Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
切换分支/标签
查找文件 普通视图历史永久链接Permalink
conv_func.cpp 40.2 KB
Newer Older
Z
adjust gpu code structure  
zhaojiaying01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

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 "operators/kernel/cl/cl-kernel-func/conv_func.h"
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
16 
#include <vector>
Z
adjust gpu code structure  
zhaojiaying01 已提交
17 18 19 20 21 
#include "framework/cl/cl_image_converter.h"
#include "framework/cl/cl_tensor.h"

namespace paddle_mobile {
namespace operators {
S
add opencl slidingwindow3*3s1 (#1799)  
StarryRain 已提交
22 
bool use_lws = true;
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
23 24 
int preferred_lws = 0;
int preferred_lws_divisor = 2;
Z
adjust gpu code structure  
zhaojiaying01 已提交
25 26 

template <>
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
27 28 
void winograd_transform_weight<4, 3>(framework::CLHelper *cl_helper,
                                     framework::CLImage *weight) {}
Z
adjust gpu code structure  
zhaojiaying01 已提交
29 30 

template <>
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
31 32 33 34 35 
void WinogradConv3x3<4, 3>(framework::CLHelper *cl_helper,
                           const ConvParam<GPU_CL> &param, bool ifRelu,
                           const framework::CLImage *biase,
                           const framework::CLImage *new_scale,
                           const framework::CLImage *new_bias) {}
Z
adjust gpu code structure  
zhaojiaying01 已提交
36
N
optimize conv7*7, compute 2 point in 1 thread. test=develop (#2028)  
NazgulLee 已提交
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 
void ConvAddBnReluPt1x2(framework::CLHelper *cl_helper,
                        const ConvParam<GPU_CL> &param, bool ifRelu,
                        const framework::CLImage *biase,
                        const framework::CLImage *new_scale,
                        const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  default_work_size[1] = (default_work_size[1] + 1) / 2;
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];
  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int offset = param.Offset();
  int input_c = reinterpret_cast<framework::CLImageConverterFolder *>(
                    param.Input()->Converter())
                    ->GetCBlock();
  int dilation = param.Dilations()[0];
  int input_width = param.Input()->dims()[3];
  int input_height = param.Input()->dims()[2];
  int output_width = param.Output()->dims()[3];
  int output_height = param.Output()->dims()[2];
Y
1. fix group logic for convolution op. 2. add pixel shuffle op for OpenCL. (#2178)  
Yanzhan Yang 已提交
62 
  int output_c = param.Output()->dims()[1];
N
optimize conv7*7, compute 2 point in 1 thread. test=develop (#2028)  
NazgulLee 已提交
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 
  int filter_channel = param.Filter()->dims()[1];
  int input_channel = param.Input()->dims()[1];
  //
  //    DLOG << " c block " << c_block;
  //    DLOG << " w " << w;
  //    DLOG << " nh " << nh;
  //    DLOG << " stride " << stride;
  //    DLOG << " offset " << offset;
  //    DLOG << " input_c " << input_c;
  //    DLOG << " dilation " << dilation;
  //    DLOG << " input width " << input_width;
  //    DLOG << " input height " << input_height;
  //    DLOG << " output width " << output_width;
  //    DLOG << " output height " << output_height;
  //    DLOG << " input dim " << param.Input()->dims();
  //    DLOG << " output dim " << param.Output()->dims();
  //    DLOG << " filter dim " << param.Filter()->dims();

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &nh);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &offset);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_c);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);

