hl_cuda_matrix.cu 19.5 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

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. */


#include "hl_base.h"
#include "hl_matrix.h"
#include "hl_matrix_ops.cuh"
#include "hl_matrix_apply.cuh"
#include "hl_sequence.h"
#include "paddle/utils/Logging.h"
22
#include "hl_device_functions.cuh"
Z
zhangjinchao01 已提交
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

DEFINE_MATRIX_UNARY_OP(Zero, a = 0);
DEFINE_MATRIX_TERNARY_PARAMETER_OP(_add, TWO_PARAMETER, c = p1*a + p2*b);
void hl_matrix_add(real *A_d,
                   real *B_d,
                   real *C_d,
                   int dimM,
                   int dimN,
                   real alpha,
                   real beta) {
  CHECK_NOTNULL(A_d);
  CHECK_NOTNULL(B_d);
  CHECK_NOTNULL(C_d);

  hl_gpu_apply_ternary_op
    <real, ternary::_add<real>, 0, 0>(ternary::_add<real>(alpha, beta),
                                      A_d,
                                      B_d,
                                      C_d,
                                      dimM,
                                      dimN,
                                      dimN,
                                      dimN,
                                      dimN);
  CHECK_SYNC("hl_matrix_add failed");
}

50
#ifdef PADDLE_TYPE_DOUBLE
Z
zhangjinchao01 已提交
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
    #define THRESHOLD   128
#else
    #define THRESHOLD   64
#endif
__device__ __forceinline__
void findMax(real* I,
             real* dfMax_s,
             int blockSize,
             int base,
             int curIdx,
             int nextIdx,
             int dimN,
             real* max) {
  dfMax_s[base] = -1.0e20;
  while (curIdx < dimN) {
    if (dfMax_s[base] < I[nextIdx]) {
      dfMax_s[base] = I[nextIdx];
    }
    nextIdx += blockSize;
    curIdx += blockSize;
  }
  __syncthreads();

  for (int stride = blockSize >> 1; stride > 0; stride >>= 1) {
    __syncthreads();
    if (base < stride) {
      nextIdx = base + stride;
      if (dfMax_s[base] < dfMax_s[nextIdx]) {
          dfMax_s[base] = dfMax_s[nextIdx];
      }
    }
  }

  if (0 == base)  {
    max[0] = dfMax_s[0];
  }
  __syncthreads();
}

__device__ __forceinline__
void subMaxAndExp(real* I,
                  real* O,
                  int curIdx,
                  int nextIdx,
                  int blockSize,
                  int dimN,
                  real max) {
  real val;
  while (curIdx < dimN) {
    val = I[nextIdx] - max;
    if (val < -THRESHOLD) {
      val = -THRESHOLD;
    }
    I[nextIdx] = val;
105
#ifndef PADDLE_TYPE_DOUBLE
Z
zhangjinchao01 已提交
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
    O[nextIdx] = __expf(val);
#else
    O[nextIdx] = exp(val);
#endif
    nextIdx += blockSize;
    curIdx += blockSize;
  }
  __syncthreads();
}

__device__ __forceinline__
void valueSum(real* O,
              real* dfMax_s,
              int blockSize,
              int base,
              int curIdx,
              int nextIdx,
              int dimN) {
  dfMax_s[base] = 0;
  while (curIdx < dimN) {
    dfMax_s[base] += O[nextIdx];
    nextIdx += blockSize;
    curIdx += blockSize;
  }
  __syncthreads();

  for (int stride = blockSize >> 1; stride > 0; stride >>= 1) {
    __syncthreads();
    if (base < stride) {
      nextIdx = base + stride;
      dfMax_s[base] += dfMax_s[nextIdx];
    }
  }
  __syncthreads();
}

__device__ __forceinline__
void divSum(real* O,
            real sum,
            int curIdx,
            int nextIdx,
            int blockSize,
            int dimN) {
  while (curIdx < dimN) {
    O[nextIdx] /= sum;
    nextIdx += blockSize;
    curIdx += blockSize;
  }
}

__device__ __forceinline__
void softmax(real* I,
             real* O,
             real* dfMax_s,
             int blockSize,
             int base,
             int curIdx,
             int nextIdx,
             int dimN) {
  __shared__ real max;

  // find the max number
  findMax(I, dfMax_s, blockSize, base, curIdx,
          nextIdx, dimN, &max);

