D3D11: Add support to compile and link compute shaders

src/compiler/translator/OutputHLSL.cpp

@@ -115,6 +115,11 @@ OutputHLSL::OutputHLSL(sh::GLenum shaderType,
     mUsesInstanceID              = false;
     mUsesVertexID                = false;
     mUsesFragDepth               = false;
+    mUsesNumWorkGroups           = false;
+    mUsesWorkGroupID             = false;
+    mUsesLocalInvocationID       = false;
+    mUsesGlobalInvocationID      = false;
+    mUsesLocalInvocationIndex    = false;
     mUsesXor                     = false;
     mUsesDiscardRewriting        = false;
     mUsesNestedBreak             = false;
@@ -540,7 +545,7 @@ void OutputHLSL::header(TInfoSinkBase &out, const BuiltInFunctionEmulator *built
             out << "#define GL_USES_FRAG_DATA\n";
         }
     }
-    else  // Vertex shader
+    else if (mShaderType == GL_VERTEX_SHADER)
     {
         out << "// Attributes\n";
         out << attributes;
@@ -631,6 +636,40 @@ void OutputHLSL::header(TInfoSinkBase &out, const BuiltInFunctionEmulator *built
             out << "\n";
         }
     }
+    else  // Compute shader
+    {
+        ASSERT(mShaderType == GL_COMPUTE_SHADER);
+        if (mUsesNumWorkGroups)
+        {
+            out << "cbuffer DriverConstants : register(b1)\n"
+                   "{\n";
+            out << "    uint3 gl_NumWorkGroups : packoffset(c0);\n";
+            out << "};\n";
+        }
+
+        // Follow built-in variables would be initialized in
+        // DynamicHLSL::generateComputeShaderLinkHLSL, if they
+        // are used in compute shader.
+        if (mUsesWorkGroupID)
+        {
+            out << "static uint3 gl_WorkGroupID = uint3(0, 0, 0);\n";
+        }
+
+        if (mUsesLocalInvocationID)
+        {
+            out << "static uint3 gl_LocalInvocationID = uint3(0, 0, 0);\n";
+        }
+
+        if (mUsesGlobalInvocationID)
+        {
+            out << "static uint3 gl_GlobalInvocationID = uint3(0, 0, 0);\n";
+        }
+
+        if (mUsesLocalInvocationIndex)
+        {
+            out << "static uint gl_LocalInvocationIndex = uint(0);\n";
+        }
+    }
 
     bool getDimensionsIgnoresBaseLevel =
         (mCompileOptions & SH_HLSL_GET_DIMENSIONS_IGNORES_BASE_LEVEL) != 0;
@@ -666,6 +705,31 @@ void OutputHLSL::header(TInfoSinkBase &out, const BuiltInFunctionEmulator *built
         out << "#define GL_USES_DEPTH_RANGE\n";
     }
 
+    if (mUsesNumWorkGroups)
+    {
+        out << "#define GL_USES_NUM_WORK_GROUPS\n";
+    }
+
+    if (mUsesWorkGroupID)
+    {
+        out << "#define GL_USES_WORK_GROUP_ID\n";
+    }
+
+    if (mUsesLocalInvocationID)
+    {
+        out << "#define GL_USES_LOCAL_INVOCATION_ID\n";
+    }
+
+    if (mUsesGlobalInvocationID)
+    {
+        out << "#define GL_USES_GLOBAL_INVOCATION_ID\n";
+    }
+
+    if (mUsesLocalInvocationIndex)
+    {
+        out << "#define GL_USES_LOCAL_INVOCATION_INDEX\n";
+    }
+
     if (mUsesXor)
     {
         out << "bool xor(bool p, bool q)\n"
@@ -779,6 +843,31 @@ void OutputHLSL::visitSymbol(TIntermSymbol *node)
             mUsesFragDepth = true;
             out << "gl_Depth";
         }
+        else if (qualifier == EvqNumWorkGroups)
+        {
+            mUsesNumWorkGroups = true;
+            out << name;
+        }
+        else if (qualifier == EvqWorkGroupID)
+        {
+            mUsesWorkGroupID = true;
+            out << name;
+        }
+        else if (qualifier == EvqLocalInvocationID)
+        {
+            mUsesLocalInvocationID = true;
+            out << name;
+        }
+        else if (qualifier == EvqGlobalInvocationID)
+        {
+            mUsesGlobalInvocationID = true;
+            out << name;
+        }
+        else if (qualifier == EvqLocalInvocationIndex)
+        {
+            mUsesLocalInvocationIndex = true;
+            out << name;
+        }
         else
         {
             out << DecorateIfNeeded(node->getName());

