[Fix, Test] (GLImpl, ShaderTranspiler): reconcile the compute work-group limits glGetIntegeri_v and glslang advertise

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -47,7 +47,10 @@ namespace MobileGL::MG_Impl::GLImpl {
             }
         }
 
-        constexpr GLint kFrontendMaxComputeUniformComponents = 1024;
+        // Shared with the glslang resource table for the same reason as the atomic-counter
+        // limits below: gl_MaxComputeUniformComponents expands from BuildTBuiltInResource.
+        constexpr GLint kFrontendMaxComputeUniformComponents =
+            static_cast<GLint>(MG_Util::ShaderTranspiler::MAX_COMPUTE_UNIFORM_COMPONENTS);
         // Every atomic-counter limit is shared with the glslang resource table
         // (BuildTBuiltInResource) through MG_Util/ShaderTranspiler/Types.h: GL 4.6 requires
         // glGetIntegerv and the gl_MaxAtomicCounter* built-in constants to agree, and the two
@@ -118,12 +121,16 @@ namespace MobileGL::MG_Impl::GLImpl {
         constexpr GLint kFrontendSubpixelBits = 4;
         constexpr GLint kFrontendMaxSamples = 4;
 
+        // The floors under GL_MAX_COMPUTE_WORK_GROUP_COUNT / _SIZE. Shared with the compile
+        // pipeline (CaptureCompileEnv floors the same driver answers at them, and
+        // BuildTBuiltInResource expands gl_MaxComputeWorkGroup* from the result), because a
+        // shader is allowed to compare the built-in constant against this query.
         constexpr GLint GetMinComputeWorkGroupCount(GLuint index) {
-            return index < 3 ? 65535 : 0;
+            return index < 3 ? static_cast<GLint>(MG_Util::ShaderTranspiler::MIN_COMPUTE_WORK_GROUP_COUNT[index]) : 0;
         }
 
         constexpr GLint GetMinComputeWorkGroupSize(GLuint index) {
-            return index < 2 ? 1024 : (index == 2 ? 64 : 0);
+            return index < 3 ? static_cast<GLint>(MG_Util::ShaderTranspiler::MIN_COMPUTE_WORK_GROUP_SIZE[index]) : 0;
         }
 
         GLint GetMaxCombinedUniformComponents(GLint maxDefaultUniformComponents, GLint maxUniformBlocks,

MobileGL/MG_Test/SanityTest.cpp

@@ -994,6 +994,68 @@ void main() {
     MG_State::pGLContext = Move(previousContext);
 }
 
+// KHR-GL43.compute_shader.max: the test queries every GL_MAX_COMPUTE_* value through the API and
+// then makes a compute shader compare the matching gl_MaxCompute* constant against it. The two
+// used to be independent tables and gl_MaxComputeWorkGroupSize.z disagreed - glslang compiled
+// against a permissive 1024 while the context advertises the 64 the GL 4.6 minimum (and every ES
+// driver) reports.
+TEST(GetterSanity, ComputeWorkGroupQueriesMatchShaderCompilerLimits) {
+    using namespace MobileGL;
+
+    auto previousContext = Move(MG_State::pGLContext);
+    auto previousBackend = Move(MG_Backend::pActiveBackendObject);
+    MG_State::pGLContext = MakeUnique<MG_State::GLState::GLContext>();
+    MG_Backend::pActiveBackendObject = MakeUnique<DynamicParameterBackend>(MG_Backend::DynamicBackendParameters{});
+
+    GLint size[3] = {0, 0, 0};
+    GLint count[3] = {0, 0, 0};
+    for (GLuint index = 0; index < 3; ++index) {
+        MG_Impl::GLImpl::GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, index, &size[index]);
+        MG_Impl::GLImpl::GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, index, &count[index]);
+    }
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    // A negative array size is a compile error, so the stage only compiles when EVERY component
+    // of both built-in constants equals what the query above reported. Two-sided by construction:
+    // a resource table that is too permissive fails it exactly like one that is too tight.
+    const String source = R"(#version 430 core
+layout(local_size_x = 1) in;
+const int mgAgree = (gl_MaxComputeWorkGroupSize == ivec3()" +
+                         std::to_string(size[0]) + ", " + std::to_string(size[1]) + ", " +
+                         std::to_string(size[2]) + R"() &&
+                     gl_MaxComputeWorkGroupCount == ivec3()" +
+                         std::to_string(count[0]) + ", " + std::to_string(count[1]) + ", " +
+                         std::to_string(count[2]) + R"()) ? 1 : -1;
+int mgProbe[mgAgree];
+void main() {
+    mgProbe[0] = 0;
+}
+)";
+    auto compiled = MG_Util::ShaderTranspiler::ShaderCompiler::CompileShader({
+        .shaderType = GL_COMPUTE_SHADER,
+        .sourceStr = source,
+    });
+    EXPECT_TRUE(compiled) << (compiled ? "" : compiled.error().log);
+
+    // The z ceiling is also what glslang checks a declared local_size_z against, so it has to
+    // reject one invocation past the advertised limit and accept the limit itself.
+    const String atLimit = "#version 430 core\nlayout(local_size_z = " + std::to_string(size[2]) +
+                           ") in;\nvoid main() {}\n";
+    const String pastLimit = "#version 430 core\nlayout(local_size_z = " + std::to_string(size[2] + 1) +
+                             ") in;\nvoid main() {}\n";
+    EXPECT_TRUE(MG_Util::ShaderTranspiler::ShaderCompiler::CompileShader({
+        .shaderType = GL_COMPUTE_SHADER,
+        .sourceStr = atLimit,
+    }));
+    EXPECT_FALSE(MG_Util::ShaderTranspiler::ShaderCompiler::CompileShader({
+        .shaderType = GL_COMPUTE_SHADER,
+        .sourceStr = pastLimit,
+    }));
+
+    MG_Backend::pActiveBackendObject = Move(previousBackend);
+    MG_State::pGLContext = Move(previousContext);
+}
+
 TEST(GetterSanity, ReportsKhrSubgroupDynamicParameters) {
     using namespace MobileGL;
 

