interpolate_v2_op.cu 86.7 KB
Newer Older
X
xiaoting 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2018 PaddlePaddle Authors. 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 <algorithm>
#include <string>
#include "paddle/fluid/operators/interpolate_v2_op.h"
15 16 17
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
#include "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
#include "paddle/fluid/platform/device/gpu/gpu_primitives.h"
18
#include "paddle/fluid/platform/fast_divmod.h"
19
#include "paddle/pten/kernels/funcs/math_cuda_utils.h"
X
xiaoting 已提交
20 21 22 23 24

namespace paddle {
namespace operators {

using framework::Tensor;
25
using platform::FastDivMod;
X
xiaoting 已提交
26 27
using DataLayout = framework::DataLayout;

28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
static inline int GetLastPow2(int n) {
  n |= (n >> 1);
  n |= (n >> 2);
  n |= (n >> 4);
  n |= (n >> 8);
  n |= (n >> 16);
  return std::max(1, n - (n >> 1));
}

inline platform::GpuLaunchConfig GetGpuLaunchConfig3D(
    const platform::CUDADeviceContext& context, int num_img, int height,
    int width) {
  const int kThreadsPerBlock = 256;
  int max_threads_per_block = context.GetMaxThreadsPerBlock();  // 1024
  int max_threads = std::min(kThreadsPerBlock, max_threads_per_block);

  int block_x = std::min(GetLastPow2(width), max_threads);
  int block_y = std::min(GetLastPow2(height), max_threads / block_x);
  int block_z = std::min(num_img, max_threads / block_x / block_y);

W
Wilber 已提交
48 49 50
  auto max_grid_dim = context.GetCUDAMaxGridDimSize();
  int grid_x = std::min<int>(max_grid_dim[0], platform::DivUp(width, block_x));
  int grid_y = std::min<int>(max_grid_dim[1], platform::DivUp(height, block_y));
51
  int grid_z =
W
Wilber 已提交
52
      std::min<int>(max_grid_dim[2], platform::DivUp(num_img, block_z * 4));
53 54 55 56 57 58 59 60 61

  const int capability = context.GetComputeCapability();
  platform::GpuLaunchConfig config;
  config.compute_capability = capability;
  config.thread_per_block = dim3(block_x, block_y, block_z);
  config.block_per_grid = dim3(grid_x, grid_y, grid_z);
  return config;
}

62 63 64 65 66 67 68 69 70 71 72
template <typename T>
__forceinline__ __device__ void PreCalculatorForLinearInterpInputIndex(
    int* in_img_idx, int* w_id, T* w1lambda, T* w2lambda, T src_w,
    const int in_img_w) {
  src_w = (src_w > 0) ? src_w : 0.f;
  *in_img_idx = static_cast<int>(src_w);
  *w_id = (*in_img_idx < in_img_w - 1) ? 1 : 0;
  *w1lambda = src_w - *in_img_idx;
  *w2lambda = 1.f - *w1lambda;
}

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
struct FastDivModForInterpolate {
 public:
  FastDivMod channels_div;
  FastDivMod output_w_div;
  FastDivMod output_wc_div;

  explicit HOSTDEVICE FastDivModForInterpolate(const int channels,
                                               const int output_w,
                                               const int outout_wc)
      : channels_div(FastDivMod(channels)),
        output_w_div(FastDivMod(output_w)),
        output_wc_div(FastDivMod(outout_wc)) {}
};

template <typename T>
__global__ void KeNearestNeighborInterpNCHWFw(
    const T* in, const size_t in_img_h, const size_t in_img_w, T* out,
    const size_t out_img_h, const size_t out_img_w, const size_t nc,
    const float ratio_h, const float ratio_w, const bool align_corners) {
  int out_img_idx = threadIdx.x + blockIdx.x * blockDim.x;
  int out_img_idy = threadIdx.y + blockIdx.y * blockDim.y;
  int nc_id = threadIdx.z + blockIdx.z * blockDim.z;
  int nc_stride = blockDim.z * gridDim.z;

  // nearest_sampling by multiple read in_addr and write to out_addr
  int in_img_idx = (align_corners)
                       ? static_cast<int>(ratio_w * out_img_idx + 0.5)
                       : static_cast<int>(ratio_w * out_img_idx);
  int in_img_idy = (align_corners)
                       ? static_cast<int>(ratio_h * out_img_idy + 0.5)
                       : static_cast<int>(ratio_h * out_img_idy);

  int in_index = (nc_id * in_img_h + in_img_idy) * in_img_w + in_img_idx;
  int in_index_stride = nc_stride * in_img_h * in_img_w;

  int out_index = (nc_id * out_img_h + out_img_idy) * out_img_w + out_img_idx;
  int out_index_stride = nc_stride * out_img_h * out_img_w;

  // prevent from multiple threads writing
  if (out_img_idx < out_img_w && out_img_idy < out_img_h) {
    while (nc_id < nc) {
      out[out_index] = in[in_index];
      in_index += in_index_stride;
      out_index += out_index_stride;
      nc_id += nc_stride;
    }
  }
}

X
xiaoting 已提交
122 123 124 125 126 127
template <typename T>
__global__ void KeNearestNeighborInterpFw(
    const T* in, const size_t in_img_h, const size_t in_img_w,
    const size_t input_h, const size_t input_w, T* out, const size_t out_img_h,
    const size_t out_img_w, const size_t output_h, const size_t output_w,
    const size_t num_channels, const float ratio_h, const float ratio_w,
128
    const bool align_corners, FastDivModForInterpolate divmods) {
X
xiaoting 已提交
129 130 131
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
132 133 134
  int in_img_size = in_img_h * in_img_w;
  int out_img_size = out_img_h * out_img_w;

X
xiaoting 已提交
135
  for (; tid < nthreads; tid += stride) {
136 137 138
    auto out_id_divmod = divmods.output_w_div.Divmod(tid);
    int out_id_h = out_id_divmod.val[0];
    int out_id_w = out_id_divmod.val[1];
X
xiaoting 已提交
139

140 141 142 143 144
    int channel_id = divmods.channels_div.Divmod(tid).val[1];
    auto outimg_id_divmod = divmods.output_wc_div.Divmod(out_id_w);
    int out_img_idy = outimg_id_divmod.val[0];
    int out_img_idx =
        divmods.channels_div.Divmod(outimg_id_divmod.val[1]).val[0];
X
xiaoting 已提交
145 146 147 148 149 150 151 152

    int in_img_idy = (align_corners)
                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    int in_img_idx = (align_corners)
                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);

