|
| 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/cross_entropy_grad_kernel.h" |
| 16 | + |
| 17 | +#ifdef __NVCC__ |
| 18 | +#include "cub/cub.cuh" |
| 19 | +#endif |
| 20 | +#ifdef __HIPCC__ |
| 21 | +#include <hipcub/hipcub.hpp> |
| 22 | +namespace cub = hipcub; |
| 23 | +#endif |
| 24 | + |
| 25 | +#include "kernels/gpudnn/softmax_gpudnn.h" |
| 26 | +#include "paddle/phi/backends/gpu/gpu_device_function.h" |
| 27 | +#include "paddle/phi/backends/gpu/gpu_dnn.h" |
| 28 | +#include "paddle/phi/common/amp_type_traits.h" |
| 29 | +#include "paddle/phi/core/kernel_registry.h" |
| 30 | +#include "paddle/phi/core/tensor_utils.h" |
| 31 | +#include "paddle/phi/core/visit_type.h" |
| 32 | +#include "paddle/phi/kernels/funcs/axis_utils.h" |
| 33 | +#include "paddle/phi/kernels/funcs/for_range.h" |
| 34 | +#include "paddle/phi/kernels/funcs/math_function.h" |
| 35 | +#include "paddle/phi/kernels/funcs/softmax.h" |
| 36 | + |
| 37 | +namespace phi { |
| 38 | + |
| 39 | +/* |
| 40 | + Vectorized wrapper of softmax with cross entropy grad hard label. |
| 41 | + Optimized with float4 vectorization for memory coalescing and improved |
| 42 | + throughput. |
| 43 | +*/ |
| 44 | +template <typename T, typename LabelT, typename LogitT> |
| 45 | +__global__ void SoftmaxWithCrossEntropyGradHardLabelVectorized( |
| 46 | + LogitT* __restrict__ logits_grad, |
| 47 | + const T* __restrict__ loss_grad, |
| 48 | + const T* __restrict__ softmax, |
| 49 | + const LabelT* __restrict__ labels, |
| 50 | + const int64_t n, |
| 51 | + const int64_t dim, |
| 52 | + const int64_t d, |
| 53 | + const int ignore_index) { |
| 54 | + // Vectorized load/store with float4 for 128-bit memory transactions |
| 55 | + constexpr int VEC_SIZE = 4; |
| 56 | + using VecT = typename phi::AlignedVector<LogitT, VEC_SIZE>; |
| 57 | + using SoftmaxVecT = typename phi::AlignedVector<T, VEC_SIZE>; |
| 58 | + |
| 59 | + int64_t tid = blockIdx.x * blockDim.x + threadIdx.x; |
| 60 | + int64_t vec_id = tid * VEC_SIZE; |
| 61 | + |
| 62 | + // Ensure we don't exceed bounds |
| 63 | + if (vec_id >= n * dim * d) return; |
| 64 | + |
| 65 | + // Compute indices for vectorized access |
| 66 | + int64_t idx_n = vec_id / (d * dim); |
| 67 | + int64_t idx_dim_start = (vec_id / d) % dim; |
| 68 | + int64_t idx_d = vec_id % d; |
| 69 | + int64_t ids = idx_n * d + idx_d; |
| 70 | + |
| 71 | + // Load label once per thread |
| 72 | + auto lbl = static_cast<int64_t>(labels[ids]); |
| 73 | + |
| 74 | + if (lbl == ignore_index) { |
| 75 | + // Vectorized zero fill for ignore_index |
| 76 | + VecT* vec_grad = reinterpret_cast<VecT*>(&logits_grad[vec_id]); |
| 77 | + VecT zero_vec; |
| 78 | +#pragma unroll |
| 79 | + for (int i = 0; i < VEC_SIZE; ++i) { |
| 80 | + zero_vec.val[i] = static_cast<LogitT>(0.0f); |
| 81 | + } |
| 82 | + *vec_grad = zero_vec; |
| 83 | + return; |
| 84 | + } |
| 85 | + |
| 86 | + // Vectorized load of softmax values |
| 87 | + SoftmaxVecT softmax_vec; |
| 88 | + const SoftmaxVecT* softmax_ptr = |
| 89 | + reinterpret_cast<const SoftmaxVecT*>(&softmax[vec_id]); |
| 90 | + softmax_vec = *softmax_ptr; |
| 91 | + |
| 92 | + // Load loss gradient (broadcast across vector elements) |
| 93 | + T loss_grad_val = loss_grad[ids]; |
| 94 | + |
| 95 | + // Vectorized computation |
| 96 | + VecT grad_vec; |
| 97 | +#pragma unroll |
| 98 | + for (int i = 0; i < VEC_SIZE; ++i) { |
| 99 | + int64_t current_dim = idx_dim_start + i; |
| 100 | + if (current_dim < dim) { // Bounds check for partial vectors |
| 101 | + float softmax_val = static_cast<float>(softmax_vec.val[i]); |
| 102 | + float grad_val; |
| 103 | + |
| 104 | + if (lbl == current_dim) { |
| 105 | + grad_val = (softmax_val - 1.0f) * static_cast<float>(loss_grad_val); |
| 106 | + } else { |
| 107 | + grad_val = softmax_val * static_cast<float>(loss_grad_val); |
