[mlir][amdgpu] Remove shared memory optimization pass (#88225)

This implementation has a number of issues and ultimately does not work
on gfx9.
* It does not reduce bank conflicts with wide memory accesses.
* It does not correctly account for when LDS bank conflicts occur on
amdgpu.
* The implementation is too fragile to be used on real-world code. For
example, the code bails out on any `memref.subview` in the root op, even
when the subview is not a user of any of the `memref.alloc` ops.

I do not see how these can be easily fixed, therefore I think it's
better to delete this code.

Author: kuhar

mlir/include/mlir/Dialect/AMDGPU/CMakeLists.txt

@@ -1,3 +1,2 @@
 add_subdirectory(IR)
-add_subdirectory(TransformOps)
 add_subdirectory(Transforms)

mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td

@@ -29,23 +29,6 @@ def AMDGPU_Dialect : Dialect {
     "gpu::GPUDialect"
   ];
   let useDefaultAttributePrinterParser = 1;
-
-  let extraClassDeclaration = [{
-    /// Return true if the given MemRefType has an integer address
-    /// space that matches the ROCDL shared memory address space or
-    /// is a gpu::AddressSpaceAttr attribute with value 'workgroup`.
-    static bool hasSharedMemoryAddressSpace(MemRefType type);
-
-    /// Return true if the given Attribute has an integer address
-    /// space that matches the ROCDL shared memory address space or
-    /// is a gpu::AddressSpaceAttr attribute with value 'workgroup`.
-    static bool isSharedMemoryAddressSpace(Attribute type);
-
-    /// Defines the MemRef memory space attribute numeric value that indicates
-    /// a memref is located in shared memory. This should correspond to the
-    /// value used in ROCDL.
-    static constexpr unsigned kSharedMemoryAddressSpace = 3;
-  }];
 }
 
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/AMDGPU/TransformOps/AMDGPUTransformOps.h

@@ -20,7 +20,7 @@ namespace mlir {
 class ConversionTarget;
 namespace amdgpu {
 
-#define GEN_PASS_DECL
+#define GEN_PASS_DECL_AMDGPUEMULATEATOMICSPASS
 #define GEN_PASS_REGISTRATION
 #include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc"
 

mlir/include/mlir/Dialect/AMDGPU/TransformOps/AMDGPUTransformOps.td

@@ -30,24 +30,4 @@ def AmdgpuEmulateAtomicsPass : Pass<"amdgpu-emulate-atomics"> {
                         "Chipset that these operations will run on">];
 }
 
-def OptimizeSharedMemory : Pass<"amdgpu-optimize-shared-memory"> {
-  let summary = "Optimizes accesses to shared memory memrefs in order to reduce bank conflicts.";
-  let description = [{
-    This pass adds a transformation and pass to the AMDGPU dialect that
-    attempts to optimize reads/writes from a memref representing GPU shared
-    memory in order to avoid bank conflicts.
-  }];
-  let dependentDialects = [
-    "memref::MemRefDialect", "vector::VectorDialect"
-  ];
-  let options = [
-    Option<"sharedMemoryLineSizeBytes", "shared-memory-line-size-bytes", "int64_t",
-           /*default=*/"128",
-           "Shared memory line size in bytes">,
-    Option<"defaultVectorSizeBits", "default-vector-size-bits", "int64_t",
-           /*default=*/"128",
-           "Default vector size in bits">,
-  ];
-}
-
 #endif // MLIR_DIALECT_AMDGPU_TRANSFORMS_PASSES_TD_

mlir/include/mlir/Dialect/AMDGPU/TransformOps/CMakeLists.txt

@@ -23,7 +23,6 @@
 #include "mlir/Conversion/MemRefToLLVM/MemRefToLLVM.h"
 #include "mlir/Conversion/NVVMToLLVM/NVVMToLLVM.h"
 #include "mlir/Conversion/UBToLLVM/UBToLLVM.h"
-#include "mlir/Dialect/AMDGPU/TransformOps/AMDGPUTransformOps.h"
 #include "mlir/Dialect/Affine/TransformOps/AffineTransformOps.h"
 #include "mlir/Dialect/Bufferization/TransformOps/BufferizationTransformOps.h"
 #include "mlir/Dialect/Func/Extensions/AllExtensions.h"
@@ -67,7 +66,6 @@ inline void registerAllExtensions(DialectRegistry &registry) {
   ub::registerConvertUBToLLVMInterface(registry);
 
   // Register all transform dialect extensions.
-  amdgpu::registerTransformDialectExtension(registry);
   affine::registerTransformDialectExtension(registry);
   bufferization::registerTransformDialectExtension(registry);
   func::registerTransformDialectExtension(registry);

mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.h

@@ -1,4 +1,3 @@
 add_subdirectory(IR)
-add_subdirectory(TransformOps)
 add_subdirectory(Transforms)
 add_subdirectory(Utils)

mlir/include/mlir/Dialect/AMDGPU/Transforms/Passes.td

@@ -43,21 +43,6 @@ void AMDGPUDialect::initialize() {
       >();
 }
 
-bool amdgpu::AMDGPUDialect::isSharedMemoryAddressSpace(Attribute memorySpace) {
-  if (!memorySpace)
-    return false;
-  if (auto intAttr = llvm::dyn_cast<IntegerAttr>(memorySpace))
-    return intAttr.getInt() == AMDGPUDialect::kSharedMemoryAddressSpace;
-  if (auto gpuAttr = llvm::dyn_cast<gpu::AddressSpaceAttr>(memorySpace))
-    return gpuAttr.getValue() == gpu::AddressSpace::Workgroup;
-  return false;
-}
-
-bool amdgpu::AMDGPUDialect::hasSharedMemoryAddressSpace(MemRefType type) {
-  Attribute memorySpace = type.getMemorySpace();
-  return isSharedMemoryAddressSpace(memorySpace);
-}
-
 //===----------------------------------------------------------------------===//
 // 8-bit float ops
 //===----------------------------------------------------------------------===//