cross_entropy_kernel.cu 55.5 KB
Newer Older
1 2
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.

L
Luo Tao 已提交
3 4 5
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
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
8

C
caoying03 已提交
9 10 11 12 13
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. */
14 15 16

#include "paddle/phi/kernels/cross_entropy_kernel.h"

17 18 19 20 21 22 23
#ifdef __NVCC__
#include "cub/cub.cuh"
#endif
#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
namespace cub = hipcub;
#endif
24

25 26 27 28
#include "paddle/fluid/operators/math/cross_entropy.h"
#include "paddle/fluid/operators/math/softmax.h"
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
#include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
29 30
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/core/kernel_registry.h"
31
#include "paddle/phi/core/tensor_utils.h"
32
#include "paddle/phi/core/visit_type.h"
33 34 35 36 37 38
#include "paddle/phi/kernels/funcs/axis_utils.h"
#include "paddle/phi/kernels/funcs/for_range.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/gpudnn/softmax_gpudnn.h"

namespace phi {
C
caoying03 已提交
39

40 41
#define ALIGN_BYTES 16

42
enum class SoftmaxMode { kSoftmax, kLogSoftmax, kCrossEntropy };
C
caoying03 已提交
43

44
// Wrapper of log function. Use log(float32) for float16
45
template <typename T>
46
static __device__ __forceinline__ T Log(T x) {
47
  using AccT = typename dtype::MPTypeTrait<T>::Type;
48
  AccT logx = std::log(static_cast<AccT>(x));
49
  return paddle::operators::math::TolerableValue<T>()(static_cast<T>(logx));
50 51 52 53 54
}

// Wrapper of exp function. Use exp(float32) for float16
template <typename T>
static __device__ __forceinline__ T Exp(T x) {
55
  using AccT = typename dtype::MPTypeTrait<T>::Type;
56
  AccT expx = std::exp(static_cast<AccT>(x));
57
  return paddle::operators::math::TolerableValue<T>()(static_cast<T>(expx));
58 59
}

60 61 62 63 64 65 66 67 68 69 70 71
template <typename Tx, typename Ty = Tx>
struct ExpAddFunctor {
  HOSTDEVICE inline ExpAddFunctor(Tx max) : max(max) {}

  HOSTDEVICE inline Ty operator()(const Tx& sum, const Tx& x) const {
    return static_cast<Ty>(sum + std::exp(x - max));
  }

 private:
  Tx max;
};

72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
/*
  Cross entropy soft label with dynamic size on axis (log2_elements is
  varibale).
  - if the input is softmax,compute loss with softmax
  - if the input is log_softmax, compute loss with log_softmax and update
  softmax
*/
template <typename T, typename VecT, bool InLogMode = false>
__global__ void CrossEntropySoftLabel(T* loss,
                                      T* softmaxwrt,
                                      const T* softmax,
                                      const T* labels,
                                      const int n,
                                      const int dim,
                                      const int d,
                                      int log2_elements) {
  const int kDimCeil = 1 << log2_elements;
  const int kVSize = sizeof(VecT) / sizeof(T);
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
#ifdef __HIPCC__
  const int kThreadPerBlock = 256;
#else
  const int kThreadPerBlock = 512;
#endif
  const int kBatchPerBlock = 1;
  const int kWarpSize = 32;  // (dim < 32) ? dim : 32;
  const int kBatchSize = 1;
  const int kThreadPerBatch = kThreadPerBlock / kBatchPerBlock;
  const int kWarpPerBatch = kThreadPerBatch / kWarpSize;

  const int kIterations = (dim + kThreadPerBatch - 1) / kThreadPerBatch;
  const int kIterationsV = (kIterations >= kVSize) ? (kIterations / kVSize) : 1;

  const int first_batch = (blockDim.y * blockIdx.x + threadIdx.y) * kBatchSize;

  T sum[kBatchSize]{static_cast<T>(0.0)};
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    int ids = first_batch + i;
    if (ids >= n * d) break;
    int idx_n = ids / d;
    int idx_d = ids % d;
#pragma unroll
    for (int it = 0; it < kIterations; ++it) {
      int idx_dim = it * kThreadPerBatch + threadIdx.x;
      int idx = idx_n * dim * d + idx_dim * d + idx_d;

      if (idx_n < n && idx_dim < dim) {
        VecT softmaxdata;
        if (InLogMode) {
          softmaxdata = reinterpret_cast<VecT*>(&softmaxwrt[idx])[0];
        } else {
          softmaxdata = reinterpret_cast<const VecT*>(&softmax[idx])[0];
        }
        VecT labelsdata = reinterpret_cast<const VecT*>(&labels[idx])[0];
        T* softmaxptr = reinterpret_cast<T*>(&softmaxdata);
        T* labelsptr = reinterpret_cast<T*>(&labelsdata);
#pragma unroll
        for (int s = 0; s < kVSize; s++) {
          if (InLogMode) {
            sum[i] -= softmaxptr[s] * labelsptr[s];
            softmaxptr[s] = Exp(softmaxptr[s]);
          } else {
            sum[i] -= Log(softmaxptr[s]) * labelsptr[s];
          }
        }
        if (InLogMode) {
          reinterpret_cast<VecT*>(&softmaxwrt[idx])[0] = softmaxdata;
        }
      }
    }
  }
  phi::WarpReduceSum<T, kBatchSize, kWarpSize>(sum);
  __syncthreads();

  __shared__ T sumshare[kWarpPerBatch][kBatchPerBlock][kBatchSize];
  if (threadIdx.x % kWarpSize == 0) {
#pragma unroll
    for (int i = 0; i < kBatchSize; i++) {
      sumshare[threadIdx.x / kWarpSize][threadIdx.y][i] = sum[i];
    }
  }
  __syncthreads();

  // write
  if (threadIdx.x == 0) {
    for (int i = 0; i < kBatchSize; i++) {
      int ids = first_batch + i;
      if (ids < n * d) {
        loss[ids] = sumshare[0][threadIdx.y][i];
        for (int s = 1; s < kWarpPerBatch; s++) {
          loss[ids] += sumshare[s][threadIdx.y][i];
        }
      }
    }
  }
}
169 170 171 172

/*
  Hard label cross entropy.
*/
173
template <typename T, typename LabelT, bool IgnoreIndex>
174 175 176 177 178 179
__global__ void CrossEntropyHardLabel(T* loss,
                                      const T* softmax,
                                      const LabelT* labels,
                                      const int n,
                                      const int dim,
                                      const int d,
180 181 182 183 184 185 186
                                      const int ignore_idx) {
  int64_t ids = blockIdx.x * blockDim.x + threadIdx.x;
  int64_t idx_n = ids / d;
  int64_t idx_d = ids % d;

  // thread ids compute loss[ids] using softmax[idx]
  if (ids < n * d) {
187
    auto lbl = static_cast<int64_t>(labels[ids]);
188 189
    assert(lbl >= 0 && lbl < dim || lbl == ignore_idx);
    if (lbl < 0 || lbl >= dim) {  // label is out of bound
190
      loss[ids] = static_cast<T>(0.0);
191
    } else {
192
      int64_t idx = idx_n * dim * d + lbl * d + idx_d;
193 194
      if (IgnoreIndex == true) {
        // IgnoreIndex is true
195
        if (lbl == ignore_idx) {
196 197 198 199
          loss[ids] = static_cast<T>(0.0);
        } else {
          loss[ids] = -Log(softmax[idx]);
        }
200
      } else {
201
        // IgnoreIndex is false
202 203 204 205 206 207 208 209 210 211 212
        loss[ids] = -Log(softmax[idx]);
      }
    }
  }
}

