Vulkan: Improve cubemap emulation seam handling

include/GLSLANG/ShaderLang.h

@@ -290,12 +290,7 @@ const ShCompileOptions SH_EMULATE_GL_BASE_VERTEX_BASE_INSTANCE = UINT64_C(1) <<
 // Emulate seamful cube map sampling for OpenGL ES2.0.  Currently only applies to the Vulkan
 // backend, as is done after samplers are moved out of structs.  Can likely be made to work on
 // the other backends as well.
-//
-// There are two variations of this.  One using subgroup operations where available, and another
-// that emulates those operations using dFdxFine and dFdyFine.  The latter is more universally
-// available, but is buggy on Nvidia.
 const ShCompileOptions SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING = UINT64_C(1) << 44;
-const ShCompileOptions SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING_WITH_SUBGROUP_OP = UINT64_C(1) << 45;
 
 // If requested, validates the AST after every transformation.  Useful for debugging.
 const ShCompileOptions SH_VALIDATE_AST = UINT64_C(1) << 46;

src/compiler/translator/TranslatorVulkan.cpp

@@ -665,11 +665,6 @@ bool TranslatorVulkan::translate(TIntermBlock *root,
 
     sink << "#version 450 core\n";
 
-    if (compileOptions & SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING_WITH_SUBGROUP_OP)
-    {
-        sink << "#extension GL_KHR_shader_subgroup_quad : require\n";
-    }
-
     // Write out default uniforms into a uniform block assigned to a specific set/binding.
     int defaultUniformCount           = 0;
     int aggregateTypesUsedForUniforms = 0;
@@ -732,12 +727,10 @@ bool TranslatorVulkan::translate(TIntermBlock *root,
 
     // Rewrite samplerCubes as sampler2DArrays.  This must be done after rewriting struct samplers
     // as it doesn't expect that.
-    if (compileOptions & (SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING |
-                          SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING_WITH_SUBGROUP_OP))
+    if (compileOptions & SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING)
     {
-        if (!RewriteCubeMapSamplersAs2DArray(
-                this, root, &getSymbolTable(), getShaderType() == GL_FRAGMENT_SHADER,
-                compileOptions & SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING_WITH_SUBGROUP_OP))
+        if (!RewriteCubeMapSamplersAs2DArray(this, root, &getSymbolTable(),
+                                             getShaderType() == GL_FRAGMENT_SHADER))
         {
             return false;
         }

src/compiler/translator/tree_ops/RewriteCubeMapSamplersAs2DArray.h

@@ -23,8 +23,7 @@ class TSymbolTable;
 ANGLE_NO_DISCARD bool RewriteCubeMapSamplersAs2DArray(TCompiler *compiler,
                                                       TIntermBlock *root,
                                                       TSymbolTable *symbolTable,
-                                                      bool isFragmentShader,
-                                                      bool useSubgroupOps);
+                                                      bool isFragmentShader);
 }  // namespace sh
 
 #endif  // COMPILER_TRANSLATOR_TREEOPS_REWRITECUBEMAPSAMPLERSAS2DARRAY_H_

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -239,7 +239,6 @@ ContextVk::ContextVk(const gl::State &state, gl::ErrorSet *errorSet, RendererVk
       mFlipYForCurrentSurface(false),
       mIsAnyHostVisibleBufferWritten(false),
       mEmulateSeamfulCubeMapSampling(false),
-      mEmulateSeamfulCubeMapSamplingWithSubgroupOps(false),
       mUseOldRewriteStructSamplers(false),
       mLastCompletedQueueSerial(renderer->nextSerial()),
       mCurrentQueueSerial(renderer->nextSerial()),
@@ -444,8 +443,7 @@ angle::Result ContextVk::initialize()
         ANGLE_TRY(synchronizeCpuGpuTime());
     }
 
