Vulkan: Emulate instanced attrib divisor

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -477,12 +477,14 @@ angle::Result ContextVk::setupDraw(const gl::Context *context,
     }
 
     // Must be called before the command buffer is started. Can call finish.
-    if (context->getStateCache().hasAnyActiveClientAttrib())
+    if (mVertexArray->getStreamingVertexAttribsMask().any())
     {
         ASSERT(firstVertexOrInvalid != -1);
-        ANGLE_TRY(mVertexArray->updateClientAttribs(context, firstVertexOrInvalid,
-                                                    vertexOrIndexCount, instanceCount,
-                                                    indexTypeOrInvalid, indices));
+        // All client attribs & any emulated buffered attribs will be updated
+        ANGLE_TRY(mVertexArray->updateStreamedAttribs(context, firstVertexOrInvalid,
+                                                      vertexOrIndexCount, instanceCount,
+                                                      indexTypeOrInvalid, indices));
+
         mGraphicsDirtyBits.set(DIRTY_BIT_VERTEX_BUFFERS);
     }
 

src/libANGLE/renderer/vulkan/ContextVk.h

@@ -213,9 +213,10 @@ class ContextVk : public ContextImpl, public vk::Context, public vk::RenderPassO
                                               GLuint relativeOffset)
     {
         invalidateVertexAndIndexBuffers();
-        mGraphicsPipelineDesc->updateVertexInput(&mGraphicsPipelineTransition,
-                                                 static_cast<uint32_t>(attribIndex), stride,
-                                                 divisor, format, relativeOffset);
+        // Set divisor to 1 for attribs with emulated divisor
+        mGraphicsPipelineDesc->updateVertexInput(
+            &mGraphicsPipelineTransition, static_cast<uint32_t>(attribIndex), stride,
+            divisor > mRenderer->getMaxVertexAttribDivisor() ? 1 : divisor, format, relativeOffset);
     }
 
     void invalidateDefaultAttribute(size_t attribIndex);

src/libANGLE/renderer/vulkan/RendererVk.cpp

@@ -1056,7 +1056,11 @@ angle::Result RendererVk::initializeDevice(DisplayVk *displayVk, uint32_t queueF
         deviceProperties.pNext = &divisorProperties;
 
         vkGetPhysicalDeviceProperties2KHR(mPhysicalDevice, &deviceProperties);
-        mMaxVertexAttribDivisor = divisorProperties.maxVertexAttribDivisor;
+        // We only store 8 bit divisor in GraphicsPipelineDesc so capping value & we emulate if
+        // exceeded
+        mMaxVertexAttribDivisor =
+            std::min(divisorProperties.maxVertexAttribDivisor,
+                     static_cast<uint32_t>(std::numeric_limits<uint8_t>::max()));
 
         createInfo.pNext = &enabledFeatures;
     }

src/libANGLE/renderer/vulkan/RendererVk.h

@@ -125,6 +125,7 @@ class RendererVk : angle::NonCopyable
         ASSERT(mFeaturesInitialized);
         return mFeatures;
     }
+    uint32_t getMaxVertexAttribDivisor() const { return mMaxVertexAttribDivisor; }
 
     bool isMockICDEnabled() const { return mEnabledICD == vk::ICD::Mock; }
 

src/libANGLE/renderer/vulkan/VertexArrayVk.h

@@ -37,12 +37,12 @@ class VertexArrayVk : public VertexArrayImpl
                              VkBuffer bufferHandle,
                              uint32_t offset);
 
-    angle::Result updateClientAttribs(const gl::Context *context,
-                                      GLint firstVertex,
-                                      GLsizei vertexOrIndexCount,
-                                      GLsizei instanceCount,
-                                      gl::DrawElementsType indexTypeOrInvalid,
-                                      const void *indices);
+    angle::Result updateStreamedAttribs(const gl::Context *context,
+                                        GLint firstVertex,
+                                        GLsizei vertexOrIndexCount,
+                                        GLsizei instanceCount,
+                                        gl::DrawElementsType indexTypeOrInvalid,
+                                        const void *indices);
 
     angle::Result handleLineLoop(ContextVk *contextVk,
                                  GLint firstVertex,
@@ -92,6 +92,11 @@ class VertexArrayVk : public VertexArrayImpl
                                         size_t indexCount,
                                         const void *sourcePointer);
 
+    const gl::AttributesMask &getStreamingVertexAttribsMask() const
+    {
+        return mStreamingVertexAttribsMask;
+    }
+
   private:
     void setDefaultPackedInput(ContextVk *contextVk, size_t attribIndex);
 
@@ -133,6 +138,9 @@ class VertexArrayVk : public VertexArrayImpl
 
     // Vulkan does not allow binding a null vertex buffer. We use a dummy as a placeholder.
     vk::BufferHelper mTheNullBuffer;
+
+    // Track client and/or emulated attribs that we have to stream their buffer contents
+    gl::AttributesMask mStreamingVertexAttribsMask;
 };
 }  // namespace rx
 

src/libANGLE/renderer/vulkan/vk_cache_utils.cpp

@@ -16,6 +16,7 @@
 #include "libANGLE/renderer/vulkan/FramebufferVk.h"
 #include "libANGLE/renderer/vulkan/ProgramVk.h"
 #include "libANGLE/renderer/vulkan/RendererVk.h"
+#include "libANGLE/renderer/vulkan/VertexArrayVk.h"
 #include "libANGLE/renderer/vulkan/vk_format_utils.h"
 #include "libANGLE/renderer/vulkan/vk_helpers.h"
 
