1 /*******************************************************************************
2 * Copyright 2016-2018 Intel Corporation
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 *******************************************************************************/
21 #include "c_types_map.hpp"
22 #include "type_helpers.hpp"
23 #include "mkldnn_thread.hpp"
25 #include "ref_softmax.hpp"
26 #include "gemm/os_blas.hpp"
29 #include "mkl_vml_functions.h"
36 template <impl::data_type_t data_type>
37 void ref_softmax_fwd_t<data_type>::execute_forward_dense() const {
38 auto src = reinterpret_cast<const data_t *>(this->input_memory(0));
39 auto dst = reinterpret_cast<data_t *>(this->memory(0));
41 int outer_size_ = utils::array_product(pd()->src_pd()->desc()->dims, pd()->desc()->softmax_axis);
43 if (outer_size_ == 1) {
44 for (int ou = 0; ou < outer_size_; ou++) {
45 const data_t *src_data = src + ou * channels_;
46 data_t *dst_data = dst + ou * channels_;
49 _max(channels_, src_data, &scalar);
50 _sub(channels_, scalar, src_data, dst_data);
51 _exp_parallel(channels_, dst_data, dst_data);
52 _sum(channels_, dst_data, &scalar);
53 _scal_parallel(channels_, data_t(1) / scalar, dst_data);
56 parallel_nd(outer_size_, [&](int ou) {
57 const data_t *src_data = src + ou * channels_;
58 data_t *dst_data = dst + ou * channels_;
61 _max(channels_, src_data, &scalar);
62 _sub(channels_, scalar, src_data, dst_data);
63 _exp(channels_, dst_data, dst_data);
64 _sum(channels_, dst_data, &scalar);
65 _scal(channels_, data_t(1) / scalar, dst_data);
70 template <impl::data_type_t data_type>
71 void ref_softmax_fwd_t<data_type>::execute_forward_generic() const {
72 auto src = reinterpret_cast<const data_t *>(this->input_memory(0));
73 auto dst = reinterpret_cast<data_t *>(this->memory(0));
75 data_t space_max_val = 0, space_denom_val = 0;
76 data_t *space_max = &space_max_val, *space_denom = &space_denom_val;
77 if (inner_size_ > 1) {
78 using namespace memory_tracking::names;
79 space_max = scratchpad().template get<data_t>(key_softmax_reduction);
80 space_denom = space_max + inner_size_;
83 const memory_desc_wrapper data_d(pd()->src_pd());
84 const size_t dim = channels_ * inner_size_;
86 int outer_size_ = utils::array_product(pd()->src_pd()->desc()->dims, pd()->desc()->softmax_axis);
88 for (int ou = 0; ou < outer_size_; ou++) {
89 utils::array_set(space_max, -FLT_MAX, inner_size_);
90 utils::array_set(space_denom, 0, inner_size_);
92 for (int c = 0; c < channels_; c++) {
93 for(int in = 0; in < inner_size_; in++) {
94 size_t off = data_d.off_l(ou * dim + c * inner_size_ + in);
95 space_max[in] = nstl::max(space_max[in], src[off]);
99 for (int c = 0; c < channels_; c++) {
100 for(int in = 0; in < inner_size_; in++) {
101 size_t off = data_d.off_l(ou * dim + c * inner_size_ + in);
102 space_denom[in] += dst[off] = exp(src[off] - space_max[in]);
106 for (int c = 0; c < channels_; c++) {
107 for (int in = 0; in < inner_size_; in++) {
108 size_t off = data_d.off_l(ou * dim + c * inner_size_ + in);
109 dst[off] /= space_denom[in];
115 template <impl::data_type_t data_type>
116 void ref_softmax_fwd_t<data_type>::_max(int n, const data_t *x,
117 data_t *max_data) const {
118 // Intel(R) C++ Compiler generates the maxps + shuffle pattern
119 // for the max search which works faster
120 #if !defined(__INTEL_COMPILER)
121 // The code below makes a compiler to generate maxps instruction
122 // rather than maxss, which is generated for the 'else' code path
123 auto max_wrapper = [](data_t a, data_t b) { return nstl::max(a, b); };
124 auto min_wrapper = [](int a, int b) { return nstl::min(a, b); };
126 constexpr int unroll_factor = 32;
127 data_t max_values[unroll_factor];
129 if (n < unroll_factor) {
130 data_t max_val = x[0];
131 for (int i = 1; i < n; i++) {
132 max_val = max_wrapper(max_val, x[i]);
134 max_data[0] = max_val;
137 for (int i = 0; i < unroll_factor; i++) {
138 max_values[i] = x[i];
140 for (int i = unroll_factor; i < n; i += unroll_factor) {
