1 /* Copyright 2018 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 16 #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_CUDNN_FUSED_CONV_REWRITER_H_ 17 #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_CUDNN_FUSED_CONV_REWRITER_H_ 18 19 #include "tensorflow/compiler/xla/service/hlo_instructions.h" 20 #include "tensorflow/compiler/xla/service/hlo_pass_interface.h" 21 22 namespace xla { 23 namespace gpu { 24 25 // Rewrite the custom call targeting cudnnConvolutionForward to 26 // cudnnConvolutionBiasActivationForward by fusing applicable point-wise 27 // operations following forward convolution. This transform must run after 28 // cudnn_conv_rewriter. 29 // It is straightforward for floating point convolutions: 30 // transforming 31 // max(0, alpha1 * conv(x, w) + alpha2 * side_input + broadcast(bias)) 32 // to 33 // cudnnConvolutionBiasActivationForward(x, w, bias, alpha1, alpha2, side) 34 // 35 // Integer convolution requires additional patterns to match CuDNN semantics: 36 // #1 from 37 // cast<int8>(clamp<-128, 127>(conv(int8_x, int8_w))) 38 // to 39 // cudnnConvolutionForward<int8>(int8_x, int8_w) 40 // or #2 from 41 // cast<float>(conv(int8_x, int8_w)) 42 // to 43 // cudnnConvolutionForward<float>(int8_x, int8_w) 44 // or #3 from 45 // cast<int8>(clamp<-128, 127>(max(0, alpha1 * 46 // cast<float>(conv(int8_x, int8_w)) + 47 // alpha2 * cast<float>(int8_side) + 48 // broadcast(bias))) 49 // to 50 // cudnnConvolutionBiasActivationForward<int8>(int8_x, int8_w, bias, alpha1, 51 // alpha2, int8_side) 52 // or #4 from 53 // max(0, alpha1 * cast<float>(conv(int8_x, int8_w)) + 54 // alpha2 * float_side + broadcast(bias)) 55 // to 56 // cudnnConvolutionBiasActivationForward<float>(int8_x, int8_w, bias, alpha1, 57 // alpha2, float_side) 58 59 class CudnnFusedConvRewriter : public HloModulePass { 60 public: name()61 absl::string_view name() const override { 62 return "cudnn-fused-convolution-rewriter"; 63 } 64 65 StatusOr<bool> Run(HloModule* module) override; 66 }; 67 68 } // namespace gpu 69 } // namespace xla 70 71 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_CUDNN_FUSED_CONV_REWRITER_H_ 72