Vulkan: Work around Nvidia depth clamping bug

include/platform/FeaturesVk.h

@@ -64,6 +64,13 @@ struct FeaturesVk : FeatureSetBase
         "The point size range reported from the API is inconsistent with the actual behavior",
         &members, "http://anglebug.com/2970"};
 
+    // On some NVIDIA drivers the depth value is not clamped to [0,1] for floating point depth
+    // buffers. This is NVIDIA bug 3171019, see http://anglebug.com/3970 for details.
+    Feature depthClamping = {
+        "depth_clamping", FeatureCategory::VulkanWorkarounds,
+        "The depth value is not clamped to [0,1] for floating point depth buffers.", &members,
+        "http://anglebug.com/3970"};
+
     // On some android devices, the memory barrier between the compute shader that converts vertex
     // attributes and the vertex shader that reads from it is ineffective.  Only known workaround is
     // to perform a flush after the conversion.  http://anglebug.com/3016

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -602,7 +602,7 @@ angle::Result ContextVk::initialize()
     }
 
     mGraphicsPipelineDesc.reset(new vk::GraphicsPipelineDesc());
-    mGraphicsPipelineDesc->initDefaults();
+    mGraphicsPipelineDesc->initDefaults(this);
 
     // Initialize current value/default attribute buffers.
     for (vk::DynamicBuffer &buffer : mDefaultAttribBuffers)

src/libANGLE/renderer/vulkan/RendererVk.cpp

@@ -1166,6 +1166,12 @@ angle::Result RendererVk::initializeDevice(DisplayVk *displayVk, uint32_t queueF
     // Initialize features and workarounds.
     initFeatures(displayVk, deviceExtensionNames);
 
+    // Enable VK_EXT_depth_clip_enable, if supported
+    if (ExtensionFound(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME, deviceExtensionNames))
+    {
+        enabledDeviceExtensions.push_back(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
+    }
+
     // Enable VK_KHR_get_memory_requirements2, if supported
     if (ExtensionFound(VK_KHR_GET_MEMORY_REQUIREMENTS_2_EXTENSION_NAME, deviceExtensionNames))
     {
@@ -1771,6 +1777,15 @@ void RendererVk::initFeatures(DisplayVk *displayVk,
     }
     ANGLE_FEATURE_CONDITION(&mFeatures, clampPointSize,
                             isNvidia && nvidiaVersion.major < uint32_t(IsWindows() ? 430 : 421));
+    // http://anglebug.com/3970#c25.
+    // The workaround requires the VK_EXT_depth_clip_enable extension and the 'depthClamp' physical
+    // device feature. This workaround caused test failures on Quadro P400/driver 418.56/Linux.
+    // Therefore, on Linux we require a major version > 418.
+    ANGLE_FEATURE_CONDITION(
+        &mFeatures, depthClamping,
+        isNvidia && mPhysicalDeviceFeatures.depthClamp &&
+            ExtensionFound(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME, deviceExtensionNames) &&
+            (!IsLinux() || nvidiaVersion.major > 418u));
 
     // Work around ineffective compute-graphics barriers on Nexus 5X.
     // TODO(syoussefi): Figure out which other vendors and driver versions are affected.

src/libANGLE/renderer/vulkan/UtilsVk.cpp

@@ -1517,7 +1517,7 @@ angle::Result UtilsVk::clearFramebuffer(ContextVk *contextVk,
     shaderParams.clearDepth = params.depthStencilClearValue.depth;
 
     vk::GraphicsPipelineDesc pipelineDesc;
-    pipelineDesc.initDefaults();
+    pipelineDesc.initDefaults(contextVk);
     pipelineDesc.setCullMode(VK_CULL_MODE_NONE);
     pipelineDesc.setColorWriteMasks(0, gl::DrawBufferMask(), gl::DrawBufferMask());
     pipelineDesc.setSingleColorWriteMask(params.colorAttachmentIndexGL, params.colorMaskFlags);
@@ -1763,7 +1763,7 @@ angle::Result UtilsVk::blitResolveImpl(ContextVk *contextVk,
         VK_COLOR_COMPONENT_A_BIT;
 
     vk::GraphicsPipelineDesc pipelineDesc;
-    pipelineDesc.initDefaults();
+    pipelineDesc.initDefaults(contextVk);
     if (blitColor)
     {
         pipelineDesc.setColorWriteMasks(
@@ -2204,7 +2204,7 @@ angle::Result UtilsVk::copyImage(ContextVk *contextVk,
     ASSERT(src->getSamples() == 1);
 
     vk::GraphicsPipelineDesc pipelineDesc;
-    pipelineDesc.initDefaults();
+    pipelineDesc.initDefaults(contextVk);
     pipelineDesc.setCullMode(VK_CULL_MODE_NONE);
     pipelineDesc.setRenderPassDesc(renderPassDesc);
     pipelineDesc.setRasterizationSamples(dest->getSamples());
@@ -2662,7 +2662,7 @@ angle::Result UtilsVk::unresolve(ContextVk *contextVk,
     ANGLE_TRY(ensureUnresolveResourcesInitialized(contextVk, function, totalBindingCount));
 
     vk::GraphicsPipelineDesc pipelineDesc;
-    pipelineDesc.initDefaults();
+    pipelineDesc.initDefaults(contextVk);
     pipelineDesc.setCullMode(VK_CULL_MODE_NONE);
     pipelineDesc.setRasterizationSamples(framebuffer->getSamples());
     pipelineDesc.setRenderPassDesc(framebuffer->getRenderPassDesc());

src/libANGLE/renderer/vulkan/vk_cache_utils.cpp

@@ -1490,7 +1490,7 @@ bool GraphicsPipelineDesc::operator==(const GraphicsPipelineDesc &other) const
 // TODO(jmadill): We should prefer using Packed GLenums. http://anglebug.com/2169
 