-    mEmulateSeamfulCubeMapSampling =
-        shouldEmulateSeamfulCubeMapSampling(&mEmulateSeamfulCubeMapSamplingWithSubgroupOps);
+    mEmulateSeamfulCubeMapSampling = shouldEmulateSeamfulCubeMapSampling();
 
     mUseOldRewriteStructSamplers = shouldUseOldRewriteStructSamplers();
 
@@ -2926,7 +2924,7 @@ vk::DescriptorSetLayoutDesc ContextVk::getDriverUniformsDescriptorSetDesc(
     return desc;
 }
 
-bool ContextVk::shouldEmulateSeamfulCubeMapSampling(bool *useSubgroupOpsOut) const
+bool ContextVk::shouldEmulateSeamfulCubeMapSampling() const
 {
     // Only allow seamful cube map sampling in non-webgl ES2.
     if (mState.getClientMajorVersion() != 2 || mState.isWebGL())
@@ -2939,15 +2937,6 @@ bool ContextVk::shouldEmulateSeamfulCubeMapSampling(bool *useSubgroupOpsOut) con
         return false;
     }
 
-    // Use subgroup ops where available.
-    constexpr VkSubgroupFeatureFlags kSeamfulCubeMapSubgroupOperations =
-        VK_SUBGROUP_FEATURE_BASIC_BIT | VK_SUBGROUP_FEATURE_BALLOT_BIT |
-        VK_SUBGROUP_FEATURE_QUAD_BIT;
-    const VkSubgroupFeatureFlags deviceSupportedOperations =
-        mRenderer->getPhysicalDeviceSubgroupProperties().supportedOperations;
-    *useSubgroupOpsOut = (deviceSupportedOperations & kSeamfulCubeMapSubgroupOperations) ==
-                         kSeamfulCubeMapSubgroupOperations;
-
     return true;
 }
 

src/libANGLE/renderer/vulkan/ContextVk.h

@@ -327,11 +327,7 @@ class ContextVk : public ContextImpl, public vk::Context, public vk::RenderPassO
 
     void updateScissor(const gl::State &glState);
 
-    bool emulateSeamfulCubeMapSampling(bool *useSubgroupOpsOut) const
-    {
-        *useSubgroupOpsOut = mEmulateSeamfulCubeMapSamplingWithSubgroupOps;
-        return mEmulateSeamfulCubeMapSampling;
-    }
+    bool emulateSeamfulCubeMapSampling() const { return mEmulateSeamfulCubeMapSampling; }
 
     bool useOldRewriteStructSamplers() const { return mUseOldRewriteStructSamplers; }
 
@@ -492,7 +488,7 @@ class ContextVk : public ContextImpl, public vk::Context, public vk::RenderPassO
 
     void waitForSwapchainImageIfNecessary();
 
-    bool shouldEmulateSeamfulCubeMapSampling(bool *useSubgroupOpsOut) const;
+    bool shouldEmulateSeamfulCubeMapSampling() const;
 
     bool shouldUseOldRewriteStructSamplers() const;
 
@@ -557,10 +553,8 @@ class ContextVk : public ContextImpl, public vk::Context, public vk::RenderPassO
     // at the end of the command buffer to make that write available to the host.
     bool mIsAnyHostVisibleBufferWritten;
 
-    // Whether this context should do seamful cube map sampling emulation, and whether subgroup
-    // operations should be used.
+    // Whether this context should do seamful cube map sampling emulation.
     bool mEmulateSeamfulCubeMapSampling;
-    bool mEmulateSeamfulCubeMapSamplingWithSubgroupOps;
 
     // Whether this context should use the old version of the
     // RewriteStructSamplers pass.

src/libANGLE/renderer/vulkan/GlslangWrapper.cpp

@@ -951,7 +951,6 @@ void GlslangWrapper::GetShaderSource(bool useOldRewriteStructSamplers,
 angle::Result GlslangWrapper::GetShaderCode(vk::Context *context,
                                             const gl::Caps &glCaps,
                                             bool enableLineRasterEmulation,
-                                            bool enableSubgroupOps,
                                             const gl::ShaderMap<std::string> &shaderSources,
                                             gl::ShaderMap<std::vector<uint32_t>> *shaderCodeOut)
 {
@@ -971,18 +970,17 @@ angle::Result GlslangWrapper::GetShaderCode(vk::Context *context,
                                                kVersionDefine, kLineRasterDefine),
                        VK_ERROR_INVALID_SHADER_NV);
 