/*
  Hard label cross entropy with exp.
  Input: log softmax
  Output: loss and exp(input)
*/
213
template <typename T, typename LabelT, bool IgnoreIndex>
214 215 216 217 218 219
__global__ void CrossEntropyExpHardLabel(T* loss,
                                         T* softmax,
                                         const LabelT* labels,
                                         const int n,
                                         const int dim,
                                         const int d,
220 221 222 223 224 225 226 227
                                         const int ignore_idx) {
  int64_t idx = blockIdx.x * blockDim.x + threadIdx.x;
  int64_t idx_n = idx / (d * dim);
  int64_t idx_dim = (idx / d) % dim;
  int64_t idx_d = idx % d;
  int64_t ids = idx_n * d + idx_d;

  if (idx < n * dim * d) {
228
    auto lbl = static_cast<int64_t>(labels[ids]);
229
    assert(lbl >= 0 && lbl < dim || lbl == ignore_idx);
230 231
    if (IgnoreIndex == true) {
      // IgnoreIndex is true
232 233
      if (idx_dim == lbl) {
        if (lbl == ignore_idx) {
234 235 236 237 238 239 240
          loss[ids] = static_cast<T>(0.0);
        } else {
          loss[ids] = -softmax[idx];
        }
      }
    } else {
      // IgnoreIndex is false
241 242
      if (lbl >= 0 && lbl < dim) {
        if (lbl == idx_dim) {
243 244 245 246 247
          loss[ids] = -softmax[idx];
        }
      } else {
        loss[ids] = static_cast<T>(0.0);
      }
248
    }
249
    softmax[idx] = Exp(softmax[idx]);
250 251 252
  }
}

253 254 255 256 257 258 259 260 261
template <typename T, typename AccT, int VecSize, class ReduceFunctor>
__device__ __forceinline__ AccT ThreadReduce(const T* input,
                                             int size,
                                             const int offset,
                                             AccT init,
                                             ReduceFunctor reducer) {
  using VecT = kps::details::VectorType<T, VecSize>;
  int tid = threadIdx.x;
  AccT val = init;
262

263 264 265 266 267 268 269 270
  if (offset > 0) {
    input -= offset;
    size += offset;
    if (tid >= offset) {
      val = reducer(val, input[tid]);
    }
    size -= blockDim.x;
    input += blockDim.x;
271
  }
272
  int remain = size % (VecSize * blockDim.x);
273

274 275
  T ins[VecSize];
  VecT* ins_vec = reinterpret_cast<VecT*>(&ins);
276

277 278 279
  // vector part
  for (; VecSize * tid < (size - remain); tid += blockDim.x) {
    *ins_vec = reinterpret_cast<const VecT*>(input)[tid];
280 281

#pragma unroll
282 283
    for (int i = 0; i < VecSize; ++i) {
      val = reducer(val, ins[i]);
284
    }
285
  }
286

287 288 289 290
  // scalar part
  tid = size - remain + threadIdx.x;
  for (; tid < size; tid += blockDim.x) {
    val = reducer(val, input[tid]);
291
  }
292 293
  return val;
}
294

295 296 297 298 299 300 301 302 303 304 305 306 307 308
template <typename T, bool IgnoreIndex>
__device__ __forceinline__ void ComputeLoss(T* loss,
                                            const T loss_value,
                                            const int label_id,
                                            const int64_t label_value,
                                            const int tid,
                                            const int vec_size,
                                            const int offset,
                                            const int ignore_index) {
  int loss_id = vec_size * tid + offset;
  if (IgnoreIndex) {
    if (label_value == loss_id) {
      if (label_value == ignore_index) {
        loss[label_id] = static_cast<T>(0.0f);
309
      } else {
310
        loss[label_id] = loss_value;
311 312 313 314 315 316 317 318 319
      }
    }
  } else {
    if (label_value == loss_id) {
      loss[label_id] = loss_value;
    }
  }
}

320 321 322 323
template <typename T,
          typename AccT,
          typename LabelT,
          int VecSize,
324 325
          bool IgnoreIndex>
__device__ __forceinline__ void VectorizedSoftmaxForwardImpl(
326 327 328 329 330 331 332
    T* loss,
    T* softmax,
    const T* logits,
    const LabelT* label,
    int size,
    const int offset,
    const phi::LogSoftmaxForwardFunctor<AccT>& func,
333 334 335 336 337
    const int ignore_index) {
  using VecT = kps::details::VectorType<T, VecSize>;
  int tid = threadIdx.x;
  int label_id = blockIdx.x;
  auto label_value = static_cast<int64_t>(label[label_id]);
338
  assert(label_value >= 0 && label_value < size || label_value == ignore_index);
339 340 341 342 343 344 345 346 347 348 349 350 351
  const bool label_valid = label_value >= 0 && label_value < size;
  int loss_id_offset = 0;

  if (offset > 0) {
    logits -= offset;
    softmax -= offset;
    size += offset;
    loss_id_offset -= offset;
    if (tid >= offset) {
      AccT log_softmax = func(static_cast<AccT>(logits[tid]));
      softmax[tid] = static_cast<T>(std::exp(log_softmax));
      // loss
      if (label_valid) {
352 353 354 355 356 357 358 359
        ComputeLoss<T, IgnoreIndex>(loss,
                                    static_cast<T>(-log_softmax),
                                    label_id,
                                    label_value,
                                    tid,
                                    1,
                                    loss_id_offset,
                                    ignore_index);
360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
      }
    }
    size -= blockDim.x;
    logits += blockDim.x;
    softmax += blockDim.x;
    loss_id_offset += blockDim.x;
  }
  int remain = size % (VecSize * blockDim.x);

  T ins[VecSize];
  T outs[VecSize];
  VecT* ins_vec = reinterpret_cast<VecT*>(&ins);
  VecT* outs_vec = reinterpret_cast<VecT*>(&outs);

  // vector part
  for (; VecSize * tid < (size - remain); tid += blockDim.x) {
    // read
    *ins_vec = reinterpret_cast<const VecT*>(logits)[tid];

#pragma unroll
    // compute
    for (int i = 0; i < VecSize; ++i) {
      AccT log_softmax = func(static_cast<AccT>(ins[i]));
      outs[i] = static_cast<T>(std::exp(log_softmax));

      // loss
      if (label_valid) {
387 388 389 390 391 392 393 394
        ComputeLoss<T, IgnoreIndex>(loss,
                                    static_cast<T>(-log_softmax),
                                    label_id,
                                    label_value,
                                    tid,
                                    VecSize,
                                    loss_id_offset + i,
                                    ignore_index);
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409
      }
    }

    // write
    reinterpret_cast<VecT*>(softmax)[tid] = *outs_vec;
  }

  // scalar part
  tid = size - remain + threadIdx.x;
  for (; tid < size; tid += blockDim.x) {
    AccT log_softmax = func(static_cast<AccT>(logits[tid]));
    softmax[tid] = static_cast<T>(std::exp(log_softmax));

    // loss
    if (label_valid) {
410 411 412 413 414 415 416
      ComputeLoss<T, IgnoreIndex>(loss,
                                  static_cast<T>(-log_softmax),
                                  label_id,
                                  label_value,
                                  tid,
                                  1,
                                  loss_id_offset,
417 418 419 420 421 422 423 424 425 426
                                  ignore_index);
    }
  }