  // sub max Value and do Exp operation
  subMaxAndExp(I, O, base, nextIdx, blockSize, dimN, max);

  // add dimN values into blockDim.x buffer
  // sum is in dfMax_s[0]
  valueSum(O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN);

  // divided by sum
  divSum(O, dfMax_s[0], curIdx, nextIdx, blockSize, dimN);
}

template<int blockSize>
__global__ void KeMatrixSoftMax(real *O, real *I, int dimN) {
  int base = threadIdx.x;
  __shared__ real dfMax_s[blockSize];
  int nextIdx = blockIdx.x * dimN + base;
  int curIdx = base;

  softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN);
}

void hl_matrix_softmax(real *A_d, real *C_d, int dimM, int dimN) {
  CHECK_NOTNULL(A_d);
  CHECK_NOTNULL(C_d);

  dim3 block(512, 1);
  dim3 grid(dimM, 1);
  KeMatrixSoftMax<512>
           <<<grid, block, 0, STREAM_DEFAULT>>>(C_d, A_d, dimN);
  CHECK_SYNC("hl_matrix_softmax failed");
}

template<int blockSize>
__global__ void KeSequenceSoftMax(real *O, real *I, const int* index) {
  int base = threadIdx.x;
  int bid = blockIdx.x;
  __shared__ real dfMax_s[blockSize];

  int start = index[bid];
  int dimN = index[bid + 1] - start;

  int nextIdx = start + base;
  int curIdx = base;

  softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN);
}

void hl_sequence_softmax_forward(real *A_d,
                                 real *C_d,
                                 const int* index,
                                 int numSequence) {
  CHECK_NOTNULL(A_d);
  CHECK_NOTNULL(C_d);

  dim3 block(512, 1);
  dim3 grid(numSequence, 1);
  KeSequenceSoftMax<512>
           <<<grid, block, 0, STREAM_DEFAULT>>>(C_d, A_d, index);
  CHECK_SYNC("hl_sequence_softmax_forward failed");
}

__global__ void KeMatrixDerivative(real *grad_d,
                                   real *output_d,
                                   real *sftmaxSum_d,
                                   int dimM,
                                   int dimN) {
  int rowIdx = blockIdx.x*blockDim.x + threadIdx.x;
  int colIdx = blockIdx.y*blockDim.y + threadIdx.y;
  int index;

  if (rowIdx < dimM && colIdx < dimN) {
    index = rowIdx*dimN + colIdx;
    grad_d[index] = output_d[index] * (grad_d[index] - sftmaxSum_d[rowIdx]);
  }
}

void hl_matrix_softmax_derivative(real *grad_d,
                                  real *output_d,
                                  real *sftmaxSum_d,
                                  int dimM,
                                  int dimN) {
  CHECK_NOTNULL(grad_d);
  CHECK_NOTNULL(output_d);
  CHECK_NOTNULL(sftmaxSum_d);

  int blocksX = (dimM + 0) / 1;
  int blocksY = (dimN + 1024 -1) / 1024;
  dim3 threads(1, 1024);
  dim3 grid(blocksX, blocksY);

  KeMatrixDerivative<<< grid, threads, 0, STREAM_DEFAULT >>>
           (grad_d, output_d, sftmaxSum_d, dimM, dimN);
  CHECK_SYNC("hl_matrix_softmax_derivative failed");
}

template<int blockSize>
__global__ void KeMatrixClassificationError(real* in_A,
                                            int* in_B,
                                            real* out_C,
                                            int dimN) {
  __shared__ real max_s[blockSize];
  __shared__ int max_l[blockSize];
H
He 已提交
273 274
  const int tid = threadIdx.x;
  const int rowId = blockIdx.x;
Z
zhangjinchao01 已提交
275 276

  max_s[tid] = -1e30f;
H
He 已提交
277 278 279 280 281 282 283
  in_A += rowId * dimN;
  real tmp;
  for (int colId = tid; colId < dimN; colId += blockSize) {
    tmp = in_A[colId];
    if (max_s[tid] < tmp) {
      max_s[tid] = tmp;
      max_l[tid] = colId;
Z
zhangjinchao01 已提交
284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299
    }
  }
  __syncthreads();

