| 1 | //===- SCFToGPU.h - Convert loop nests to GPU kernels -----------*- C++ -*-===// |
| 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 | #ifndef MLIR_CONVERSION_SCFTOGPU_SCFTOGPU_H_ |
| 9 | #define MLIR_CONVERSION_SCFTOGPU_SCFTOGPU_H_ |
| 10 | |
| 11 | #include "mlir/Support/LLVM.h" |
| 12 | |
| 13 | namespace mlir { |
| 14 | class ConversionTarget; |
| 15 | class MLIRContext; |
| 16 | class Value; |
| 17 | class Operation; |
| 18 | class RewritePatternSet; |
| 19 | |
| 20 | namespace affine { |
| 21 | class AffineForOp; |
| 22 | } // namespace affine |
| 23 | |
| 24 | namespace scf { |
| 25 | class ForOp; |
| 26 | } // namespace scf |
| 27 | |
| 28 | /// Convert a perfect affine loop nest with the outermost loop identified by |
| 29 | /// `forOp` into a gpu::Launch operation. Map `numBlockDims` outer loops to |
| 30 | /// GPU blocks and `numThreadDims` to GPU threads. The bounds of the loops that |
| 31 | /// are mapped should be independent of the induction variables of the other |
| 32 | /// mapped loops. |
| 33 | /// |
| 34 | /// No check on the size of the block or grid, or on the validity of |
| 35 | /// parallelization is performed, it is under the responsibility of the caller |
| 36 | /// to strip-mine the loops and to perform the dependence analysis before |
| 37 | /// calling the conversion. |
| 38 | |
| 39 | // TODO: Consider removing this in favor of affine.for -> affine.parallel |
| 40 | // detection followed by an affine.parallel -> scf.parallel -> gpu.launch |
| 41 | // conversion |
| 42 | LogicalResult convertAffineLoopNestToGPULaunch(affine::AffineForOp forOp, |
| 43 | unsigned numBlockDims, |
| 44 | unsigned numThreadDims); |
| 45 | |
| 46 | /// Adds the conversion pattern from `scf.parallel` to `gpu.launch` to the |
| 47 | /// provided pattern list. |
| 48 | void populateParallelLoopToGPUPatterns(RewritePatternSet &patterns); |
| 49 | |
| 50 | /// Configures the rewrite target such that only `scf.parallel` operations that |
| 51 | /// are not rewritten by the provided patterns are legal. |
| 52 | void configureParallelLoopToGPULegality(ConversionTarget &target); |
| 53 | |
| 54 | /// Clean up after applyPartialConversion/applyFullConversion call. |
| 55 | void finalizeParallelLoopToGPUConversion(Operation *op); |
| 56 | |
| 57 | } // namespace mlir |
| 58 | |
| 59 | #endif // MLIR_CONVERSION_SCFTOGPU_SCFTOGPU_H_ |
| 60 | |