src/compiler/translator/OutputHLSL.h

@@ -174,6 +174,11 @@ class OutputHLSL : public TIntermTraverser
     bool mUsesInstanceID;
     bool mUsesVertexID;
     bool mUsesFragDepth;
+    bool mUsesNumWorkGroups;
+    bool mUsesWorkGroupID;
+    bool mUsesLocalInvocationID;
+    bool mUsesGlobalInvocationID;
+    bool mUsesLocalInvocationIndex;
     bool mUsesXor;
     bool mUsesDiscardRewriting;
     bool mUsesNestedBreak;

src/libANGLE/angletypes.h

@@ -42,7 +42,8 @@ PrimitiveType GetPrimitiveType(GLenum drawMode);
 enum SamplerType
 {
     SAMPLER_PIXEL,
-    SAMPLER_VERTEX
+    SAMPLER_VERTEX,
+    SAMPLER_COMPUTE
 };
 
 struct Rectangle
@@ -288,6 +289,11 @@ inline DestT *GetImplAs(SrcT *src)
     return GetAs<DestT>(src->getImplementation());
 }
 
+template <typename DestT, typename SrcT>
+inline DestT *SafeGetImplAs(SrcT *src)
+{
+    return src != nullptr ? GetAs<DestT>(src->getImplementation()) : nullptr;
+}
 }
 
 #include "angletypes.inl"

src/libANGLE/renderer/d3d/DynamicHLSL.cpp

@@ -764,6 +764,79 @@ void DynamicHLSL::generateShaderLinkHLSL(const gl::ContextState &data,
     *pixelHLSL  = pixelStream.str();
 }
 
+std::string DynamicHLSL::generateComputeShaderLinkHLSL(const gl::ProgramState &programData) const
+{
+    const gl::Shader *computeShaderGL = programData.getAttachedComputeShader();
+    std::stringstream computeStream;
+    std::string translatedSource = computeShaderGL->getTranslatedSource();
+    computeStream << translatedSource;
+
+    bool usesWorkGroupID = translatedSource.find("GL_USES_WORK_GROUP_ID") != std::string::npos;
+    bool usesLocalInvocationID =
+        translatedSource.find("GL_USES_LOCAL_INVOCATION_ID") != std::string::npos;
+    bool usesGlobalInvocationID =
+        translatedSource.find("GL_USES_GLOBAL_INVOCATION_ID") != std::string::npos;
+    bool usesLocalInvocationIndex =
+        translatedSource.find("GL_USES_LOCAL_INVOCATION_INDEX") != std::string::npos;
+
+    computeStream << "\nstruct CS_INPUT\n{\n";
+    if (usesWorkGroupID)
+    {
+        computeStream << "    uint3 dx_WorkGroupID : "
+                      << "SV_GroupID;\n";
+    }
+
+    if (usesLocalInvocationID)
+    {
+        computeStream << "    uint3 dx_LocalInvocationID : "
+                      << "SV_GroupThreadID;\n";
+    }
+
+    if (usesGlobalInvocationID)
+    {
+        computeStream << "    uint3 dx_GlobalInvocationID : "
+                      << "SV_DispatchThreadID;\n";
+    }
+
+    if (usesLocalInvocationIndex)
+    {
+        computeStream << "    uint dx_LocalInvocationIndex : "
+                      << "SV_GroupIndex;\n";
+    }
+
+    computeStream << "};\n\n";
+
+    const sh::WorkGroupSize &localSize = computeShaderGL->getWorkGroupSize();
+    computeStream << "[numthreads(" << localSize[0] << ", " << localSize[1] << ", " << localSize[2]
+                  << ")]\n";
+
+    computeStream << "void main(CS_INPUT input)\n"
+                  << "{\n";
+
+    if (usesWorkGroupID)
+    {
+        computeStream << "    gl_WorkGroupID = input.dx_WorkGroupID;\n";
+    }
+    if (usesLocalInvocationID)
+    {
+        computeStream << "    gl_LocalInvocationID = input.dx_LocalInvocationID;\n";
+    }
+    if (usesGlobalInvocationID)
+    {
+        computeStream << "    gl_GlobalInvocationID = input.dx_GlobalInvocationID;\n";
+    }
+    if (usesLocalInvocationIndex)
+    {
+        computeStream << "    gl_LocalInvocationIndex = input.dx_LocalInvocationIndex;\n";
+    }
+
+    computeStream << "\n"
+                  << "    gl_main();\n"
+                  << "}\n";
+
+    return computeStream.str();
+}
+
 std::string DynamicHLSL::generateGeometryShaderPreamble(const VaryingPacking &varyingPacking,
                                                         const BuiltinVaryingsD3D &builtinsD3D) const
 {