  // invalid label, write once
  if (!label_valid && threadIdx.x == 0) {
    loss[label_id] = static_cast<T>(0.0f);
  }
}

427 428 429 430
template <typename T,
          typename AccT,
          typename LabelT,
          int VecSize,
431 432
          bool IgnoreIndex>
__device__ __forceinline__ void ScalarSoftmaxForwardImpl(
433 434 435 436 437 438 439
    T* loss,
    T* softmax,
    const T* logits,
    const LabelT* label,
    const int size,
    const phi::LogSoftmaxForwardFunctor<AccT>& func,
    const int ignore_index) {
440 441 442 443
  int tid = threadIdx.x;
  int remain = size % (VecSize * blockDim.x);
  int label_id = blockIdx.x;
  auto label_value = static_cast<int64_t>(label[label_id]);
444
  assert(label_value >= 0 && label_value < size || label_value == ignore_index);
445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
  const bool label_valid = label_value >= 0 && label_value < size;

  // main part
  for (; tid < (size - remain); tid += VecSize * blockDim.x) {
    T ins[VecSize];

#pragma unroll
    for (int i = 0; i < VecSize; ++i) {
      ins[i] = logits[tid + i * blockDim.x];
    }
#pragma unroll
    for (int i = 0; i < VecSize; ++i) {
      AccT log_softmax = func(static_cast<AccT>(ins[i]));
      softmax[tid + i * blockDim.x] = static_cast<T>(std::exp(log_softmax));
      // loss
      if (label_valid) {
461 462 463 464 465 466 467
        ComputeLoss<T, IgnoreIndex>(loss,
                                    static_cast<T>(-log_softmax),
                                    label_id,
                                    label_value,
                                    tid,
                                    VecSize,
                                    i,
468 469 470 471 472 473 474 475 476 477 478
                                    ignore_index);
      }
    }
  }

  // tail part
  for (; tid < size; tid += blockDim.x) {
    AccT log_softmax = func(static_cast<AccT>(logits[tid]));
    softmax[tid] = static_cast<T>(std::exp(log_softmax));
    // loss
    if (label_valid) {
479 480 481 482 483 484 485 486
      ComputeLoss<T, IgnoreIndex>(loss,
                                  static_cast<T>(-log_softmax),
                                  label_id,
                                  label_value,
                                  tid,
                                  1,
                                  0,
                                  ignore_index);
487 488 489 490 491 492 493 494 495
    }
  }

  // invalid label, write once
  if (!label_valid && threadIdx.x == 0) {
    loss[label_id] = static_cast<T>(0.0f);
  }
}

496 497 498 499
template <typename T,
          typename AccT,
          typename LabelT,
          int VecSize,
500
          bool IgnoreIndex>
501 502 503
__global__ void VectorizedSoftmaxForward(T* loss,
                                         T* softmax,
                                         const T* logits,
504
                                         const LabelT* label,
505 506
                                         const int high_dim,
                                         const int mid_dim,
507 508 509 510 511 512 513 514 515 516 517 518
                                         const int ignore_index) {
  using VecT = kps::details::VectorType<T, VecSize>;

  // each block deal with one batch
  logits += blockIdx.x * mid_dim;
  softmax += blockIdx.x * mid_dim;

  const int input_offset = ((uint64_t)logits) % ALIGN_BYTES / sizeof(T);
  const int output_offset = ((uint64_t)softmax) % ALIGN_BYTES / sizeof(T);

  // 1. reduce max
  AccT max = ThreadReduce<T, AccT, VecSize, kps::MaxFunctor<AccT>>(
519 520 521 522
      logits,
      mid_dim,
      input_offset,
      -std::numeric_limits<AccT>::infinity(),
523 524 525 526 527 528
      kps::MaxFunctor<AccT>());
  max = kps::details::BlockXReduce<AccT, kps::MaxFunctor<AccT>>(
      max, kps::MaxFunctor<AccT>());

  // 2. reduce sum
  AccT sum = ThreadReduce<T, AccT, VecSize, ExpAddFunctor<AccT>>(
529 530 531 532
      logits,
      mid_dim,
      input_offset,
      static_cast<AccT>(0),
533 534 535 536 537
      ExpAddFunctor<AccT>(max));
  sum = kps::details::BlockXReduce<AccT, kps::AddFunctor<AccT>>(
      sum, kps::AddFunctor<AccT>());

  // 3. softmax
538
  phi::LogSoftmaxForwardFunctor<AccT> func(max, sum);
539 540
  if (input_offset == output_offset) {
    VectorizedSoftmaxForwardImpl<T, AccT, LabelT, VecSize, IgnoreIndex>(
541 542 543 544 545 546 547
        loss,
        softmax,
        logits,
        label,
        mid_dim,
        input_offset,
        func,
548 549 550 551 552 553 554
        ignore_index);
  } else {
    ScalarSoftmaxForwardImpl<T, AccT, LabelT, VecSize, IgnoreIndex>(
        loss, softmax, logits, label, mid_dim, func, ignore_index);
  }
}

555
/*
556 557 558 559 560 561 562 563 564
Core function of softmax with cross entropy forward soft label.
The computation includes
  - Compute maximum of batch: maxvalue_{i} = max_j src_{i,j}
  - Compute sum of exp batch: s_{i} = sum_{j}{ exp(src_{i,j} - maxvalue_{i} }
  - Compute: sum of - sum_{j}{ label_{i,j} * (src_{i,j} - maxvalue_{i} -
log(sum[i]))}
One warp (32 threads) is used to compute 1 or 2 batch (kBatchSize).
For reduction max (sum), firstly compute max (sum) to one warp, then use shuffle
api to compute max (sum) in one warp.
565
*/
566 567 568 569 570 571 572 573 574
template <typename T, typename VecT, typename AccT, int Log2Elements>
__global__ void WarpSoftmaxForwardSoftLabel(T* loss,
                                            T* softmax,
                                            const T* src,
                                            const T* label,
                                            const int batch_size,
                                            const int stride,
                                            const int element_count) {
  const bool LogMode = true;
575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638

  constexpr int kDimCeil = 1 << Log2Elements;
  constexpr int kWarpSize = (kDimCeil < 32) ? kDimCeil : 32;
  constexpr int kVSize = sizeof(VecT) / sizeof(T);
  constexpr int kIterations = kDimCeil / kWarpSize;
  constexpr int kIterationsV =
      (kIterations >= kVSize) ? (kIterations / kVSize) : 1;
  constexpr int kBatchSize = (kDimCeil <= 128) ? 2 : 1;

  int first_batch = (blockDim.y * blockIdx.x + threadIdx.y) * kBatchSize;
  int local_batches = batch_size - first_batch;
  if (local_batches > kBatchSize) {
    local_batches = kBatchSize;
  }

  // read data from global memory
  VecT srcdata[kBatchSize][kIterationsV];
  VecT labeldata[kBatchSize][kIterationsV];

  for (int i = 0; i < kBatchSize; ++i) {
    const VecT* src_v =
        reinterpret_cast<const VecT*>(&src[(first_batch + i) * stride]);
    const VecT* label_v =
        reinterpret_cast<const VecT*>(&label[(first_batch + i) * stride]);

    // max index to read
    int idx_max = (i < local_batches) ? element_count : 0;
    int idx_max_v = idx_max / kVSize;

