[Refactor] (MG_Config, DirectVulkan): route the blended depth-write quirk through the FeaturesTable as MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE instead of an ad-hoc getenv

.github/scripts/fetch-trace-fixture-lfs.sh

@@ -9,7 +9,23 @@ fi
 case_name="$1"
 fixture_dir="${2:-tools/trace_replay/fixtures}"
 python_bin="${PYTHON:-python3}"
-mirror_base="${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE:-https://repo.miawa.cn/mgl/tools/trace_replay/fixtures}"
+# Fixture mirrors, tried in order before falling back to Git LFS. Override the
+# whole list with MOBILEGL_TRACE_FIXTURE_MIRROR_BASES (whitespace separated);
+# MOBILEGL_TRACE_FIXTURE_MIRROR_BASE still works and is tried first.
+default_mirror_bases=(
+  "https://git.hit.moe/swung0x48/MobileGL/media/branch/dev/tools/trace_replay/fixtures"
+  "https://repo.miawa.cn/mgl/tools/trace_replay/fixtures"
+)
+if [ -n "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASES:-}" ]; then
+  read -r -a mirror_bases <<< "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASES}"
+else
+  mirror_bases=("${default_mirror_bases[@]}")
+fi
+if [ -n "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE:-}" ]; then
+  mirror_bases=("${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE}" "${mirror_bases[@]}")
+fi
+# Optional bearer token for mirrors that require authentication (private Gitea).
+mirror_token="${MOBILEGL_TRACE_FIXTURE_MIRROR_TOKEN:-}"
 download_attempts="${MOBILEGL_TRACE_FIXTURE_DOWNLOAD_ATTEMPTS:-5}"
 retry_delay="${MOBILEGL_TRACE_FIXTURE_RETRY_DELAY:-2}"
 
@@ -30,7 +46,8 @@ fixture_list="$("${python_bin}" tools/trace_replay/trace_cases.py \
   --format fixture-files \
   --case "${case_name}" \
   --fixture-root "${fixture_dir}")"
-mapfile -t files <<< "${fixture_list}"
+# Strip CR so the script also works when python emits CRLF (Git Bash on Windows).
+mapfile -t files < <(printf '%s\n' "${fixture_list}" | tr -d '\r')
 
 include="$(IFS=,; echo "${files[*]}")"
 if [ "${case_name}" = "OpenRA" ]; then
@@ -106,6 +123,7 @@ fetch_file_from_mirror() {
   local attempt
   local partial_size
   local curl_status
+  local curl_auth
 
   metadata="$(get_lfs_metadata "${file}")" || return 1
   read -r expected_oid expected_size <<< "${metadata}"
@@ -136,7 +154,11 @@ fetch_file_from_mirror() {
       echo "Starting mirror download for ${file} (attempt ${attempt}/${download_attempts})"
     fi
 
-    if curl -L --fail --show-error --continue-at - --output "${tmp_file}" "${url}"; then
+    curl_auth=()
+    if [ -n "${mirror_token}" ]; then
+      curl_auth=(--header "Authorization: token ${mirror_token}")
+    fi
+    if curl -L --fail --show-error --continue-at - "${curl_auth[@]}" --output "${tmp_file}" "${url}"; then
       if verify_fixture_file "${tmp_file}" "${file}" "${expected_oid}" "${expected_size}"; then
         mv "${tmp_file}" "${file}"
         return 0
@@ -184,10 +206,19 @@ fetch_from_mirror() {
   for file in "${files[@]}"; do
     local name
     local url
+    local base
+    local fetched=0
     name="$(basename "${file}")"
-    url="${mirror_base%/}/${name}"
-    echo "Fetching trace fixture from mirror: ${url}"
-    if ! fetch_file_from_mirror "${file}" "${url}"; then
+    for base in "${mirror_bases[@]}"; do
+      url="${base%/}/${name}"
+      echo "Fetching trace fixture from mirror: ${url}"
+      if fetch_file_from_mirror "${file}" "${url}"; then
+        fetched=1
+        break
+      fi
+      echo "Mirror did not serve ${name}; trying the next mirror" >&2
+    done
+    if [ "${fetched}" -ne 1 ]; then
       return 1
     fi
   done
@@ -196,7 +227,7 @@ fetch_from_mirror() {
 if fetch_from_mirror; then
   echo "Fetched trace fixture files for ${case_name} from mirror: ${include}"
 else
-  echo "Mirror fetch failed for ${case_name}; falling back to Git LFS: ${include}"
+  echo "All mirrors failed for ${case_name}; falling back to Git LFS: ${include}"
   git lfs install --local
   git lfs pull --include="${include}" --exclude=""
 fi

MobileGL/Config.h

@@ -20,6 +20,15 @@ namespace MobileGL::MG_Config {
 
     extern BackendType ActiveBackendType;
 
+    // Tri-state override for device-specific quirks: Auto lets the detected device decide,
+    // ForceOn/ForceOff bypass the detection in either direction. ForceOn only bypasses the
+    // device gate - each quirk keeps its structural safety checks.
+    enum class QuirkOverride : Uint8 {
+        Auto = 0,
+        ForceOn,
+        ForceOff,
+    };
+
     // Feature toggles parsed once from environment variables in MG_ConfigLoader::Init()
     // (ConfigLoader.cpp), before the accepted-env map is destroyed. All Bool fields share
     // one truthy rule: the variable is set, non-empty, not "0", and not "false"
@@ -67,6 +76,19 @@ namespace MobileGL::MG_Config {
         // explicitly request a core profile via EGL_CONTEXT_OPENGL_PROFILE_MASK / a >=3.1
         // version request.
         Bool RelaxedSemantics = false;
+        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN: overrides the shader-source quirk that
+        // rewrites the recognized workgroup prefix-scan template on Qualcomm devices with
+        // subgroups wider than 32 lanes (see ShaderSourceProcessor's quirk registry).
+        QuirkOverride SubgroupPrefixScanQuirk = QuirkOverride::Auto;
+        // MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE: overrides the DirectVulkan quirk that
+        // strips depth writes from blended pipelines on drivers without cross-pipeline
+        // vertex position invariance (see VulkanRenderer's PipelineFactory setup). Auto
+        // detects Qualcomm.
+        QuirkOverride MagmaDisableBlendedDepthWriteQuirk = QuirkOverride::Auto;
+        // MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS: leave the Vulkan robustBufferAccess device
+        // feature off. It is enabled by default to match GL's defined out-of-range fetch
+        // behavior; this escape hatch exists to measure or dodge its GPU cost on a device.
+        Bool DisableRobustBufferAccess = false;
     };
     extern FeaturesTable Features;
 } // namespace MobileGL::MG_Config

MobileGL/ConfigLoader.cpp

@@ -86,6 +86,17 @@ namespace MobileGL::MG_ConfigLoader {
         return it != acceptedEnvVariablesMap->end() && IsTruthyValue(it->second);
     }
 
+    // Quirk overrides are tri-state: an unset variable keeps device auto-detection, a truthy
+    // value forces the quirk on, anything else set ("0", "false", "") forces it off.
+    inline MG_Config::QuirkOverride QueryEnvQuirkOverride(const String& key) {
+        auto it = acceptedEnvVariablesMap->find(key);
+        if (it == acceptedEnvVariablesMap->end()) {
+            return MG_Config::QuirkOverride::Auto;
+        }
+        return IsTruthyValue(it->second) ? MG_Config::QuirkOverride::ForceOn
+                                         : MG_Config::QuirkOverride::ForceOff;
+    }
+
     inline Uint32 QueryEnvUint32(const String& key, Uint32 defaultValue, Uint32 minValue, Uint32 maxValue) {
         auto it = acceptedEnvVariablesMap->find(key);
         if (it == acceptedEnvVariablesMap->end()) {
@@ -123,6 +134,10 @@ namespace MobileGL::MG_ConfigLoader {
         features.TraceSkipAutodestroy = QueryEnvFlag("MOBILEGL_TRACE_SKIP_AUTODESTROY");
         features.DisableUboRing = QueryEnvFlag("MOBILEGL_DISABLE_UBO_RING");
         features.RelaxedSemantics = QueryEnvFlag("MOBILEGL_RELAXED_SEMANTICS");
+        features.SubgroupPrefixScanQuirk = QueryEnvQuirkOverride("MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN");
+        features.MagmaDisableBlendedDepthWriteQuirk =
+            QueryEnvQuirkOverride("MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE");
+        features.DisableRobustBufferAccess = QueryEnvFlag("MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS");
     }
 
     inline void InitBackendType() {

MobileGL/MG_Backend/BackendObject.h

@@ -230,6 +230,21 @@ namespace MobileGL {
             void (*SetSwapInterval)(Int interval);
         };
 
+        // Coarse GPU vendor identity for gating device-specific quirks. Detected from the
+        // Vulkan physical-device vendorID or the GLES GL_VENDOR/GL_RENDERER strings; stays
+        // Unknown when detection is inconclusive, in which case auto-gated quirks stay off.
+        enum class GpuVendorKind : Uint8 {
+            Unknown = 0,
+            Qualcomm,
+            Arm,
+            Nvidia,
+            Amd,
+            Intel,
+            ImgTec,
+            // Software rasterizers (llvmpipe/lavapipe, SwiftShader).
+            Software,
+        };
+
         struct DynamicBackendParameters {
             SizeT UniformBufferOffsetAlignment = 256;
             // GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT. 1.0 means the backend cannot filter anisotropically,
@@ -272,6 +287,9 @@ namespace MobileGL {
             Int MaxUniformBlockSize = 16384;
             Int MaxImageUnits = 8;
             Int MaxCombinedImageUniforms = 8;
+            Int MaxVertexImageUniforms = 0;
+            Int MaxGeometryImageUniforms = 0;
+            Int MaxFragmentImageUniforms = 8;
             Int MaxComputeImageUniforms = 8;
             Int MaxDrawBuffers = 8;
             Int MaxColorAttachments = 8;
@@ -288,6 +306,7 @@ namespace MobileGL {
             Uint32 SubgroupSupportedStages = 0;
             Uint32 SubgroupSupportedFeatures = 0;
             Bool SubgroupQuadOperationsInAllStages = false;
+            GpuVendorKind GpuVendor = GpuVendorKind::Unknown;
         };
 
         enum class WindowBackend {

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -826,7 +826,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                           E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object,
                                           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_direct_state_access,
+                                          E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access,
                                           E_GL_ARB_multi_draw_indirect, E_GL_ARB_indirect_parameters,
                                           E_GL_ARB_shader_draw_parameters, E_GL_ARB_gpu_shader5, E_GL_ARB_multi_bind,
                                           E_GL_ARB_shading_language_420pack, E_GL_ARB_vertex_attrib_binding,
@@ -947,6 +947,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return m_dynamicParameters;
     }
 
+    void BackendObject_DirectGLES::ApplyGLESCapabilitiesForTesting(
+        const MG_External::GLESCapabilities& capabilities) {
+        m_GLESCapabilities = capabilities;
+        UpdateDynamicBackendParameters();
+    }
+
     void BackendObject_DirectGLES::UpdateDynamicBackendParameters() {
         m_dynamicParameters.UniformBufferOffsetAlignment = m_GLESCapabilities.UniformBufferOffsetAlignment;
         m_dynamicParameters.MaxTextureMaxAnisotropy = m_GLESCapabilities.MaxTextureMaxAnisotropy;
@@ -1003,9 +1009,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
         m_dynamicParameters.MaxUniformBlockSize = m_GLESCapabilities.MaxUniformBlockSize;
         const Int maxSupportedTextureUnits =
             static_cast<Int>(MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS);
-        m_dynamicParameters.MaxImageUnits = std::min(m_GLESCapabilities.MaxImageUnits, maxSupportedTextureUnits);
-        m_dynamicParameters.MaxCombinedImageUniforms = m_GLESCapabilities.MaxCombinedImageUniforms;
-        m_dynamicParameters.MaxComputeImageUniforms = m_GLESCapabilities.MaxComputeImageUniforms;
+        m_dynamicParameters.MaxImageUnits =
+            std::max(std::min(m_GLESCapabilities.MaxImageUnits, maxSupportedTextureUnits), 0);
+        m_dynamicParameters.MaxCombinedImageUniforms = std::max(m_GLESCapabilities.MaxCombinedImageUniforms, 0);
+        const auto clampStageImageUniforms = [this](Int stageLimit) {
+            return std::min({std::max(stageLimit, 0), m_dynamicParameters.MaxImageUnits,
+                             m_dynamicParameters.MaxCombinedImageUniforms});
+        };
+        m_dynamicParameters.MaxVertexImageUniforms =
+            clampStageImageUniforms(m_GLESCapabilities.MaxVertexImageUniforms);
+        m_dynamicParameters.MaxGeometryImageUniforms =
+            clampStageImageUniforms(m_GLESCapabilities.MaxGeometryImageUniforms);
+        m_dynamicParameters.MaxFragmentImageUniforms =
+            clampStageImageUniforms(m_GLESCapabilities.MaxFragmentImageUniforms);
+        m_dynamicParameters.MaxComputeImageUniforms =
+            clampStageImageUniforms(m_GLESCapabilities.MaxComputeImageUniforms);
         m_dynamicParameters.MaxDrawBuffers = m_GLESCapabilities.MaxDrawBuffers;
         m_dynamicParameters.MaxColorAttachments = m_GLESCapabilities.MaxColorAttachments;
         m_dynamicParameters.MaxClipDistances = m_GLESCapabilities.MaxClipDistances;
@@ -1017,6 +1035,32 @@ namespace MobileGL::MG_Backend::DirectGLES {
         m_dynamicParameters.ViewportSubpixelBits = m_GLESCapabilities.ViewportSubpixelBits;
         m_dynamicParameters.SupportsWideLines =
             m_GLESCapabilities.AliasedLineWidthRangeMax > 1.0f || m_GLESCapabilities.SmoothLineWidthRangeMax > 1.0f;
+
+        const auto containsAny = [](const String& haystack, std::initializer_list<const char*> needles) {
+            return std::any_of(needles.begin(), needles.end(), [&](const char* needle) {
+                return haystack.find(needle) != String::npos;
+            });
+        };
+        const String vendorAndRenderer =
+            m_GLESCapabilities.GLESVendorString + " " + m_GLESCapabilities.GLESRendererString;
+        if (containsAny(vendorAndRenderer, {"llvmpipe", "SwiftShader", "softpipe"})) {
+            // Check software rasterizers first: ANGLE-on-llvmpipe reports both.
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Software;
+        } else if (containsAny(vendorAndRenderer, {"Qualcomm", "Adreno"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Qualcomm;
+        } else if (containsAny(vendorAndRenderer, {"Mali", "ARM"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Arm;
+        } else if (containsAny(vendorAndRenderer, {"NVIDIA"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Nvidia;
+        } else if (containsAny(vendorAndRenderer, {"AMD", "Radeon"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Amd;
+        } else if (containsAny(vendorAndRenderer, {"Intel"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Intel;
+        } else if (containsAny(vendorAndRenderer, {"Imagination", "PowerVR"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::ImgTec;
+        } else {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Unknown;
+        }
     }
 
     const MG_External::GLESFunctionsTable& BackendObject_DirectGLES::GetGLESFunctions() const {

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.h

@@ -41,6 +41,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         const MG_External::GLESFunctionsTable& GetGLESFunctions() const;
         const MG_External::EGLFunctionsTable& GetEGLFunctions() const;
+        void ApplyGLESCapabilitiesForTesting(const MG_External::GLESCapabilities& capabilities);
 
     private:
         void UpdateDynamicBackendParameters();

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -504,7 +504,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                           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_direct_state_access,
+                                          E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access,
                                           E_GL_ARB_shader_draw_parameters, E_GL_ARB_gpu_shader_int64, E_GL_KHR_debug,
                                           E_GL_ARB_gpu_shader5, E_GL_ARB_multi_bind, E_GL_ARB_shading_language_420pack,
                                           E_GL_ARB_vertex_attrib_binding, E_GL_ARB_shader_image_size};
@@ -746,9 +746,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_dynamicParameters.MaxTextureBufferSize = m_vulkanCaps.MaxTextureBufferSize;
         m_dynamicParameters.MaxUniformBufferBindings = m_vulkanCaps.MaxUniformBufferBindings;
         m_dynamicParameters.MaxUniformBlockSize = m_vulkanCaps.MaxUniformBlockSize;
-        m_dynamicParameters.MaxImageUnits = std::min(m_vulkanCaps.MaxImageUnits, maxSupportedTextureUnits);
-        m_dynamicParameters.MaxCombinedImageUniforms = m_vulkanCaps.MaxCombinedImageUniforms;
-        m_dynamicParameters.MaxComputeImageUniforms = m_vulkanCaps.MaxComputeImageUniforms;
+        m_dynamicParameters.MaxImageUnits =
+            std::max(std::min(m_vulkanCaps.MaxImageUnits, maxSupportedTextureUnits), 0);
+        m_dynamicParameters.MaxCombinedImageUniforms = std::max(m_vulkanCaps.MaxCombinedImageUniforms, 0);
+        const Int maxPerStageImageUniforms =
+            std::min(m_dynamicParameters.MaxImageUnits, m_dynamicParameters.MaxCombinedImageUniforms);
+        // Vulkan uses one descriptor limit for every stage, but non-compute stores/atomics are
+        // optional device features. VulkanRenderer enables each feature whenever the physical
+        // device reports it, so these are the exact limits the logical device can compile and run.
+        m_dynamicParameters.MaxVertexImageUniforms =
+            m_vulkanCaps.SupportsVertexPipelineStoresAndAtomics ? maxPerStageImageUniforms : 0;
+        m_dynamicParameters.MaxGeometryImageUniforms =
+            m_vulkanCaps.SupportsVertexPipelineStoresAndAtomics && m_vulkanCaps.SupportsGeometryShader
+                ? maxPerStageImageUniforms
+                : 0;
+        m_dynamicParameters.MaxFragmentImageUniforms =
+            m_vulkanCaps.SupportsFragmentStoresAndAtomics ? maxPerStageImageUniforms : 0;
+        m_dynamicParameters.MaxComputeImageUniforms =
+            std::min(std::max(m_vulkanCaps.MaxComputeImageUniforms, 0), maxPerStageImageUniforms);
         const Int maxSupportedDrawBuffers =
             static_cast<Int>(MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS);
         m_dynamicParameters.MaxDrawBuffers = std::min(m_vulkanCaps.MaxDrawBuffers, maxSupportedDrawBuffers);
@@ -779,5 +794,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     m_vulkanCaps.MaxShaderStorageBlockSize,
                     m_dynamicParameters.MaxShaderStorageBlockSize);
         }
+        switch (m_vulkanCaps.VendorId) {
+        case 0x5143u: // VK_VENDOR_ID: Qualcomm
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Qualcomm;
+            break;
+        case 0x13B5u: // ARM
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Arm;
+            break;
+        case 0x10DEu: // NVIDIA
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Nvidia;
+            break;
+        case 0x1002u: // AMD
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Amd;
+            break;
+        case 0x8086u: // Intel
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Intel;
+            break;
+        case 0x1010u: // Imagination
+            m_dynamicParameters.GpuVendor = GpuVendorKind::ImgTec;
+            break;
+        case 0x10005u: // Mesa software (lavapipe)
+        case 0x1AE0u:  // Google (SwiftShader)
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Software;
+            break;
+        default:
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Unknown;
+            break;
+        }
     }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -8,6 +8,7 @@
 
 #include "PipelineFactory.h"
 
+
 namespace MobileGL::MG_Backend::DirectVulkan {
     static const char* PrimitiveTopologyToString(VkPrimitiveTopology topology) {
         switch (topology) {
@@ -108,6 +109,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                   "vkCreatePipelineCache");
     }
 
+    void PipelineFactory::SetSuppressBlendedDepthWrite(Bool enabled) {
+        s_suppressBlendedDepthWrite = enabled;
+    }
+
     PipelineFactory::~PipelineFactory() {
         DestroyAll();
         if (m_pipelineCache != VK_NULL_HANDLE) {
@@ -258,6 +263,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (Uint32 i = 0; i < payload.colorAttachmentCount; ++i) {
             colorAttachments[i] = payload.colorBlendAttachments[i];
         }
+        // Suppress depth writes on blended pipelines when the active driver cannot keep
+        // vertex positions invariant across the pipelines of a multi-pass depth-equality
+        // chain (see SetSuppressBlendedDepthWrite). Blended draws that write depth are rare
+        // and the equality-dependent prepass pattern is exactly the case that breaks.
+        if (s_suppressBlendedDepthWrite && depthStencil.depthWriteEnable == VK_TRUE) {
+            for (Uint32 i = 0; i < payload.colorAttachmentCount; ++i) {
+                if (colorAttachments[i].blendEnable == VK_TRUE) {
+                    depthStencil.depthWriteEnable = VK_FALSE;
+                    break;
+                }
+            }
+        }
         VkPipelineColorBlendStateCreateInfo blend{VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO};
         blend.logicOpEnable = payload.logicOpEnable ? VK_TRUE : VK_FALSE;
         blend.logicOp = payload.logicOp;

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -62,6 +62,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkPipeline GetOrCreatePipeline(const PipelineCreatePayload& payload);
         void DestroyAll();
 
+        // Driver quirk: suppress depth writes on blended pipelines. Multi-pass depth-equality
+        // rendering (a blended prepass writes depth that later passes re-test with an
+        // equality-inclusive compare on the re-rasterized geometry) requires cross-pipeline
+        // position invariance that some mobile compilers do not provide, even with the
+        // SPIR-V Invariant decoration; whole primitives then drop out of the later passes.
+        // Set at renderer initialization based on the active driver.
+        static void SetSuppressBlendedDepthWrite(Bool enabled);
+
     private:
         VkPipeline CreatePipeline(const PipelineCreatePayload& payload) const;
 
@@ -70,5 +78,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkPipelineCache m_pipelineCache = VK_NULL_HANDLE;
         UnorderedMap<HashType, VkPipeline> m_cache;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline Bool s_suppressBlendedDepthWrite = false;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -1197,6 +1197,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return TextureTarget::Unknown;
             }
         }
+
+        Bool IsFloatStorageImageUniformType(GLenum uniformType) {
+            switch (uniformType) {
+            case GL_IMAGE_1D:
+            case GL_IMAGE_2D:
+            case GL_IMAGE_3D:
+            case GL_IMAGE_2D_RECT:
+            case GL_IMAGE_CUBE:
+            case GL_IMAGE_BUFFER:
+            case GL_IMAGE_1D_ARRAY:
+            case GL_IMAGE_2D_ARRAY:
+            case GL_IMAGE_CUBE_MAP_ARRAY:
+            case GL_IMAGE_2D_MULTISAMPLE:
+            case GL_IMAGE_2D_MULTISAMPLE_ARRAY:
+                return true;
+            default:
+                return false;
+            }
+        }
     } // namespace
 
     VkShaderStageFlagBits ProgramFactory::ToVkStage(ShaderStage stage) {
@@ -1218,6 +1237,105 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    VkFormat ProgramFactory::ConvertSpirvImageFormatToVkFormat(SpvImageFormat format) {
+        switch (format) {
+        case SpvImageFormatUnknown: return VK_FORMAT_UNDEFINED;
+        case SpvImageFormatRgba32f: return VK_FORMAT_R32G32B32A32_SFLOAT;
+        case SpvImageFormatRgba16f: return VK_FORMAT_R16G16B16A16_SFLOAT;
+        case SpvImageFormatR32f: return VK_FORMAT_R32_SFLOAT;
+        case SpvImageFormatRgba8: return VK_FORMAT_R8G8B8A8_UNORM;
+        case SpvImageFormatRgba8Snorm: return VK_FORMAT_R8G8B8A8_SNORM;
+        case SpvImageFormatRg32f: return VK_FORMAT_R32G32_SFLOAT;
+        case SpvImageFormatRg16f: return VK_FORMAT_R16G16_SFLOAT;
+        case SpvImageFormatR11fG11fB10f: return VK_FORMAT_B10G11R11_UFLOAT_PACK32;
+        case SpvImageFormatR16f: return VK_FORMAT_R16_SFLOAT;
+        case SpvImageFormatRgba16: return VK_FORMAT_R16G16B16A16_UNORM;
+        case SpvImageFormatRgb10A2: return VK_FORMAT_A2R10G10B10_UNORM_PACK32;
+        case SpvImageFormatRg16: return VK_FORMAT_R16G16_UNORM;
+        case SpvImageFormatRg8: return VK_FORMAT_R8G8_UNORM;
+        case SpvImageFormatR16: return VK_FORMAT_R16_UNORM;
+        case SpvImageFormatR8: return VK_FORMAT_R8_UNORM;
+        case SpvImageFormatRgba16Snorm: return VK_FORMAT_R16G16B16A16_SNORM;
+        case SpvImageFormatRg16Snorm: return VK_FORMAT_R16G16_SNORM;
+        case SpvImageFormatRg8Snorm: return VK_FORMAT_R8G8_SNORM;
+        case SpvImageFormatR16Snorm: return VK_FORMAT_R16_SNORM;
+        case SpvImageFormatR8Snorm: return VK_FORMAT_R8_SNORM;
+        case SpvImageFormatRgba32i: return VK_FORMAT_R32G32B32A32_SINT;
+        case SpvImageFormatRgba16i: return VK_FORMAT_R16G16B16A16_SINT;
+        case SpvImageFormatRgba8i: return VK_FORMAT_R8G8B8A8_SINT;
+        case SpvImageFormatR32i: return VK_FORMAT_R32_SINT;
+        case SpvImageFormatRg32i: return VK_FORMAT_R32G32_SINT;
+        case SpvImageFormatRg16i: return VK_FORMAT_R16G16_SINT;
+        case SpvImageFormatRg8i: return VK_FORMAT_R8G8_SINT;
+        case SpvImageFormatR16i: return VK_FORMAT_R16_SINT;
+        case SpvImageFormatR8i: return VK_FORMAT_R8_SINT;
+        case SpvImageFormatRgba32ui: return VK_FORMAT_R32G32B32A32_UINT;
+        case SpvImageFormatRgba16ui: return VK_FORMAT_R16G16B16A16_UINT;
+        case SpvImageFormatRgba8ui: return VK_FORMAT_R8G8B8A8_UINT;
+        case SpvImageFormatR32ui: return VK_FORMAT_R32_UINT;
+        case SpvImageFormatRgb10a2ui: return VK_FORMAT_A2R10G10B10_UINT_PACK32;
+        case SpvImageFormatRg32ui: return VK_FORMAT_R32G32_UINT;
+        case SpvImageFormatRg16ui: return VK_FORMAT_R16G16_UINT;
+        case SpvImageFormatRg8ui: return VK_FORMAT_R8G8_UINT;
+        case SpvImageFormatR16ui: return VK_FORMAT_R16_UINT;
+        case SpvImageFormatR8ui: return VK_FORMAT_R8_UINT;
+        case SpvImageFormatR64ui: return VK_FORMAT_R64_UINT;
+        case SpvImageFormatR64i: return VK_FORMAT_R64_SINT;
+        case SpvImageFormatMax: return VK_FORMAT_UNDEFINED;
+        }
+        return VK_FORMAT_UNDEFINED;
+    }
+
+    SamplerNumericDomain ProgramFactory::UniformTypeToSamplerNumericDomain(GLenum glType) {
+        switch (glType) {
+        case GL_INT_SAMPLER_1D:
+        case GL_INT_SAMPLER_2D:
+        case GL_INT_SAMPLER_3D:
+        case GL_INT_SAMPLER_CUBE:
+        case GL_INT_SAMPLER_2D_RECT:
+        case GL_INT_SAMPLER_1D_ARRAY:
+        case GL_INT_SAMPLER_2D_ARRAY:
+        case GL_INT_SAMPLER_BUFFER:
+        case GL_INT_SAMPLER_2D_MULTISAMPLE:
+        case GL_INT_SAMPLER_2D_MULTISAMPLE_ARRAY:
+        case GL_INT_SAMPLER_CUBE_MAP_ARRAY:
+            return SamplerNumericDomain::SignedInteger;
+        case GL_UNSIGNED_INT_SAMPLER_1D:
+        case GL_UNSIGNED_INT_SAMPLER_2D:
+        case GL_UNSIGNED_INT_SAMPLER_3D:
+        case GL_UNSIGNED_INT_SAMPLER_CUBE:
+        case GL_UNSIGNED_INT_SAMPLER_2D_RECT:
+        case GL_UNSIGNED_INT_SAMPLER_1D_ARRAY:
+        case GL_UNSIGNED_INT_SAMPLER_2D_ARRAY:
+        case GL_UNSIGNED_INT_SAMPLER_BUFFER:
+        case GL_UNSIGNED_INT_SAMPLER_2D_MULTISAMPLE:
+        case GL_UNSIGNED_INT_SAMPLER_2D_MULTISAMPLE_ARRAY:
+        case GL_UNSIGNED_INT_SAMPLER_CUBE_MAP_ARRAY:
+            return SamplerNumericDomain::UnsignedInteger;
+        case GL_SAMPLER_1D:
+        case GL_SAMPLER_2D:
+        case GL_SAMPLER_3D:
+        case GL_SAMPLER_CUBE:
+        case GL_SAMPLER_2D_RECT:
+        case GL_SAMPLER_1D_ARRAY:
+        case GL_SAMPLER_2D_ARRAY:
+        case GL_SAMPLER_BUFFER:
+        case GL_SAMPLER_2D_MULTISAMPLE:
+        case GL_SAMPLER_2D_MULTISAMPLE_ARRAY:
+        case GL_SAMPLER_CUBE_MAP_ARRAY:
+        case GL_SAMPLER_1D_SHADOW:
+        case GL_SAMPLER_2D_SHADOW:
+        case GL_SAMPLER_CUBE_SHADOW:
+        case GL_SAMPLER_2D_RECT_SHADOW:
+        case GL_SAMPLER_1D_ARRAY_SHADOW:
+        case GL_SAMPLER_2D_ARRAY_SHADOW:
+        case GL_SAMPLER_CUBE_MAP_ARRAY_SHADOW:
+            return SamplerNumericDomain::Float;
+        default:
+            return SamplerNumericDomain::Unknown;
+        }
+    }
+
     ProgramFactory::HashType ProgramFactory::ComputeHash(const MG_State::GLState::ProgramObject& program,
                                                          CompileOptionFlags flags) const {
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
@@ -1465,6 +1583,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.samplerNameByBinding.assign(m_maxBindings, String());
         entry.samplerUniformLocationByBinding.assign(m_maxBindings, -1);
         entry.samplerTextureTargetByBinding.assign(m_maxBindings, TextureTarget::Texture2D);
+        entry.samplerNumericDomainByBinding.assign(m_maxBindings, SamplerNumericDomain::Unknown);
+        entry.storageImageFormatByBinding.assign(m_maxBindings, VK_FORMAT_UNDEFINED);
+        entry.storageImageUsesBindingFormatByBinding.assign(m_maxBindings, false);
         entry.storageBlockNameByBinding.assign(m_maxBindings, String());
         entry.storageBlockIndexByBinding.assign(m_maxBindings, -1);
         entry.globalUboBinding = -1;
@@ -1592,10 +1713,54 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     continue;
                 }
 
-                const TextureTarget target = UniformTypeToTextureTarget(program.GetUniformType(static_cast<Uint>(location)));
+                const GLenum uniformType = program.GetUniformType(static_cast<Uint>(location));
+
+                if (descriptorKind == DescriptorBindingKind::StorageImage) {
+                    const VkFormat reflectedFormat =
+                        ConvertSpirvImageFormatToVkFormat(sampler->image.image_format);
+                    VkFormat& existingFormat = entry.storageImageFormatByBinding[binding];
+                    MOBILEGL_ASSERT(existingFormat == VK_FORMAT_UNDEFINED ||
+                                        reflectedFormat == VK_FORMAT_UNDEFINED ||
+                                        existingFormat == reflectedFormat,
+                                    "ProgramFactory::ReflectLayout: storage image binding %u ('%s') has "
+                                    "conflicting reflected formats (%d vs %d)",
+                                    binding, uniformName.c_str(), static_cast<Int>(existingFormat),
+                                    static_cast<Int>(reflectedFormat));
+                    if (existingFormat == VK_FORMAT_UNDEFINED) {
+                        existingFormat = reflectedFormat;
+                    }
+                    if (m_unformattedFloatStorageImagesEnabled &&
+                        existingFormat == VK_FORMAT_UNDEFINED &&
+                        IsFloatStorageImageUniformType(uniformType)) {
+                        entry.storageImageUsesBindingFormatByBinding[binding] = true;
+                    } else if (reflectedFormat != VK_FORMAT_UNDEFINED) {
+                        // A typed declaration in any stage wins for the entire binding. This is
+                        // required when another stage reaches the same image through an atomic
+                        // path and therefore could not be made formatless.
+                        entry.storageImageUsesBindingFormatByBinding[binding] = false;
+                    }
+                }
+
+                const TextureTarget target = UniformTypeToTextureTarget(uniformType);
                 MOBILEGL_ASSERT(target != TextureTarget::Unknown,
                                 "ProgramFactory::ReflectLayout: failed to resolve texture target for '%s'",
                                 uniformName.c_str());
+                if (descriptorKind == DescriptorBindingKind::CombinedImageSampler) {
+                    const SamplerNumericDomain numericDomain = UniformTypeToSamplerNumericDomain(uniformType);
+                    MOBILEGL_ASSERT(numericDomain != SamplerNumericDomain::Unknown,
+                                    "ProgramFactory::ReflectLayout: failed to resolve sampler numeric domain "
+                                    "for '%s' (uniformType=0x%x)",
+                                    uniformName.c_str(), uniformType);
+                    MOBILEGL_ASSERT(entry.samplerNumericDomainByBinding[binding] ==
+                                            SamplerNumericDomain::Unknown ||
+                                        entry.samplerNumericDomainByBinding[binding] == numericDomain,
+                                    "ProgramFactory::ReflectLayout: sampler binding %u ('%s') has conflicting "
+                                    "numeric domains (%d vs %d)",
+                                    binding, uniformName.c_str(),
+                                    static_cast<Int>(entry.samplerNumericDomainByBinding[binding]),
+                                    static_cast<Int>(numericDomain));
+                    entry.samplerNumericDomainByBinding[binding] = numericDomain;
+                }
                 MOBILEGL_ASSERT(entry.samplerUniformLocationByBinding[binding] < 0 || location < 0 ||
                                     entry.samplerUniformLocationByBinding[binding] == location,
                                 "ProgramFactory::ReflectLayout: texture binding %u maps to conflicting uniform locations (%d vs %d)",
@@ -1643,6 +1808,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
             } else if (kind == DescriptorBindingKind::StorageImage) {
                 layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
+                entry.hasStorageImages = true;
             } else {
                 layoutBinding.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
             }
@@ -1697,6 +1863,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 moduleSpirvs[i] = spv;
             }
 
+            // GL apps depend on cross-program position invariance for multi-pass equality
+            // depth tests (MC 26.3's OIT re-draws the cloud geometry with GEQUAL against the
+            // depth its own first pass wrote); decorate Position outputs Invariant so
+            // per-pipeline compilers cannot vary the position math between passes.
+            {
+                Vector<Uint> invariantSpirv;
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::DecoratePositionInvariantForVulkan(
+                        moduleSpirvs[i], invariantSpirv)) {
+                    moduleSpirvs[i] = std::move(invariantSpirv);
+                } else {
+                    MGLOG_W("ProgramFactory: position-invariant decoration failed for program %u; "
+                            "keeping the original module",
+                            program.GetExternalIndex());
+                }
+            }
+
             // glslang's relaxed-Vulkan mode aliases GL's zero-based gl_InstanceID to Vulkan's
             // gl_InstanceIndex, which wrongly includes the draw's baseInstance. Rebase vertex-stage
             // loads to (InstanceIndex - BaseInstance) so shaders observe GL semantics. Reflection
@@ -1721,6 +1903,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     }
                 }
             }
+
+            // When Vulkan can legally access storage images without a statically declared
+            // format, let GL's glBindImageTexture format select the runtime image view. This
+            // provides desktop-driver-compatible behavior for packs such as iterationRP, whose
+            // float image qualifier can disagree with the bound render-target format. Integer
+            // storage images remain formatted so r32ui/r32i bit-reinterpretation paths keep the
+            // exact descriptor format required by their shader operations.
+            if (m_unformattedFloatStorageImagesEnabled) {
+                Vector<Uint> unformattedSpirv;
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
+                        moduleSpirvs[i], unformattedSpirv)) {
+                    moduleSpirvs[i] = std::move(unformattedSpirv);
+                } else {
+                    MGLOG_E("ProgramFactory: failed to make float storage images unformatted for program %u",
+                            program.GetExternalIndex());
+                }
+            }
         }
 
         const Bool remapOk = RemapDescriptorBindingsForVulkan(moduleSpirvs, m_maxBindings, moduleSpirvs);

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -14,8 +14,16 @@
 #include "MG_State/GLState/TextureState/TextureEnum.h"
 
