Deprecate Vulkan support

Removes Vulkan support and related passes.
Lowers maintenance cost and allows future GPU pipeline cleanup.

This GPU execution path has showed limitation and has been largely
unused for a long time now. For Nvidia, native CUDA backend is more
complete. For Intel, various dedicated backends are  available and
will become the primary way to execute on Intel GPU.

.github/workflows/tpp-llvm.yml

@@ -30,14 +30,3 @@ jobs:
               ${{ env.SRUN }} --partition=a100,v100 --time=0:30:00 -- \
               'KIND=RelWithDebInfo COMPILER=clang GPU=cuda \
               ${{ github.workspace }}/scripts/buildkite/build_llvm.sh'
-
-  TPP-MLIR-LLVM-Vulkan:
-    runs-on: pcl-tiergarten
-    steps:
-      - uses: actions/checkout@v4
-      - name: LLVM Vulkan
-        run: |-
-              GPU=vulkan scripts/buildkite/check_llvm.sh || \
-              ${{ env.SRUN }} --partition=spr-all --time=0:30:00 -- \
-              'KIND=RelWithDebInfo COMPILER=clang GPU=vulkan \
-              ${{ github.workspace }}/scripts/buildkite/build_llvm.sh'

CMakeLists.txt

@@ -13,7 +13,7 @@ project(tpp-dialect LANGUAGES CXX C)
 set(CMAKE_CXX_STANDARD 17 CACHE STRING "C++ standard to conform to")
 
 set(TPP_GPU "" CACHE STRING "Enables GPU runtime (default: '')")
-set_property(CACHE TPP_GPU PROPERTY STRINGS "" "cuda" "vulkan")
+set_property(CACHE TPP_GPU PROPERTY STRINGS "" "cuda")
 
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
   message(STATUS "TPP-MLIR out-of-tree build.")

benchmarks/driver.py

@@ -565,7 +565,7 @@ def verifyStats(self):
         "--init-type", type=str, help="Random initializer type"
     )
     parser.add_argument(
-        "--gpu", type=str, help="Target GPU backend for lowering (cuda,vulkan)"
+        "--gpu", type=str, help="Target GPU backend for lowering (cuda)"
     )
     args = parser.parse_args()
 

benchmarks/harness/controller.py

@@ -263,7 +263,7 @@ def verifyStats(self):
         help="Random initializer type (default: normal)",
     )
     parser.add_argument(
-        "--gpu", type=str, help="Target GPU backend for lowering (cuda,vulkan)"
+        "--gpu", type=str, help="Target GPU backend for lowering (cuda)"
     )
     args = parser.parse_args()
 

include/TPP/PassBundles.td

@@ -171,13 +171,4 @@ def GpuToCuda : Pass<"gpu-to-cuda", "ModuleOp"> {
   ];
 }
 
-def GpuToVulkan : Pass<"gpu-to-vulkan", "ModuleOp"> {
-  let summary = "Lower generic GPU operations to Vulkan backend";
-  let dependentDialects = ["arith::ArithDialect",
-                           "memref::MemRefDialect",
-                           "func::FuncDialect",
-                           "gpu::GPUDialect",
-                           "spirv::SPIRVDialect"];
-}
-
 #endif // TPP_DIALECT_TPP_PASSBUNDLES

include/TPP/Passes.td

@@ -290,7 +290,7 @@ def SetSPIRVCapabilities : Pass<"tpp-set-spirv-capabilities", "ModuleOp"> {
   let summary = "Set SPIR-V capabilities.";
   let options = [
     Option<"clientAPI", "client-api", "std::string",
-            /*default=*/"\"vulkan\"",
+            /*default=*/"\"opencl\"",
            "The client API to use for capabilities">,
   ];
 }
@@ -299,30 +299,11 @@ def SetSPIRVAbiAttribute : Pass<"set-spirv-abi-attr", "gpu::GPUModuleOp"> {
   let summary = "Set SPIR-V ABI attribute.";
   let options = [
     Option<"clientAPI", "client-api", "std::string",
-            /*default=*/"\"vulkan\"",
+            /*default=*/"\"opencl\"",
            "The client API to use for ABI attribute">,
   ];
 }
 
