File size: 5,323 Bytes
23bd7af |
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 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 |
/* coding=utf-8
* Copyright (c) 2020, NVIDIA CORPORATION. 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.
*/
/*This code is copied fron NVIDIA apex:
* https://github.com/NVIDIA/apex
* with minor changes. */
#include <torch/extension.h>
#include <vector>
#include <cassert>
#include "compat.h"
namespace {
void compute_n1_n2(
at::Tensor input,
at::IntArrayRef normalized_shape,
int& n1,
int& n2) {
int idiff = input.ndimension() - normalized_shape.size();
n2 = 1;
for (int i = 0; i < (int)normalized_shape.size(); ++i) {
assert( input.sizes()[i+idiff] == normalized_shape[i] );
n2 *= normalized_shape[i];
}
n1 = 1;
for (int i = 0; i < idiff; ++i) {
n1 *= input.sizes()[i];
}
}
void check_args(
at::IntArrayRef normalized_shape,
at::Tensor gamma,
at::Tensor beta
)
{
TORCH_CHECK(!gamma.defined() || gamma.sizes().equals(normalized_shape));
TORCH_CHECK(!beta.defined() || beta.sizes().equals(normalized_shape));
}
void check_args(
at::Tensor input,
at::IntArrayRef normalized_shape,
int& n1,
int& n2
)
{
int64_t normalized_ndim = normalized_shape.size();
if (normalized_ndim < 1) {
std::stringstream ss;
ss << "Expected normalized_shape to be at least 1-dimensional, i.e., "
<< "containing at least one element, but got normalized_shape="
<< normalized_shape;
throw std::runtime_error(ss.str());
}
auto input_shape = input.sizes();
auto input_ndim = input.dim();
if (input_ndim < normalized_ndim ||
!input_shape.slice(input_ndim - normalized_ndim).equals(normalized_shape)) {
std::stringstream ss;
ss << "Given normalized_shape=" << normalized_shape
<< ", expected input with shape [*";
for (auto size : normalized_shape) {
ss << ", " << size;
}
ss << "], but got input of size" << input_shape;
throw std::runtime_error(ss.str());
}
compute_n1_n2(input,normalized_shape,n1,n2);
}
void check_args(
at::Tensor input,
at::IntArrayRef normalized_shape,
at::Tensor gamma,
at::Tensor beta,
int& n1,
int& n2
)
{
check_args(input,normalized_shape,n1,n2);
check_args(normalized_shape,gamma,beta);
}
}
void cuda_layer_norm(
at::Tensor* output,
at::Tensor* mean,
at::Tensor* invvar,
at::Tensor* input,
int n1,
int n2,
at::IntArrayRef normalized_shape,
at::Tensor* gamma,
at::Tensor* beta,
double epsilon);
#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
std::vector<at::Tensor> layer_norm_affine(
at::Tensor input,
at::IntArrayRef normalized_shape,
at::Tensor gamma,
at::Tensor beta,
double epsilon) {
CHECK_INPUT(input);
CHECK_INPUT(gamma);
CHECK_INPUT(beta);
int n1, n2;
check_args(input, normalized_shape, gamma, beta, n1, n2);
at::Tensor output = at::empty_like(
input, gamma.options().dtype(gamma.scalar_type()));
at::Tensor mean = at::empty(
{n1}, input.options().dtype(at::ScalarType::Float));
at::Tensor invvar = at::empty_like(mean);
cuda_layer_norm(&output, &mean, &invvar, &input, n1, n2,
normalized_shape, &gamma, &beta, epsilon);
return {output, mean, invvar};
}
void cuda_layer_norm_gradient(
at::Tensor* dout,
at::Tensor* mean,
at::Tensor* invvar,
at::Tensor* input,
int n1,
int n2,
at::IntArrayRef normalized_shape,
at::Tensor* gamma,
at::Tensor* beta,
double epsilon,
at::Tensor* grad_input,
at::Tensor* grad_gamma,
at::Tensor* grad_beta
);
std::vector<at::Tensor> layer_norm_gradient_affine(
at::Tensor dout,
at::Tensor mean,
at::Tensor invvar,
at::Tensor input,
at::IntArrayRef normalized_shape,
at::Tensor gamma,
at::Tensor beta,
double epsilon) {
CHECK_INPUT(dout);
CHECK_INPUT(mean);
CHECK_INPUT(invvar);
CHECK_INPUT(input);
CHECK_INPUT(gamma);
CHECK_INPUT(beta);
int n1, n2;
check_args(input, normalized_shape, gamma, beta, n1, n2);
at::Tensor grad_input = at::empty_like(input);
at::Tensor grad_gamma = at::empty_like(gamma);
at::Tensor grad_beta = at::empty_like(beta);
cuda_layer_norm_gradient(&dout, &mean, &invvar, &input, n1, n2,
normalized_shape, &gamma, &beta, epsilon,
&grad_input, &grad_gamma, &grad_beta);
return {grad_input, grad_gamma, grad_beta};
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward_affine", &layer_norm_affine,
"LayerNorm forward (CUDA)");
m.def("backward_affine", &layer_norm_gradient_affine,
"LayerNorm backward (CUDA)");
}
|