samskalicky commented on a change in pull request #15921: dynamic custom operator support URL: https://github.com/apache/incubator-mxnet/pull/15921#discussion_r334231752
########## File path: example/extensions/lib_custom_op/gemm_lib.cc ########## @@ -0,0 +1,240 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * Copyright (c) 2019 by Contributors + * \file gemm_lib.cc + * \brief Sample 2D gemm custom operator implementation library file + */ + +#include <iostream> +#include "lib_api.h" + +// main matrix multiplication routine +void gemm(const float* A, const float* B, float* C, + const unsigned n, const unsigned k, const unsigned m) { + unsigned i, j, kk; + for (i = 0; i < n; i++) { + for (j = 0; j < m; j++) { + C[i*m+j] = 0; + for (kk = 0; kk < k; kk++) { + C[i*m+j] += A[i*k+kk] * B[kk*m+j]; + } + } + } +} + +void transpose(const float* A, float* At, const unsigned n, const unsigned m) { + unsigned i, j; + for (i = 0; i < n; i++) { + for (j = 0; j < m; j++) { + At[i*m+j] = A[j*n+i]; + } + } +} + +/* + * Executes C = A * B + * inputs[0] = A; inputs[1] = B; outputs[0] = C + */ +MXReturnValue forward(std::map<std::string, std::string> attrs, + std::vector<MXTensor> inputs, + std::vector<MXTensor> outputs, + OpResource res) { + // simple example of using runtime data type + if (inputs[0].dtype == kFloat32) { + typedef float DType; + // extract data pointers from tensors + DType* A = inputs[0].data<DType>(); + DType* B = inputs[1].data<DType>(); + DType* C = outputs[0].data<DType>(); + // set tensor shapes + unsigned n = inputs[0].shape[0]; + unsigned k = inputs[0].shape[1]; + unsigned m = inputs[1].shape[1]; + + gemm(A, B, C, n, k, m); + } + return MX_SUCCESS; +} + +/* + * Executes dA = dC * B.T; Executes dB = A.T * dC + ***** gradient inputs + * inputs[0] = dC + ***** original inputs + * inputs[1] = A; inputs[2] = B + ***** original outputs + * inputs[3] = C + ***** gradient outputs + * outputs[0] = dA; outputs[1] = dB + */ +MXReturnValue backward(std::map<std::string, std::string> attrs, + std::vector<MXTensor> inputs, + std::vector<MXTensor> outputs, + OpResource res) { + // extract data pointers from tensors + float* dC = inputs[0].data<float>(); + float* A = inputs[1].data<float>(); + float* B = inputs[2].data<float>(); + float* dA = outputs[0].data<float>(); + float* dB = outputs[1].data<float>(); + // set tensor shapes + unsigned n = inputs[1].shape[0]; + unsigned k = inputs[1].shape[1]; + unsigned m = inputs[2].shape[1]; + + float *At = new float[k*n]; Review comment: @wkcn remember in this PR there is only CPU operator support. The next PR will have GPU support and we’ll make sure to distinguish cpu/gpu allocations. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