src/libANGLE/renderer/d3d/DynamicHLSL.h

@@ -115,6 +115,7 @@ class DynamicHLSL : angle::NonCopyable
                                 const BuiltinVaryingsD3D &builtinsD3D,
                                 std::string *pixelHLSL,
                                 std::string *vertexHLSL) const;
+    std::string generateComputeShaderLinkHLSL(const gl::ProgramState &programData) const;
 
     std::string generateGeometryShaderPreamble(const gl::VaryingPacking &varyingPacking,
                                                const BuiltinVaryingsD3D &builtinsD3D) const;

src/libANGLE/renderer/d3d/ProgramD3D.h

@@ -44,6 +44,7 @@ struct D3DUniform : angle::NonCopyable
     unsigned int elementCount() const { return std::max(1u, arraySize); }
     bool isReferencedByVertexShader() const;
     bool isReferencedByFragmentShader() const;
+    bool isReferencedByComputeShader() const;
 
     // Duplicated from the GL layer
     GLenum type;
@@ -60,6 +61,7 @@ struct D3DUniform : angle::NonCopyable
     // Register information.
     unsigned int vsRegisterIndex;
     unsigned int psRegisterIndex;
+    unsigned int csRegisterIndex;
     unsigned int registerCount;
 
     // Register "elements" are used for uniform structs in ES3, to appropriately identify single
@@ -70,14 +72,22 @@ struct D3DUniform : angle::NonCopyable
 
 struct D3DUniformBlock
 {
-    D3DUniformBlock() : vsRegisterIndex(GL_INVALID_INDEX), psRegisterIndex(GL_INVALID_INDEX) {}
+    D3DUniformBlock()
+        : vsRegisterIndex(GL_INVALID_INDEX),
+          psRegisterIndex(GL_INVALID_INDEX),
+          csRegisterIndex(GL_INVALID_INDEX)
+    {
+    }
 
     bool vertexStaticUse() const { return vsRegisterIndex != GL_INVALID_INDEX; }
 
     bool fragmentStaticUse() const { return psRegisterIndex != GL_INVALID_INDEX; }
 
+    bool computeStaticUse() const { return csRegisterIndex != GL_INVALID_INDEX; }
+
     unsigned int vsRegisterIndex;
     unsigned int psRegisterIndex;
+    unsigned int csRegisterIndex;
 };
 
 struct D3DVarying final
@@ -166,6 +176,9 @@ class ProgramD3D : public ProgramImpl
                                                     GLenum drawMode,
                                                     ShaderExecutableD3D **outExecutable,
                                                     gl::InfoLog *infoLog);
+    gl::Error getComputeExecutable(const gl::ContextState &data,
+                                   ShaderExecutableD3D **outExecutable,
+                                   gl::InfoLog *infoLog);
 
     LinkResult link(const gl::ContextState &data,
                     const gl::VaryingPacking &packing,
@@ -236,8 +249,18 @@ class ProgramD3D : public ProgramImpl
 
     void setUniformBlockBinding(GLuint uniformBlockIndex, GLuint uniformBlockBinding) override;
 
-    const UniformStorageD3D &getVertexUniformStorage() const { return *mVertexUniformStorage; }
-    const UniformStorageD3D &getFragmentUniformStorage() const { return *mFragmentUniformStorage; }
+    const UniformStorageD3D &getVertexUniformStorage() const
+    {
+        return *mVertexUniformStorage.get();
+    }
+    const UniformStorageD3D &getFragmentUniformStorage() const
+    {
+        return *mFragmentUniformStorage.get();
+    }
+    const UniformStorageD3D &getComputeUniformStorage() const
+    {
+        return *mComputeUniformStorage.get();
+    }
 
     unsigned int getSerial() const;
 
@@ -352,6 +375,7 @@ class ProgramD3D : public ProgramImpl
                             GLenum targetUniformType);
 
     LinkResult compileProgramExecutables(const gl::ContextState &data, gl::InfoLog &infoLog);
+    LinkResult compileComputeExecutable(gl::InfoLog &infoLog);
 
     void gatherTransformFeedbackVaryings(const gl::VaryingPacking &varyings,
                                          const BuiltinInfo &builtins);
@@ -363,15 +387,16 @@ class ProgramD3D : public ProgramImpl
     void reset();
     void ensureUniformBlocksInitialized();
 
-    void initUniformBlockInfo();
+    void initUniformBlockInfo(const gl::Shader *shader);
     size_t getUniformBlockInfo(const sh::InterfaceBlock &interfaceBlock);
 
     RendererD3D *mRenderer;
     DynamicHLSL *mDynamicHLSL;
 