  for (int stride = blockSize/2; stride > 0; stride = stride/2) {
    if (tid < stride) {
      if (max_s[tid] < max_s[tid + stride]) {
        max_s[tid] = max_s[tid + stride];
        max_l[tid] = max_l[tid + stride];
      }
    }
    __syncthreads();
  }
  __syncthreads();

  if (tid == 0) {
H
He 已提交
300
    out_C[rowId] = (max_l[0] == in_B[rowId] ? 0 : 1.0f);
Z
zhangjinchao01 已提交
301 302 303 304 305 306 307 308 309 310 311 312
  }
}

void hl_matrix_classification_error(real* A_d,
                                    int* B_d,
                                    real* C_d,
                                    int dimM,
                                    int dimN) {
  CHECK_NOTNULL(A_d);
  CHECK_NOTNULL(B_d);
  CHECK_NOTNULL(C_d);

H
He 已提交
313 314 315
  // each sample is calculated by one block
  KeMatrixClassificationError<1024><<< dimM, 1024, 0, STREAM_DEFAULT >>>
    (A_d, B_d, C_d, dimN);
Z
zhangjinchao01 已提交
316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625
  CHECK_SYNC("hl_matrix_classification_error");
}

__global__ void KeMatrixCrossEntropy(real* O,
                                     real* E,
                                     int* label,
                                     int dimM,
                                     int dimN) {
  int index = blockIdx.x * blockDim.x + threadIdx.x;
  int newBase;
  if (index < dimM) {
    newBase = label[index];
    newBase = newBase % dimN;
    E[index] = -log(O[index * dimN + newBase]);
  }
}

void hl_matrix_cross_entropy(real* A_d,
                             real* C_d,
                             int* label_d,
                             int dimM,
                             int dimN) {
  CHECK_NOTNULL(A_d);
  CHECK_NOTNULL(C_d);

  int blocks = (dimM + 1024 - 1) / 1024;
  dim3 threads(1024, 1);
  dim3 grid(blocks, 1);
  KeMatrixCrossEntropy<<< grid, threads, 0, STREAM_DEFAULT >>>
           (A_d, C_d, label_d, dimM, dimN);
  CHECK_SYNC("hl_matrix_cross_entropy failed");
}

__global__ void KeMatrixCrossEntropyBp(real* grad_d,
                                       real* output_d,
                                       int* label_d,
                                       int dimM,
                                       int dimN) {
  int rowIdx = blockIdx.x*blockDim.x + threadIdx.x;
  int colIdx = blockIdx.y*blockDim.y + threadIdx.y;
  int index;
  if (rowIdx < dimM && colIdx < dimN) {
    index = rowIdx*dimN + colIdx;
    if (label_d[rowIdx] == colIdx) {
      grad_d[index] -= 1.0f / output_d[index];
    }
  }
}

void hl_matrix_cross_entropy_bp(real* grad_d,
                                real* output_d,
                                int* label_d,
                                int dimM,
                                int dimN) {
  CHECK_NOTNULL(grad_d);
  CHECK_NOTNULL(output_d);
  CHECK_NOTNULL(label_d);

  int blocksX = (dimM + 0)/1;
  int blocksY = (dimN + 1024 -1) / 1024;
  dim3 threads(1, 1024);
  dim3 grid(blocksX, blocksY);
  KeMatrixCrossEntropyBp<<< grid, threads, 0, STREAM_DEFAULT >>>
           (grad_d, output_d, label_d, dimM, dimN);
  CHECK_SYNC("hl_matrix_cross_entropy_bp failed");
}

void hl_matrix_zero_mem(real* data, int num) {
  hl_gpu_apply_unary_op(
        unary::Zero<real>(), data, 1, num, num);
}

__global__ void KeParamReluForward(real* output,
                                   real* input,
                                   real* w,
                                   int width,
                                   int height,
                                   int partial_sum) {
  int tx = blockIdx.x * blockDim.x + threadIdx.x;
  int ty = blockIdx.y * blockDim.y + threadIdx.y;
  if (tx < width && ty < height) {
    int index = ty * width + tx;
    output[index] = input[index] > 0 ? input[index] :
        input[index] * w[tx / partial_sum];
  }
}