    // read data
    for (int it = 0; it < kIterationsV; ++it) {
      int src_idx = threadIdx.x + it * kWarpSize;
      if (src_idx < idx_max_v) {
        srcdata[i][it] = src_v[src_idx];
        labeldata[i][it] = label_v[src_idx];
      } else {
#pragma unroll
        for (int s = 0; s < kVSize; s++) {
          reinterpret_cast<T*>(&srcdata[i][it])[s] =
              -std::numeric_limits<AccT>::max();
          reinterpret_cast<T*>(&labeldata[i][it])[s] = 0.0;
        }
      }
    }
  }

  // compute max value
  AccT max_value[kBatchSize];
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    max_value[i] = -std::numeric_limits<AccT>::infinity();
#pragma unroll
    for (int it = 0; it < kIterationsV; ++it) {
      T* srcptr_v = reinterpret_cast<T*>(&srcdata[i][it]);
      T valmax = srcptr_v[0];
#pragma unroll
      for (int s = 1; s < kVSize; ++s) {
        valmax = (valmax > srcptr_v[s]) ? valmax : srcptr_v[s];
      }
      max_value[i] = (max_value[i] > static_cast<AccT>(valmax))
                         ? max_value[i]
                         : static_cast<AccT>(valmax);
    }
  }
639
  phi::WarpReduceMax<AccT, kBatchSize, kWarpSize>(max_value);
640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658

  // compute sum
  AccT sum[kBatchSize]{0.0};
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
#pragma unroll
    for (int it = 0; it < kIterationsV; ++it) {
      T* srcptr_v = reinterpret_cast<T*>(&srcdata[i][it]);
#pragma unroll
      for (int s = 0; s < kVSize; ++s) {
        if (LogMode) {
          sum[i] += std::exp(static_cast<AccT>(srcptr_v[s]) - max_value[i]);
        } else {
          srcptr_v[s] = std::exp(static_cast<AccT>(srcptr_v[s]) - max_value[i]);
          sum[i] += static_cast<AccT>(srcptr_v[s]);
        }
      }
    }
  }
659
  phi::WarpReduceSum<AccT, kBatchSize, kWarpSize>(sum);
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701

  // log_softmax and loss
  AccT sumloss[kBatchSize]{0.0};
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    if (i >= local_batches) break;

    VecT* softmax_v =
        reinterpret_cast<VecT*>(&softmax[(first_batch + i) * stride]);

    // max index to write
    int idx_max = (i < local_batches) ? element_count : 0;
    int idx_max_v = idx_max / kVSize;

    if (LogMode) {
      sum[i] = std::log(sum[i]);
    }
#pragma unroll
    for (int it = 0; it < kIterationsV; ++it) {
      T* srcvp = reinterpret_cast<T*>(&srcdata[i][it]);
      T* labelvp = reinterpret_cast<T*>(&labeldata[i][it]);
      VecT tmpv;
      T* tmpvp = reinterpret_cast<T*>(&tmpv);
#pragma unroll
      for (int s = 0; s < kVSize; ++s) {
        if (LogMode) {
          AccT logsoftmax = static_cast<AccT>(srcvp[s]) - max_value[i] - sum[i];
          sumloss[i] -= logsoftmax * static_cast<AccT>(labelvp[s]);
          tmpvp[s] = std::exp(logsoftmax);
        } else {
          tmpvp[s] = static_cast<AccT>(srcvp[s]) / sum[i];
        }
      }

      int idx = threadIdx.x + it * kWarpSize;
      if (idx < idx_max_v) {
        softmax_v[idx] = tmpv;
      }
    }
  }

  // loss
702
  phi::WarpReduceSum<AccT, kBatchSize, kWarpSize>(sumloss);
703 704 705 706 707 708 709

  for (int i = 0; i < kBatchSize; i++) {
    if (i >= local_batches) break;
    loss[first_batch + i] = sumloss[i];
  }
}

710 711 712 713 714
#define SOFTMAX_WARP_FORWARD_SOFT_CASE(Log2Elements, VecT, AccT)           \
  case Log2Elements:                                                       \
    WarpSoftmaxForwardSoftLabel<T, VecT, AccT, Log2Elements>               \
        <<<blocks, threads, 0, stream>>>(                                  \
            loss, softmax, src, label, batch_size, stride, element_count); \
715 716 717 718 719 720
    break;

/*
  Wrapper of softmax with cross entropy forward soft label.
*/
template <typename T>
721 722 723 724 725 726 727 728 729
void SwitchWarpSoftmaxForwardSoftLabel(const int blocks,
                                       const dim3 threads,
                                       gpuStream_t stream,
                                       T* loss,
                                       T* softmax,
                                       const T* src,
                                       const T* label,
                                       const int batch_size,
                                       const int stride,
730 731
                                       const int element_count,
                                       const int log2_elements) {
732
  using AccT = typename dtype::MPTypeTrait<T>::Type;
733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749
  switch (log2_elements) {
    SOFTMAX_WARP_FORWARD_SOFT_CASE(0, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(1, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(2, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(3, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(4, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(5, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(6, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(7, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(8, T, AccT);
    SOFTMAX_WARP_FORWARD_SOFT_CASE(9, T, AccT);
    default:
      break;
  }
}

template <typename T>
750 751 752 753 754 755 756 757 758 759
static void SoftmaxWithCrossEntropySoftLabel(const GPUContext& dev_ctx,
                                             const int rank,
                                             const int axis,
                                             const T* logits_data,
                                             const T* labels_data,
                                             T* softmax_data,
                                             T* loss_data,
                                             int N,
                                             int dim,
                                             int D) {
760 761 762 763 764 765 766 767 768 769 770 771 772 773
#ifdef __HIPCC__
  constexpr int kMaxBlockDim = 256;
#else
  constexpr int kMaxBlockDim = 512;
#endif
  int64_t block_dim = dim >= kMaxBlockDim
                          ? kMaxBlockDim
                          : (1 << static_cast<int>(std::log2(dim)));

  int64_t grid_dim = N * D;
  constexpr int max_dim = 320;

  const int kDimLog2 = static_cast<int>(Log2Ceil(dim));
  const int kDimCeil = 1 << kDimLog2;
774
  auto stream = dev_ctx.stream();
775 776 777 778 779 780 781 782 783 784 785 786

  if (D == 1 && dim <= max_dim) {
    int kWarpSize = (kDimCeil < 32) ? kDimCeil : 32;
    int batches_per_warp = (kDimCeil <= 128) ? 2 : 1;

    // use 128 threads per block to maximimize gpu utilization
    constexpr int threads_per_block = 128;
    int warps_per_block = (threads_per_block / kWarpSize);
    int batches_per_block = warps_per_block * batches_per_warp;
    int blocks = (N + batches_per_block - 1) / batches_per_block;
    dim3 threads(kWarpSize, warps_per_block, 1);

787 788 789 790 791 792 793 794 795 796 797
    SwitchWarpSoftmaxForwardSoftLabel<T>(blocks,
                                         threads,
                                         stream,
                                         loss_data,
                                         softmax_data,
                                         logits_data,
                                         labels_data,
                                         N,
                                         dim,
                                         dim,
                                         kDimLog2);
798 799 800 801

  } else {
    ScopedTensorDescriptor desc;
    std::vector<int> tensor_dims = {N, dim, D, 1};
802
    GPUDNNDataLayout layout = GPUDNNDataLayout::kNCHW;
803 804 805 806 807 808
#ifdef PADDLE_WITH_HIP
    miopenTensorDescriptor_t descp = desc.descriptor<T>(layout, tensor_dims);
#else
    cudnnTensorDescriptor_t descp = desc.descriptor<T>(layout, tensor_dims);
#endif