  if (param.Filter()->dims()[2] == 3 && param.Filter()->dims()[3] == 3) {
    if (filter_channel != input_channel) {
      if (filter_channel != 1) {
        status = clSetKernelArg(kernel, index++, sizeof(int), &filter_channel);
        CL_CHECK_ERRORS(status);
        int has_group = 1;
        status = clSetKernelArg(kernel, index++, sizeof(int), &has_group);
        CL_CHECK_ERRORS(status);
      }
    } else {
      status = clSetKernelArg(kernel, index++, sizeof(int), &filter_channel);
      CL_CHECK_ERRORS(status);
      int has_group = 0;
      status = clSetKernelArg(kernel, index++, sizeof(int), &has_group);
      CL_CHECK_ERRORS(status);
    }
  }
  //  DLOG<<"default_work_size"<<default_work_size[0]<<"
  //  "<<default_work_size[1]<<" "<<default_work_size[2];
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
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 
  auto kernel_work_size = cl_helper->KernelWorkSize(kernel);
  auto tmp0 = default_work_size.data()[0];
  auto tmp1 = default_work_size.data()[1];
  auto tmp2 = default_work_size.data()[2];
  int max_work_size = static_cast<const uint32_t>(kernel_work_size);
  if (preferred_lws_divisor > 1) {
    max_work_size /= preferred_lws_divisor;
  }
  if (preferred_lws > 0 && preferred_lws <= max_work_size) {
    max_work_size = preferred_lws;
  }
  while (tmp1 > max_work_size && max_work_size > 0) {
    tmp1 = tmp1 % 2 == 0 ? tmp1 / 2 : 1;
  }
  while (tmp2 * tmp1 > max_work_size && max_work_size > 0) {
    tmp2 = tmp2 % 2 == 0 ? tmp2 / 2 : 1;
  }
  while (tmp0 * tmp1 * tmp2 > max_work_size && max_work_size > 0) {
    tmp0 = tmp0 % 2 == 0 ? tmp0 / 2 : 1;
  }
  const size_t local_work_size[3] = {static_cast<const uint32_t>(tmp0),
                                     static_cast<const uint32_t>(tmp1),
                                     static_cast<const uint32_t>(tmp2)};
  if (max_work_size > 0 && use_lws) {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
N
optimize conv7*7, compute 2 point in 1 thread. test=develop (#2028)  
NazgulLee 已提交
193 194 195 
  CL_CHECK_ERRORS(status);
}
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
196 
void ConvAddBnRelu(framework::CLHelper *cl_helper,
Z
adjust gpu code structure  
zhaojiaying01 已提交
197 
                   const ConvParam<GPU_CL> &param, bool ifRelu,
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
198 199 200 201 202 
                   const framework::CLImage *biase,
                   const framework::CLImage *new_scale,
                   const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
Z
adjust gpu code structure  
zhaojiaying01 已提交
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];
  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int offset = param.Offset();
  int input_c = reinterpret_cast<framework::CLImageConverterFolder *>(
                    param.Input()->Converter())
                    ->GetCBlock();
  int dilation = param.Dilations()[0];
  int input_width = param.Input()->dims()[3];
  int input_height = param.Input()->dims()[2];
  int output_width = param.Output()->dims()[3];
  int output_height = param.Output()->dims()[2];
Y
1. fix group logic for convolution op. 2. add pixel shuffle op for OpenCL. (#2178)  
Yanzhan Yang 已提交
220 
  int output_c = param.Output()->dims()[1];
Y
support group conv in opencl (#1776)  
Yanzhan Yang 已提交
221 222 
  int filter_channel = param.Filter()->dims()[1];
  int input_channel = param.Input()->dims()[1];
Z
adjust gpu code structure  
zhaojiaying01 已提交
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 

  //  DLOG << " c block " << c_block;
  //  DLOG << " w " << w;
  //  DLOG << " nh " << nh;
  //  DLOG << " stride " << stride;
  //  DLOG << " offset " << offset;
  //  DLOG << " input_c " << input_c;
  //  DLOG << " dilation " << dilation;
  //  DLOG << " input width " << input_width;
  //  DLOG << " input height " << input_height;
  //  DLOG << " output width " << output_width;
  //  DLOG << " output height " << output_height;
  //  DLOG << " input dim " << param.Input()->dims();
  //  DLOG << " output dim " << param.Output()->dims();
  //  DLOG << " filter dim " << param.Filter()->dims();

  cl_int status;
  int index = 0;

  if (param.Filter()->dims()[2] == 1 && param.Filter()->dims()[3] == 1) {
    status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
    CL_CHECK_ERRORS(status);

    int maped_w = maptofactor(w, 4);
    status = clSetKernelArg(kernel, index++, sizeof(int), &maped_w);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &nh);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
    CL_CHECK_ERRORS(status);

    if (biase) {
      auto bias_mem = biase->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
      CL_CHECK_ERRORS(status);
    }

    if (new_scale && new_bias) {
      auto new_scale_mem = new_scale->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
      CL_CHECK_ERRORS(status);

      auto new_bias_mem = new_bias->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
      CL_CHECK_ERRORS(status);
    }

