|
| 1 | +// Copyright (c) 2025 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/median_grad_kernel.h" |
| 16 | + |
| 17 | +#include <math.h> |
| 18 | +#include "paddle/phi/backends/cpu/cpu_context.h" |
| 19 | +#include "paddle/phi/core/kernel_registry.h" |
| 20 | +#include "paddle/phi/kernels/funcs/math_function.h" |
| 21 | +#include "paddle/phi/kernels/funcs/nanmedian_utils.h" |
| 22 | + |
| 23 | +namespace phi { |
| 24 | + |
| 25 | +template <typename T> |
| 26 | +void CalcMedianMinGrad(int64_t pre_dim, |
| 27 | + int64_t stride, |
| 28 | + const int64_t* m_data, |
| 29 | + T* dx_data, |
| 30 | + const T* dout_data) { |
| 31 | + int64_t i = 0; |
| 32 | + int64_t offset = 0; |
| 33 | + for (i = 0; i < pre_dim; i++) { |
| 34 | + if (m_data[i] >= 0) { |
| 35 | + dx_data[offset + m_data[i]] = dout_data[i]; |
| 36 | + } |
| 37 | + offset += stride; |
| 38 | + } |
| 39 | +} |
| 40 | + |
| 41 | +template <typename T> |
| 42 | +void CalcMedianGradEvenly(int64_t pre_dim, |
| 43 | + int64_t stride, |
| 44 | + const DenseTensor& x, |
| 45 | + const T* m_data, |
| 46 | + const int64_t* m_index, |
| 47 | + T* dx_data, |
| 48 | + const T* dout_data) { |
| 49 | + int64_t i = 0, j = 0; |
| 50 | + int64_t offset = 0; |
| 51 | + std::vector<int64_t> data_index; |
| 52 | + const T* x_data = x.data<T>(); |
| 53 | + for (i = 0; i < pre_dim; i++) { |
| 54 | + data_index.clear(); |
| 55 | + for (j = 0; j < stride; j++) { |
| 56 | + if ((m_data[i] == x_data[offset + j]) || |
| 57 | + (isnan(static_cast<float>(m_data[i])) && |
| 58 | + isnan(static_cast<float>(x_data[offset + j])))) { |
| 59 | + data_index.push_back(offset + j); |
| 60 | + } |
| 61 | + } |
| 62 | + if (data_index.size() == 0) { |
| 63 | + if (m_index[2 * i] == m_index[2 * i + 1]) { |
| 64 | + dx_data[offset + m_index[2 * i]] = dout_data[i]; |
| 65 | + } else { |
| 66 | + dx_data[offset + m_index[2 * i]] = dout_data[i] / static_cast<T>(2.0); |
| 67 | + dx_data[offset + m_index[2 * i + 1]] = |
| 68 | + dout_data[i] / static_cast<T>(2.0); |
| 69 | + } |
| 70 | + } else { |
| 71 | + for (j = 0; j < data_index.size(); j++) { |
| 72 | + dx_data[data_index[j]] = |
| 73 | + dout_data[i] / static_cast<T>(data_index.size()); |
| 74 | + } |
| 75 | + } |
| 76 | + |
| 77 | + offset += stride; |
| 78 | + } |
| 79 | +} |
| 80 | + |
| 81 | +template <typename T, typename Context> |
| 82 | +void CalcMedianGradKernel_CPU(const Context& dev_ctx, |
| 83 | + const DenseTensor& x, |
| 84 | + const DenseTensor& median_data, |
| 85 | + const DenseTensor& median_index, |
| 86 | + const DenseTensor& out_grad, |
| 87 | + const std::string& mode, |
| 88 | + const bool evenly, |
| 89 | + DenseTensor* x_grad) { |
| 90 | + T* dx_data = dev_ctx.template Alloc<T>(x_grad); |
| 91 | + if (!dx_data) return; |
| 92 | + |
| 93 | + phi::funcs::SetConstant<Context, T> set_zero; |
| 94 | + set_zero(dev_ctx, x_grad, static_cast<T>(0)); |
| 95 | + |
| 96 | + const int64_t* m_index = median_index.data<int64_t>(); |
| 97 | + const T* m_data = median_data.data<T>(); |
| 98 | + const T* dout_data = out_grad.data<T>(); |
| 99 | + int64_t numel = x.numel(); |
| 100 | + auto x_dim = x.dims(); |
| 101 | + int64_t rank = x_dim.size(); |
| 102 | + int64_t stride = x_dim[static_cast<int>(rank - 1)]; |
| 103 | + int64_t pre_dim = numel / stride; |
| 104 | + if (!evenly) { |
| 105 | + CalcMedianMinGrad(pre_dim, stride, m_index, dx_data, dout_data); |
| 106 | + } else { |
| 107 | + CalcMedianGradEvenly( |
| 108 | + pre_dim, stride, x, m_data, m_index, dx_data, dout_data); |
| 109 | + } |
| 110 | +} |
| 111 | + |
| 112 | +template <typename T, typename Context> |
| 113 | +void MedianGradKernel(const Context& dev_ctx, |
| 114 | + const DenseTensor& x, |
| 115 | + const DenseTensor& median_data, |
| 116 | + const DenseTensor& median_index, |
| 117 | + const DenseTensor& out_grad, |
| 118 | + const IntArray& axes, |
| 119 | + bool keepdim UNUSED, |
| 120 | + const std::string& mode, |
| 121 | + DenseTensor* x_grad) { |
| 122 | + if (x_grad && x_grad->numel() == 0) { |
| 123 | + dev_ctx.template Alloc<T>(x_grad); |
| 124 | + return; |
| 125 | + } |
| 126 | + bool evenly = (axes.size() != 1 || mode == "avg"); |
| 127 | + DenseTensor tmp_x; |
| 128 | + auto rank = x.dims().size(); |
| 129 | + if ((axes.size() == 0) || rank <= 1) { |
| 130 | + tmp_x = x; |
| 131 | + tmp_x.Resize({x.numel()}); |
| 132 | + CalcMedianGradKernel_CPU<T, Context>(dev_ctx, |
| 133 | + tmp_x, |
| 134 | + median_data, |
| 135 | + median_index, |
| 136 | + out_grad, |
| 137 | + mode, |
| 138 | + evenly, |
| 139 | + x_grad); |
| 140 | + } else { |
| 141 | + funcs::PreprocessMedianKernel<T, Context>(dev_ctx, x, axes, &tmp_x); |
| 142 | + |
| 143 | + DenseTensor tmp_x_grad; |
| 144 | + tmp_x_grad.Resize(x_grad->dims()); |
| 145 | + CalcMedianGradKernel_CPU<T, Context>(dev_ctx, |
| 146 | + tmp_x, |
| 147 | + median_data, |
| 148 | + median_index, |
| 149 | + out_grad, |
| 150 | + mode, |
| 151 | + evenly, |
| 152 | + &tmp_x_grad); |
| 153 | + |
| 154 | + dev_ctx.template Alloc<T>(x_grad); |
| 155 | + funcs::PostprocessMedianGradKernel<T, Context>( |
| 156 | + dev_ctx, &tmp_x_grad, axes, x_grad); |
| 157 | + } |
| 158 | +} |
| 159 | + |
| 160 | +} // namespace phi |
| 161 | + |
| 162 | +PD_REGISTER_KERNEL(median_grad, |
| 163 | + CPU, |
| 164 | + ALL_LAYOUT, |
| 165 | + phi::MedianGradKernel, |
| 166 | + float, |
| 167 | + double, |
| 168 | + int, |
| 169 | + int64_t) {} |
0 commit comments