-        return GetShaderCodeImpl(context, glCaps, enableSubgroupOps, patchedSources, shaderCodeOut);
+        return GetShaderCodeImpl(context, glCaps, patchedSources, shaderCodeOut);
     }
     else
     {
-        return GetShaderCodeImpl(context, glCaps, enableSubgroupOps, shaderSources, shaderCodeOut);
+        return GetShaderCodeImpl(context, glCaps, shaderSources, shaderCodeOut);
     }
 }
 
 // static
 angle::Result GlslangWrapper::GetShaderCodeImpl(vk::Context *context,
                                                 const gl::Caps &glCaps,
-                                                bool enableSubgroupOps,
                                                 const gl::ShaderMap<std::string> &shaderSources,
                                                 gl::ShaderMap<std::vector<uint32_t>> *shaderCodeOut)
 {
@@ -1018,11 +1016,6 @@ angle::Result GlslangWrapper::GetShaderCodeImpl(vk::Context *context,
         glslang::TShader *shader = shaders[shaderType];
         shader->setStringsWithLengths(&shaderString, &shaderLength, 1);
         shader->setEntryPoint("main");
-        if (enableSubgroupOps)
-        {
-            // Enable SPIR-V 1.3 if to be able to use subgroup operations.
-            shader->setEnvTarget(glslang::EShTargetSpv, glslang::EShTargetSpv_1_3);
-        }
 
         bool result = shader->parse(&builtInResources, 450, ECoreProfile, false, false, messages);
         if (!result)

src/libANGLE/renderer/vulkan/GlslangWrapper.h

@@ -30,14 +30,12 @@ class GlslangWrapper
     static angle::Result GetShaderCode(vk::Context *context,
                                        const gl::Caps &glCaps,
                                        bool enableLineRasterEmulation,
-                                       bool enableSubgroupOps,
                                        const gl::ShaderMap<std::string> &shaderSources,
                                        gl::ShaderMap<std::vector<uint32_t>> *shaderCodesOut);
 
   private:
     static angle::Result GetShaderCodeImpl(vk::Context *context,
                                            const gl::Caps &glCaps,
-                                           bool enableSubgroupOps,
                                            const gl::ShaderMap<std::string> &shaderSources,
                                            gl::ShaderMap<std::vector<uint32_t>> *shaderCodesOut);
 };

src/libANGLE/renderer/vulkan/ProgramVk.cpp

@@ -304,16 +304,9 @@ angle::Result ProgramVk::ShaderInfo::initShaders(ContextVk *contextVk,
 {
     ASSERT(!valid());
 
-    bool useSubgroupOpsWithSeamfulCubeMapEmulation = false;
-    bool emulateSeamfulCubeMapSampling =
-        contextVk->emulateSeamfulCubeMapSampling(&useSubgroupOpsWithSeamfulCubeMapEmulation);
-    bool useSubgroupOps =
-        emulateSeamfulCubeMapSampling && useSubgroupOpsWithSeamfulCubeMapEmulation;
-
     gl::ShaderMap<std::vector<uint32_t>> shaderCodes;
-    ANGLE_TRY(GlslangWrapper::GetShaderCode(contextVk, contextVk->getCaps(),
-                                            enableLineRasterEmulation, useSubgroupOps,
-                                            shaderSources, &shaderCodes));
+    ANGLE_TRY(GlslangWrapper::GetShaderCode(
+        contextVk, contextVk->getCaps(), enableLineRasterEmulation, shaderSources, &shaderCodes));
 
     for (const gl::ShaderType shaderType : gl::AllShaderTypes())
     {
@@ -1475,8 +1468,7 @@ angle::Result ProgramVk::updateTexturesDescriptorSet(ContextVk *contextVk)
 
     const gl::ActiveTextureArray<vk::TextureUnit> &activeTextures = contextVk->getActiveTextures();
 