153 154
    out[tid] = in[out_id_h * input_w + in_img_idy * in_img_w * num_channels +
                  in_img_idx * num_channels + channel_id];
X
xiaoting 已提交
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
template <typename T>
__global__ void KeNearestNeighbor3DInterpFw(
    const T* in, const size_t in_img_d, const size_t in_img_h,
    const size_t in_img_w, const size_t input_h, const size_t input_w, T* out,
    const size_t out_img_d, const size_t out_img_h, const size_t out_img_w,
    const size_t output_h, const size_t output_w, const size_t num_channels,
    const float ratio_d, const float ratio_h, const float ratio_w,
    const bool align_corners, const DataLayout data_layout) {
  int nthreads = output_h * output_w;  // ncdhw
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idt, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idt = (out_id_w % out_img_size) / out_img_h / out_img_w;
      out_img_idy = ((out_id_w % out_img_size) / out_img_w) % out_img_h;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idt = out_id_w / (out_img_h * out_img_w * num_channels);
      out_img_idy = out_id_w % (out_img_h * out_img_w * num_channels) /
                    (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idt = (align_corners)
                         ? static_cast<int>(ratio_d * out_img_idt + 0.5)
                         : static_cast<int>(ratio_d * out_img_idt);

    int in_img_idy = (align_corners)
                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    int in_img_idx = (align_corners)
                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);

    if (data_layout == DataLayout::kNCHW) {
      out[tid] = in[out_id_h * input_w + channel_id * in_img_size +
                    in_img_idt * in_img_h * in_img_w + in_img_idy * in_img_w +
                    in_img_idx];
    } else {
      out[tid] = in[out_id_h * input_w +
                    in_img_idt * in_img_h * in_img_w * num_channels +
                    in_img_idy * in_img_w * num_channels +
                    in_img_idx * num_channels + channel_id];
    }
  }
}

X
xiaoting 已提交
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
template <typename T>
__global__ void KeNearestNeighborInterpBw(
    T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h,
    const size_t input_w, const T* out, const size_t out_img_h,
    const size_t out_img_w, const size_t output_h, const size_t output_w,
    const size_t num_channels, const float ratio_h, const float ratio_w,
    const bool align_corners, const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idy = (out_id_w % out_img_size) / out_img_w;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idy = out_id_w / (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idy = (align_corners)
                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    int in_img_idx = (align_corners)
                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);

    T* in_pos;
    if (data_layout == DataLayout::kNCHW) {
      in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
                   in_img_idy * in_img_w + in_img_idx];
    } else {
      in_pos = &in[out_id_h * input_w + in_img_idy * in_img_w * num_channels +
                   in_img_idx * num_channels + channel_id];
    }
    const T out_pos = out[out_id_h * output_w + out_id_w];
    platform::CudaAtomicAdd(in_pos, out_pos);
  }
}

260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316
template <typename T>
__global__ void KeNearestNeighbor3DInterpBw(
    T* in, const size_t in_img_d, const size_t in_img_h, const size_t in_img_w,
    const size_t input_h, const size_t input_w, const T* out,
    const size_t out_img_d, const size_t out_img_h, const size_t out_img_w,
    const size_t output_h, const size_t output_w, const size_t num_channels,
    const float ratio_d, const float ratio_h, const float ratio_w,
    const bool align_corners, const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idt, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idt = (out_id_w % out_img_size) / out_img_h / out_img_w;
      out_img_idy = ((out_id_w % out_img_size) / out_img_w) % out_img_h;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idt = out_id_w / (out_img_h * out_img_w * num_channels);
      out_img_idy = out_id_w % (out_img_h * out_img_w * num_channels) /
                    (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idt = (align_corners)
                         ? static_cast<int>(ratio_d * out_img_idt + 0.5)
                         : static_cast<int>(ratio_d * out_img_idt);
    int in_img_idy = (align_corners)
                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    int in_img_idx = (align_corners)
                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);

    T* in_pos;
    if (data_layout == DataLayout::kNCHW) {
      in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
                   in_img_idt * in_img_h * in_img_w + in_img_idy * in_img_w +
                   in_img_idx];
    } else {
      in_pos = &in[out_id_h * input_w +
                   in_img_idt * in_img_h * in_img_w * num_channels +
                   in_img_idy * in_img_w * num_channels +
                   in_img_idx * num_channels + channel_id];
    }
    const T out_pos = out[out_id_h * output_w + out_id_w];
    platform::CudaAtomicAdd(in_pos, out_pos);
  }
}

X
xiaoting 已提交
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
template <typename T>
__global__ void KeLinearInterpFw(const T* in, const size_t in_img_w,
                                 const size_t input_w, T* out,
                                 const size_t out_img_w, const size_t output_h,
                                 const size_t output_w,
                                 const size_t num_channels, const float ratio_w,
                                 const bool align_corners, const int align_mode,
                                 const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  bool align_flag = (align_mode == 0 && !align_corners);
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idx = align_flag
                         ? static_cast<int>(ratio_w * (out_img_idx + 0.5) - 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);
    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;  // w
    int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;  // w_id

    T src_w = ratio_w * (out_img_idx + 0.5) - 0.5;
    src_w = (src_w > 0) ? src_w : 0;
    T w1lambda =
        align_flag ? src_w - in_img_idx : ratio_w * out_img_idx - in_img_idx;
    T w2lambda = 1.f - w1lambda;

    if (data_layout == DataLayout::kNCHW) {
      const T* in_pos =
          &in[out_id_h * out_id_w + channel_id * in_img_size + in_img_idx];
      // linear interpolation
      out[out_id_h * output_w + out_id_w] =
          w2lambda * in_pos[0] + w1lambda * in_pos[w_id];

    } else {
      const T* in_pos =
          &in[out_id_h * input_w + in_img_idx * num_channels + channel_id];
      // linear interpolation
      out[out_id_h * output_w + out_id_w] =
          w2lambda * in_pos[0] + w1lambda * in_pos[w_id * num_channels];
    }
  }
}

template <typename T>
__global__ void KeLinearInterpBw(T* in, const size_t in_img_w,
                                 const size_t input_w, const T* out,
                                 const size_t out_img_w, const size_t output_h,
                                 const size_t output_w,
                                 const size_t num_channels, const T ratio_w,
                                 const bool align_corners, const int align_mode,
                                 const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  bool align_flag = (align_mode == 0 && !align_corners);
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idx = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5
                                : ratio_w * out_img_idx;
    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;  // w
    int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;  // w_id

    T src_w = ratio_w * (out_img_idx + 0.5) - 0.5;
    src_w = (src_w > 0) ? src_w : 0;
    T w1lambda =
        align_flag ? src_w - in_img_idx : ratio_w * out_img_idx - in_img_idx;
    T w2lambda = 1.f - w1lambda;