 #include <Includes.h>
+#include <spirv_reflect.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
+    enum class SamplerNumericDomain : Uint8 {
+        Unknown = 0,
+        Float,
+        SignedInteger,
+        UnsignedInteger,
+    };
+
     class ProgramFactory {
     public:
         enum class DescriptorBindingKind : Uint8 {
@@ -54,8 +62,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<String> samplerNameByBinding;
             Vector<Int> samplerUniformLocationByBinding;
             Vector<TextureTarget> samplerTextureTargetByBinding;
+            Vector<SamplerNumericDomain> samplerNumericDomainByBinding;
+            Vector<VkFormat> storageImageFormatByBinding;
+            Vector<Bool> storageImageUsesBindingFormatByBinding;
             Vector<String> storageBlockNameByBinding;
             Vector<Int> storageBlockIndexByBinding;
+            // Set once during ReflectLayout so the per-draw path can skip the whole
+            // storage-image preparation for the overwhelming majority of programs.
+            Bool hasStorageImages = false;
             Int globalUboBinding = -1;
             Uint32 activeVertexInputLocationMask = 0;
             Array<GLenum, kMaxVertexInputLocations> vertexInputTypes{};
@@ -82,8 +96,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 samplerNameByBinding = std::move(other.samplerNameByBinding);
                 samplerUniformLocationByBinding = std::move(other.samplerUniformLocationByBinding);
                 samplerTextureTargetByBinding = std::move(other.samplerTextureTargetByBinding);
+                samplerNumericDomainByBinding = std::move(other.samplerNumericDomainByBinding);
+                storageImageFormatByBinding = std::move(other.storageImageFormatByBinding);
+                storageImageUsesBindingFormatByBinding =
+                    std::move(other.storageImageUsesBindingFormatByBinding);
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
+                hasStorageImages = other.hasStorageImages;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -95,6 +114,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
+                other.hasStorageImages = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -118,8 +138,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 samplerNameByBinding = std::move(other.samplerNameByBinding);
                 samplerUniformLocationByBinding = std::move(other.samplerUniformLocationByBinding);
                 samplerTextureTargetByBinding = std::move(other.samplerTextureTargetByBinding);
+                samplerNumericDomainByBinding = std::move(other.samplerNumericDomainByBinding);
+                storageImageFormatByBinding = std::move(other.storageImageFormatByBinding);
+                storageImageUsesBindingFormatByBinding =
+                    std::move(other.storageImageUsesBindingFormatByBinding);
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
+                hasStorageImages = other.hasStorageImages;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -131,6 +156,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
+                other.hasStorageImages = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -167,9 +193,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         };
 
         explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
-                                Bool shaderDrawParametersEnabled = false)
+                                Bool shaderDrawParametersEnabled = false,
+                                Bool unformattedFloatStorageImagesEnabled = false)
             : m_device(device), m_maxBindings(maxBindings), m_config(config),
-              m_shaderDrawParametersEnabled(shaderDrawParametersEnabled) {
+              m_shaderDrawParametersEnabled(shaderDrawParametersEnabled),
+              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled) {
             VkProgramObject::s_device = device;
         }
         ~ProgramFactory() = default;
@@ -180,6 +208,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
 
         static VkShaderStageFlagBits ToVkStage(ShaderStage stage);
+        static VkFormat ConvertSpirvImageFormatToVkFormat(SpvImageFormat format);
+        static SamplerNumericDomain UniformTypeToSamplerNumericDomain(GLenum glType);
 
     private:
         struct ProgramLookupCache {
@@ -206,6 +236,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True when the device enabled shaderDrawParameters; gates the InstanceIndex rebase pass
         // (which needs the DrawParameters capability / gl_BaseInstance builtin).
         Bool m_shaderDrawParametersEnabled = false;
+        // True only when the logical device enabled both
+        // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
+        Bool m_unformattedFloatStorageImagesEnabled = false;
         mutable ProgramLookupCache m_lastLookup;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -13,6 +13,7 @@
 #include "MG_State/GLState/ProgramState/ProgramObject.h"
 #include "MG_State/GLState/TextureState/TextureObject2D.h"
 #include "MG_State/GLState/TextureState/TextureObjectBuffer.h"
+#include "MG_Util/Converters/GLToMG/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToStr/FramebufferEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
 #include "MG_Util/Metrics/TextureMetrics.h"
@@ -78,6 +79,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return uniformUnit >= 0 ? uniformUnit : 0;
     }
 
+    VkFormat UniformManager::ResolveStorageImageViewFormat(VkFormat reflectedFormat, GLenum bindingFormat,
+                                                           VkFormat resourceFormat, Bool useBindingFormat) {
+        if (useBindingFormat) {
+            const TextureInternalFormat bindingInternalFormat =
+                MG_Util::ConvertGLEnumToTextureInternalFormat(bindingFormat);
+            return MG_Util::ConvertTextureInternalFormatToVkEnum(bindingInternalFormat);
+        }
+        return reflectedFormat != VK_FORMAT_UNDEFINED ? reflectedFormat : resourceFormat;
+    }
+
     Bool UniformManager::Initialize(VkDevice device, VkBufferManager* bufferManager,
                                              ProgramFactory* programFactory,
                                              VkDeviceSize minUniformBufferOffsetAlignment, Uint32 frameCount,
@@ -276,6 +287,55 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                             "ResolveSamplerDescriptor: invalid sampled image layout=%d for textureId=%d, binding=%u",
                             static_cast<Int>(resource->layout), texture->GetExternalIndex(), binding);
         }
+
+        MOBILEGL_ASSERT(binding < programObj.samplerNumericDomainByBinding.size(),
+                        "ResolveSamplerDescriptor: sampler numeric-domain binding %u out of range", binding);
+        const SamplerNumericDomain numericDomain = programObj.samplerNumericDomainByBinding[binding];
+        // Vulkan forbids linear filtering and anisotropy for integer sampled-image formats.
+        // Some desktop GL shader packs deliberately bit-read a mutable float texture through a
+        // usampler and still leave the texture's ordinary linear parameters in place; texelFetch
+        // ignores filtering, so a nearest VkSampler preserves the operation while keeping the
+        // descriptor valid.
+        const Bool forceNearestFiltering = numericDomain == SamplerNumericDomain::SignedInteger ||
+                                           numericDomain == SamplerNumericDomain::UnsignedInteger;
+        SamplerResolveMemo* viewFormatMemo =
+            binding < m_samplerResolveMemo.size() ? &m_samplerResolveMemo[binding] : nullptr;
+        VkFormat sampledViewFormat;
+        if (viewFormatMemo != nullptr && viewFormatMemo->viewFormatValid &&
+            viewFormatMemo->viewFormatSource == resource->format &&
+            viewFormatMemo->viewFormatDomain == numericDomain) {
+            sampledViewFormat = viewFormatMemo->viewFormat;
+        } else {
+            sampledViewFormat =
+                VkTextureManager::ResolveSampledImageViewFormat(resource->format, numericDomain);
+            if (viewFormatMemo != nullptr) {
+                viewFormatMemo->viewFormatSource = resource->format;
+                viewFormatMemo->viewFormatDomain = numericDomain;
+                viewFormatMemo->viewFormat = sampledViewFormat;
+                viewFormatMemo->viewFormatValid = true;
+            }
+        }
+        if (sampledViewFormat == VK_FORMAT_UNDEFINED) {
+            MGLOG_E("ResolveSamplerDescriptor: no compatible sampled view for binding=%u ('%s') "
+                    "textureId=%d imageFormat=%d numericDomain=%d",
+                    binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
+                    static_cast<Int>(resource->format), static_cast<Int>(numericDomain));
+            return false;
+        }
+        // No reinterpretation requested: bind the depth-or-color aspect view the sync above
+        // already produced instead of re-entering GetOrCreateSampledImageView's sync path.
+        const VkImageView sampledImageView =
+            sampledViewFormat == resource->format
+                ? resource->sampledView
+                : m_textureManager->GetOrCreateSampledImageView(*texture, sampledViewFormat);
+        if (sampledImageView == VK_NULL_HANDLE) {
+            MGLOG_E("ResolveSamplerDescriptor: failed to resolve sampled view for binding=%u ('%s') "
+                    "textureId=%d imageFormat=%d viewFormat=%d numericDomain=%d",
+                    binding, programObj.samplerNameByBinding[binding].c_str(), texture->GetExternalIndex(),
+                    static_cast<Int>(resource->format), static_cast<Int>(sampledViewFormat),
+                    static_cast<Int>(numericDomain));
+            return false;
+        }
         // Skip GetOrCreateSampler's per-draw key hash + map lookup when this binding's
         // sampler object and texture (both by lifetime id + version) are unchanged from the
         // last draw that resolved it: the resulting sampler key, and therefore the VkSampler
@@ -291,23 +351,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Uint64 textureLifetimeId = texture->GetLifetimeId();
             const Uint16 textureParamsVersion = texture->GetTextureParamsVersion();
             if (memo.valid && memo.samplerLifetimeId == samplerLifetimeId && memo.samplerVersion == samplerVersion &&
-                memo.textureLifetimeId == textureLifetimeId && memo.textureParamsVersion == textureParamsVersion) {
+                memo.textureLifetimeId == textureLifetimeId && memo.textureParamsVersion == textureParamsVersion &&
+                memo.forceNearestFiltering == forceNearestFiltering) {
                 resolvedSampler = memo.sampler;
             } else {
-                resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture);
+                resolvedSampler =
+                    m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
                 memo.samplerLifetimeId = samplerLifetimeId;
                 memo.samplerVersion = samplerVersion;
                 memo.textureLifetimeId = textureLifetimeId;
                 memo.textureParamsVersion = textureParamsVersion;
+                memo.forceNearestFiltering = forceNearestFiltering;
                 memo.sampler = resolvedSampler;
                 memo.valid = true;
             }
         } else {
-            resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture);
+            resolvedSampler =
+                m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
         }
         outImageInfo = {
             .sampler = resolvedSampler,
-            .imageView = resource->sampledView != VK_NULL_HANDLE ? resource->sampledView : resource->fullView,
+            .imageView = sampledImageView,
             .imageLayout = resource->layout,
         };
         return outImageInfo.sampler != VK_NULL_HANDLE;
@@ -572,9 +636,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         const Uint32 mipLevel = static_cast<Uint32>(std::max<GLint>(0, imageBinding.Level));
-        VkImageView view = m_textureManager->GetOrCreateViewAtMipLevel(*imageBinding.Texture, mipLevel);
+        MOBILEGL_ASSERT(binding < programObj.storageImageFormatByBinding.size(),
+                        "ResolveStorageImageDescriptor: storage image format binding %u out of range", binding);
+        MOBILEGL_ASSERT(binding < programObj.storageImageUsesBindingFormatByBinding.size(),
+                        "ResolveStorageImageDescriptor: storage image format policy binding %u out of range",
+                        binding);
+        const VkFormat reflectedFormat = programObj.storageImageFormatByBinding[binding];
+        const Bool useBindingFormat = programObj.storageImageUsesBindingFormatByBinding[binding];
+        const VkFormat viewFormat = ResolveStorageImageViewFormat(
+            reflectedFormat, imageBinding.Format, resource->format, useBindingFormat);
+        if (viewFormat == VK_FORMAT_UNDEFINED) {
+            MGLOG_E("ResolveStorageImageDescriptor: unsupported glBindImageTexture format=0x%x "
+                    "for binding=%u imageUnit=%d textureId=%d bindingPolicy=%s",
+                    imageBinding.Format, binding, imageUnit, imageBinding.Texture->GetExternalIndex(),
+                    useBindingFormat ? "true" : "false");
+            return false;
+        }
+        const VkImageView view = m_textureManager->GetOrCreateStorageImageView(
+            *imageBinding.Texture, mipLevel, viewFormat, imageBinding.Layered != GL_FALSE, imageBinding.Layer);
         if (view == VK_NULL_HANDLE) {
-            view = resource->fullView;
+            MGLOG_E("ResolveStorageImageDescriptor: failed to resolve storage view textureId=%d mip=%u "
+                    "bindingFormat=0x%x imageFormat=%d reflectedFormat=%d selectedFormat=%d bindingPolicy=%s",
+                    imageBinding.Texture->GetExternalIndex(), mipLevel, imageBinding.Format,
+                    static_cast<Int>(resource->format), static_cast<Int>(reflectedFormat),
+                    static_cast<Int>(viewFormat),
+                    useBindingFormat ? "true" : "false");
+            return false;
         }
         outImageInfo.sampler = VK_NULL_HANDLE;
         outImageInfo.imageView = view;
@@ -632,6 +719,50 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
+    Bool UniformManager::CollectStorageImageTextures(
+        const MG_State::GLState::ProgramObject& program,
+        const ProgramFactory::VkProgramObject& programObj,
+        Vector<MG_State::GLState::ITextureObject*>& outTextures) const {
+        outTextures.clear();
+        MOBILEGL_ASSERT(MG_State::pGLContext != nullptr,
+                        "CollectStorageImageTextures: GL context is null");
+
+        const Uint32 bindingCount =
+            std::min<Uint32>(m_maxBindings, static_cast<Uint32>(programObj.bindingKinds.size()));
+        for (Uint32 binding = 0; binding < bindingCount; ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::StorageImage) {
+                continue;
+            }
+            if (binding >= programObj.samplerUniformLocationByBinding.size()) {
+                MGLOG_E("CollectStorageImageTextures: binding %u has no uniform-location mapping", binding);
+                return false;
+            }
+
+            const Int location = programObj.samplerUniformLocationByBinding[binding];
+            if (location < 0) {
+                MGLOG_E("CollectStorageImageTextures: binding %u has no image uniform location", binding);
+                return false;
+            }
+            const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
+            if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+                MGLOG_E("CollectStorageImageTextures: image unit %d is invalid for binding %u",
+                        imageUnit, binding);
+                return false;
+            }
+
+            auto* texture = MG_State::pGLContext->GetImageTextureBinding(imageUnit).Texture.get();
+            if (texture == nullptr) {
+                MGLOG_E("CollectStorageImageTextures: image unit %d is unbound for binding %u",
+                        imageUnit, binding);
+                return false;
+            }
+            if (std::find(outTextures.begin(), outTextures.end(), texture) == outTextures.end()) {
+                outTextures.push_back(texture);
+            }
+        }
+        return true;
+    }
+
     Bool UniformManager::ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                                      const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                      UboBindResult& out) const {

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -42,6 +42,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
                                     const ProgramFactory::VkProgramObject& programObj,
                                     Vector<MG_State::GLState::ITextureObject*>& outTextures);
+        Bool CollectStorageImageTextures(const MG_State::GLState::ProgramObject& program,
+                                         const ProgramFactory::VkProgramObject& programObj,
+                                         Vector<MG_State::GLState::ITextureObject*>& outTextures) const;
         Bool BindProgramUniformBuffers(VkCommandBuffer commandBuffer,
                                        const MG_State::GLState::ProgramObject& program,
                                        const ProgramFactory::VkProgramObject& programObj,
@@ -49,6 +52,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
                                        const SamplerBindingOverride* samplerBindingOverride = nullptr);
 
+        // Pure format-policy helper kept public for host regression tests. Formatted storage
+        // images use their shader qualifier; transformed float images use glBindImageTexture's
+        // format and never silently fall back to the backing image format.
+        static VkFormat ResolveStorageImageViewFormat(VkFormat reflectedFormat, GLenum bindingFormat,
+                                                      VkFormat resourceFormat, Bool useBindingFormat);
+
     private:
         struct DescriptorPoolBucket {
             VkDescriptorPool handle = VK_NULL_HANDLE;
@@ -165,9 +174,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkSampler sampler = VK_NULL_HANDLE;
             Uint16 samplerVersion = 0;
             Uint16 textureParamsVersion = 0;
+            Bool forceNearestFiltering = false;
             Bool valid = false;
+            // ResolveSampledImageViewFormat is pure in (image format, numeric domain), but a
+            // domain mismatch walks a ~184-entry format table. Memo the resolution per binding
+            // so a reinterpreted sampler pays that scan once, not once per draw.
+            VkFormat viewFormatSource = VK_FORMAT_UNDEFINED;
+            SamplerNumericDomain viewFormatDomain = SamplerNumericDomain::Unknown;
+            VkFormat viewFormat = VK_FORMAT_UNDEFINED;
+            Bool viewFormatValid = false;
         };
         mutable Vector<SamplerResolveMemo> m_samplerResolveMemo;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan
-

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -66,6 +66,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Vector<SizeT> bindingBaseOffsets;
         Vector<Uint32> bindingAttributeLocations;
         Vector<Bool> bindingUsesClientMemory;
+        Vector<VertexStreamConversion> bindingConversions;
         Uint32 unsupportedAttribMask = 0;
 
         for (Uint32 location = 0; location < MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS; ++location) {
@@ -74,8 +75,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
 
-            const auto vkFormat = ToVkVertexFormat(attr.Type, attr.Size, attr.Normalized, attr.IsInteger, attr.IsBgra);
-            if (vkFormat == VK_FORMAT_UNDEFINED) {
+            const VkFormat sourceVkFormat =
+                ToVkVertexFormat(attr.Type, attr.Size, attr.Normalized, attr.IsInteger, attr.IsBgra);
+            if (sourceVkFormat == VK_FORMAT_UNDEFINED) {
                 MGLOG_E("Unsupported vertex attribute layout (location=%u, type=%s, size=%d): the array is "
                         "enabled but cannot be mapped to a VkFormat",
                         location, MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
@@ -83,6 +85,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
 
+            VkFormat vkFormat = sourceVkFormat;
+            VertexStreamConversion conversion = VertexStreamConversion::None;
+            if (!SupportsVertexBufferFormat(vkFormat)) {
+                if (IsScaledIntegerVertexFormat(vkFormat)) {
+                    const VkFormat fallbackFormat = ToFloat32VertexFormat(attr.Size);
+                    if (fallbackFormat != VK_FORMAT_UNDEFINED && SupportsVertexBufferFormat(fallbackFormat)) {
+                        vkFormat = fallbackFormat;
+                        conversion = VertexStreamConversion::ScaledIntegerToFloat32;
+                        MGLOG_W("Vertex attribute location=%u format=%d lacks "
+                                "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT; using float32 stream format=%d "
+                                "(type=%s size=%d normalized=%s integer=%s)",
+                                location, static_cast<Int>(sourceVkFormat), static_cast<Int>(vkFormat),
+                                MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size,
+                                attr.Normalized ? "true" : "false", attr.IsInteger ? "true" : "false");
+                    }
+                }
+
+                if (conversion == VertexStreamConversion::None) {
+                    MGLOG_E("Unsupported Vulkan vertex format (location=%u, format=%d, type=%s, size=%d): "
+                            "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT is unavailable and no semantic fallback exists",
+                            location, static_cast<Int>(sourceVkFormat),
+                            MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
+                    unsupportedAttribMask |= (1u << location);
+                    continue;
+                }
+            }
+
             const SizeT attribByteSize = GetAttributeByteSize(attr.Type, attr.Size, attr.IsBgra);
             if (attribByteSize == 0) {
                 MGLOG_E("Vertex attribute with unknown component size (location=%u, type=%s): the array is "
@@ -92,8 +121,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
 
-            const Uint32 stride =
+            const Uint32 sourceStride =
                 attr.Stride > 0 ? static_cast<Uint32>(attr.Stride) : static_cast<Uint32>(attribByteSize);
+            const Bool packedAttribute = attr.Type == DataType::Int2101010Rev ||
+                                         attr.Type == DataType::Uint2101010Rev;
+            const SizeT requiredAlignment = packedAttribute ? attribByteSize : GetComponentSize(attr.Type);
+            // For a client-memory array attr.Offset holds the raw client pointer, and the
+            // draw path re-uploads the data to a 16-aligned transient slice with attribute
+            // offset 0, so only the stride can violate Vulkan's fetch alignment there.
+            const Bool clientMemoryAttribute = attr.Buffer == nullptr;
+            if (conversion == VertexStreamConversion::None && requiredAlignment > 1 &&
+                ((sourceStride % requiredAlignment) != 0 ||
+                 (!clientMemoryAttribute && (attr.Offset % requiredAlignment) != 0))) {
+                // GL accepts arbitrary byte strides and offsets. Core Vulkan vertex fetches do not
+                // unless VK_EXT_legacy_vertex_attributes is available, so deinterleave this one
+                // attribute into a tightly packed transient stream without changing its format.
+                conversion = VertexStreamConversion::Repack;
+                MGLOG_W("Vertex attribute location=%u uses Vulkan-incompatible alignment "
+                        "(offset=%zu stride=%u required=%zu); using a tightly packed stream",
+                        location, attr.Offset, sourceStride, requiredAlignment);
+            }
+
+            Uint32 stride = sourceStride;
+            if (conversion == VertexStreamConversion::Repack) {
+                stride = static_cast<Uint32>(attribByteSize);
+            } else if (conversion == VertexStreamConversion::ScaledIntegerToFloat32) {
+                stride = static_cast<Uint32>(attr.Size * static_cast<Int>(sizeof(Float)));
+            }
             const VkVertexInputRate inputRate =
                 (attr.Divisor == 0) ? VK_VERTEX_INPUT_RATE_VERTEX : VK_VERTEX_INPUT_RATE_INSTANCE;
 
@@ -103,6 +157,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             bindingBaseOffsets.push_back(attr.Buffer ? attr.Offset : 0);
             bindingAttributeLocations.push_back(location);
             bindingUsesClientMemory.push_back(attr.Buffer == nullptr);
+            bindingConversions.push_back(conversion);
             builder.AddBinding(binding, stride, inputRate);
             builder.AddAttribute(location, binding, vkFormat, 0);
         }
@@ -117,6 +172,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.bindingBaseOffsets = std::move(bindingBaseOffsets);
         entry.bindingAttributeLocations = std::move(bindingAttributeLocations);
         entry.bindingUsesClientMemory = std::move(bindingUsesClientMemory);
+        entry.bindingConversions = std::move(bindingConversions);
         entry.unsupportedAttribMask = unsupportedAttribMask;
         entry.state = state;
         entry.state.pVertexBindingDescriptions = entry.bindings.empty() ? nullptr : entry.bindings.data();
@@ -282,4 +338,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const SizeT componentSize = GetComponentSize(type);
         return componentSize == 0 ? 0 : componentSize * static_cast<SizeT>(size);
     }
+
+    Bool VertexInputStateFactory::IsScaledIntegerVertexFormat(VkFormat format) {
+        switch (format) {
+        case VK_FORMAT_R8_USCALED:
+        case VK_FORMAT_R8_SSCALED:
+        case VK_FORMAT_R8G8_USCALED:
+        case VK_FORMAT_R8G8_SSCALED:
+        case VK_FORMAT_R8G8B8_USCALED:
+        case VK_FORMAT_R8G8B8_SSCALED:
+        case VK_FORMAT_R8G8B8A8_USCALED:
+        case VK_FORMAT_R8G8B8A8_SSCALED:
+        case VK_FORMAT_R16_USCALED:
+        case VK_FORMAT_R16_SSCALED:
+        case VK_FORMAT_R16G16_USCALED:
+        case VK_FORMAT_R16G16_SSCALED:
+        case VK_FORMAT_R16G16B16_USCALED:
+        case VK_FORMAT_R16G16B16_SSCALED:
+        case VK_FORMAT_R16G16B16A16_USCALED:
+        case VK_FORMAT_R16G16B16A16_SSCALED:
+            return true;
+        default:
+            return false;
+        }
+    }
+
+    VkFormat VertexInputStateFactory::ToFloat32VertexFormat(Int componentCount) {
+        switch (componentCount) {
+        case 1: return VK_FORMAT_R32_SFLOAT;
+        case 2: return VK_FORMAT_R32G32_SFLOAT;
+        case 3: return VK_FORMAT_R32G32B32_SFLOAT;
+        case 4: return VK_FORMAT_R32G32B32A32_SFLOAT;
+        default: return VK_FORMAT_UNDEFINED;
+        }
+    }
+
+    Bool VertexInputStateFactory::SupportsVertexBufferFormat(VkFormat format) const {
+        if (m_physicalDevice == VK_NULL_HANDLE || format == VK_FORMAT_UNDEFINED) {
+            return false;
+        }
+        VkFormatProperties properties{};
+        vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &properties);
+        return (properties.bufferFeatures & VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT) != 0;
+    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -19,6 +19,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     public:
         using HashType = Uint64;
 
+        enum class VertexStreamConversion : Uint8 {
+            None = 0,
+            Repack,
+            ScaledIntegerToFloat32,
+        };
+
         struct BackendVertexInputState {
             HashType hash = 0;
             Vector<VkVertexInputBindingDescription> bindings;
@@ -27,6 +33,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<SizeT> bindingBaseOffsets;
             Vector<Uint32> bindingAttributeLocations;
             Vector<Bool> bindingUsesClientMemory;
+            Vector<VertexStreamConversion> bindingConversions;
             // Locations whose array is ENABLED but whose GL format has no VkFormat mapping. They are
             // absent from `attributes`, so without this mask the draw path cannot tell them apart from
             // a genuinely disabled array and would silently feed the shader the current attribute value.
@@ -36,8 +43,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             };
         };
 
-        explicit VertexInputStateFactory(const VulkanRendererConfig& config):
-            m_config(config) {}
+        VertexInputStateFactory(const VulkanRendererConfig& config, VkPhysicalDevice physicalDevice):
+            m_config(config), m_physicalDevice(physicalDevice) {}
         ~VertexInputStateFactory() = default;
         VertexInputStateFactory(const VertexInputStateFactory&) = delete;
 
@@ -56,8 +63,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     private:
         static VkFormat ToVkVertexFormat(DataType type, Int size, Bool normalized, Bool isInteger, Bool isBgra = false);
+        static Bool IsScaledIntegerVertexFormat(VkFormat format);
+        static VkFormat ToFloat32VertexFormat(Int componentCount);
+        Bool SupportsVertexBufferFormat(VkFormat format) const;
 
         const VulkanRendererConfig& m_config;
+        VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
         UnorderedMap<HashType, BackendVertexInputState> m_cache;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferObject.cpp

@@ -157,6 +157,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
+    Bool VkBufferObject::Invalidate(VkDeviceSize size, VkDeviceSize offset) {
+        MOBILEGL_ASSERT(IsValid(), "VkBufferObject::Invalidate called on invalid buffer");
+        MOBILEGL_ASSERT(IsMapped(), "VkBufferObject::Invalidate requires mapped memory");
+        MOBILEGL_ASSERT(offset <= m_size, "VkBufferObject::Invalidate offset out of range");
+
+        const VkDeviceSize resolvedSize = size == VK_WHOLE_SIZE ? m_size - offset : size;
+        MOBILEGL_ASSERT(offset + resolvedSize <= m_size, "VkBufferObject::Invalidate range out of bounds");
+        if (resolvedSize == 0) {
+            return true;
+        }
+
+        const VkResult result = vmaInvalidateAllocation(m_allocator, m_allocation, offset, resolvedSize);
+        if (result != VK_SUCCESS) {
+            MGLOG_E("VkBufferObject::Invalidate failed: vmaInvalidateAllocation returned %d", result);
+            return false;
+        }
+        return true;
+    }
+
     BufferSlice VkBufferObject::GetSlice(VkDeviceSize offset, VkDeviceSize size) const {
         MOBILEGL_ASSERT(offset <= m_size, "VkBufferObject::GetSlice offset out of range");
         const VkDeviceSize resolvedSize = (size == VK_WHOLE_SIZE) ? (m_size - offset) : size;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferObject.h

@@ -44,6 +44,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void* Map();
         void Unmap();
         Bool Upload(const void* data, VkDeviceSize size, VkDeviceSize offset = 0);
+        Bool Invalidate(VkDeviceSize size = VK_WHOLE_SIZE, VkDeviceSize offset = 0);
 
         VkBuffer GetHandle() const { return m_buffer; }
         VkDeviceSize GetSize() const { return m_size; }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp

@@ -65,8 +65,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
-    Float VkSamplerManager::ResolveEffectiveMaxAnisotropy(const MG_State::GLState::SamplerObject& sampler) const {
+    Float VkSamplerManager::ResolveEffectiveMaxAnisotropy(const MG_State::GLState::SamplerObject& sampler,
+                                                           Bool forceNearestFiltering) const {
         if (!m_samplerAnisotropySupported) return 1.0f;
+        if (forceNearestFiltering) return 1.0f;
         // VUID-VkSamplerCreateInfo-anisotropyEnable-01071/01072: anisotropy requires both filters to
         // be LINEAR and the value to sit within [1, limits.maxSamplerAnisotropy].
         if (sampler.GetMinFilter() != SamplerFilterMode::Linear ||
@@ -90,10 +92,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     Uint64 VkSamplerManager::BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
-                                             const MG_State::GLState::ITextureObject& texture) const {
+                                             const MG_State::GLState::ITextureObject& texture,
+                                             Bool forceNearestFiltering) const {
         MOBILEGL_ASSERT(m_config != nullptr, "VkSamplerManager::BuildSamplerKey: m_config is null");
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config->CacheVersion));
 
+        XXHASH_VERIFY(XXH64_update(m_hashState, &forceNearestFiltering, sizeof(forceNearestFiltering)));
+
         const auto minFilter = sampler.GetMinFilter();
         XXHASH_VERIFY(XXH64_update(m_hashState, &minFilter, sizeof(minFilter)));
         const auto magFilter = sampler.GetMagFilter();
@@ -115,7 +120,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // 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);
+        const auto maxAnisotropy = ResolveEffectiveMaxAnisotropy(sampler, forceNearestFiltering);
         XXHASH_VERIFY(XXH64_update(m_hashState, &maxAnisotropy, sizeof(maxAnisotropy)));
         const auto compareMode = sampler.GetCompareMode();
         XXHASH_VERIFY(XXH64_update(m_hashState, &compareMode, sizeof(compareMode)));
@@ -127,8 +132,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     VkSampler VkSamplerManager::GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
-                                                   const MG_State::GLState::ITextureObject& texture) {
-        const Uint64 key = BuildSamplerKey(sampler, texture);
+                                                   const MG_State::GLState::ITextureObject& texture,
+                                                   Bool forceNearestFiltering) {
+        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering);
         auto it = m_samplers.find(key);
         if (it != m_samplers.end()) {
             return it->second.handle;
@@ -136,16 +142,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkSamplerCreateInfo samplerInfo{};
         samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
-        samplerInfo.magFilter = ToVkFilter(sampler.GetMagFilter());
-        samplerInfo.minFilter = ToVkFilter(sampler.GetMinFilter());
-        samplerInfo.mipmapMode = ToVkMipmapMode(sampler.GetMipmapMode());
+        samplerInfo.magFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMagFilter());
+        samplerInfo.minFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMinFilter());
+        samplerInfo.mipmapMode = forceNearestFiltering ? VK_SAMPLER_MIPMAP_MODE_NEAREST
+                                                       : ToVkMipmapMode(sampler.GetMipmapMode());
         samplerInfo.addressModeU = ToVkAddressMode(sampler.GetWrapS());
         samplerInfo.addressModeV = ToVkAddressMode(sampler.GetWrapT());
         samplerInfo.addressModeW = ToVkAddressMode(sampler.GetWrapR());
         samplerInfo.mipLodBias = sampler.GetLodBias();
         // Must use the same resolver as BuildSamplerKey - a divergence would either collide two
         // different samplers or silently create duplicates.
-        const Float maxAnisotropy = ResolveEffectiveMaxAnisotropy(sampler);
+        const Float maxAnisotropy = ResolveEffectiveMaxAnisotropy(sampler, forceNearestFiltering);
         samplerInfo.anisotropyEnable = maxAnisotropy > 1.0f ? VK_TRUE : VK_FALSE;
         samplerInfo.maxAnisotropy = maxAnisotropy;
         samplerInfo.compareEnable = sampler.GetCompareMode() == SamplerCompareMode::CompareToTexture ? VK_TRUE : VK_FALSE;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.h

@@ -34,7 +34,8 @@ public:
     void Shutdown();
 
     VkSampler GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
-                                 const MG_State::GLState::ITextureObject& texture);
+                                 const MG_State::GLState::ITextureObject& texture,
+                                 Bool forceNearestFiltering = false);
 
 private:
     struct SamplerCacheEntry {
@@ -44,7 +45,8 @@ private:
     };
 
     Uint64 BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
-                           const MG_State::GLState::ITextureObject& texture) const;
+                           const MG_State::GLState::ITextureObject& texture,
+                           Bool forceNearestFiltering) const;
     static VkFilter ToVkFilter(SamplerFilterMode mode);
     static VkSamplerMipmapMode ToVkMipmapMode(SamplerMipmapMode mode);
     static VkSamplerAddressMode ToVkAddressMode(SamplerWrapMode mode);
@@ -57,7 +59,8 @@ private:
     // the sampler filters linearly both ways, otherwise the GL request clamped to the device limit.
     // GL happily carries GL_TEXTURE_MAX_ANISOTROPY on a NEAREST sampler (Blaze3D's blocks do exactly
     // that) while Vulkan forbids anisotropyEnable there, so the GL value must never be forwarded raw.
-    Float ResolveEffectiveMaxAnisotropy(const MG_State::GLState::SamplerObject& sampler) const;
+    Float ResolveEffectiveMaxAnisotropy(const MG_State::GLState::SamplerObject& sampler,
+                                        Bool forceNearestFiltering) const;
 
     VkDevice m_device = VK_NULL_HANDLE;
     const VulkanRendererConfig* m_config = nullptr;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -8,6 +8,8 @@
 
 #include "VkTextureManager.h"
 
+#include "ProgramFactory.h"
+
 #include "MG_State/GLState/Core.h"
 #include "MG_Util/Converters/MGToStr/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
@@ -17,6 +19,7 @@
 #include <cstdlib>
 #include <cstring>
 #include <memory>
+#include <vulkan/utility/vk_format_utils.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
     // Compute shaders may legally sample framebuffer-attached textures (the GL feedback-loop rule
@@ -63,6 +66,59 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                target == TextureUploadTarget::ProxyTexture2DMultisampleArray;
     }
 
+    static Bool IsMutableStorageImageFormat(VkFormat format) {
+        if (!vkuFormatIsColor(format) || vkuFormatIsCompressed(format)) {
+            return false;
+        }
+
+        // These are the uncompressed color compatibility classes covered by the core GLSL/SPIR-V
+        // storage-image formats. OpenGL mutable texture storage uses image-format compatibility by
+        // size, so a shader may legally reinterpret (for example) RGBA16_UNORM storage as rgba16f. Vulkan
+        // requires the image to be mutable and the view formats to share this exact compatibility
+        // class for the equivalent operation.
+        switch (vkuFormatCompatibilityClass(format)) {
+        case VKU_FORMAT_COMPATIBILITY_CLASS_8BIT:
+        case VKU_FORMAT_COMPATIBILITY_CLASS_16BIT:
+        case VKU_FORMAT_COMPATIBILITY_CLASS_32BIT:
+        case VKU_FORMAT_COMPATIBILITY_CLASS_64BIT:
+        case VKU_FORMAT_COMPATIBILITY_CLASS_128BIT:
+            return true;
+        default:
+            return false;
+        }
+    }
+
+    static Bool HasMatchingColorComponentLayout(VkFormat lhs, VkFormat rhs) {
+        const VKU_FORMAT_INFO lhsInfo = vkuGetFormatInfo(lhs);
+        const VKU_FORMAT_INFO rhsInfo = vkuGetFormatInfo(rhs);
+        if (lhsInfo.component_count == 0 || lhsInfo.component_count != rhsInfo.component_count ||
+            lhsInfo.texel_block_size != rhsInfo.texel_block_size ||
+            lhsInfo.texels_per_block != 1 || rhsInfo.texels_per_block != 1) {
+            return false;
+        }
+        for (Uint32 component = 0; component < lhsInfo.component_count; ++component) {
+            if (lhsInfo.components[component].type != rhsInfo.components[component].type ||
+                lhsInfo.components[component].size != rhsInfo.components[component].size) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    static Bool FormatMatchesSamplerNumericDomain(VkFormat format, SamplerNumericDomain numericDomain) {
+        switch (numericDomain) {
+        case SamplerNumericDomain::Float:
+            return vkuFormatIsSampledFloat(format);
+        case SamplerNumericDomain::SignedInteger:
+            return vkuFormatIsSINT(format);
+        case SamplerNumericDomain::UnsignedInteger:
+            return vkuFormatIsUINT(format);
+        case SamplerNumericDomain::Unknown:
+            return true;
+        }
+        return false;
+    }
+
     static Bool TryResolveSampleCountFlagBits(Int requestedSamples, VkSampleCountFlagBits& outSampleCount) {
         switch (requestedSamples) {
         case 1:
@@ -767,6 +823,180 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return perMipSampledView;
     }
 
