|
| 1 | +// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +// |
| 3 | +// Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +// you may not use this file except in compliance with the License. |
| 5 | +// You may obtain a copy of the License at |
| 6 | +// |
| 7 | +// http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +// |
| 9 | +// Unless required by applicable law or agreed to in writing, software |
| 10 | +// distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +// See the License for the specific language governing permissions and |
| 13 | +// limitations under the License. |
| 14 | + |
| 15 | +#include "paddle/phi/kernels/sgd_kernel.h" |
| 16 | +#include "paddle/fluid/operators/jit/kernels.h" |
| 17 | +#include "paddle/phi/backends/cpu/cpu_context.h" |
| 18 | +#include "paddle/phi/core/kernel_registry.h" |
| 19 | +#include "paddle/phi/kernels/funcs/eigen/common.h" |
| 20 | + |
| 21 | +namespace phi { |
| 22 | + |
| 23 | +template <typename T> |
| 24 | +void sgd_dense_param_dense_grad_impl(const DenseTensor& param, |
| 25 | + const DenseTensor& learning_rate, |
| 26 | + const DenseTensor& grad, |
| 27 | + DenseTensor* param_out) { |
| 28 | + const auto sz = param_out->numel(); |
| 29 | + paddle::operators::jit::sgd_attr_t attr(1, sz, 1, sz, 1); |
| 30 | + const T* lr = learning_rate.data<T>(); |
| 31 | + const T* param_data = param.data<T>(); |
| 32 | + const T* grad_data = grad.data<T>(); |
| 33 | + int64_t rows_idx = 0; |
| 34 | + T* out_data = param_out->data<T>(); |
| 35 | + |
| 36 | + auto sgd = |
| 37 | + paddle::operators::jit::KernelFuncs<paddle::operators::jit::SgdTuple<T>, |
| 38 | + phi::CPUPlace>::Cache() |
| 39 | + .At(attr); |
| 40 | + sgd(lr, param_data, grad_data, &rows_idx, out_data, &attr); |
| 41 | +} |
| 42 | + |
| 43 | +template <> |
| 44 | +void sgd_dense_param_dense_grad_impl<phi::dtype::bfloat16>( |
| 45 | + const DenseTensor& param, |
| 46 | + const DenseTensor& learning_rate, |
| 47 | + const DenseTensor& grad, |
| 48 | + DenseTensor* param_out) { |
| 49 | + auto p = EigenVector<phi::dtype::bfloat16>:: Flatten(param); |
| 50 | + auto g = EigenVector<phi::dtype::bfloat16>:: Flatten(grad); |
| 51 | + auto o = EigenVector<phi::dtype::bfloat16>:: Flatten(*param_out); |
| 52 | + const auto* lr = learning_rate. data<phi::dtype::bfloat16>(); |
| 53 | + |
| 54 | + o = p - lr[0] * g; |
| 55 | +} |
| 56 | + |
| 57 | +template <typename T> |
| 58 | +void sgd_dense_param_sparse_grad_impl(const DenseTensor& param, |
| 59 | + const DenseTensor& learning_rate, |
| 60 | + const SelectedRows& grad, |
| 61 | + DenseTensor* param_out) { |
| 62 | + const auto& grad_value = grad.value(); |
| 63 | + const auto& grad_rows = grad.rows(); |
| 64 | + const T* param_data = param.data<T>(); |
| 65 | + const T* grad_data = grad_value.data<T>(); |
| 66 | + const T* lr = learning_rate.data<T>(); |
| 67 | + const int64_t* rows_data = grad_rows.data(); |
| 68 | + T* out_data = param_out->data<T>(); |
| 69 | + |
| 70 | + paddle::operators::jit::sgd_attr_t attr; |
| 71 | + attr.param_height = param_out->dims()[0]; |
| 72 | + attr.param_width = param_out->numel() / attr.param_height; |
| 73 | + attr.grad_height = grad_rows.size(); // note: it is not grad->height() |
| 74 | + attr.grad_width = grad_value.numel() / attr.grad_height; |
| 75 | + attr.selected_rows_size = grad_rows.size(); |
| 76 | + |
| 77 | + auto sgd = |
| 78 | + paddle::operators::jit::KernelFuncs<paddle::operators::jit::SgdTuple<T>, |
| 79 | + phi::CPUPlace>::Cache() |
| 80 | + .At(attr); |
| 81 | + sgd(lr, param_data, grad_data, rows_data, out_data, &attr); |
| 82 | +} |
| 83 | + |
| 84 | +template <> |
| 85 | +void sgd_dense_param_sparse_grad_impl<phi::dtype::bfloat16>( |
| 86 | + const DenseTensor& param, |
| 87 | + const DenseTensor& learning_rate, |
| 88 | + const SelectedRows& grad, |
| 89 | + DenseTensor* param_out) { |
| 90 | + const auto& grad_value = grad.value(); |
| 91 | + const auto& grad_rows = grad.rows(); |
| 92 | + const auto grad_height = grad.height(); |
| 93 | + const int64_t grad_val_height = static_cast<int64_t>(grad_rows.size()); |
| 94 | + const auto grad_width = grad_value.numel() / grad_val_height; |
| 95 | + |
| 96 | + const auto* grad_data = grad_value. data<phi::dtype::bfloat16>(); |
| 97 | + auto* out_data = param_out-> data<phi::dtype::bfloat16>(); |
| 98 | + const auto* lr = learning_rate. data<phi::dtype::bfloat16>(); |
