/
softmax_grad_kernel.cc
91 lines (75 loc) · 3.22 KB
/
softmax_grad_kernel.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
Licensed 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. */
#include "paddle/phi/kernels/sparse/softmax_grad_kernel.h"
#include "paddle/fluid/platform/cpu_info.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/visit_type.h"
#include "paddle/phi/kernels/funcs/cpu_vec.h"
#include "paddle/phi/kernels/sparse/empty_kernel.h"
namespace plt = paddle::platform;
namespace phi {
namespace sparse {
template <typename T, typename Context>
void SoftmaxCsrGradKernel(const Context& dev_ctx,
const SparseCsrTensor& out,
const SparseCsrTensor& dout,
int axis,
SparseCsrTensor* dx) {
PADDLE_ENFORCE_EQ(axis,
-1,
phi::errors::Unimplemented(
"SparseCsrTensor only support axis=-1 for softmax, "
"which is faster when reading data by row (axis=-1)"));
EmptyLikeCsrKernel<T, Context>(dev_ctx, dout, dx);
auto out_dim = out.dims();
int rows = 1;
for (int i = 0; i < out_dim.size() - 1; ++i) {
rows *= out_dim[i];
}
const DenseTensor& out_crows = out.non_zero_crows();
const DenseTensor& out_values = out.non_zero_elements();
const DenseTensor& dout_values = dout.non_zero_elements();
DenseTensor* dx_values = dx->mutable_non_zero_elements();
int row_first = 0;
int row_nnz = 0;
const T* out_data = out_values.data<T>();
const T* dout_data = dout_values.data<T>();
T* dx_data = dx_values->data<T>();
// dx = (dout - sum(dout * out)) * out
PD_VISIT_INTEGRAL_TYPES(
out.non_zero_crows().dtype(), "SoftmaxCsrGradKernel", ([&] {
const data_t* out_crows_data = out_crows.data<data_t>();
for (int i = 0; i < rows; ++i) {
row_first = static_cast<int>(out_crows_data[i]);
row_nnz = static_cast<int>(out_crows_data[i + 1] - out_crows_data[i]);
out_data = out_data + row_first;
dout_data = dout_data + row_first;
dx_data = dx_data + row_first;
T sum = 0;
phi::funcs::vec_mul_reduce<T, plt::avx>(
row_nnz, dout_data, out_data, &sum);
phi::funcs::vec_add_bias<T, plt::avx>(
row_nnz, static_cast<T>(-1) * sum, dout_data, dx_data);
phi::funcs::vec_mul<T, plt::avx>(row_nnz, dx_data, out_data, dx_data);
}
}));
}
} // namespace sparse
} // namespace phi
PD_REGISTER_KERNEL(softmax_csr_grad,
CPU,
ALL_LAYOUT,
phi::sparse::SoftmaxCsrGradKernel,
float,
double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_CSR);
}