[DX] View instancing follow-up: 4-view test (#1051)

include/API/Device.h

@@ -98,6 +98,10 @@ struct TraditionalRasterPipelineCreateDesc {
   std::optional<ShaderContainer> DS;
   std::optional<ShaderContainer> GS;
   ShaderContainer PS;
+  // View-instancing fan-out (1 == no view instancing). When > 1, backends
+  // that implement view instancing should route view N to render-target
+  // array slice N.
+  uint32_t ViewInstanceCount = 1;
 
   void setShader(Stages Stage, ShaderContainer &&SC) {
     switch (Stage) {
@@ -253,7 +257,8 @@ createRenderTargetFromCPUBuffer(Device &Dev, const CPUBuffer &Buf);
 
 // Creates a depth/stencil texture matching the dimensions of a render target.
 llvm::Expected<std::unique_ptr<Texture>>
-createDefaultDepthStencilTarget(Device &Dev, uint32_t Width, uint32_t Height);
+createDefaultDepthStencilTarget(Device &Dev, uint32_t Width, uint32_t Height,
+                                uint32_t ArraySize = 1);
 
 llvm::Expected<std::unique_ptr<offloadtest::Buffer>>
 createBufferWithData(Device &Dev, std::string Name,

include/API/Enums.h

@@ -16,6 +16,7 @@ enum class ResourceKind {
   StructuredBuffer,
   ByteAddressBuffer,
   Texture2D,
+  Texture2DArray,
   RWBuffer,
   RWStructuredBuffer,
   RWByteAddressBuffer,

include/API/FormatConversion.h

@@ -21,6 +21,8 @@
 
 #include "llvm/Support/Error.h"
 
+#include <algorithm>
+
 namespace offloadtest {
 
 // Bridge for code that still describes textures as DataFormat + Channels (e.g.
@@ -139,20 +141,28 @@ validateTextureDescMatchesCPUBuffer(const TextureCreateDesc &Desc,
         "TextureCreateDesc mip levels %u does not match CPUBuffer mip "
         "levels %d.",
         Desc.MipLevels, Buf.OutputProps.MipLevels);
+  if (Desc.ArraySize !=
+      static_cast<uint32_t>(std::max(1, Buf.OutputProps.ArraySize)))
+    return llvm::createStringError(
+        std::errc::invalid_argument,
+        "TextureCreateDesc array size %u does not match CPUBuffer array "
+        "size %d.",
+        Desc.ArraySize, Buf.OutputProps.ArraySize);
   const uint32_t TexelSize = getFormatSizeInBytes(Desc.Fmt);
   if (Buf.Stride > 0 && static_cast<uint32_t>(Buf.Stride) != TexelSize)
     return llvm::createStringError(
         std::errc::invalid_argument,
         "CPUBuffer stride %d does not match texture format element size %u.",
         Buf.Stride, TexelSize);
-  const uint64_t ExpectedSize =
-      static_cast<uint64_t>(Desc.Width) * Desc.Height * TexelSize;
+  const uint64_t ExpectedSize = static_cast<uint64_t>(Desc.Width) *
+                                Desc.Height * Desc.ArraySize * TexelSize;
   if (static_cast<uint64_t>(Buf.size()) != ExpectedSize)
     return llvm::createStringError(
         std::errc::invalid_argument,
         "CPUBuffer size %u does not match expected size %llu "
-        "(width %u * height %u * element size %u).",
-        Buf.size(), ExpectedSize, Desc.Width, Desc.Height, TexelSize);
+        "(width %u * height %u * array size %u * element size %u).",
+        Buf.size(), ExpectedSize, Desc.Width, Desc.Height, Desc.ArraySize,
+        TexelSize);
   return llvm::Error::success();
 }
 

include/API/Texture.h

@@ -71,6 +71,10 @@ struct TextureCreateDesc {
   uint32_t Width;
   uint32_t Height;
   uint32_t MipLevels;
+  // Texture array slice count (1 == plain Texture2D). Render targets used
+  // with view instancing require this to match the pipeline's
+  // ViewInstanceCount.
+  uint32_t ArraySize = 1;
   // Clear value for render target or depth/stencil textures.
   // How and when this is applied depends on the backend:
   // - DX uses it as an optimized clear hint at resource creation time

include/Support/Pipeline.h

@@ -81,6 +81,7 @@ static inline DescriptorKind getDescriptorKind(ResourceKind RK) {
   case ResourceKind::StructuredBuffer:
   case ResourceKind::ByteAddressBuffer:
   case ResourceKind::Texture2D:
+  case ResourceKind::Texture2DArray:
   case ResourceKind::AccelerationStructure:
     return DescriptorKind::SRV;
 
@@ -145,6 +146,7 @@ struct OutputProperties {
   int Width;
   int Depth;
   int MipLevels = 1;
+  int ArraySize = 1;
 };
 
 static inline uint32_t getFormatSize(DataFormat Format) {
@@ -246,6 +248,7 @@ struct Resource {
     case ResourceKind::Buffer:
     case ResourceKind::RWBuffer:
     case ResourceKind::Texture2D:
+    case ResourceKind::Texture2DArray:
     case ResourceKind::RWTexture2D:
     case ResourceKind::Sampler:
     case ResourceKind::SampledTexture2D:
@@ -273,6 +276,7 @@ struct Resource {
     case ResourceKind::RWByteAddressBuffer:
     case ResourceKind::ConstantBuffer:
     case ResourceKind::Texture2D:
+    case ResourceKind::Texture2DArray:
     case ResourceKind::RWTexture2D:
     case ResourceKind::SampledTexture2D:
     case ResourceKind::AccelerationStructure:
@@ -294,6 +298,7 @@ struct Resource {
     case ResourceKind::AccelerationStructure:
       return false;
     case ResourceKind::Texture2D:
+    case ResourceKind::Texture2DArray:
     case ResourceKind::RWTexture2D:
     case ResourceKind::SampledTexture2D:
       return true;
@@ -356,6 +361,7 @@ struct Resource {
     case ResourceKind::StructuredBuffer:
     case ResourceKind::ByteAddressBuffer:
     case ResourceKind::Texture2D:
+    case ResourceKind::Texture2DArray:
     case ResourceKind::ConstantBuffer:
     case ResourceKind::Sampler:
     case ResourceKind::SampledTexture2D:
@@ -541,6 +547,12 @@ struct Pipeline {
   DispatchParametersSet DispatchParameters;
   AccelerationStructureDescs AccelStructs;
 
+  // Number of view instances to render in a single draw (D3D12 view
+  // instancing). Default 1 = no view instancing. When > 1 the render target
+  // must be a Texture2DArray whose ArraySize is at least this value, and the
+  // backend routes view N to array slice N of the render/depth target.
+  uint32_t ViewInstanceCount = 1;
+
   uint32_t getVertexCount() const {
     if (DispatchParameters.VertexCount)
       return *DispatchParameters.VertexCount;
@@ -829,6 +841,7 @@ template <> struct ScalarEnumerationTraits<offloadtest::ResourceKind> {
     ENUM_CASE(StructuredBuffer);
     ENUM_CASE(ByteAddressBuffer);
     ENUM_CASE(Texture2D);
+    ENUM_CASE(Texture2DArray);
     ENUM_CASE(RWBuffer);
     ENUM_CASE(RWStructuredBuffer);
     ENUM_CASE(RWByteAddressBuffer);