You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2018/05/11 18:52:02 UTC
[GitHub] piiswrong commented on a change in pull request #10104:
[WIP][MXNET-107] Fused LSTM implementation for CPU
piiswrong commented on a change in pull request #10104: [WIP][MXNET-107] Fused LSTM implementation for CPU
URL: https://github.com/apache/incubator-mxnet/pull/10104#discussion_r187701481
##########
File path: src/operator/rnn_impl.hpp
##########
@@ -0,0 +1,454 @@
+/*
+ * 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) 2015 by Contributors
+ * \file rnn_impl.hpp
+ * \brief
+ * \author Shu Zhang(shu.zhang@intel.com)
+*/
+#ifndef MXNET_OPERATOR_RNN_IMPL_HPP_
+#define MXNET_OPERATOR_RNN_IMPL_HPP_
+
+#include <dmlc/logging.h>
+#include <dmlc/parameter.h>
+#include <mxnet/operator.h>
+#include <algorithm>
+#include <map>
+#include <vector>
+#include <string>
+#include <utility>
+#include "./math.h"
+#include "./math_functions-inl.h"
+#include "./operator_common.h"
+#include "./mshadow_op.h"
+#include "./linalg.h"
+
+template<typename DType>
+inline DType sigmoid(DType x) {
+ return 1.0f / (1.0f + exp(-x));
+}
+
+template<typename DType>
+void LstmForwardTrainingSingleLayer(DType* ws,
+ DType* rs,
+ bool state_outputs,
+ bool bid,
+ const int T,
+ const int N,
+ const int I,
+ const int H,
+ const Tensor<cpu, 2, DType> &x,
+ const Tensor<cpu, 2, DType> &hx,
+ const Tensor<cpu, 2, DType> &cx,
+ const Tensor<cpu, 3, DType> &y,
+ DType* w_ptr,
+ DType* b_ptr,
+ DType* hy_ptr,
+ DType* cy_ptr) {
+ using namespace mshadow;
+ const Tensor<cpu, 2, DType> wx(w_ptr, Shape2(H * 4, I));
+ const Tensor<cpu, 2, DType> wh(w_ptr + I * H * 4, Shape2(H * 4, H));
+ const Tensor<cpu, 2, DType> bx(b_ptr, Shape2(4, H));
+ const Tensor<cpu, 2, DType> bh(b_ptr + H * 4, Shape2(4, H));
+ const Tensor<cpu, 2, DType> yx_flat(ws, Shape2(T * N, 4 * H));
+ const Tensor<cpu, 2, DType> yh_flat(ws + T * N * H * 4, Shape2(N, 4 * H));
+ const Tensor<cpu, 4, DType> yx(yx_flat.dptr_, Shape4(T, N, 4, H));
+ const Tensor<cpu, 3, DType> yh(yh_flat.dptr_, Shape3(N, 4, H));
+ Tensor<cpu, 2, DType> h(yh_flat.dptr_ + N * H * 4, Shape2(N, H));
+ DType *c_ptr = bid ? rs + T * N * H * 7 : rs;
+ Tensor<cpu, 3, DType> c(c_ptr, Shape3(T, N, H));
+ Tensor<cpu, 4, DType> ifgo(c_ptr + T * N * H, Shape4(T, N, H, 4));
+
+ const int offset = bid ? H : 0;
+ const DType alpha = 1.0;
+ const DType beta = 0.0;
+ const int cell_size = N * H;
+ linalg_gemm(x, wx, yx_flat, alpha, beta, false, true);
+
+ for (int i = 0; i < T; ++i) {
+ int t = bid ? T - 1 - i : i;
+ linalg_gemm(i ? h : hx, wh, yh_flat, alpha, beta, false, true);
+ #pragma omp parallel for
+ for (int jk = 0; jk < cell_size; ++jk) {
+ int j = jk / H;
+ int k = jk % H;
+ DType it = sigmoid<DType>(yx[t][j][0][k] + yh[j][0][k] + bx[0][k] + bh[0][k]);
+ DType ft = sigmoid<DType>(yx[t][j][1][k] + yh[j][1][k] + bx[1][k] + bh[1][k]);
+ DType gt = tanh(yx[t][j][2][k] + yh[j][2][k] + bx[2][k] + bh[2][k]);
+ DType ot = sigmoid<DType>(yx[t][j][3][k] + yh[j][3][k] + bx[3][k] + bh[3][k]);
+ DType ct = (i ? c[i-1][j][k] : cx[j][k]) * ft + it * gt;
+ DType ht = ot * tanh(ct);
+ h[j][k] = ht;
+ // reserve
+ y[t][j][k + offset] = ht;
+ c[i][j][k] = ct;
+ ifgo[i][j][k][0] = it;
+ ifgo[i][j][k][1] = ft;
+ ifgo[i][j][k][2] = gt;
+ ifgo[i][j][k][3] = ot;
+ if (i == T - 1 && state_outputs) {
+ hy_ptr[jk] = ht;
+ cy_ptr[jk] = ct;
+ }
+ }
+ }
+}
+
+template <typename DType>
+void LstmForwardTraining(DType* ws,
+ DType* rs,
+ bool state_outputs,
+ const int L,
+ const int D,
+ const int T,
+ const int N,
+ const int I,
+ const int H,
+ DType* x_ptr,
+ DType* hx_ptr,
+ DType* cx_ptr,
+ DType* w_ptr,
+ DType* b_ptr,
+ DType* y_ptr,
+ DType* hy_ptr,
+ DType* cy_ptr) {
+ const int total_layers = D * L;
+ Tensor<cpu, 3, DType> hx(hx_ptr, Shape3(total_layers, N, H));
+ Tensor<cpu, 3, DType> cx(cx_ptr, Shape3(total_layers, N, H));
+ const int b_size = 2 * H * 4;
+ const int r_size = D * T * N * H * 6;
+ const int y_offset = T * N * H * 5;
+ const int cell_size = N * H;
+ int idx = 0; // state & cell state's idx;
+ for (int i = 0; i < L; ++i) {
+ const int input_size = i ? H * D : I;
+ const int w_size = (input_size + H) * H * 4;
+ Tensor<cpu, 2, DType> x(x_ptr, Shape2(T * N, input_size));
+ Tensor<cpu, 3, DType> y(rs + y_offset, Shape3(T, N, H * D));
+ LstmForwardTrainingSingleLayer<DType>(ws, rs, state_outputs, false, T, N, input_size, H, x,
+ hx[idx], cx[idx], y, w_ptr, b_ptr, hy_ptr, cy_ptr);
+ if (D == 2) {
+ w_ptr += w_size;
+ b_ptr += b_size;
+ ++idx;
+ if (state_outputs) {
+ hy_ptr += cell_size;
+ cy_ptr += cell_size;
+ }
+ LstmForwardTrainingSingleLayer<DType>(ws, rs, state_outputs, true, T, N, input_size, H, x,
+ hx[idx], cx[idx], y, w_ptr, b_ptr, hy_ptr, cy_ptr);
+ }
+ if (i != L - 1) {
+ w_ptr += w_size;
+ b_ptr += b_size;
+ x_ptr = y.dptr_;
+ rs += r_size;
+ ++idx;
+ if (state_outputs) {
+ hy_ptr += cell_size;
+ cy_ptr += cell_size;
+ }
+ }
+ }
+ memcpy(y_ptr, rs + y_offset, T * N * H * D * sizeof(DType));
Review comment:
why is copy needed?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services