-    bool useSubgroupOps                = false;
-    bool emulateSeamfulCubeMapSampling = contextVk->emulateSeamfulCubeMapSampling(&useSubgroupOps);
+    bool emulateSeamfulCubeMapSampling = contextVk->emulateSeamfulCubeMapSampling();
     bool useOldRewriteStructSamplers   = contextVk->useOldRewriteStructSamplers();
 
     std::unordered_map<std::string, uint32_t> mappedSamplerNameToBindingIndex;

src/libANGLE/renderer/vulkan/ShaderVk.cpp

@@ -40,17 +40,9 @@ std::shared_ptr<WaitableCompileEvent> ShaderVk::compile(const gl::Context *conte
         compileOptions |= SH_CLAMP_POINT_SIZE;
     }
 
-    bool useSubgroupOps = false;
-    if (contextVk->emulateSeamfulCubeMapSampling(&useSubgroupOps))
+    if (contextVk->emulateSeamfulCubeMapSampling())
     {
-        if (useSubgroupOps)
-        {
-            compileOptions |= SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING_WITH_SUBGROUP_OP;
-        }
-        else
-        {
-            compileOptions |= SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING;
-        }
+        compileOptions |= SH_EMULATE_SEAMFUL_CUBE_MAP_SAMPLING;
     }
 
     if (contextVk->useOldRewriteStructSamplers())

src/tests/gl_tests/CubeMapTextureTest.cpp

@@ -46,6 +46,8 @@ class CubeMapTextureTest : public ANGLETest
 
     void testTearDown() override { glDeleteProgram(mProgram); }
 
+    void runSampleCoordinateTransformTest(const char *shader);
+
     GLuint mProgram;
     GLint mColorLocation;
 };
@@ -113,9 +115,7 @@ TEST_P(CubeMapTextureTest, RenderToFacesConsecutively)
     EXPECT_GL_NO_ERROR();
 }
 
-// Verify that cube map sampling follows the rules that map cubemap coordinates to coordinates
-// within each face.  See section 3.7.5 of GLES2.0 (Cube Map Texture Selection).
-TEST_P(CubeMapTextureTest, SampleCoordinateTransform)
+void CubeMapTextureTest::runSampleCoordinateTransformTest(const char *shader)
 {
     // Fails to compile the shader.  anglebug.com/3776
     ANGLE_SKIP_TEST_IF(IsOpenGL() && IsIntel() && IsWindows());
@@ -182,42 +182,7 @@ TEST_P(CubeMapTextureTest, SampleCoordinateTransform)
     glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, fboTex, 0);
     EXPECT_GL_NO_ERROR();
 
-    // Create a program that samples from 6x4 directions of the cubemap, draw and verify that the
-    // colors match the right color from |faceColors|.
-    constexpr char kFS[] = R"(precision mediump float;
-
-uniform samplerCube texCube;
-
-const mat4 coordInSection = mat4(
-    vec4(-0.5, -0.5, 0, 0),
-    vec4( 0.5, -0.5, 0, 0),
-    vec4(-0.5,  0.5, 0, 0),
-    vec4( 0.5,  0.5, 0, 0)
-);
-
-void main()
-{
-    vec3 coord;
-    if (gl_FragCoord.x < 2.0)
-    {
-        coord.x = gl_FragCoord.x < 1.0 ? 1.0 : -1.0;
-        coord.zy = coordInSection[int(gl_FragCoord.y)].xy;
-    }
-    else if (gl_FragCoord.x < 4.0)
-    {
-        coord.y = gl_FragCoord.x < 3.0 ? 1.0 : -1.0;
-        coord.xz = coordInSection[int(gl_FragCoord.y)].xy;
-    }
-    else
-    {
-        coord.z = gl_FragCoord.x < 5.0 ? 1.0 : -1.0;
-        coord.xy = coordInSection[int(gl_FragCoord.y)].xy;
-    }
-
-    gl_FragColor = textureCube(texCube, coord);
-})";
-
-    ANGLE_GL_PROGRAM(program, essl1_shaders::vs::Simple(), kFS);
+    ANGLE_GL_PROGRAM(program, essl1_shaders::vs::Simple(), shader);
     glUseProgram(program);
 