| 99 | + |
| 100 | + for (size_t i = 0; i < grad_rows.size(); ++i) { |
| 101 | + PADDLE_ENFORCE_LT( |
| 102 | + grad_rows[i], |
| 103 | + grad_height, |
| 104 | + phi::errors::OutOfRange( |
| 105 | + "Grad rows index value should be less than grad height." |
| 106 | + "Got [%s], but expected less than [%s]", |
| 107 | + grad_rows[i], |
| 108 | + grad_height)); |
| 109 | + const int64_t row = grad_rows[i]; |
| 110 | + for (int64_t j = 0; j < grad_width; ++j) { |
| 111 | + out_data[row * grad_width + j] -= lr[0] * grad_data[i * grad_width + j]; |
| 112 | + } |
| 113 | + } |
| 114 | +} |
| 115 | + |
| 116 | +template <typename T, typename Context> |
| 117 | +void SGDDenseKernel(const Context& dev_ctx, |
| 118 | + const DenseTensor& param, |
| 119 | + const DenseTensor& learning_rate, |
| 120 | + const DenseTensor& grad, |
| 121 | + paddle::optional<const DenseTensor&> master_param, |
| 122 | + bool multi_precision, |
| 123 | + DenseTensor* param_out, |
| 124 | + DenseTensor* master_param_out) { |
| 125 | + dev_ctx.template Alloc<T>(param_out); |
| 126 | + sgd_dense_param_dense_grad_impl<T>(param, learning_rate, grad, param_out); |
| 127 | +} |
| 128 | + |
| 129 | +template <typename T, typename Context> |
| 130 | +void SGDDenseParamSparseGradKernel( |
| 131 | + const Context& dev_ctx, |
| 132 | + const DenseTensor& param, |
| 133 | + const DenseTensor& learning_rate, |
| 134 | + const SelectedRows& grad, |
| 135 | + paddle::optional<const DenseTensor&> master_param, |
| 136 | + bool multi_precision, |
| 137 | + DenseTensor* param_out, |
| 138 | + DenseTensor* master_param_out) { |
| 139 | + dev_ctx.template Alloc<T>(param_out); |
| 140 | + sgd_dense_param_sparse_grad_impl<T>(param, learning_rate, grad, param_out); |
| 141 | +} |
| 142 | + |
| 143 | +template <typename T, typename Context> |
| 144 | +void SGDSparseParamSparseGradKernel( |
| 145 | + const Context& dev_ctx, |
| 146 | + const SelectedRows& param, |
| 147 | + const DenseTensor& learning_rate, |
| 148 | + const SelectedRows& grad, |
| 149 | + paddle::optional<const SelectedRows&> master_param, |
| 150 | + bool multi_precision, |
| 151 | + SelectedRows* param_out, |
| 152 | + SelectedRows* master_param_out) { |
| 153 | + // for distributed training, a sparse var may be empty, |
| 154 | + // just skip updating. |
| 155 | + if (grad.rows().size() == 0) { |
| 156 | + return; |
| 157 | + } |
| 158 | + |
| 159 | + auto param_row_width = param.value().dims()[1]; |
| 160 | + auto grad_row_width = grad.value().dims()[1]; |
| 161 | + PADDLE_ENFORCE_EQ( |
| 162 | + param_row_width, |
| 163 | + grad_row_width, |
| 164 | + phi::errors::InvalidArgument( |
| 165 | + "The param_row in SgdOP should have the same size with grad_row. " |
| 166 | + "But received param_row's width is [%s], and grad_row's width is " |
| 167 | + "[%s]", |
| 168 | + param_row_width, |
| 169 | + grad_row_width)); |
| 170 | + |
| 171 | + const auto* lr = learning_rate.data<T>(); |
| 172 | + const auto* grad_data = grad.value().data<T>(); |
| 173 | + auto* out_data = param_out->mutable_value()->data<T>(); |
| 174 | + for (size_t i = 0; i < grad.rows().size(); i++) { |
| 175 | + int64_t id_index = param_out->AutoGrownIndex(grad.rows()[i], false); |
| 176 | + PADDLE_ENFORCE_GE( |
| 177 | + id_index, |
| 178 | + static_cast<int64_t>(0), |
| 179 | + phi::errors::InvalidArgument( |
| 180 | + "The id in SgdOp should be >= 0. But recevied id_index is [%s]", |
| 181 | + id_index)); |
| 182 | + for (int64_t j = 0; j < grad_row_width; j++) { |
| 183 | + out_data[id_index * grad_row_width + j] -= |
| 184 | + lr[0] * grad_data[i * grad_row_width + j]; |
| 185 | + } |
| 186 | + } |
| 187 | +} |
| 188 | + |
| 189 | +} // namespace phi |
| 190 | + |
| 191 | +PD_REGISTER_KERNEL(sgd, |
| 192 | + CPU, |
| 193 | + ALL_LAYOUT, |
| 194 | + phi::SGDDenseKernel, |
| 195 | + phi::dtype::bfloat16, |
| 196 | + float, |
| 197 | + double) {} |
| 198 | + |
| 199 | +PD_REGISTER_KERNEL(sgd_dense_param_sparse_grad, |
| 200 | + CPU, |
| 201 | + ALL_LAYOUT, |
| 202 | + phi::SGDDenseParamSparseGradKernel, |
| 203 | + phi::dtype::bfloat16, |
| 204 | + float, |
| 205 | + double) {} |
| 206 | + |
| 207 | +PD_REGISTER_KERNEL(sgd_sparse_param_sparse_grad, |
| 208 | + CPU, |
| 209 | + ALL_LAYOUT, |
| 210 | + phi::SGDSparseParamSparseGradKernel, |
| 211 | + phi::dtype::bfloat16, |
| 212 | + float, |
| 213 | + double) {} |
0 commit comments