLESVersion.Minor);
             advertisedExtensions = JoinAdvertisedExtensions(MG_Backend::DirectGLES::BuildAdvertisedExtensions(
-                summary.caps.SupportsDisjointTimerQuery, summary.caps.SupportsTextureFilterAnisotropy));
+                summary.caps.SupportsDisjointTimerQuery, summary.caps.SupportsTextureFilterAnisotropy,
+                summary.caps.SupportsDrawIndirect,
+                summary.caps.SupportsDrawIndirect && summary.caps.SupportsBaseInstance));
         }
         AppendMobileGLReportedRows(builder, MG_Backend::DirectGLES::GetRendererIdentity(), backendApiVersionString,
                                    advertisedExtensions);
@@ -1461,6 +1463,8 @@ namespace MobileGL::MG_Util::SelfTest {
             Bool shaderSubgroupUsable = false;
             Bool timerQueriesSupported = false;
             Bool samplerAnisotropySupported = false;
+            Bool drawIndirectFirstInstanceSupported = false;
+            Bool shaderDrawParametersSupported = false;
         };
     } // namespace
 
@@ -1744,6 +1748,7 @@ namespace MobileGL::MG_Util::SelfTest {
         VkPhysicalDeviceFeatures features{};
         vkGetPhysicalDeviceFeaturesFn(physicalDevice, &features);
         summary.samplerAnisotropySupported = features.samplerAnisotropy == VK_TRUE;
+        summary.drawIndirectFirstInstanceSupported = features.drawIndirectFirstInstance == VK_TRUE;
         if (features.multiDrawIndirect == VK_TRUE) {
             builder.Pass("multiDrawIndirect", "indirect multi-draw batches run as single native commands");
         } else {
@@ -1910,6 +1915,7 @@ namespace MobileGL::MG_Util::SelfTest {
             builder.Warn("shaderDrawParameters",
                          "unavailable; shaders using gl_DrawID/gl_BaseInstance will not work");
         }
+        summary.shaderDrawParametersSupported = shaderDrawParameters;
 
         Bool provokingVertexLast = false;
         Bool transformFeedbackPreservesProvokingVertex = false;
@@ -2108,7 +2114,8 @@ namespace MobileGL::MG_Util::SelfTest {
             backendApiVersionString = MG_Backend::DirectVulkan::FormatBackendAPIVersionString(
                 summary.deviceName, summary.apiVersionString, summary.driverVersionString);
             advertisedExtensions = JoinAdvertisedExtensions(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(
-                summary.shaderSubgroupUsable, summary.timerQueriesSupported, summary.samplerAnisotropySupported));
+                summary.shaderSubgroupUsable, summary.timerQueriesSupported, summary.samplerAnisotropySupported,
+                summary.drawIndirectFirstInstanceSupported && summary.shaderDrawParametersSupported));
         }
         AppendMobileGLReportedRows(builder, MG_Backend::DirectVulkan::GetRendererIdentity(), backendApiVersionString,
                                    advertisedExtensions);

MobileGL/MG_Util/ShaderTranspiler/EsslBuiltinFunctionNames.h

@@ -61,7 +61,7 @@ namespace MobileGL {
                 "imageAtomicXor", "imageLoad", "imageSize", "imageStore", "imulExtended",
                 "intBitsToFloat", "interpolateAtCentroid", "interpolateAtOffset",
                 "interpolateAtSample", "inverse", "inversesqrt", "isinf", "isnan",
-                "ldexp", "length", "lessThan", "lessThanEqual", "log", "log2",
+                "ldexp", "length", "length_squared", "lessThan", "lessThanEqual", "log", "log2",
                 "matrixCompMult", "max", "max3", "memoryBarrier",
                 "memoryBarrierAtomicCounter", "memoryBarrierBuffer", "memoryBarrierImage",
                 "memoryBarrierShared", "mid3", "min", "min3", "mix", "mod", "modf",

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameSamplerFunctionParameterPass.cpp

@@ -19,23 +19,27 @@ namespace MobileGL {
     namespace MG_Util {
         namespace ShaderTranspiler {
             namespace {
-                constexpr const char* kConflictingName = "sampler";
-                constexpr const char* kCompatName = "MGL_COMPAT_sampler";
+                const char* GetCompatName(StringView name) {
+                    if (name == "sampler") return "MGL_COMPAT_sampler";
+                    if (name == "new") return "MGL_COMPAT_new";
+                    return nullptr;
+                }
 
