Support VK_KHR_dynamic_rendering_local_read

Fixes #3341

qrenderdoc/Windows/PipelineState/VulkanPipelineStateViewer.cpp

@@ -41,6 +41,8 @@ Q_DECLARE_METATYPE(CombinedSamplerData);
 
 namespace
 {
+constexpr uint32_t VK_ATTACHMENT_UNUSED = ~0U;
+
 QString getTextureRenderSamples(const TextureDescription *tex, const VKPipe::RenderPass &renderpass)
 {
   const uint32_t texSamples = tex ? tex->msSamp : 1;
@@ -2614,12 +2616,32 @@ void VulkanPipelineStateViewer::setState()
         {
           slotname = QFormatStr("Color %1").arg(a.localIdx);
 
+          // With dynamic rendering, the API references the framebuffer index everywhere, for
+          // example when specifying blend state for attachments or with vkCmdClearAttachments. As
+          // such, RenderDoc shows the same index in Color attachments (i.e. fbIdx == localIdx) to
+          // avoid confusion, even when VK_KHR_dynamic_rendering_local_read maps these attachments
+          // to different "locations" used by the shader.  In that case, the mapped location is
+          // shown besides the attachment index.
+          uint32_t location = a.localIdx;
+          if(a.fbIdx < rp.colorAttachmentLocations.count())
+          {
+            location = rp.colorAttachmentLocations[a.fbIdx];
+            if(location == VK_ATTACHMENT_UNUSED)
+            {
+              slotname += QFormatStr(" [disabled]");
+            }
+            else
+            {
+              slotname += QFormatStr(" [location %1]").arg(location);
+            }
+          }
+
           if(state.fragmentShader.reflection != NULL)
           {
             const rdcarray<SigParameter> &outSig = state.fragmentShader.reflection->outputSignature;
             for(int s = 0; s < outSig.count(); s++)
             {
-              if(outSig[s].regIndex == (uint32_t)a.localIdx &&
+              if(outSig[s].regIndex == location &&
                  (outSig[s].systemValue == ShaderBuiltin::Undefined ||
                   outSig[s].systemValue == ShaderBuiltin::ColorOutput))
               {
@@ -2819,6 +2841,22 @@ void VulkanPipelineStateViewer::setState()
 
       if(showNode(usedSlot, /*filledSlot*/ true))
       {
+        QString writemask = QFormatStr("%1%2%3%4")
+                                .arg((blend.writeMask & 0x1) == 0 ? lit("_") : lit("R"))
+                                .arg((blend.writeMask & 0x2) == 0 ? lit("_") : lit("G"))
+                                .arg((blend.writeMask & 0x4) == 0 ? lit("_") : lit("B"))
+                                .arg((blend.writeMask & 0x8) == 0 ? lit("_") : lit("A"));
+
+        // With VK_KHR_dynamic_rendering_local_read, if a color attachment is mapped to
+        // VK_ATTACHMENT_UNUSED, it is implicitly disabled.  The Slot name in the "Render Pass"
+        // pane already tags the attachment with [disabled], so the write mask is simply set to
+        // ____ here.
+        if(i < rp.colorAttachmentLocations.count() &&
+           rp.colorAttachmentLocations[i] == VK_ATTACHMENT_UNUSED)
+        {
+          writemask = lit("____");
+        }
+
         RDTreeWidgetItem *node = new RDTreeWidgetItem(
             {i, blend.enabled ? tr("True") : tr("False"),
 
@@ -2828,11 +2866,7 @@ void VulkanPipelineStateViewer::setState()
              ToQStr(blend.alphaBlend.source), ToQStr(blend.alphaBlend.destination),
              ToQStr(blend.alphaBlend.operation),
 
