1 /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_GEMM_REWRITER_H_ 16 #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_GEMM_REWRITER_H_ 17 18 #include "absl/types/optional.h" 19 #include "tensorflow/compiler/xla/service/hlo_instructions.h" 20 #include "tensorflow/compiler/xla/service/hlo_module.h" 21 #include "tensorflow/compiler/xla/service/hlo_pass_interface.h" 22 23 namespace xla { 24 namespace gpu { 25 26 // cuBLAS GEMM in the most general form can run the following operation: 27 // 28 // (kAdd 29 // (kMultiply (kDot A B) alpha) 30 // (kMultiply C beta)) 31 // 32 // where A, B, C are matrixes and `alpha` and `beta` are host constants. 33 // The additional requirement is that C has no other users (otherwise, 34 // it does not make sense to fuse it inside the custom call). 35 // 36 // Both multiplication and addition can be avoided (equivalent to setting 37 // `alpha` to one and `beta` to zero). 38 // 39 // This pass pattern-matches the most general form of this instruction 40 // (we assume transposes are already folded), and rewrites it into a custom call 41 // where (A, B, C) are three operands respectively, and `alpha` and `beta` are 42 // stored in the backend config. 43 class GemmRewriter : public HloModulePass { 44 public: name()45 absl::string_view name() const override { return "cublas-gemm-rewriter"; } 46 47 StatusOr<bool> Run(HloModule* module) override; 48 }; 49 50 } // namespace gpu 51 } // namespace xla 52 53 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_GEMM_REWRITER_H_ 54