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &offset);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_c);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &w);
    CL_CHECK_ERRORS(status);

    const size_t work_size[3] = {
        static_cast<const uint32_t>(default_work_size.data()[0]),
        static_cast<const uint32_t>(maped_w),
        static_cast<const uint32_t>(default_work_size.data()[2])};
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 
    auto kernel_work_size = cl_helper->KernelWorkSize(kernel);
    auto tmp0 = work_size[0];
    auto tmp1 = work_size[1];
    auto tmp2 = work_size[2];
    int max_work_size = static_cast<const uint32_t>(kernel_work_size);
    if (preferred_lws_divisor > 1) {
      max_work_size /= preferred_lws_divisor;
    }
    if (preferred_lws > 0 && preferred_lws <= max_work_size) {
      max_work_size = preferred_lws;
    }
    while (tmp1 > max_work_size && max_work_size > 0) {
      tmp1 = tmp1 % 2 == 0 ? tmp1 / 2 : 1;
    }
    while (tmp2 * tmp1 > max_work_size && max_work_size > 0) {
      tmp2 = tmp2 % 2 == 0 ? tmp2 / 2 : 1;
    }
    while (tmp0 * tmp1 * tmp2 > max_work_size && max_work_size > 0) {
      tmp0 = tmp0 % 2 == 0 ? tmp0 / 2 : 1;
    }
    const size_t local_work_size[3] = {static_cast<const uint32_t>(tmp0),
                                       static_cast<const uint32_t>(tmp1),
                                       static_cast<const uint32_t>(tmp2)};
    if (max_work_size > 0 && use_lws) {
S
add opencl slidingwindow3*3s1 (#1799)  
StarryRain 已提交
334 335 336 337 338 339 340 341 
      status = clEnqueueNDRangeKernel(cl_helper->CLCommandQueue(), kernel,
                                      default_work_size.size(), NULL, work_size,
                                      local_work_size, 0, NULL, NULL);
    } else {
      status = clEnqueueNDRangeKernel(cl_helper->CLCommandQueue(), kernel,
                                      default_work_size.size(), NULL, work_size,
                                      NULL, 0, NULL, NULL);
    }
Z
adjust gpu code structure  
zhaojiaying01 已提交
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 
    CL_CHECK_ERRORS(status);
  } else {
    status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &w);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &nh);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
    CL_CHECK_ERRORS(status);

    if (biase) {
      auto bias_mem = biase->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
      CL_CHECK_ERRORS(status);
    }

    if (new_scale && new_bias) {
      auto new_scale_mem = new_scale->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
      CL_CHECK_ERRORS(status);

      auto new_bias_mem = new_bias->GetCLImage();
      status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
      CL_CHECK_ERRORS(status);
    }

    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &offset);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_c);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
    CL_CHECK_ERRORS(status);

