| 1 | //===- Utils.cpp - GPU transforms utils -----------------------------------===// |
|---|---|
| 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 | // |
| 9 | // Implements GPU dialect transforms utils. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/GPU/Utils/GPUUtils.h" |
| 14 | #include "llvm/Support/ErrorHandling.h" |
| 15 | #include "llvm/Support/MathExtras.h" |
| 16 | |
| 17 | namespace mlir::gpu { |
| 18 | |
| 19 | vector::CombiningKind convertReductionKind(gpu::AllReduceOperation mode) { |
| 20 | switch (mode) { |
| 21 | #define MAP_CASE(X) \ |
| 22 | case gpu::AllReduceOperation::X: \ |
| 23 | return vector::CombiningKind::X |
| 24 | |
| 25 | MAP_CASE(ADD); |
| 26 | MAP_CASE(MUL); |
| 27 | MAP_CASE(MINUI); |
| 28 | MAP_CASE(MINSI); |
| 29 | MAP_CASE(MINNUMF); |
| 30 | MAP_CASE(MAXSI); |
| 31 | MAP_CASE(MAXUI); |
| 32 | MAP_CASE(MAXNUMF); |
| 33 | MAP_CASE(AND); |
| 34 | MAP_CASE(OR); |
| 35 | MAP_CASE(XOR); |
| 36 | MAP_CASE(MINIMUMF); |
| 37 | MAP_CASE(MAXIMUMF); |
| 38 | MAP_CASE(MINIMUMNUMF); |
| 39 | MAP_CASE(MAXIMUMNUMF); |
| 40 | |
| 41 | #undef MAP_CASE |
| 42 | } |
| 43 | |
| 44 | llvm_unreachable("Vector and GPU reduction kinds should match 1:1"); |
| 45 | } |
| 46 | |
| 47 | FailureOr<SubgroupReduceClusterInfo> |
| 48 | getSubgroupReduceClusterInfo(SubgroupReduceOp op, unsigned subgroupSize) { |
| 49 | assert(llvm::isPowerOf2_32(subgroupSize)); |
| 50 | |
| 51 | std::optional<uint32_t> clusterSize = op.getClusterSize(); |
| 52 | assert(!clusterSize || |
| 53 | llvm::isPowerOf2_32(*clusterSize)); // Verifier should've caught this. |
| 54 | if (clusterSize && *clusterSize > subgroupSize) |
| 55 | return op.emitOpError() |
| 56 | << "cluster size "<< *clusterSize |
| 57 | << " is greater than subgroup size "<< subgroupSize; |
| 58 | unsigned effectiveClusterSize = clusterSize.value_or(u&: subgroupSize); |
| 59 | |
| 60 | auto clusterStride = op.getClusterStride(); |
| 61 | assert(llvm::isPowerOf2_32(clusterStride)); // Verifier should've caught this. |
| 62 | if (clusterStride >= subgroupSize) |
| 63 | return op.emitOpError() |
| 64 | << "cluster stride "<< clusterStride |
| 65 | << " is not less than subgroup size "<< subgroupSize; |
| 66 | |
| 67 | return SubgroupReduceClusterInfo{.clusterStride: clusterStride, .clusterSize: effectiveClusterSize, |
| 68 | .subgroupSize: subgroupSize}; |
| 69 | } |
| 70 | } // namespace mlir::gpu |
| 71 |