-                Bool IsNamedSamplerFunctionParameter(spvtools::opt::IRContext* context,
-                                                     spvtools::opt::Instruction& nameInst) {
+                const char* GetConflictingFunctionParameterCompatName(spvtools::opt::IRContext* context,
+                                                                       spvtools::opt::Instruction& nameInst) {
                     if (nameInst.opcode() != spv::Op::OpName || nameInst.NumInOperands() < 2) {
-                        return false;
+                        return nullptr;
                     }
 
-                    if (nameInst.GetInOperand(1).AsString() != kConflictingName) {
-                        return false;
+                    const char* compatName = GetCompatName(nameInst.GetInOperand(1).AsString());
+                    if (compatName == nullptr) {
+                        return nullptr;
                     }
 
                     auto* defUseMgr = context->get_def_use_mgr();
                     const Uint32 targetId = nameInst.GetSingleWordInOperand(0);
                     const auto* target = defUseMgr->GetDef(targetId);
-                    return target != nullptr && target->opcode() == spv::Op::OpFunctionParameter;
+                    return target != nullptr && target->opcode() == spv::Op::OpFunctionParameter ? compatName : nullptr;
                 }
             } // namespace
 
@@ -44,12 +48,13 @@ namespace MobileGL {
                 auto* irContext = context();
 
                 for (auto& debugInst : irContext->debugs2()) {
-                    if (!IsNamedSamplerFunctionParameter(irContext, debugInst)) {
+                    const char* compatName = GetConflictingFunctionParameterCompatName(irContext, debugInst);
+                    if (compatName == nullptr) {
                         continue;
                     }
 
                     debugInst.SetInOperand(
-                        1, spvtools::utils::MakeVector<spvtools::opt::Operand::OperandData>(kCompatName));
+                        1, spvtools::utils::MakeVector<spvtools::opt::Operand::OperandData>(compatName));
                     modified = true;
                 }
 

MobileGL/MG_Util/Types.h

@@ -52,6 +52,22 @@ namespace MobileGL {
     inline UniquePtr<T> MakeUnique(Args&&... args) {
         return std::make_unique<T>(std::forward<Args>(args)...);
     }
+    // RAII owner for the one-shot XXH64 state used by the Vulkan cache hashers.
+    // The previous `static inline XXH64_state_t*` form allocated five states per
+    // process and never called XXH64_freeState; a destructor here is independent of
+    // Vulkan/glslang teardown, so it is safe at static destruction time.
+    class XXH64State {
+    public:
+        XXH64State() : m_state(XXH64_createState()) {}
+        ~XXH64State() { XXH64_freeState(m_state); }
+        XXH64State(const XXH64State&) = delete;
+        XXH64State& operator=(const XXH64State&) = delete;
+
+        XXH64_state_t* Get() const { return m_state; }
+
+    private:
+        XXH64_state_t* m_state = nullptr;
+    };
     using SizeT = std::size_t;
     template <typename T, SizeT N>
     using Array = std::array<T, N>;

tools/trace_replay/README.md

@@ -11,6 +11,8 @@ The bundled fixtures cover:
   ![Minecraft 1.21.4 startup golden](fixtures/minecraft-1.21.4-startup.0000092195.png)
 - minecraft-1.21.4-main-menu: captured from Minecraft 1.21.4's main menu.
   ![Minecraft 1.21.4 main menu golden](fixtures/minecraft-1.21.4-main-menu.0000481787.png)
+- minecraft-1.21.11-main-menu: captured from Minecraft 1.21.11's main menu on a Pixel 8 Pro through FCL MobileGL.
+  ![Minecraft 1.21.11 main menu golden](fixtures/minecraft-1.21.11-main-menu.0000205347.png)
 - minecraft-1.17-main-menu-854: captured from Minecraft 1.17's 854x480 main menu through FCL MobileGL capture.
   ![Minecraft 1.17 854x480 main menu golden](fixtures/minecraft-1.17-main-menu-854.0000117757.png)
 - minecraft-1.21.4-in-world: captured from Minecraft 1.21.4 after entering a singleplayer world.

tools/trace_replay/trace_cases.json

@@ -40,6 +40,13 @@
       "target_call": 481787,
       "timeout_seconds": 180
     },
+    {
+      "name": "minecraft-1.21.11-main-menu",
+      "trace_archive": "minecraft-1.21.11-main-menu.tgz",
+      "golden": "minecraft-1.21.11-main-menu.0000205347.png",
+      "target_call": 205347,
+      "timeout_seconds": 180
+    },
     {
       "name": "minecraft-1.17-main-menu-854",
       "trace_archive": "minecraft-1.17-main-menu-854.tgz",