    T* in_pos;
    if (data_layout == DataLayout::kNCHW) {
      in_pos = &in[out_id_h * input_w + channel_id * in_img_size + in_img_idx];
    } else {
      in_pos = &in[out_id_h * input_w + in_img_idx * num_channels + channel_id];
    }
    const T* out_pos = &out[out_id_w];

    if (data_layout == DataLayout::kNCHW) {
      platform::CudaAtomicAdd(&in_pos[0], w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos[w_id], w1lambda * out_pos[0]);
    } else {
      platform::CudaAtomicAdd(&in_pos[0], w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos[w_id * num_channels],
                              w1lambda * out_pos[0]);
    }
  }
}

template <typename T>
431 432 433 434 435 436 437 438 439 440
__global__ void KeBilinearInterpNCHWFw(const T* in, const size_t in_img_h,
                                       const size_t in_img_w, T* out,
                                       const size_t out_img_h,
                                       const size_t out_img_w, const size_t nc,
                                       const float ratio_h, const float ratio_w,
                                       const T align_type_value) {
  int out_img_idx = threadIdx.x + blockIdx.x * blockDim.x;
  int out_img_idy = threadIdx.y + blockIdx.y * blockDim.y;
  int nc_id = threadIdx.z + blockIdx.z * blockDim.z;
  int nc_stride = blockDim.z * gridDim.z;
X
xiaoting 已提交
441

442 443 444 445
  int in_img_idx, in_img_idy, h_id, w_id;
  T h1lambda, w1lambda, h2lambda, w2lambda;
  T src_w = ratio_w * (out_img_idx + align_type_value) - align_type_value;
  T src_h = ratio_h * (out_img_idy + align_type_value) - align_type_value;
X
xiaoting 已提交
446

447 448 449 450
  PreCalculatorForLinearInterpInputIndex(&in_img_idx, &w_id, &w1lambda,
                                         &w2lambda, src_w, in_img_w);
  PreCalculatorForLinearInterpInputIndex(&in_img_idy, &h_id, &h1lambda,
                                         &h2lambda, src_h, in_img_h);
X
xiaoting 已提交
451

452 453
  int in_index = (nc_id * in_img_h + in_img_idy) * in_img_w + in_img_idx;
  int in_index_stride = nc_stride * in_img_h * in_img_w;
X
xiaoting 已提交
454

455 456
  int out_index = (nc_id * out_img_h + out_img_idy) * out_img_w + out_img_idx;
  int out_index_stride = nc_stride * out_img_h * out_img_w;
X
xiaoting 已提交
457

458 459 460 461 462
  // prevent from multiple threads writing
  if (out_img_idx < out_img_w && out_img_idy < out_img_h) {
    while (nc_id < nc) {
      const T* in_pos = &in[in_index];
      out[out_index] =
X
xiaoting 已提交
463 464 465 466
          h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[w_id]) +
          h1lambda * (w2lambda * in_pos[h_id * in_img_w] +
                      w1lambda * in_pos[h_id * in_img_w + w_id]);

467 468 469
      in_index += in_index_stride;
      out_index += out_index_stride;
      nc_id += nc_stride;
X
xiaoting 已提交
470 471 472 473 474
    }
  }
}

template <typename T>
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
__global__ void KeBilinearInterpFw(
    const T* in, const size_t in_img_h, const size_t in_img_w,
    const size_t input_h, const size_t input_w, T* out, const size_t out_img_h,
    const size_t out_img_w, const size_t output_h, const size_t output_w,
    const size_t num_channels, const float ratio_h, const float ratio_w,
    const T align_type_value, FastDivModForInterpolate divmods) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;

  for (; tid < nthreads; tid += stride) {
    auto out_id_divmod = divmods.output_w_div.Divmod(tid);
    int out_id_h = out_id_divmod.val[0];
    int out_id_w = out_id_divmod.val[1];

    int channel_id = divmods.channels_div.Divmod(tid).val[1];
    auto outimg_id_divmod = divmods.output_wc_div.Divmod(out_id_w);
    int out_img_idy = outimg_id_divmod.val[0];
    int out_img_idx =
        divmods.channels_div.Divmod(outimg_id_divmod.val[1]).val[0];

    int in_img_idx, in_img_idy, h_id, w_id;
    T h1lambda, w1lambda, h2lambda, w2lambda;
    T src_w = ratio_w * (out_img_idx + align_type_value) - align_type_value;
    T src_h = ratio_h * (out_img_idy + align_type_value) - align_type_value;

    PreCalculatorForLinearInterpInputIndex(&in_img_idx, &w_id, &w1lambda,
                                           &w2lambda, src_w, in_img_w);
    PreCalculatorForLinearInterpInputIndex(&in_img_idy, &h_id, &h1lambda,
                                           &h2lambda, src_h, in_img_h);

    // bilinear interpolation
    const T* in_pos =
        &in[out_id_h * input_w + in_img_idy * in_img_w * num_channels +
            in_img_idx * num_channels + channel_id];
    out[tid] =
        h2lambda *
            (w2lambda * in_pos[0] + w1lambda * in_pos[w_id * num_channels]) +
        h1lambda *
            (w2lambda * in_pos[h_id * in_img_w * num_channels] +
             w1lambda *
                 in_pos[h_id * in_img_w * num_channels + w_id * num_channels]);
  }
518
}
X
xiaoting 已提交
519

520 521 522 523 524 525 526 527 528 529 530 531 532
/* Calculate the minimum of partial elements in a block */
template <typename T>
__inline__ __device__ T PartialBlockMin(T val, size_t threads_num_in_block,
                                        unsigned mask) {
  __shared__ T shared[WARP_SIZE];
  __shared__ T shared_last_val;
  __shared__ int shared_last_idx;
  int lane = threadIdx.x & 0x1f;
  int wid = threadIdx.x >> 5;
  int threshold = (threads_num_in_block & (-WARP_SIZE));

  if (threadIdx.x < threshold) {
    shared_last_idx = (threshold >> 5) - 1;
533
    val = pten::funcs::warpReduceMin(val, mask);
534 535
    if (lane == 0) {
      shared[wid] = val;
X
xiaoting 已提交
536
    }
537 538 539 540 541 542 543
  } else {
    shared_last_val = std::numeric_limits<T>::max();
    platform::CudaAtomicMin(&shared_last_val, val);
    shared[wid] = shared_last_val;
    shared_last_idx = wid;
  }
  __syncthreads();
X
xiaoting 已提交
544

545 546 547
  if (threadIdx.x < threshold) {
    val = (lane <= shared_last_idx) ? shared[lane]
                                    : std::numeric_limits<T>::max();
548
    val = pten::funcs::warpReduceMin(val, mask);
549 550 551 552 553 554 555 556
    shared_last_val = val;
  }
  __syncthreads();
  if (threadIdx.x >= threshold) {
    val = shared_last_val;
  }
  return val;
}
X
xiaoting 已提交
557