+    VkImageView VkTextureManager::GetOrCreateSampledImageView(MG_State::GLState::ITextureObject& texture,
+                                                               VkFormat format) {
+        TextureResource* resource = SyncTextureAndGetDescriptor(texture);
+        if (resource == nullptr || resource->image == VK_NULL_HANDLE ||
+            resource->sampledView == VK_NULL_HANDLE) {
+            return VK_NULL_HANDLE;
+        }
+
+        if (format == VK_FORMAT_UNDEFINED || format == resource->format) {
+            return resource->sampledView;
+        }
+        if (!AreSampledImageViewFormatsCompatible(resource->format, format)) {
+            MGLOG_E("%s: incompatible sampled image view format=%d for textureId=%d imageFormat=%d",
+                    __func__, static_cast<Int>(format), texture.GetExternalIndex(),
+                    static_cast<Int>(resource->format));
+            return VK_NULL_HANDLE;
+        }
+        if ((resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
+            MGLOG_E("%s: textureId=%d needs mutable image format=%d for sampled view format=%d",
+                    __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
+                    static_cast<Int>(format));
+            return VK_NULL_HANDLE;
+        }
+
+        const TextureResource::SampledImageViewKey key{
+            .baseMipLevel = resource->sampledBaseMipLevel,
+            .levelCount = resource->sampledLevelCount,
+            .viewType = resource->viewType,
+            .format = format,
+        };
+        const auto existing = resource->alternateSampledViews.find(key);
+        if (existing != resource->alternateSampledViews.end()) {
+            return existing->second;
+        }
+
+        VkFormatProperties formatProperties{};
+        vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
+        if ((formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT) == 0) {
+            MGLOG_E("%s: sampled image view format=%d lacks VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT "
+                    "for textureId=%d (available=0x%x)",
+                    __func__, static_cast<Int>(format), texture.GetExternalIndex(),
+                    static_cast<Uint32>(formatProperties.optimalTilingFeatures));
+            return VK_NULL_HANDLE;
+        }
+
+        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
+        const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
+        const VkImageView view = CreateImageView(
+            resource->image, format, VK_IMAGE_ASPECT_COLOR_BIT, resource->viewType,
+            resource->sampledBaseMipLevel, resource->sampledLevelCount, 0, resource->arrayLayers,
+            &sampledComponents, VK_IMAGE_USAGE_SAMPLED_BIT);
+        if (view == VK_NULL_HANDLE) {
+            MGLOG_E("%s: failed to create sampled image view textureId=%d imageFormat=%d viewFormat=%d",
+                    __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
+                    static_cast<Int>(format));
+            return VK_NULL_HANDLE;
+        }
+
+        resource->alternateSampledViews.emplace(key, view);
+        MGLOG_D("%s: created sampled image view textureId=%d imageFormat=%d viewFormat=%d mip=[%u,%u)",
+                __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
+                static_cast<Int>(format), resource->sampledBaseMipLevel,
+                resource->sampledBaseMipLevel + resource->sampledLevelCount);
+        return view;
+    }
+
+    VkImageView VkTextureManager::GetOrCreateStorageImageView(MG_State::GLState::ITextureObject& texture,
+                                                               Uint32 mipLevel, VkFormat format,
+                                                               Bool layered, Int32 layer) {
+        TextureResource* resource = SyncTextureAndGetDescriptor(texture);
+        if (resource == nullptr || resource->image == VK_NULL_HANDLE || mipLevel >= resource->mipLevels ||
+            resource->sampleCount != VK_SAMPLE_COUNT_1_BIT ||
+            (resource->aspect & VK_IMAGE_ASPECT_COLOR_BIT) == 0) {
+            return VK_NULL_HANDLE;
+        }
+
+        if (format == VK_FORMAT_UNDEFINED) {
+            format = resource->format;
+        }
+        if (!AreStorageImageViewFormatsCompatible(resource->format, format)) {
+            MGLOG_E("%s: incompatible storage image view format=%d for textureId=%d imageFormat=%d",
+                    __func__, static_cast<Int>(format), texture.GetExternalIndex(),
+                    static_cast<Int>(resource->format));
+            return VK_NULL_HANDLE;
+        }
+        if (format != resource->format &&
+            (resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
+            MGLOG_E("%s: textureId=%d needs mutable image format=%d for storage view format=%d",
+                    __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
+                    static_cast<Int>(format));
+            return VK_NULL_HANDLE;
+        }
+
+        Uint32 baseArrayLayer = 0;
+        Uint32 layerCount = resource->arrayLayers;
+        VkImageViewType viewType = resource->viewType;
+        if (!layered) {
+            switch (resource->viewType) {
+            case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
+                viewType = VK_IMAGE_VIEW_TYPE_1D;
+                break;
+            case VK_IMAGE_VIEW_TYPE_2D_ARRAY:
+            case VK_IMAGE_VIEW_TYPE_CUBE:
+            case VK_IMAGE_VIEW_TYPE_CUBE_ARRAY:
+                viewType = VK_IMAGE_VIEW_TYPE_2D;
+                break;
+            case VK_IMAGE_VIEW_TYPE_3D:
+                MGLOG_E("%s: non-layered 3D storage views are unsupported for textureId=%d",
+                        __func__, texture.GetExternalIndex());
+                return VK_NULL_HANDLE;
+            default:
+                break;
+            }
+
+            if (viewType != resource->viewType) {
+                if (layer < 0 || static_cast<Uint32>(layer) >= resource->arrayLayers) {
+                    MGLOG_E("%s: storage image layer=%d is out of range for textureId=%d arrayLayers=%u",
+                            __func__, layer, texture.GetExternalIndex(), resource->arrayLayers);
+                    return VK_NULL_HANDLE;
+                }
+                baseArrayLayer = static_cast<Uint32>(layer);
+                layerCount = 1;
+            }
+        }
+
+        const Bool isFullResourceView = baseArrayLayer == 0 && layerCount == resource->arrayLayers &&
+                                        viewType == resource->viewType;
+        if (format == resource->format && isFullResourceView) {
+            return GetOrCreateViewAtMipLevel(texture, mipLevel);
+        }
+
+        const TextureResource::StorageImageViewKey key{
+            .mipLevel = mipLevel,
+            .baseArrayLayer = baseArrayLayer,
+            .layerCount = layerCount,
+            .viewType = viewType,
+            .format = format,
+        };
+        auto it = resource->storageImageViews.find(key);
+        if (it != resource->storageImageViews.end()) {
+            return it->second;
+        }
+
+        VkFormatFeatureFlags requiredFormatFeatures = VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+        if (format != resource->format &&
+            (format == VK_FORMAT_R32_UINT || format == VK_FORMAT_R32_SINT)) {
+            requiredFormatFeatures |= VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT;
+        }
+        VkFormatProperties formatProperties{};
+        vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
+        if ((formatProperties.optimalTilingFeatures & requiredFormatFeatures) != requiredFormatFeatures) {
+            MGLOG_E("%s: storage image view format=%d lacks required features=0x%x for textureId=%d "
+                    "(available=0x%x)",
+                    __func__, static_cast<Int>(format), static_cast<Uint32>(requiredFormatFeatures),
+                    texture.GetExternalIndex(), static_cast<Uint32>(formatProperties.optimalTilingFeatures));
+            return VK_NULL_HANDLE;
+        }
+
+        const VkImageView view = CreateImageView(resource->image, format, VK_IMAGE_ASPECT_COLOR_BIT, viewType,
+                                                 mipLevel, 1, baseArrayLayer, layerCount, nullptr,
+                                                 VK_IMAGE_USAGE_STORAGE_BIT);
+        if (view == VK_NULL_HANDLE) {
+            MGLOG_E("%s: failed to create storage image view for textureId=%d mip=%u imageFormat=%d viewFormat=%d",
+                    __func__, texture.GetExternalIndex(), mipLevel, static_cast<Int>(resource->format),
+                    static_cast<Int>(format));
+            return VK_NULL_HANDLE;
+        }
+        resource->storageImageViews.emplace(key, view);
+        MGLOG_D("%s: created storage image view textureId=%d mip=%u imageFormat=%d viewFormat=%d",
+                __func__, texture.GetExternalIndex(), mipLevel, static_cast<Int>(resource->format),
+                static_cast<Int>(format));
+        return view;
+    }
+
     void VkTextureManager::UpdateTrackedImageLayout(MG_State::GLState::ITextureObject* texture, VkImageLayout newLayout) {
         MOBILEGL_ASSERT(texture != nullptr, "UpdateTrackedImageLayout: texture is null");
         auto it = m_textureResources.find(MakeTextureIdentity(texture));
@@ -910,6 +1140,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    Bool VkTextureManager::NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const {
+        const auto it = m_textureResources.find(MakeTextureIdentity(&texture));
+        if (it == m_textureResources.end()) {
+            return true;
+        }
+        const TextureResource& resource = it->second;
+        if (resource.image == VK_NULL_HANDLE || resource.layout != VK_IMAGE_LAYOUT_GENERAL) {
+            return true;
+        }
+        // Mirror SyncTexture's cross-draw skip condition: any version drift means the sync
+        // path may upload or rebuild, both of which need the render pass ended first.
+        const auto* mipTexture = MG_State::GLState::AsMipmapTexture(&texture);
+        const Uint32 mipLevelCount = mipTexture != nullptr ? mipTexture->GetMipmapLevelCount() : 0u;
+        return resource.syncedContentVersion != texture.GetContentVersion() ||
+               resource.syncedTextureParamsVersion != texture.GetTextureParamsVersion() ||
+               resource.syncedMipLevelCount != mipLevelCount;
+    }
+
     Bool VkTextureManager::TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image,
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
@@ -1085,6 +1333,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return false;
         }
 
+        const VkImageAspectFlags aspect = GetAspectMaskForFormat(format);
+        VkFormatProperties formatProperties{};
+        vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
+        const Bool supportsStorageImage =
+            !isMultisampleTexture &&
+            (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
+            (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT) != 0;
+        VkImageCreateFlags imageCreateFlags = shapeInfo.imageFlags;
+        if (supportsStorageImage && IsMutableStorageImageFormat(format) &&
+            m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
+            imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+        }
+
         const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
                                 resource.extent.width == static_cast<Uint32>(texelSize.x()) &&
                                 resource.extent.height == static_cast<Uint32>(texelSize.y()) &&
@@ -1092,6 +1353,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 resource.arrayLayers == shapeInfo.arrayLayers &&
                                 resource.viewType == shapeInfo.viewType &&
                                 resource.sampleCount == resolvedSampleCount &&
+                                resource.imageCreateFlags == imageCreateFlags &&
                                 resource.mipLevels == backingMipLevels;
         if (compatible) {
             if (resource.perMipViews.size() != backingMipLevels) {
@@ -1112,6 +1374,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.arrayLayers == shapeInfo.arrayLayers &&
             resource.viewType == shapeInfo.viewType &&
             resource.sampleCount == resolvedSampleCount &&
+            resource.imageCreateFlags == imageCreateFlags &&
             resolvedSampleCount == VK_SAMPLE_COUNT_1_BIT &&
             resource.mipLevels < backingMipLevels &&
             resource.layout != VK_IMAGE_LAYOUT_UNDEFINED;
@@ -1123,26 +1386,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             DeferResourceRelease(Move(resource));
         }
 
-        auto aspect = GetAspectMaskForFormat(format);
-
         VkImageCreateInfo imageInfo{};
         imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
-    imageInfo.flags = shapeInfo.imageFlags;
-    imageInfo.imageType = shapeInfo.imageType;
+        imageInfo.flags = imageCreateFlags;
+        imageInfo.imageType = shapeInfo.imageType;
         imageInfo.extent.width = static_cast<Uint32>(texelSize.x());
         imageInfo.extent.height = static_cast<Uint32>(texelSize.y());
-    imageInfo.extent.depth = shapeInfo.depth;
+        imageInfo.extent.depth = shapeInfo.depth;
         imageInfo.mipLevels = backingMipLevels;
-    imageInfo.arrayLayers = shapeInfo.arrayLayers;
+        imageInfo.arrayLayers = shapeInfo.arrayLayers;
         imageInfo.format = format;
         imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
         imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-        VkFormatProperties formatProperties{};
-        vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
-        const Bool supportsStorageImage =
-            !isMultisampleTexture &&
-            (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
-            (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT) != 0;
         imageInfo.usage = VK_IMAGE_USAGE_SAMPLED_BIT |
                           (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
                           ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
@@ -1153,15 +1408,35 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
         }
         imageInfo.samples = resolvedSampleCount;
-        if (isMultisampleTexture) {
+        if (isMultisampleTexture || (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
             VkImageFormatProperties imageFormatProperties{};
-            const VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
+            VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
                 m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage,
                 imageInfo.flags, &imageFormatProperties);
+            if (imageFormatResult != VK_SUCCESS && !isMultisampleTexture &&
+                (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+                // Losing reinterpreted views only degrades the formatless-image feature for
+                // this texture; failing creation would lose the texture entirely, so retry
+                // as a plain immutable-format image.
+                MGLOG_W("%s: mutable image format=%d is unsupported for textureId=%d; creating "
+                        "without VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT (format reinterpretation "
+                        "will be unavailable for it)",
+                        __func__, static_cast<Int>(format), texture.GetExternalIndex());
+                // Remember the verdict so later syncs of same-format textures neither retry
+                // the probe nor flag-mismatch against this image and recreate it.
+                m_mutableFormatUnsupported.insert(format);
+                imageInfo.flags &= ~VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+                imageCreateFlags = imageInfo.flags;
+                imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
+                    m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage,
+                    imageInfo.flags, &imageFormatProperties);
+            }
             if (imageFormatResult != VK_SUCCESS ||
-                (imageFormatProperties.sampleCounts & resolvedSampleCount) == 0) {
-                MGLOG_D("%s: sampleCount=%d is unsupported for textureId=%d target=%s format=%d usage=0x%x",
-                        __func__, texture.GetSamples(), texture.GetExternalIndex(),
+                (isMultisampleTexture && (imageFormatProperties.sampleCounts & resolvedSampleCount) == 0)) {
+                MGLOG_D("%s: image flags=0x%x sampleCount=%d are unsupported for textureId=%d target=%s "
+                        "format=%d usage=0x%x",
+                        __func__, static_cast<Uint32>(imageInfo.flags), texture.GetSamples(),
+                        texture.GetExternalIndex(),
                         MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
                         static_cast<Int>(format), static_cast<Uint32>(imageInfo.usage));
                 return false;
@@ -1188,6 +1463,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         resource.aspect = aspect;
         resource.viewType = shapeInfo.viewType;
         resource.sampleCount = resolvedSampleCount;
+        resource.imageCreateFlags = imageCreateFlags;
         resource.syncedTextureParamsVersion = 0;
 
         if (preservedResource) {
@@ -1203,7 +1479,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkTextureManager::DeferResourceRelease(TextureResource&& resource) {
         if (resource.image == VK_NULL_HANDLE && resource.fullView == VK_NULL_HANDLE &&
             resource.sampledView == VK_NULL_HANDLE &&
-            resource.perMipViews.empty() && resource.perMipSampledViews.empty()) {
+            resource.perMipViews.empty() && resource.perMipSampledViews.empty() &&
+            resource.attachmentViews.empty() && resource.alternateSampledViews.empty() &&
+            resource.storageImageViews.empty()) {
             return;
         }
 
@@ -1299,6 +1577,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 sampledView = VK_NULL_HANDLE;
             }
         }
+        for (const auto& [_, sampledView] : resource.alternateSampledViews) {
+            DeferViewRelease(sampledView);
+        }
+        resource.alternateSampledViews.clear();
 
         const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
         const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
@@ -1324,7 +1606,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                   VkImageViewType viewType, Uint32 baseMipLevel, Uint32 levelCount,
                                                   Uint32 baseArrayLayer,
                                                   Uint32 layerCount,
-                                                  const VkComponentMapping* components) const {
+                                                  const VkComponentMapping* components,
+                                                  VkImageUsageFlags viewUsage) const {
         VkImageViewCreateInfo viewInfo{};
         viewInfo.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
         viewInfo.image = image;
@@ -1340,6 +1623,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         viewInfo.subresourceRange.baseArrayLayer = baseArrayLayer;
         viewInfo.subresourceRange.layerCount = layerCount;
 
+        VkImageViewUsageCreateInfo usageInfo{};
+        if (viewUsage != 0) {
+            usageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO;
+            usageInfo.usage = viewUsage;
+            viewInfo.pNext = &usageInfo;
+        }
+
         VkImageView view = VK_NULL_HANDLE;
         VK_VERIFY(vkCreateImageView(m_device, &viewInfo, nullptr, &view), "vkCreateImageView(texture)");
         return view;
@@ -1665,4 +1955,64 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         return imageAspect;
     }
+
+    VkFormat VkTextureManager::ResolveSampledImageViewFormat(VkFormat imageFormat,
+                                                              SamplerNumericDomain numericDomain) {
+        // Depth/stencil images always sample through the existing depth-aspect sampledView.
+        // Combined formats (D24S8, D32FS8) are multi-numeric, so vkuFormatIsSampledFloat is
+        // false for them by design, yet their depth aspect reads as float in every GL depth
+        // texture mode; Vulkan also forbids reinterpreting them through color-class views.
+        // Integer domains keep the same view (pre-reinterpretation behavior for stencil-index
+        // style access) rather than failing the draw.
+        if (vkuFormatIsDepthOrStencil(imageFormat)) {
+            return imageFormat;
+        }
+        if (imageFormat == VK_FORMAT_UNDEFINED || numericDomain == SamplerNumericDomain::Unknown ||
+            FormatMatchesSamplerNumericDomain(imageFormat, numericDomain)) {
+            return imageFormat;
+        }
+        if (!IsMutableStorageImageFormat(imageFormat)) {
+            return VK_FORMAT_UNDEFINED;
+        }
+
+        // Preserve component ordering and bit widths. This selects R32_UINT for an R32_SFLOAT
+        // texture sampled by a usampler rather than an arbitrary member (such as
+        // R8G8B8A8_UINT) of Vulkan's broad 32-bit compatibility class.
+        for (Int candidateValue = static_cast<Int>(VK_FORMAT_R4G4_UNORM_PACK8);
+             candidateValue <= static_cast<Int>(VK_FORMAT_ASTC_12x12_SRGB_BLOCK);
+             ++candidateValue) {
+            const VkFormat candidate = static_cast<VkFormat>(candidateValue);
+            if (!IsMutableStorageImageFormat(candidate) ||
+                !FormatMatchesSamplerNumericDomain(candidate, numericDomain) ||
+                !HasMatchingColorComponentLayout(imageFormat, candidate) ||
+                !AreSampledImageViewFormatsCompatible(imageFormat, candidate)) {
+                continue;
+            }
+
+            // If an integer backing is intentionally bit-read through a float sampler, require
+            // a true floating-point view. Normalized/scaled views satisfy OpTypeFloat but apply
+            // an unrelated numeric conversion to those bits.
+            if (numericDomain == SamplerNumericDomain::Float && !vkuFormatIsSFLOAT(candidate)) {
+                continue;
+            }
+            return candidate;
+        }
+        return VK_FORMAT_UNDEFINED;
+    }
+
+    Bool VkTextureManager::AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat) {
+        if (imageFormat == viewFormat) {
+            return true;
+        }
+        return IsMutableStorageImageFormat(imageFormat) && IsMutableStorageImageFormat(viewFormat) &&
+               vkuFormatCompatibilityClass(imageFormat) == vkuFormatCompatibilityClass(viewFormat);
+    }
+
+    Bool VkTextureManager::AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat) {
+        if (imageFormat == viewFormat) {
+            return true;
+        }
+        return IsMutableStorageImageFormat(imageFormat) && IsMutableStorageImageFormat(viewFormat) &&
+               vkuFormatCompatibilityClass(imageFormat) == vkuFormatCompatibilityClass(viewFormat);
+    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -13,12 +13,15 @@
 #include <MG_State/GLState/TextureState/TextureObject.h>
 #include <vk_mem_alloc.h>
 #include <unordered_map>
+#include <unordered_set>
 
 namespace MobileGL::MG_State::GLState {
 class ITextureObject;
 }
 
 namespace MobileGL::MG_Backend::DirectVulkan {
+enum class SamplerNumericDomain : Uint8;
+
 class VkTextureManager {
 public:
     // Monotonic epoch bumped whenever a texture VkImage is (re)created. The render-pass
@@ -78,6 +81,61 @@ public:
             }
         };
 
+        struct StorageImageViewKey {
+            Uint32 mipLevel = 0;
+            Uint32 baseArrayLayer = 0;
+            Uint32 layerCount = 1;
+            VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
+            VkFormat format = VK_FORMAT_UNDEFINED;
+
+            Bool operator==(const StorageImageViewKey& other) const {
+                return mipLevel == other.mipLevel &&
+                       baseArrayLayer == other.baseArrayLayer &&
+                       layerCount == other.layerCount &&
+                       viewType == other.viewType &&
+                       format == other.format;
+            }
+        };
+
+        struct SampledImageViewKey {
+            Uint32 baseMipLevel = 0;
+            Uint32 levelCount = 1;
+            VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
+            VkFormat format = VK_FORMAT_UNDEFINED;
+
+            Bool operator==(const SampledImageViewKey& other) const {
+                return baseMipLevel == other.baseMipLevel &&
+                       levelCount == other.levelCount &&
+                       viewType == other.viewType &&
+                       format == other.format;
+            }
+        };
+
+        struct SampledImageViewKeyHash {
+            SizeT operator()(const SampledImageViewKey& key) const {
+                SizeT hash = std::hash<Uint32>{}(key.baseMipLevel);
+                hash ^= std::hash<Uint32>{}(key.levelCount) + 0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.viewType)) +
+                        0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.format)) +
+                        0x9e3779b9u + (hash << 6) + (hash >> 2);
+                return hash;
+            }
+        };
+
+        struct StorageImageViewKeyHash {
+            SizeT operator()(const StorageImageViewKey& key) const {
+                SizeT hash = std::hash<Uint32>{}(key.mipLevel);
+                hash ^= std::hash<Uint32>{}(key.baseArrayLayer) + 0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(key.layerCount) + 0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.viewType)) +
+                        0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.format)) +
+                        0x9e3779b9u + (hash << 6) + (hash >> 2);
+                return hash;
+            }
+        };
+
         VkImage image = VK_NULL_HANDLE;
         VmaAllocation allocation = nullptr;
         VkImageView fullView = VK_NULL_HANDLE;
@@ -85,6 +143,8 @@ public:
         Vector<VkImageView> perMipViews;
         Vector<VkImageView> perMipSampledViews;
         UnorderedMap<AttachmentViewKey, VkImageView, AttachmentViewKeyHash> attachmentViews;
+        UnorderedMap<SampledImageViewKey, VkImageView, SampledImageViewKeyHash> alternateSampledViews;
+        UnorderedMap<StorageImageViewKey, VkImageView, StorageImageViewKeyHash> storageImageViews;
         VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
         VkExtent2D extent = {0, 0};
         Uint32 depth = 1;
@@ -96,6 +156,7 @@ public:
         VkImageAspectFlags aspect = VK_IMAGE_ASPECT_NONE;
         VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
+        VkImageCreateFlags imageCreateFlags = 0;
         Uint16 syncedTextureParamsVersion = 0;
         // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
         // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
@@ -115,6 +176,8 @@ public:
             std::swap(this->perMipViews, that.perMipViews);
             std::swap(this->perMipSampledViews, that.perMipSampledViews);
             std::swap(this->attachmentViews, that.attachmentViews);
+            std::swap(this->alternateSampledViews, that.alternateSampledViews);
+            std::swap(this->storageImageViews, that.storageImageViews);
             std::swap(this->layout, that.layout);
             std::swap(this->extent, that.extent);
             std::swap(this->depth, that.depth);
@@ -126,6 +189,7 @@ public:
             std::swap(this->aspect, that.aspect);
             std::swap(this->viewType, that.viewType);
             std::swap(this->sampleCount, that.sampleCount);
+            std::swap(this->imageCreateFlags, that.imageCreateFlags);
             std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
             std::swap(this->syncedContentVersion, that.syncedContentVersion);
             std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
@@ -153,6 +217,16 @@ public:
                     vkDestroyImageView(s_device, attachmentView, nullptr);
                 }
             }
+            for (const auto& [_, sampledView] : alternateSampledViews) {
+                if (sampledView != VK_NULL_HANDLE) {
+                    vkDestroyImageView(s_device, sampledView, nullptr);
+                }
+            }
+            for (const auto& [_, storageImageView] : storageImageViews) {
+                if (storageImageView != VK_NULL_HANDLE) {
+                    vkDestroyImageView(s_device, storageImageView, nullptr);
+                }
+            }
             if (image != VK_NULL_HANDLE && allocation != nullptr) {
                 vmaDestroyImage(s_allocator, image, allocation);
             }
@@ -161,6 +235,8 @@ public:
             perMipViews.clear();
             perMipSampledViews.clear();
             attachmentViews.clear();
+            alternateSampledViews.clear();
+            storageImageViews.clear();
             image = VK_NULL_HANDLE;
             allocation = nullptr;
             layout = VK_IMAGE_LAYOUT_UNDEFINED;
@@ -174,6 +250,7 @@ public:
             aspect = VK_IMAGE_ASPECT_NONE;
             viewType = VK_IMAGE_VIEW_TYPE_2D;
             sampleCount = VK_SAMPLE_COUNT_1_BIT;
+            imageCreateFlags = 0;
             syncedTextureParamsVersion = 0;
             syncedContentVersion = 0;
             syncedMipLevelCount = 0;
@@ -198,6 +275,9 @@ public:
                                                     Uint32 baseArrayLayer, Uint32 layerCount,
                                                     VkImageViewType viewType);
     VkImageView GetOrCreateSampledViewAtMipLevel(MG_State::GLState::ITextureObject& texture, Uint32 mipLevel);
+    VkImageView GetOrCreateSampledImageView(MG_State::GLState::ITextureObject& texture, VkFormat format);
+    VkImageView GetOrCreateStorageImageView(MG_State::GLState::ITextureObject& texture, Uint32 mipLevel,
+                                            VkFormat format, Bool layered, Int32 layer);
     void UpdateTrackedImageLayout(MG_State::GLState::ITextureObject* texture, VkImageLayout newLayout);
     void UpdateTrackedImageLayoutAfterAttachmentWrite(VkCommandBuffer commandBuffer,
                                                       MG_State::GLState::ITextureObject* texture,
@@ -205,8 +285,16 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Non-mutating probe for the per-draw storage-image fast path: true when preparing this
+    // texture as a storage image may need work that is illegal inside a render pass (resource
+    // creation, dirty-content upload, or a layout transition to GENERAL). Unknown state reports
+    // true - a false positive merely ends the render pass, a false negative would skip a barrier.
+    Bool NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const;
 
     static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect);
+    static VkFormat ResolveSampledImageViewFormat(VkFormat imageFormat, SamplerNumericDomain numericDomain);
+    static Bool AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
+    static Bool AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
 
     static Bool TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image, VkImageLayout& trackedLayout,
                                VkImageLayout newLayout, VkPipelineStageFlags srcStageMask,
@@ -255,7 +343,8 @@ private:
                                 VkImageViewType viewType, Uint32 baseMipLevel, Uint32 levelCount,
                                 Uint32 baseArrayLayer,
                                 Uint32 layerCount,
-                                const VkComponentMapping* components = nullptr) const;
+                                const VkComponentMapping* components = nullptr,
+                                VkImageUsageFlags viewUsage = 0) const;
     Bool UploadDirtyMipLevels(MG_State::GLState::TextureObjectMipmap &mipmapTexture,
                       TextureUploadTarget uploadTarget,
                       TextureResource &outResource);
@@ -296,6 +385,9 @@ private:
         TextureResource* resource = nullptr;
     };
     Vector<DrawSyncedTexture> m_drawSyncedThisDraw;
+    // Formats whose mutable-image probe failed on this device; their images are created
+    // without MUTABLE_FORMAT_BIT so repeat syncs neither re-probe nor flag-mismatch.
+    std::unordered_set<VkFormat> m_mutableFormatUnsupported;
     std::unordered_map<TextureIdentity, WeakPtr<MG_State::GLState::ITextureObject>, TextureIdentityHash> m_aliveObjects;
     std::unordered_map<TextureIdentity, TextureResource, TextureIdentityHash> m_textureResources;
     Vector<Vector<TextureResource>> m_deferredReleases;

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -51,6 +51,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint32 instanceCount = 1;
         Uint32 firstVertex = 0;
         Uint32 firstInstance = 0;
+        // Indexed-draw metadata for bounding vertex-stream conversion. baseVertex is the
+        // draw's base-vertex offset; indexRangeIsExactView is true only when the draw
+        // fetches exactly the indices its IndexBufferView describes (direct DrawElements;
+        // multi/indirect forms leave it false because the CPU cannot bound their ranges).
+        Int32 baseVertex = 0;
+        Bool indexRangeIsExactView = false;
     };
 
     struct DrawIndexedCmdParam {
@@ -165,6 +171,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                               GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth);
         void GenerateMipmap(GLenum target);
         void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels);
+        static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
+        static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
+                                          GLsizei width, GLsizei height, GLenum destinationFormat,
+                                          GLenum destinationType, SizeT destinationRowStride,
+                                          Uint8* destinationPixels);
         void GetTexImage(GLenum target, GLint level, GLenum format, GLenum type, GLvoid* pixels);
         void GetTextureImage(const SharedPtr<MG_State::GLState::ITextureObject>& texture,
                              TextureUploadTarget uploadTarget, GLint level, GLenum format, GLenum type,
@@ -365,6 +376,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool m_samplerAnisotropyFeatureEnabled = false;
         Bool m_shaderDrawParametersExtensionEnabled = false;
         Bool m_shaderDrawParametersFeatureEnabled = false;
+        Bool m_unformattedFloatStorageImagesEnabled = false;
         // fillModeNonSolid gates VK_POLYGON_MODE_LINE/_POINT (glPolygonMode); independentBlend gates
         // per-draw-buffer color write masks (glColorMaski). Both are cached at device creation and
         // drive a runtime fallback when the device lacks them.
@@ -437,9 +449,64 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Per-draw scratch buffers (clear keeps capacity) — these paths run for every
         // draw call and must not allocate.
         Vector<MG_State::GLState::ITextureObject*> m_sampledTexturesScratch;
+        Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
         Vector<VkBuffer> m_vertexBuffersScratch;
         Vector<VkDeviceSize> m_vertexOffsetsScratch;
         Vector<VkVertexInputAttributeDescription> m_patchedAttributesScratch;
+        Vector<Float> m_vertexConversionScratch;
+        Vector<Uint8> m_vertexRepackScratch;
+
+        struct ConvertedVertexStreamKey {
+            const MG_State::GLState::BufferObject* buffer = nullptr;
+            Uint64 changeSerial = 0;
+            SizeT baseOffset = 0;
+            Uint32 sourceStride = 0;
+            DataType type = DataType::Float32;
+            Int size = 0;
+            Bool normalized = false;
+            Bool isInteger = false;
+            VertexInputStateFactory::VertexStreamConversion conversion =
+                VertexInputStateFactory::VertexStreamConversion::None;
+
+            Bool operator==(const ConvertedVertexStreamKey& other) const {
+                return buffer == other.buffer && changeSerial == other.changeSerial &&
+                       baseOffset == other.baseOffset && sourceStride == other.sourceStride &&
+                       type == other.type && size == other.size && normalized == other.normalized &&
+                       isInteger == other.isInteger && conversion == other.conversion;
+            }
+        };
+
+        struct ConvertedVertexStreamKeyHash {
+            SizeT operator()(const ConvertedVertexStreamKey& key) const {
+                SizeT hash = std::hash<const void*>{}(key.buffer);
+                auto combine = [&hash](SizeT value) {
+                    hash ^= value + static_cast<SizeT>(0x9e3779b97f4a7c15ull) + (hash << 6) + (hash >> 2);
+                };
+                combine(std::hash<Uint64>{}(key.changeSerial));
+                combine(std::hash<SizeT>{}(key.baseOffset));
+                combine(std::hash<Uint32>{}(key.sourceStride));
+                combine(std::hash<Uint32>{}(static_cast<Uint32>(key.type)));
+                combine(std::hash<Int>{}(key.size));
+                combine(std::hash<Bool>{}(key.normalized));
+                combine(std::hash<Bool>{}(key.isInteger));
+                combine(std::hash<Uint32>{}(static_cast<Uint32>(key.conversion)));
+                return hash;
+            }
+        };
+
+        struct ConvertedVertexStream {
+            BufferSlice slice;
+            // Number of source elements the cached slice covers. A draw needing a prefix of
+            // this range reuses the slice (converted streams are tightly packed); a draw
+            // needing more reconverts and replaces the entry, so per (buffer, layout) a
+            // frame converts at most the largest range any draw asked for.
+            SizeT elementCount = 0;
+            // Pins the source buffer for the frame so its heap address cannot be reused by
+            // a new BufferObject while this pointer-keyed entry is alive.
+            SharedPtr<const MG_State::GLState::BufferObject> sourcePin;
+        };
+        UnorderedMap<ConvertedVertexStreamKey, ConvertedVertexStream, ConvertedVertexStreamKeyHash>
+            m_convertedVertexStreams;
 
         void CreateInstance();
         VkResult SetupDebugMessenger();
@@ -462,10 +529,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const RenderPassEntry& renderPassEntry);
         VkPipeline GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj);
         void DestroyComputePipelines();
+        Bool PrepareStorageImageTextures(
+            VkCommandBuffer commandBuffer,
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj);
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,
-                                        const DrawCmdParam& drawParams);
+                                        const DrawCmdParam& drawParams,
+                                        const IndexBufferView* pIndexBufferView);
         Bool UploadAndBindIndexBuffer(FrameContext::FrameData& frame,
                                      const MG_State::GLState::VertexArrayObject& vao,
                                       const IndexBufferView* pIndexBufferView = nullptr);

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -998,8 +998,8 @@ DECLARE_GL_FUNCTION_HEAD(void, ShaderStorageBlockBinding, GLuint program, GLuint
 DECLARE_GL_FUNCTION_STUB_HEAD(void, TextureView, GLuint texture, GLenum target, GLuint origtexture, GLenum internalformat, GLuint minlevel, GLuint numlevels, GLuint minlayer, GLuint numlayers) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, TextureView, texture, target, origtexture, internalformat, minlevel, numlevels, minlayer, numlayers)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribLFormat, GLuint attribindex, GLint size, GLenum type, GLuint relativeoffset) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribLFormat, attribindex, size, type, relativeoffset)
 DECLARE_GL_FUNCTION_HEAD(void, BufferStorage, GLenum target, GLsizeiptr size, const void* data, GLbitfield flags) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BufferStorage, target, size, data, flags)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearTexImage, GLuint texture, GLint level, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearTexImage, texture, level, format, type, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearTexImage, GLuint texture, GLint level, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearTexImage, texture, level, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, type, data)
 DECLARE_GL_FUNCTION_HEAD(void, BindBuffersBase, GLenum target, GLuint first, GLsizei count, const GLuint* buffers) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBuffersBase, target, first, count, buffers)
 DECLARE_GL_FUNCTION_HEAD(void, BindBuffersRange, GLenum target, GLuint first, GLsizei count, const GLuint* buffers, const GLintptr* offsets, const GLsizeiptr* sizes) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBuffersRange, target, first, count, buffers, offsets, sizes)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTextures, GLuint first, GLsizei count, const GLuint* textures) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BindTextures, first, count, textures)
@@ -1063,7 +1063,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage1D, GLuint texture,
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage3D, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, imageSize, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CopyTextureSubImage1D, GLuint texture, GLint level, GLint xoffset, GLint x, GLint y, GLsizei width) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CopyTextureSubImage1D, texture, level, xoffset, x, y, width)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, CopyTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CopyTextureSubImage2D, texture, level, xoffset, yoffset, x, y, width, height)
+DECLARE_GL_FUNCTION_HEAD(void, CopyTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CopyTextureSubImage2D, texture, level, xoffset, yoffset, x, y, width, height)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CopyTextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLint x, GLint y, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CopyTextureSubImage3D, texture, level, xoffset, yoffset, zoffset, x, y, width, height)
 DECLARE_GL_FUNCTION_HEAD(void, TextureParameterf, GLuint texture, GLenum pname, GLfloat param) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureParameterf, texture, pname, param)
 DECLARE_GL_FUNCTION_HEAD(void, TextureParameterfv, GLuint texture, GLenum pname, const GLfloat* param) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureParameterfv, texture, pname, param)