809
    auto handle = dev_ctx.cudnn_handle();
810 811 812 813

#ifdef PADDLE_WITH_HIP
    auto mode = axis == rank - 1 ? MIOPEN_SOFTMAX_MODE_INSTANCE
                                 : MIOPEN_SOFTMAX_MODE_CHANNEL;
814 815 816 817 818 819 820 821 822 823
    PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::miopenSoftmaxForward_V2(
        handle,
        paddle::platform::CudnnDataType<T>::kOne(),
        descp,
        logits_data,
        paddle::platform::CudnnDataType<T>::kZero(),
        descp,
        softmax_data,
        MIOPEN_SOFTMAX_LOG,
        mode));
824 825 826
#else
    auto mode = axis == rank - 1 ? CUDNN_SOFTMAX_MODE_INSTANCE
                                 : CUDNN_SOFTMAX_MODE_CHANNEL;
827 828 829 830 831 832 833 834 835
    PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSoftmaxForward(
        handle,
        CUDNN_SOFTMAX_LOG,
        mode,
        paddle::platform::CudnnDataType<T>::kOne(),
        descp,
        logits_data,
        paddle::platform::CudnnDataType<T>::kZero(),
        descp,
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855
        softmax_data));
#endif

    const int kDimLog2 = static_cast<int>(Log2Ceil(dim));
    const int kDimCeil = 1 << kDimLog2;
#ifdef __HIPCC__
    int kThreadPerBlock = 256;
#else
    int kThreadPerBlock = 512;
#endif

    int kBatchPerBlock = 1;
    int blocks = (N * D + kBatchPerBlock - 1) / kBatchPerBlock;
    dim3 threads(kThreadPerBlock / kBatchPerBlock, kBatchPerBlock, 1);

    CrossEntropySoftLabel<T, T, true><<<blocks, threads, 0, stream>>>(
        loss_data, softmax_data, NULL, labels_data, N, dim, D, kDimLog2);
  }
}

856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899
/*
  Core function of softmax with cross entropy forward
    - softmax, SoftmaxMode=kSoftmax
    - log softmax, SoftmaxMode=kLogSoftmax
    - softmax with cross entropy hard label, SoftmaxMode=kCrossEntropy
  The computation includes
    - Compute max value: maxvalue_{i} = max_j src_{i,j}
    - Compute sum of exp: s_{i} = sum_{j}{e^{src_{i,j} - maxvalue_{i}}}
    - Compute: softmax_{i,j} = e^{src_{i,j} - maxvalue_{i}} / s_{i}
    - Compute: logsoftmax_{i,j} = src_{i,j} - maxvalue_{i} - log(s_{i})
    - Compute: loss_{i} = -logsoftmax[i,label[i]] (Hard label)
  This computation results from following formula:
    softmax_{i,j} = e^{src_{i,j}} / sum_{j}{e^{src_{i,j}}}
                  = e^{src_{i,j} - maxvalue_{i}}
                    / sum_{j}{e^{src_{i,j} - maxvalue_{i}}}
                  = e^{src_{i,j} - maxvalue_{i}} / s_{i}
    logsoftmax_{i,j} = log(softmax_{i,j})
                     = src_{i,j} - maxvalue_{i} - log(s_{i})
  One warp (32 threads) is used to compute 1 or 2 batch (kBatchSize).
  For reduction max (sum), firstly compute max (sum) to one warp, then use
  shuffle api to compute max (sum) in one warp.
*/
template <typename T,
          typename LabelT,
          typename VecT,
          typename AccT,
          int Log2Elements,
          SoftmaxMode mode,
          bool IgnoreIndex>
__global__ void WarpSoftmaxForward(T* loss,
                                   T* softmax,
                                   const T* src,
                                   const LabelT* label,
                                   const int batch_size,
                                   const int stride,
                                   const int element_count,
                                   const int ignore_index) {
  constexpr int kDimCeil = 1 << Log2Elements;
  constexpr int kWarpSize = (kDimCeil < 32) ? kDimCeil : 32;
  constexpr int kVSize = sizeof(VecT) / sizeof(T);
  constexpr int kIterations = kDimCeil / kWarpSize;
  constexpr int kIterationsV =
      (kIterations >= kVSize) ? (kIterations / kVSize) : 1;
  constexpr int kBatchSize = (kDimCeil <= 128) ? 2 : 1;
S
sneaxiy 已提交
900

901 902 903 904 905 906 907 908
  int first_batch = (blockDim.y * blockIdx.x + threadIdx.y) * kBatchSize;

  // max index to read
  int idx_max_v[kBatchSize];
#pragma unroll
  for (int i = 0; i < kBatchSize; i++) {
    int idx_max = ((i + first_batch) < batch_size) ? element_count : 0;
    idx_max_v[i] = idx_max / kVSize;
909 910
  }

911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943
  // read data from global memory
  AccT srcdata[kBatchSize][kIterationsV][kVSize];

#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
// read data to srcdata: - KVSize==1, - KVSize>1
#pragma unroll
    for (int it = 0; it < kIterationsV; ++it) {
      int src_idx = threadIdx.x + it * kWarpSize;
      if (kVSize == 1) {
        if (src_idx < idx_max_v[i]) {
          srcdata[i][it][0] =
              static_cast<AccT>(src[(first_batch + i) * stride + src_idx]);
        } else {
          srcdata[i][it][0] = -std::numeric_limits<AccT>::infinity();
        }
      } else {
        const VecT* src_v =
            reinterpret_cast<const VecT*>(&src[(first_batch + i) * stride]);
        if (src_idx < idx_max_v[i]) {
          VecT srctmp = src_v[src_idx];
          const T* srcinptr = reinterpret_cast<const T*>(&srctmp);
#pragma unroll
          for (int s = 0; s < kVSize; s++) {
            srcdata[i][it][s] = static_cast<AccT>(srcinptr[s]);
          }
        } else {
#pragma unroll
          for (int s = 0; s < kVSize; s++) {
            srcdata[i][it][s] = -std::numeric_limits<AccT>::infinity();
          }
        }
      }
944 945 946
    }
  }

947 948 949 950 951 952 953 954 955
  // compute max value: maxvalue_{i} = max_j src_{i,j}
  AccT max_value[kBatchSize];
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    // it = 0
    AccT valmax = srcdata[i][0][0];
#pragma unroll
    for (int s = 1; s < kVSize; ++s) {
      valmax = (valmax > srcdata[i][0][s]) ? valmax : srcdata[i][0][s];
956
    }
957
    max_value[i] = valmax;
958

959 960 961 962 963 964 965 966 967 968
// it = 1, 2, ...
#pragma unroll
    for (int it = 1; it < kIterationsV; ++it) {
      AccT valmax = srcdata[i][it][0];
#pragma unroll
      for (int s = 1; s < kVSize; ++s) {
        valmax = (valmax > srcdata[i][it][s]) ? valmax : srcdata[i][it][s];
      }
      max_value[i] = (max_value[i] > valmax) ? max_value[i] : valmax;
    }
969
  }
970
  phi::WarpReduceMax<AccT, kBatchSize, kWarpSize>(max_value);
971