-             QFormatStr("%1%2%3%4")
-                 .arg((blend.writeMask & 0x1) == 0 ? lit("_") : lit("R"))
-                 .arg((blend.writeMask & 0x2) == 0 ? lit("_") : lit("G"))
-                 .arg((blend.writeMask & 0x4) == 0 ? lit("_") : lit("B"))
-                 .arg((blend.writeMask & 0x8) == 0 ? lit("_") : lit("A"))});
+             writemask});
 
         if(!usedSlot)
           setInactiveRow(node);
@@ -4188,6 +4222,40 @@ void VulkanPipelineStateViewer::exportHTML(QXmlStreamWriter &xml, const VKPipe::
       xml.writeEndElement();
     }
 
+    if(!pass.renderpass.colorAttachmentLocations.isEmpty())
+    {
+      QList<QVariantList> locations;
+
+      for(int i = 0; i < pass.renderpass.colorAttachmentLocations.count(); i++)
+        locations.push_back({pass.renderpass.colorAttachmentLocations[i]});
+
+      m_Common.exportHTMLTable(xml,
+                               {
+                                   tr("Color Attachment Location"),
+                               },
+                               locations);
+
+      xml.writeStartElement(lit("p"));
+      xml.writeEndElement();
+    }
+
+    if(!pass.renderpass.colorAttachmentInputIndices.isEmpty())
+    {
+      QList<QVariantList> inputIndices;
+
+      for(int i = 0; i < pass.renderpass.colorAttachmentInputIndices.count(); i++)
+        inputIndices.push_back({pass.renderpass.colorAttachmentInputIndices[i]});
+
+      m_Common.exportHTMLTable(xml,
+                               {
+                                   tr("Color Attachment Input Index"),
+                               },
+                               inputIndices);
+
+      xml.writeStartElement(lit("p"));
+      xml.writeEndElement();
+    }
+
     if(!pass.renderpass.resolveAttachments.isEmpty())
     {
       QList<QVariantList> resolves;
@@ -4213,6 +4281,22 @@ void VulkanPipelineStateViewer::exportHTML(QXmlStreamWriter &xml, const VKPipe::
       xml.writeEndElement();
     }
 
+    if(!pass.renderpass.isDepthInputAttachmentIndexImplicit)
+    {
+      xml.writeStartElement(lit("p"));
+      xml.writeCharacters(
+          tr("Depth Input Attachment Index: %1").arg(pass.renderpass.depthInputAttachmentIndex));
+      xml.writeEndElement();
+    }
+
+    if(!pass.renderpass.isStencilInputAttachmentIndexImplicit)
+    {
+      xml.writeStartElement(lit("p"));
+      xml.writeCharacters(
+          tr("Stencil Input Attachment Index: %1").arg(pass.renderpass.stencilInputAttachmentIndex));
+      xml.writeEndElement();
+    }
+
     if(pass.renderpass.depthstencilResolveAttachment >= 0)
     {
       xml.writeStartElement(lit("p"));

renderdoc/api/replay/vk_pipestate.h

@@ -859,6 +859,30 @@ samples used to render this subpass.
 If the subpass is not internally multisampled, tileOnlyMSAASampleCount is set to 0.
 )");
   uint32_t tileOnlyMSAASampleCount = 0;
+
+  DOCUMENT(R"(The color index->location mapping set up by dynamic rendering local read.
+
+:type: List[int]
+)");
+  rdcarray<uint32_t> colorAttachmentLocations;
+
+  DOCUMENT(R"(The color index->input index mapping set up by dynamic rendering local read.
+
+:type: List[int]
+)");
+  rdcarray<uint32_t> colorAttachmentInputIndices;
+
+  DOCUMENT("Whether or not depth input attachment index is implicit (dynamic rendering).");
+  bool isDepthInputAttachmentIndexImplicit = true;
+
+  DOCUMENT("Whether or not stencil  input attachment index is implicit (dynamic rendering).");
+  bool isStencilInputAttachmentIndexImplicit = true;
+
+  DOCUMENT("Depth input attachment index if explicit (dynamic rendering).");
+  uint32_t depthInputAttachmentIndex = 0x1234ABCD;
+
+  DOCUMENT("Stencil input attachment index if explicit (dynamic rendering).");
+  uint32_t stencilInputAttachmentIndex = 0x1234ABCD;
 };
 