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -1175,10 +1175,9 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxFragmentInputComponents;
             return;
         case GL_MAX_FRAGMENT_IMAGE_UNIFORMS:
-            // TODO: Track per-stage image uniform limits separately instead of reusing the compute/backend stage cap.
             *params = MG_Backend::pActiveBackendObject
-                          ? MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxComputeImageUniforms
-                          : MG_Backend::DynamicBackendParameters{}.MaxComputeImageUniforms;
+                          ? MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxFragmentImageUniforms
+                          : MG_Backend::DynamicBackendParameters{}.MaxFragmentImageUniforms;
             return;
         case GL_MAX_FRAGMENT_UNIFORM_COMPONENTS:
             *params = kFrontendMaxFragmentUniformComponents;
@@ -1208,7 +1207,9 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxGeometryTextureImageUnits;
             return;
         case GL_MAX_GEOMETRY_IMAGE_UNIFORMS:
-            *params = 0;
+            *params = MG_Backend::pActiveBackendObject
+                          ? MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxGeometryImageUniforms
+                          : MG_Backend::DynamicBackendParameters{}.MaxGeometryImageUniforms;
             return;
         case GL_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
             *params = kFrontendMaxGeometryTotalOutputComponents;
@@ -1277,7 +1278,9 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxVertexAtomicCounters;
             return;
         case GL_MAX_VERTEX_IMAGE_UNIFORMS:
-            *params = 0;
+            *params = MG_Backend::pActiveBackendObject
+                          ? MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxVertexImageUniforms
+                          : MG_Backend::DynamicBackendParameters{}.MaxVertexImageUniforms;
             return;
         case GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS:
             *params = 16; // TODO

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -470,6 +470,223 @@ namespace MobileGL::MG_Impl::GLImpl {
         return textureObject;
     }
 
+    namespace {
+        void RecordClearTextureError(const char* caller, ErrorCode code, const String& message) {
+            MG_State::pGLContext->RecordError(
+                code, MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, message));
+        }
+
+        SharedPtr<MG_State::GLState::TextureObjectMipmap> GetClearTextureObject(GLuint texture, GLint level,
+                                                                                const char* caller) {
+            if (texture == 0) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        "Clear texture operations require a non-zero texture name.");
+                return nullptr;
+            }
+
+            auto textureObject = GetTextureObjectByName(texture, caller);
+            if (!textureObject) return nullptr;
+            if (textureObject->GetStorageType() != TextureStorageType::Mipmap) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        "Buffer textures cannot be cleared with glClearTexImage.");
+                return nullptr;
+            }
+
+            auto mipmapTexture = std::static_pointer_cast<MG_State::GLState::TextureObjectMipmap>(textureObject);
+            if (level < 0) {
+                RecordClearTextureError(caller, ErrorCode::InvalidValue,
+                                        std::format("Texture level {} is negative.", level));
+                return nullptr;
+            }
+            // ARB_clear_texture: clearing an image that was never defined by TexImage*/
+            // TexStorage* is INVALID_OPERATION, not INVALID_VALUE.
+            if (static_cast<Uint>(level) >= mipmapTexture->GetMipmapLevelCount()) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        std::format("Texture level {} is not defined.", level));
+                return nullptr;
+            }
+            return mipmapTexture;
+        }
+
+        Bool BuildClearPixel(const SharedPtr<MG_State::GLState::TextureObjectMipmap>& textureObject,
+                             GLenum format, GLenum type, const void* data, Vector<Uint8>& clearPixel) {
+            const TextureInputFormat inputFormat = MG_Util::ConvertGLEnumToTextureInputFormat(format);
+            const TexturePixelDataType inputType = MG_Util::ConvertGLEnumToTexturePixelDataType(type);
+            if (!TextureImpl::ValidateTextureInputFormat(inputFormat) ||
+                !TextureImpl::ValidateTexturePixelDataType(inputType) ||
+                !TextureImpl::ValidateTextureInternalFormatCompatibleWithInput(
+                    inputFormat, textureObject->GetFormat(), inputType)) {
+                return false;
+            }
+
+            clearPixel.clear();
+            if (data == nullptr) {
+                // ARB_clear_texture defines a null clear value as all zeroes. Keeping the
+                // pattern empty lets the region writer use a fast memset path.
+                return true;
+            }
+
+            PixelStoreParameters clearPixelStore{};
+            clearPixelStore.Alignment = 1;
+            SizeT clearPixelSize = 0;
+            void* converted = MG_Util::PixelStoreProcessor::ProcessTexturePixelsDataUnpack(
+                data, clearPixelStore, textureObject->GetFormat(), inputFormat, inputType,
+                {1, 1, 1}, false, clearPixelSize);
+            if (!converted || clearPixelSize == 0) {
+                if (converted) free(converted);
+                return false;
+            }
+
+            clearPixel.resize(clearPixelSize);
+            Memcpy(clearPixel.data(), converted, clearPixelSize);
+            free(converted);
+            return true;
+        }
+
+        // Writes the clear into the CPU shadow and marks the whole level dirty, exactly like
+        // TexSubImage*_State does. Shared limitation of the level-granular shadow sync: the
+        // shadow does not reflect GPU-side writes (FBO rendering, imageStore), so a PARTIAL
+        // clear of a GPU-written level re-uploads stale shadow bytes outside the region on
+        // the next sync. Full-level clears (glClearTexImage, or a sub-clear covering the
+        // level) rewrite the entire shadow and are always correct.
+        Bool ClearMipmapRegion(const SharedPtr<MG_State::GLState::TextureObjectMipmap>& textureObject,
+                               TextureUploadTarget uploadTarget, GLint level,
+                               GLint xoffset, GLint yoffset, GLint zoffset,
+                               GLsizei width, GLsizei height, GLsizei depth,
+                               const Vector<Uint8>& clearPixel, const char* caller) {
+            const IntVec3 texelSize = textureObject->GetMipmapTexelSize(uploadTarget, static_cast<Uint>(level));
+            if (texelSize.x() <= 0 || texelSize.y() <= 0 || texelSize.z() <= 0) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        "The requested texture level has no storage.");
+                return false;
+            }
+            if (xoffset < 0 || yoffset < 0 || zoffset < 0 ||
+                width < 0 || height < 0 || depth < 0 ||
+                width > texelSize.x() - xoffset ||
+                height > texelSize.y() - yoffset ||
+                depth > texelSize.z() - zoffset) {
+                RecordClearTextureError(caller, ErrorCode::InvalidValue,
+                                        "The clear region lies outside the requested texture level.");
+                return false;
+            }
+            if (width == 0 || height == 0 || depth == 0) return true;
+
+            const SizeT texelCount = static_cast<SizeT>(texelSize.x()) *
+                                     static_cast<SizeT>(texelSize.y()) *
+                                     static_cast<SizeT>(texelSize.z());
+            const SizeT byteSize = textureObject->GetMipmapByteSize(uploadTarget, static_cast<Uint>(level));
+            if (byteSize == 0 || byteSize % texelCount != 0) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        "The requested texture storage cannot be cleared.");
+                return false;
+            }
+
+            const SizeT bytesPerTexel = byteSize / texelCount;
+            if (!clearPixel.empty() && clearPixel.size() != bytesPerTexel) {
+                RecordClearTextureError(
+                    caller, ErrorCode::InvalidOperation,
+                    std::format("Converted clear value is {} bytes, but the texture stores {} bytes per texel.",
+                                clearPixel.size(), bytesPerTexel));
+                return false;
+            }
+
+            auto* destination = static_cast<Uint8*>(
+                textureObject->MapMipmapData(uploadTarget, static_cast<Uint>(level)));
+            if (!destination) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
+                                        "The requested texture level could not be mapped.");
+                return false;
+            }
+
+            const SizeT fullRowBytes = static_cast<SizeT>(texelSize.x()) * bytesPerTexel;
+            const SizeT fullSliceBytes = static_cast<SizeT>(texelSize.y()) * fullRowBytes;
+            const SizeT clearRowBytes = static_cast<SizeT>(width) * bytesPerTexel;
+            Uint8* firstClearRow = nullptr;
+
+            for (GLsizei z = 0; z < depth; ++z) {
+                for (GLsizei y = 0; y < height; ++y) {
+                    Uint8* row = destination +
+                                 static_cast<SizeT>(zoffset + z) * fullSliceBytes +
+                                 static_cast<SizeT>(yoffset + y) * fullRowBytes +
+                                 static_cast<SizeT>(xoffset) * bytesPerTexel;
+                    if (firstClearRow) {
+                        Memcpy(row, firstClearRow, clearRowBytes);
+                        continue;
+                    }
+
+                    firstClearRow = row;
+                    if (clearPixel.empty()) {
+                        Memset(row, 0, clearRowBytes);
+                        continue;
+                    }
+
+                    Memcpy(row, clearPixel.data(), bytesPerTexel);
+                    SizeT filled = bytesPerTexel;
+                    while (filled < clearRowBytes) {
+                        const SizeT copySize = std::min(filled, clearRowBytes - filled);
+                        Memcpy(row + filled, row, copySize);
+                        filled += copySize;
+                    }
+                }
+            }
+
+            textureObject->MarkStorageDirty(uploadTarget, static_cast<Uint>(level), true);
+            return true;
+        }
+    } // namespace
+
+    void ClearTexImage(GLuint texture, GLint level, GLenum format, GLenum type, const void* data) {
+        auto textureObject = GetClearTextureObject(texture, level, __func__);
+        if (!textureObject) return;
+
+        Vector<Uint8> clearPixel;
+        if (!BuildClearPixel(textureObject, format, type, data, clearPixel)) return;
+
+        for (TextureUploadTarget uploadTarget : textureObject->GetUploadTargets()) {
+            const IntVec3 size = textureObject->GetMipmapTexelSize(uploadTarget, static_cast<Uint>(level));
+            if (!ClearMipmapRegion(textureObject, uploadTarget, level, 0, 0, 0,
+                                   size.x(), size.y(), size.z(), clearPixel, __func__)) {
+                return;
+            }
+        }
+    }
+
+    void ClearTexSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset,
+                          GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type,
+                          const void* data) {
+        auto textureObject = GetClearTextureObject(texture, level, __func__);
+        if (!textureObject) return;
+
+        Vector<Uint8> clearPixel;
+        if (!BuildClearPixel(textureObject, format, type, data, clearPixel)) return;
+
+        const auto& uploadTargets = textureObject->GetUploadTargets();
+        if (textureObject->GetTarget() == TextureTarget::TextureCubeMap) {
+            if (zoffset < 0 || depth < 0 ||
+                static_cast<SizeT>(zoffset) > uploadTargets.size() ||
+                static_cast<SizeT>(depth) > uploadTargets.size() - static_cast<SizeT>(zoffset)) {
+                RecordClearTextureError(__func__, ErrorCode::InvalidValue,
+                                        "The cube-map clear region selects invalid faces.");
+                return;
+            }
+            for (GLsizei face = 0; face < depth; ++face) {
+                if (!ClearMipmapRegion(textureObject, uploadTargets[static_cast<SizeT>(zoffset + face)], level,
+                                       xoffset, yoffset, 0, width, height, 1, clearPixel, __func__)) {
+                    return;
+                }
+            }
+            return;
+        }
+
+        if (uploadTargets.empty()) {
+            RecordClearTextureError(__func__, ErrorCode::InvalidOperation,
+                                    "The requested texture has no upload target.");
+            return;
+        }
+        ClearMipmapRegion(textureObject, uploadTargets.front(), level, xoffset, yoffset, zoffset,
+                          width, height, depth, clearPixel, __func__);
+    }
+
     Bool ValidateTextureParameterForTarget(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
                                            GLenum pname, GLint param, const char* caller) {
         const auto target = textureObject->GetTarget();
@@ -3818,6 +4035,27 @@ namespace MobileGL::MG_Impl::GLImpl {
         CopyTexSubImage2D_Backend(target, level, xoffset, yoffset, x, y, width, height);
     }
 
+    void CopyTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y,
+                               GLsizei width, GLsizei height) {
+        auto textureObject = GetTextureObjectByName(texture, __func__);
+        if (!textureObject) return;
+        // GL 4.6 sec. 8.8: the 2D form only accepts these effective targets; cube maps must
+        // go through CopyTextureSubImage3D with the face as a layer.
+        const auto target = textureObject->GetTarget();
+        if (target != TextureTarget::Texture2D && target != TextureTarget::Texture1DArray &&
+            target != TextureTarget::TextureRectangle) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "CopyTextureSubImage2D requires a 2D, 1D-array, or "
+                                             "rectangle texture."));
+            return;
+        }
+        WithTemporarilyBoundNamedTexture(textureObject, [&](GLenum glTarget) {
+            CopyTexSubImage2D_Backend(glTarget, level, xoffset, yoffset, x, y, width, height);
+        });
+    }
+
     void CopyTexSubImage1D(GLenum target, GLint level, GLint xoffset, GLint x, GLint y, GLsizei width) {
         CopyTexSubImage1D_State(target, level, xoffset, x, y, width);
     }

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.h

@@ -11,6 +11,9 @@
 
 namespace MobileGL::MG_Impl::GLImpl {
     /* @INSERTION_POINT:FUNCTION_DECLARATION@ */
+    void ClearTexImage(GLuint texture, GLint level, GLenum format, GLenum type, const void* data);
+    void ClearTexSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
+                          GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* data);
     void BindImageTexture(GLuint unit, GLuint texture, GLint level, GLboolean layered, GLint layer, GLenum access,
                           GLenum format);
     void GenerateMipmap(GLenum target);
@@ -95,6 +98,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                            GLsizei width, GLsizei height);
     void CopyTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width,
                            GLsizei height);
+    void CopyTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y,
+                               GLsizei width, GLsizei height);
     void CopyTexSubImage1D(GLenum target, GLint level, GLint xoffset, GLint x, GLint y, GLsizei width);
     void CopyTexImage2D(GLenum target, GLint level, GLenum internalformat, GLint x, GLint y, GLsizei width,
                         GLsizei height, GLint border);

MobileGL/MG_Test/BackendLoader/BackendLoaderTest.cpp

@@ -27,6 +27,13 @@ namespace {
     struct FakeDriverState {
         // Behavior knobs, configured per test before running the probe.
         GLint maxVertexSsboBlocks = 4;
+        GLint glesMajorVersion = 3;
+        GLint glesMinorVersion = 1;
+        GLint maxVertexImageUniforms = 2;
+        GLint maxGeometryImageUniforms = 3;
+        GLint maxFragmentImageUniforms = 4;
+        GLint maxComputeImageUniforms = 5;
+        bool maxGeometryImageUniformsQueried = false;
         // Emulates ANGLE-on-Vulkan: the draw reads the indirect command's
         // baseInstance word and exposes it through gl_InstanceID.
         bool drawLeaksBaseInstanceWord = false;
@@ -88,13 +95,26 @@ namespace {
             case GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS:
                 *data = g_fake.maxVertexSsboBlocks;
                 break;
+            case GL_MAX_VERTEX_IMAGE_UNIFORMS:
+                *data = g_fake.maxVertexImageUniforms;
+                break;
+            case GL_MAX_GEOMETRY_IMAGE_UNIFORMS:
+                g_fake.maxGeometryImageUniformsQueried = true;
+                *data = g_fake.maxGeometryImageUniforms;
+                break;
+            case GL_MAX_FRAGMENT_IMAGE_UNIFORMS:
+                *data = g_fake.maxFragmentImageUniforms;
+                break;
+            case GL_MAX_COMPUTE_IMAGE_UNIFORMS:
+                *data = g_fake.maxComputeImageUniforms;
+                break;
             // FillInGLESCapabilities reads the context version before running the
             // baseInstance probe, which requires ES >= 3.1.
             case GL_MAJOR_VERSION:
-                *data = 3;
+                *data = g_fake.glesMajorVersion;
                 break;
             case GL_MINOR_VERSION:
-                *data = 1;
+                *data = g_fake.glesMinorVersion;
                 break;
             case GL_NUM_EXTENSIONS:
                 *data = static_cast<GLint>(g_fake.extensions.size());
@@ -417,6 +437,31 @@ TEST(IndirectInstanceIdProbe, FillInCapabilitiesWiresProbeResult) {
     ExpectProbeReleasedAllObjects();
 }
 
+TEST(ImageUniformCapabilities, QueriesRealPerStageLimitsAndConservativelyGatesGeometry) {
+    const auto funcs = MakeFakeGLESFunctions();
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    MobileGL::MG_External::GLESCapabilities es31Caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(es31Caps, funcs));
+    EXPECT_EQ(es31Caps.MaxVertexImageUniforms, g_fake.maxVertexImageUniforms);
+    EXPECT_EQ(es31Caps.MaxGeometryImageUniforms, 0);
+    EXPECT_EQ(es31Caps.MaxFragmentImageUniforms, g_fake.maxFragmentImageUniforms);
+    EXPECT_EQ(es31Caps.MaxComputeImageUniforms, g_fake.maxComputeImageUniforms);
+    EXPECT_FALSE(g_fake.maxGeometryImageUniformsQueried);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 2;
+    MobileGL::MG_External::GLESCapabilities es32Caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(es32Caps, funcs));
+    EXPECT_EQ(es32Caps.MaxVertexImageUniforms, g_fake.maxVertexImageUniforms);
+    EXPECT_EQ(es32Caps.MaxGeometryImageUniforms, g_fake.maxGeometryImageUniforms);
+    EXPECT_EQ(es32Caps.MaxFragmentImageUniforms, g_fake.maxFragmentImageUniforms);
+    EXPECT_EQ(es32Caps.MaxComputeImageUniforms, g_fake.maxComputeImageUniforms);
+    EXPECT_TRUE(g_fake.maxGeometryImageUniformsQueried);
+}
+
 // The extension string is what apps gate on (LWJGL builds GLCapabilities from it), so advertising
 // it on a driver that cannot filter anisotropically would leave them silently on trilinear.
 TEST(TextureAnisotropyCapabilities, ExtensionIsAdvertisedOnlyWhenTheHostDriverSupportsIt) {

MobileGL/MG_Test/Program/ProgramTest.cpp

@@ -7,6 +7,7 @@
 // End of Source File Header
 
 #include <gtest/gtest.h>
+#include <spirv_reflect.h>
 #include <cstring>
 #include <vector>
 
@@ -1325,6 +1326,88 @@ TEST_F(ProgramTest, CompileAndLinkWithExplicitVertexIn) {
                                        << "\")";
 }
 
+TEST_F(ProgramTest, InactiveExplicitVertexBindingsDoNotReserveLocations) {
+    const char* vertexSource = R"(#version 430 compatibility
+
+in vec3 Position;
+in vec2 UV0;
+in vec3 vaPosition;
+
+void main() {
+    gl_Position = vec4(vaPosition, 1.0);
+}
+)";
+    const char* fragmentSource = R"(#version 430 compatibility
+
+out vec4 fragColor;
+
+void main() {
+    fragColor = vec4(1.0);
+}
+)";
+
+    GLuint vertexShader = CreateShader(GL_VERTEX_SHADER);
+    ShaderSource(vertexShader, 1, &vertexSource, nullptr);
+    CompileShader(vertexShader);
+    GLint compileStatus = GL_FALSE;
+    GetShaderiv(vertexShader, GL_COMPILE_STATUS, &compileStatus);
+    ASSERT_EQ(compileStatus, GL_TRUE);
+
+    GLuint fragmentShader = CreateShader(GL_FRAGMENT_SHADER);
+    ShaderSource(fragmentShader, 1, &fragmentSource, nullptr);
+    CompileShader(fragmentShader);
+    GetShaderiv(fragmentShader, GL_COMPILE_STATUS, &compileStatus);
+    ASSERT_EQ(compileStatus, GL_TRUE);
+
+    GLuint program = CreateProgram();
+    AttachShader(program, vertexShader);
+    AttachShader(program, fragmentShader);
+
+    // Iris binds these canonical names before linking every program. Its compatibility
+    // transformer can inject both declarations even when the shader pack instead reads
+    // vaPosition. Inactive API bindings must not consume locations during the link.
+    BindAttribLocation(program, 0, "Position");
+    BindAttribLocation(program, 1, "UV0");
+    LinkProgram(program);
+
+    GLint linkStatus = GL_FALSE;
+    GetProgramiv(program, GL_LINK_STATUS, &linkStatus);
+    ASSERT_EQ(linkStatus, GL_TRUE);
+
+    EXPECT_EQ(GetAttribLocation(program, "Position"), -1);
+    EXPECT_EQ(GetAttribLocation(program, "UV0"), -1);
+    EXPECT_EQ(GetAttribLocation(program, "vaPosition"), 0);
+
+    auto programObject = MG_State::pGLContext->GetProgramObject(program);
+    ASSERT_NE(programObject, nullptr);
+    const Int vertexIndex = programObject->GetShaderIndexByStage(ShaderStage::Vertex);
+    ASSERT_GE(vertexIndex, 0);
+    const auto& spirvs = programObject->GetGeneratedSpirv();
+    ASSERT_LT(static_cast<SizeT>(vertexIndex), spirvs.size());
+
+    const auto& vertexSpirv = spirvs[vertexIndex];
+    spv_reflect::ShaderModule reflection(vertexSpirv.size() * sizeof(Uint), vertexSpirv.data());
+    ASSERT_EQ(reflection.GetResult(), SPV_REFLECT_RESULT_SUCCESS);
+
+    uint32_t inputCount = 0;
+    ASSERT_EQ(reflection.EnumerateInputVariables(&inputCount, nullptr), SPV_REFLECT_RESULT_SUCCESS);
+    Vector<SpvReflectInterfaceVariable*> inputs(inputCount);
+    ASSERT_EQ(reflection.EnumerateInputVariables(&inputCount, inputs.data()), SPV_REFLECT_RESULT_SUCCESS);
+
+    Uint32 userInputCount = 0;
+    Uint32 locationMask = 0;
+    for (const auto* input : inputs) {
+        if (input == nullptr || (input->decoration_flags & SPV_REFLECT_DECORATION_BUILT_IN) != 0) {
+            continue;
+        }
+        ASSERT_LT(input->location, 32u);
+        locationMask |= 1u << input->location;
+        ++userInputCount;
+    }
+    EXPECT_EQ(userInputCount, 1u);
+    EXPECT_EQ(locationMask, 0x1u);
+}
+
 TEST_F(ProgramTest, CompileAndLinkWithExplicitFragmentOut) {
     char infoLog[1024] = "";
 

MobileGL/MG_Test/Program/ProgramUtilTest.cpp

@@ -8,7 +8,9 @@
 
 #include <gtest/gtest.h>
 
+#include <cstring>
 #include <string>
+#include <utility>
 
 #include "Includes.h"
 #include "Init.h"
@@ -92,6 +94,120 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
     EXPECT_EQ(exactSamplerNameCount, 1u);
 }
 
+TEST_F(ProgramUtilTest, UnformattedFloatStorageImagesKeepIntegerAtomicImagesTyped) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const String source = R"(#version 430 core
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(rgba16, binding = 0) uniform image2D floatImage;
+layout(r32ui, binding = 1) uniform uimage2D atomicImage;
+
+void main() {
+    ivec2 coordinate = ivec2(gl_GlobalInvocationID.xy);
+    imageStore(floatImage, coordinate, imageLoad(floatImage, coordinate));
+    imageAtomicAdd(atomicImage, coordinate, 1u);
+}
+)";
+
+    ShaderAttrib shaderAttrib{.shaderType = GL_COMPUTE_SHADER, .sourceStr = source};
+    auto shaderResult = ShaderCompiler::CompileShader(shaderAttrib);
+    ASSERT_TRUE(shaderResult) << shaderResult.error().log;
+
+    ProgramAttrib programAttrib{.shaders = {shaderResult.value()}};
+    auto programResult = ShaderCompiler::LinkProgram(programAttrib);
+    ASSERT_TRUE(programResult) << programResult.error().log;
+
+    ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_COMPUTE_SHADER}, .program = *programResult.value()};
+    auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+    ASSERT_TRUE(binaryResult) << binaryResult.error().log;
+    ASSERT_EQ(binaryResult->size(), 1u);
+    const auto& inputBinary = binaryResult->front();
+
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String inputText;
+    ASSERT_TRUE(tools.Disassemble(inputBinary, &inputText));
+    EXPECT_NE(inputText.find("2D 0 0 0 2 Rgba16"), String::npos) << inputText;
+    EXPECT_NE(inputText.find("2D 0 0 0 2 R32ui"), String::npos) << inputText;
+    EXPECT_EQ(inputText.find("StorageImageReadWithoutFormat"), String::npos) << inputText;
+    EXPECT_EQ(inputText.find("StorageImageWriteWithoutFormat"), String::npos) << inputText;
+
+    Vector<Uint32> outputBinary;
+    ASSERT_TRUE(ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(inputBinary, outputBinary));
+
+    String outputText;
+    ASSERT_TRUE(tools.Disassemble(outputBinary, &outputText));
+    EXPECT_EQ(outputText.find("2D 0 0 0 2 Rgba16"), String::npos) << outputText;
+    EXPECT_NE(outputText.find("2D 0 0 0 2 Unknown"), String::npos) << outputText;
+    EXPECT_NE(outputText.find("2D 0 0 0 2 R32ui"), String::npos) << outputText;
+
+    const auto countOccurrences = [](const String& text, const String& needle) {
+        SizeT count = 0;
+        for (SizeT offset = 0; (offset = text.find(needle, offset)) != String::npos;
+             offset += needle.size()) {
+            ++count;
+        }
+        return count;
+    };
+    EXPECT_EQ(countOccurrences(outputText, "OpCapability StorageImageReadWithoutFormat"), 1u)
+        << outputText;
+    EXPECT_EQ(countOccurrences(outputText, "OpCapability StorageImageWriteWithoutFormat"), 1u)
+        << outputText;
+    EXPECT_TRUE(tools.Validate(outputBinary));
+
+    Vector<Uint32> secondOutputBinary;
+    ASSERT_TRUE(ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(outputBinary, secondOutputBinary));
+    EXPECT_EQ(secondOutputBinary, outputBinary);
+}
+
+TEST_F(ProgramUtilTest, UnformattedFloatStorageImagesKeepFloatAtomicImageTypesTyped) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const String spirvText = R"(
+               OpCapability Shader
+               OpCapability StorageImageExtendedFormats
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint GLCompute %main "main"
+               OpExecutionMode %main LocalSize 1 1 1
+               OpDecorate %target DescriptorSet 0
+               OpDecorate %target Binding 0
+       %void = OpTypeVoid
+      %float = OpTypeFloat 32
+        %int = OpTypeInt 32 1
+      %v2int = OpTypeVector %int 2
+      %image = OpTypeImage %float 2D 0 0 0 2 R32f
+%imageUniformPtr = OpTypePointer UniformConstant %image
+%imageTexelPtr = OpTypePointer Image %float
+ %mainType = OpTypeFunction %void
+       %zero = OpConstant %int 0
+ %coordinate = OpConstantComposite %v2int %zero %zero
+     %target = OpVariable %imageUniformPtr UniformConstant
+       %main = OpFunction %void None %mainType
+      %entry = OpLabel
+   %texelPtr = OpImageTexelPointer %imageTexelPtr %target %coordinate %zero
+               OpReturn
+               OpFunctionEnd
+)";
+
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    Vector<Uint32> inputBinary;
+    ASSERT_TRUE(tools.Assemble(spirvText, &inputBinary));
+
+    Vector<Uint32> outputBinary;
+    ASSERT_TRUE(ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(inputBinary, outputBinary));
+
+    String outputText;
+    ASSERT_TRUE(tools.Disassemble(outputBinary, &outputText));
+    EXPECT_NE(outputText.find("2D 0 0 0 2 R32f"), String::npos) << outputText;
+    EXPECT_EQ(outputText.find("StorageImageReadWithoutFormat"), String::npos) << outputText;
+    EXPECT_EQ(outputText.find("StorageImageWriteWithoutFormat"), String::npos) << outputText;
+    String validationDiagnostics;
+    tools.SetMessageConsumer([&validationDiagnostics](spv_message_level_t, const char*,
+                                                       const spv_position_t&, const char* message) {
+        validationDiagnostics += message;
+    });
+    EXPECT_TRUE(tools.Validate(outputBinary)) << validationDiagnostics;
+}
+
 TEST_F(ProgramUtilTest, PreprocessLegacyVertexShaderModernizesGlmarkStyleSource) {
     using namespace MG_Util::ShaderTranspiler;
 
@@ -1383,3 +1499,150 @@ void main() {
     EXPECT_NE(source.find("shared float sharedScratch[8];"), String::npos);
     EXPECT_NE(source.find("layout(std140) uniform Blk"), String::npos);
 }
+
+namespace {
+    String MakeLinearSubgroupPrefixScanShader() {
+        return R"(#version 460 core
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+layout(local_size_x = 1024) in;
+shared float prefixSumCache[64];
+
+layout(std430, binding = 0) writeonly buffer OutputBuffer {
+    float outputValues[];
+};
+
+void main() {
+    float importance = 1.0f;
+    float prefixSum = subgroupInclusiveAdd(importance);
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u) prefixSumCache[gl_SubgroupID] = prefixSum;
+    barrier();
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+    for (uint i = 0; i < loopLength; i++) {
+        if ((gl_SubgroupID & (1u << i)) > 0u) {
+            prefixSum += prefixSumCache[(gl_SubgroupID >> i << i) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u) prefixSumCache[gl_SubgroupID] = prefixSum;
+        }
+        barrier();
+    }
+    if (gl_LocalInvocationID.x == uint(1024 - 1)) prefixSumCache[0] = prefixSum;
+    barrier();
+    float sum = prefixSumCache[0];
+    float warp = (prefixSum - importance) / sum - float(gl_LocalInvocationID.x + 1u) / float(1024);
+    outputValues[gl_GlobalInvocationID.x] = warp;
+}
+)";
+    }
+} // namespace
+
+TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanUsesSharedMemoryAndProducesValidSpirv) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = MakeLinearSubgroupPrefixScanShader();
+    ASSERT_TRUE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
+
+    EXPECT_NE(source.find("shared float prefixSumCache[1024]"), String::npos) << source;
+    EXPECT_NE(source.find("mglVirtualSubgroupInvocation"), String::npos) << source;
+    EXPECT_NE(source.find("for (uint mglPrefixLane"), String::npos) << source;
+    EXPECT_EQ(source.find("subgroupInclusiveAdd"), String::npos) << source;
+    EXPECT_EQ(source.find("gl_Subgroup"), String::npos) << source;
+
+    const String onceRewritten = source;
+    EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
+    EXPECT_EQ(source, onceRewritten);
+
+    ShaderAttrib shaderAttrib{.shaderType = GL_COMPUTE_SHADER, .sourceStr = source};
+    auto shaderResult = ShaderCompiler::CompileShader(shaderAttrib);
+    ASSERT_TRUE(shaderResult) << shaderResult.error().log << "\nsource:\n" << source;
+
+    ProgramAttrib programAttrib{.shaders = {shaderResult.value()}};
+    auto programResult = ShaderCompiler::LinkProgram(programAttrib);
+    ASSERT_TRUE(programResult) << programResult.error().log;
+
+    ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_COMPUTE_SHADER}, .program = *programResult.value()};
+    auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+    ASSERT_TRUE(binaryResult) << binaryResult.error().log;
+    ASSERT_EQ(binaryResult->size(), 1u);
+
+    String validationDiagnostics;
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    tools.SetMessageConsumer([&](spv_message_level_t, const char*, const spv_position_t&, const char* message) {
+        validationDiagnostics += message;
+        validationDiagnostics += '\n';
+    });
+    EXPECT_TRUE(tools.Validate(binaryResult->front())) << validationDiagnostics;
+
+    String spirvText;
+    ASSERT_TRUE(tools.Disassemble(binaryResult->front(), &spirvText));
+    EXPECT_EQ(spirvText.find("OpGroupNonUniform"), String::npos) << spirvText;
+}
+
+TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsOtherStagesAndSubgroupWidths) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const String original = MakeLinearSubgroupPrefixScanShader();
+    for (const auto& [stage, subgroupSize] :
+         {std::pair{ShaderStage::Compute, Uint32{32}}, std::pair{ShaderStage::Fragment, Uint32{64}},
+          std::pair{ShaderStage::Compute, Uint32{96}}}) {
+        String source = original;
+        EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(stage, subgroupSize, source));
+        EXPECT_EQ(source, original);
+    }
+}
+
+TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsPartialOrUnsafeTemplateMatches) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const auto expectUnchanged = [](String source) {
+        const String original = source;
+        EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
+        EXPECT_EQ(source, original);
+    };
+
+    String wrongLocalSize = MakeLinearSubgroupPrefixScanShader();
+    wrongLocalSize.replace(wrongLocalSize.find("local_size_x = 1024"), std::strlen("local_size_x = 1024"),
+                           "local_size_x = 512");
+    expectUnchanged(std::move(wrongLocalSize));
+
+    String cacheHasAnotherUse = MakeLinearSubgroupPrefixScanShader();
+    cacheHasAnotherUse.insert(cacheHasAnotherUse.find("float importance"), "prefixSumCache[0] = 0.0f;\n    ");
+    expectUnchanged(std::move(cacheHasAnotherUse));
+
+    String extraSubgroupBuiltin = MakeLinearSubgroupPrefixScanShader();
+    extraSubgroupBuiltin.insert(extraSubgroupBuiltin.find("float importance"),
+                                "uvec4 extraMask = gl_SubgroupEqMask;\n    ");
+    expectUnchanged(std::move(extraSubgroupBuiltin));
+
+    String alteredBarrier = MakeLinearSubgroupPrefixScanShader();
+    alteredBarrier.replace(alteredBarrier.find("barrier();"), std::strlen("barrier();"), "memoryBarrierShared();");
+    expectUnchanged(std::move(alteredBarrier));
+
+    String nestedScan = MakeLinearSubgroupPrefixScanShader();
+    nestedScan.insert(nestedScan.find("float prefixSum ="), "if (importance > 0.0f) {\n    ");
+    const SizeT consumerEnd = nestedScan.find(';', nestedScan.find("float warp ="));
+    ASSERT_NE(consumerEnd, String::npos);
+    nestedScan.insert(consumerEnd + 1, "\n    }");
+    expectUnchanged(std::move(nestedScan));
+
+    // ARB/NV spellings of lane-width-sensitive builtins must block the rewrite exactly
+    // like their KHR counterparts.
+    String arbSubgroupBuiltin = MakeLinearSubgroupPrefixScanShader();
+    arbSubgroupBuiltin.insert(arbSubgroupBuiltin.find("float importance"),
+                              "uint arbLane = gl_SubGroupInvocationARB;\n    ");
+    expectUnchanged(std::move(arbSubgroupBuiltin));
+
+    String arbBallotCall = MakeLinearSubgroupPrefixScanShader();
+    arbBallotCall.insert(arbBallotCall.find("float importance"),
+                         "uint64_t arbMask = ballotARB(true);\n    ");
+    expectUnchanged(std::move(arbBallotCall));
+
+    String nvWarpBuiltin = MakeLinearSubgroupPrefixScanShader();
+    nvWarpBuiltin.insert(nvWarpBuiltin.find("float importance"),
+                         "uint warpSize = gl_WarpSizeNV;\n    ");
+    expectUnchanged(std::move(nvWarpBuiltin));
+
+    String nvShuffleCall = MakeLinearSubgroupPrefixScanShader();
+    nvShuffleCall.insert(nvShuffleCall.find("float importance"),
+                         "float other = shuffleNV(1.0f, 0u, 32u);\n    ");
+    expectUnchanged(std::move(nvShuffleCall));
+}

MobileGL/MG_Test/SanityTest.cpp

@@ -24,9 +24,12 @@
 #include <MG_State/GLState/Core.h>
 #include <MG_State/GLState/FramebufferState/FramebufferObject.h>
 #include <MG_State/GLState/TextureState/TextureState.h>
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+#include <MG_Backend/DirectVulkan/Renderer/UniformManager.h>
 #include <MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h>
 #include <MG_Backend/DirectVulkan/Renderer/VkTextureManager.h>
 #include <MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h>
+#include <MG_Util/Math/HalfFloat.h>
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 #include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
 #include <MG_Util/Debug/Log.h>
@@ -460,6 +463,58 @@ TEST(DirectVulkanSanity, ClampsAdvertisedTextureAndDrawBufferLimitsToFrontendSta
     EXPECT_EQ(lowParams.MaxColorAttachments, 6);
 }
 