    status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
    CL_CHECK_ERRORS(status);
Y
support group conv in opencl (#1776)  
Yanzhan Yang 已提交
402 
    if (param.Filter()->dims()[2] == 3 && param.Filter()->dims()[3] == 3) {
S
fix “CL_INVALID_KERNEL_ARGS ” error， test=develop (#2213)  
StarryRain 已提交
403 404 405 406 407 
      // normal conv
      if (param.Filter()->dims()[0] == param.Output()->dims()[1] &&
          param.Filter()->dims()[1] == param.Input()->dims()[1]) {
        status = clSetKernelArg(kernel, index++, sizeof(int), &output_c);
        CL_CHECK_ERRORS(status);
Y
1. fix group logic for convolution op. 2. add pixel shuffle op for OpenCL. (#2178)  
Yanzhan Yang 已提交
408 409 
        status = clSetKernelArg(kernel, index++, sizeof(int), &filter_channel);
        CL_CHECK_ERRORS(status);
S
fix “CL_INVALID_KERNEL_ARGS ” error， test=develop (#2213)  
StarryRain 已提交
410 
        int group = 1;
Y
1. fix group logic for convolution op. 2. add pixel shuffle op for OpenCL. (#2178)  
Yanzhan Yang 已提交
411 412 
        status = clSetKernelArg(kernel, index++, sizeof(int), &group);
        CL_CHECK_ERRORS(status);
S
fix “CL_INVALID_KERNEL_ARGS ” error， test=develop (#2213)  
StarryRain 已提交
413 414 415 416 
      } else if (!(param.Filter()->dims()[0] == param.Input()->dims()[1] &&
                   param.Filter()->dims()[1] == 1)) {  // not depwise
        status = clSetKernelArg(kernel, index++, sizeof(int), &output_c);
        CL_CHECK_ERRORS(status);
Y
support group conv in opencl (#1776)  
Yanzhan Yang 已提交
417 418 
        status = clSetKernelArg(kernel, index++, sizeof(int), &filter_channel);
        CL_CHECK_ERRORS(status);
S
fix “CL_INVALID_KERNEL_ARGS ” error， test=develop (#2213)  
StarryRain 已提交
419 
        int group = input_channel / filter_channel;
Y
1. fix group logic for convolution op. 2. add pixel shuffle op for OpenCL. (#2178)  
Yanzhan Yang 已提交
420 
        status = clSetKernelArg(kernel, index++, sizeof(int), &group);
Y
support group conv in opencl (#1776)  
Yanzhan Yang 已提交
421 422 423 424 
        CL_CHECK_ERRORS(status);
      }
    }
Z
adjust gpu code structure  
zhaojiaying01 已提交
425 
    status = clEnqueueNDRangeKernel(
J
optimize GPU conv performance and structure (#1595)  
Jiaying Zhao 已提交
426 
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
Z
adjust gpu code structure  
zhaojiaying01 已提交
427 428 429 430 431 
        default_work_size.data(), NULL, 0, NULL, NULL);
    CL_CHECK_ERRORS(status);
  }
}
Z
add OpenCL depthwise3x3s1 kernel  
zhaojiaying01 已提交
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 
void DWConvAddBnRelu(framework::CLHelper *cl_helper,
                     const ConvParam<GPU_CL> &param, bool ifRelu,
                     const framework::CLImage *biase,
                     const framework::CLImage *new_scale,
                     const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];
  int w_blk_size = 2;
  int w_blk = (w + w_blk_size - 1) / w_blk_size;

  default_work_size[1] = w_blk;
  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int pad = param.Paddings()[0];
  int dilation = param.Dilations()[0];

  int input_channel = param.Input()->dims()[1];
  int input_height = param.Input()->dims()[2];
  int input_width = param.Input()->dims()[3];

  int output_height = param.Output()->dims()[2];
  int output_width = param.Output()->dims()[3];

  //  DLOG << " w " << w;
  //  DLOG << " nh " << nh;
  //  DLOG << " stride " << stride;
  //  DLOG << " dilation " << dilation;
  //  DLOG << " input width " << input_width;
  //  DLOG << " input height " << input_height;
  //  DLOG << " output width " << output_width;
  //  DLOG << " output height " << output_height;
  //  DLOG << " input dim " << param.Input()->dims();
  //  DLOG << " output dim " << param.Output()->dims();
  //  DLOG << " filter dim " << param.Filter()->dims();