MobileGL/MG_Util/ShaderTranspiler/CompileEnv.cpp

@@ -28,6 +28,9 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
         HashValue(state, env.maxComputeWorkGroupSize[0]);
         HashValue(state, env.maxComputeWorkGroupSize[1]);
         HashValue(state, env.maxComputeWorkGroupSize[2]);
+        HashValue(state, env.maxComputeWorkGroupCount[0]);
+        HashValue(state, env.maxComputeWorkGroupCount[1]);
+        HashValue(state, env.maxComputeWorkGroupCount[2]);
         HashValue(state, env.maxComputeWorkGroupInvocations);
         HashValue(state, env.backend);
         // DynamicBackendParameters is a plain aggregate of scalars; hashing its object
@@ -51,19 +54,23 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
             env->advertisedExtensions = activeBackend->GetRendererInfo().RendererGLInfo.Extensions;
         }
 
-        // GL_MAX_COMPUTE_WORK_GROUP_SIZE. This is a REAL driver call on DirectGLES; it must
-        // happen here, on the context thread, and exactly once per context. The frontend
-        // minimum is the floor, matching what GL_Getter reports.
-        // TODO: Share these exposed compute limit helpers with GL_Getter.cpp instead of duplicating the frontend minima.
-        constexpr Uint kFrontendMinComputeWorkGroupSizes[3] = {1024, 1024, 64};
+        // GL_MAX_COMPUTE_WORK_GROUP_SIZE / _COUNT. These are REAL driver calls on DirectGLES; they
+        // must happen here, on the context thread, and exactly once per context. The frontend
+        // minimum is the floor, matching what GL_Getter reports - both sides now floor at the
+        // shared MIN_COMPUTE_WORK_GROUP_* constants rather than at their own copy of them.
         for (Uint index = 0; index < 3; ++index) {
-            Int backendValue = 0;
+            Int backendSize = 0;
+            Int backendCount = 0;
             if (MG_Backend::gBackendFunctionsTable.GL.GetIntegeri_v) {
                 MG_Backend::gBackendFunctionsTable.GL.GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, index,
-                                                                    &backendValue);
+                                                                    &backendSize);
+                MG_Backend::gBackendFunctionsTable.GL.GetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, index,
+                                                                    &backendCount);
             }
             env->maxComputeWorkGroupSize[index] =
-                std::max(static_cast<Uint>(std::max(backendValue, 0)), kFrontendMinComputeWorkGroupSizes[index]);
+                std::max(static_cast<Uint>(std::max(backendSize, 0)), MIN_COMPUTE_WORK_GROUP_SIZE[index]);
+            env->maxComputeWorkGroupCount[index] =
+                std::max(static_cast<Uint>(std::max(backendCount, 0)), MIN_COMPUTE_WORK_GROUP_COUNT[index]);
         }
 
         constexpr Uint64 kFrontendMaxComputeWorkGroupInvocations = 1024;