@@ -673,7 +674,6 @@ angle::Result GraphicsPipelineDesc::initializePipeline(
     VkPipelineVertexInputDivisorStateCreateInfoEXT divisorState = {};
     divisorState.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT;
     divisorState.pVertexBindingDivisors = divisorDesc.data();
-
     for (size_t attribIndexSizeT : activeAttribLocationsMask)
     {
         const uint32_t attribIndex = static_cast<uint32_t>(attribIndexSizeT);
@@ -885,10 +885,6 @@ void GraphicsPipelineDesc::updateVertexInput(GraphicsPipelineTransitionBits *tra
 {
     vk::PackedAttribDesc &packedAttrib = mVertexInputAttribs.attribs[attribIndex];
 
-    // TODO: Handle the case where the divisor overflows the field that holds it.
-    // http://anglebug.com/2672
-    ASSERT(divisor <= std::numeric_limits<decltype(packedAttrib.divisor)>::max());
-
     SetBitField(packedAttrib.stride, stride);
     SetBitField(packedAttrib.divisor, divisor);
 

src/libANGLE/renderer/vulkan/vk_cache_utils.h

@@ -38,7 +38,7 @@ using RefCountedPipelineLayout      = RefCounted<PipelineLayout>;
 // fewer bits. For example, boolean values could be represented by a single bit instead
 // of a uint8_t. However at the current time there are concerns about the portability
 // of bitfield operators, and complexity issues with using bit mask operations. This is
-// something likely we will want to investigate as the Vulkan implementation progresses.
+// something we will likely want to investigate as the Vulkan implementation progresses.
 //
 // Second implementation note: the struct packing is also a bit fragile, and some of the
 // packing requirements depend on using alignas and field ordering to get the result of
@@ -164,8 +164,6 @@ static_assert(sizeof(AttachmentOpsArray) == 20, "Size check failed");
 struct PackedAttribDesc final
 {
     uint8_t format;
-
-    // TODO(http://anglebug.com/2672): Emulate divisors greater than UBYTE_MAX.
     uint8_t divisor;
 
     // Can only take 11 bits on NV.
@@ -315,9 +313,9 @@ constexpr size_t kPackedInputAssemblyAndColorBlendStateSize =
 static_assert(kPackedInputAssemblyAndColorBlendStateSize == 56, "Size check failed");
 
 constexpr size_t kGraphicsPipelineDescSumOfSizes =
-    kVertexInputAttributesSize + kPackedInputAssemblyAndColorBlendStateSize +
-    kPackedRasterizationAndMultisampleStateSize + kPackedDepthStencilStateSize +
-    kRenderPassDescSize + sizeof(VkViewport) + sizeof(VkRect2D);
+    kVertexInputAttributesSize + kRenderPassDescSize + kPackedRasterizationAndMultisampleStateSize +
+    kPackedDepthStencilStateSize + kPackedInputAssemblyAndColorBlendStateSize + sizeof(VkViewport) +
+    sizeof(VkRect2D);
 
 // Number of dirty bits in the dirty bit set.
 constexpr size_t kGraphicsPipelineDirtyBitBytes = 4;

src/libANGLE/renderer/vulkan/vk_caps_utils.cpp

@@ -82,10 +82,9 @@ void RendererVk::ensureCapsInitialized() const
 
     mNativeExtensions.vertexHalfFloat = true;
 
-    // TODO: Enable this always and emulate instanced draws if any divisor exceeds the maximum
-    // supported.  http://anglebug.com/2672
-    mNativeExtensions.instancedArraysANGLE = mMaxVertexAttribDivisor > 1;
-    mNativeExtensions.instancedArraysEXT   = mMaxVertexAttribDivisor > 1;
+    // Enabled in HW if VK_EXT_vertex_attribute_divisor available, otherwise emulated
+    mNativeExtensions.instancedArraysANGLE = true;
+    mNativeExtensions.instancedArraysEXT   = true;
 
     // Only expose robust buffer access if the physical device supports it.
     mNativeExtensions.robustBufferAccessBehavior =

src/tests/deqp_support/deqp_gles3_test_expectations.txt

@@ -555,16 +555,6 @@
 3219 VULKAN : dEQP-GLES3.functional.negative_api.shader.link_program = FAIL
 3219 VULKAN : dEQP-GLES3.functional.negative_api.shader.use_program = FAIL
 
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_elements_instanced.attribute_divisor.2*_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_elements_instanced.attribute_divisor.4_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_elements_instanced.mixed.2*_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_elements_instanced.mixed.4_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_arrays_instanced.attribute_divisor.2*_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_arrays_instanced.attribute_divisor.4_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_arrays_instanced.mixed.2*_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.draw_arrays_instanced.mixed.4_instances = FAIL
-2672 VULKAN : dEQP-GLES3.functional.instanced.types* = FAIL
-
 // Polygon offset:
 3678 VULKAN : dEQP-GLES3.functional.polygon_offset.float32_result_depth_clamp = FAIL
 3678 VULKAN : dEQP-GLES3.functional.polygon_offset.float32_factor_1_slope = FAIL

src/tests/gl_tests/InstancingTest.cpp

@@ -588,7 +588,7 @@ TEST_P(InstancingTestES3, LargestDivisor)
         << "Vertex attrib divisor read was not the same that was passed in.";
 }
 
-ANGLE_INSTANTIATE_TEST(InstancingTestES3, ES3_OPENGL(), ES3_OPENGLES(), ES3_D3D11());
+ANGLE_INSTANTIATE_TEST(InstancingTestES3, ES3_OPENGL(), ES3_OPENGLES(), ES3_D3D11(), ES3_VULKAN());
 
 ANGLE_INSTANTIATE_TEST(InstancingTestES31, ES31_OPENGL(), ES31_OPENGLES(), ES31_D3D11());