-def GpuVulkanAbi : Pass<"gpu-vulkan-abi", "ModuleOp"> {
-  let summary = "Rewrite GPU kernels to comply with Vulkan ABI.";
-  let description = [{
-    This pass rewrites GPU kernels and the GPU function launches
-    to be compatible with Vulkan calling convention.
-
-    The rewrite is focused only on ensuring GPU kernel prototype
-    compatibility with Vulkan ABI.Thus, it is assumed that the kernel
-    operations have been already preserved in a separate SPIR-V module.
-    The original GPU kernel logic is thrown away to make interface
-    adaptation easier.
-  }];
-  let options = [
-    Option<"use64bitIndex", "use-64bit-index",
-           "bool", /*default=*/"false",
-           "Use 64-bit integers to convert index types">
-  ];
-}
-
 def DecomposeAggregatedOps : Pass<"decompose-aggregated-ops", "func::FuncOp"> {
   let summary = "Decompose aggregated operations.";
   let description = [{
@@ -467,7 +448,7 @@ def TppRunnerWrapper : Pass<"tpp-runner-wrapper", "ModuleOp">{
            "The type of the kernel function.">,
     Option<"backend", "backend", "std::string",
             /*default=*/"\"cpu\"",
-           "Kernel target device backend (cpu, cuda, vulkan).">,
+           "Kernel target device backend (cpu, cuda, intel).">,
     Option<"offloadToDevice", "offload-on-device", "bool",
             /*default=*/"true",
            "Offload kernel arguments to the target device.">,

lib/TPP/DefaultPipeline.cpp

@@ -181,8 +181,6 @@ struct DefaultPipeline : public tpp::impl::DefaultPipelineBase<DefaultPipeline>,
     pm.addNestedPass<func::FuncOp>(createCSEPass());
     pm.addPass(createReconcileUnrealizedCastsPass());
 
-    pm.addPass(createConvertVulkanLaunchFuncToVulkanCallsPass());
-
     // Anything useful has been lowered by now.
     // Cleanup IR by removing any dead symbols.
     // This step aims to avoid errors caused by frontend leftovers.

lib/TPP/GPU/CMakeLists.txt

@@ -3,10 +3,8 @@ add_mlir_library(TPPGPU
   GpuPipeline.cpp
   GpuConversion.cpp
   GpuToCuda.cpp
-  GpuToVulkan.cpp
   SetSPIRVCapabilities.cpp
   SetSPIRVAbiAttribute.cpp
-  GpuVulkanAbi.cpp
   LinalgToGpu.cpp
   GpuDataTransfer.cpp
   GpuInlineConstants.cpp
@@ -46,17 +44,11 @@ if (TPP_GPU MATCHES "cuda")
   )
 endif()
 
-if (TPP_GPU MATCHES "vulkan")
-  target_compile_definitions(obj.TPPGPU
-    PRIVATE
-    TPP_VULKAN_ENABLE=1
-  )
-
+if (DEFINED ${TPP_GPU})
   target_link_libraries(TPPGPU
     PUBLIC
     MLIRSPIRVDialect
     MLIRSPIRVTransforms
-    MLIRGPUToVulkanTransforms
     MLIRGPUToSPIRV
   )
 endif()

lib/TPP/GPU/GpuPipeline.cpp

@@ -84,14 +84,12 @@ namespace {
 
 enum class GpuType {
   Cuda,
-  Vulkan,
   Intel,
 };
 
 GpuType parseGpuOption(StringRef gpuStr) {
   auto type = llvm::StringSwitch<std::optional<GpuType>>(gpuStr)
                   .CaseLower("cuda", GpuType::Cuda)
-                  .CaseLower("vulkan", GpuType::Vulkan)
                   .CaseLower("intel", GpuType::Intel)
                   .Default(std::nullopt);
   assert(type && "Unsupported GPU backend");
@@ -115,7 +113,6 @@ GpuOptions getGpuOptions(GpuType gpuType) {
     options.features = "+ptx60";
     break;
   }
-  case GpuType::Vulkan:
   case GpuType::Intel: {
     // No options needed at the moment.
     break;
@@ -196,8 +193,7 @@ struct GpuPipeline : public tpp::impl::GpuPipelineBase<GpuPipeline>,
     // Preprocess and bufferize as further conversion requires memref
     // abstraction.
     pm.addPass(createLowerPacksAndUnPacks());
-    bool dealloc = gpuType == GpuType::Vulkan;
-    pm.addPass(createBufferize(BufferizeOptions{dealloc}));
+    pm.addPass(createBufferize(BufferizeOptions{/*dealloc=*/false}));
     pm.addPass(createConvertForAllToParallelOp());
     pm.addPass(createCleanup());
 
@@ -211,16 +207,11 @@ struct GpuPipeline : public tpp::impl::GpuPipelineBase<GpuPipeline>,
     case GpuType::Cuda: {
       // Perform explicit GPU data transfers only for CUDA as the unified
       // memory is not currently used here.
-      // Vulkan runner assumes usage of GPU unified memory.
       pm.addNestedPass<func::FuncOp>(createGpuDataTransfer());
       pm.addPass(createGpuToCuda(GpuToCudaOptions{
           gpuOptions.triple, gpuOptions.chip, gpuOptions.features}));
       break;
     }
-    case GpuType::Vulkan: {
-      pm.addPass(createGpuToVulkan());
-      break;
-    }
     case GpuType::Intel:
       pm.addPass(xegpu::createXeGPUFoldAliasOps());