+TEST(DirectVulkanSanity, GatesPerStageImageUniformLimitsOnPhysicalDeviceFeatures) {
+    using namespace MobileGL;
+
+    MG_Backend::DirectVulkan::BackendObject_DirectVulkan backend;
+    MG_External::VulkanCapabilities caps;
+    caps.MaxImageUnits = 12;
+    caps.MaxCombinedImageUniforms = 10;
+    caps.MaxComputeImageUniforms = 9;
+    caps.SupportsVertexPipelineStoresAndAtomics = true;
+    caps.SupportsFragmentStoresAndAtomics = true;
+    caps.SupportsGeometryShader = false;
+    backend.ApplyVulkanCapabilitiesForTesting(caps);
+
+    const auto& withoutGeometry = backend.GetDynamicParameters();
+    EXPECT_EQ(withoutGeometry.MaxVertexImageUniforms, 10);
+    EXPECT_EQ(withoutGeometry.MaxGeometryImageUniforms, 0);
+    EXPECT_EQ(withoutGeometry.MaxFragmentImageUniforms, 10);
+    EXPECT_EQ(withoutGeometry.MaxComputeImageUniforms, 9);
+
+    caps.SupportsGeometryShader = true;
+    backend.ApplyVulkanCapabilitiesForTesting(caps);
+    EXPECT_EQ(backend.GetDynamicParameters().MaxGeometryImageUniforms, 10);
+
+    caps.SupportsVertexPipelineStoresAndAtomics = false;
+    caps.SupportsFragmentStoresAndAtomics = false;
+    backend.ApplyVulkanCapabilitiesForTesting(caps);
+    EXPECT_EQ(backend.GetDynamicParameters().MaxVertexImageUniforms, 0);
+    EXPECT_EQ(backend.GetDynamicParameters().MaxGeometryImageUniforms, 0);
+    EXPECT_EQ(backend.GetDynamicParameters().MaxFragmentImageUniforms, 0);
+    EXPECT_EQ(backend.GetDynamicParameters().MaxComputeImageUniforms, 9);
+}
+
+TEST(DirectGLESSanity, PreservesHostPerStageImageUniformLimits) {
+    using namespace MobileGL;
+
+    MG_Backend::DirectGLES::BackendObject_DirectGLES backend;
+    MG_External::GLESCapabilities caps;
+    caps.MaxImageUnits = 8;
+    caps.MaxCombinedImageUniforms = 16;
+    caps.MaxVertexImageUniforms = 2;
+    caps.MaxGeometryImageUniforms = 3;
+    caps.MaxFragmentImageUniforms = 4;
+    caps.MaxComputeImageUniforms = 5;
+    backend.ApplyGLESCapabilitiesForTesting(caps);
+
+    const auto& params = backend.GetDynamicParameters();
+    EXPECT_EQ(params.MaxVertexImageUniforms, 2);
+    EXPECT_EQ(params.MaxGeometryImageUniforms, 3);
+    EXPECT_EQ(params.MaxFragmentImageUniforms, 4);
+    EXPECT_EQ(params.MaxComputeImageUniforms, 5);
+}
+
 TEST(DirectVulkanSanity, AdvertisesSubgroupOnlyWhenVulkanReportsUsableSupport) {
     using namespace MobileGL;
 
@@ -582,6 +637,54 @@ TEST(GetterSanity, ClampsMaxVertexAttribsToCurrentValueStorageCapacity) {
     MG_State::pGLContext.reset();
 }
 
+TEST(GetterSanity, PerStageImageUniformQueriesMatchShaderCompilerLimits) {
+    using namespace MobileGL;
+
+    MG_Backend::DynamicBackendParameters params;
+    params.MaxImageUnits = 8;
+    params.MaxCombinedImageUniforms = 8;
+    params.MaxVertexImageUniforms = 1;
+    params.MaxGeometryImageUniforms = 2;
+    params.MaxFragmentImageUniforms = 3;
+    params.MaxComputeImageUniforms = 4;
+    MG_Backend::pActiveBackendObject = MakeUnique<DynamicParameterBackend>(params);
+
+    GLint reported = -1;
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_VERTEX_IMAGE_UNIFORMS, &reported);
+    EXPECT_EQ(reported, 1);
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_GEOMETRY_IMAGE_UNIFORMS, &reported);
+    EXPECT_EQ(reported, 2);
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_FRAGMENT_IMAGE_UNIFORMS, &reported);
+    EXPECT_EQ(reported, 3);
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &reported);
+    EXPECT_EQ(reported, 4);
+
+    const String vertexImageStore = R"(#version 430 core
+layout(r32ui, binding = 0) uniform uimage2D targetImages[gl_MaxVertexImageUniforms];
+void main() {
+    imageStore(targetImages[0], ivec2(0), uvec4(1));
+    gl_Position = vec4(0.0, 0.0, 0.0, 1.0);
+}
+)";
+    auto supported = MG_Util::ShaderTranspiler::ShaderCompiler::CompileShader({
+        .shaderType = GL_VERTEX_SHADER,
+        .sourceStr = vertexImageStore,
+    });
+    EXPECT_TRUE(supported) << (supported ? "" : supported.error().log);
+
+    params.MaxVertexImageUniforms = 0;
+    MG_Backend::pActiveBackendObject = MakeUnique<DynamicParameterBackend>(params);
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_VERTEX_IMAGE_UNIFORMS, &reported);
+    EXPECT_EQ(reported, 0);
+    auto unsupported = MG_Util::ShaderTranspiler::ShaderCompiler::CompileShader({
+        .shaderType = GL_VERTEX_SHADER,
+        .sourceStr = vertexImageStore,
+    });
+    EXPECT_FALSE(unsupported);
+
+    MG_Backend::pActiveBackendObject.reset();
+}
+
 TEST(GetterSanity, ReportsKhrSubgroupDynamicParameters) {
     using namespace MobileGL;
 
@@ -629,6 +732,95 @@ TEST(DirectVulkanSanity, CommandMemoryBarrierMakesIndirectDrawCommandsVisible) {
     EXPECT_EQ(storageOnlyBarrier.dstAccessMask & VK_ACCESS_INDIRECT_COMMAND_READ_BIT, 0u);
 }
 
+TEST(DirectVulkanSanity, ReadbackUsesTheSourceFormatTexelSize) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+
+    EXPECT_EQ(VulkanRenderer::GetReadbackTexelSize(VK_FORMAT_R8G8B8A8_UNORM), 4u);
+    EXPECT_EQ(VulkanRenderer::GetReadbackTexelSize(VK_FORMAT_R16G16B16A16_SFLOAT), 8u);
+    EXPECT_EQ(VulkanRenderer::GetReadbackTexelSize(VK_FORMAT_R32G32B32A32_SFLOAT), 16u);
+}
+
+TEST(DirectVulkanSanity, ReadbackConvertsRgba8AndRgba16fPixels) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+    using MobileGL::MG_Util::EncodeFloatToHalfBits;
+
+    const MobileGL::Uint8 rgba8[] = {17, 34, 51, 68, 85, 102, 119, 136};
+    MobileGL::Uint8 rgba8Result[sizeof(rgba8)]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        rgba8, VK_FORMAT_R8G8B8A8_UNORM, 2, 1, GL_RGBA, GL_UNSIGNED_BYTE,
+        sizeof(rgba8Result), rgba8Result));
+    EXPECT_TRUE(std::equal(std::begin(rgba8), std::end(rgba8), std::begin(rgba8Result)));
+
+    const MobileGL::Uint8 bgra8[] = {51, 34, 17, 68};
+    MobileGL::Uint8 bgra8Result[4]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        bgra8, VK_FORMAT_B8G8R8A8_UNORM, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE,
+        sizeof(bgra8Result), bgra8Result));
+    const MobileGL::Uint8 expectedBgra8[] = {17, 34, 51, 68};
+    EXPECT_TRUE(std::equal(std::begin(expectedBgra8), std::end(expectedBgra8), std::begin(bgra8Result)));
+
+    const MobileGL::Uint16 rgba16f[] = {
+        EncodeFloatToHalfBits(-0.25f), EncodeFloatToHalfBits(0.5f), EncodeFloatToHalfBits(1.5f),
+        EncodeFloatToHalfBits(1.0f), EncodeFloatToHalfBits(0.25f), EncodeFloatToHalfBits(0.0f),
+        EncodeFloatToHalfBits(1.0f), EncodeFloatToHalfBits(0.5f),
+        EncodeFloatToHalfBits(0.75f), EncodeFloatToHalfBits(0.125f), EncodeFloatToHalfBits(-1.0f),
+        EncodeFloatToHalfBits(2.0f), EncodeFloatToHalfBits(1.0f), EncodeFloatToHalfBits(0.75f),
+        EncodeFloatToHalfBits(0.25f), EncodeFloatToHalfBits(0.0f),
+    };
+    constexpr MobileGL::SizeT kDestinationRowStride = 12;
+    MobileGL::Uint8 rgba16fResult[kDestinationRowStride * 2];
+    std::fill(std::begin(rgba16fResult), std::end(rgba16fResult), 0xCD);
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(rgba16f), VK_FORMAT_R16G16B16A16_SFLOAT,
+        2, 2, GL_RGBA, GL_UNSIGNED_BYTE, kDestinationRowStride, rgba16fResult));
+    const MobileGL::Uint8 expectedRgba16fRow0[] = {0, 128, 255, 255, 64, 0, 255, 128};
+    const MobileGL::Uint8 expectedRgba16fRow1[] = {191, 32, 0, 255, 255, 191, 64, 0};
+    EXPECT_TRUE(std::equal(std::begin(expectedRgba16fRow0), std::end(expectedRgba16fRow0),
+                           std::begin(rgba16fResult)));
+    EXPECT_TRUE(std::equal(std::begin(expectedRgba16fRow1), std::end(expectedRgba16fRow1),
+                           std::begin(rgba16fResult) + kDestinationRowStride));
+    EXPECT_TRUE(std::all_of(std::begin(rgba16fResult) + 8,
+                            std::begin(rgba16fResult) + kDestinationRowStride,
+                            [](MobileGL::Uint8 value) { return value == 0xCD; }));
+
+    MobileGL::Float rgba16fFloatResult[16]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(rgba16f), VK_FORMAT_R16G16B16A16_SFLOAT,
+        2, 2, GL_RGBA, GL_FLOAT, sizeof(MobileGL::Float) * 8,
+        reinterpret_cast<MobileGL::Uint8*>(rgba16fFloatResult)));
+    EXPECT_FLOAT_EQ(rgba16fFloatResult[0], -0.25f);
+    EXPECT_FLOAT_EQ(rgba16fFloatResult[1], 0.5f);
+    EXPECT_FLOAT_EQ(rgba16fFloatResult[2], 1.5f);
+    EXPECT_FLOAT_EQ(rgba16fFloatResult[3], 1.0f);
+}
+
+TEST(DirectVulkanSanity, ReadbackDecodesSingleChannel32BitFormats) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+
+    // The reinterpretation feature makes R32F/R32UI-class images common readback sources
+    // (iterationRP custom images). Missing channels take GL defaults: 0 for GB, 1 for alpha.
+    const MobileGL::Float r32f[] = {0.75f, -2.0f};
+    MobileGL::Float r32fResult[8]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(r32f), VK_FORMAT_R32_SFLOAT,
+        2, 1, GL_RGBA, GL_FLOAT, sizeof(MobileGL::Float) * 8,
+        reinterpret_cast<MobileGL::Uint8*>(r32fResult)));
+    EXPECT_FLOAT_EQ(r32fResult[0], 0.75f);
+    EXPECT_FLOAT_EQ(r32fResult[1], 0.0f);
+    EXPECT_FLOAT_EQ(r32fResult[2], 0.0f);
+    EXPECT_FLOAT_EQ(r32fResult[3], 1.0f);
+    EXPECT_FLOAT_EQ(r32fResult[4], -2.0f);
+
+    const MobileGL::Uint32 r32ui[] = {12345u};
+    MobileGL::Float r32uiResult[4]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(r32ui), VK_FORMAT_R32_UINT,
+        1, 1, GL_RGBA, GL_FLOAT, sizeof(MobileGL::Float) * 4,
+        reinterpret_cast<MobileGL::Uint8*>(r32uiResult)));
+    EXPECT_FLOAT_EQ(r32uiResult[0], 12345.0f);
+    EXPECT_FLOAT_EQ(r32uiResult[3], 1.0f);
+}
+
 TEST(DirectVulkanSanity, DrawIndexedIndirectCommandMatchesGlAndVulkanLayout) {
     using namespace MobileGL::MG_Backend::DirectVulkan;
 
@@ -675,6 +867,193 @@ TEST(DirectVulkanSanity, SampledDepthStencilViewUsesSingleDepthAspect) {
               VK_IMAGE_ASPECT_DEPTH_BIT);
 }
 
+TEST(DirectVulkanSanity, SpirvStorageImageFormatsMapToVulkanFormats) {
+    using MobileGL::MG_Backend::DirectVulkan::ProgramFactory;
+
+    struct FormatCase {
+        SpvImageFormat spirv;
+        VkFormat vulkan;
+    };
+    const FormatCase cases[] = {
+        {SpvImageFormatUnknown, VK_FORMAT_UNDEFINED},
+        {SpvImageFormatRgba32f, VK_FORMAT_R32G32B32A32_SFLOAT},
+        {SpvImageFormatRgba16f, VK_FORMAT_R16G16B16A16_SFLOAT},
+        {SpvImageFormatR32f, VK_FORMAT_R32_SFLOAT},
+        {SpvImageFormatRgba8, VK_FORMAT_R8G8B8A8_UNORM},
+        {SpvImageFormatRgba8Snorm, VK_FORMAT_R8G8B8A8_SNORM},
+        {SpvImageFormatRg32f, VK_FORMAT_R32G32_SFLOAT},
+        {SpvImageFormatRg16f, VK_FORMAT_R16G16_SFLOAT},
+        {SpvImageFormatR11fG11fB10f, VK_FORMAT_B10G11R11_UFLOAT_PACK32},
+        {SpvImageFormatR16f, VK_FORMAT_R16_SFLOAT},
+        {SpvImageFormatRgba16, VK_FORMAT_R16G16B16A16_UNORM},
+        {SpvImageFormatRgb10A2, VK_FORMAT_A2R10G10B10_UNORM_PACK32},
+        {SpvImageFormatRg16, VK_FORMAT_R16G16_UNORM},
+        {SpvImageFormatRg8, VK_FORMAT_R8G8_UNORM},
+        {SpvImageFormatR16, VK_FORMAT_R16_UNORM},
+        {SpvImageFormatR8, VK_FORMAT_R8_UNORM},
+        {SpvImageFormatRgba16Snorm, VK_FORMAT_R16G16B16A16_SNORM},
+        {SpvImageFormatRg16Snorm, VK_FORMAT_R16G16_SNORM},
+        {SpvImageFormatRg8Snorm, VK_FORMAT_R8G8_SNORM},
+        {SpvImageFormatR16Snorm, VK_FORMAT_R16_SNORM},
+        {SpvImageFormatR8Snorm, VK_FORMAT_R8_SNORM},
+        {SpvImageFormatRgba32i, VK_FORMAT_R32G32B32A32_SINT},
+        {SpvImageFormatRgba16i, VK_FORMAT_R16G16B16A16_SINT},
+        {SpvImageFormatRgba8i, VK_FORMAT_R8G8B8A8_SINT},
+        {SpvImageFormatR32i, VK_FORMAT_R32_SINT},
+        {SpvImageFormatRg32i, VK_FORMAT_R32G32_SINT},
+        {SpvImageFormatRg16i, VK_FORMAT_R16G16_SINT},
+        {SpvImageFormatRg8i, VK_FORMAT_R8G8_SINT},
+        {SpvImageFormatR16i, VK_FORMAT_R16_SINT},
+        {SpvImageFormatR8i, VK_FORMAT_R8_SINT},
+        {SpvImageFormatRgba32ui, VK_FORMAT_R32G32B32A32_UINT},
+        {SpvImageFormatRgba16ui, VK_FORMAT_R16G16B16A16_UINT},
+        {SpvImageFormatRgba8ui, VK_FORMAT_R8G8B8A8_UINT},
+        {SpvImageFormatR32ui, VK_FORMAT_R32_UINT},
+        {SpvImageFormatRgb10a2ui, VK_FORMAT_A2R10G10B10_UINT_PACK32},
+        {SpvImageFormatRg32ui, VK_FORMAT_R32G32_UINT},
+        {SpvImageFormatRg16ui, VK_FORMAT_R16G16_UINT},
+        {SpvImageFormatRg8ui, VK_FORMAT_R8G8_UINT},
+        {SpvImageFormatR16ui, VK_FORMAT_R16_UINT},
+        {SpvImageFormatR8ui, VK_FORMAT_R8_UINT},
+        {SpvImageFormatR64ui, VK_FORMAT_R64_UINT},
+        {SpvImageFormatR64i, VK_FORMAT_R64_SINT},
+    };
+
+    for (const auto& testCase : cases) {
+        EXPECT_EQ(ProgramFactory::ConvertSpirvImageFormatToVkFormat(testCase.spirv), testCase.vulkan)
+            << "SpvImageFormat=" << static_cast<int>(testCase.spirv);
+    }
+}
+
+TEST(DirectVulkanSanity, MutableStorageImageViewsUseVulkanCompatibilityClasses) {
+    using MobileGL::MG_Backend::DirectVulkan::VkTextureManager;
+
+    EXPECT_TRUE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_UINT));
+    EXPECT_TRUE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_UINT, VK_FORMAT_R32_SINT));
+    EXPECT_TRUE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R16G16B16A16_UNORM, VK_FORMAT_R16G16B16A16_SFLOAT));
+    EXPECT_TRUE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R8G8B8A8_UINT));
+    EXPECT_TRUE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT));
+    EXPECT_FALSE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R16G16B16A16_SFLOAT));
+    EXPECT_FALSE(VkTextureManager::AreStorageImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_D32_SFLOAT));
+}
+
+TEST(DirectVulkanSanity, StorageImageViewFormatUsesBindingOnlyForFormatlessFloatPolicy) {
+    using MobileGL::MG_Backend::DirectVulkan::UniformManager;
+
+    EXPECT_EQ(UniformManager::ResolveStorageImageViewFormat(
+                  VK_FORMAT_UNDEFINED, GL_RGBA16F, VK_FORMAT_R16G16B16A16_UNORM, true),
+              VK_FORMAT_R16G16B16A16_SFLOAT);
+    EXPECT_EQ(UniformManager::ResolveStorageImageViewFormat(
+                  VK_FORMAT_UNDEFINED, GL_RGBA16, VK_FORMAT_R16G16B16A16_SFLOAT, true),
+              VK_FORMAT_R16G16B16A16_UNORM);
+    EXPECT_EQ(UniformManager::ResolveStorageImageViewFormat(
+                  VK_FORMAT_R32_UINT, GL_RGBA16F, VK_FORMAT_R32_SFLOAT, false),
+              VK_FORMAT_R32_UINT);
+    EXPECT_EQ(UniformManager::ResolveStorageImageViewFormat(
+                  VK_FORMAT_UNDEFINED, GL_RGBA16F, VK_FORMAT_R32_SFLOAT, false),
+              VK_FORMAT_R32_SFLOAT);
+    EXPECT_EQ(UniformManager::ResolveStorageImageViewFormat(
+                  VK_FORMAT_UNDEFINED, GL_NONE, VK_FORMAT_R16G16B16A16_SFLOAT, true),
+              VK_FORMAT_UNDEFINED);
+}
+
+TEST(DirectVulkanSanity, ProgramObjectMovePreservesStorageImageFormatPolicy) {
+    using MobileGL::MG_Backend::DirectVulkan::ProgramFactory;
+
+    ProgramFactory::VkProgramObject source;
+    source.storageImageFormatByBinding = {VK_FORMAT_UNDEFINED, VK_FORMAT_R32_UINT};
+    source.storageImageUsesBindingFormatByBinding = {true, false};
+
+    ProgramFactory::VkProgramObject moved(std::move(source));
+    ASSERT_EQ(moved.storageImageFormatByBinding.size(), 2u);
+    ASSERT_EQ(moved.storageImageUsesBindingFormatByBinding.size(), 2u);
+    EXPECT_EQ(moved.storageImageFormatByBinding[0], VK_FORMAT_UNDEFINED);
+    EXPECT_EQ(moved.storageImageFormatByBinding[1], VK_FORMAT_R32_UINT);
+    EXPECT_TRUE(moved.storageImageUsesBindingFormatByBinding[0]);
+    EXPECT_FALSE(moved.storageImageUsesBindingFormatByBinding[1]);
+
+    ProgramFactory::VkProgramObject assigned;
+    assigned = std::move(moved);
+    ASSERT_EQ(assigned.storageImageFormatByBinding.size(), 2u);
+    ASSERT_EQ(assigned.storageImageUsesBindingFormatByBinding.size(), 2u);
+    EXPECT_TRUE(assigned.storageImageUsesBindingFormatByBinding[0]);
+    EXPECT_FALSE(assigned.storageImageUsesBindingFormatByBinding[1]);
+}
+
+TEST(DirectVulkanSanity, SamplerUniformTypesPreserveTheirNumericDomain) {
+    using namespace MobileGL::MG_Backend::DirectVulkan;
+
+    EXPECT_EQ(ProgramFactory::UniformTypeToSamplerNumericDomain(GL_SAMPLER_2D),
+              SamplerNumericDomain::Float);
+    EXPECT_EQ(ProgramFactory::UniformTypeToSamplerNumericDomain(GL_SAMPLER_CUBE_MAP_ARRAY_SHADOW),
+              SamplerNumericDomain::Float);
+    EXPECT_EQ(ProgramFactory::UniformTypeToSamplerNumericDomain(GL_INT_SAMPLER_2D_ARRAY),
+              SamplerNumericDomain::SignedInteger);
+    EXPECT_EQ(ProgramFactory::UniformTypeToSamplerNumericDomain(GL_UNSIGNED_INT_SAMPLER_2D),
+              SamplerNumericDomain::UnsignedInteger);
+    EXPECT_EQ(ProgramFactory::UniformTypeToSamplerNumericDomain(GL_IMAGE_2D),
+              SamplerNumericDomain::Unknown);
+}
+
+TEST(DirectVulkanSanity, SampledViewFormatMatchesSamplerNumericDomainWithoutChangingComponentLayout) {
+    using namespace MobileGL::MG_Backend::DirectVulkan;
+
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R32_SFLOAT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_R32_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R32_SFLOAT, SamplerNumericDomain::SignedInteger),
+              VK_FORMAT_R32_SINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R32_UINT, SamplerNumericDomain::Float),
+              VK_FORMAT_R32_SFLOAT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R16G16B16A16_SFLOAT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_R16G16B16A16_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R8G8B8A8_UNORM, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_R8G8B8A8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_R32_UINT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_R32_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_B10G11R11_UFLOAT_PACK32, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_UNDEFINED);
+
+    // Depth/stencil formats never resolve through color-class reinterpretation; they pass
+    // through unchanged so the existing depth-aspect sampled view is used. Combined
+    // depth-stencil formats are multi-numeric (vkuFormatIsSampledFloat is false for them),
+    // so without the passthrough a plain sampler2D/sampler2DShadow on GL_DEPTH24_STENCIL8
+    // would resolve to UNDEFINED and the draw would be dropped.
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D24_UNORM_S8_UINT, SamplerNumericDomain::Float),
+              VK_FORMAT_D24_UNORM_S8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D32_SFLOAT_S8_UINT, SamplerNumericDomain::Float),
+              VK_FORMAT_D32_SFLOAT_S8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D32_SFLOAT, SamplerNumericDomain::Float),
+              VK_FORMAT_D32_SFLOAT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D24_UNORM_S8_UINT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_D24_UNORM_S8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D32_SFLOAT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_D32_SFLOAT);
+
+    EXPECT_TRUE(VkTextureManager::AreSampledImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_UINT));
+    EXPECT_FALSE(VkTextureManager::AreSampledImageViewFormatsCompatible(
+        VK_FORMAT_R32_SFLOAT, VK_FORMAT_R16G16B16A16_UINT));
+}
+
 TEST(RenderStateSanity, ProvokingVertexUpdatesStateAndValidatesEnum) {
     using namespace MobileGL;
 

MobileGL/MG_Test/Texture/TextureTest.cpp

@@ -158,6 +158,39 @@ namespace {
         auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
         return static_cast<const Uint8*>(mipmapObject->MapMipmapData(TextureUploadTarget::Texture2D, level));
     }
+
+    class ScopedTextureBackendFunctionsOverride {
+    public:
+        ScopedTextureBackendFunctionsOverride(): m_snapshot(MG_Backend::gBackendFunctionsTable) {}
+        ~ScopedTextureBackendFunctionsOverride() { MG_Backend::gBackendFunctionsTable = m_snapshot; }
+
+    private:
+        MG_Backend::GlobalBackendFunctionsTable m_snapshot;
+    };
+
+    struct CopyTexSubImage2DCall {
+        Bool Called = false;
+        GLenum Target = GL_NONE;
+        GLint Level = -1;
+        GLint XOffset = -1;
+        GLint YOffset = -1;
+        GLint X = -1;
+        GLint Y = -1;
+        GLsizei Width = -1;
+        GLsizei Height = -1;
+        GLuint BoundTexture = 0;
+    } g_copyTexSubImage2DCall;
+
+    void RecordCopyTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y,
+                                 GLsizei width, GLsizei height) {
+        g_copyTexSubImage2DCall = {
+            true, target, level, xoffset, yoffset, x, y, width, height,
+            MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit())
+                .GetBindingSlot(TextureTarget::Texture2D)
+                .GetBoundObject()
+                ->GetExternalIndex(),
+        };
+    }
 } // namespace
 
 TEST_F(TextureTest, CreateTexturesCreatesObjectsWithoutBinding) {
@@ -178,6 +211,156 @@ TEST_F(TextureTest, CreateTexturesCreatesObjectsWithoutBinding) {
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
 
+TEST_F(TextureTest, ClearTexImageNullClearsWholeNamedTextureAndMarksStorageDirty) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+
+    const Uint8 initialPixels[] = {
+        1, 2, 3, 4,
+        5, 6, 7, 8,
+        9, 10, 11, 12,
+        13, 14, 15, 16,
+    };
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 2, 2, 0,
+                                GL_RGBA, GL_UNSIGNED_BYTE, initialPixels);
+
+    const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+    auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
+    mipmapObject->MarkStorageDirty(TextureUploadTarget::Texture2D, 0, false);
+
+    MG_Impl::GLImpl::ClearTexImage(texture, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const Uint8* stored = GetBoundTexture2DLevelBytes(texture);
+    ASSERT_NE(stored, nullptr);
+    const Uint8 zeros[sizeof(initialPixels)] = {};
+    EXPECT_EQ(std::memcmp(stored, zeros, sizeof(zeros)), 0);
+    EXPECT_TRUE(mipmapObject->IsStorageDirty(TextureUploadTarget::Texture2D, 0));
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+TEST_F(TextureTest, ClearTexImageRepeatsConvertedClearPixel) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 2, 2, 0,
+                                GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const Uint8 clearPixel[] = {17, 34, 51, 68};
+    MG_Impl::GLImpl::ClearTexImage(texture, 0, GL_RGBA, GL_UNSIGNED_BYTE, clearPixel);
+
+    const Uint8* stored = GetBoundTexture2DLevelBytes(texture);
+    ASSERT_NE(stored, nullptr);
+    const Uint8 expected[] = {
+        17, 34, 51, 68,
+        17, 34, 51, 68,
+        17, 34, 51, 68,
+        17, 34, 51, 68,
+    };
+    EXPECT_EQ(std::memcmp(stored, expected, sizeof(expected)), 0);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+TEST_F(TextureTest, ClearTexSubImageClearsOnlyRequestedRectangle) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+    Uint8 initialPixels[3 * 2 * 4];
+    std::memset(initialPixels, 0x7f, sizeof(initialPixels));
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 3, 2, 0,
+                                GL_RGBA, GL_UNSIGNED_BYTE, initialPixels);
+
+    MG_Impl::GLImpl::ClearTexSubImage(texture, 0, 1, 0, 0, 1, 2, 1,
+                                     GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const Uint8* stored = GetBoundTexture2DLevelBytes(texture);
+    ASSERT_NE(stored, nullptr);
+    for (Int y = 0; y < 2; ++y) {
+        for (Int x = 0; x < 3; ++x) {
+            for (Int channel = 0; channel < 4; ++channel) {
+                EXPECT_EQ(stored[(y * 3 + x) * 4 + channel], x == 1 ? 0 : 0x7f);
+            }
+        }
+    }
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+TEST_F(TextureTest, CopyTextureSubImage2DUsesNamedObjectAndRestoresBinding) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyTexSubImage2D = RecordCopyTexSubImage2D;
+    g_copyTexSubImage2DCall = {};
+
+    GLuint namedTexture = 0;
+    GLuint boundTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D, 1, &namedTexture);
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_2D, 1, &boundTexture);
+    MG_Impl::GLImpl::BindTextureUnit(0, boundTexture);
+
+    const auto boundBefore = MG_State::pGLContext->GetTextureUnitObject(0)
+                                 .GetBindingSlot(TextureTarget::Texture2D)
+                                 .GetBoundObject();
+    MG_Impl::GLImpl::CopyTextureSubImage2D(namedTexture, 2, 3, 4, 5, 6, 7, 8);
+
+    EXPECT_TRUE(g_copyTexSubImage2DCall.Called);
+    EXPECT_EQ(g_copyTexSubImage2DCall.Target, GL_TEXTURE_2D);
+    EXPECT_EQ(g_copyTexSubImage2DCall.Level, 2);
+    EXPECT_EQ(g_copyTexSubImage2DCall.XOffset, 3);
+    EXPECT_EQ(g_copyTexSubImage2DCall.YOffset, 4);
+    EXPECT_EQ(g_copyTexSubImage2DCall.X, 5);
+    EXPECT_EQ(g_copyTexSubImage2DCall.Y, 6);
+    EXPECT_EQ(g_copyTexSubImage2DCall.Width, 7);
+    EXPECT_EQ(g_copyTexSubImage2DCall.Height, 8);
+    EXPECT_EQ(g_copyTexSubImage2DCall.BoundTexture, namedTexture);
+    EXPECT_EQ(MG_State::pGLContext->GetTextureUnitObject(0)
+                  .GetBindingSlot(TextureTarget::Texture2D)
+                  .GetBoundObject(),
+              boundBefore);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+TEST_F(TextureTest, CopyTextureSubImage2DRejectsCubeMapTargets) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyTexSubImage2D = RecordCopyTexSubImage2D;
+    g_copyTexSubImage2DCall = {};
+
+    GLuint cubeTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_CUBE_MAP, 1, &cubeTexture);
+    MG_Impl::GLImpl::CopyTextureSubImage2D(cubeTexture, 0, 0, 0, 0, 0, 1, 1);
+
+    // GL 4.6 sec. 8.8: the 2D form only accepts 2D/1D-array/rectangle effective targets.
+    EXPECT_FALSE(g_copyTexSubImage2DCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+}
+
+TEST_F(TextureTest, ClearTexImageErrorContracts) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 2, 2, 0,
+                                GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    // Zero texture name is INVALID_OPERATION (ARB_clear_texture).
+    MG_Impl::GLImpl::ClearTexImage(0, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+
+    // A negative level is INVALID_VALUE...
+    MG_Impl::GLImpl::ClearTexImage(texture, -1, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_VALUE));
+
+    // ...but clearing a level that was never defined is INVALID_OPERATION.
+    MG_Impl::GLImpl::ClearTexImage(texture, 5, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+
+    // A clear region outside the level is INVALID_VALUE.
+    MG_Impl::GLImpl::ClearTexSubImage(texture, 0, 1, 1, 0, 4, 4, 1,
+                                      GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_VALUE));
+
+    // An invalid pixel-transfer format is INVALID_ENUM from the shared validators.
+    MG_Impl::GLImpl::ClearTexImage(texture, 0, GL_NONE, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_ENUM));
+}
+
 // GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT is float state that must answer every numeric query: GetFloatv
 // is authoritative and GetIntegerv would otherwise fall through to its INVALID_ENUM default.
 TEST_F(TextureTest, MaxTextureMaxAnisotropyIsAnsweredFromTheBackendLimit) {

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp

@@ -865,6 +865,9 @@ namespace MobileGL::MG_Util::BackendLoader {
         GLint maxUniformBlockSize = 16384;
         GLint maxImageUnits = 8;
         GLint maxCombinedImageUniforms = 8;
+        GLint maxVertexImageUniforms = 0;
+        GLint maxGeometryImageUniforms = 0;
+        GLint maxFragmentImageUniforms = 8;
         GLint maxComputeImageUniforms = 8;
         GLint maxDrawBuffers = 8;
         GLint maxColorAttachments = 8;
@@ -904,7 +907,16 @@ namespace MobileGL::MG_Util::BackendLoader {
         glesFuncs.glGetIntegerv(GL_MAX_UNIFORM_BLOCK_SIZE, &maxUniformBlockSize);
         glesFuncs.glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
         glesFuncs.glGetIntegerv(GL_MAX_COMBINED_IMAGE_UNIFORMS, &maxCombinedImageUniforms);
+        glesFuncs.glGetIntegerv(GL_MAX_VERTEX_IMAGE_UNIFORMS, &maxVertexImageUniforms);
+        glesFuncs.glGetIntegerv(GL_MAX_FRAGMENT_IMAGE_UNIFORMS, &maxFragmentImageUniforms);
         glesFuncs.glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+        // Geometry shaders and their image-uniform query are core only in ES 3.2. DirectGLES
+        // emits ESSL 3.10 on an ES 3.1 context, so reporting zero there is both legal and an
+        // accurate description of what the backend compiler can consume.
+        if (caps.GLESVersion.Major > 3 ||
+            (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 2)) {
+            glesFuncs.glGetIntegerv(GL_MAX_GEOMETRY_IMAGE_UNIFORMS, &maxGeometryImageUniforms);
+        }
         glesFuncs.glGetIntegerv(GL_MAX_DRAW_BUFFERS, &maxDrawBuffers);
         glesFuncs.glGetIntegerv(GL_MAX_COLOR_ATTACHMENTS, &maxColorAttachments);
         glesFuncs.glGetIntegerv(GL_MAX_CLIP_DISTANCES, &maxClipDistances);
@@ -955,6 +967,9 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.MaxUniformBlockSize = maxUniformBlockSize;
         caps.MaxImageUnits = maxImageUnits;
         caps.MaxCombinedImageUniforms = maxCombinedImageUniforms;
+        caps.MaxVertexImageUniforms = maxVertexImageUniforms;
+        caps.MaxGeometryImageUniforms = maxGeometryImageUniforms;
+        caps.MaxFragmentImageUniforms = maxFragmentImageUniforms;
         caps.MaxComputeImageUniforms = maxComputeImageUniforms;
         caps.MaxDrawBuffers = maxDrawBuffers;
         caps.MaxColorAttachments = maxColorAttachments;
@@ -1000,6 +1015,9 @@ namespace MobileGL::MG_Util::BackendLoader {
         MGLOG_I("    GL_MAX_UNIFORM_BLOCK_SIZE: %d", caps.MaxUniformBlockSize);
         MGLOG_I("    GL_MAX_IMAGE_UNITS: %d", caps.MaxImageUnits);
         MGLOG_I("    GL_MAX_COMBINED_IMAGE_UNIFORMS: %d", caps.MaxCombinedImageUniforms);
+        MGLOG_I("    GL_MAX_VERTEX_IMAGE_UNIFORMS: %d", caps.MaxVertexImageUniforms);
+        MGLOG_I("    GL_MAX_GEOMETRY_IMAGE_UNIFORMS: %d", caps.MaxGeometryImageUniforms);
+        MGLOG_I("    GL_MAX_FRAGMENT_IMAGE_UNIFORMS: %d", caps.MaxFragmentImageUniforms);
         MGLOG_I("    GL_MAX_COMPUTE_IMAGE_UNIFORMS: %d", caps.MaxComputeImageUniforms);
         MGLOG_I("    GL_MAX_DRAW_BUFFERS: %d", caps.MaxDrawBuffers);
         MGLOG_I("    GL_MAX_COLOR_ATTACHMENTS: %d", caps.MaxColorAttachments);

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.h

@@ -1102,6 +1102,9 @@ namespace MobileGL {
             Int MaxUniformBlockSize = 16384;
             Int MaxImageUnits = 8;
             Int MaxCombinedImageUniforms = 8;
+            Int MaxVertexImageUniforms = 0;
+            Int MaxGeometryImageUniforms = 0;
+            Int MaxFragmentImageUniforms = 8;
             Int MaxComputeImageUniforms = 8;
             Int MaxDrawBuffers = 8;
             Int MaxColorAttachments = 8;

MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp

@@ -121,6 +121,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.VulkanAPIVersion = DecodeApiVersion(p.apiVersion);
         caps.DeviceName = p.deviceName;
         caps.DriverVersionString = DecodeDriverVersion(p.driverVersion);
+        caps.VendorId = p.vendorID;
         caps.UniformBufferOffsetAlignment = static_cast<int>(p.limits.minUniformBufferOffsetAlignment);
         caps.AliasedLineWidthRangeMin = p.limits.lineWidthRange[0];
         caps.AliasedLineWidthRangeMax = p.limits.lineWidthRange[1];
@@ -174,6 +175,10 @@ namespace MobileGL::MG_Util::BackendLoader {
         VkPhysicalDeviceFeatures supportedFeatures{};
         vkGetPhysicalDeviceFeatures(physicalDevice, &supportedFeatures);
         caps.SupportsWideLines = supportedFeatures.wideLines == VK_TRUE;
+        caps.SupportsVertexPipelineStoresAndAtomics =
+            supportedFeatures.vertexPipelineStoresAndAtomics == VK_TRUE;
+        caps.SupportsFragmentStoresAndAtomics = supportedFeatures.fragmentStoresAndAtomics == VK_TRUE;
+        caps.SupportsGeometryShader = supportedFeatures.geometryShader == VK_TRUE;
         caps.MaxShaderStorageBlockSize = static_cast<SizeT>(p.limits.maxStorageBufferRange);
         const Bool supportsShaderSubgroup = vk.vkGetPhysicalDeviceProperties2 &&
                                             HasUsableShaderSubgroupSupport(subgroupProps);
@@ -206,6 +211,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.VulkanAPIVersion = DecodeApiVersion(properties.apiVersion);
         caps.DeviceName = properties.deviceName;
         caps.DriverVersionString = DecodeDriverVersion(properties.driverVersion);
+        caps.VendorId = properties.vendorID;
         caps.UniformBufferOffsetAlignment = static_cast<int>(properties.limits.minUniformBufferOffsetAlignment);
         caps.AliasedLineWidthRangeMin = properties.limits.lineWidthRange[0];
         caps.AliasedLineWidthRangeMax = properties.limits.lineWidthRange[1];
@@ -256,6 +262,11 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.ViewportBoundsRangeMax = properties.limits.viewportBoundsRange[1];
         caps.ViewportSubpixelBits = static_cast<Int>(properties.limits.viewportSubPixelBits);
         caps.SupportsWideLines = false;
+        // This helper only receives properties, not VkPhysicalDeviceFeatures. Leave optional
+        // stage writes disabled rather than inferring them from descriptor limits alone.
+        caps.SupportsVertexPipelineStoresAndAtomics = false;
+        caps.SupportsFragmentStoresAndAtomics = false;
+        caps.SupportsGeometryShader = false;
         caps.MaxShaderStorageBlockSize = static_cast<SizeT>(properties.limits.maxStorageBufferRange);
         caps.SupportsShaderSubgroup = false;
         caps.SubgroupSize = 0;

MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.h

@@ -15,6 +15,8 @@ namespace MobileGL {
             Version VulkanAPIVersion{1, 0, 0};
             String DeviceName;
             String DriverVersionString;
+            // VkPhysicalDeviceProperties::vendorID, for device-quirk vendor gating.
+            Uint32 VendorId = 0;
             Int UniformBufferOffsetAlignment = 256;
             Float AliasedLineWidthRangeMin = 1.0f;
             Float AliasedLineWidthRangeMax = 1.0f;
@@ -67,6 +69,12 @@ namespace MobileGL {
             Float ViewportBoundsRangeMax = 0.0f;
             Int ViewportSubpixelBits = 0;
             Bool SupportsWideLines = false;
+            // Storage-image descriptors are limited per stage by
+            // maxPerStageDescriptorStorageImages, but writes/atomics outside compute additionally
+            // require these core Vulkan features to be enabled on the logical device.
+            Bool SupportsVertexPipelineStoresAndAtomics = false;
+            Bool SupportsFragmentStoresAndAtomics = false;
+            Bool SupportsGeometryShader = false;
             SizeT MaxShaderStorageBlockSize = 128 * 1024 * 1024;
             Bool SupportsShaderSubgroup = false;
             Uint32 SubgroupSize = 0;

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.cpp

@@ -6,6 +6,10 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header
 
+#define SPV_ENABLE_UTILITY_CODE
+#include "glslang/SPIRV/spirv.hpp11"
+#undef SPV_ENABLE_UTILITY_CODE
+
 #include "ShaderCompiler.h"
 
 #include "SpirvPasses/EliminateFloatEqualsZeroPass.h"
@@ -19,14 +23,16 @@
 #include "spirv-tools/optimizer.hpp"
 
 #include "ShaderSourceProcessor.h"
+#include <MG_Backend/BackendObjects.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToGlslang/ProgramEnumConverter.h>
+#include <cstdlib>
 
 namespace MobileGL {
     namespace MG_Util {
         namespace ShaderTranspiler {
-            TBuiltInResource& GetTBuiltInResourceInstance() {
-                static TBuiltInResource Resources{};
+            TBuiltInResource BuildTBuiltInResource() {
+                TBuiltInResource Resources{};
                 Resources.maxLights = 32;
                 Resources.maxClipPlanes = 6;
                 Resources.maxTextureUnits = 32;
@@ -121,6 +127,22 @@ namespace MobileGL {
                 Resources.maxTaskWorkGroupSizeZ_NV = 1;
                 Resources.maxMeshViewCountNV = 4;
 
+                // Resource checking must describe the same backend contract exposed through
+                // glGetIntegerv. Keeping this copy local also avoids racing on a process-global
+                // TBuiltInResource when Iris compiles shaders concurrently.
+                const MG_Backend::DynamicBackendParameters fallbackParameters{};
+                const auto& activeBackend = MG_Backend::pActiveBackendObject;
+                const auto& dynamicParameters =
+                    activeBackend ? activeBackend->GetDynamicParameters() : fallbackParameters;
+                Resources.maxImageUnits = dynamicParameters.MaxImageUnits;
+                Resources.maxCombinedImageUnitsAndFragmentOutputs =
+                    dynamicParameters.MaxImageUnits + dynamicParameters.MaxDrawBuffers;
+                Resources.maxVertexImageUniforms = dynamicParameters.MaxVertexImageUniforms;
+                Resources.maxGeometryImageUniforms = dynamicParameters.MaxGeometryImageUniforms;
+                Resources.maxFragmentImageUniforms = dynamicParameters.MaxFragmentImageUniforms;
+                Resources.maxComputeImageUniforms = dynamicParameters.MaxComputeImageUniforms;
+                Resources.maxCombinedImageUniforms = dynamicParameters.MaxCombinedImageUniforms;
+
                 Resources.limits.nonInductiveForLoops = true;
                 Resources.limits.whileLoops = true;
                 Resources.limits.doWhileLoops = true;
@@ -166,7 +188,8 @@ namespace MobileGL {
                 tshader->setAutoMapLocations(true);
                 tshader->setAutoMapBindings(true);
                 tshader->setGlobalUniformBlockName(GLOBAL_UBO_NAME);
-                if (!tshader->parse(&GetTBuiltInResourceInstance(), 460, ECoreProfile,
+                auto resources = BuildTBuiltInResource();
+                if (!tshader->parse(&resources, 460, ECoreProfile,
                                     /*forceDefaultVersionAndProfile: */ false,
                                     /*forwardCompatible: */ true, EShMsgDefault)) {
                     ResultInfo r;
@@ -322,6 +345,205 @@ namespace MobileGL {
                 return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
             }
 
+            bool ShaderCompiler::DecoratePositionInvariantForVulkan(const Vector<Uint32>& inputBinary,
+                                                                    Vector<uint32_t>& outputBinary) {
+                static constexpr Uint32 kHeaderWords = 5;
+                static constexpr Uint32 kOpDecorate = 71;
+                static constexpr Uint32 kOpMemberDecorate = 72;
+                static constexpr Uint32 kDecorationInvariant = 18;
+                static constexpr Uint32 kDecorationBuiltIn = 11;
+                static constexpr Uint32 kBuiltInPosition = 0;
+                if (inputBinary.size() < kHeaderWords) {
+                    return false;
+                }
+
+                // First pass: find targets that already carry Invariant so we never duplicate.
+                struct MemberKey {
+                    Uint32 id;
+                    Uint32 member;
+                    bool operator==(const MemberKey& o) const { return id == o.id && member == o.member; }
+                };
+                Vector<Uint32> invariantIds;
+                Vector<MemberKey> invariantMembers;
+                for (SizeT i = kHeaderWords; i < inputBinary.size();) {
+                    const Uint32 word0 = inputBinary[i];
+                    const Uint32 opcode = word0 & 0xFFFFu;
+                    const Uint32 length = word0 >> 16;
+                    if (length == 0 || i + length > inputBinary.size()) {
+                        return false;
+                    }
+                    if (opcode == kOpDecorate && length >= 3 && inputBinary[i + 2] == kDecorationInvariant) {
+                        invariantIds.push_back(inputBinary[i + 1]);
+                    } else if (opcode == kOpMemberDecorate && length >= 4 &&
+                               inputBinary[i + 3] == kDecorationInvariant) {
+                        invariantMembers.push_back({inputBinary[i + 1], inputBinary[i + 2]});
+                    }
+                    i += length;
+                }
+
+                outputBinary.clear();
+                outputBinary.reserve(inputBinary.size() + 8);
+                outputBinary.insert(outputBinary.end(), inputBinary.begin(), inputBinary.begin() + kHeaderWords);
+                for (SizeT i = kHeaderWords; i < inputBinary.size();) {
+                    const Uint32 word0 = inputBinary[i];
+                    const Uint32 opcode = word0 & 0xFFFFu;
+                    const Uint32 length = word0 >> 16;
+                    outputBinary.insert(outputBinary.end(), inputBinary.begin() + i,
+                                        inputBinary.begin() + i + length);
+                    if (opcode == kOpDecorate && length == 4 &&
+                        inputBinary[i + 2] == kDecorationBuiltIn && inputBinary[i + 3] == kBuiltInPosition) {
+                        const Uint32 target = inputBinary[i + 1];
+                        if (std::find(invariantIds.begin(), invariantIds.end(), target) == invariantIds.end()) {
+                            outputBinary.push_back((3u << 16) | kOpDecorate);
+                            outputBinary.push_back(target);
+                            outputBinary.push_back(kDecorationInvariant);
+                        }
+                    } else if (opcode == kOpMemberDecorate && length == 5 &&
+                               inputBinary[i + 3] == kDecorationBuiltIn && inputBinary[i + 4] == kBuiltInPosition) {
+                        const MemberKey key{inputBinary[i + 1], inputBinary[i + 2]};
+                        if (std::find(invariantMembers.begin(), invariantMembers.end(), key) ==
+                            invariantMembers.end()) {
+                            outputBinary.push_back((4u << 16) | kOpMemberDecorate);
+                            outputBinary.push_back(key.id);
+                            outputBinary.push_back(key.member);
+                            outputBinary.push_back(kDecorationInvariant);
+                        }
+                    }
+                    i += length;
+                }
+                return true;
+            }
+
+            bool ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
+                const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary) {
+                constexpr SizeT kSpirvHeaderWordCount = 5;
+                outputBinary.clear();
+                if (inputBinary.size() < kSpirvHeaderWordCount || inputBinary[0] != spv::MagicNumber) {
+                    return false;
+                }
+
+                Vector<Uint32> floatTypeIds;
+                Vector<Uint32> resultTypeById(inputBinary[3], 0);
+                Vector<Uint32> pointerPointeeTypeById(inputBinary[3], 0);
+                Bool hasReadWithoutFormatCapability = false;
+                Bool hasWriteWithoutFormatCapability = false;
+                SizeT capabilityInsertOffset = kSpirvHeaderWordCount;
+
+                for (SizeT offset = kSpirvHeaderWordCount; offset < inputBinary.size();) {
+                    const Uint32 instructionWord = inputBinary[offset];
+                    const Uint32 wordCount = instructionWord >> 16u;
+                    const auto opcode = static_cast<spv::Op>(instructionWord & 0xffffu);
+                    if (wordCount == 0 || offset + wordCount > inputBinary.size()) {
+                        return false;
+                    }
+
+                    if (opcode == spv::Op::OpCapability && wordCount >= 2) {
+                        capabilityInsertOffset = offset + wordCount;
+                        const auto capability = static_cast<spv::Capability>(inputBinary[offset + 1]);
+                        hasReadWithoutFormatCapability |=
+                            capability == spv::Capability::StorageImageReadWithoutFormat;
+                        hasWriteWithoutFormatCapability |=
+                            capability == spv::Capability::StorageImageWriteWithoutFormat;
+                    } else if (opcode == spv::Op::OpTypeFloat && wordCount >= 3) {
+                        floatTypeIds.push_back(inputBinary[offset + 1]);
+                    } else if (opcode == spv::Op::OpTypePointer && wordCount >= 4) {
+                        const Uint32 pointerTypeId = inputBinary[offset + 1];
+                        if (pointerTypeId >= pointerPointeeTypeById.size()) {
+                            return false;
+                        }
+                        pointerPointeeTypeById[pointerTypeId] = inputBinary[offset + 3];
+                    }
+
+                    bool hasResult = false;
+                    bool hasResultType = false;
+                    spv::HasResultAndType(opcode, &hasResult, &hasResultType);
+                    if (hasResult && hasResultType && wordCount >= 3) {
+                        const Uint32 resultTypeId = inputBinary[offset + 1];
+                        const Uint32 resultId = inputBinary[offset + 2];
+                        if (resultId >= resultTypeById.size()) {
+                            return false;
+                        }
+                        resultTypeById[resultId] = resultTypeId;
+                    }
+                    offset += wordCount;
+                }
+
+                // OpImageTexelPointer is the bridge to image atomic instructions. Vulkan requires
+                // those image types to retain an atomic-compatible declared format, so exclude only
+                // the exact image types used by an atomic path rather than disabling formatless
+                // access for unrelated float images in the same module.
+                Vector<Uint32> atomicImageTypeIds;
+                for (SizeT offset = kSpirvHeaderWordCount; offset < inputBinary.size();) {
+                    const Uint32 instructionWord = inputBinary[offset];
+                    const Uint32 wordCount = instructionWord >> 16u;
+                    const auto opcode = static_cast<spv::Op>(instructionWord & 0xffffu);
+                    if (opcode == spv::Op::OpImageTexelPointer && wordCount >= 6) {
+                        const Uint32 imageId = inputBinary[offset + 3];
+                        if (imageId >= resultTypeById.size()) {
+                            return false;
+                        }
+                        Uint32 imageTypeId = resultTypeById[imageId];
+                        if (imageTypeId < pointerPointeeTypeById.size() &&
+                            pointerPointeeTypeById[imageTypeId] != 0) {
+                            imageTypeId = pointerPointeeTypeById[imageTypeId];
+                        }
+                        if (imageTypeId != 0 &&
+                            std::find(atomicImageTypeIds.begin(), atomicImageTypeIds.end(), imageTypeId) ==
+                                atomicImageTypeIds.end()) {
+                            atomicImageTypeIds.push_back(imageTypeId);
+                        }
+                    }
+                    offset += wordCount;
+                }
+
+                outputBinary = inputBinary;
+                Bool hasFloatStorageImage = false;
+                for (SizeT offset = kSpirvHeaderWordCount; offset < outputBinary.size();) {
+                    const Uint32 instructionWord = outputBinary[offset];
+                    const Uint32 wordCount = instructionWord >> 16u;
+                    const auto opcode = static_cast<spv::Op>(instructionWord & 0xffffu);
+
+                    // OpTypeImage operands are: result id, sampled type, dim, depth, arrayed,
+                    // multisampled, sampled, image format, and an optional access qualifier.
+                    if (opcode == spv::Op::OpTypeImage && wordCount >= 9) {
+                        const Uint32 imageTypeId = outputBinary[offset + 1];
+                        const Uint32 sampledTypeId = outputBinary[offset + 2];
+                        const Uint32 sampled = outputBinary[offset + 7];
+                        const Bool hasFloatSampledType =
+                            std::find(floatTypeIds.begin(), floatTypeIds.end(), sampledTypeId) != floatTypeIds.end();
+                        const Bool usedByAtomic =
+                            std::find(atomicImageTypeIds.begin(), atomicImageTypeIds.end(), imageTypeId) !=
+                            atomicImageTypeIds.end();
+                        if (sampled == 2 && hasFloatSampledType && !usedByAtomic) {
+                            outputBinary[offset + 8] = static_cast<Uint32>(spv::ImageFormat::Unknown);
+                            hasFloatStorageImage = true;
+                        }
+                    }
+                    offset += wordCount;
+                }
+
+                if (!hasFloatStorageImage) {
+                    return true;
+                }
+
+                Vector<Uint32> addedCapabilities;
+                const Uint32 capabilityInstruction =
+                    (2u << 16u) | static_cast<Uint32>(spv::Op::OpCapability);
+                if (!hasReadWithoutFormatCapability) {
+                    addedCapabilities.push_back(capabilityInstruction);
+                    addedCapabilities.push_back(
+                        static_cast<Uint32>(spv::Capability::StorageImageReadWithoutFormat));
+                }
+                if (!hasWriteWithoutFormatCapability) {
+                    addedCapabilities.push_back(capabilityInstruction);
+                    addedCapabilities.push_back(
+                        static_cast<Uint32>(spv::Capability::StorageImageWriteWithoutFormat));
+                }
+                outputBinary.insert(outputBinary.begin() + static_cast<std::ptrdiff_t>(capabilityInsertOffset),
+                                    addedCapabilities.begin(), addedCapabilities.end());
+                return true;
+            }
+
             Result<String> ShaderCompiler::DecompileShader(SpvcSession& session) {
                 spvc_compiler_options options;
                 session.CreateOptions(&options);

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.h

@@ -39,6 +39,21 @@ namespace MobileGL {
                 // which wrongly includes baseInstance).
                 static bool RebaseInstanceIndexForVulkan(const Vector<Uint32>& inputBinary,
                                                          Vector<uint32_t>& outputBinary);
+                // Adds the Invariant decoration to every Position builtin output. GL apps
+                // routinely rely on cross-program position invariance for multi-pass
+                // equality depth tests (e.g. GEQUAL re-draws of the same geometry), and
+                // mobile drivers that optimize per-pipeline break that without the
+                // decoration. DirectVulkan only.
+                static bool DecoratePositionInvariantForVulkan(const Vector<Uint32>& inputBinary,
+                                                               Vector<uint32_t>& outputBinary);
+                // Replaces the declared format of float storage images with Unknown and adds the
+                // matching SPIR-V capabilities. DirectVulkan uses this only when both Vulkan
+                // shaderStorageImage*WithoutFormat features are enabled, allowing the
+                // glBindImageTexture format to select the descriptor view at runtime. Integer
+                // storage images deliberately keep their declared format for GL-compatible bit
+                // reinterpretation paths (for example, R32F storage accessed as r32ui).
+                static bool UseUnformattedFloatStorageImagesForVulkan(
+                    const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary);
                 static Result<String> DecompileShader(SpvcSession& session);
             };
         } // namespace ShaderTranspiler

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.cpp

@@ -10,10 +10,15 @@
 
 #include <algorithm>
 #include <cctype>
+#include <initializer_list>
+#include <utility>
+#include <Config.h>
 #include <MG_Backend/BackendObjects.h>
 
 namespace {
     using MobileGL::SizeT;
+    using MobileGL::String;
+    using MobileGL::Vector;
 
     bool IsIdentifierChar(char ch) {
         return (ch >= '0' && ch <= '9') || (ch >= 'A' && ch <= 'Z') || (ch >= 'a' && ch <= 'z') || ch == '_';
@@ -91,6 +96,389 @@ namespace {
         return masked;
     }
 
+    struct CodeToken {
+        String text;
+        SizeT begin = 0;
+        SizeT end = 0;
+    };
+
+    Vector<CodeToken> TokenizeCode(const String& source) {
+        const String masked = MaskCommentsAndQuotedText(source);
+        Vector<CodeToken> tokens;
+        tokens.reserve(source.size() / 4);
+
+        SizeT pos = 0;
+        while (pos < masked.size()) {
+            const char ch = masked[pos];
+            if (std::isspace(static_cast<unsigned char>(ch))) {
+                ++pos;
+                continue;
+            }
+
+            const SizeT begin = pos;
+            if (IsIdentifierStart(ch)) {
+                ++pos;
+                while (pos < masked.size() && IsIdentifierChar(masked[pos])) {
+                    ++pos;
+                }
+            } else if (std::isdigit(static_cast<unsigned char>(ch))) {
+                ++pos;
+                while (pos < masked.size()) {
+                    const char numberChar = masked[pos];
+                    if (!IsIdentifierChar(numberChar) && numberChar != '.') {
+                        break;
+                    }
+                    ++pos;
+                }
+            } else {
+                ++pos;
+                if (pos < masked.size()) {
+                    const String twoChars = masked.substr(begin, 2);
+                    if (twoChars == "==" || twoChars == "!=" || twoChars == "<=" || twoChars == ">=" ||
+                        twoChars == "+=" || twoChars == "-=" || twoChars == "<<" || twoChars == ">>" ||
+                        twoChars == "++" || twoChars == "--" || twoChars == "&&" || twoChars == "||") {
+                        ++pos;
+                    }
+                }
+            }
+
+            tokens.push_back(CodeToken{source.substr(begin, pos - begin), begin, pos});
+        }
+        return tokens;
+    }
+
+    bool IsIdentifierToken(const CodeToken& token) {
+        if (token.text.empty() || !IsIdentifierStart(token.text.front())) {
+            return false;
+        }
+        return std::all_of(token.text.begin() + 1, token.text.end(), IsIdentifierChar);
+    }
+
+    class TokenCursor {
+    public:
+        TokenCursor(const Vector<CodeToken>& tokens, SizeT position) : m_tokens(tokens), m_position(position) {}
+
+        bool Consume(const char* expected) {
+            if (m_position >= m_tokens.size() || m_tokens[m_position].text != expected) {
+                return false;
+            }
+            ++m_position;
+            return true;
+        }
+
+        bool ConsumeAnyIdentifier(String& identifier) {
+            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position])) {
+                return false;
+            }
+            identifier = m_tokens[m_position++].text;
+            return true;
+        }
+
+        bool ConsumeAnyIdentifier() {
+            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position])) {
+                return false;
+            }
+            ++m_position;
+            return true;
+        }
+
+        bool ConsumeIdentifier(const String& expected) {
+            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position]) ||
+                m_tokens[m_position].text != expected) {
+                return false;
+            }
+            ++m_position;
+            return true;
+        }
+
+        SizeT Position() const { return m_position; }
+
+    private:
+        const Vector<CodeToken>& m_tokens;
+        SizeT m_position;
+    };
+
+    SizeT CountToken(const Vector<CodeToken>& tokens, const String& tokenText) {
+        return static_cast<SizeT>(std::count_if(tokens.begin(), tokens.end(),
+                                                [&](const CodeToken& token) { return token.text == tokenText; }));
+    }
+
+    bool HasIdentifierWithPrefixOutsideAllowed(const Vector<CodeToken>& tokens, const String& prefix,
+                                               std::initializer_list<const char*> allowedIdentifiers) {
+        return std::any_of(tokens.begin(), tokens.end(), [&](const CodeToken& token) {
+            if (!IsIdentifierToken(token) || !token.text.starts_with(prefix)) {
+                return false;
+            }
+            return std::none_of(allowedIdentifiers.begin(), allowedIdentifiers.end(),
+                                [&](const char* allowed) { return token.text == allowed; });
+        });
+    }
+
+    bool MatchTokenSequence(const Vector<CodeToken>& tokens, SizeT position,
+                            std::initializer_list<const char*> expected) {
+        if (position + expected.size() > tokens.size()) {
+            return false;
+        }
+        for (const char* token : expected) {
+            if (tokens[position++].text != token) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    struct LinearPrefixScanMatch {
+        SizeT sharedArraySizeBegin = 0;
+        SizeT sharedArraySizeEnd = 0;
+        SizeT scanBegin = 0;
+        SizeT scanEnd = 0;
+        String cache;
+        String importance;
+        String prefixSum;
+        String loopLength;
+        String loopIndex;
+        String sum;
+    };
+
+    bool ParseLinearPrefixScanTemplate(const Vector<CodeToken>& tokens, LinearPrefixScanMatch& match) {
+        // The workaround deliberately recognizes one complete algorithm, not merely the
+        // subgroupInclusiveAdd token. Changing scratch storage is only safe when that storage is
+        // private to this scan and the workgroup has exactly 1024 X invocations.
+        SizeT localSizeDeclarationCount = 0;
+        for (SizeT i = 0; i < tokens.size(); ++i) {
+            if (MatchTokenSequence(tokens, i, {"layout", "(", "local_size_x", "=", "1024", ")", "in", ";"})) {
+                ++localSizeDeclarationCount;
+            }
+        }
+        if (localSizeDeclarationCount != 1) {
+            return false;
+        }
+
+        SizeT sharedDeclarationIndex = String::npos;
+        SizeT sharedDeclarationCount = 0;
+        String cacheName;
+        for (SizeT i = 0; i + 6 < tokens.size(); ++i) {
+            if (tokens[i].text != "shared" || tokens[i + 1].text != "float" || !IsIdentifierToken(tokens[i + 2]) ||
+                tokens[i + 3].text != "[" || tokens[i + 4].text != "64" || tokens[i + 5].text != "]" ||
+                tokens[i + 6].text != ";") {
+                continue;
+            }
+            ++sharedDeclarationCount;
+            sharedDeclarationIndex = i;
+            cacheName = tokens[i + 2].text;
+        }
+        if (sharedDeclarationCount != 1) {
+            return false;
+        }
+
+        SizeT scanTokenIndex = String::npos;
+        SizeT scanCount = 0;
+        for (SizeT i = 0; i + 7 < tokens.size(); ++i) {
+            if (tokens[i].text == "float" && IsIdentifierToken(tokens[i + 1]) && tokens[i + 2].text == "=" &&
+                tokens[i + 3].text == "subgroupInclusiveAdd" && tokens[i + 4].text == "(" &&
+                IsIdentifierToken(tokens[i + 5]) && tokens[i + 6].text == ")" && tokens[i + 7].text == ";") {
+                ++scanCount;
+                scanTokenIndex = i;
+            }
+        }
+        if (scanCount != 1 || sharedDeclarationIndex >= scanTokenIndex) {
+            return false;
+        }
+
+        TokenCursor cursor(tokens, scanTokenIndex);
+        String prefixSum;
+        String importance;
+        String loopLength;
+        String loopIndex;
+        String sum;
+        if (!cursor.Consume("float") || !cursor.ConsumeAnyIdentifier(prefixSum) || !cursor.Consume("=") ||
+            !cursor.Consume("subgroupInclusiveAdd") || !cursor.Consume("(") ||
+            !cursor.ConsumeAnyIdentifier(importance) || !cursor.Consume(")") || !cursor.Consume(";") ||
+            !cursor.Consume("if") || !cursor.Consume("(") || !cursor.Consume("gl_SubgroupInvocationID") ||
+            !cursor.Consume("==") || !cursor.Consume("gl_SubgroupSize") || !cursor.Consume("-") ||
+            !cursor.Consume("1u") || !cursor.Consume(")") || !cursor.ConsumeIdentifier(cacheName) ||
+            !cursor.Consume("[") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume("]") || !cursor.Consume("=") ||
+            !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume(";") || !cursor.Consume("barrier") ||
+            !cursor.Consume("(") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("uint") ||
+            !cursor.ConsumeAnyIdentifier(loopLength) || !cursor.Consume("=") || !cursor.Consume("uint") ||
+            !cursor.Consume("(") || !cursor.Consume("findMSB") || !cursor.Consume("(") ||
+            !cursor.Consume("gl_NumSubgroups") || !cursor.Consume(")") || !cursor.Consume(")") ||
+            !cursor.Consume(";") || !cursor.ConsumeIdentifier(loopLength) || !cursor.Consume("+=") ||
+            !cursor.Consume("uint") || !cursor.Consume("(") || !cursor.Consume("gl_NumSubgroups") ||
+            !cursor.Consume("-") || !cursor.Consume("(") || !cursor.Consume("1u") || !cursor.Consume("<<") ||
+            !cursor.Consume("(") || !cursor.ConsumeIdentifier(loopLength) || !cursor.Consume("-") ||
+            !cursor.Consume("1u") || !cursor.Consume(")") || !cursor.Consume(")") || !cursor.Consume(">") ||
+            !cursor.Consume("0u") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("for") ||
+            !cursor.Consume("(") || !cursor.Consume("uint") || !cursor.ConsumeAnyIdentifier(loopIndex) ||
+            !cursor.Consume("=") || !cursor.Consume("0") || !cursor.Consume(";") ||
+            !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("<") || !cursor.ConsumeIdentifier(loopLength) ||
+            !cursor.Consume(";") || !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("++") ||
+            !cursor.Consume(")") || !cursor.Consume("{") || !cursor.Consume("if") || !cursor.Consume("(") ||
+            !cursor.Consume("(") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume("&") || !cursor.Consume("(") ||
+            !cursor.Consume("1u") || !cursor.Consume("<<") || !cursor.ConsumeIdentifier(loopIndex) ||
+            !cursor.Consume(")") || !cursor.Consume(")") || !cursor.Consume(">") || !cursor.Consume("0u") ||
+            !cursor.Consume(")") || !cursor.Consume("{") || !cursor.ConsumeIdentifier(prefixSum) ||
+            !cursor.Consume("+=") || !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") ||
+            !cursor.Consume("(") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume(">>") ||
+            !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("<<") || !cursor.ConsumeIdentifier(loopIndex) ||
+            !cursor.Consume(")") || !cursor.Consume("-") || !cursor.Consume("1u") || !cursor.Consume("]") ||
+            !cursor.Consume(";") || !cursor.Consume("if") || !cursor.Consume("(") ||
+            !cursor.Consume("gl_SubgroupInvocationID") || !cursor.Consume("==") || !cursor.Consume("gl_SubgroupSize") ||
+            !cursor.Consume("-") || !cursor.Consume("1u") || !cursor.Consume(")") ||
+            !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") || !cursor.Consume("gl_SubgroupID") ||
+            !cursor.Consume("]") || !cursor.Consume("=") || !cursor.ConsumeIdentifier(prefixSum) ||
+            !cursor.Consume(";") || !cursor.Consume("}") || !cursor.Consume("barrier") || !cursor.Consume("(") ||
+            !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("}") || !cursor.Consume("if") ||
+            !cursor.Consume("(") || !cursor.Consume("gl_LocalInvocationID") || !cursor.Consume(".") ||
+            !cursor.Consume("x") || !cursor.Consume("==") || !cursor.Consume("uint") || !cursor.Consume("(") ||
+            !cursor.Consume("1024") || !cursor.Consume("-") || !cursor.Consume("1") || !cursor.Consume(")") ||
+            !cursor.Consume(")") || !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") ||
+            !cursor.Consume("0") || !cursor.Consume("]") || !cursor.Consume("=") ||
+            !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume(";") || !cursor.Consume("barrier") ||
+            !cursor.Consume("(") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("float") ||
+            !cursor.ConsumeAnyIdentifier(sum) || !cursor.Consume("=") || !cursor.ConsumeIdentifier(cacheName) ||
+            !cursor.Consume("[") || !cursor.Consume("0") || !cursor.Consume("]") || !cursor.Consume(";")) {
+            return false;
+        }
+        const SizeT scanEndToken = cursor.Position() - 1;
+
+        // Require the scan's immediate consumer as well. This makes the match specific to a
+        // linear distribution warp, and avoids changing unrelated prefix scans which may rely on
+        // the implementation's native subgroup partitioning.
+        if (!cursor.Consume("float") || !cursor.ConsumeAnyIdentifier() || !cursor.Consume("=") ||
+            !cursor.Consume("(") || !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume("-") ||
+            !cursor.ConsumeIdentifier(importance) || !cursor.Consume(")") || !cursor.Consume("/") ||
+            !cursor.ConsumeIdentifier(sum) || !cursor.Consume("-") || !cursor.Consume("float") ||
+            !cursor.Consume("(") || !cursor.Consume("gl_LocalInvocationID") || !cursor.Consume(".") ||
+            !cursor.Consume("x") || !cursor.Consume("+") || !cursor.Consume("1u") || !cursor.Consume(")") ||
+            !cursor.Consume("/") || !cursor.Consume("float") || !cursor.Consume("(") || !cursor.Consume("1024") ||
+            !cursor.Consume(")") || !cursor.Consume(";")) {
+            return false;
+        }
+
+        // No other use may share the scratch array, and no additional subgroup operation or
+        // builtin may silently retain native-64 semantics after this module becomes virtual-32.
+        if (CountToken(tokens, cacheName) != 6 || CountToken(tokens, "subgroupInclusiveAdd") != 1 ||
+            CountToken(tokens, "gl_SubgroupInvocationID") != 2 || CountToken(tokens, "gl_SubgroupSize") != 2 ||
+            CountToken(tokens, "gl_SubgroupID") != 4 || CountToken(tokens, "gl_NumSubgroups") != 2 ||
+            CountToken(tokens, "gl_LocalInvocationID") != 2 || CountToken(tokens, "barrier") != 3 ||
+            CountToken(tokens, "findMSB") != 1 ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "subgroup", {"subgroupInclusiveAdd"}) ||
+            HasIdentifierWithPrefixOutsideAllowed(
+                tokens, "gl_Subgroup",
+                {"gl_SubgroupInvocationID", "gl_SubgroupSize", "gl_SubgroupID", "gl_NumSubgroups"}) ||
+            // ARB/NV spellings of lane-width-sensitive builtins and functions
+            // (gl_SubGroupSizeARB, ballotARB, gl_WarpSizeNV, shuffleNV, ...) must block the
+            // rewrite just like their KHR counterparts: they would silently keep native-width
+            // semantics in a module rewritten to the virtual 32-lane model.
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_SubGroup", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_Warp", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_Thread", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_SMID", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "ballot", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "shuffle", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "readInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "readFirstInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "anyInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "allInvocations", {})) {
+            return false;
+        }
+
+        // The scan must be at the top level of the sole main() body. Its existing barriers already
+        // require uniform control flow; this check prevents us from introducing extra barriers in
+        // a nested branch or loop.
+        SizeT mainOpenBrace = String::npos;
+        SizeT mainCloseBrace = String::npos;
+        SizeT mainCount = 0;
+        for (SizeT i = 0; i + 4 < tokens.size(); ++i) {
+            if (!MatchTokenSequence(tokens, i, {"void", "main", "(", ")", "{"})) {
+                continue;
+            }
+            ++mainCount;
+            mainOpenBrace = i + 4;
+            int depth = 1;
+            for (SizeT j = mainOpenBrace + 1; j < tokens.size(); ++j) {
+                if (tokens[j].text == "{")
+                    ++depth;
+                else if (tokens[j].text == "}" && --depth == 0) {
+                    mainCloseBrace = j;
+                    break;
+                }
+            }
+        }
+        if (mainCount != 1 || mainCloseBrace == String::npos || scanTokenIndex <= mainOpenBrace ||
+            scanEndToken >= mainCloseBrace) {
+            return false;
+        }
+        int depthAtScan = 1;
+        for (SizeT i = mainOpenBrace + 1; i < scanTokenIndex; ++i) {
+            if (tokens[i].text == "{")
+                ++depthAtScan;
+            else if (tokens[i].text == "}")
+                --depthAtScan;
+        }
+        if (depthAtScan != 1) {
+            return false;
+        }
+
+        constexpr const char* injectedNames[] = {"mglPrefixScanLane",  "mglVirtualSubgroupInvocation",
+                                                 "mglVirtualSubgroup", "mglVirtualSubgroupBase",
+                                                 "mglPrefixLane",      "mglVirtualSubgroupCount"};
+        for (const char* injectedName : injectedNames) {
+            if (CountToken(tokens, injectedName) != 0) {
+                return false;
+            }
+        }
+
+        match.sharedArraySizeBegin = tokens[sharedDeclarationIndex + 4].begin;
+        match.sharedArraySizeEnd = tokens[sharedDeclarationIndex + 4].end;
+        match.scanBegin = tokens[scanTokenIndex].begin;
+        match.scanEnd = tokens[scanEndToken].end;
+        match.cache = std::move(cacheName);
+        match.importance = std::move(importance);
+        match.prefixSum = std::move(prefixSum);
+        match.loopLength = std::move(loopLength);
+        match.loopIndex = std::move(loopIndex);
+        match.sum = std::move(sum);
+        return true;
+    }
+
+    String BuildLinearPrefixScanReplacement(const LinearPrefixScanMatch& match) {
+        String replacement;
+        replacement.reserve(1800);
+        replacement += "uint mglPrefixScanLane = gl_LocalInvocationID.x;\n";
+        replacement += "uint mglVirtualSubgroupInvocation = mglPrefixScanLane & 31u;\n";
+        replacement += "uint mglVirtualSubgroup = mglPrefixScanLane >> 5u;\n";
+        replacement += "const uint mglVirtualSubgroupCount = 32u;\n";
+        replacement += match.cache + "[mglPrefixScanLane] = " + match.importance + ";\n";
+        replacement += "barrier();\n";
+        replacement += "float " + match.prefixSum + " = 0.0f;\n";
+        replacement += "uint mglVirtualSubgroupBase = mglVirtualSubgroup << 5u;\n";
+        replacement += "for (uint mglPrefixLane = mglVirtualSubgroupBase; "
+                       "mglPrefixLane <= mglPrefixScanLane; ++mglPrefixLane) {\n";
+        replacement += match.prefixSum + " += " + match.cache + "[mglPrefixLane];\n";
+        replacement += "}\n";
+        replacement += "barrier();\n";
+        replacement += "if (mglVirtualSubgroupInvocation == 31u) " + match.cache +
+                       "[mglVirtualSubgroup] = " + match.prefixSum + ";\n";
+        replacement += "barrier();\n";
+        replacement += "uint " + match.loopLength + " = uint(findMSB(mglVirtualSubgroupCount));\n";
+        replacement +=
+            match.loopLength + " += uint(mglVirtualSubgroupCount - (1u << (" + match.loopLength + " - 1u)) > 0u);\n";
+        replacement += "for (uint " + match.loopIndex + " = 0u; " + match.loopIndex + " < " + match.loopLength +
+                       "; ++" + match.loopIndex + ") {\n";
+        replacement += "if ((mglVirtualSubgroup & (1u << " + match.loopIndex + ")) > 0u) {\n";
+        replacement += match.prefixSum + " += " + match.cache + "[(mglVirtualSubgroup >> " + match.loopIndex + " << " +
+                       match.loopIndex + ") - 1u];\n";
+        replacement += "if (mglVirtualSubgroupInvocation == 31u) " + match.cache +
+                       "[mglVirtualSubgroup] = " + match.prefixSum + ";\n";
+        replacement += "}\nbarrier();\n}\n";
+        replacement += "if (mglPrefixScanLane == 1023u) " + match.cache + "[0] = " + match.prefixSum + ";\n";
+        replacement += "barrier();\n";
+        replacement += "float " + match.sum + " = " + match.cache + "[0];";
+        return replacement;
+    }
+
     void SkipDirectiveWhitespace(const MobileGL::String& source, SizeT& pos, SizeT lineEnd) {
         while (pos < lineEnd && std::isspace(static_cast<unsigned char>(source[pos]))) {
             pos++;
@@ -576,6 +964,114 @@ namespace {
 namespace MobileGL {
     namespace MG_Util {
         namespace ShaderTranspiler {
+            Bool RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage stage, Uint32 nativeSubgroupSize,
+                                                          String& source) {
+                constexpr Uint32 capturedSubgroupSize = 32;
+                if (stage != ShaderStage::Compute || nativeSubgroupSize <= capturedSubgroupSize ||
+                    nativeSubgroupSize % capturedSubgroupSize != 0) {
+                    return false;
+                }
+
+                // Vulkan subgroup widths are powers of two. Keep the workaround restricted to
+                // wider widths which are a power-of-two multiple of the captured 32-lane model.
+                const Uint32 subgroupScale = nativeSubgroupSize / capturedSubgroupSize;
+                if ((subgroupScale & (subgroupScale - 1u)) != 0u) {
+                    return false;
+                }
+
+                const Vector<CodeToken> tokens = TokenizeCode(source);
+                LinearPrefixScanMatch match;
+                if (!ParseLinearPrefixScanTemplate(tokens, match)) {
+                    // Diagnosability: when the trigger op is present but the template no longer
+                    // matches (e.g. the pack shipped a new shader revision), the affected device
+                    // silently falls back to the driver's miscompiled path. Make that visible.
+                    if (CountToken(tokens, "subgroupInclusiveAdd") > 0) {
+                        MGLOG_W("%s: subgroupInclusiveAdd present but the linear prefix-scan template "
+                                "did not match; the wide-subgroup rewrite was NOT applied",
+                                __func__);
+                    }
+                    return false;
+                }
+
+                const String replacement = BuildLinearPrefixScanReplacement(match);
+                source.replace(match.scanBegin, match.scanEnd - match.scanBegin, replacement);
+                // The declaration occurs before the replaced scan, so its original offsets remain
+                // valid after the first replacement.
+                source.replace(match.sharedArraySizeBegin, match.sharedArraySizeEnd - match.sharedArraySizeBegin,
+                               "1024");
+                return true;
+            }
+
+            namespace {
+                struct ShaderSourceQuirkContext {
+                    ShaderStage stage = ShaderStage::Unknown;
+                    BackendType backend = BackendType::Unknown;
+                    MG_Backend::GpuVendorKind vendor = MG_Backend::GpuVendorKind::Unknown;
+                    Uint32 subgroupSize = 0;
+                };
+
+                // Device-quirk registry. Every entry is a narrowly scoped source rewrite that
+                // works around a specific driver defect. A quirk runs when its env override
+                // forces it on, or when the override is Auto and DeviceApplies matches the
+                // detected device. ForceOn bypasses only the device gate - each Apply keeps
+                // its own structural safety checks. Add new per-device workarounds here
+                // instead of open-coding them in PreprocessShaderSource.
+                struct ShaderSourceQuirk {
+                    const char* name;
+                    MG_Config::QuirkOverride (*GetOverride)();
+                    Bool (*DeviceApplies)(const ShaderSourceQuirkContext&);
+                    Bool (*Apply)(const ShaderSourceQuirkContext&, String&);
+                };
+
+                constexpr ShaderSourceQuirk kShaderSourceQuirks[] = {
+                    {
+                        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN
+                        "subgroup-prefix-scan-rewrite",
+                        [] { return MG_Config::Features.SubgroupPrefixScanQuirk; },
+                        [](const ShaderSourceQuirkContext& ctx) {
+                            // Qualcomm's Vulkan driver miscompiles the recognized float
+                            // InclusiveScan pattern for native subgroups wider than the
+                            // captured 32 lanes; other vendors compile it correctly and
+                            // should keep their native scan.
+                            return ctx.backend == BackendType::DirectVulkan &&
+                                   ctx.vendor == MG_Backend::GpuVendorKind::Qualcomm;
+                        },
+                        [](const ShaderSourceQuirkContext& ctx, String& source) {
+                            return RewriteLinearSubgroupPrefixScanForVulkan(ctx.stage, ctx.subgroupSize,
+                                                                            source);
+                        },
+                    },
+                };
+
+                void ApplyShaderSourceQuirks(ShaderStage stage, String& source) {
+                    const auto& activeBackend = MG_Backend::pActiveBackendObject;
+                    if (!activeBackend) {
+                        return;
+                    }
+                    const auto& dynamicParameters = activeBackend->GetDynamicParameters();
+                    const ShaderSourceQuirkContext quirkContext{
+                        stage,
+                        activeBackend->GetBackendType(),
+                        dynamicParameters.GpuVendor,
+                        dynamicParameters.SubgroupSize,
+                    };
+                    for (const ShaderSourceQuirk& quirk : kShaderSourceQuirks) {
+                        const MG_Config::QuirkOverride quirkOverride = quirk.GetOverride();
+                        if (quirkOverride == MG_Config::QuirkOverride::ForceOff) {
+                            continue;
+                        }
+                        if (quirkOverride == MG_Config::QuirkOverride::Auto &&
+                            !quirk.DeviceApplies(quirkContext)) {
+                            continue;
+                        }
+                        if (quirk.Apply(quirkContext, source)) {
+                            MGLOG_I("ApplyShaderSourceQuirks: applied '%s'%s", quirk.name,
+                                    quirkOverride == MG_Config::QuirkOverride::ForceOn ? " (forced on)" : "");
+                        }
+                    }
+                }
+            } // namespace
+
             void PreprocessShaderSource(ShaderStage stage, String& source) {
                 // Normalize while the inspector's source span still refers to the untouched input. Later passes
                 // remove comments and directives, so any subsequent insertion re-inspects the current source.
@@ -631,6 +1127,8 @@ namespace MobileGL {
                 RenameBuiltinShadowingFunction(source, "max3", "mg_max3");
                 ModernizeLegacyGLSL(stage, source);
                 InjectDepthRangeBuiltinShim(stage, source);
+
+                ApplyShaderSourceQuirks(stage, source);
             }
 