141 int offset = min_wrapper(i, n - unroll_factor);
142 for (int j = 0; j < unroll_factor; j++) {
143 max_values[j] = max_wrapper(max_values[j], x[offset + j]);
146 data_t max_val = max_values[0];
147 for (int i = 1; i < unroll_factor; i++) {
148 max_val = max_wrapper(max_val, max_values[i]);
150 max_data[0] = max_val;
153 for (int c = 1; c < n; ++c)
154 max_data[0] = nstl::max(max_data[0], x[c]);
158 template <impl::data_type_t data_type>
159 void ref_softmax_fwd_t<data_type>::_sub(int n, data_t alpha, const data_t *x,
161 constexpr int unroll_factor = 32;
162 int tail = n % unroll_factor;
163 for (int i = 0; i < n - tail; i += unroll_factor) {
165 for (int j = 0; j < unroll_factor; j++) {
166 y[i + j] = x[i + j] - alpha;
170 for (int i = n - tail; i < n; i++) {
175 template <impl::data_type_t data_type>
176 void ref_softmax_fwd_t<data_type>::_exp_parallel(int n, const data_t *a, data_t *r) const {
178 if (data_type == data_type::f32) {
183 parallel_nd(n, [&](int c) { r[c] = expf(a[c]); });
186 template <impl::data_type_t data_type>
187 void ref_softmax_fwd_t<data_type>::_exp(int n, const data_t *a, data_t *r) const {
188 for (int c = 0; c < n; c++)
192 template <impl::data_type_t data_type>
193 void ref_softmax_fwd_t<data_type>::_sum(int n, const data_t *x,
194 data_t *sum_data) const {
196 // Here we are summing x's eg. e^z , which are positives
197 // so we can use BLAS ASUM
198 if (data_type == data_type::f32) {
199 sum_data[0] = cblas_sasum(n, x, 1);
203 data_t tsum = static_cast<data_t>(0);
204 PRAGMA_OMP_SIMD(reduction(+ : tsum))
205 for (int c = 0; c < n; ++c)
210 template <impl::data_type_t data_type>
211 void ref_softmax_fwd_t<data_type>::_scal_parallel(int n, data_t alpha, data_t *x) const {
213 if (data_type == data_type::f32) {
214 cblas_sscal(n, alpha, x, 1);
218 parallel_nd(n, [&](int c) { x[c] *= alpha; });
221 template <impl::data_type_t data_type>
222 void ref_softmax_fwd_t<data_type>::_scal(int n, data_t alpha, data_t *x) const {
223 for (int c = 0; c < n; c++)
227 template struct ref_softmax_fwd_t<data_type::f32>;
230 // NC/NCHW softmax for along final axe (1 for NC, 3 for NCHW)
231 template <impl::data_type_t data_type>
232 void ref_softmax_bwd_t<data_type>::execute_backward_dense() const {
233 auto data = reinterpret_cast<const data_t *>(this->input_memory(0));
234 auto diff_dst = reinterpret_cast<const data_t *>(this->input_memory(1));
235 auto diff_src = reinterpret_cast<data_t *>(this->memory(0));
237 parallel_nd(outer_size_, [&](int ou) {
239 size_t off = channels_*ou;
240 for (int c = 0; c < channels_; c++) {
241 size_t loff = off + c;
242 data_t ldata = data[loff];
243 sbr += diff_dst[loff]*ldata;
244 diff_src[loff] = ldata;
247 for(int c=0; c < channels_ ; ++c) {
248 size_t loff = off + c;
249 diff_src[loff] *= (diff_dst[loff] - sbr);
254 template <impl::data_type_t data_type>
255 void ref_softmax_bwd_t<data_type>::execute_backward_generic() const {
256 const size_t dim = channels_ * inner_size_;
257 auto data = reinterpret_cast<const data_t *>(this->input_memory(0));
258 auto diff_dst = reinterpret_cast<const data_t *>(this->input_memory(1));
259 auto diff_src = reinterpret_cast<data_t *>(this->memory(0));
260 const memory_desc_wrapper diff_d(pd()->diff_src_pd());
261 const memory_desc_wrapper data_d(pd()->dst_pd());
263 parallel_nd(outer_size_, [&](int ou) {
264 for (int in = 0; in < inner_size_; in++) {
266 for (int c = 0; c < channels_; c++) {
267 size_t off_diff = diff_d.off_l(ou * dim + c * inner_size_ + in);
268 size_t off_data = diff_d.off_l(ou * dim + c * inner_size_ + in);
269 sbr += diff_dst[off_diff]*data[off_data];
272 for(int c=0; c < channels_ ; ++c) {
273 size_t off_diff = diff_d.off_l(ou * dim + c * inner_size_ + in);
274 size_t off_data = data_d.off_l(ou * dim + c * inner_size_ + in);
275 diff_src[off_diff] = data[off_data]*(diff_dst[off_diff] - sbr);
281 template struct ref_softmax_bwd_t<data_type::f32>;
287 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s