972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991
  // compute sum: s_{i} = sum_{j}{ exp(src_{i,j} - maxvalue_{i} }
  AccT sum[kBatchSize];
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    // it = 0
    if (mode == SoftmaxMode::kLogSoftmax ||
        mode == SoftmaxMode::kCrossEntropy) {
      sum[i] = std::exp(srcdata[i][0][0] - max_value[i]);
    } else {
      srcdata[i][0][0] = std::exp(srcdata[i][0][0] - max_value[i]);
      sum[i] = srcdata[i][0][0];
    }
#pragma unroll
    for (int s = 1; s < kVSize; ++s) {
      if (mode == SoftmaxMode::kLogSoftmax ||
          mode == SoftmaxMode::kCrossEntropy) {
        sum[i] += std::exp(srcdata[i][0][s] - max_value[i]);
      } else {
        srcdata[i][0][s] = std::exp(srcdata[i][0][s] - max_value[i]);
        sum[i] += srcdata[i][0][s];
992
      }
993
    }
994

995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006
// it = 1, 2, ...
#pragma unroll
    for (int it = 1; it < kIterationsV; ++it) {
#pragma unroll
      for (int s = 0; s < kVSize; ++s) {
        if (mode == SoftmaxMode::kLogSoftmax ||
            mode == SoftmaxMode::kCrossEntropy) {
          sum[i] += std::exp(srcdata[i][it][s] - max_value[i]);
        } else {
          srcdata[i][it][s] = std::exp(srcdata[i][it][s] - max_value[i]);
          sum[i] += srcdata[i][it][s];
        }
1007
      }
1008 1009 1010
    }
  }
  phi::WarpReduceSum<AccT, kBatchSize, kWarpSize>(sum);
1011

1012 1013 1014 1015 1016 1017 1018
// write data
#pragma unroll
  for (int i = 0; i < kBatchSize; ++i) {
    if (mode == SoftmaxMode::kLogSoftmax ||
        mode == SoftmaxMode::kCrossEntropy) {
      sum[i] = std::log(sum[i]);
    }
1019

1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035
#pragma unroll
    for (int it = 0; it < kIterationsV; ++it) {
      int idx = threadIdx.x + it * kWarpSize;
      if (kVSize == 1) {  // kVSize==1
        if (idx < idx_max_v[i]) {
          if (mode == SoftmaxMode::kLogSoftmax) {  // log softmax
            softmax[(first_batch + i) * stride + idx] =
                srcdata[i][it][0] - max_value[i] - sum[i];
            // softmax with cross entropy hard label
          } else if (mode == SoftmaxMode::kCrossEntropy) {
            AccT logsoftmax = srcdata[i][it][0] - max_value[i] - sum[i];
            // softmax
            softmax[(first_batch + i) * stride + idx] = std::exp(logsoftmax);
            // label
            int loss_idx = (threadIdx.x + it * kWarpSize) * kVSize;
            auto lbl = static_cast<int64_t>(label[first_batch + i]);
1036
            assert(lbl >= 0 && lbl < element_count || lbl == ignore_index);
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079
            if (IgnoreIndex == true) {
              // IgnoreIndex is true
              if (lbl == loss_idx) {
                if (lbl != ignore_index) {
                  loss[first_batch + i] = -logsoftmax;
                } else {
                  loss[first_batch + i] = static_cast<T>(0.0);
                }
              }
            } else {
              // IgnoreIndex is false
              if (lbl >= 0 && lbl < element_count) {
                if (lbl == loss_idx) {
                  loss[first_batch + i] = -logsoftmax;
                }
              } else {
                loss[first_batch + i] = static_cast<T>(0.0);
              }
            }
          } else {  // softmax
            softmax[(first_batch + i) * stride + idx] =
                srcdata[i][it][0] / sum[i];
          }
        } else {
          break;
        }
      } else {  // KVSize>1
        VecT* softmax_v =
            reinterpret_cast<VecT*>(&softmax[(first_batch + i) * stride]);
        VecT tmpdata;
        T* tmpptr = reinterpret_cast<T*>(&tmpdata);
#pragma unroll
        for (int s = 0; s < kVSize; ++s) {
          if (mode == SoftmaxMode::kLogSoftmax) {  // log softmax
            tmpptr[s] = srcdata[i][it][s] - max_value[i] - sum[i];
            // softmax with cross entropy hard label
          } else if (mode == SoftmaxMode::kCrossEntropy) {
            AccT logsoftmax = srcdata[i][it][s] - max_value[i] - sum[i];
            // softmax
            tmpptr[s] = std::exp(logsoftmax);
            // label
            int loss_idx = (threadIdx.x + it * kWarpSize) * kVSize + s;
            auto lbl = static_cast<int64_t>(label[first_batch + i]);
1080
            assert(lbl >= 0 && lbl < element_count || lbl == ignore_index);
1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101
            if (IgnoreIndex == true) {
              // IgnoreIndex is true
              if (lbl == loss_idx && lbl != ignore_index) {
                loss[first_batch + i] = -logsoftmax;
              }
            } else {
              // IgnoreIndex is false
              if (lbl >= 0 && lbl < element_count) {
                if (lbl == loss_idx) {
                  loss[first_batch + i] = -logsoftmax;
                }
              } else {
                loss[first_batch + i] = static_cast<T>(0.0);
              }
            }
          } else {  // softmax
            tmpptr[s] = srcdata[i][it][s] / sum[i];
          }
        }
        if (idx < idx_max_v[i]) {
          softmax_v[idx] = tmpdata;
1102
        } else {
1103
          break;
1104
        }
1105 1106
      }
    }
1107 1108
  }
}
1109

1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120
#define SOFTMAX_WARP_FORWARD_CASE(Log2Elements, LabelT, VecT, AccT)            \
  case Log2Elements:                                                           \
    WarpSoftmaxForward<T, LabelT, VecT, AccT, Log2Elements, mode, IgnoreIndex> \
        <<<blocks, threads, 0, stream>>>(loss,                                 \
                                         softmax,                              \
                                         src,                                  \
                                         label,                                \
                                         batch_size,                           \
                                         stride,                               \
                                         element_count,                        \
                                         ignore_index);                        \
1121
    break;
1122

1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136
/*
  Wrapper of softmax with cross entropy forward hard label.
*/
template <typename T, typename LabelT, SoftmaxMode mode, bool IgnoreIndex>
void SwitchWarpSoftmaxForward(T* loss,
                              T* softmax,
                              const T* src,
                              const LabelT* label,
                              const int batch_size,
                              const int stride,
                              const int element_count,
                              const int ignore_index,
                              gpuStream_t stream) {
  using AccT = typename dtype::MPTypeTrait<T>::Type;
1137

1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
  // use 128 threads per block to maximimize gpu utilization
  const int log2_elements = static_cast<int>(Log2Ceil(element_count));
  const int kDimCeil = 1 << log2_elements;
  int kWarpSize = (kDimCeil < 32) ? kDimCeil : 32;
  int batches_per_warp = (kDimCeil <= 128) ? 2 : 1;
  constexpr int threads_per_block = 128;
  int warps_per_block = (threads_per_block / kWarpSize);
  int batches_per_block = warps_per_block * batches_per_warp;
  int blocks = (batch_size + batches_per_block - 1) / batches_per_block;
  dim3 threads(kWarpSize, warps_per_block, 1);
1148

1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163
  switch (log2_elements) {
    SOFTMAX_WARP_FORWARD_CASE(0, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(1, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(2, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(3, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(4, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(5, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(6, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(7, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(8, LabelT, T, AccT);
    SOFTMAX_WARP_FORWARD_CASE(9, LabelT, T, AccT);
    default:
      break;
  }
}
1164