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &nh);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 
  auto kernel_work_size = cl_helper->KernelWorkSize(kernel);
  auto tmp0 = default_work_size.data()[0];
  auto tmp1 = default_work_size.data()[1];
  auto tmp2 = default_work_size.data()[2];
  int max_work_size = static_cast<const uint32_t>(kernel_work_size);
  if (preferred_lws_divisor > 1) {
    max_work_size /= preferred_lws_divisor;
  }
  if (preferred_lws > 0 && preferred_lws <= max_work_size) {
    max_work_size = preferred_lws;
  }
  while (tmp1 > max_work_size && max_work_size > 0) {
    tmp1 = tmp1 % 2 == 0 ? tmp1 / 2 : 1;
  }
  while (tmp2 * tmp1 > max_work_size && max_work_size > 0) {
    tmp2 = tmp2 % 2 == 0 ? tmp2 / 2 : 1;
  }
  while (tmp0 * tmp1 * tmp2 > max_work_size && max_work_size > 0) {
    tmp0 = tmp0 % 2 == 0 ? tmp0 / 2 : 1;
  }
  const size_t local_work_size[3] = {static_cast<const uint32_t>(tmp0),
                                     static_cast<const uint32_t>(tmp1),
                                     static_cast<const uint32_t>(tmp2)};
  if (max_work_size > 0 && use_lws) {
S
add opencl slidingwindow3*3s1 (#1799)  
StarryRain 已提交
558 559 560 561 562 563 564 565 566 
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
Z
add OpenCL depthwise3x3s1 kernel  
zhaojiaying01 已提交
567 568 569 
  CL_CHECK_ERRORS(status);
}
S
add opencl slidingwindow3*3s1 (#1799)  
StarryRain 已提交
570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 
void SWConvAddBnRelu(framework::CLHelper *cl_helper,
                     const ConvParam<GPU_CL> &param, bool ifRelu,
                     const framework::CLImage *biase,
                     const framework::CLImage *new_scale,
                     const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];

  int w_blk_size = 5;
  int w_blk = (w + w_blk_size - 1) / w_blk_size;
  default_work_size[1] = w_blk;

  int h_blk_size = 1;
  int h_blk = (nh + h_blk_size - 1) / h_blk_size;
  default_work_size[2] = h_blk;

  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int pad = param.Paddings()[0];
  int dilation = param.Dilations()[0];

  int input_channel = param.Input()->dims()[1];
  int input_height = param.Input()->dims()[2];
  int input_width = param.Input()->dims()[3];
  int output_height = param.Output()->dims()[2];
  int output_width = param.Output()->dims()[3];

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &h_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 
  auto kernel_work_size = cl_helper->KernelWorkSize(kernel);
  auto tmp0 = default_work_size.data()[0];
  auto tmp1 = default_work_size.data()[1];
  auto tmp2 = default_work_size.data()[2];
  int max_work_size = static_cast<const uint32_t>(kernel_work_size);
  if (preferred_lws_divisor > 1) {
    max_work_size /= preferred_lws_divisor;
  }
  if (preferred_lws > 0 && preferred_lws <= max_work_size) {
    max_work_size = preferred_lws;
  }
  while (tmp1 > max_work_size && max_work_size > 0) {
    tmp1 = tmp1 % 2 == 0 ? tmp1 / 2 : 1;
  }
  while (tmp2 * tmp1 > max_work_size && max_work_size > 0) {
    tmp2 = tmp2 % 2 == 0 ? tmp2 / 2 : 1;
  }
  while (tmp0 * tmp1 * tmp2 > max_work_size && max_work_size > 0) {
    tmp0 = tmp0 % 2 == 0 ? tmp0 / 2 : 1;
  }
  const size_t local_work_size[3] = {static_cast<const uint32_t>(tmp0),
                                     static_cast<const uint32_t>(tmp1),
                                     static_cast<const uint32_t>(tmp2)};
  if (max_work_size > 0 && use_lws) {
S
add opencl slidingwindow3*3s1 (#1799)  
StarryRain 已提交
686 687 688 689 690 691 692 693 694 695 
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
  CL_CHECK_ERRORS(status);
}
S
add opencl depthwise_conv_trans_op (#1949)  
StarryRain 已提交
696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 

void DWConvTransposeAddBnRelu(framework::CLHelper *cl_helper,
                              const ConvTransposeParam<GPU_CL> &param,
                              bool ifRelu, const framework::CLImage *biase,
                              const framework::CLImage *new_scale,
                              const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];

  int w_blk_size = 1;
  int w_blk = (w + w_blk_size - 1) / w_blk_size;
  default_work_size[1] = w_blk;

  int h_blk_size = 1;
  int h_blk = (nh + h_blk_size - 1) / h_blk_size;
  default_work_size[2] = h_blk;

