| 1 | //===- IndexedAccessOpInterfaceImpl.cpp -----------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // Implement IndexedAccessOpInterface on GPU dialect operations that have |
| 9 | // %memref[%i0, %i1, ...] arguments to allow them to be manipulated by |
| 10 | // generic memref-dialect passes. |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/GPU/Transforms/IndexedAccessOpInterfaceImpl.h" |
| 14 | |
| 15 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 16 | #include "mlir/Dialect/MemRef/IR/MemoryAccessOpInterfaces.h" |
| 17 | #include "mlir/IR/Dialect.h" |
| 18 | #include "mlir/IR/Operation.h" |
| 19 | #include "mlir/IR/PatternMatch.h" |
| 20 | |
| 21 | using namespace mlir; |
| 22 | using namespace mlir::memref; |
| 23 | using namespace mlir::gpu; |
| 24 | |
| 25 | /// Given a GPU matrix type that will be loaded or stored, the leading dimension |
| 26 | /// of the matrix in memory, and whether or not the matrix is transposed, |
| 27 | /// compute the size of the linear memory that the load/store spans as |
| 28 | /// dC + leadingDim * (dR - 1) where dR and dC are the non-contiguous and |
| 29 | /// contiguous matrix dimensions, respectively (we get to the dX-1th row and |
| 30 | /// then access the first dY elements of it). |
| 31 | static int64_t get1DAccessSize(MMAMatrixType matrixType, int64_t leadingDim, |
| 32 | bool transpose) { |
| 33 | assert(matrixType.getShape().size() == 2 && "expected matrices to be 2D" ); |
| 34 | |
| 35 | int64_t c = matrixType.getShape()[1]; |
| 36 | int64_t r = matrixType.getShape()[0]; |
| 37 | if (transpose) |
| 38 | std::swap(a&: c, b&: r); |
| 39 | return c + leadingDim * (r - 1); |
| 40 | } |
| 41 | |
| 42 | namespace { |
| 43 | struct SubgroupMmaLoadMatrixOpImpl final |
| 44 | : IndexedAccessOpInterface::ExternalModel<SubgroupMmaLoadMatrixOpImpl, |
| 45 | SubgroupMmaLoadMatrixOp> { |
| 46 | TypedValue<MemRefType> getAccessedMemref(Operation *op) const { |
| 47 | return cast<SubgroupMmaLoadMatrixOp>(Val: op).getSrcMemref(); |
| 48 | } |
| 49 | |
| 50 | Operation::operand_range getIndices(Operation *op) const { |
| 51 | return cast<SubgroupMmaLoadMatrixOp>(Val: op).getIndices(); |
| 52 | } |
| 53 | |
| 54 | /// This returns a 1-D shape so that it's clear that both linearization and |
| 55 | /// folding in expand/collapse_shape operations are allowed. |
| 56 | SmallVector<int64_t> getAccessedShape(Operation *op) const { |
| 57 | auto loadOp = cast<SubgroupMmaLoadMatrixOp>(Val: op); |
| 58 | return {get1DAccessSize(matrixType: cast<MMAMatrixType>(Val: loadOp.getRes().getType()), |
| 59 | leadingDim: loadOp.getLeadDimension().getZExtValue(), |
| 60 | transpose: loadOp.getTranspose().value_or(u: false))}; |
| 61 | } |
| 62 | |
| 63 | std::optional<SmallVector<Value>> |
| 64 | updateMemrefAndIndices(Operation *op, RewriterBase &rewriter, Value newMemref, |
| 65 | ValueRange newIndices) const { |
| 66 | auto loadOp = cast<SubgroupMmaLoadMatrixOp>(Val: op); |
| 67 | rewriter.modifyOpInPlace(root: loadOp, callable: [&]() { |
| 68 | loadOp.getSrcMemrefMutable().assign(value: newMemref); |
| 69 | loadOp.getIndicesMutable().assign(values: newIndices); |
| 70 | }); |
| 71 | return std::nullopt; |
| 72 | } |
| 73 | |
| 74 | bool hasInboundsIndices(Operation *) const { return true; } |
| 75 | }; |
| 76 | |
| 77 | struct SubgroupMmaStoreMatrixOpImpl final |
| 78 | : IndexedAccessOpInterface::ExternalModel<SubgroupMmaStoreMatrixOpImpl, |
| 79 | SubgroupMmaStoreMatrixOp> { |
| 80 | TypedValue<MemRefType> getAccessedMemref(Operation *op) const { |
| 81 | return cast<SubgroupMmaStoreMatrixOp>(Val: op).getDstMemref(); |
| 82 | } |
| 83 | |
| 84 | Operation::operand_range getIndices(Operation *op) const { |
| 85 | return cast<SubgroupMmaStoreMatrixOp>(Val: op).getIndices(); |
| 86 | } |
| 87 | |
| 88 | /// This returns a 1-D shape so that it's clear that both linearization and |
| 89 | /// folding in expand/collapse_shape operations are allowed. |
| 90 | SmallVector<int64_t> getAccessedShape(Operation *op) const { |
| 91 | auto storeOp = cast<SubgroupMmaStoreMatrixOp>(Val: op); |
| 92 | return {get1DAccessSize(matrixType: storeOp.getSrc().getType(), |
| 93 | leadingDim: storeOp.getLeadDimension().getZExtValue(), |
| 94 | transpose: storeOp.getTranspose().value_or(u: false))}; |
| 95 | } |
| 96 | |
| 97 | std::optional<SmallVector<Value>> |
| 98 | updateMemrefAndIndices(Operation *op, RewriterBase &rewriter, Value newMemref, |
| 99 | ValueRange newIndices) const { |
| 100 | auto storeOp = cast<SubgroupMmaStoreMatrixOp>(Val: op); |
| 101 | rewriter.modifyOpInPlace(root: storeOp, callable: [&]() { |
| 102 | storeOp.getDstMemrefMutable().assign(value: newMemref); |
| 103 | storeOp.getIndicesMutable().assign(values: newIndices); |
| 104 | }); |
| 105 | return std::nullopt; |
| 106 | } |
| 107 | |
| 108 | bool hasInboundsIndices(Operation *) const { return true; } |
| 109 | }; |
| 110 | } // namespace |
| 111 | |
| 112 | void mlir::gpu::registerIndexedAccessOpInterfaceExternalModels( |
| 113 | DialectRegistry ®istry) { |
| 114 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, gpu::GPUDialect *dialect) { |
| 115 | SubgroupMmaLoadMatrixOp::attachInterface<SubgroupMmaLoadMatrixOpImpl>(context&: *ctx); |
| 116 | SubgroupMmaStoreMatrixOp::attachInterface<SubgroupMmaStoreMatrixOpImpl>( |
| 117 | context&: *ctx); |
| 118 | }); |
| 119 | } |
| 120 | |