 DOCUMENT("Describes a framebuffer object and its attachments.");

renderdoc/driver/vulkan/extension_support.md

@@ -144,6 +144,7 @@ Maintainers can update this file by updating vk.xml in this folder and running `
 * `VK_KHR_draw_indirect_count`
 * `VK_KHR_driver_properties`
 * `VK_KHR_dynamic_rendering`
+* `VK_KHR_dynamic_rendering_local_read`
 * `VK_KHR_external_fence_capabilities`
 * `VK_KHR_external_fence_fd`
 * `VK_KHR_external_fence_win32`
@@ -241,9 +242,9 @@ KHR extensions will definitely be implemented at some point, though KHR extensio
 ## KHR Extensions
 
 * `VK_KHR_cooperative_matrix`
-* `VK_KHR_dynamic_rendering_local_read`
 * `VK_KHR_maintenance5`
 * `VK_KHR_maintenance6`
+* `VK_KHR_maintenance7`
 * `VK_KHR_map_memory2`
 * `VK_KHR_shader_expect_assume`
 * `VK_KHR_shader_float_controls2`

renderdoc/driver/vulkan/vk_common.h

@@ -1053,6 +1053,8 @@ enum class VulkanChunk : uint32_t
   vkCmdSetVertexInputEXT,
   vkCmdBeginRendering,
   vkCmdEndRendering,
+  vkCmdSetRenderingAttachmentLocationsKHR,
+  vkCmdSetRenderingInputAttachmentIndicesKHR,
   vkCmdSetFragmentShadingRateKHR,
   vkSetDeviceMemoryPriorityEXT,
   vkCmdSetAttachmentFeedbackLoopEnableEXT,
@@ -1335,6 +1337,7 @@ DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceDescriptorIndexingProperties)
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceDiscardRectanglePropertiesEXT);
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceDriverProperties);
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceDynamicRenderingFeatures);
+DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceDynamicRenderingLocalReadFeaturesKHR);
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceExtendedDynamicState2FeaturesEXT);
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceExtendedDynamicState3FeaturesEXT);
 DECLARE_REFLECTION_STRUCT(VkPhysicalDeviceExtendedDynamicState3PropertiesEXT);
@@ -1519,8 +1522,10 @@ DECLARE_REFLECTION_STRUCT(VkRayTracingShaderGroupCreateInfoKHR);
 DECLARE_REFLECTION_STRUCT(VkRefreshCycleDurationGOOGLE);
 DECLARE_REFLECTION_STRUCT(VkReleaseSwapchainImagesInfoEXT);
 DECLARE_REFLECTION_STRUCT(VkRenderingAttachmentInfo);
+DECLARE_REFLECTION_STRUCT(VkRenderingAttachmentLocationInfoKHR);
 DECLARE_REFLECTION_STRUCT(VkRenderingFragmentDensityMapAttachmentInfoEXT);
 DECLARE_REFLECTION_STRUCT(VkRenderingFragmentShadingRateAttachmentInfoKHR);
+DECLARE_REFLECTION_STRUCT(VkRenderingInputAttachmentIndexInfoKHR);
 DECLARE_REFLECTION_STRUCT(VkRenderingInfo);
 DECLARE_REFLECTION_STRUCT(VkRenderPassAttachmentBeginInfo);
 DECLARE_REFLECTION_STRUCT(VkRenderPassBeginInfo);
@@ -1775,6 +1780,7 @@ DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceDescriptorIndexingProperties)
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceDiscardRectanglePropertiesEXT);
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceDriverProperties);
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceDynamicRenderingFeatures);
+DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceDynamicRenderingLocalReadFeaturesKHR);
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceExtendedDynamicState2FeaturesEXT);
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceExtendedDynamicState3FeaturesEXT);
 DECLARE_DESERIALISE_TYPE(VkPhysicalDeviceExtendedDynamicState3PropertiesEXT);
@@ -1957,8 +1963,10 @@ DECLARE_DESERIALISE_TYPE(VkRayTracingPipelineInterfaceCreateInfoKHR);
 DECLARE_DESERIALISE_TYPE(VkRayTracingShaderGroupCreateInfoKHR);
 DECLARE_DESERIALISE_TYPE(VkReleaseSwapchainImagesInfoEXT);
 DECLARE_DESERIALISE_TYPE(VkRenderingAttachmentInfo);
+DECLARE_DESERIALISE_TYPE(VkRenderingAttachmentLocationInfoKHR);
 DECLARE_DESERIALISE_TYPE(VkRenderingFragmentDensityMapAttachmentInfoEXT);
 DECLARE_DESERIALISE_TYPE(VkRenderingFragmentShadingRateAttachmentInfoKHR);
+DECLARE_DESERIALISE_TYPE(VkRenderingInputAttachmentIndexInfoKHR);
 DECLARE_DESERIALISE_TYPE(VkRenderingInfo);
 DECLARE_DESERIALISE_TYPE(VkRenderPassAttachmentBeginInfo);
 DECLARE_DESERIALISE_TYPE(VkRenderPassBeginInfo);

