[Metal] Use common resource creation, readbacks, and InvocationState

include/API/Buffer.h

@@ -70,14 +70,16 @@ struct BufferCreateDesc {
                             false};
   }
 
-  static BufferCreateDesc scratchBuffer() {
+  static BufferCreateDesc gpuOnlyStorage() {
     return BufferCreateDesc{MemoryLocation::GpuOnly,
                             MemoryBacking::Automatic,
                             BufferUsage::Storage,
                             BufferShaderAccessType::Raw,
                             {},
                             false};
   }
+
+  static BufferCreateDesc scratchBuffer() { return gpuOnlyStorage(); }
 };
 
 class Buffer {

include/API/Device.h

@@ -20,6 +20,7 @@
 #include "API/Capabilities.h"
 #include "API/CommandBuffer.h"
 #include "API/RenderPass.h"
+#include "API/Sampler.h"
 #include "API/ShaderBindingTable.h"
 #include "API/Texture.h"
 
@@ -346,6 +347,9 @@ class Device {
   virtual llvm::Expected<std::unique_ptr<Texture>>
   createTexture(std::string Name, const TextureCreateDesc &Desc) = 0;
 
+  virtual llvm::Expected<std::unique_ptr<Sampler>>
+  createSampler(std::string Name, const SamplerCreateDesc &Desc) = 0;
+
   virtual llvm::Expected<std::unique_ptr<MemoryHeap>>
   createMemoryHeap(std::string Name, size_t SizeInBytes) = 0;
 
@@ -354,6 +358,12 @@ class Device {
   virtual uint32_t
   getTextureUploadRowStrideInBytes(const TextureCreateDesc &Desc) const = 0;
 
+  // The layout an upload buffer must have to feed createTextureWithData /
+  // copyBufferToTexture for the given texture description. Encodes per-mip
+  // offsets, row pitch, and total size in the backend's required alignment.
+  virtual TextureUploadLayout
+  getTextureUploadLayout(const TextureCreateDesc &Desc) const = 0;
+
   virtual llvm::Expected<std::unique_ptr<RenderPass>>
   createRenderPass(const RenderPassDesc &Desc) = 0;
 
@@ -382,7 +392,8 @@ class Device {
   createBLAS(const AccelerationStructureSizes &Sizes) = 0;
 
   virtual llvm::Expected<std::unique_ptr<AccelerationStructure>>
-  createTLAS(const AccelerationStructureSizes &Sizes) = 0;
+  createTLAS(const AccelerationStructureSizes &Sizes,
+             uint32_t InstanceCount) = 0;
 
   virtual ~Device() = 0;
 
@@ -421,6 +432,9 @@ createBufferWithData(Device &Dev, std::string Name,
                      size_t SizeInBytes, ComputeEncoder *Encoder,
                      std::unique_ptr<offloadtest::Buffer> *OutUploadBuffer);
 
+// Create a texture and upload `Data` (tightly-packed across mip levels) into
+// it via a staging buffer recorded on `Encoder`. The staging buffer is handed
+// back through `OutUploadBuffer` and must outlive command-buffer submission.
 llvm::Expected<std::unique_ptr<offloadtest::Texture>>
 createTextureWithData(Device &Dev, std::string Name,
                       const TextureCreateDesc &Desc, const void *Data,

include/API/Sampler.h

@@ -0,0 +1,67 @@
+//===- Sampler.h - Offload API Texture ------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef OFFLOADTEST_API_SAMPLER_H
+#define OFFLOADTEST_API_SAMPLER_H
+
+#include "API/API.h"
+#include "API/Resources.h"
+
+namespace offloadtest {
+
+enum class FilterMode { Nearest, Linear };
+
+enum class AddressMode { Clamp, Repeat, Mirror, Border, MirrorOnce };
+
+enum class CompareFunction {
+  Never,
+  Less,
+  Equal,
+  LessEqual,
+  Greater,
+  NotEqual,
+  GreaterEqual,
+  Always
+};
+
+enum class SamplerKind { Sampler, SamplerComparison };
+
+struct SamplerCreateDesc {
+  FilterMode MinFilter = FilterMode::Linear;
+  FilterMode MagFilter = FilterMode::Linear;
+  AddressMode Address = AddressMode::Clamp;
+  float MinLOD = 0.0f;
+  float MaxLOD = std::numeric_limits<float>::max();
+  float MipLODBias = 0.0f;
+  CompareFunction ComparisonOp = CompareFunction::Never;
+  SamplerKind Kind = SamplerKind::Sampler;
+};
+
+class Sampler {
+  GPUAPI API;
+
+public:
+  virtual ~Sampler();
+  Sampler(const Sampler &) = delete;
+  // Sampler(Sampler &&) = delete;
+  Sampler &operator=(const Sampler &) = delete;
+  // Sampler &operator=(Sampler &&) = delete;
+
+  GPUAPI getAPI() const { return API; }
+  virtual const SamplerCreateDesc &getDesc() const = 0;
+
+protected:
+  explicit Sampler(GPUAPI API) : API(API) {}
+};
+
+} // namespace offloadtest
+
+#endif // OFFLOADTEST_API_SAMPLER_H

include/API/Texture.h

@@ -16,6 +16,7 @@
 #include "API/Resources.h"
 