| 108 | + } |
| 109 | + |
| 110 | + grad_vec.val[i] = static_cast<LogitT>(grad_val); |
| 111 | + } else { |
| 112 | + grad_vec.val[i] = static_cast<LogitT>(0.0f); |
| 113 | + } |
| 114 | + } |
| 115 | + |
| 116 | + // Vectorized store |
| 117 | + VecT* grad_ptr = reinterpret_cast<VecT*>(&logits_grad[vec_id]); |
| 118 | + *grad_ptr = grad_vec; |
| 119 | +} |
| 120 | + |
| 121 | +/* |
| 122 | + Specialized kernel for dimensions not divisible by vector size |
| 123 | + Uses warp-level primitives for better performance on irregular sizes |
| 124 | +*/ |
| 125 | +template <typename T, typename LabelT, typename LogitT> |
| 126 | +__global__ void SoftmaxWithCrossEntropyGradHardLabelWarp( |
| 127 | + LogitT* __restrict__ logits_grad, |
| 128 | + const T* __restrict__ loss_grad, |
| 129 | + const T* __restrict__ softmax, |
| 130 | + const LabelT* __restrict__ labels, |
| 131 | + const int64_t n, |
| 132 | + const int64_t dim, |
| 133 | + const int64_t d, |
| 134 | + const int ignore_index) { |
| 135 | + const int warps_per_block = 4; |
| 136 | + const int threads_per_warp = 32; |
| 137 | + const int threads_per_block = warps_per_block * threads_per_warp; |
| 138 | + |
| 139 | + int tid = blockIdx.x * threads_per_block + threadIdx.x; |
| 140 | + int warp_id = threadIdx.x / threads_per_warp; |
| 141 | + int lane_id = threadIdx.x % threads_per_warp; |
| 142 | + |
| 143 | + // Process multiple elements per thread using warp-level parallelism |
| 144 | + int64_t elements_per_thread = |
| 145 | + (n * dim * d + gridDim.x * threads_per_block - 1) / |
| 146 | + (gridDim.x * threads_per_block); |
| 147 | + |
| 148 | + for (int e = 0; e < elements_per_thread; ++e) { |
| 149 | + int64_t idx = tid + e * gridDim.x * threads_per_block; |
| 150 | + if (idx >= n * dim * d) break; |
| 151 | + |
| 152 | + int64_t idx_n = idx / (d * dim); |
| 153 | + int64_t idx_dim = (idx / d) % dim; |
| 154 | + int64_t idx_d = idx % d; |
| 155 | + int64_t ids = idx_n * d + idx_d; |
| 156 | + |
| 157 | + auto lbl = static_cast<int64_t>(labels[ids]); |
| 158 | + |
| 159 | + if (lbl == ignore_index) { |
| 160 | + logits_grad[idx] = static_cast<LogitT>(0.0f); |
| 161 | + } else if (lbl == idx_dim) { |
| 162 | + logits_grad[idx] = |
| 163 | + static_cast<LogitT>((static_cast<float>(softmax[idx]) - 1.0f) * |
| 164 | + static_cast<float>(loss_grad[ids])); |
| 165 | + } else { |
| 166 | + logits_grad[idx] = |
| 167 | + static_cast<LogitT>(static_cast<float>(softmax[idx]) * |
| 168 | + static_cast<float>(loss_grad[ids])); |
| 169 | + } |
| 170 | + } |
| 171 | +} |
| 172 | + |
| 173 | +/* |
| 174 | + Optimized kernel selector based on problem size and alignment |
| 175 | +*/ |
| 176 | +template <typename T, typename LabelT, typename LogitT> |
| 177 | +void LaunchOptimizedCrossEntropyGradKernel(const GPUContext& dev_ctx, |
| 178 | + LogitT* logits_grad, |
| 179 | + const T* loss_grad, |
| 180 | + const T* softmax, |
| 181 | + const LabelT* labels, |
| 182 | + const int64_t n, |
| 183 | + const int64_t dim, |
| 184 | + const int64_t d, |
| 185 | + const int ignore_index) { |
| 186 | + const int64_t total_elements = n * dim * d; |
| 187 | + auto stream = dev_ctx.stream(); |
| 188 | + |
| 189 | + // Check alignment for vectorized kernel |
| 190 | + bool is_aligned = (reinterpret_cast<uintptr_t>(logits_grad) % 16 == 0) && |
| 191 | + (reinterpret_cast<uintptr_t>(softmax) % 16 == 0) && |
| 192 | + (total_elements % 4 == 0); |
| 193 | + |
| 194 | + if (is_aligned && total_elements >= 1024) { |
| 195 | + // Use vectorized kernel for aligned, large problems |
| 196 | + constexpr int VEC_SIZE = 4; |
| 197 | + const int threads_per_block = 256; |
| 198 | + const int vec_elements = total_elements / VEC_SIZE; |
| 199 | + const int blocks = |