558 559 560 561 562 563 564 565 566 567 568
template <typename T>
__global__ void KeBilinearInterpBwShareMemory(
    T* in, const int in_h, const int in_w, const T* __restrict__ out,
    const int out_h, const int out_w, const int n, const int num_channels,
    float ratio_h, float ratio_w, const T align_type_value, bool is_nchw) {
  __shared__ T s_data[2][1024];
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  int in_chw = in_h * in_w * num_channels;
  int out_chw = num_channels * out_h * out_w;
  int nthreads = n * out_chw;
X
xiaoting 已提交
569

570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / out_chw;
    int out_id_w = tid % out_chw;
    const int in_img_size = in_h * in_w;
    const int out_img_size = out_h * out_w;
    T value = out[out_id_h * out_chw + out_id_w];

    int channel_id = out_id_w / out_img_size;
    int out_img_idy = (out_id_w % out_img_size) / out_w;
    int out_img_idx = tid % out_w;

    int in_img_idx, in_img_idy, w_id, h_id;
    T w1lambda, h1lambda, w2lambda, h2lambda;
    T src_w = ratio_w * (out_img_idx + align_type_value) - align_type_value;
    T src_h = ratio_h * (out_img_idy + align_type_value) - align_type_value;
585 586 587 588 589

    PreCalculatorForLinearInterpInputIndex(&in_img_idx, &w_id, &w1lambda,
                                           &w2lambda, src_w, in_w);
    PreCalculatorForLinearInterpInputIndex(&in_img_idy, &h_id, &h1lambda,
                                           &h2lambda, src_h, in_h);
590 591 592 593 594 595 596 597 598 599 600 601

    // top_left_index is just input_index.
    int input_index = out_id_h * in_chw + channel_id * in_img_size +
                      in_img_idy * in_w + in_img_idx;
    int top_right_index = input_index + w_id;
    int bot_left_index = input_index + h_id * in_w;
    int bot_right_index = input_index + h_id * in_w + w_id;
    int in_top_min_index, in_bot_min_index;

    s_data[0][threadIdx.x] = 0.f;
    s_data[1][threadIdx.x] = 0.f;
    int remain = nthreads - (tid & (-blockDim.x));
602 603 604 605
    int in_top_max_index =
        pten::funcs::blockReduceMax(top_right_index, FINAL_MASK);
    int in_bot_max_index =
        pten::funcs::blockReduceMax(bot_right_index, FINAL_MASK);
606 607

    if (remain > blockDim.x) {
608 609 610
      in_top_min_index = pten::funcs::blockReduceMin(input_index, FINAL_MASK);
      in_bot_min_index =
          pten::funcs::blockReduceMin(bot_left_index, FINAL_MASK);
X
xiaoting 已提交
611
    } else {
612 613
      in_top_min_index = PartialBlockMin(input_index, remain, FINAL_MASK);
      in_bot_min_index = PartialBlockMin(bot_left_index, remain, FINAL_MASK);
X
xiaoting 已提交
614
    }
615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634
    int upper_limit_share_idx = (in_top_max_index - in_top_min_index) >
                                        (in_bot_max_index - in_bot_min_index)
                                    ? (in_top_max_index - in_top_min_index)
                                    : (in_bot_max_index - in_bot_min_index);
    if (h_id != 0) {
      platform::CudaAtomicAdd(&s_data[0][input_index - in_top_min_index],
                              h2lambda * w2lambda * value);
      platform::CudaAtomicAdd(&s_data[0][top_right_index - in_top_min_index],
                              h2lambda * w1lambda * value);
      platform::CudaAtomicAdd(&s_data[1][bot_left_index - in_bot_min_index],
                              h1lambda * w2lambda * value);
      platform::CudaAtomicAdd(&s_data[1][bot_right_index - in_bot_min_index],
                              h1lambda * w1lambda * value);
    } else {
      platform::CudaAtomicAdd(&s_data[0][top_right_index - in_top_min_index],
                              (h2lambda + h1lambda) * w1lambda * value);
      platform::CudaAtomicAdd(&s_data[1][bot_left_index - in_bot_min_index],
                              (h1lambda + h2lambda) * w2lambda * value);
    }
    __syncthreads();
X
xiaoting 已提交
635

636 637 638 639 640 641 642 643
    if (threadIdx.x <= upper_limit_share_idx) {
      platform::CudaAtomicAdd(&in[in_top_min_index + threadIdx.x],
                              s_data[0][threadIdx.x]);
      platform::CudaAtomicAdd(&in[in_bot_min_index + threadIdx.x],
                              s_data[1][threadIdx.x]);
    }
  }
}
X
xiaoting 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
template <typename T>
__global__ void KeBilinearInterpBw(T* in, const int in_h, const int in_w,
                                   const T* __restrict__ out, const int out_h,
                                   const int out_w, const int n,
                                   const int num_channels, float ratio_h,
                                   float ratio_w, const T align_type_value,
                                   bool is_nchw) {
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  int in_chw = in_h * in_w * num_channels;
  int out_chw = num_channels * out_h * out_w;
  int nthreads = n * out_chw;

  if (is_nchw) {
    for (; tid < nthreads; tid += stride) {
      int out_id_h = tid / out_chw;
      int out_id_w = tid % out_chw;
      const int in_img_size = in_h * in_w;
      const int out_img_size = out_h * out_w;
      T value = out[out_id_h * out_chw + out_id_w];

      int channel_id = out_id_w / out_img_size;
      int out_img_idy = (out_id_w % out_img_size) / out_w;
      int out_img_idx = tid % out_w;
      int in_img_idx, in_img_idy, w_id, h_id;
      T w1lambda, h1lambda, w2lambda, h2lambda;

      T src_w = ratio_w * (out_img_idx + align_type_value) - align_type_value;
      T src_h = ratio_h * (out_img_idy + align_type_value) - align_type_value;
674 675 676 677 678

      PreCalculatorForLinearInterpInputIndex(&in_img_idx, &w_id, &w1lambda,
                                             &w2lambda, src_w, in_w);
      PreCalculatorForLinearInterpInputIndex(&in_img_idy, &h_id, &h1lambda,
                                             &h2lambda, src_h, in_h);
679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704