1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180
template <typename T, typename LabelT, bool IgnoreIndex>
void LaunchVectorizedSoftmaxForward(T* loss,
                                    T* softmax,
                                    const T* logits,
                                    const LabelT* label,
                                    const int high_dim,
                                    const int mid_dim,
                                    const int ignore_index,
                                    gpuStream_t stream) {
  using AccT = typename dtype::MPTypeTrait<T>::Type;
  constexpr int vec_size = sizeof(float4) / sizeof(T);
  const int max_num_threads = 1024;
  int max_block_size = std::min(mid_dim / vec_size, max_num_threads);
  if (vec_size > 1) {
    max_block_size /= 2;
  }
1181

1182 1183 1184 1185 1186 1187 1188
  int block_size = 1;
  while (block_size < max_block_size) {
    block_size *= 2;
  }
  block_size = std::max(block_size, kps::details::kWarpSize);
  dim3 grids(high_dim);
  dim3 blocks(block_size);
1189 1190 1191
  VectorizedSoftmaxForward<T, AccT, LabelT, vec_size, IgnoreIndex>
      <<<grids, blocks, 0, stream>>>(
          loss, softmax, logits, label, high_dim, mid_dim, ignore_index);
1192
}
1193

1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
/*
  Wrapper of softmax with cross entropy hard label.
  - SwitchWarpSoftmaxForward for small size when axis == -1
  - LaunchVectorizedSoftmaxForward for large size when axis == -1
  - cudnn function for axis != -1
*/
template <typename T, typename LabelT, bool IgnoreIndex>
static void SoftmaxWithCrossEntropyHardLabel(const GPUContext& dev_ctx,
                                             int rank,
                                             int axis,
                                             const T* logits_data,
                                             const LabelT* labels_data,
                                             T* loss_data,
                                             T* softmax_data,
                                             int N,
                                             int dim,
                                             int D,
                                             const int ignore_index) {
  auto stream = dev_ctx.stream();
  constexpr int max_dim = 320;
  if (D == 1) {
    if (dim <= max_dim) {  // small size
      const SoftmaxMode mode = SoftmaxMode::kCrossEntropy;
      SwitchWarpSoftmaxForward<T, LabelT, mode, IgnoreIndex>(loss_data,
                                                             softmax_data,
                                                             logits_data,
                                                             labels_data,
                                                             N,
                                                             dim,
                                                             dim,
                                                             ignore_index,
                                                             stream);
    } else {  // large size
      LaunchVectorizedSoftmaxForward<T, LabelT, IgnoreIndex>(loss_data,
                                                             softmax_data,
                                                             logits_data,
                                                             labels_data,
                                                             N,
                                                             dim,
                                                             ignore_index,
                                                             stream);
S
sneaxiy 已提交
1235
    }
1236 1237 1238 1239 1240 1241 1242 1243 1244
  } else {
    ScopedTensorDescriptor desc;
    std::vector<int> tensor_dims = {N, dim, D, 1};
    GPUDNNDataLayout layout = GPUDNNDataLayout::kNCHW;
#ifdef PADDLE_WITH_HIP
    miopenTensorDescriptor_t descp = desc.descriptor<T>(layout, tensor_dims);
#else
    cudnnTensorDescriptor_t descp = desc.descriptor<T>(layout, tensor_dims);
#endif
C
caoying03 已提交
1245

1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
    auto handle = dev_ctx.cudnn_handle();

#ifdef PADDLE_WITH_HIP
    auto mode = axis == rank - 1 ? MIOPEN_SOFTMAX_MODE_INSTANCE
                                 : MIOPEN_SOFTMAX_MODE_CHANNEL;
    PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::miopenSoftmaxForward_V2(
        handle,
        paddle::platform::CudnnDataType<T>::kOne(),
        descp,
        logits_data,
        paddle::platform::CudnnDataType<T>::kZero(),
        descp,
        softmax_data,
        MIOPEN_SOFTMAX_LOG,
        mode));
#else
    auto mode = axis == rank - 1 ? CUDNN_SOFTMAX_MODE_INSTANCE
                                 : CUDNN_SOFTMAX_MODE_CHANNEL;
    PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSoftmaxForward(
        handle,
        CUDNN_SOFTMAX_LOG,
        mode,
        paddle::platform::CudnnDataType<T>::kOne(),
        descp,
        logits_data,
        paddle::platform::CudnnDataType<T>::kZero(),
        descp,
        softmax_data));
#endif
    int threads = 128;
    int blocks = (N * dim * D + threads - 1) / threads;
    // compute cross entropy, input is log softmax
1278 1279 1280
    CrossEntropyExpHardLabel<T, LabelT, IgnoreIndex>
        <<<blocks, threads, 0, stream>>>(
            loss_data, softmax_data, labels_data, N, dim, D, ignore_index);
1281
  }
1282
}
1283

1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321
template <typename T, typename LabelT>
void CrossEntropyWithSoftmaxCUDAKernel(const GPUContext& dev_ctx,
                                       const DenseTensor& logits,
                                       const DenseTensor& label,
                                       bool soft_label,
                                       bool use_softmax,
                                       bool numeric_stable_mode,
                                       int ignore_index,
                                       int axis,
                                       DenseTensor* softmax,
                                       DenseTensor* loss) {
  PADDLE_ENFORCE_EQ(
      dev_ctx.GetPlace().GetType(),
      AllocationType::GPU,
      phi::errors::Unavailable("softmax_with_cross_entropy operator's "
                               "CUDA kernel only runs on GPU device."));

  // do not with softmax op, and input is softmax
  if (!use_softmax) {
    DenseTensor* softmax_out = softmax;
    const DenseTensor* softmax = &logits;
    const DenseTensor& labels = label;

    const int rank = softmax->dims().size();
    const int axis_v = phi::funcs::CanonicalAxis(axis, rank);
    const int axis_dim = softmax->dims()[axis_v];

    const int n = phi::funcs::SizeToAxis(axis_v, softmax->dims());
    const int d = phi::funcs::SizeFromAxis(axis_v, softmax->dims());

    auto* softmax_out_data = dev_ctx.template Alloc<T>(softmax_out);
    auto* loss_data = dev_ctx.template Alloc<T>(loss);

    phi::funcs::SetConstant<GPUContext, T> set_constant;
    set_constant(dev_ctx, loss, static_cast<T>(0));
    if (axis_dim == 1) {
      set_constant(dev_ctx, softmax_out, static_cast<T>(1));
      return;
Z
Zeng Jinle 已提交
1322
    }
C
caoying03 已提交
1323

1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344
    DenseTensor softmax_2d(*softmax);
    softmax_2d.Resize({n, d});
    DenseTensor labels_2d(labels);
    labels_2d.Resize({n, labels.numel() / n});
    DenseTensor loss_2d(*loss);
    loss_2d.Resize({n, 1});
    DenseTensor softmax_out_2d(*softmax_out);
    softmax_out_2d.Resize({n, d});

    // math::CrossEntropyFunctor support axis is the last
    if (axis_v == -1) {
      paddle::operators::math::CrossEntropyFunctor<GPUContext, T>()(
          dev_ctx,
          &loss_2d,
          &softmax_2d,
          &labels_2d,
          soft_label,
          ignore_index,
          axis_dim);
      return;
    }
1345

1346 1347 1348 1349
    // if axis is not the last, we need a new impliment
    if (soft_label) {
      auto* logits_data = softmax->data<T>();
      auto* labels_data = labels.data<T>();
1350