-    std::vector<VertexExecutable *> mVertexExecutables;
-    std::vector<PixelExecutable *> mPixelExecutables;
-    std::vector<ShaderExecutableD3D *> mGeometryExecutables;
+    std::vector<std::unique_ptr<VertexExecutable>> mVertexExecutables;
+    std::vector<std::unique_ptr<PixelExecutable>> mPixelExecutables;
+    std::vector<std::unique_ptr<ShaderExecutableD3D>> mGeometryExecutables;
+    std::unique_ptr<ShaderExecutableD3D> mComputeExecutable;
 
     std::string mVertexHLSL;
     angle::CompilerWorkaroundsD3D mVertexWorkarounds;
@@ -389,13 +414,16 @@ class ProgramD3D : public ProgramImpl
     bool mUsesPointSize;
     bool mUsesFlatInterpolation;
 
-    UniformStorageD3D *mVertexUniformStorage;
-    UniformStorageD3D *mFragmentUniformStorage;
+    std::unique_ptr<UniformStorageD3D> mVertexUniformStorage;
+    std::unique_ptr<UniformStorageD3D> mFragmentUniformStorage;
+    std::unique_ptr<UniformStorageD3D> mComputeUniformStorage;
 
     std::vector<Sampler> mSamplersPS;
     std::vector<Sampler> mSamplersVS;
+    std::vector<Sampler> mSamplersCS;
     GLuint mUsedVertexSamplerRange;
     GLuint mUsedPixelSamplerRange;
+    GLuint mUsedComputeSamplerRange;
     bool mDirtySamplerMapping;
 
     // Cache for getPixelExecutableForFramebuffer

src/libANGLE/renderer/d3d/RendererD3D.h

@@ -75,6 +75,7 @@ enum ShaderType
     SHADER_VERTEX,
     SHADER_PIXEL,
     SHADER_GEOMETRY,
+    SHADER_COMPUTE,
     SHADER_TYPE_MAX
 };
 

src/libANGLE/renderer/d3d/ShaderD3D.cpp

@@ -29,6 +29,9 @@ const char *GetShaderTypeString(GLenum type)
         case GL_FRAGMENT_SHADER:
             return "FRAGMENT";
 
+        case GL_COMPUTE_SHADER:
+            return "COMPUTE";
+
         default:
             UNREACHABLE();
             return "";
@@ -192,7 +195,7 @@ bool ShaderD3D::postTranslateCompile(gl::Compiler *compiler, std::string *infoLo
     mUsesPointSize             = translatedSource.find("GL_USES_POINT_SIZE") != std::string::npos;
     mUsesPointCoord            = translatedSource.find("GL_USES_POINT_COORD") != std::string::npos;
     mUsesDepthRange            = translatedSource.find("GL_USES_DEPTH_RANGE") != std::string::npos;
-    mUsesFragDepth = translatedSource.find("GL_USES_FRAG_DEPTH") != std::string::npos;
+    mUsesFragDepth             = translatedSource.find("GL_USES_FRAG_DEPTH") != std::string::npos;
     mUsesDiscardRewriting =
         translatedSource.find("ANGLE_USES_DISCARD_REWRITING") != std::string::npos;
     mUsesNestedBreak  = translatedSource.find("ANGLE_USES_NESTED_BREAK") != std::string::npos;

src/libANGLE/renderer/d3d/d3d11/Renderer11.cpp

@@ -3460,15 +3460,14 @@ gl::Error Renderer11::loadExecutable(const void *function,
     {
         case SHADER_VERTEX:
         {
-            ID3D11VertexShader *vertexShader      = NULL;
-            ID3D11GeometryShader *streamOutShader = NULL;
+            ID3D11VertexShader *vertexShader      = nullptr;
+            ID3D11GeometryShader *streamOutShader = nullptr;
 
             HRESULT result = mDevice->CreateVertexShader(function, length, NULL, &vertexShader);
             ASSERT(SUCCEEDED(result));
             if (FAILED(result))
             {
-                return gl::Error(GL_OUT_OF_MEMORY, "Failed to create vertex shader, result: 0x%X.",
-                                 result);
+                return gl::OutOfMemory() << "Failed to create vertex shader, " << result;
             }
 
             if (!streamOutVaryings.empty())
@@ -3496,8 +3495,7 @@ gl::Error Renderer11::loadExecutable(const void *function,
                 ASSERT(SUCCEEDED(result));
                 if (FAILED(result))
                 {
-                    return gl::Error(GL_OUT_OF_MEMORY,
-                                     "Failed to create steam output shader, result: 0x%X.", result);
+                    return gl::OutOfMemory() << "Failed to create steam output shader, " << result;
                 }
             }
 