 // Initialize PSO states, it is consistent with initial value of gl::State
-void GraphicsPipelineDesc::initDefaults()
+void GraphicsPipelineDesc::initDefaults(const ContextVk *contextVk)
 {
     // Set all vertex input attributes to default, the default format is Float
     angle::FormatID defaultFormat = GetCurrentValueFormatID(gl::VertexAttribType::Float);
@@ -1503,8 +1503,9 @@ void GraphicsPipelineDesc::initDefaults()
         SetBitField(packedAttrib.offset, 0);
     }
 
-    mRasterizationAndMultisampleStateInfo.bits.subpass                    = 0;
-    mRasterizationAndMultisampleStateInfo.bits.depthClampEnable           = 0;
+    mRasterizationAndMultisampleStateInfo.bits.subpass = 0;
+    mRasterizationAndMultisampleStateInfo.bits.depthClampEnable =
+        contextVk->getFeatures().depthClamping.enabled ? VK_TRUE : VK_FALSE;
     mRasterizationAndMultisampleStateInfo.bits.rasterizationDiscardEnable = 0;
     SetBitField(mRasterizationAndMultisampleStateInfo.bits.polygonMode, VK_POLYGON_MODE_FILL);
     SetBitField(mRasterizationAndMultisampleStateInfo.bits.cullMode, VK_CULL_MODE_BACK_BIT);
@@ -1824,6 +1825,21 @@ angle::Result GraphicsPipelineDesc::initializePipeline(
         *pNextPtr                                = &provokingVertexState;
         pNextPtr                                 = &provokingVertexState.pNext;
     }
+
+    // When depth clamping is used, depth clipping is automatically disabled.
+    // When the 'depthClamping' feature is enabled, we'll be using depth clamping
+    // to work around a driver issue, not as an alternative to depth clipping. Therefore we need to
+    // explicitly re-enable depth clipping.
+    VkPipelineRasterizationDepthClipStateCreateInfoEXT depthClipState = {};
+    depthClipState.sType =
+        VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT;
+    if (contextVk->getFeatures().depthClamping.enabled)
+    {
+        depthClipState.depthClipEnable = VK_TRUE;
+        *pNextPtr                      = &depthClipState;
+        pNextPtr                       = &depthClipState.pNext;
+    }
+
     VkPipelineRasterizationStateStreamCreateInfoEXT rasterStreamState = {};
     rasterStreamState.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_STREAM_CREATE_INFO_EXT;
     if (contextVk->getFeatures().supportsTransformFeedbackExtension.enabled)

src/libANGLE/renderer/vulkan/vk_cache_utils.h

@@ -574,7 +574,7 @@ class GraphicsPipelineDesc final
     size_t hash() const;
     bool operator==(const GraphicsPipelineDesc &other) const;
 
-    void initDefaults();
+    void initDefaults(const ContextVk *contextVk);
 
     // For custom comparisons.
     template <typename T>

src/tests/deqp_support/deqp_gles3_test_expectations.txt

@@ -533,14 +533,16 @@
 3677 VULKAN : dEQP-GLES3.functional.fragment_out.random.* = FAIL
 
 // NVIDIA Depth/Stencil textures:
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.fbo.depth.depth_write_clamp.depth_component32f = FAIL
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.fbo.depth.depth_write_clamp.depth32f_stencil8 = FAIL
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.fbo.depth.depth_test_clamp.depth_component32f = FAIL
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.fbo.depth.depth_test_clamp.depth32f_stencil8 = FAIL
+// Fails on Quadro P400, driver 418.56, Linux
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.fbo.depth.depth_write_clamp.depth_component32f = FAIL
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.fbo.depth.depth_write_clamp.depth32f_stencil8 = FAIL
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.fbo.depth.depth_test_clamp.depth_component32f = FAIL
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.fbo.depth.depth_test_clamp.depth32f_stencil8 = FAIL
 
 // Polygon offset:
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.polygon_offset.float32_result_depth_clamp = FAIL
-3970 VULKAN NVIDIA : dEQP-GLES3.functional.polygon_offset.float32_factor_1_slope = FAIL
+// Fails on Quadro P400, driver 418.56, Linux
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.polygon_offset.float32_result_depth_clamp = FAIL
+3970 VULKAN NVIDIA LINUX : dEQP-GLES3.functional.polygon_offset.float32_factor_1_slope = FAIL
 
 // Transform Feedback
 4666 VULKAN NVIDIA : dEQP-GLES3.functional.transform_feedback.array_element.interleaved.* = SKIP