      T* in_pos = &in[out_id_h * in_chw + channel_id * in_img_size +
                      in_img_idy * in_w + in_img_idx];
      platform::CudaAtomicAdd(&in_pos[0], h2lambda * w2lambda * value);
      platform::CudaAtomicAdd(&in_pos[w_id], h2lambda * w1lambda * value);
      platform::CudaAtomicAdd(&in_pos[h_id * in_w],
                              h1lambda * w2lambda * value);
      platform::CudaAtomicAdd(&in_pos[h_id * in_w + w_id],
                              h1lambda * w1lambda * value);
    }
  } else {
    for (; tid < nthreads; tid += stride) {
      int out_id_h = tid / out_chw;
      int out_id_w = tid % out_chw;
      const int in_img_size = in_h * in_w;
      const int out_img_size = out_h * out_w;
      T value = out[out_id_h * out_chw + out_id_w];

      int out_img_idy = out_id_w / (out_w * num_channels);
      int out_img_idx = out_id_w % (out_w * num_channels) / num_channels;
      int channel_id = tid % num_channels;

      int in_img_idx, in_img_idy, w_id, h_id;
      T w1lambda, h1lambda, w2lambda, h2lambda;
      T src_w = ratio_w * (out_img_idx + align_type_value) - align_type_value;
      T src_h = ratio_h * (out_img_idy + align_type_value) - align_type_value;
705 706 707 708 709

      PreCalculatorForLinearInterpInputIndex(&in_img_idx, &w_id, &w1lambda,
                                             &w2lambda, src_w, in_w);
      PreCalculatorForLinearInterpInputIndex(&in_img_idy, &h_id, &h1lambda,
                                             &h2lambda, src_h, in_h);
710 711 712 713

      T* in_pos = &in[out_id_h * in_chw + in_img_idy * in_w * num_channels +
                      in_img_idx * num_channels + channel_id];
      platform::CudaAtomicAdd(&in_pos[0], h2lambda * w2lambda * value);
X
xiaoting 已提交
714
      platform::CudaAtomicAdd(&in_pos[w_id * num_channels],
715 716 717
                              h2lambda * w1lambda * value);
      platform::CudaAtomicAdd(&in_pos[h_id * in_w * num_channels],
                              h1lambda * w2lambda * value);
X
xiaoting 已提交
718
      platform::CudaAtomicAdd(
719 720
          &in_pos[h_id * in_w * num_channels + w_id * num_channels],
          h1lambda * w1lambda * value);
X
xiaoting 已提交
721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 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 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179
    }
  }
}

template <typename T>
__global__ void KeTrilinearInterpFw(
    const T* in, const size_t in_img_d, const size_t in_img_h,
    const size_t in_img_w, const size_t input_h, const size_t input_w, T* out,
    const size_t out_img_d, const size_t out_img_h, const size_t out_img_w,
    const size_t output_h, const size_t output_w, const size_t num_channels,
    const float ratio_d, const float ratio_h, const float ratio_w,
    const bool align_corners, const int align_mode,
    const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  bool align_flag = (align_mode == 0 && !align_corners);
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idt, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idt = (out_id_w % out_img_size) / out_img_h / out_img_w;
      out_img_idy = ((out_id_w % out_img_size) / out_img_w) % out_img_h;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idt = out_id_w / (out_img_h * out_img_w * num_channels);
      out_img_idy = out_id_w % (out_img_h * out_img_w * num_channels) /
                    (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idt = align_flag
                         ? static_cast<int>(ratio_d * (out_img_idt + 0.5) - 0.5)
                         : static_cast<int>(ratio_d * out_img_idt);
    in_img_idt = (in_img_idt > 0) ? in_img_idt : 0;
    int d_id = (in_img_idt < in_img_d - 1) ? 1 : 0;
    T src_d = ratio_d * (out_img_idt + 0.5) - 0.5;
    src_d = (src_d > 0) ? src_d : 0;
    T d1lambda =
        align_flag ? src_d - in_img_idt : ratio_d * out_img_idt - in_img_idt;
    T d2lambda = 1.f - d1lambda;

    int in_img_idy = align_flag
                         ? static_cast<int>(ratio_h * (out_img_idy + 0.5) - 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    in_img_idy = (in_img_idy > 0) ? in_img_idy : 0;
    int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0;
    T src_h = ratio_h * (out_img_idy + 0.5) - 0.5;
    src_h = (src_h > 0) ? src_h : 0;
    T h1lambda =
        align_flag ? src_h - in_img_idy : ratio_h * out_img_idy - in_img_idy;
    T h2lambda = 1.f - h1lambda;

    int in_img_idx = align_flag
                         ? static_cast<int>(ratio_w * (out_img_idx + 0.5) - 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);
    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;
    int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;
    T src_w = ratio_w * (out_img_idx + 0.5) - 0.5;
    src_w = (src_w > 0) ? src_w : 0;
    T w1lambda =
        align_flag ? src_w - in_img_idx : ratio_w * out_img_idx - in_img_idx;
    T w2lambda = 1.f - w1lambda;

    if (data_layout == DataLayout::kNCHW) {
      int in_pos1_idx = out_id_h * input_w + channel_id * in_img_size +
                        (in_img_idt * in_img_h + in_img_idy) * in_img_w +
                        in_img_idx;
      const T* in_pos1 = &in[in_pos1_idx];
      int in_pos2_idx = in_pos1_idx + d_id * in_img_h * in_img_w;
      const T* in_pos2 = &in[in_pos2_idx];

      // trilinear interpolation
      out[out_id_h * output_w + out_id_w] =
          d2lambda *
              (h2lambda * (w2lambda * in_pos1[0] + w1lambda * in_pos1[w_id]) +
               h1lambda * (w2lambda * in_pos1[h_id * in_img_w] +
                           w1lambda * in_pos1[h_id * in_img_w + w_id])) +
          d1lambda *
              (h2lambda * (w2lambda * in_pos2[0] + w1lambda * in_pos2[w_id]) +
               h1lambda * (w2lambda * in_pos2[h_id * in_img_w] +
                           w1lambda * in_pos2[h_id * in_img_w + w_id]));

    } else {
      int in_pos1_idx = out_id_h * input_w +
                        in_img_idt * in_img_h * in_img_w * num_channels +
                        in_img_idy * in_img_w * num_channels +
                        in_img_idx * num_channels + channel_id;
      const T* in_pos1 = &in[in_pos1_idx];
      int in_pos2_idx = in_pos1_idx + d_id * in_img_h * in_img_w * num_channels;
      const T* in_pos2 = &in[in_pos2_idx];