@@ -3507,14 +3505,13 @@ gl::Error Renderer11::loadExecutable(const void *function,
         break;
         case SHADER_PIXEL:
         {
-            ID3D11PixelShader *pixelShader = NULL;
+            ID3D11PixelShader *pixelShader = nullptr;
 
             HRESULT result = mDevice->CreatePixelShader(function, length, NULL, &pixelShader);
             ASSERT(SUCCEEDED(result));
             if (FAILED(result))
             {
-                return gl::Error(GL_OUT_OF_MEMORY, "Failed to create pixel shader, result: 0x%X.",
-                                 result);
+                return gl::OutOfMemory() << "Failed to create pixel shader, " << result;
             }
 
             *outExecutable = new ShaderExecutable11(function, length, pixelShader);
@@ -3522,19 +3519,32 @@ gl::Error Renderer11::loadExecutable(const void *function,
         break;
         case SHADER_GEOMETRY:
         {
-            ID3D11GeometryShader *geometryShader = NULL;
+            ID3D11GeometryShader *geometryShader = nullptr;
 
             HRESULT result = mDevice->CreateGeometryShader(function, length, NULL, &geometryShader);
             ASSERT(SUCCEEDED(result));
             if (FAILED(result))
             {
-                return gl::Error(GL_OUT_OF_MEMORY,
-                                 "Failed to create geometry shader, result: 0x%X.", result);
+                return gl::OutOfMemory() << "Failed to create geometry shader, " << result;
             }
 
             *outExecutable = new ShaderExecutable11(function, length, geometryShader);
         }
         break;
+        case SHADER_COMPUTE:
+        {
+            ID3D11ComputeShader *computeShader = nullptr;
+
+            HRESULT result = mDevice->CreateComputeShader(function, length, NULL, &computeShader);
+            ASSERT(SUCCEEDED(result));
+            if (FAILED(result))
+            {
+                return gl::OutOfMemory() << "Failed to create compute shader, " << result;
+            }
+
+            *outExecutable = new ShaderExecutable11(function, length, computeShader);
+        }
+        break;
         default:
             UNREACHABLE();
             return gl::Error(GL_INVALID_OPERATION);
@@ -3564,6 +3574,9 @@ gl::Error Renderer11::compileToExecutable(gl::InfoLog &infoLog,
         case SHADER_GEOMETRY:
             profileStream << "gs";
             break;
+        case SHADER_COMPUTE:
+            profileStream << "cs";
+            break;
         default:
             UNREACHABLE();
             return gl::Error(GL_INVALID_OPERATION);

src/libANGLE/renderer/d3d/d3d11/ShaderExecutable11.cpp

@@ -17,27 +17,42 @@ ShaderExecutable11::ShaderExecutable11(const void *function, size_t length, ID3D
     : ShaderExecutableD3D(function, length)
 {
     mPixelExecutable = executable;
-    mVertexExecutable = NULL;
-    mGeometryExecutable = NULL;
-    mStreamOutExecutable = NULL;
+    mVertexExecutable    = nullptr;
+    mGeometryExecutable  = nullptr;
+    mStreamOutExecutable = nullptr;
+    mComputeExecutable   = nullptr;
 }
 
 ShaderExecutable11::ShaderExecutable11(const void *function, size_t length, ID3D11VertexShader *executable, ID3D11GeometryShader *streamOut)
     : ShaderExecutableD3D(function, length)
 {
     mVertexExecutable = executable;
-    mPixelExecutable = NULL;
-    mGeometryExecutable = NULL;
+    mPixelExecutable     = nullptr;
+    mGeometryExecutable  = nullptr;
     mStreamOutExecutable = streamOut;
+    mComputeExecutable   = nullptr;
 }
 
 ShaderExecutable11::ShaderExecutable11(const void *function, size_t length, ID3D11GeometryShader *executable)
     : ShaderExecutableD3D(function, length)
 {
     mGeometryExecutable = executable;
-    mVertexExecutable = NULL;
-    mPixelExecutable = NULL;
-    mStreamOutExecutable = NULL;
+    mVertexExecutable    = nullptr;
+    mPixelExecutable     = nullptr;
+    mStreamOutExecutable = nullptr;
+    mComputeExecutable   = nullptr;
+}
+
+ShaderExecutable11::ShaderExecutable11(const void *function,
+                                       size_t length,
+                                       ID3D11ComputeShader *executable)
+    : ShaderExecutableD3D(function, length)
+{
+    mComputeExecutable   = executable;
+    mPixelExecutable     = nullptr;
+    mVertexExecutable    = nullptr;
+    mGeometryExecutable  = nullptr;
+    mStreamOutExecutable = nullptr;
 }
 
