[Feat, Test] (MG_Backend/DirectVulkan, MG_Test, MG_IntegrationTest): rasterize the viewport gl_ViewportIndex selects, instead of collapsing all sixteen onto viewport 0

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -206,6 +206,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         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.viewportCount, sizeof(payload.viewportCount)));
         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)));
@@ -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

@@ -42,6 +42,14 @@ 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;

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) {
@@ -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

@@ -151,6 +151,12 @@ 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
@@ -400,6 +406,12 @@ 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 +446,9 @@ 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;

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -437,12 +437,24 @@ 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) {
+        // Rounded to integers on purpose: MobileGL advertises GL_VIEWPORT_SUBPIXEL_BITS = 0, so
+        // the fractional rectangle glViewportIndexedf can store is exact as STATE and snapped
+        // when it rasterizes.
+        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 +489,14 @@ 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 &&
@@ -4879,6 +4899,7 @@ 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 +5342,71 @@ 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.
@@ -5761,7 +5844,7 @@ 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;
     }
 
@@ -6201,7 +6284,8 @@ 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).
@@ -6220,6 +6304,7 @@ 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);
@@ -12216,6 +12301,28 @@ 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

@@ -547,6 +547,12 @@ 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 +836,11 @@ 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 +1131,22 @@ 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,