  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int pad = param.Paddings()[0];
  int dilation = param.Dilations()[0];

  int input_channel = param.Input()->dims()[1];
  int input_height = param.Input()->dims()[2];
  int input_width = param.Input()->dims()[3];

  int output_height = param.Output()->dims()[2];
  int output_width = param.Output()->dims()[3];

  int filter_height = param.Filter()->dims()[2];
  int filter_width = param.Filter()->dims()[3];

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &h_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_height);
  CL_CHECK_ERRORS(status);

  if (default_work_size.data()[1] % 60 == 0 && use_lws) {
    const size_t local_work_size[3] = {static_cast<const uint32_t>(1),
                                       static_cast<const uint32_t>(60),
                                       static_cast<const uint32_t>(1)};
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
  CL_CHECK_ERRORS(status);
}
S
add opencl conv_trans_op (#2031)  
StarryRain 已提交
815 816 817 818 819 
void ConvTransposeAddBnRelu_b(framework::CLHelper *cl_helper,
                              const ConvTransposeParam<GPU_CL> &param,
                              bool ifRelu, const framework::CLImage *biase,
                              const framework::CLImage *new_scale,
                              const framework::CLImage *new_bias) {
S
add opencl depthwise_conv_trans_op (#1949)  
StarryRain 已提交
820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 
  auto kernel = cl_helper->KernelAt(0);
  const auto *input = param.Input();
  auto *output = param.Output();
  auto *filter = param.Filter();
  const int n = input->dims()[0];
  const int input_c = input->dims()[1];
  const int input_c_block = (input_c + 3) / 4;
  const int input_width = input->dims()[3];
  const int input_height = input->dims()[2];
  const int output_c = output->dims()[1];
  const int output_c_block = (output_c + 3) / 4;
  const int output_width = output->dims()[3];
  const int output_height = output->dims()[2];

  auto inputImage = input->GetCLImage();
  auto outputImage = output->GetCLImage();
  auto filterImage = filter->GetCLImage();

  cl_int status;
  status = clSetKernelArg(kernel, 0, sizeof(int), &input_c_block);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 1, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 2, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 3, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 4, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 5, sizeof(cl_mem), &inputImage);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 6, sizeof(cl_mem), &filterImage);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, 7, sizeof(cl_mem), &outputImage);
  CL_CHECK_ERRORS(status);

  const size_t work_size[3] = {(size_t)output_c_block, (size_t)input_width,
                               (size_t)(n * input_height)};

  DLOG << "conv transpose " << input_c_block << input_width << input_height
       << output_width << output_height << work_size[0] << work_size[1]
       << work_size[2];

  clEnqueueNDRangeKernel(cl_helper->CLCommandQueue(), kernel, 3, NULL,
                         work_size, NULL, 0, NULL, NULL);
}
S
add opencl conv_trans_op (#2031)  
StarryRain 已提交
866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 
void ConvTransposeAddBnRelu(framework::CLHelper *cl_helper,
                            const ConvTransposeParam<GPU_CL> &param,
                            bool ifRelu, const framework::CLImage *biase,
                            const framework::CLImage *new_scale,
                            const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];

  int w_blk_size = 1;
  int w_blk = (w + w_blk_size - 1) / w_blk_size;
  default_work_size[1] = w_blk;

  int h_blk_size = 1;
  int h_blk = (nh + h_blk_size - 1) / h_blk_size;
  default_work_size[2] = h_blk;

  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int pad = param.Paddings()[0];
  int dilation = param.Dilations()[0];

  int input_channel = param.Input()->dims()[1];
  int input_height = param.Input()->dims()[2];
  int input_width = param.Input()->dims()[3];

  int output_height = param.Output()->dims()[2];
  int output_width = param.Output()->dims()[3];

  int filter_height = param.Filter()->dims()[2];
  int filter_width = param.Filter()->dims()[3];