 ShaderExecutable11::~ShaderExecutable11()
@@ -46,6 +61,7 @@ ShaderExecutable11::~ShaderExecutable11()
     SafeRelease(mPixelExecutable);
     SafeRelease(mGeometryExecutable);
     SafeRelease(mStreamOutExecutable);
+    SafeRelease(mComputeExecutable);
 }
 
 ID3D11VertexShader *ShaderExecutable11::getVertexShader() const
@@ -68,6 +84,11 @@ ID3D11GeometryShader *ShaderExecutable11::getStreamOutShader() const
     return mStreamOutExecutable;
 }
 
+ID3D11ComputeShader *ShaderExecutable11::getComputeShader() const
+{
+    return mComputeExecutable;
+}
+
 UniformStorage11::UniformStorage11(Renderer11 *renderer, size_t initialSize)
     : UniformStorageD3D(initialSize),
       mConstantBuffer(NULL)

src/libANGLE/renderer/d3d/d3d11/ShaderExecutable11.h

@@ -23,6 +23,7 @@ class ShaderExecutable11 : public ShaderExecutableD3D
     ShaderExecutable11(const void *function, size_t length, ID3D11PixelShader *executable);
     ShaderExecutable11(const void *function, size_t length, ID3D11VertexShader *executable, ID3D11GeometryShader *streamOut);
     ShaderExecutable11(const void *function, size_t length, ID3D11GeometryShader *executable);
+    ShaderExecutable11(const void *function, size_t length, ID3D11ComputeShader *executable);
 
     virtual ~ShaderExecutable11();
 
@@ -30,12 +31,14 @@ class ShaderExecutable11 : public ShaderExecutableD3D
     ID3D11VertexShader *getVertexShader() const;
     ID3D11GeometryShader *getGeometryShader() const;
     ID3D11GeometryShader *getStreamOutShader() const;
+    ID3D11ComputeShader *getComputeShader() const;
 
   private:
     ID3D11PixelShader *mPixelExecutable;
     ID3D11VertexShader *mVertexExecutable;
     ID3D11GeometryShader *mGeometryExecutable;
     ID3D11GeometryShader *mStreamOutExecutable;
+    ID3D11ComputeShader *mComputeExecutable;
 };
 
 class UniformStorage11 : public UniformStorageD3D

src/libANGLE/renderer/d3d/d3d11/renderer11_utils.cpp

@@ -830,6 +830,84 @@ size_t GetMaximumPixelTextureUnits(D3D_FEATURE_LEVEL featureLevel)
     }
 }
 
+std::array<GLuint, 3> GetMaxComputeWorkGroupCount(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return {{D3D11_CS_DISPATCH_MAX_THREAD_GROUPS_PER_DIMENSION,
+                     D3D11_CS_DISPATCH_MAX_THREAD_GROUPS_PER_DIMENSION,
+                     D3D11_CS_DISPATCH_MAX_THREAD_GROUPS_PER_DIMENSION}};
+            break;
+        default:
+            return {{0, 0, 0}};
+    }
+}
+
+std::array<GLuint, 3> GetMaxComputeWorkGroupSize(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return {{D3D11_CS_THREAD_GROUP_MAX_X, D3D11_CS_THREAD_GROUP_MAX_Y,
+                     D3D11_CS_THREAD_GROUP_MAX_Z}};
+            break;
+        default:
+            return {{0, 0, 0}};
+    }
+}
+
+size_t GetMaxComputeWorkGroupInvocations(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return D3D11_CS_THREAD_GROUP_MAX_THREADS_PER_GROUP;
+        default:
+            return 0;
+    }
+}
+
+size_t GetMaximumComputeUniformVectors(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return D3D11_REQ_CONSTANT_BUFFER_ELEMENT_COUNT;
+        default:
+            return 0;
+    }
+}
+
+size_t GetMaximumComputeUniformBlocks(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return D3D11_COMMONSHADER_CONSTANT_BUFFER_API_SLOT_COUNT -
+                   d3d11::RESERVED_CONSTANT_BUFFER_SLOT_COUNT;
+        default:
+            return 0;
+    }
+}
+
+size_t GetMaximumComputeTextureUnits(D3D_FEATURE_LEVEL featureLevel)
+{
+    switch (featureLevel)
+    {
+        case D3D_FEATURE_LEVEL_11_1:
+        case D3D_FEATURE_LEVEL_11_0:
+            return D3D11_COMMONSHADER_SAMPLER_SLOT_COUNT;
+        default:
+            return 0;
+    }
+}
+
 int GetMinimumTexelOffset(D3D_FEATURE_LEVEL featureLevel)
 {
     switch (featureLevel)
@@ -1160,6 +1238,18 @@ void GenerateCaps(ID3D11Device *device, ID3D11DeviceContext *deviceContext, cons
     caps->minProgramTexelOffset = GetMinimumTexelOffset(featureLevel);
     caps->maxProgramTexelOffset = GetMaximumTexelOffset(featureLevel);
 