| 200 | + (vec_elements + threads_per_block - 1) / threads_per_block; |
| 201 | + |
| 202 | + SoftmaxWithCrossEntropyGradHardLabelVectorized<T, LabelT, LogitT> |
| 203 | + <<<blocks, threads_per_block, 0, stream>>>( |
| 204 | + logits_grad, loss_grad, softmax, labels, n, dim, d, ignore_index); |
| 205 | + } else { |
| 206 | + // Use warp-specialized kernel for irregular sizes |
| 207 | + const int warps_per_block = 4; |
| 208 | + const int threads_per_block = warps_per_block * 32; |
| 209 | + const int blocks = |
| 210 | + std::min(1024, |
| 211 | + static_cast<int>((total_elements + threads_per_block - 1) / |
| 212 | + threads_per_block)); |
| 213 | + |
| 214 | + SoftmaxWithCrossEntropyGradHardLabelWarp<T, LabelT, LogitT> |
| 215 | + <<<blocks, threads_per_block, 0, stream>>>( |
| 216 | + logits_grad, loss_grad, softmax, labels, n, dim, d, ignore_index); |
| 217 | + } |
| 218 | +} |
| 219 | + |
| 220 | +template <typename T, typename LabelT> |
| 221 | +void CrossEntropyWithSoftmaxBwdWithDowncastGPUKernel( |
| 222 | + const GPUContext& dev_ctx, |
| 223 | + const DenseTensor& label, |
| 224 | + const DenseTensor& softmax, |
| 225 | + const DenseTensor& loss_grad, |
| 226 | + int axis, |
| 227 | + DenseTensor* logits_grad) { |
| 228 | + // PADDLE_ENFORCE_EQ( |
| 229 | + // dev_ctx.GetPlace().GetType(), |
| 230 | + // phi::AllocationType::GPU, |
| 231 | + // common::errors::Unavailable("softmax_with_cross_entropy operator's " |
| 232 | + // "CUDA kernel only runs on GPU device.")); |
| 233 | + |
| 234 | + using LogitT = phi::bfloat16; |
| 235 | + const T* loss_grad_data = loss_grad.data<T>(); |
| 236 | + DenseTensor* logit_grad = logits_grad; |
| 237 | + |
| 238 | + LogitT* logit_grad_data = nullptr; |
| 239 | + logit_grad_data = dev_ctx.template Alloc<LogitT>(logit_grad); |
| 240 | + |
| 241 | + const int rank = logit_grad->dims().size(); |
| 242 | + const int axis_v = phi::funcs::CanonicalAxis(axis, rank); |
| 243 | + int axis_dim = logit_grad->dims()[axis_v]; |
| 244 | + |
| 245 | + const int64_t n = phi::funcs::SizeToAxis(axis_v, logit_grad->dims()); |
| 246 | + const int64_t d = phi::funcs::SizeFromAxis(axis_v, logit_grad->dims()); |
| 247 | + const int64_t remain = d / axis_dim; |
| 248 | + |
| 249 | + const T* softmax_data = softmax.data<T>(); |
| 250 | + const auto* label_data = label.data<LabelT>(); |
| 251 | + |
| 252 | + // Launch optimized kernel with automatic selection |
| 253 | + LaunchOptimizedCrossEntropyGradKernel<T, LabelT, LogitT>(dev_ctx, |
| 254 | + logit_grad_data, |
| 255 | + loss_grad_data, |
| 256 | + softmax_data, |
| 257 | + label_data, |
| 258 | + n, |
| 259 | + axis_dim, |
| 260 | + remain, |
| 261 | + -100); |
| 262 | +} |
| 263 | + |
| 264 | +template <typename T, typename Context> |
| 265 | +void CrossEntropyWithSoftmaxBwdWithDowncastKernel(const Context& dev_ctx, |
| 266 | + const DenseTensor& label, |
| 267 | + const DenseTensor& softmax, |
| 268 | + const DenseTensor& loss_grad, |
| 269 | + DenseTensor* logits_grad) { |
| 270 | + constexpr int axis = -1; |
| 271 | + if (logits_grad->numel() == 0) { |
| 272 | + dev_ctx. template Alloc<phi::bfloat16>(logits_grad); |
| 273 | + return; |
| 274 | + } |
| 275 | + auto dtype = label.dtype(); |
| 276 | + PD_VISIT_INTEGRAL_TYPES( |
| 277 | + dtype, "CrossEntropyWithSoftmaxBwdWithDowncastGPUKernel", ([&] { |
| 278 | + CrossEntropyWithSoftmaxBwdWithDowncastGPUKernel<T, data_t>( |
| 279 | + dev_ctx, label, softmax, loss_grad, axis, logits_grad); |
| 280 | + })); |
| 281 | +} |
| 282 | + |
| 283 | +} // namespace phi |
| 284 | + |
| 285 | +PD_REGISTER_PLUGIN_KERNEL(cross_entropy_with_softmax_bwd_w_downcast, |
| 286 | + metax_gpu, |
| 287 | + ALL_LAYOUT, |
| 288 | + phi::CrossEntropyWithSoftmaxBwdWithDowncastKernel, |
| 289 | + float, |
| 290 | + double, |
| 291 | + phi::float16) {} |
0 commit comments