             Bool RetargetLegacyVersionDirectiveTo460(String& source) {

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.h

@@ -21,6 +21,18 @@ namespace MobileGL {
         namespace ShaderTranspiler {
             void PreprocessShaderSource(ShaderStage stage, String& source);
 
+            // Some desktop-captured compute shaders build a workgroup-wide linear prefix scan
+            // from subgroupInclusiveAdd plus a shared array of subgroup totals. Qualcomm's
+            // Vulkan driver miscompiles that exact float InclusiveScan path for native subgroups
+            // wider than the capture's 32 lanes. For the narrowly recognized, uniform-control-
+            // flow template, replace the subgroup-local scan with a shared-memory, strict
+            // left-fold over virtual 32-lane segments. Returns true only when the complete safe
+            // template was recognized and rewritten. PreprocessShaderSource reaches this through
+            // its device-quirk registry: by default only on detected Qualcomm Vulkan devices,
+            // overridable either way with MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN=1/0. The explicit
+            // entry point exists for deterministic tests.
+            Bool RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage stage, Uint32 nativeSubgroupSize, String& source);
+
             // Rewrites a "#version 330 core" directive that PreprocessShaderSource normalized down
             // from a legacy desktop version back up to "#version 460 core". Returns false (leaving
             // the source untouched) for anything else: ES, compatibility, or an already-modern
@@ -30,4 +42,4 @@ namespace MobileGL {
             Bool RetargetLegacyVersionDirectiveTo460(String& source);
         } // namespace ShaderTranspiler
     } // namespace MG_Util
-} // namespace MobileGL
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/glslang/TMglGlslIoResolver.cpp

@@ -63,7 +63,17 @@ namespace MobileGL {
     void TMglGlslIoResolver::reserverStorageSlot(glslang::TVarEntryInfo& ent, TInfoSink& infoSink) {
         const glslang::TType& type = ent.symbol->getType();
         const glslang::TString& name = ent.symbol->getAccessName();
-        if (currentStage == EShLangVertex && type.getQualifier().isPipeInput()) {
+        // OpenGL assigns generic vertex attribute locations only to active inputs. glslang gathers
+        // both live and dead declarations before mapping, so allowing the default collector to
+        // reserve a dead vertex input would make it consume a location that an active input should
+        // reuse. Other stage interfaces still need the default cross-stage matching behavior.
+        if (!ent.live && currentStage == EShLangVertex && type.getQualifier().isPipeInput()) {
+            return;
+        }
+        // glBindAttribLocation only affects active inputs in the linked program. Applying an API
+        // binding to an inactive declaration would reserve its slot in glslang's collector and
+        // incorrectly push an active, automatically mapped input to a different location.
+        if (ent.live && currentStage == EShLangVertex && type.getQualifier().isPipeInput()) {
             auto it = m_explicitVertexIns.find(name.c_str());
             if (it != m_explicitVertexIns.end()) {
                 auto& writableType = ent.symbol->getWritableType();
@@ -94,6 +104,13 @@ namespace MobileGL {
         TDefaultGlslIoResolver::reserverStorageSlot(ent, infoSink);
     }
 
+    int TMglGlslIoResolver::resolveInOutLocation(EShLanguage stage, glslang::TVarEntryInfo& ent) {
+        if (!ent.live && stage == EShLangVertex && ent.symbol->getType().getQualifier().isPipeInput()) {
+            return ent.newLocation = -1;
+        }
+        return TDefaultGlslIoResolver::resolveInOutLocation(stage, ent);
+    }
+
     void TMglGlslIoResolver::reserverResourceSlot(glslang::TVarEntryInfo& ent, TInfoSink& infoSink) {
         const glslang::TType& type = ent.symbol->getType();
         if (m_explicitOpaqueUniformBindings != nullptr && type.getBasicType() == glslang::EbtSampler &&

MobileGL/MG_Util/ShaderTranspiler/glslang/TMglGlslIoResolver.h

@@ -37,6 +37,7 @@ namespace MobileGL {
                                  opaqueUniformBindings) {}
         void reserverStorageSlot(glslang::TVarEntryInfo& ent, TInfoSink& infoSink) override;
         void reserverResourceSlot(glslang::TVarEntryInfo& ent, TInfoSink& infoSink) override;
+        int resolveInOutLocation(EShLanguage stage, glslang::TVarEntryInfo& ent) override;
         int resolveUniformLocation(EShLanguage stage, glslang::TVarEntryInfo& ent) override;
 
     protected:

tools/trace_replay/README.md

@@ -93,6 +93,15 @@ The bundled fixtures cover:
 - minecraft-1.21.4-fabric-iris-iterationt-nodsa-in-world: captured from Minecraft 1.21.4 Fabric with Sodium, Iris, and
   iterationT after entering a singleplayer world, with Iris' DSA path disabled.
   ![Minecraft 1.21.4 Fabric Iris iterationT no-DSA in-world golden](fixtures/minecraft-1.21.4-fabric-iris-iterationt-nodsa-in-world.0000115019.png)
+- minecraft-1.21.4-fabric-iris-iterationrp-in-world: captured from Minecraft 1.21.4 Fabric with Sodium, Iris, and
+  iterationRP after entering a singleplayer world, framing the iterationRP name overlay over a lake with far-shore
+  tree reflections. iterationRP's temporal auto-exposure makes a single-frame trim overexpose and drop the overlay,
+  so the fixture is a prefix trace (all calls up to the target frame) that replays the temporal state. The pack also
+  gates an NVIDIA-only shadow path (`subgroupPartitionNV`, `GL_NV_shader_subgroup_partitioned`) on the GL vendor
+  string, so the capture reports a masked vendor and the trace carries the portable `subgroupShuffleXor` path that
+  non-NVIDIA GPUs take.
+  The trace archive and golden are not committed yet (the repository's Git LFS quota rejects new objects with
+  `GH009`); the case stays registered and its fixture files are hydrated from the trace fixture mirror.
 - minecraft-1.21.4-fabric-iris-photon-v1.1-in-world: captured from Minecraft 1.21.4 Fabric with Sodium, Iris, and
   Photon v1.1 after entering a singleplayer world.
   ![Minecraft 1.21.4 Fabric Iris Photon v1.1 in-world golden](fixtures/minecraft-1.21.4-fabric-iris-photon-v1.1-in-world.0000159866.png)

tools/trace_replay/renderdoc_trace_capture/agents/openai.yaml

@@ -1,4 +0,0 @@
-interface:
-  display_name: "Capture RenderDoc Trace Frame"
-  short_description: "Capture exact Android Vulkan/GLES trace frames"
-  default_prompt: "Use $renderdoc-capture-trace-frame to capture and validate an exact Android retrace frame."

tools/trace_replay/run_android_retrace_local.py

@@ -18,17 +18,16 @@ SUMMARY_HTML = "mobilegl-android-retrace-overview.html"
 DEFAULT_ANGLE_VARIANT = "ec889e6ea831"
 BLISS_ANGLE_VARIANT = "90a62123d794"
 BLISS_CASE = "minecraft-1.21.4-fabric-iris-bliss-in-world"
+TRACE_APK_DIR = ROOT / "android-plugin" / "app" / "build" / "outputs" / "apk" / "trace" / "debug"
 
 BACKENDS = {
     "DirectGLES": {
-        "apk": ROOT / "android-plugin" / "app" / "build" / "outputs" / "apk" / "esprytTrace" / "debug" / "MobileGL-EsprytTrace-debug.apk",
-        "package": "top.mobilegl.plugin.espryt.trace",
-        "use_angle": True,
+        "package": "top.mobilegl.plugin.trace",
+        "use_angle": False,
         "use_pbuffer": False,
     },
     "DirectVulkan": {
-        "apk": ROOT / "android-plugin" / "app" / "build" / "outputs" / "apk" / "magmaTrace" / "debug" / "MobileGL-MagmaTrace-debug.apk",
-        "package": "top.mobilegl.plugin.magma.trace",
+        "package": "top.mobilegl.plugin.trace",
         "use_angle": False,
         "use_pbuffer": False,
     },
@@ -45,6 +44,11 @@ def is_lfs_pointer(path):
     return path.exists() and path.read_bytes()[:80].startswith(b"version https://git-lfs.github.com/spec/v1")
 
 
+def find_trace_apk():
+    candidates = list(TRACE_APK_DIR.glob("MobileGL-plugin-trace-release-*.apk"))
+    return max(candidates, key=lambda path: path.stat().st_mtime) if candidates else None
+
+
 def bash_path(path):
     path = Path(path).resolve()
     drive = path.drive.rstrip(":").lower()
@@ -117,9 +121,13 @@ def render_summary():
 
 def run_case(case, backend):
     backend_info = BACKENDS[backend]
+    apk = find_trace_apk()
     trace_archive = FIXTURES / case["trace_archive"]
     golden = FIXTURES / case["golden"]
     alternate = FIXTURES / case["alternate_golden"] if case.get("alternate_golden") else None
+    if apk is None:
+        mark_skipped(case, backend, f"SKIPPED_MISSING_APK: no trace APK found under {TRACE_APK_DIR}")
+        return 2
     if not trace_archive.exists() or is_lfs_pointer(trace_archive):
         mark_skipped(case, backend, "SKIPPED_LFS_POINTER: trace archive is missing or still an LFS pointer")
         copy_goldens(case, backend)
@@ -134,7 +142,7 @@ def run_case(case, backend):
         "C:/Program Files/Git/bin/bash.exe",
         "android-plugin/trace-replay-ci.sh",
         "--apk-file",
-        bash_path(backend_info["apk"]),
+        bash_path(apk),
         "--package",
         backend_info["package"],
         "--backend",

tools/trace_replay/run_macos_window_retrace_local.py

@@ -19,7 +19,12 @@ RESULT_ROOT = ROOT / ".trace-work" / "macos-window-retrace-result"
 WORK_ROOT = ROOT / ".trace-work" / "macos-window-retrace-work"
 SUMMARY_DIR = ROOT / ".trace-work" / "macos-window-retrace-summary"
 SUMMARY_HTML = "mobilegl-macos-window-vulkan-retrace-overview.html"
-DEFAULT_MIRROR_BASE = "https://repo.miawa.cn/mgl/tools/trace_replay/fixtures"
+# Fixture mirrors, tried in order before Git LFS.
+DEFAULT_MIRROR_BASES = [
+    "https://git.hit.moe/swung0x48/MobileGL/media/branch/dev/tools/trace_replay/fixtures",
+    "https://repo.miawa.cn/mgl/tools/trace_replay/fixtures",
+]
+DEFAULT_MIRROR_BASE = DEFAULT_MIRROR_BASES[0]
 BACKENDS = ("DirectVulkan",)
 CASES = load_trace_cases()
 
@@ -100,34 +105,35 @@ def default_vulkan_icd(mobilegl_library):
     return ""
 
 
-def download_fixture(path, mirror_base):
+def download_fixture(path, mirror_bases):
+    """Try each mirror in order; return on the first that serves a good file."""
+    if isinstance(mirror_bases, str):
+        mirror_bases = [mirror_bases]
     path.parent.mkdir(parents=True, exist_ok=True)
-    url = f"{mirror_base.rstrip('/')}/{path.name}"
     tmp = path.with_suffix(path.suffix + ".tmp")
-    command = [
-        "curl",
-        "-L",
-        "--fail",
-        "--retry",
-        "3",
-        "--retry-delay",
-        "2",
-        "--continue-at",
-        "-",
-        "-o",
-        str(tmp),
-        url,
-    ]
-    result = subprocess.run(command, text=True, capture_output=True)
-    if result.returncode != 0:
-        return path, False, result.stderr.strip() or result.stdout.strip()
-    tmp.replace(path)
-    if is_bad_fixture(path):
-        return path, False, "downloaded fixture is empty or still an LFS pointer"
-    return path, True, ""
+    token = os.environ.get("MOBILEGL_TRACE_FIXTURE_MIRROR_TOKEN", "")
+    errors = []
+    for mirror_base in mirror_bases:
+        url = f"{mirror_base.rstrip('/')}/{path.name}"
+        command = ["curl", "-L", "--fail", "--retry", "3", "--retry-delay", "2",
+                   "--continue-at", "-"]
+        if token:
+            command += ["--header", f"Authorization: token {token}"]
+        command += ["-o", str(tmp), url]
+        result = subprocess.run(command, text=True, capture_output=True)
+        if result.returncode != 0:
+            errors.append(f"{mirror_base}: {result.stderr.strip() or result.stdout.strip()}")
+            tmp.unlink(missing_ok=True)
+            continue
+        tmp.replace(path)
+        if is_bad_fixture(path):
+            errors.append(f"{mirror_base}: downloaded fixture is empty or still an LFS pointer")
+            continue
+        return path, True, ""
+    return path, False, "; ".join(errors)
 
 
-def hydrate_fixtures(cases, fetch, mirror_base, download_jobs):
+def hydrate_fixtures(cases, fetch, mirror_bases, download_jobs):
     required = []
     seen = set()
     for case in cases:
@@ -149,7 +155,7 @@ def hydrate_fixtures(cases, fetch, mirror_base, download_jobs):
     print(f"fixtures: downloading {len(required)} file(s) with {download_jobs} worker(s)", flush=True)
     failures = []
     with concurrent.futures.ThreadPoolExecutor(max_workers=download_jobs) as executor:
-        futures = [executor.submit(download_fixture, path, mirror_base) for path in required]
+        futures = [executor.submit(download_fixture, path, mirror_bases) for path in required]
         for future in concurrent.futures.as_completed(futures):
             path, ok, message = future.result()
             if ok:
@@ -434,7 +440,9 @@ def parse_args():
     parser.add_argument("--keep-results", action="store_true", help="Do not clear previous result/work roots.")
     parser.add_argument("--no-render", action="store_true", help="Do not render the HTML summary.")
     parser.add_argument("--fetch-fixtures", action=argparse.BooleanOptionalAction, default=True, help="Hydrate missing fixtures before running.")
-    parser.add_argument("--fixture-mirror-base", default=os.environ.get("MOBILEGL_TRACE_FIXTURE_MIRROR_BASE", DEFAULT_MIRROR_BASE))
+    parser.add_argument("--fixture-mirror-base", action="append", dest="fixture_mirror_bases",
+                        help="Fixture mirror base URL; repeatable, tried in order. "
+                             "Defaults to the hit.moe mirror then the miawa mirror.")
     parser.add_argument("--download-jobs", type=int, default=4, help="Parallel fixture download count.")
     parser.add_argument("--continue-after-fatal", action="store_true", help="Continue launching later cases after a fatal native-window replay.")
     return parser.parse_args()
@@ -463,7 +471,10 @@ def main():
         print(f"missing MobileGL library: {mobilegl_library}", file=sys.stderr)
         return 2
 
-    if not hydrate_fixtures(selected_cases, args.fetch_fixtures, args.fixture_mirror_base, max(1, args.download_jobs)):
+    mirror_bases = args.fixture_mirror_bases or [
+        b for b in os.environ.get("MOBILEGL_TRACE_FIXTURE_MIRROR_BASES", "").split() or []
+    ] or ([os.environ["MOBILEGL_TRACE_FIXTURE_MIRROR_BASE"]] if os.environ.get("MOBILEGL_TRACE_FIXTURE_MIRROR_BASE") else DEFAULT_MIRROR_BASES)
+    if not hydrate_fixtures(selected_cases, args.fetch_fixtures, mirror_bases, max(1, args.download_jobs)):
         return 2
 
     if not args.keep_results:

tools/trace_replay/skills/README.md

@@ -0,0 +1,20 @@
+# MobileGL trace-replay skills
+
+Task-focused skills for capturing, replaying, debugging, and authoring MobileGL
+apitrace fixtures. Each skill is a self-contained package:
+
+- `SKILL.md` — the skill (frontmatter `name` + `description`, then the body). The
+  directory name equals the frontmatter `name`.
+- `agents/openai.yaml` — OpenAI agent descriptor (`display_name`,
+  `short_description`, `default_prompt`).
+- `scripts/` and/or `references/` — bundled tooling and supporting docs, when the
+  skill has them.
+
+## Skills
+
+| Skill | What it does |
+| --- | --- |
+| [trace-fixture-authoring-on-android-fcl](trace-fixture-authoring-on-android-fcl/SKILL.md) | Capture an on-device Android apitrace from FCL's MobileGL renderers (DirectGLES / Magma / SimpleFPEWrapper), mark the defect frame, and pull `full.trace`. |
+| [renderdoc-debug-on-trace-replay](renderdoc-debug-on-trace-replay/SKILL.md) | Capture and validate an exact frame from a MobileGL retrace on a connected Android device with RenderDoc / rdc-cli. |
+| [mismatch-retrace-debugging](mismatch-retrace-debugging/SKILL.md) | Localize the first divergent render pass and draw call when a fixture replays correctly in a golden environment but renders differently on a target backend. |
+| [trace-fixture-authoring](trace-fixture-authoring/SKILL.md) | Author a deterministic trace-replay fixture — trim, golden, package under the size budget, register in `trace_cases.json`, and validate on Linux and Android. |

tools/trace_replay/skills/mismatch-retrace-debugging/SKILL.md

@@ -1,3 +1,8 @@
+---
+name: mismatch-retrace-debugging
+description: Localize the first divergent render pass and draw call when a MobileGL apitrace fixture replays correctly in a golden environment but renders differently under mobilegl_trace_replay, Android trace replay, or another backend. Use to binary-search pass/draw endpoints, diff GL state around the first bad call, and classify the fault as a vertex/VS, fragment/FS, or framebuffer/composition mismatch.
+---
+
 # Mismatch retrace debugging
 
 Use this when an apitrace fixture replays correctly on one environment but

tools/trace_replay/skills/mismatch-retrace-debugging/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "MobileGL Mismatch Retrace Debugging"
+  short_description: "Localize the first divergent draw in a mismatching MobileGL retrace"
+  default_prompt: "Use $mismatch-retrace-debugging to find the first divergent render pass and draw call in a MobileGL retrace mismatch."

tools/trace_replay/skills/renderdoc-debug-on-trace-replay/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: renderdoc-capture-trace-frame
+name: renderdoc-debug-on-trace-replay
 description: Capture and validate an exact frame from a MobileGL apitrace retrace on a connected Android device with RenderDoc/rdc-cli. Use for DirectVulkan or DirectGLES trace replay, mapping a target API call to an eglSwapBuffers frame, producing an .rdc plus a complete command manifest, checking capture stability, or troubleshooting Android TargetControl timing and replay failures.
 ---
 
@@ -19,20 +19,20 @@ adb -s SERIAL shell pm path top.mobilegl.plugin.trace
 rdc doctor
 ```
 
-4. Pass the unpacked `trace.trace`, its golden PNG, the fixture target call, backend, and output path to `tools/trace_replay/capture_android_retrace.py`.
+4. Pass the unpacked `trace.trace`, its golden PNG, the fixture target call, backend, and output path to `tools/trace_replay/skills/renderdoc-debug-on-trace-replay/scripts/capture_android_retrace.py`.
 
 ## Capture
 
 Let the tool infer the zero-based target swap from `eglSwapBuffers` calls:
 
 ```powershell
-python tools/trace_replay/capture_android_retrace.py --trace .trace-work/case/trace.trace --golden tools/trace_replay/fixtures/case.0002667619.png --target-call 2667619 --backend DirectVulkan --output captures/case-vulkan.rdc --serial SERIAL --json
+python tools/trace_replay/skills/renderdoc-debug-on-trace-replay/scripts/capture_android_retrace.py --trace .trace-work/case/trace.trace --golden tools/trace_replay/fixtures/case.0002667619.png --target-call 2667619 --backend DirectVulkan --output captures/case-vulkan.rdc --serial SERIAL --json
 ```
 
 Change only the backend and output for GLES:
 