MobileGL/MG_Util/ShaderTranspiler/CompileEnv.h

@@ -12,6 +12,18 @@
 #include <MG_Backend/BackendObject.h>
 
 namespace MobileGL::MG_Util::ShaderTranspiler {
+    // GL_MAX_COMPUTE_WORK_GROUP_COUNT / _SIZE core minimums (GL 4.6 core table 23.45), in ONE
+    // place because three separate readers have to agree on them: CaptureCompileEnv (which floors
+    // the backend's answer at them), GL_Getter (which answers the same query the same way) and
+    // BuildTBuiltInResource (whose gl_MaxComputeWorkGroup* constants a shader compares against
+    // the query - KHR-GL43.compute_shader.max does exactly that). They used to be three copies,
+    // and the z one disagreed: glslang compiled against 1024 while the context advertised 64.
+    inline constexpr Uint MIN_COMPUTE_WORK_GROUP_COUNT[3] = {65535, 65535, 65535};
+    inline constexpr Uint MIN_COMPUTE_WORK_GROUP_SIZE[3] = {1024, 1024, 64};
+    // GL_MAX_COMPUTE_UNIFORM_COMPONENTS, the same invariant with no backend input: the number
+    // glGetIntegerv answers and the number gl_MaxComputeUniformComponents expands to.
+    inline constexpr Int MAX_COMPUTE_UNIFORM_COMPONENTS = 1024;
+
     // everything outside (stage, source) this reads - advertised extensions and backend limits -
     // so the transformation is a pure function of its three arguments and can run on a worker
     // thread.
@@ -34,7 +46,13 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
     struct CompileEnv {
         // --- compute limits: the ONLY former real-driver read in the pipeline ---
         // GL_MAX_COMPUTE_WORK_GROUP_SIZE, already max()'d with the frontend minimum.
-        Uint maxComputeWorkGroupSize[3] = {1024, 1024, 64};
+        Uint maxComputeWorkGroupSize[3] = {MIN_COMPUTE_WORK_GROUP_SIZE[0], MIN_COMPUTE_WORK_GROUP_SIZE[1],
+                                           MIN_COMPUTE_WORK_GROUP_SIZE[2]};
+        // GL_MAX_COMPUTE_WORK_GROUP_COUNT, likewise. Carried for the same reason the size is:
+        // gl_MaxComputeWorkGroupCount expands from it at parse time, so the compile pipeline
+        // needs the number the context advertises without reaching back to the live backend.
+        Uint maxComputeWorkGroupCount[3] = {MIN_COMPUTE_WORK_GROUP_COUNT[0], MIN_COMPUTE_WORK_GROUP_COUNT[1],
+                                            MIN_COMPUTE_WORK_GROUP_COUNT[2]};
         // GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, likewise.
         Uint64 maxComputeWorkGroupInvocations = 1024;
 

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.cpp

@@ -83,14 +83,7 @@ namespace MobileGL {
                 Resources.minProgramTexelOffset = -8;
                 Resources.maxProgramTexelOffset = 7;
                 Resources.maxClipDistances = 8;
-                Resources.maxComputeWorkGroupCountX = 65535;
-                Resources.maxComputeWorkGroupCountY = 65535;
-                Resources.maxComputeWorkGroupCountZ = 65535;
-                Resources.maxComputeWorkGroupSizeX = 1024;
-                Resources.maxComputeWorkGroupSizeY = 1024;
-                // TODO: Drive glslang compute resource limits from the active backend instead of this permissive cap.
-                Resources.maxComputeWorkGroupSizeZ = 1024;
-                Resources.maxComputeUniformComponents = 1024;
+                Resources.maxComputeUniformComponents = MAX_COMPUTE_UNIFORM_COMPONENTS;
                 Resources.maxComputeTextureImageUnits = 16;
                 Resources.maxComputeImageUniforms = 8;
                 Resources.maxComputeAtomicCounters = MAX_ATOMIC_COUNTERS_PER_STAGE;
@@ -179,6 +172,25 @@ namespace MobileGL {
                 Resources.maxFragmentImageUniforms = dynamicParameters.MaxFragmentImageUniforms;
                 Resources.maxComputeImageUniforms = dynamicParameters.MaxComputeImageUniforms;
                 Resources.maxCombinedImageUniforms = dynamicParameters.MaxCombinedImageUniforms;
+                Resources.maxComputeTextureImageUnits = dynamicParameters.MaxComputeTextureImageUnits;
+
+                // The compute work-group limits are the env's, not the backend parameters': they
+                // are the only ones that come from a REAL indexed driver query, which
+                // CaptureCompileEnv already issued once on the GL thread and floored at the core
+                // minimum exactly as GL_Getter does. Reading the same snapshot here is what makes
+                // gl_MaxComputeWorkGroupSize and glGetIntegeri_v agree by construction
+                // (KHR-GL43.compute_shader.max compares them); the z component was 1024 here
+                // against the 64 every ES driver reports. A null env is the standalone/test entry
+                // point, which has no context to have queried one - the core minimums stand, which
+                // is what a default-constructed CompileEnv carries anyway.
+                const Uint* maxWorkGroupSize = env ? env->maxComputeWorkGroupSize : MIN_COMPUTE_WORK_GROUP_SIZE;
+                const Uint* maxWorkGroupCount = env ? env->maxComputeWorkGroupCount : MIN_COMPUTE_WORK_GROUP_COUNT;
+                Resources.maxComputeWorkGroupSizeX = static_cast<int>(maxWorkGroupSize[0]);
+                Resources.maxComputeWorkGroupSizeY = static_cast<int>(maxWorkGroupSize[1]);
+                Resources.maxComputeWorkGroupSizeZ = static_cast<int>(maxWorkGroupSize[2]);
+                Resources.maxComputeWorkGroupCountX = static_cast<int>(maxWorkGroupCount[0]);
+                Resources.maxComputeWorkGroupCountY = static_cast<int>(maxWorkGroupCount[1]);
+                Resources.maxComputeWorkGroupCountZ = static_cast<int>(maxWorkGroupCount[2]);
 
                 Resources.limits.nonInductiveForLoops = true;
                 Resources.limits.whileLoops = true;