void hl_param_relu_forward(real* output,
                           real* input,
                           real* w,
                           int width,
                           int height,
                           int partial_sum) {
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(input);
  CHECK_NOTNULL(w);
  dim3 threads(16, 16);
  int blockX = (width + 16 - 1) / 16;
  int blockY = (height + 16 -1) / 16;
  dim3 grid(blockX, blockY);
  KeParamReluForward<<<grid, threads, 0, STREAM_DEFAULT>>>
    (output, input, w, width, height, partial_sum);
  CHECK_SYNC("hl_param_relu_forward failed");
}

template<int blockSize>
__global__ void KeParamReluBackWardW(real* grad_w,
                                     real* grad_o,
                                     real* input,
                                     int width,
                                     int height,
                                     int partial_sum) {
  const int tid = threadIdx.x;
  __shared__ real temp[blockSize];
  grad_o += partial_sum * blockIdx.x;
  input += partial_sum * blockIdx.x;
  real tmp = 0.0;
  for (int index = tid; index < partial_sum * height; index += blockSize) {
    int row = index / partial_sum;
    int offset = row * width + (index - row * partial_sum);
    if (input[offset] < 0) {
      tmp += grad_o[offset] * input[offset];
    }
  }
  temp[tid] = tmp;
  __syncthreads();
  for (int s = blockSize / 2; s > 0; s >>= 1) {
    if (tid < s) {
      temp[tid] += temp[tid + s];
    }
    __syncthreads();
  }
  if (tid == 0) {
    grad_w[blockIdx.x] += temp[0];
  }
}

void hl_param_relu_backward_w(real* grad_w,
                              real* grad_o,
                              real* input,
                              int width,
                              int height,
                              int partial_sum) {
  CHECK_NOTNULL(grad_w);
  CHECK_NOTNULL(grad_o);
  CHECK_NOTNULL(input);
  const int blockSize = 1024;
  int grid_num = width / partial_sum;
  dim3 threads(blockSize, 1);
  dim3 grid(grid_num, 1);
  KeParamReluBackWardW<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
    (grad_w, grad_o, input, width, height, partial_sum);
  CHECK_SYNC("hl_param_relu_backward_w failed");
}

__global__ void KeParamReluBackwardDiff(real* grad_o,
                                        real* input,
                                        real* w,
                                        real* diff,
                                        int width,
                                        int height,
                                        int partial_sum) {
  int tx = blockIdx.x * blockDim.x + threadIdx.x;
  int ty = blockIdx.y * blockDim.y + threadIdx.y;
  if (tx < width && ty < height) {
    int index = ty * width + tx;
    diff[index] += grad_o[index] * (input[index] > 0 ? 1 : w[tx / partial_sum]);
  }
}

void hl_param_relu_backward_diff(real* grad_o,
                                 real* data,
                                 real* w,
                                 real* diff,
                                 int width,
                                 int height,
                                 int partial_sum) {
  CHECK_NOTNULL(grad_o);
  CHECK_NOTNULL(data);
  CHECK_NOTNULL(w);
  CHECK_NOTNULL(diff);
  dim3 threads(16, 16);
  int blockX = (width + 16 - 1) / 16;
  int blockY = (height + 16 -1) / 16;
  dim3 grid(blockX, blockY);
  KeParamReluBackwardDiff<<<grid, threads, 0, STREAM_DEFAULT>>>
      (grad_o, data, w, diff, width, height, partial_sum);
  CHECK_SYNC("hl_param_relu_backward_diff failed");
}

template<int blockSize>
__global__ void KeCosSim(real* output,
                         real* input1,
                         real* input2,
                         int width,
                         int input1_height,
                         int input2_height,
                         real scale) {
  const int ty = blockIdx.y;
  int tid = threadIdx.x;

  __shared__ real xx[blockSize];
  __shared__ real yy[blockSize];
  __shared__ real xy[blockSize];

  xx[tid] = 0.0;
  yy[tid] = 0.0;
  xy[tid] = 0.0;
  __syncthreads();

  input1 += ty * width;
  if (input2_height > 1) {
    input2 += ty * width;
  }
  for (int index = tid; index < width; index += blockSize) {
    real x = input1[index];
    real y = input2[index];
    xx[tid] += x * x;
    yy[tid] += y * y;
    xy[tid] += x * y;
  }
  __syncthreads();

  for (int s = blockSize / 2; s > 0; s >>= 1) {
    if (tid < s) {
      xx[tid] += xx[tid + s];
      yy[tid] += yy[tid + s];
      xy[tid] += xy[tid + s];
    }
    __syncthreads();
  }
  if (tid == 0) {
    output[ty] = scale * xy[0] / (sqrt(xx[0]) * sqrt(yy[0]));
  }
}

void hl_cossim(real* output,
               real* input1,
               real* input2,
               int width,
               int input1_height,
               int input2_height,
               real scale) {
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(input1);
  CHECK_NOTNULL(input2);
  const int blockSize = 256;
  dim3 threads(blockSize, 1);
  dim3 grid(1, input1_height);

