remove gpu pipelining

bytedance · Jun 23, 2024 · fdcbf80 · fdcbf80
1 parent 0c7ed3e
commit fdcbf80
Show file tree

Hide file tree

Showing 6 changed files with 1 addition and 252 deletions.
diff --git a/compiler/include/byteir/Dialect/GPU/Passes.h b/compiler/include/byteir/Dialect/GPU/Passes.h
@@ -23,7 +23,6 @@
 #include "byteir/Dialect/GPU/Transforms/GPUDistributeSharedMemoryCopy.h"
 #include "byteir/Dialect/GPU/Transforms/GPUDistributeToWarp.h"
 #include "byteir/Dialect/GPU/Transforms/GPUPackSharedMemoryAlloc.h"
-#include "byteir/Dialect/GPU/Transforms/GPUPipelining.h"
 #include "byteir/Dialect/GPU/Transforms/GPUTensorCoreVectorization.h"
 #include "byteir/Dialect/GPU/Transforms/GPUVectorToGPU.h"
 #include "byteir/Dialect/GPU/Transforms/OptimizeVectorTransfer.h"

diff --git a/compiler/include/byteir/Dialect/GPU/Passes.td b/compiler/include/byteir/Dialect/GPU/Passes.td
@@ -103,20 +103,6 @@ def GPUPackSharedMemoryAlloc : Pass<"gpu-pack-shared-memory-alloc", "func::FuncO
   let constructor = "mlir::createGPUPackSharedMemoryAllocPass()";
 }
 
-//===----------------------------------------------------------------------===//
-// GPUPipelining
-//===----------------------------------------------------------------------===//
-def GPUPipelining : Pass<"gpu-pipelining", "func::FuncOp"> {
-  let summary = "Pipelining async copy and mma oprations to improve performance.";
-  let constructor = "mlir::createGPUPipeliningPass()";
-  let options = [
-    Option<"stages", "stages", "int64_t", /*default=*/"0", "the number of stages for pipelining">,
-  ];
-  let dependentDialects = [
-    "NVVM::NVVMDialect",
-  ];
-}
-
 //===----------------------------------------------------------------------===//
 // GPUVectorToGPU
 //===----------------------------------------------------------------------===//

diff --git a/compiler/include/byteir/Dialect/GPU/Transforms/GPUPipelining.h b/compiler/include/byteir/Dialect/GPU/Transforms/GPUPipelining.h
diff --git a/compiler/lib/Dialect/GPU/Transforms/CMakeLists.txt b/compiler/lib/Dialect/GPU/Transforms/CMakeLists.txt
@@ -5,7 +5,6 @@ add_mlir_dialect_library(ByteIRGPUPasses
   GPUDistributeToWarp.cpp
   GPUTensorCoreVectorization.cpp
   GPUPackSharedMemoryAlloc.cpp
-  GPUPipelining.cpp
   GPUVectorToGPU.cpp
   OptimizeVectorTransfer.cpp
   RemoveTrivialLoops.cpp

diff --git a/compiler/lib/Dialect/GPU/Transforms/GPUPipelining.cpp b/compiler/lib/Dialect/GPU/Transforms/GPUPipelining.cpp
diff --git a/compiler/lib/Pipelines/LinalgMemrefOpt.cpp b/compiler/lib/Pipelines/LinalgMemrefOpt.cpp
@@ -56,9 +56,7 @@ void addGemmOptPasses(OpPassManager &pm) {
       anchoredPM.addPass(createCanonicalizerPass());
       anchoredPM.addPass(createCSEPass());
       anchoredPM.addPass(createCanonicalizerPass());
-      // anchoredPM.addPass(createGPUPipeliningPass());
-      // anchoredPM.addPass(createCSEPass());
-      // anchoredPM.addPass(createCanonicalizerPass());
+
       anchoredPM.addPass(createGPUDistributeToWarpPass());
       anchoredPM.addPass(createRemoveTrivialLoopsPass());
       anchoredPM.addPass(createGPUTensorCoreVectorizationPass());
@@ -87,7 +85,6 @@ void addGemmOptPasses(OpPassManager &pm) {
       pm.addPass(memref::createFoldMemRefAliasOpsPass());
     }
 
-    // anchoredPM.addPass(createGPUPipeliningPass());
     {
       OpPassManager anchoredPM(func::FuncOp::getOperationName());
       anchoredPM.addPass(createGPUPackSharedMemoryAllocPass());