      // trilinear interpolation
      out[out_id_h * output_w + out_id_w] =
          d2lambda *
              (h2lambda * (w2lambda * in_pos1[0] +
                           w1lambda * in_pos1[w_id * num_channels]) +
               h1lambda * (w2lambda * in_pos1[h_id * in_img_w * num_channels] +
                           w1lambda * in_pos1[h_id * in_img_w * num_channels +
                                              w_id * num_channels])) +
          d1lambda *
              (h2lambda * (w2lambda * in_pos2[0] +
                           w1lambda * in_pos2[w_id * num_channels]) +
               h1lambda * (w2lambda * in_pos2[h_id * in_img_w * num_channels] +
                           w1lambda * in_pos2[h_id * in_img_w * num_channels +
                                              w_id * num_channels]));
    }
  }
}

template <typename T>
__global__ void KeTrilinearInterpBw(
    T* in, const size_t in_img_d, const size_t in_img_h, const size_t in_img_w,
    const size_t input_h, const size_t input_w, const T* out,
    const size_t out_img_d, const size_t out_img_h, const size_t out_img_w,
    const size_t output_h, const size_t output_w, const size_t num_channels,
    const T ratio_d, const T ratio_h, const T ratio_w, const bool align_corners,
    const int align_mode, const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  bool align_flag = (align_mode == 0 && !align_corners);
  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idt, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idt = (out_id_w % out_img_size) / out_img_h / out_img_w;
      out_img_idy = ((out_id_w % out_img_size) / out_img_w) % out_img_h;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idt = out_id_w / (out_img_h * out_img_w * num_channels);
      out_img_idy = out_id_w % (out_img_h * out_img_w * num_channels) /
                    (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    int in_img_idt = align_flag
                         ? static_cast<int>(ratio_d * (out_img_idt + 0.5) - 0.5)
                         : static_cast<int>(ratio_d * out_img_idt);
    in_img_idt = (in_img_idt > 0) ? in_img_idt : 0;
    int d_id = (in_img_idt < in_img_d - 1) ? 1 : 0;
    T src_d = ratio_d * (out_img_idt + 0.5) - 0.5;
    src_d = (src_d > 0) ? src_d : 0;
    T d1lambda =
        align_flag ? src_d - in_img_idt : ratio_d * out_img_idt - in_img_idt;
    T d2lambda = 1.f - d1lambda;

    int in_img_idy = align_flag
                         ? static_cast<int>(ratio_h * (out_img_idy + 0.5) - 0.5)
                         : static_cast<int>(ratio_h * out_img_idy);
    in_img_idy = (in_img_idy > 0) ? in_img_idy : 0;
    int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0;
    T src_h = ratio_h * (out_img_idy + 0.5) - 0.5;
    src_h = (src_h > 0) ? src_h : 0;
    T h1lambda =
        align_flag ? src_h - in_img_idy : ratio_h * out_img_idy - in_img_idy;
    T h2lambda = 1.f - h1lambda;

    int in_img_idx = align_flag
                         ? static_cast<int>(ratio_w * (out_img_idx + 0.5) - 0.5)
                         : static_cast<int>(ratio_w * out_img_idx);
    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;
    int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;
    T src_w = ratio_w * (out_img_idx + 0.5) - 0.5;
    src_w = (src_w > 0) ? src_w : 0;
    T w1lambda =
        align_flag ? src_w - in_img_idx : ratio_w * out_img_idx - in_img_idx;
    T w2lambda = 1.f - w1lambda;

    if (data_layout == DataLayout::kNCHW) {
      int in_pos1_idx = out_id_h * input_w + channel_id * in_img_size +
                        (in_img_idt * in_img_h + in_img_idy) * in_img_w +
                        in_img_idx;
      T* in_pos1 = &in[in_pos1_idx];
      int in_pos2_idx = in_pos1_idx + d_id * in_img_h * in_img_w;
      T* in_pos2 = &in[in_pos2_idx];

      const T* out_pos = &out[out_id_h * output_w + out_id_w];

      // trilinear interpolation grad
      platform::CudaAtomicAdd(&in_pos1[0],
                              d2lambda * h2lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos1[w_id],
                              d2lambda * h2lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos1[h_id * in_img_w],
                              d2lambda * h1lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos1[h_id * in_img_w + w_id],
                              d2lambda * h1lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[0],
                              d1lambda * h2lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[w_id],
                              d1lambda * h2lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[h_id * in_img_w],
                              d1lambda * h1lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[h_id * in_img_w + w_id],
                              d1lambda * h1lambda * w1lambda * out_pos[0]);
    } else {
      int in_pos1_idx = out_id_h * input_w +
                        in_img_idt * in_img_h * in_img_w * num_channels +
                        in_img_idy * in_img_w * num_channels +
                        in_img_idx * num_channels + channel_id;
      T* in_pos1 = &in[in_pos1_idx];
      int in_pos2_idx = in_pos1_idx + d_id * in_img_h * in_img_w * num_channels;
      T* in_pos2 = &in[in_pos2_idx];

      const T* out_pos = &out[out_id_h * output_w + out_id_w];

      // trilinear interpolation grad
      platform::CudaAtomicAdd(&in_pos1[0],
                              d2lambda * h2lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos1[w_id * num_channels],
                              d2lambda * h2lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos1[h_id * in_img_w * num_channels],
                              d2lambda * h1lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(
          &in_pos1[h_id * in_img_w * num_channels + w_id * num_channels],
          d2lambda * h1lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[0],
                              d1lambda * h2lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[w_id * num_channels],
                              d1lambda * h2lambda * w1lambda * out_pos[0]);
      platform::CudaAtomicAdd(&in_pos2[h_id * in_img_w * num_channels],
                              d1lambda * h1lambda * w2lambda * out_pos[0]);
      platform::CudaAtomicAdd(
          &in_pos2[h_id * in_img_w * num_channels + w_id * num_channels],
          d1lambda * h1lambda * w1lambda * out_pos[0]);
    }
  }
}

template <typename T>
__device__ __forceinline__ static T Kecubic_interp(const T x0, const T x1,
                                                   const T x2, const T x3,
                                                   T t) {
  T coeffs[4];
  T a = -0.75;
  T x_1 = t;
  T x_2 = 1.0 - t;
  coeffs[0] = cubic_convolution2<T>(x_1 + 1.0, a);
  coeffs[1] = cubic_convolution1<T>(x_1, a);
  coeffs[2] = cubic_convolution1<T>(x_2, a);
  coeffs[3] = cubic_convolution2<T>(x_2 + 1.0, a);
  return x0 * coeffs[0] + x1 * coeffs[1] + x2 * coeffs[2] + x3 * coeffs[3];
}

template <typename T>
__global__ void KeBicubicInterpFw(
    const T* in, const size_t in_img_h, const size_t in_img_w,
    const size_t input_h, const size_t input_w, T* out, const size_t out_img_h,
    const size_t out_img_w, const size_t output_h, const size_t output_w,
    const size_t num_channels, const float ratio_h, const float ratio_w,
    const bool align_corners, const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;

  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idy, out_img_idx;