     GLint texCubeLocation = glGetUniformLocation(program, "texCube");
@@ -261,6 +226,93 @@ void main()
     EXPECT_GL_NO_ERROR();
 }
 
+// Verify that cube map sampling follows the rules that map cubemap coordinates to coordinates
+// within each face.  See section 3.7.5 of GLES2.0 (Cube Map Texture Selection).
+TEST_P(CubeMapTextureTest, SampleCoordinateTransform)
+{
+    // Create a program that samples from 6x4 directions of the cubemap, draw and verify that the
+    // colors match the right color from |faceColors|.
+    constexpr char kFS[] = R"(precision mediump float;
+
+uniform samplerCube texCube;
+
+const mat4 coordInSection = mat4(
+    vec4(-0.5, -0.5, 0, 0),
+    vec4( 0.5, -0.5, 0, 0),
+    vec4(-0.5,  0.5, 0, 0),
+    vec4( 0.5,  0.5, 0, 0)
+);
+
+void main()
+{
+    vec3 coord;
+    if (gl_FragCoord.x < 2.0)
+    {
+        coord.x = gl_FragCoord.x < 1.0 ? 1.0 : -1.0;
+        coord.zy = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+    else if (gl_FragCoord.x < 4.0)
+    {
+        coord.y = gl_FragCoord.x < 3.0 ? 1.0 : -1.0;
+        coord.xz = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+    else
+    {
+        coord.z = gl_FragCoord.x < 5.0 ? 1.0 : -1.0;
+        coord.xy = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+
+    gl_FragColor = textureCube(texCube, coord);
+})";
+
+    runSampleCoordinateTransformTest(kFS);
+}
+
+// On Android Vulkan, unequal x and y derivatives cause this test to fail.
+TEST_P(CubeMapTextureTest, SampleCoordinateTransformGrad)
+{
+    ANGLE_SKIP_TEST_IF(IsAndroid() && IsVulkan());  // anglebug.com/3814
+    ANGLE_SKIP_TEST_IF(IsD3D11());                  // anglebug.com/3856
+    ANGLE_SKIP_TEST_IF(!IsGLExtensionEnabled("GL_EXT_shader_texture_lod"));
+
+    constexpr char kFS[] = R"(#extension GL_EXT_shader_texture_lod : require
+precision mediump float;
+
+uniform samplerCube texCube;
+
+const mat4 coordInSection = mat4(
+    vec4(-0.5, -0.5, 0, 0),
+    vec4( 0.5, -0.5, 0, 0),
+    vec4(-0.5,  0.5, 0, 0),
+    vec4( 0.5,  0.5, 0, 0)
+);
+
+void main()
+{
+    vec3 coord;
+    if (gl_FragCoord.x < 2.0)
+    {
+        coord.x = gl_FragCoord.x < 1.0 ? 1.0 : -1.0;
+        coord.zy = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+    else if (gl_FragCoord.x < 4.0)
+    {
+        coord.y = gl_FragCoord.x < 3.0 ? 1.0 : -1.0;
+        coord.xz = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+    else
+    {
+        coord.z = gl_FragCoord.x < 5.0 ? 1.0 : -1.0;
+        coord.xy = coordInSection[int(gl_FragCoord.y)].xy;
+    }
+
+    gl_FragColor = textureCubeGradEXT(texCube, coord,
+                                      vec3(10.0, 10.0, 0.0), vec3(0.0, 10.0, 10.0));
+})";
+
+    runSampleCoordinateTransformTest(kFS);
+}
+
 // Use this to select which configurations (e.g. which renderer, which GLES major version) these
 // tests should be run against.
 ANGLE_INSTANTIATE_TEST(CubeMapTextureTest,