  KeCosSim<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
    (output, input1, input2, width, input1_height, input2_height, scale);
  CHECK_SYNC("hl_cossim failed");
}

template<int blockSize>
__global__ void KeCosSimDerivative(real* grad,
                                   real* output,
                                   real* prevOutX,
                                   real* prevOutY,
                                   real* prevGradX,
                                   real* prevGradY,
                                   int width,
                                   int input1_height,
                                   int input2_height,
                                   real scale) {
  const int ty = blockIdx.y;
  int tid = threadIdx.x;

  __shared__ real xx[blockSize];
  __shared__ real yy[blockSize];
  __shared__ real xy[blockSize];

  xx[tid] = 0.0;
  yy[tid] = 0.0;
  xy[tid] = 0.0;
  __syncthreads();

  prevOutX += ty * width;
  prevGradX += ty * width;
  if (input2_height > 1) {
    prevOutY += ty * width;
    prevGradY += ty * width;
  }
  for (int index = tid; index < width; index += blockSize) {
    real x = prevOutX[index];
    real y = prevOutY[index];
    xx[tid] += x * x;
    yy[tid] += y * y;
    xy[tid] += x * y;
  }
  __syncthreads();

  for (int s = blockSize / 2; s > 0; s >>= 1) {
    if (tid < s) {
      xx[tid] += xx[tid + s];
      yy[tid] += yy[tid + s];
      xy[tid] += xy[tid + s];
    }
    __syncthreads();
  }
  if (xy[0] == 0) {
    real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0]));
    for (int index = tid; index < width; index += blockSize) {
      prevGradX[index] +=
        scale * grad[ty] * prevOutY[index] * reciprocal;
      if (input2_height > 1) {
        prevGradY[index] +=
          scale * grad[ty] * prevOutX[index] * reciprocal;
      } else {
626
        paddle::paddleAtomicAdd(prevGradY + index,
Z
zhangjinchao01 已提交
627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642
          scale * grad[ty] * prevOutX[index] * reciprocal);
      }
    }
  } else {
    real reciprocalXY = 1.0 / xy[0];
    real reciprocalSquareSumX = 1.0 / xx[0];
    real reciprocalSquareSumY = 1.0 / yy[0];
    for (int index = tid; index < width; index += blockSize) {
      prevGradX[index] += output[ty] * grad[ty] *
        (prevOutY[index] * reciprocalXY -
         prevOutX[index] * reciprocalSquareSumX);
      if (input2_height > 1) {
        prevGradY[index] += output[ty] * grad[ty] *
          (prevOutX[index] * reciprocalXY -
           prevOutY[index] * reciprocalSquareSumY);
      } else {
643
        paddle::paddleAtomicAdd(prevGradY + index, output[ty] * grad[ty] *
Z
zhangjinchao01 已提交
644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675
          (prevOutX[index] * reciprocalXY -
           prevOutY[index] * reciprocalSquareSumY));
      }
    }
  }
}


void hl_cossim_derivative(real* grad,
                          real* output,
                          real* prevOutX,
                          real* prevOutY,
                          real* prevGradX,
                          real* prevGradY,
                          int width,
                          int input1_height,
                          int input2_height,
                          real scale) {
  CHECK_NOTNULL(grad);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(prevOutX);
  CHECK_NOTNULL(prevOutY);
  CHECK_NOTNULL(prevGradX);
  CHECK_NOTNULL(prevGradY);
  const int blockSize = 256;
  dim3 threads(blockSize, 1);
  dim3 grid(1, input1_height);
  KeCosSimDerivative<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
    (grad, output, prevOutX, prevOutY, prevGradX, prevGradY, width,
        input1_height, input2_height, scale);
  CHECK_SYNC("hl_cossim_derivate failed");
}