    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idy = (out_id_w % out_img_size) / out_img_w;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idy = out_id_w / (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    T in_img_idy = align_corners
                       ? static_cast<T>(ratio_h * out_img_idy)
                       : static_cast<T>(ratio_h * (out_img_idy + 0.5) - 0.5);
    int input_y = floorf(in_img_idy);
    const T y_t = in_img_idy - input_y;

    T in_img_idx = align_corners
                       ? static_cast<T>(ratio_w * out_img_idx)
                       : static_cast<T>(ratio_w * (out_img_idx + 0.5) - 0.5);
    int input_x = floorf(in_img_idx);
    const T x_t = in_img_idx - input_x;

    T coefficients[4];
    const T* in_pos_0;
    const T* in_pos_1;
    const T* in_pos_2;
    const T* in_pos_3;
    int access_x_0;
    if (data_layout == DataLayout::kNCHW) {
      for (int k = 0; k < 4; k++) {
        int access_y =
            max(min(input_y - 1 + k, static_cast<int>(in_img_h - 1)), 0);
        access_x_0 = max(min(input_x - 1, static_cast<int>(in_img_w - 1)), 0);
        int access_x_1 =
            max(min(input_x + 0, static_cast<int>(in_img_w - 1)), 0);
        int access_x_2 =
            max(min(input_x + 1, static_cast<int>(in_img_w - 1)), 0);
        int access_x_3 =
            max(min(input_x + 2, static_cast<int>(in_img_w - 1)), 0);

        in_pos_0 = &in[out_id_h * input_w + channel_id * in_img_size +
                       access_y * in_img_w + access_x_0];
        in_pos_1 = &in[out_id_h * input_w + channel_id * in_img_size +
                       access_y * in_img_w + access_x_1];
        in_pos_2 = &in[out_id_h * input_w + channel_id * in_img_size +
                       access_y * in_img_w + access_x_2];
        in_pos_3 = &in[out_id_h * input_w + channel_id * in_img_size +
                       access_y * in_img_w + access_x_3];

        coefficients[k] = Kecubic_interp<T>(in_pos_0[0], in_pos_1[0],
                                            in_pos_2[0], in_pos_3[0], x_t);
      }

      out[out_id_h * output_w + out_id_w] =
          Kecubic_interp<T>(coefficients[0], coefficients[1], coefficients[2],
                            coefficients[3], y_t);

    } else {
      for (int k = 0; k < 4; k++) {
        int access_y =
            max(min(input_y - 1 + k, static_cast<int>((in_img_h - 1))), 0);
        int access_x_0 =
            max(min(input_x - 1, static_cast<int>((in_img_w - 1))), 0);
        int access_x_1 =
            max(min(input_x + 0, static_cast<int>((in_img_w - 1))), 0);
        int access_x_2 =
            max(min(input_x + 1, static_cast<int>((in_img_w - 1))), 0);
        int access_x_3 =
            max(min(input_x + 2, static_cast<int>((in_img_w - 1))), 0);

        const T* in_pos_0 =
            &in[out_id_h * input_w + access_y * in_img_w * num_channels +
                access_x_0 * num_channels + channel_id];
        const T* in_pos_1 =
            &in[out_id_h * input_w + access_y * in_img_w * num_channels +
                access_x_1 * num_channels + channel_id];
        const T* in_pos_2 =
            &in[out_id_h * input_w + access_y * in_img_w * num_channels +
                access_x_2 * num_channels + channel_id];
        const T* in_pos_3 =
            &in[out_id_h * input_w + access_y * in_img_w * num_channels +
                access_x_3 * num_channels + channel_id];

        coefficients[k] = Kecubic_interp(in_pos_0[0], in_pos_1[0], in_pos_2[0],
                                         in_pos_3[0], x_t);
      }

      out[out_id_h * output_w + out_id_w] =
          static_cast<T>(Kecubic_interp(coefficients[0], coefficients[1],
                                        coefficients[2], coefficients[3], y_t));
    }
  }
}

template <typename T>
__global__ void KeBicubicInterpBw(
    T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h,
    const size_t input_w, const T* out, const size_t out_img_h,
    const size_t out_img_w, const size_t output_h, const size_t output_w,
    const size_t num_channels, const float ratio_h, const float ratio_w,
    const bool align_corners, const DataLayout data_layout) {
  int nthreads = output_h * output_w;
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;

  for (; tid < nthreads; tid += stride) {
    int out_id_h = tid / output_w;
    int out_id_w = tid % output_w;
    int in_img_size = input_w / num_channels;
    int out_img_size = output_w / num_channels;

    int channel_id, out_img_idy, out_img_idx;
    if (data_layout == DataLayout::kNCHW) {
      channel_id = out_id_w / out_img_size;
      out_img_idy = (out_id_w % out_img_size) / out_img_w;
      out_img_idx = tid % out_img_w;
    } else {
      out_img_idy = out_id_w / (out_img_w * num_channels);
      out_img_idx = out_id_w % (out_img_w * num_channels) / num_channels;
      channel_id = tid % num_channels;
    }

    T in_img_idy = align_corners
                       ? static_cast<T>(ratio_h * out_img_idy)
                       : static_cast<T>(ratio_h * (out_img_idy + 0.5) - 0.5);
    int input_y = floorf(in_img_idy);
    const T y_t = in_img_idy - input_y;

    T in_img_idx = align_corners
                       ? static_cast<T>(ratio_w * out_img_idx)
                       : static_cast<T>(ratio_w * (out_img_idx + 0.5) - 0.5);
    int input_x = floorf(in_img_idx);

    const T x_t = in_img_idx - input_x;

    T x_coeffs[4];
    T y_coeffs[4];

    get_cubic_upsample_coefficients(x_coeffs, x_t);
    get_cubic_upsample_coefficients(y_coeffs, y_t);

    const T* out_pos = &out[out_id_h * output_w + out_id_w];
    T* in_pos;

    for (int i = 0; i < 4; i++) {
      for (int j = 0; j < 4; j++) {
        int access_y = max(min(static_cast<int>(input_y - 1 + j),
                               static_cast<int>(in_img_h - 1)),
                           0);
        int access_x = max(min(static_cast<int>(input_x - 1 + i),
                               static_cast<int>(in_img_w - 1)),
                           0);
        if (data_layout == DataLayout::kNCHW) {
          in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
                       access_y * in_img_w + access_x];
        } else {
          in_pos = &in[out_id_h * input_w + access_y * in_img_w * num_channels +
                       access_x * num_channels + channel_id];
        }
        platform::CudaAtomicAdd(&in_pos[0],
                                (out_pos[0] * y_coeffs[j] * x_coeffs[i]));
      }
    }
  }
}

template <typename T>
static void Interpolate1DCUDAFwd(const framework::ExecutionContext& ctx,
                                 const Tensor& input, Tensor* output) {
  auto* input_data = input.data<T>();