 #include "llvm/ADT/BitmaskEnum.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/Error.h"
@@ -105,12 +106,6 @@ inline llvm::Error validateTextureCreateDesc(const TextureCreateDesc &Desc) {
         std::errc::not_supported,
         "DepthStencil combined with Storage is not yet supported.");
 
-  // Depth formats require DepthStencil usage; non-depth formats forbid it.
-  if (IsDepth && !IsDS)
-    return llvm::createStringError(
-        std::errc::invalid_argument,
-        "Depth format '%s' requires DepthStencil usage.",
-        getFormatName(Desc.Fmt).data());
   if (!IsDepth && IsDS)
     return llvm::createStringError(
         std::errc::invalid_argument,
@@ -154,6 +149,24 @@ struct TileShape {
   uint32_t Depth = 1;
 };
 
+struct SubresourceFootprint {
+  uint64_t Offset = 0; // Byte offset of this subresource in the buffer.
+  uint32_t RowPitchInBytes = 0; // Destination row stride (may include padding).
+  uint32_t RowSizeInBytes = 0;  // Tightly-packed bytes per row to copy.
+  uint32_t NumRows = 0;         // Number of rows in this subresource.
+};
+
+struct TextureUploadLayout {
+  llvm::SmallVector<SubresourceFootprint> Subresources; // One entry per mip.
+  uint64_t TotalSizeInBytes = 0;
+};
+
+// Compute a tightly-packed upload layout (no row or subresource padding) for
+// the given texture description. Suitable for backends whose buffer-to-texture
+// copy consumes a tightly-packed staging buffer (e.g. Vulkan, Metal).
+TextureUploadLayout
+computeTightTextureUploadLayout(const TextureCreateDesc &Desc);
+
 class Texture {
   GPUAPI API;
 

include/Support/Pipeline.h

@@ -16,6 +16,7 @@
 #include "API/AccelerationStructure.h"
 #include "API/Enums.h"
 #include "API/Resources.h"
+#include "API/Sampler.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/Support/Error.h"
@@ -140,24 +141,7 @@ static inline DescriptorKind getDescriptorKind(ResourceKind RK) {
   llvm_unreachable("All cases handled");
 }
 
-enum class FilterMode { Nearest, Linear };
-
-enum class AddressMode { Clamp, Repeat, Mirror, Border, MirrorOnce };
-
-enum class CompareFunction {
-  Never,
-  Less,
-  Equal,
-  LessEqual,
-  Greater,
-  NotEqual,
-  GreaterEqual,
-  Always
-};
-
-enum class SamplerKind { Sampler, SamplerComparison };
-
-struct Sampler {
+struct YAMLSampler {
   std::string Name;
   FilterMode MinFilter = FilterMode::Linear;
   FilterMode MagFilter = FilterMode::Linear;
@@ -270,7 +254,7 @@ struct Resource {
   DirectXBinding DXBinding;
   std::optional<VulkanBinding> VKBinding;
   CPUBuffer *BufferPtr = nullptr;
-  Sampler *SamplerPtr = nullptr;
+  YAMLSampler *SamplerPtr = nullptr;
   bool HasCounter = false;
   std::optional<uint32_t> TilesMapped;
   bool IsReserved = false;
@@ -630,7 +614,7 @@ struct Pipeline {
   IOBindings Bindings;
   llvm::SmallVector<PushConstantBlock> PushConstants;
   llvm::SmallVector<CPUBuffer> Buffers;
-  llvm::SmallVector<Sampler> Samplers;
+  llvm::SmallVector<YAMLSampler> Samplers;
   llvm::SmallVector<Result> Results;
   llvm::SmallVector<DescriptorSet> Sets;
   DispatchParametersSet DispatchParameters;
@@ -671,7 +655,7 @@ struct Pipeline {
     return nullptr;
   }
 
-  Sampler *getSampler(llvm::StringRef Name) {
+  YAMLSampler *getSampler(llvm::StringRef Name) {
     for (auto &S : Samplers)
       if (Name == S.Name)
         return &S;
@@ -712,7 +696,7 @@ struct Pipeline {
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::DescriptorSet)
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::Resource)
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::CPUBuffer)
-LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::Sampler)
+LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::YAMLSampler)
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::Shader)
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::dx::RootParameter)
 LLVM_YAML_IS_SEQUENCE_VECTOR(offloadtest::Result)
@@ -744,8 +728,8 @@ template <> struct MappingTraits<offloadtest::CPUBuffer> {
   static void mapping(IO &I, offloadtest::CPUBuffer &R);
 };
 
-template <> struct MappingTraits<offloadtest::Sampler> {
-  static void mapping(IO &I, offloadtest::Sampler &S);
+template <> struct MappingTraits<offloadtest::YAMLSampler> {
+  static void mapping(IO &I, offloadtest::YAMLSampler &S);
 };
 
 template <> struct MappingTraits<offloadtest::Result> {