renderdoc/driver/vulkan/vk_core.cpp

@@ -1490,6 +1490,10 @@ static const VkExtensionProperties supportedExtensions[] = {
         VK_KHR_DYNAMIC_RENDERING_EXTENSION_NAME,
         VK_KHR_DYNAMIC_RENDERING_SPEC_VERSION,
     },
+    {
+        VK_KHR_DYNAMIC_RENDERING_LOCAL_READ_EXTENSION_NAME,
+        VK_KHR_DYNAMIC_RENDERING_LOCAL_READ_SPEC_VERSION,
+    },
     {
         VK_KHR_EXTERNAL_FENCE_EXTENSION_NAME,
         VK_KHR_EXTERNAL_FENCE_SPEC_VERSION,
@@ -4230,6 +4234,10 @@ bool WrappedVulkan::ProcessChunk(ReadSerialiser &ser, VulkanChunk chunk)
     case VulkanChunk::vkCmdBeginRendering:
       return Serialise_vkCmdBeginRendering(ser, VK_NULL_HANDLE, NULL);
     case VulkanChunk::vkCmdEndRendering: return Serialise_vkCmdEndRendering(ser, VK_NULL_HANDLE);
+    case VulkanChunk::vkCmdSetRenderingAttachmentLocationsKHR:
+      return Serialise_vkCmdSetRenderingAttachmentLocationsKHR(ser, VK_NULL_HANDLE, NULL);
+    case VulkanChunk::vkCmdSetRenderingInputAttachmentIndicesKHR:
+      return Serialise_vkCmdSetRenderingInputAttachmentIndicesKHR(ser, VK_NULL_HANDLE, NULL);
 
     case VulkanChunk::vkCmdSetFragmentShadingRateKHR:
       return Serialise_vkCmdSetFragmentShadingRateKHR(ser, VK_NULL_HANDLE, NULL, NULL);

renderdoc/driver/vulkan/vk_core.h

@@ -2787,6 +2787,15 @@ class WrappedVulkan : public IFrameCapturer
 
   IMPLEMENT_FUNCTION_SERIALISED(void, vkCmdEndRendering, VkCommandBuffer commandBuffer);
 