+    // Compute shader limits
+    caps->maxComputeWorkGroupCount = GetMaxComputeWorkGroupCount(featureLevel);
+    caps->maxComputeWorkGroupSize  = GetMaxComputeWorkGroupSize(featureLevel);
+    caps->maxComputeWorkGroupInvocations =
+        static_cast<GLuint>(GetMaxComputeWorkGroupInvocations(featureLevel));
+    caps->maxComputeUniformComponents =
+        static_cast<GLuint>(GetMaximumComputeUniformVectors(featureLevel)) * 4;
+    caps->maxComputeUniformBlocks =
+        static_cast<GLuint>(GetMaximumComputeUniformBlocks(featureLevel));
+    caps->maxComputeTextureImageUnits =
+        static_cast<GLuint>(GetMaximumComputeTextureUnits(featureLevel));
+
     // Aggregate shader limits
     caps->maxUniformBufferBindings = caps->maxVertexUniformBlocks + caps->maxFragmentUniformBlocks;
     caps->maxUniformBlockSize = GetMaximumConstantBufferSize(featureLevel);
@@ -1175,6 +1265,9 @@ void GenerateCaps(ID3D11Device *device, ID3D11DeviceContext *deviceContext, cons
                                                static_cast<GLint64>(caps->maxVertexUniformComponents);
     caps->maxCombinedFragmentUniformComponents = (static_cast<GLint64>(caps->maxFragmentUniformBlocks) * static_cast<GLint64>(caps->maxUniformBlockSize / 4)) +
                                                  static_cast<GLint64>(caps->maxFragmentUniformComponents);
+    caps->maxCombinedComputeUniformComponents =
+        static_cast<GLuint>(caps->maxComputeUniformBlocks * (caps->maxUniformBlockSize / 4) +
+                            caps->maxComputeUniformComponents);
     caps->maxVaryingComponents =
         static_cast<GLuint>(GetMaximumVertexOutputVectors(featureLevel)) * 4;
     caps->maxVaryingVectors            = static_cast<GLuint>(GetMaximumVertexOutputVectors(featureLevel));

src/tests/gl_tests/ComputeShaderTest.cpp

@@ -190,6 +190,41 @@ TEST_P(ComputeShaderTest, AttachmentCount)
     EXPECT_GL_NO_ERROR();
 }
 
+// Access all compute shader special variables.
+TEST_P(ComputeShaderTest, AccessAllSpecialVariables)
+{
+    const std::string csSource =
+        "#version 310 es\n"
+        "layout(local_size_x=4, local_size_y=3, local_size_z=2) in;\n"
+        "void main()\n"
+        "{\n"
+        "    uvec3 temp1 = gl_NumWorkGroups;\n"
+        "    uvec3 temp2 = gl_WorkGroupSize;\n"
+        "    uvec3 temp3 = gl_WorkGroupID;\n"
+        "    uvec3 temp4 = gl_LocalInvocationID;\n"
+        "    uvec3 temp5 = gl_GlobalInvocationID;\n"
+        "    uint  temp6 = gl_LocalInvocationIndex;\n"
+        "}\n";
+
+    ANGLE_GL_COMPUTE_PROGRAM(program, csSource);
+}
+
+// Access part compute shader special variables.
+TEST_P(ComputeShaderTest, AccessPartSpecialVariables)
+{
+    const std::string csSource =
+        "#version 310 es\n"
+        "layout(local_size_x=4, local_size_y=3, local_size_z=2) in;\n"
+        "void main()\n"
+        "{\n"
+        "    uvec3 temp1 = gl_WorkGroupSize;\n"
+        "    uvec3 temp2 = gl_WorkGroupID;\n"
+        "    uint  temp3 = gl_LocalInvocationIndex;\n"
+        "}\n";
+
+    ANGLE_GL_COMPUTE_PROGRAM(program, csSource);
+}
+
 // Check that it is not possible to create a compute shader when the context does not support ES
 // 3.10
 TEST_P(ComputeShaderTestES3, NotSupported)
@@ -199,7 +234,7 @@ TEST_P(ComputeShaderTestES3, NotSupported)
     EXPECT_GL_ERROR(GL_INVALID_ENUM);
 }
 