 ```powershell
-python tools/trace_replay/capture_android_retrace.py --trace .trace-work/case/trace.trace --golden tools/trace_replay/fixtures/case.0002667619.png --target-call 2667619 --backend DirectGLES --output captures/case-gles.rdc --serial SERIAL --json
+python tools/trace_replay/skills/renderdoc-debug-on-trace-replay/scripts/capture_android_retrace.py --trace .trace-work/case/trace.trace --golden tools/trace_replay/fixtures/case.0002667619.png --target-call 2667619 --backend DirectGLES --output captures/case-gles.rdc --serial SERIAL --json
 ```
 
 Use `--target-swap N` when the mapping is already known. Use `--capture-frame N` only to override the backend rule deliberately.

tools/trace_replay/skills/renderdoc-debug-on-trace-replay/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "RenderDoc Debug on Trace Replay"
+  short_description: "Capture and debug an exact MobileGL trace-replay frame in RenderDoc"
+  default_prompt: "Use $renderdoc-debug-on-trace-replay to capture and validate an exact Android trace-replay frame in RenderDoc."

tools/trace_replay/skills/renderdoc-debug-on-trace-replay/references/android-renderdoc-troubleshooting.md

@@ -2,7 +2,7 @@
 
 ## TargetControl timing
 
-- Start `tools/trace_replay/queue_android_frame.py` before launching `TraceReplayActivity`. A fast retrace can pass the requested frame before a late client connects.
+- Start `tools/trace_replay/skills/renderdoc-debug-on-trace-replay/scripts/queue_android_frame.py` before launching `TraceReplayActivity`. A fast retrace can pass the requested frame before a late client connects.
 - Keep TargetControl connected until `NewCapture` arrives. A queued request alone is not sufficient evidence that the RDC finished.
 - Drain the asynchronous `RegisterAPI` and `CapturableWindowCount` messages before calling `QueueCapture`; otherwise `NewCapture` can be lost.
 - Do not use the daemon-backed `rdc script` path for a capture that may exceed 30 seconds. Its outer RPC times out even when the device later writes a valid RDC. The repository helper imports the RenderDoc module discovered by `rdc` directly and has an independent capture timeout.

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/SKILL.md

@@ -0,0 +1,137 @@
+---
+name: trace-fixture-authoring-on-android-fcl
+description: Capture an on-device Android apitrace from FCL's MobileGL renderers. Use when preparing a reproducible MobileGL DirectGLES, Magma (DirectVulkan), or SimpleFPEWrapper rendering trace, marking the frame of a visual defect, pulling the resulting full.trace, or turning a device capture into a replay fixture.
+---
+
+# MobileGL Android trace capture
+
+## Overview
+
+Use FCL's Android `egltrace.so` wrapper, not Perfetto. When enabled before
+launch, it records the complete EGL/GL call stream to `full.trace`. The game's
+**MobileGL Trace → Capture** control marks the next swap frame in
+`capture-result.json`; it does not start or stop recording and does not produce
+a one-frame trace by itself.
+
+Run commands from the FoldCraftLauncher repository root:
+
+```sh
+export REPO="$PWD"
+export CAPTURE="$REPO/MobileGL/tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts"
+export SERIAL=<adb-device-serial> # omit --serial only if exactly one device is attached
+```
+
+The capture scripts are bundled inside this skill under `scripts/`; they
+auto-detect the FoldCraftLauncher repository root from their own location, so
+`--repo` only needs to be passed for a non-standard checkout layout.
+
+## Prerequisites
+
+- Use an FCL build containing `MobileGLTraceCapture` and the in-game Capture
+  menu entry.
+- Select one of these renderers: MobileGL (DirectGLES), MobileGL Magma
+  (DirectVulkan), or SimpleFPEWrapper. MobileGlues is not supported by this
+  capture wrapper.
+- Install `adb` and make it available on `PATH`; authorize USB debugging.
+- Build the wrapper with Android NDK, CMake, Ninja, Python 3, and the checked
+  out in-tree `MobileGL/3rdparty/apitrace` submodule.
+
+Confirm the attached device and ABI before building. The wrapper ABI must match
+the device process ABI.
+
+```sh
+adb devices -l
+adb -s "$SERIAL" shell getprop ro.product.cpu.abi
+```
+
+Use `arm64-v8a` for the usual `arm64-v8a` result; use the matching NDK ABI for
+other devices.
+
+## Build and install the wrapper
+
+Build once per ABI or after changing apitrace/wrapper sources:
+
+```sh
+python3 "$CAPTURE/build_android_egltrace.py" --abi arm64-v8a
+```
+
+This generates `egltrace.so` under the skill's `scripts/out/` directory. Push
+it and write FCL's enable sentinel before launching the game:
+
+```sh
+python3 "$CAPTURE/adb_capture.py" --serial "$SERIAL" install-wrapper
+python3 "$CAPTURE/adb_capture.py" --serial "$SERIAL" enable
+```
+
+The device-side control directory is `/sdcard/FCL/mobilegl-trace`. FCL copies
+the shared `egltrace.so` into its private files directory at launch, replaces
+the renderer's EGL library with it, and forwards to the real MobileGL library.
+
+## Capture a reproduction
+
+1. Start FCL after the wrapper and enable sentinel are in place. Select the
+   intended supported MobileGL renderer and launch the game.
+2. Trace mode forces the game to `854x480`; account for that when reproducing
+   and comparing output.
+3. Reproduce the issue. Start close to the target scene because tracing starts
+   when the game launches and trace files can grow rapidly.
+4. At the desired visual state, open FCL's right-side game menu and press
+   **MobileGL Trace → Capture**. Let at least one frame present afterward.
+5. Exit the game cleanly, then pull the latest session:
+
+```sh
+python3 "$CAPTURE/adb_capture.py" --serial "$SERIAL" pull-latest
+```
+
+The default local result is:
+
+```text
+.trace-work/pulled-mobilegl-captures/capture-YYYYMMDD-HHMMSS-<renderer>/
+  full.trace
+  capture-status.json
+  capture-result.json
+```
+
+`capture-result.json` must show `"status": "captured"`. Its `targetFrame`
+is the one-based swap count used by `gltrim`; `zeroBasedFrame` is included for
+tools that use zero-based indexing.
+
+## Diagnose setup failures
+
+Inspect the active device session directly:
+
+```sh
+adb -s "$SERIAL" shell cat /sdcard/FCL/mobilegl-trace/latest-session.txt
+adb -s "$SERIAL" shell ls -lh /sdcard/FCL/mobilegl-trace
+adb -s "$SERIAL" shell cat /sdcard/FCL/mobilegl-trace/capture-*/capture-status.json
+adb -s "$SERIAL" shell cat /sdcard/FCL/mobilegl-trace/capture-*/capture-result.json
+```
+
+If `capture-status.json` reports a missing `egltrace.so`, rebuild/push the
+correct ABI and relaunch. If `capture-result.json` is absent, Capture was
+pressed without an active trace session, or no subsequent `eglSwapBuffers`
+occurred. The menu button itself only writes `capture-once.request`.
+
+Disable tracing when finished; otherwise the next supported MobileGL launch
+will trace again:
+
+```sh
+python3 "$CAPTURE/adb_capture.py" --serial "$SERIAL" disable
+```
+
+## Create a replay fixture (optional)
+
+Keep the raw `full.trace` until replay validation succeeds. To frame-trim and
+package the marked frame for MobileGL trace replay, use the existing helper:
+
+```sh
+python3 "$CAPTURE/package_capture_fixture.py" \
+  --serial "$SERIAL" \
+  --case <case-name> \
+  --apitrace <path-to-in-tree-apitrace>
+```
+
+It pulls the latest capture if necessary, uses `capture-result.json` to select
+the frame, runs `apitrace gltrim`, creates a golden image, and enforces the
+fixture archive-size limit. Follow `../trace-fixture-authoring/SKILL.md` for
+deterministic scene setup, verification, and registry changes.

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Trace Fixture Authoring on Android (FCL)"
+  short_description: "Capture and package a MobileGL trace fixture on Android FCL"
+  default_prompt: "Use $trace-fixture-authoring-on-android-fcl to capture a MobileGL trace on my Android device and package it into a replay fixture."

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/.gitignore

@@ -0,0 +1,3 @@
+# Build output produced by build_android_egltrace.py (ABI-specific, regenerated).
+out/
+__pycache__/

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/adb_capture.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+import argparse
+import subprocess
+from pathlib import Path
+
+
+REMOTE_ROOT = "/sdcard/FCL/mobilegl-trace"
+SCRIPT_DIR = Path(__file__).resolve().parent
+DEFAULT_WRAPPER = SCRIPT_DIR / "out" / "egltrace.so"
+
+
+def adb(serial, args):
+    cmd = ["adb"]
+    if serial:
+        cmd += ["-s", serial]
+    cmd += args
+    print("+", " ".join(cmd), flush=True)
+    subprocess.run(cmd, check=True)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--serial")
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    install = sub.add_parser("install-wrapper")
+    install.add_argument("--wrapper", default=str(DEFAULT_WRAPPER))
+
+    sub.add_parser("enable")
+    sub.add_parser("disable")
+    sub.add_parser("capture-once")
+
+    pull = sub.add_parser("pull-latest")
+    pull.add_argument("--output", default=".trace-work/pulled-mobilegl-captures")
+
+    args = parser.parse_args()
+    serial = args.serial
+
+    if args.command == "install-wrapper":
+        wrapper = Path(args.wrapper)
+        if not wrapper.exists():
+            raise SystemExit(f"missing wrapper: {wrapper}")
+        adb(serial, ["shell", "mkdir", "-p", REMOTE_ROOT])
+        adb(serial, ["push", str(wrapper), f"{REMOTE_ROOT}/egltrace.so"])
+    elif args.command == "enable":
+        adb(serial, ["shell", "mkdir", "-p", REMOTE_ROOT])
+        adb(serial, ["shell", f"printf enabled > {REMOTE_ROOT}/enable"])
+    elif args.command == "disable":
+        adb(serial, ["shell", "rm", "-f", f"{REMOTE_ROOT}/enable"])
+    elif args.command == "capture-once":
+        adb(serial, ["shell", "mkdir", "-p", REMOTE_ROOT])
+        adb(serial, ["shell", f"date +%s%3N > {REMOTE_ROOT}/capture-once.request"])
+    elif args.command == "pull-latest":
+        tmp = subprocess.check_output((["adb"] + (["-s", serial] if serial else []) +
+                                      ["shell", "cat", f"{REMOTE_ROOT}/latest-session.txt"]),
+                                      text=True, encoding="utf-8", errors="replace").strip()
+        if not tmp:
+            raise SystemExit("no latest-session.txt on device")
+        output = Path(args.output)
+        output.mkdir(parents=True, exist_ok=True)
+        adb(serial, ["pull", tmp, str(output / Path(tmp).name)])
+
+
+if __name__ == "__main__":
+    main()

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/android_egltrace/CMakeLists.txt

@@ -0,0 +1,191 @@
+cmake_minimum_required(VERSION 3.22.1)
+
+project(mobilegl_android_egltrace)
+
+set(APITRACE_ROOT "" CACHE PATH "Path to apitrace source tree")
+set(PATCHED_EGLTRACE_CPP "" CACHE FILEPATH "Generated and patched egltrace.cpp")
+set(PATCHED_GLPROC_EGL_CPP "" CACHE FILEPATH "Patched glproc_egl.cpp")
+if(NOT EXISTS "${APITRACE_ROOT}/wrappers")
+    message(FATAL_ERROR "APITRACE_ROOT must point to apitrace")
+endif()
+if(NOT EXISTS "${PATCHED_EGLTRACE_CPP}")
+    message(FATAL_ERROR "PATCHED_EGLTRACE_CPP is required")
+endif()
+if(NOT EXISTS "${PATCHED_GLPROC_EGL_CPP}")
+    message(FATAL_ERROR "PATCHED_GLPROC_EGL_CPP is required")
+endif()
+
+set(APITRACE_BINARY_DIR "${CMAKE_CURRENT_BINARY_DIR}/apitrace")
+set(APITRACE_VERSION "mobilegl-capture")
+
+find_package(Python3 REQUIRED)
+find_package(Threads REQUIRED)
+
+include("${APITRACE_ROOT}/cmake/ConvenienceLibrary.cmake")
+
+set(BUILD_TESTING OFF CACHE BOOL "" FORCE)
+set(ENABLE_STATIC_SNAPPY ON CACHE BOOL "" FORCE)
+set(DOC_INSTALL_DIR "doc" CACHE PATH "" FORCE)
+set(HAVE_X86 OFF CACHE BOOL "" FORCE)
+set(ZLIB_FOUND OFF CACHE BOOL "" FORCE)
+set(PNG_FOUND OFF CACHE BOOL "" FORCE)
+set(Snappy_FOUND OFF CACHE BOOL "" FORCE)
+set(BROTLIDEC_FOUND OFF CACHE BOOL "" FORCE)
+set(BROTLIENC_FOUND OFF CACHE BOOL "" FORCE)
+set(ZSTD_FOUND OFF CACHE BOOL "" FORCE)
+set(CMAKE_EXECUTABLE_FORMAT "MobileGLAndroid" CACHE INTERNAL "" FORCE)
+
+add_custom_target(check)
+add_subdirectory("${APITRACE_ROOT}/thirdparty" "${APITRACE_BINARY_DIR}/thirdparty")
+
+set(APITRACE_GENERATED_DIR "${APITRACE_BINARY_DIR}/generated")
+file(MAKE_DIRECTORY "${APITRACE_GENERATED_DIR}")
+configure_file("${APITRACE_ROOT}/version.h.in" "${APITRACE_GENERATED_DIR}/version.h" @ONLY)
+
+add_custom_command(
+        OUTPUT
+        "${APITRACE_GENERATED_DIR}/glproc.hpp"
+        "${APITRACE_GENERATED_DIR}/glproc.cpp"
+        COMMAND ${Python3_EXECUTABLE}
+        "${APITRACE_ROOT}/dispatch/glproc.py"
+        "${APITRACE_GENERATED_DIR}/glproc.hpp"
+        "${APITRACE_GENERATED_DIR}/glproc.cpp"
+        DEPENDS
+        "${APITRACE_ROOT}/dispatch/glproc.py"
+        "${APITRACE_ROOT}/dispatch/dispatch.py"
+        "${APITRACE_ROOT}/specs/wglapi.py"
+        "${APITRACE_ROOT}/specs/glxapi.py"
+        "${APITRACE_ROOT}/specs/cglapi.py"
+        "${APITRACE_ROOT}/specs/eglapi.py"
+        "${APITRACE_ROOT}/specs/glapi.py"
+        "${APITRACE_ROOT}/specs/gltypes.py"
+        "${APITRACE_ROOT}/specs/stdapi.py")
+
+add_library(apitrace_os STATIC
+        "${APITRACE_ROOT}/lib/os/os_backtrace.cpp"
+        "${APITRACE_ROOT}/lib/os/os_crtdbg.cpp"
+        "${APITRACE_ROOT}/lib/os/os_posix.cpp")
+target_include_directories(apitrace_os PUBLIC
+        "${APITRACE_ROOT}/compat"
+        "${APITRACE_ROOT}/thirdparty"
+        "${APITRACE_ROOT}/lib/os"
+        "${APITRACE_ROOT}/lib/trace")
+target_link_libraries(apitrace_os PUBLIC Threads::Threads)
+
+add_library(glproc STATIC
+        "${APITRACE_GENERATED_DIR}/glproc.cpp"
+        "${PATCHED_GLPROC_EGL_CPP}")
+target_include_directories(glproc PUBLIC
+        "${APITRACE_GENERATED_DIR}"
+        "${APITRACE_ROOT}/wrappers"
+        "${APITRACE_ROOT}/dispatch"
+        "${APITRACE_ROOT}/lib/os"
+        "${APITRACE_ROOT}/thirdparty/khronos")
+target_link_libraries(glproc PUBLIC apitrace_os dl)
+
+add_library(highlight STATIC "${APITRACE_ROOT}/lib/highlight/highlight.cpp")
+target_include_directories(highlight PUBLIC "${APITRACE_ROOT}/lib/highlight")
+
+add_library(guids STATIC "${APITRACE_ROOT}/lib/guids/guids.cpp")
+target_include_directories(guids PUBLIC
+        "${APITRACE_ROOT}/lib/guids"
+        "${APITRACE_ROOT}/lib/os")
+
+add_library(common STATIC
+        "${APITRACE_ROOT}/lib/trace/trace_callset.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_dump.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_fast_callset.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_read.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_zlib.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_brotli.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_snappy.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_zstd.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_file_zstd_seekable.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_model.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_option.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_ostream_snappy.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_ostream_zlib.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_ostream_zstd.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_parser.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_parser_flags.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_parser_loop.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_profiler.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_writer.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_writer_local.cpp"
+        "${APITRACE_ROOT}/lib/trace/trace_writer_model.cpp")
+target_include_directories(common PUBLIC
+        "${APITRACE_ROOT}/compat"
+        "${APITRACE_ROOT}/thirdparty"
+        "${APITRACE_ROOT}/lib/guids"
+        "${APITRACE_ROOT}/lib/highlight"
+        "${APITRACE_ROOT}/lib/os"
+        "${APITRACE_ROOT}/lib/trace"
+        "${APITRACE_ROOT}/lib/ubjson")
+target_link_libraries(common PUBLIC
+        guids
+        highlight
+        apitrace_os
+        Snappy::snappy
+        ZLIB::ZLIB
+        PkgConfig::BROTLIDEC
+        PkgConfig::ZSTD
+        zstd_seekable)
+
+add_convenience_library(trace
+        "${APITRACE_ROOT}/wrappers/memtrace.hpp"
+        "${APITRACE_ROOT}/wrappers/memtrace.cpp")
+target_include_directories(trace PUBLIC
+        "${APITRACE_ROOT}/thirdparty/crc32c")
+target_link_libraries(trace
+        common
+        guids
+        crc32c)
+
+add_library(glhelpers STATIC
+        "${APITRACE_ROOT}/helpers/glfeatures.cpp"
+        "${APITRACE_ROOT}/helpers/eglsize.cpp")
+target_include_directories(glhelpers PUBLIC
+        "${APITRACE_GENERATED_DIR}"
+        "${APITRACE_ROOT}/dispatch"
+        "${APITRACE_ROOT}/helpers"
+        "${APITRACE_ROOT}/lib/os"
+        "${APITRACE_ROOT}/thirdparty/khronos")
+target_link_libraries(glhelpers PUBLIC glproc apitrace_os)
+
+add_convenience_library(gltrace_common
+        "${APITRACE_ROOT}/wrappers/glcaps.cpp"
+        "${APITRACE_ROOT}/wrappers/config.cpp"
+        "${APITRACE_ROOT}/wrappers/gltrace_arrays.cpp"
+        "${APITRACE_ROOT}/wrappers/gltrace_state.cpp"
+        "${APITRACE_ROOT}/wrappers/glmemshadow.hpp"
+        "${APITRACE_ROOT}/wrappers/glmemshadow.cpp"
+        "${APITRACE_ROOT}/wrappers/gltrace_unpack_compressed.hpp"
+        "${APITRACE_ROOT}/wrappers/gltrace_unpack_compressed.cpp")
+add_dependencies(gltrace_common glproc)
+target_include_directories(gltrace_common PUBLIC
+        "${APITRACE_ROOT}/wrappers")
+target_link_libraries(gltrace_common
+        glhelpers
+        trace)
+
+add_library(egltrace SHARED
+        "${PATCHED_EGLTRACE_CPP}"
+        "${APITRACE_ROOT}/wrappers/dlsym.cpp"
+        "${PATCHED_GLPROC_EGL_CPP}")
+add_dependencies(egltrace glproc)
+set_target_properties(egltrace PROPERTIES PREFIX "")
+target_compile_definitions(egltrace PRIVATE -DEGLTRACE=1)
+target_include_directories(egltrace PRIVATE
+        "${APITRACE_ROOT}/wrappers"
+        "${APITRACE_GENERATED_DIR}"
+        "${APITRACE_ROOT}/helpers"
+        "${APITRACE_ROOT}/dispatch"
+        "${APITRACE_ROOT}/lib/os"
+        "${APITRACE_ROOT}/lib/trace"
+        "${APITRACE_ROOT}/thirdparty/khronos")
+target_link_libraries(egltrace
+        gltrace_common
+        glproc
+        Threads::Threads
+        dl)

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/build_android_egltrace.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+import argparse
+import os
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+# This script is bundled inside the trace-fixture-authoring-on-android-fcl skill at
+# <FCL>/MobileGL/tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/.
+SCRIPT_DIR = Path(__file__).resolve().parent
+DEFAULT_REPO = SCRIPT_DIR.parents[5]  # -> FoldCraftLauncher repo root
+DEFAULT_OUTPUT = SCRIPT_DIR / "out" / "egltrace.so"
+
+
+def run(cmd, cwd=None):
+    print("+", " ".join(str(part) for part in cmd), flush=True)
+    subprocess.run(cmd, cwd=cwd, check=True)
+
+
+def find_ndk(repo):
+    for key in ("ANDROID_NDK_HOME", "ANDROID_NDK_ROOT"):
+        value = os.environ.get(key)
+        if value:
+            return Path(value)
+    for key in ("ANDROID_HOME", "ANDROID_SDK_ROOT"):
+        value = os.environ.get(key)
+        if value:
+            ndk_root = Path(value) / "ndk"
+            if ndk_root.exists():
+                versions = sorted([p for p in ndk_root.iterdir() if p.is_dir()])
+                if versions:
+                    return versions[-1]
+    local = repo / "local.properties"
+    if local.exists():
+        sdk = None
+        ndk = None
+        for line in local.read_text(encoding="utf-8", errors="ignore").splitlines():
+            if line.startswith("sdk.dir="):
+                sdk = Path(line.split("=", 1)[1].replace("\\:", ":"))
+            if line.startswith("ndk.dir="):
+                ndk = Path(line.split("=", 1)[1].replace("\\:", ":"))
+        if ndk:
+            return ndk
+        if sdk:
+            ndk_root = sdk / "ndk"
+            if ndk_root.exists():
+                versions = sorted([p for p in ndk_root.iterdir() if p.is_dir()])
+                if versions:
+                    return versions[-1]
+    raise SystemExit("Android NDK not found; set ANDROID_NDK_HOME or local.properties sdk.dir/ndk.dir")
+
+
+def generate_and_patch(repo, build_dir):
+    wrapper_dir = repo / "MobileGL" / "3rdparty" / "apitrace" / "wrappers"
+    generated = build_dir / "patched" / "egltrace.cpp"
+    generated.parent.mkdir(parents=True, exist_ok=True)
+    with generated.open("w", encoding="utf-8", newline="\n") as out:
+        subprocess.run([sys.executable, str(wrapper_dir / "egltrace.py")], cwd=wrapper_dir, stdout=out, check=True)
+
+    text = generated.read_text(encoding="utf-8")
+    helper = r'''
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/stat.h>
+#include <time.h>
+#include <unistd.h>
+
+static unsigned long long mobilegl_capture_swap_count = 0;
+
+static long long mobilegl_capture_time_ms(void) {
+    struct timespec ts;
+    clock_gettime(CLOCK_REALTIME, &ts);
+    return (long long) ts.tv_sec * 1000LL + ts.tv_nsec / 1000000LL;
+}
+
+static int mobilegl_capture_exists(const char *path) {
+    return path != NULL && path[0] != '\0' && access(path, F_OK) == 0;
+}
+
+static void mobilegl_capture_record_request(void) {
+    ++mobilegl_capture_swap_count;
+    const char *request = getenv("MOBILEGL_TRACE_CAPTURE_REQUEST_FILE");
+    const char *output = getenv("MOBILEGL_TRACE_CAPTURE_FRAME_FILE");
+    if (!mobilegl_capture_exists(request) || output == NULL || output[0] == '\0') {
+        return;
+    }
+    unlink(request);
+    FILE *file = fopen(output, "w");
+    if (file == NULL) {
+        return;
+    }
+    const char *trace_file = getenv("TRACE_FILE");
+    fprintf(file,
+            "{\n"
+            "  \"status\": \"captured\",\n"
+            "  \"swapCount\": %llu,\n"
+            "  \"targetFrame\": %llu,\n"
+            "  \"zeroBasedFrame\": %llu,\n"
+            "  \"capturedAtMs\": %lld,\n"
+            "  \"traceFile\": \"%s\"\n"
+            "}\n",
+            mobilegl_capture_swap_count,
+            mobilegl_capture_swap_count,
+            mobilegl_capture_swap_count == 0 ? 0 : mobilegl_capture_swap_count - 1,
+            mobilegl_capture_time_ms(),
+            trace_file == NULL ? "" : trace_file);
+    fclose(file);
+}
+'''
+    insert_at = text.find("#include")
+    if insert_at < 0:
+        raise SystemExit("generated egltrace.cpp has no include block")
+    next_block = text.find("\n\n", insert_at)
+    text = text[:next_block] + "\n" + helper + text[next_block:]
+
+    needle = "EGLBoolean EGLAPIENTRY eglSwapBuffers(EGLDisplay dpy, EGLSurface surface)"
+    start = text.find(needle)
+    if start < 0:
+        raise SystemExit("generated egltrace.cpp has no eglSwapBuffers wrapper to patch")
+    brace = text.find("{", start)
+    if brace < 0:
+        raise SystemExit("eglSwapBuffers wrapper has no function body")
+    text = text[:brace + 1] + "\n    mobilegl_capture_record_request();" + text[brace + 1:]
+    generated.write_text(text, encoding="utf-8", newline="\n")
+    return generated
+
+
+def patch_glproc_egl(repo, build_dir):
+    source = repo / "MobileGL" / "3rdparty" / "apitrace" / "wrappers" / "glproc_egl.cpp"
+    patched = build_dir / "patched" / "glproc_egl.cpp"
+    patched.parent.mkdir(parents=True, exist_ok=True)
+    text = source.read_text(encoding="utf-8")
+    text = text.replace('#include "dlopen.hpp"\n', '#include "dlopen.hpp"\n#include <stdlib.h>\n')
+    needle = """void *
+_getPublicProcAddress(const char *procName)
+{
+    void *proc;
+
+"""
+    replacement = """void *
+_getPublicProcAddress(const char *procName)
+{
+    void *proc;
+
+    static void *traceLibGL = NULL;
+    static bool triedTraceLibGL = false;
+    if (!triedTraceLibGL) {
+        triedTraceLibGL = true;
+        const char *traceLibGLName = getenv("TRACE_LIBGL");
+        if (traceLibGLName && traceLibGLName[0]) {
+            traceLibGL = _dlopen(traceLibGLName, RTLD_GLOBAL | RTLD_LAZY | RTLD_DEEPBIND);
+        }
+    }
+    if (traceLibGL) {
+        proc = dlsym(traceLibGL, procName);
+        if (proc) {
+            return proc;
+        }
+    }
+
+"""
+    if needle not in text:
+        raise SystemExit("glproc_egl.cpp patch point not found")
+    text = text.replace(needle, replacement, 1)
+    patched.write_text(text, encoding="utf-8", newline="\n")
+    return patched
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--repo", default=str(DEFAULT_REPO), help="FoldCraftLauncher repo root")
+    parser.add_argument("--abi", default="arm64-v8a")
+    parser.add_argument("--android-platform", default="android-23")
+    parser.add_argument("--build-dir", default=".trace-work/build-android-egltrace")
+    parser.add_argument("--output", default=str(DEFAULT_OUTPUT))
+    args = parser.parse_args()
+
+    repo = Path(args.repo).resolve()
+    ndk = find_ndk(repo)
+    build_dir = (repo / args.build_dir / args.abi).resolve()
+    apitrace = repo / "MobileGL" / "3rdparty" / "apitrace"
+    source_dir = SCRIPT_DIR / "android_egltrace"
+    toolchain = ndk / "build" / "cmake" / "android.toolchain.cmake"
+    if not apitrace.exists():
+        raise SystemExit(f"missing apitrace checkout: {apitrace}")
+    if not source_dir.exists():
+        raise SystemExit(f"missing wrapper CMake project: {source_dir}")
+    if not toolchain.exists():
+        raise SystemExit(f"missing Android toolchain: {toolchain}")
+    cache = build_dir / "CMakeCache.txt"
+    if cache.exists() and "CMAKE_GENERATOR:INTERNAL=Ninja" not in cache.read_text(encoding="utf-8", errors="ignore"):
+        shutil.rmtree(build_dir)
+    elif cache.exists() and str(source_dir).replace("\\", "/") not in cache.read_text(encoding="utf-8", errors="ignore").replace("\\", "/"):
+        shutil.rmtree(build_dir)
+
+    ninja = shutil.which("ninja")
+    if ninja is None:
+        cmake_ninjas = sorted((Path(os.environ.get("ANDROID_HOME", "")) / "cmake").glob("*/bin/ninja.exe"))
+        ninja = str(cmake_ninjas[-1]) if cmake_ninjas else None
+    if ninja is None:
+        raise SystemExit("ninja not found; install Ninja or Android SDK CMake")
+
+    patched_egltrace = generate_and_patch(repo, build_dir)
+    patched_glproc_egl = patch_glproc_egl(repo, build_dir)
+
+    run([
+        "cmake", "-G", "Ninja", "-S", str(source_dir), "-B", str(build_dir),
+        "-DCMAKE_BUILD_TYPE=Release",
+        f"-DCMAKE_TOOLCHAIN_FILE={toolchain}",
+        f"-DCMAKE_MAKE_PROGRAM={ninja}",
+        f"-DANDROID_ABI={args.abi}",
+        f"-DANDROID_PLATFORM={args.android_platform}",
+        f"-DAPITRACE_ROOT={apitrace}",
+        f"-DPATCHED_EGLTRACE_CPP={patched_egltrace}",
+        f"-DPATCHED_GLPROC_EGL_CPP={patched_glproc_egl}",
+    ])
+    run(["cmake", "--build", str(build_dir), "--target", "egltrace", "--parallel"])
+
+    output = Path(args.output)
+    if not output.is_absolute():
+        output = repo / output
+    output = output.resolve()
+    output.parent.mkdir(parents=True, exist_ok=True)
+    candidates = list(build_dir.rglob("egltrace.so"))
+    if not candidates:
+        raise SystemExit("egltrace.so was not produced")
+    shutil.copy2(candidates[0], output)
+    print(output)
+
+
+if __name__ == "__main__":
+    main()

tools/trace_replay/skills/trace-fixture-authoring-on-android-fcl/scripts/package_capture_fixture.py

@@ -0,0 +1,168 @@
+#!/usr/bin/env python3
+import argparse
+import json
+import re
+import shutil
+import subprocess
+import tarfile
+from pathlib import Path
+
+DEFAULT_MAX_ARCHIVE_BYTES = 20 * 1024 * 1024
+# Bundled under the skill at .../skills/trace-fixture-authoring-on-android-fcl/scripts/.
+SCRIPT_DIR = Path(__file__).resolve().parent
+DEFAULT_REPO = SCRIPT_DIR.parents[5]  # -> FoldCraftLauncher repo root
+
+
+def run(cmd, cwd=None, capture=False):
+    print("+", " ".join(str(part) for part in cmd), flush=True)
+    if capture:
+        return subprocess.check_output(cmd, cwd=cwd, text=True, encoding="utf-8", errors="replace")
+    subprocess.run(cmd, cwd=cwd, check=True)
+    return ""
+
+
+def adb(args, serial=None):
+    cmd = ["adb"]
+    if serial:
+        cmd += ["-s", serial]
+    cmd += args
+    return run(cmd, capture=True)
+
+
+def pull_latest(serial, dest):
+    latest = adb(["shell", "cat", "/sdcard/FCL/mobilegl-trace/latest-session.txt"], serial).strip()
+    if not latest:
+        raise SystemExit("device has no /sdcard/FCL/mobilegl-trace/latest-session.txt")
+    dest.mkdir(parents=True, exist_ok=True)
+    local = dest / Path(latest).name
+    if local.exists():
+        shutil.rmtree(local)
+    adb(["pull", latest, str(local)], serial)
+    return local
+
+
+def choose_target_frame(capture_dir, explicit_frame):
+    if explicit_frame is not None:
+        return explicit_frame
+    result = capture_dir / "capture-result.json"
+    if not result.exists():
+        raise SystemExit(f"missing {result}; press the FCL capture button or create capture-once.request first")
+    data = json.loads(result.read_text(encoding="utf-8"))
+    if "targetFrame" not in data:
+        raise SystemExit(f"{result} has no targetFrame")
+    return int(data["targetFrame"])
+
+
+def choose_snapshot(golden_dir):
+    pngs = sorted(golden_dir.glob("*.png"))
+    if not pngs:
+        raise SystemExit(f"no snapshots produced in {golden_dir}")
+    def call_no(path):
+        match = re.search(r"\.(\d+)\.png$", path.name)
+        return int(match.group(1)) if match else -1
+    return max(pngs, key=call_no)
+
+
+def choose_target_call(explicit_call, golden):
+    if explicit_call is not None:
+        return explicit_call
+    if golden is None:
+        return None
+    match = re.search(r"\.(\d+)\.png$", golden.name)
+    return int(match.group(1)) if match else None
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--repo", default=str(DEFAULT_REPO), help="FoldCraftLauncher repo root")
+    parser.add_argument("--serial", help="adb serial; when set, pull latest capture from device")
+    parser.add_argument("--capture-dir", help="local capture directory; defaults to pulled latest")
+    parser.add_argument("--pull-root", default=".trace-work/pulled-mobilegl-captures")
+    parser.add_argument("--case", required=True)
+    parser.add_argument("--target-frame", type=int)
+    parser.add_argument("--target-call", type=int)
+    parser.add_argument("--golden", help="existing golden PNG, normally produced by Android replay")
+    parser.add_argument("--skip-desktop-golden", action="store_true",
+                        help="skip apitrace replay --headless; requires --golden and --target-call")
+    parser.add_argument("--apitrace", default="apitrace")
+    parser.add_argument("--fixtures-dir", default="MobileGL/tools/trace_replay/fixtures")
+    parser.add_argument("--width", type=int, default=854)
+    parser.add_argument("--height", type=int, default=480)
+    parser.add_argument("--ssim-threshold", default="0.99")
+    parser.add_argument("--max-archive-bytes", type=int, default=DEFAULT_MAX_ARCHIVE_BYTES)
+    args = parser.parse_args()
+
+    repo = Path(args.repo).resolve()
+    if args.capture_dir:
+        capture_dir = Path(args.capture_dir).resolve()
+    elif args.serial:
+        capture_dir = pull_latest(args.serial, repo / args.pull_root)
+    else:
+        raise SystemExit("pass --capture-dir or --serial")
+
+    full_trace = capture_dir / "full.trace"
+    if not full_trace.exists():
+        raise SystemExit(f"missing trace: {full_trace}")
+    target_frame = choose_target_frame(capture_dir, args.target_frame)
+    work = capture_dir / "fixture-work"
+    if work.exists():
+        shutil.rmtree(work)
+    work.mkdir(parents=True)
+
+    frames_txt = work / "frames.txt"
+    frames_txt.write_text(run([args.apitrace, "dump", "--calls=frame", str(full_trace)], capture=True), encoding="utf-8")
+
+    trimmed = work / "trace.trace"
+    run([args.apitrace, "gltrim", "-f", str(target_frame), "--output", str(trimmed), str(full_trace)])
+
+    supplied_golden = Path(args.golden).resolve() if args.golden else None
+    target_call = choose_target_call(args.target_call, supplied_golden)
+    if args.skip_desktop_golden:
+        if supplied_golden is None or target_call is None:
+            raise SystemExit("--skip-desktop-golden requires --golden and --target-call")
+        golden = supplied_golden
+    else:
+        golden_dir = work / "golden"
+        golden_dir.mkdir()
+        prefix = golden_dir / f"{args.case}."
+        run([args.apitrace, "replay", "--headless", "--snapshot-prefix", str(prefix), "--call-nos", str(trimmed)])
+        golden = choose_snapshot(golden_dir)
+        target_call = choose_target_call(args.target_call, golden)
+        if target_call is None:
+            raise SystemExit(f"cannot infer target call from {golden}")
+
+    fixtures = (repo / args.fixtures_dir).resolve()
+    fixtures.mkdir(parents=True, exist_ok=True)
+    archive_root = work / "archive"
+    archive_root.mkdir()
+    shutil.copy2(trimmed, archive_root / "trace.trace")
+    tgz = fixtures / f"{args.case}.tgz"
+    with tarfile.open(tgz, "w:gz") as tar:
+        tar.add(archive_root / "trace.trace", arcname="trace.trace")
+    archive_size = tgz.stat().st_size
+    if archive_size > args.max_archive_bytes:
+        raise SystemExit(
+            f"{tgz} is {archive_size} bytes, over the {args.max_archive_bytes} byte fixture limit; "
+            "choose an earlier/smaller frame and re-run gltrim"
+        )
+    golden_out = fixtures / f"{args.case}.{target_call:010d}.png"
+    shutil.copy2(golden, golden_out)
+
+    manifest = {
+        "name": args.case,
+        "trace_archive": tgz.name,
+        "trace_file": "trace.trace",
+        "golden": golden_out.name,
+        "target_call": target_call,
+        "width": args.width,
+        "height": args.height,
+        "ssim_threshold": float(args.ssim_threshold),
+        "archive_size": archive_size,
+    }
+    manifest_path = capture_dir / f"{args.case}.fixture.json"
+    manifest_path.write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+    print(json.dumps(manifest, indent=2))
+
+
+if __name__ == "__main__":
+    main()

tools/trace_replay/skills/trace-fixture-authoring/SKILL.md

@@ -1,3 +1,8 @@
+---
+name: trace-fixture-authoring
+description: Author a deterministic MobileGL trace-replay fixture from a captured apitrace - build the in-tree apitrace fork, capture a reproducible scene, frame-trim with gltrim, generate and verify a golden image, package under the archive-size budget, register the case in trace_cases.json, and validate on Linux and Android. Use when adding or re-trimming a trace_replay regression fixture.
+---
+
 # Trace fixture authoring
 
 ## Variables
@@ -75,10 +80,19 @@ Minecraft specifics that keep the capture deterministic and small:
   `doMobSpawning`, `randomTickSpeed 0`, a fixed `DayTime`, and the player
   `Rotation` that frames the intended subject. The camera snaps to the saved
   rotation on world join, so composition is edited in the save, not in-game.
-- `options.txt`: `pauseOnLostFocus:false`, a low `maxFps` (10 works), and a
-  small `renderDistance` (3). Frame rate and render distance are the two main
-  levers on trace size; a ~35 s in-world session at 10 fps lands well under
-  the archive budget after repack.
+- `options.txt`: `pauseOnLostFocus:false`, a low `maxFps` (10 works), a small
+  `renderDistance` (3), and the capture resolution pinned to `WIDTH` x `HEIGHT`
+  (854x480) via `overrideWidth`/`overrideHeight` (or `--width`/`--height`).
+  These are the main levers on fixture size: a low frame rate keeps the full
+  trace short, a small render distance keeps per-frame geometry down, and the
+  854x480 resolution keeps every render target the frame references small (a
+  trimmed frame's framebuffer/attachment textures scale with resolution
+  squared). A ~35 s in-world session at 10 fps and 854x480 lands well under the
+  archive budget after repack.
+- `maxFps` has a practical floor: Minecraft ignores values below ~10 and falls
+  back to unlimited/vsync (a `maxFps:1` capture rendered ~60 fps and ballooned
+  the trace). 10 is as low as this lever goes, so do not count on a lower frame
+  rate to shrink the frame count further.
 - Enter the world non-interactively with `--quickPlaySingleplayer <world>` so
   every capture takes the same path from boot to gameplay.
 - Keep the game window UNFOCUSED for the whole capture (focus the desktop
@@ -111,6 +125,61 @@ For Java:
 An `@argfile` with the full JVM+game command line keeps the invocation
 reproducible across recaptures.
 
+NEVER put a real credential on the traced command line. apitrace records the
+traced process's argv into the trace as a `process.commandLine` property, so
+anything passed there - `--accessToken`, session tokens, API keys - is embedded
+in the trace and ships inside the committed fixture. Minecraft never validates
+`--accessToken` for singleplayer, so pass a placeholder (`--accessToken 0`);
+`--username`/`--uuid` are public and may stay real. Before packaging, grep the
+UNCOMPRESSED trace for the secret to confirm it is absent:
+
+```sh
+"$APITRACE" repack "$WORK/$CASE/trace.trace" /tmp/plain.trace   # decompress
+grep -ac "<secret-prefix>" /tmp/plain.trace                     # must be 0
+```
+
+If a secret has already been captured, it can be scrubbed in place instead of
+recapturing: apitrace's snappy container is `[length][raw snappy]` chunks with
+no checksum, and a high-entropy secret is stored as literal bytes, so replacing
+those bytes with an EQUAL-LENGTH filler keeps the container valid and leaves the
+GL call stream byte-identical. Blank every maximal run of the secret (it splits
+across chunks), then verify: frame count unchanged, the decompressed trace no
+longer contains the secret, and the replayed target frame still matches the
+golden. Treat any already-pushed trace as leaked regardless - rotate the
+credential, since a force-push does not purge the LFS object from the remote.
+
+Watch for vendor-gated shader paths. Shader packs branch on the GL vendor that
+Iris injects (`MC_GL_VENDOR_NVIDIA` / `_AMD` / ...) and compile a
+vendor-exclusive path, so capturing on an NVIDIA card can bake NVIDIA-only GLSL
+into the fixture (iterationRP selects `subgroupPartitionNV` /
+`GL_NV_shader_subgroup_partitioned` instead of the portable
+`subgroupShuffleXor`). Iris resolves the `#ifdef` before `glShaderSource`, so
+only the taken branch is in the trace and the fixture cannot replay on the
+mobile GPUs MobileGL targets. Rather than hunting for a second GPU (the Windows
+per-app GPU preference does NOT change which OpenGL ICD is loaded), mask the
+vendor at capture time with apitrace's own config - point `GLTRACE_CONF` at a
+file containing:
+
+```
+GL_VENDOR = "NoVIDIA (MobileGL spoof)"
+GL_RENDERER = "NoVIDIA (MobileGL spoof)"
+```
+
+The wrapper then returns that from `glGetString`, so the pack compiles the
+portable path while still running on the fast driver. Pick a string that does
+NOT contain the real vendor name as a substring (Iris matches by substring, so
+"Not NVIDIA ..." would still match) and that is self-describing, so nobody later
+mistakes the trace for a capture on different hardware. Afterwards, grep the
+decoded trace to confirm the vendor-exclusive symbols are gone:
+
+```sh
+"$APITRACE" dump "$WORK/$CASE/full.trace" | grep -c subgroupPartitionNV   # must be 0
+```
+
+Software rasterisers are not a substitute here: llvmpipe exposes no
+`GL_KHR_shader_subgroup` at all, and packs that use subgroup ops unguarded
+cannot run on it in any vendor configuration.
+
 Keep `full.trace` until both backends are validated.
 
 Persistent-mapped buffers: apps may legally write a `GL_MAP_PERSISTENT_BIT`
@@ -172,12 +241,37 @@ name"-style retrace warnings. If content is missing from the trimmed trace
 but present in the full trace, the fix belongs in `3rdparty/apitrace`'s
 frametrim, not in the fixture.
 
+Temporal shaders (auto-exposure / eye adaptation, TAA, temporal reflections -
+e.g. the iterationRP shader pack) break a single-frame `gltrim -f`: the target
+frame reads its predecessors' feedback buffers, which the isolated frame no
+longer contains, so a mid-sequence frame replays overexposed to white (and any
+timed name/version overlay the pack draws in its first seconds silently drops).
+The symptom is a trimmed frame that looks blown-out or washed while the same
+frame of `full.trace` renders correctly, and it gets worse the later the frame.
+When a single-frame trim of such a pack cannot be made to render correctly, keep
+the temporal history instead of the dependency slice: select an early in-world
+target frame and trim a PREFIX with `apitrace trim --calls=0-<target-swap-call>`
+(it preserves call numbers, so `target_call` is just that swap call). The prefix
+replays every frame up to the target, so its temporal buffers are correct.
+Prefer the earliest frame that already shows the intended subject - fewer lead-in
+frames means a smaller archive and a faster CI replay. This deviates from the
+single-frame rule deliberately; note it in the README entry.
+
 ## Generate golden
 
 Generate frame snapshots from the trimmed trace, then choose the snapshot that
 matches the selected frame. The target call used by replay registration must
 come from the trimmed trace, not from a call-filtered full-trace selection.
 
+Generate the golden with the same GL stack the scene was captured on. A headless
+software renderer (llvmpipe) is fine for vanilla and light packs, but heavy
+ray-traced shader packs (compute-driven atmosphere LUTs, screen-space tracing -
+e.g. iterationRP) render as solid black or blown-out white under llvmpipe. Drive
+the golden from a real GPU instead: on Windows a stock `glretrace.exe` (an
+upstream apitrace release works for replay even on an in-tree-fork trace) replays
+the trace on the discrete GPU and snapshots the target call. Read the resulting
+PNG back and confirm the subject actually rendered before trusting it as golden.
+
 ```sh
 mkdir -p "$WORK/$CASE/golden"
 "$APITRACE" replay --headless \
@@ -255,6 +349,15 @@ Check the final archive size. The committed fixture archive should be less than
 with a shorter run or a lower frame rate / render distance instead of adding
 call-based filtering.
 
+Some packs have an irreducible size floor: a large static lookup table baked
+into the pack (iterationRP ships a ~17 MiB half-float atmosphere LUT that the
+target frame samples) lands in the trace once and does not compress, so every
+variant - single frame, prefix, or full - sits near the same size regardless of
+frame count. When the floor alone exceeds the budget, neither a lower frame rate
+nor fewer frames helps; confirm the fixture is worth the exception and record the
+measured size in the case's README entry rather than chasing an unreachable
+target.
+
 ```sh
 du -h "$REPO/tools/trace_replay/fixtures/$CASE.tgz"
 tar -tzf "$REPO/tools/trace_replay/fixtures/$CASE.tgz"

tools/trace_replay/skills/trace-fixture-authoring/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "MobileGL Trace Fixture Authoring"
+  short_description: "Author and register a MobileGL trace-replay fixture"
+  default_prompt: "Use $trace-fixture-authoring to author and register a MobileGL trace replay fixture."

tools/trace_replay/trace_cases.json

@@ -276,12 +276,18 @@
     },
     {
       "name": "improved-transparency-minecraft-26.3",
-      "ci": false,
       "trace_archive": "improved-transparency-minecraft-26.3.tgz",
       "golden": "improved-transparency-minecraft-26.3.0002667619.png",
       "target_call": 2667619,
-      "timeout_seconds": 1800,
-      "coherent_as_flush": true
+      "timeout_seconds": 1800
+    },
+    {
+      "name": "minecraft-1.21.4-fabric-iris-iterationrp-in-world",
+      "ci": false,
+      "trace_archive": "minecraft-1.21.4-fabric-iris-iterationrp-in-world.tgz",
+      "golden": "minecraft-1.21.4-fabric-iris-iterationrp-in-world.0000202020.png",
+      "target_call": 202020,
+      "timeout_seconds": 1800
     }
   ]
 }