+  // VK_KHR_dynamic_rendering_local_read
+
+  IMPLEMENT_FUNCTION_SERIALISED(void, vkCmdSetRenderingAttachmentLocationsKHR,
+                                VkCommandBuffer commandBuffer,
+                                const VkRenderingAttachmentLocationInfoKHR *pLocationInfo);
+  IMPLEMENT_FUNCTION_SERIALISED(void, vkCmdSetRenderingInputAttachmentIndicesKHR,
+                                VkCommandBuffer commandBuffer,
+                                const VkRenderingInputAttachmentIndexInfoKHR *pLocationInfo);
+
   // VK_KHR_fragment_shading_rate
 
   IMPLEMENT_FUNCTION_SERIALISED(void, vkCmdSetFragmentShadingRateKHR, VkCommandBuffer commandBuffer,

renderdoc/driver/vulkan/vk_hookset_defs.h

@@ -522,6 +522,7 @@
   DeclExt(EXT_multisampled_render_to_single_sampled);  \
   DeclExt(EXT_vertex_input_dynamic_state);             \
   DeclExt(KHR_dynamic_rendering);                      \
+  DeclExt(KHR_dynamic_rendering_local_read);           \
   DeclExt(KHR_fragment_shading_rate);                  \
   DeclExt(EXT_attachment_feedback_loop_layout);        \
   DeclExt(EXT_pageable_device_local_memory);           \
@@ -651,6 +652,7 @@
   CheckExt(EXT_multisampled_render_to_single_sampled, VKXX);  \
   CheckExt(EXT_vertex_input_dynamic_state, VKXX);             \
   CheckExt(KHR_dynamic_rendering, VK13);                      \
+  CheckExt(KHR_dynamic_rendering_local_read, VKXX);           \
   CheckExt(KHR_fragment_shading_rate, VKXX);                  \
   CheckExt(EXT_attachment_feedback_loop_layout, VKXX);        \
   CheckExt(EXT_pageable_device_local_memory, VKXX);           \
@@ -941,6 +943,8 @@
   HookInitExtension(EXT_vertex_input_dynamic_state || EXT_shader_object, CmdSetVertexInputEXT);      \
   HookInitPromotedExtension(KHR_dynamic_rendering, CmdBeginRendering, KHR);                          \
   HookInitPromotedExtension(KHR_dynamic_rendering, CmdEndRendering, KHR);                            \
+  HookInitExtension(KHR_dynamic_rendering_local_read, CmdSetRenderingAttachmentLocationsKHR);        \
+  HookInitExtension(KHR_dynamic_rendering_local_read, CmdSetRenderingInputAttachmentIndicesKHR);     \
   HookInitExtension(KHR_fragment_shading_rate, CmdSetFragmentShadingRateKHR);                        \
   HookInitExtension(EXT_pageable_device_local_memory, SetDeviceMemoryPriorityEXT);                   \
   HookInitExtension(EXT_swapchain_maintenance1, ReleaseSwapchainImagesEXT);                          \
@@ -1743,6 +1747,10 @@
   HookDefine2(void, vkCmdBeginRendering, VkCommandBuffer, commandBuffer, const VkRenderingInfo *,    \
               pRenderingInfo);                                                                       \
   HookDefine1(void, vkCmdEndRendering, VkCommandBuffer, commandBuffer);                              \
+  HookDefine2(void, vkCmdSetRenderingAttachmentLocationsKHR, VkCommandBuffer, commandBuffer,         \
+              const VkRenderingAttachmentLocationInfoKHR *, pLocationInfo);                          \
+  HookDefine2(void, vkCmdSetRenderingInputAttachmentIndicesKHR, VkCommandBuffer, commandBuffer,      \
+              const VkRenderingInputAttachmentIndexInfoKHR *, pInputAttachmentIndexInfo);            \
   HookDefine3(void, vkCmdSetFragmentShadingRateKHR, VkCommandBuffer, commandBuffer,                  \
               const VkExtent2D *, pFragmentSize, const VkFragmentShadingRateCombinerOpKHR *,         \
               combinerOps);                                                                          \