-ANGLE_INSTANTIATE_TEST(ComputeShaderTest, ES31_OPENGL(), ES31_OPENGLES());
+ANGLE_INSTANTIATE_TEST(ComputeShaderTest, ES31_OPENGL(), ES31_OPENGLES(), ES31_D3D11());
 ANGLE_INSTANTIATE_TEST(ComputeShaderTestES3, ES3_OPENGL(), ES3_OPENGLES());
 
 }  // namespace

src/tests/gl_tests/ProgramBinaryTest.cpp

@@ -335,6 +335,69 @@ TEST_P(ProgramBinaryES3Test, UniformBlockBindingNoDraw)
 
 ANGLE_INSTANTIATE_TEST(ProgramBinaryES3Test, ES3_D3D11(), ES3_OPENGL(), ES3_OPENGLES());
 
+class ProgramBinaryES31Test : public ANGLETest
+{
+  protected:
+    ProgramBinaryES31Test()
+    {
+        setWindowWidth(128);
+        setWindowHeight(128);
+        setConfigRedBits(8);
+        setConfigGreenBits(8);
+        setConfigBlueBits(8);
+        setConfigAlphaBits(8);
+    }
+};
+
+// Tests that saving and loading a program attached with computer shader.
+TEST_P(ProgramBinaryES31Test, ProgramBinaryWithComputeShader)
+{
+    // We can't run the test if no program binary formats are supported.
+    GLint binaryFormatCount = 0;
+    glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &binaryFormatCount);
+    if (binaryFormatCount == 0)
+    {
+        std::cout << "Test skipped because no program binary formats available." << std::endl;
+        return;
+    }
+
+    const std::string &computeShader =
+        "#version 310 es\n"
+        "layout(local_size_x=4, local_size_y=3, local_size_z=1) in;\n"
+        "uniform block {\n"
+        "    vec2 f;\n"
+        "};\n"
+        "uniform vec2 g;\n"
+        "uniform highp sampler2D tex;\n"
+        "void main() {\n"
+        "    vec4 color = texture(tex, f + g);\n"
+        "}";
+
+    ANGLE_GL_COMPUTE_PROGRAM(program, computeShader);
+
+    // Read back the binary.
+    GLint programLength = 0;
+    glGetProgramiv(program.get(), GL_PROGRAM_BINARY_LENGTH, &programLength);
+    ASSERT_GL_NO_ERROR();
+
+    GLsizei readLength  = 0;
+    GLenum binaryFormat = GL_NONE;
+    std::vector<uint8_t> binary(programLength);
+    glGetProgramBinary(program.get(), programLength, &readLength, &binaryFormat, binary.data());
+    ASSERT_GL_NO_ERROR();
+
+    EXPECT_EQ(static_cast<GLsizei>(programLength), readLength);
+
+    // Load a new program with the binary.
+    ANGLE_GL_BINARY_ES3_PROGRAM(binaryProgram, binary, binaryFormat);
+
+    // TODO(Xinghua): add dispatch support when available.
+
+    ASSERT_GL_NO_ERROR();
+}
+
+ANGLE_INSTANTIATE_TEST(ProgramBinaryES31Test, ES31_D3D11(), ES31_OPENGL(), ES31_OPENGLES());
+
 class ProgramBinaryTransformFeedbackTest : public ANGLETest
 {
   protected:

src/tests/test_utils/angle_test_configs.cpp

@@ -491,6 +491,21 @@ PlatformParameters ES3_D3D11_FL10_1()
     return PlatformParameters(3, 0, egl_platform::D3D11_FL10_1());
 }
 
+PlatformParameters ES31_D3D11()
+{
+    return PlatformParameters(3, 1, egl_platform::D3D11());
+}
+
+PlatformParameters ES31_D3D11_FL11_1()
+{
+    return PlatformParameters(3, 1, egl_platform::D3D11_FL11_1());
+}
+
+PlatformParameters ES31_D3D11_FL11_0()
+{
+    return PlatformParameters(3, 1, egl_platform::D3D11_FL11_0());
+}
+
 PlatformParameters ES3_D3D11_WARP()
 {
     return PlatformParameters(3, 0, egl_platform::D3D11_WARP());

src/tests/test_utils/angle_test_configs.h

@@ -113,6 +113,9 @@ PlatformParameters ES3_D3D11();
 PlatformParameters ES3_D3D11_FL11_1();
 PlatformParameters ES3_D3D11_FL11_0();
 PlatformParameters ES3_D3D11_FL10_1();
+PlatformParameters ES31_D3D11();
+PlatformParameters ES31_D3D11_FL11_1();
+PlatformParameters ES31_D3D11_FL11_0();
 
 PlatformParameters ES3_D3D11_WARP();
 PlatformParameters ES3_D3D11_FL11_1_WARP();