1351 1352
      const int kDimLog2 = static_cast<int>(Log2Ceil(axis_dim));
      const int kDimCeil = 1 << kDimLog2;
1353
#ifdef __HIPCC__
1354
      int kThreadPerBlock = 256;
1355
#else
1356
      int kThreadPerBlock = 512;
1357
#endif
1358 1359 1360 1361
      int kBatchPerBlock = 1;
      int blocks = (n * d + kBatchPerBlock - 1) / kBatchPerBlock;
      dim3 threads(kThreadPerBlock / kBatchPerBlock, kBatchPerBlock, 1);

1362 1363 1364 1365 1366 1367 1368 1369 1370
      CrossEntropySoftLabel<T, T, false>
          <<<blocks, threads, 0, dev_ctx.stream()>>>(loss_data,
                                                     NULL,
                                                     logits_data,
                                                     labels_data,
                                                     n,
                                                     axis_dim,
                                                     d / axis_dim,
                                                     kDimLog2);
1371 1372 1373 1374 1375 1376
    } else {  // HardLabel
      auto* logits_data = softmax->data<T>();
      auto* labels_data = labels.data<LabelT>();
      int threads = 128;
      int blocks = (n * d / axis_dim + threads - 1) / threads;
      if (ignore_index >= 0 && ignore_index < axis_dim) {
1377 1378 1379 1380 1381 1382 1383 1384
        CrossEntropyHardLabel<T, LabelT, true>
            <<<blocks, threads, 0, dev_ctx.stream()>>>(loss_data,
                                                       logits_data,
                                                       labels_data,
                                                       n,
                                                       axis_dim,
                                                       d / axis_dim,
                                                       ignore_index);
1385
      } else {
1386 1387 1388 1389 1390 1391 1392 1393
        CrossEntropyHardLabel<T, LabelT, false>
            <<<blocks, threads, 0, dev_ctx.stream()>>>(loss_data,
                                                       logits_data,
                                                       labels_data,
                                                       n,
                                                       axis_dim,
                                                       d / axis_dim,
                                                       ignore_index);
1394 1395 1396
      }
    }

1397 1398 1399 1400
    // cause of input is softmax
    // copy to output softmax, directly
    phi::Copy<GPUContext>(
        dev_ctx, *softmax, dev_ctx.GetPlace(), false, softmax_out);
1401

1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
    return;
  }

  const int rank = logits.dims().size();
  const int axis_v = phi::funcs::CanonicalAxis(axis, rank);
  int axis_dim = logits.dims()[axis_v];

  const int64_t n = phi::funcs::SizeToAxis(axis_v, logits.dims());
  const int64_t d = phi::funcs::SizeFromAxis(axis_v, logits.dims());

  auto* softmax_data = dev_ctx.template Alloc<T>(softmax);
  auto* loss_data = dev_ctx.template Alloc<T>(loss);

  if (axis_dim == 1) {
    phi::funcs::SetConstant<GPUContext, T> set_constant;
    set_constant(dev_ctx, softmax, static_cast<T>(1));
    set_constant(dev_ctx, loss, static_cast<T>(0));
    return;
  }

  if (soft_label) {
    auto* logits_data = logits.data<T>();
    auto* labels_data = label.data<T>();
    SoftmaxWithCrossEntropySoftLabel<T>(dev_ctx,
                                        rank,
                                        axis_v,
                                        logits_data,
                                        labels_data,
                                        softmax_data,
                                        loss_data,
                                        n,
                                        axis_dim,
                                        d / axis_dim);
  } else {
    if (!numeric_stable_mode) {
      // CUDNN kernel only suppoer 2-D tensor and perfome softmax on last dim
      DenseTensor logits_2d(logits);
      logits_2d.Resize({n, d});
      DenseTensor softmax_2d(*softmax);
      softmax_2d.Resize({n, d});
      DenseTensor labels_2d(label);
      labels_2d.Resize({n, label.numel() / n});
      DenseTensor loss_2d(*loss);
      loss_2d.Resize({n, 1});
      paddle::operators::math::SoftmaxCUDNNFunctor<T, GPUContext>()(
          dev_ctx, &logits_2d, &softmax_2d);
      paddle::operators::math::CrossEntropyFunctor<GPUContext, T>()(
          dev_ctx,
          &loss_2d,
          &softmax_2d,
          &labels_2d,
          false,
          ignore_index,
          axis_dim);
1456
    } else {
1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483
      auto* logits_data = logits.data<T>();
      auto* labels_data = label.data<LabelT>();
      if (ignore_index >= 0 && ignore_index < axis_dim) {
        SoftmaxWithCrossEntropyHardLabel<T, LabelT, true>(dev_ctx,
                                                          rank,
                                                          axis_v,
                                                          logits_data,
                                                          labels_data,
                                                          loss_data,
                                                          softmax_data,
                                                          n,
                                                          axis_dim,
                                                          d / axis_dim,
                                                          ignore_index);
      } else {
        SoftmaxWithCrossEntropyHardLabel<T, LabelT, false>(dev_ctx,
                                                           rank,
                                                           axis_v,
                                                           logits_data,
                                                           labels_data,
                                                           loss_data,
                                                           softmax_data,
                                                           n,
                                                           axis_dim,
                                                           d / axis_dim,
                                                           ignore_index);
      }
1484
    }
C
caoying03 已提交
1485
  }
1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516
}

template <typename T, typename Context>
void CrossEntropyWithSoftmaxKernel(const Context& dev_ctx,
                                   const DenseTensor& logits,
                                   const DenseTensor& label,
                                   bool soft_label,
                                   bool use_softmax,
                                   bool numeric_stable_mode,
                                   int ignore_index,
                                   int axis,
                                   DenseTensor* softmax,
                                   DenseTensor* loss) {
  auto dtype = label.dtype();
  if (soft_label) {
    PADDLE_ENFORCE_EQ(
        dtype,
        paddle::experimental::CppTypeToDataType<T>::Type(),
        phi::errors::InvalidArgument("The Input(Label) should be with the "
                                     "same data type as Input(Logits)."));
    CrossEntropyWithSoftmaxCUDAKernel<T, T>(dev_ctx,
                                            logits,
                                            label,
                                            soft_label,
                                            use_softmax,
                                            numeric_stable_mode,
                                            ignore_index,
                                            axis,
                                            softmax,
                                            loss);
  } else {
1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529
    PD_VISIT_INTEGRAL_TYPES(dtype, "CrossEntropyWithSoftmaxCUDAKernel", ([&] {
                              CrossEntropyWithSoftmaxCUDAKernel<T, data_t>(
                                  dev_ctx,
                                  logits,
                                  label,
                                  soft_label,
                                  use_softmax,
                                  numeric_stable_mode,
                                  ignore_index,
                                  axis,
                                  softmax,
                                  loss);
                            }));
1530 1531
  }
}
C
caoying03 已提交
1532

1533
}  // namespace phi
C
caoying03 已提交
1534

1535
#ifdef PADDLE_WITH_HIP
1536 1537 1538 1539 1540 1541
PD_REGISTER_KERNEL(cross_entropy_with_softmax,
                   GPU,
                   ALL_LAYOUT,
                   phi::CrossEntropyWithSoftmaxKernel,
                   float,
                   phi::dtype::float16) {}
1542
#else
1543 1544 1545 1546 1547 1548 1549
PD_REGISTER_KERNEL(cross_entropy_with_softmax,
                   GPU,
                   ALL_LAYOUT,
                   phi::CrossEntropyWithSoftmaxKernel,
                   float,
                   double,
                   phi::dtype::float16) {}
1550
#endif