[Codegen] Query #iree_gpu.target for shared memory limit (iree-org#…

…18184) Signed-off-by: nithinsubbiah <[email protected]>
nod-ai · Aug 10, 2024 · b06bf6a · b06bf6a
1 parent e36aa78
commit b06bf6a
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 26 deletions.
diff --git a/compiler/src/iree/compiler/Codegen/Common/GPU/GPUCheckResourceUsage.cpp b/compiler/src/iree/compiler/Codegen/Common/GPU/GPUCheckResourceUsage.cpp
@@ -50,7 +50,7 @@ static int shapedTypeStaticSize(
 }
 
 /// Returns success if the total shared memory allocation size is less than the
-/// limit set by limit.
+/// limit.
 static LogicalResult checkGPUAllocationSize(
     mlir::FunctionOpInterface funcOp, unsigned limit,
     std::function<unsigned(mlir::FunctionOpInterface)> getIndexBitwidth) {
@@ -93,15 +93,14 @@ class GPUCheckResourceUsagePass final
     : public impl::GPUCheckResourceUsagePassBase<GPUCheckResourceUsagePass> {
 public:
   explicit GPUCheckResourceUsagePass(
-      std::function<unsigned(mlir::FunctionOpInterface)> getSharedMemoryLimit,
       std::function<unsigned(mlir::FunctionOpInterface)> getIndexBitwidth)
-      : getSharedMemoryLimit(getSharedMemoryLimit),
-        getIndexBitwidth(getIndexBitwidth) {}
+      : getIndexBitwidth(getIndexBitwidth) {}
 
   void runOnOperation() override {
     FunctionOpInterface funcOp = getOperation();
+    IREE::GPU::TargetAttr target = getGPUTargetAttr(funcOp);
     unsigned limit =
-        getSharedMemoryLimit ? getSharedMemoryLimit(funcOp) : 64 * 1024;
+        target ? target.getWgp().getMaxWorkgroupMemoryBytes() : 64 * 1024;
     if (failed(checkGPUAllocationSize(funcOp, limit,
                                       getIndexBitwidth
                                           ? getIndexBitwidth
@@ -111,18 +110,15 @@ class GPUCheckResourceUsagePass final
   }
 
 private:
-  std::function<unsigned(mlir::FunctionOpInterface)> getSharedMemoryLimit;
   std::function<unsigned(mlir::FunctionOpInterface)> getIndexBitwidth;
 };
 
 } // namespace
 
 std::unique_ptr<InterfacePass<FunctionOpInterface>>
 createGPUCheckResourceUsagePass(
-    std::function<unsigned(mlir::FunctionOpInterface)> getSharedMemoryLimit,
     std::function<unsigned(mlir::FunctionOpInterface)> getIndexBitwidth) {
-  return std::make_unique<GPUCheckResourceUsagePass>(getSharedMemoryLimit,
-                                                     getIndexBitwidth);
+  return std::make_unique<GPUCheckResourceUsagePass>(getIndexBitwidth);
 }
 
 } // namespace mlir::iree_compiler
diff --git a/compiler/src/iree/compiler/Codegen/Common/GPU/Passes.h b/compiler/src/iree/compiler/Codegen/Common/GPU/Passes.h
@@ -87,8 +87,6 @@ LogicalResult gpuDistributeSharedMemoryCopy(mlir::FunctionOpInterface funcOp);
 // get the index size.
 std::unique_ptr<InterfacePass<FunctionOpInterface>>
 createGPUCheckResourceUsagePass(
-    std::function<unsigned(mlir::FunctionOpInterface)> getSharedMemoryLimit =
-        nullptr,
     std::function<unsigned(mlir::FunctionOpInterface)> getIndexBitwidth =
         nullptr);
 

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp
@@ -992,13 +992,8 @@ static void addLowerToLLVMGPUPasses(OpPassManager &modulePassManager,
   // Run checks on shared memory usage.
   funcPassManager
       .addPass([&]() {
-        auto getSharedMemoryLimit = [](mlir::FunctionOpInterface entryPoint) {
-          IREE::GPU::TargetAttr target = getGPUTargetAttr(entryPoint);
-          return target.getWgp().getMaxWorkgroupMemoryBytes();
-        };
         auto getIndexBitwidth = [](mlir::FunctionOpInterface) { return 64; };
-        return createGPUCheckResourceUsagePass(getSharedMemoryLimit,
-                                               getIndexBitwidth);
+        return createGPUCheckResourceUsagePass(getIndexBitwidth);
       })
       // SCF -> CF
       .addPass(createConvertSCFToCFPass)

diff --git a/compiler/src/iree/compiler/Codegen/SPIRV/Passes.cpp b/compiler/src/iree/compiler/Codegen/SPIRV/Passes.cpp
@@ -183,18 +183,11 @@ static void addMemRefLoweringPasses(OpPassManager &modulePassManager) {
 
       .addPass(createPadDynamicAlloc);
 
-  // Check to make sure we are not exceeding shared memory usage limit.
-  auto getSharedMemoryLimit = [](mlir::FunctionOpInterface fn) {
-    IREE::GPU::TargetAttr target = getGPUTargetAttr(fn);
-    return target.getWgp().getMaxWorkgroupMemoryBytes();
-  };
   // TODO: query this from the target.
   auto getIndexBitwidth = [](mlir::FunctionOpInterface) { return 32; };
   funcPassManager
-      .addPass([&]() {
-        return createGPUCheckResourceUsagePass(getSharedMemoryLimit,
-                                               getIndexBitwidth);
-      })
+      .addPass(
+          [&]() { return createGPUCheckResourceUsagePass(getIndexBitwidth); })
 
       // Fold load/store from/to subview ops into the original memref when
       // possible. In SPIR-V we don't use memref descriptor so it's not possible