[Refactor] (ShaderTranspiler, MG_Backend, MG_Util): gate the subgroup prefix-scan rewrite behind a generic device-quirk registry with GPU vendor detection and MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN override, warn on template mismatch, and block ARB/NV subgroup spellings

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 +76,10 @@ 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;
     };
     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 +134,7 @@ 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");
     }
 
     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,
@@ -291,6 +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

@@ -1035,6 +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/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -794,5 +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_Test/Program/ProgramUtilTest.cpp

@@ -1623,4 +1623,26 @@ TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsPartialOrUnsafeTem
     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_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];
@@ -210,6 +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];

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;

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.cpp

@@ -12,6 +12,7 @@
 #include <cctype>
 #include <initializer_list>
 #include <utility>
+#include <Config.h>
 #include <MG_Backend/BackendObjects.h>
 
 namespace {
@@ -365,7 +366,21 @@ namespace {
             HasIdentifierWithPrefixOutsideAllowed(tokens, "subgroup", {"subgroupInclusiveAdd"}) ||
             HasIdentifierWithPrefixOutsideAllowed(
                 tokens, "gl_Subgroup",
-                {"gl_SubgroupInvocationID", "gl_SubgroupSize", "gl_SubgroupID", "gl_NumSubgroups"})) {
+                {"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;
         }
 
@@ -967,6 +982,14 @@ namespace MobileGL {
                 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;
                 }
 
@@ -979,6 +1002,76 @@ namespace MobileGL {
                 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.
@@ -1035,12 +1128,7 @@ namespace MobileGL {
                 ModernizeLegacyGLSL(stage, source);
                 InjectDepthRangeBuiltinShim(stage, source);
 
-                const auto& activeBackend = MG_Backend::pActiveBackendObject;
+                ApplyShaderSourceQuirks(stage, source);
-                if (stage == ShaderStage::Compute && activeBackend &&
-                    activeBackend->GetBackendType() == BackendType::DirectVulkan) {
-                    RewriteLinearSubgroupPrefixScanForVulkan(stage, activeBackend->GetDynamicParameters().SubgroupSize,
-                                                             source);
-                }
             }
 
             Bool RetargetLegacyVersionDirectiveTo460(String& source) {

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.h

@@ -27,8 +27,10 @@ namespace MobileGL {
             // 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. DirectVulkan calls this through
+            // template was recognized and rewritten. PreprocessShaderSource reaches this through
-            // PreprocessShaderSource; the explicit entry point exists for deterministic tests.
+            // 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