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &h_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);
S
Optimal OPENCL conv_trans_3x3s2 Performance (#2015)  
StarryRain 已提交
920
S
add opencl conv_trans_op (#2031)  
StarryRain 已提交
921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 
  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_height);
  CL_CHECK_ERRORS(status);

  if (default_work_size.data()[1] % 60 == 0 && use_lws) {
    const size_t local_work_size[3] = {static_cast<const uint32_t>(1),
                                       static_cast<const uint32_t>(60),
                                       static_cast<const uint32_t>(1)};
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
  CL_CHECK_ERRORS(status);
}
S
Optimal OPENCL conv_trans_3x3s2 Performance (#2015)  
StarryRain 已提交
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 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 1080 1081 1082 1083 1084 1085 
void ConvTranspose3x3s2AddBnRelu(framework::CLHelper *cl_helper,
                                 const ConvTransposeParam<GPU_CL> &param,
                                 bool ifRelu, const framework::CLImage *biase,
                                 const framework::CLImage *new_scale,
                                 const framework::CLImage *new_bias) {
  auto kernel = cl_helper->KernelAt(0);
  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];

  int w_blk_size = 5;
  int w_blk = (w + w_blk_size - 1 + 5) / w_blk_size / 2 * 2;
  default_work_size[1] = w_blk;

  int h_blk_size = 1;
  int h_blk = (nh + h_blk_size - 1) / h_blk_size;
  default_work_size[2] = h_blk;

  auto input = param.Input()->GetCLImage();
  auto filter = param.Filter()->GetCLImage();

  auto output = param.Output()->GetCLImage();
  int stride = param.Strides()[0];
  int pad = param.Paddings()[0];
  int dilation = param.Dilations()[0];

  int input_channel = param.Input()->dims()[1];
  int input_height = param.Input()->dims()[2];
  int input_width = param.Input()->dims()[3];

  int output_height = param.Output()->dims()[2];
  int output_width = param.Output()->dims()[3];

  int filter_height = param.Filter()->dims()[2];
  int filter_width = param.Filter()->dims()[3];

  cl_int status;
  int index = 0;

  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &h_blk);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  if (biase) {
    auto bias_mem = biase->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
    CL_CHECK_ERRORS(status);
  }

  if (new_scale && new_bias) {
    auto new_scale_mem = new_scale->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
    CL_CHECK_ERRORS(status);

    auto new_bias_mem = new_bias->GetCLImage();
    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
    CL_CHECK_ERRORS(status);
  }

  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
  CL_CHECK_ERRORS(status);
  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_height);
  CL_CHECK_ERRORS(status);
S
get opencl kernel_work_group_size (#2123)  
StarryRain 已提交
1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 
  auto kernel_work_size = cl_helper->KernelWorkSize(kernel);
  auto tmp0 = default_work_size.data()[0];
  auto tmp1 = default_work_size.data()[1];
  auto tmp2 = default_work_size.data()[2];
  int max_work_size = static_cast<const uint32_t>(kernel_work_size);
  if (preferred_lws_divisor > 1) {
    max_work_size /= preferred_lws_divisor;
  }
  if (preferred_lws > 0 && preferred_lws <= max_work_size) {
    max_work_size = preferred_lws;
  }
  while (tmp1 > max_work_size && max_work_size > 0) {
    tmp1 = tmp1 % 2 == 0 ? tmp1 / 2 : 1;
  }
  while (tmp2 * tmp1 > max_work_size && max_work_size > 0) {
    tmp2 = tmp2 % 2 == 0 ? tmp2 / 2 : 1;
  }
  while (tmp0 * tmp1 * tmp2 > max_work_size && max_work_size > 0) {
    tmp0 = tmp0 % 2 == 0 ? tmp0 / 2 : 1;
  }
  const size_t local_work_size[3] = {static_cast<const uint32_t>(tmp0),
                                     static_cast<const uint32_t>(tmp1),
                                     static_cast<const uint32_t>(tmp2)};
  if (max_work_size > 0 && use_lws) {
S
Optimal OPENCL conv_trans_3x3s2 Performance (#2015)  
StarryRain 已提交
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), local_work_size, 0, NULL, NULL);
  } else {
    status = clEnqueueNDRangeKernel(
        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
        default_work_size.data(), NULL, 0, NULL, NULL);
  }
  CL_CHECK_ERRORS(status);
}
Z
adjust gpu code structure  
zhaojiaying01 已提交
1120 
}  // namespace operators
Z
format code style  
zhaojiaying01 已提交
1121 
}  // namespace paddle_mobile