  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_w = ctx.Attr<int>("out_w");

  auto list_new_shape_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
1180
  float scale_w = -1;
X
xiaoting 已提交
1181 1182 1183 1184 1185 1186 1187 1188 1189 1190
  if (list_new_shape_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_shape_tensor);
    out_w = new_size[0];
  } else {
    auto scale_tensor = ctx.Input<Tensor>("Scale");
    auto scale = ctx.Attr<std::vector<float>>("scale");
    if (scale_tensor != nullptr) {
      auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
      scale_w = scale_data[0];
K
Kqnonrime 已提交
1191 1192 1193 1194 1195 1196
      PADDLE_ENFORCE_EQ(
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
X
xiaoting 已提交
1197 1198 1199
    } else {
      if (scale.size() > 0) {
        scale_w = scale[0];
K
Kqnonrime 已提交
1200 1201 1202 1203 1204 1205
        PADDLE_ENFORCE_EQ(
            scale_w > 0, true,
            platform::errors::InvalidArgument(
                "The scale_w in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_w));
X
xiaoting 已提交
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 1235 1236
      }
    }
    if (scale_w > 0.) {
      out_w = static_cast<int>(in_w * scale_w);
    }
    auto out_size = ctx.Input<Tensor>("OutSize");
    if (out_size != nullptr) {
      Tensor sizes;
      framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
      auto size_data = sizes.data<int>();
      out_w = size_data[0];
    }
  }
  PADDLE_ENFORCE_GT(out_w, 0, platform::errors::InvalidArgument(
                                  "out_w in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));
  framework::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {n, c, out_w};
  } else {
    dim_out = {n, out_w, c};
  }
  auto output_data = output->mutable_data<T>(dim_out, ctx.GetPlace());

  if (in_w == out_w) {
    framework::TensorCopy(input, ctx.GetPlace(), output);
    return;
  }

  float ratio_w = 0.f;
  if (out_w > 1) {
1237 1238 1239
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
1240
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1.0) / (out_w - 1.0)
1241
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
1242 1243
  }

X
xiaoting 已提交
1244 1245 1246
  int64_t in_cw = c * in_w;
  int64_t out_cw = c * out_w;
  auto pixelNum = n * out_cw;
X
xiaoting 已提交
1247 1248

  platform::GpuLaunchConfig config =
1249
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
1250 1251

  if ("linear" == interp_method) {
1252
    KeLinearInterpFw<T><<<config.block_per_grid, config.thread_per_block, 0,
X
xiaoting 已提交
1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276
                          ctx.cuda_device_context().stream()>>>(
        input_data, in_w, in_cw, output_data, out_w, n, out_cw, c, ratio_w,
        align_corners, align_mode, data_layout);
  }
}

template <typename T>
static void Interpolate2DCUDAFwd(const framework::ExecutionContext& ctx,
                                 const Tensor& input, Tensor* output) {
  auto* input_data = input.data<T>();

  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_h = ctx.Attr<int>("out_h");
  int out_w = ctx.Attr<int>("out_w");

  auto list_new_shape_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
1277 1278
  float scale_w = -1;
  float scale_h = -1;
X
xiaoting 已提交
1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295
  if (list_new_shape_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_shape_tensor);
    out_h = new_size[0];
    out_w = new_size[1];
  } else {
    auto scale_tensor = ctx.Input<Tensor>("Scale");
    auto scale = ctx.Attr<std::vector<float>>("scale");
    if (scale_tensor != nullptr) {
      auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
      if (scale_data.size() > 1) {
        scale_h = scale_data[0];
        scale_w = scale_data[1];
      } else {
        scale_h = scale_data[0];
        scale_w = scale_data[0];
      }
K
Kqnonrime 已提交
1296

X
xiaoting 已提交
1297
      PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      PADDLE_ENFORCE_EQ(
          scale_h > 0, true,
          platform::errors::InvalidArgument(
              "The scale_h in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
X
xiaoting 已提交
1309 1310 1311 1312
    } else {
      if (scale.size() > 1) {
        scale_w = scale[1];
        scale_h = scale[0];
K
Kqnonrime 已提交
1313

X
xiaoting 已提交
1314
        PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
            scale_w > 0, true,
            platform::errors::InvalidArgument(
                "The scale_w in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_w));
        PADDLE_ENFORCE_EQ(
            scale_h > 0, true,
            platform::errors::InvalidArgument(
                "The scale_h in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_h));
X
xiaoting 已提交
1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363
      }
    }
    if (scale_w > 0. && scale_h > 0.) {
      out_h = static_cast<int>(in_h * scale_h);
      out_w = static_cast<int>(in_w * scale_w);
    }
    auto out_size = ctx.Input<Tensor>("OutSize");
    if (out_size != nullptr) {
      Tensor sizes;
      framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
      auto size_data = sizes.data<int>();
      out_h = size_data[0];
      out_w = size_data[1];
    }
  }
  PADDLE_ENFORCE_GT(out_h, 0, platform::errors::InvalidArgument(
                                  "out_h in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));
  PADDLE_ENFORCE_GT(out_w, 0, platform::errors::InvalidArgument(
                                  "out_w in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));

  framework::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {n, c, out_h, out_w};
  } else {
    dim_out = {n, out_h, out_w, c};
  }
  auto output_data = output->mutable_data<T>(dim_out, ctx.GetPlace());

  if (in_h == out_h && in_w == out_w) {
    framework::TensorCopy(input, ctx.GetPlace(), output);
    return;
  }

  float ratio_h = 0.f;
  float ratio_w = 0.f;
  if (out_h > 1) {
1364 1365 1366
    float new_scale_h = 0.f;
    new_scale_h = (scale_h > 0) ? static_cast<float>(1. / scale_h)
                                : static_cast<float>(in_h) / out_h;
X
xiaoting 已提交
1367
    ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
1368
                              : static_cast<float>(new_scale_h);
X
xiaoting 已提交
1369 1370
  }
  if (out_w > 1) {
1371 1372 1373
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
1374
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
1375
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
1376 1377
  }

X
xiaoting 已提交
1378 1379 1380 1381
  int64_t in_hw = in_h * in_w;
  int64_t out_hw = out_h * out_w;
  int64_t in_chw = c * in_hw;
  int64_t out_chw = c * out_hw;
X
xiaoting 已提交
1382

X
xiaoting 已提交
1383
  auto pixelNum = n * out_chw;
X
xiaoting 已提交
1384 1385

  platform::GpuLaunchConfig config =
1386
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
1387 1388

  if ("nearest" == interp_method) {
1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407
    if (data_layout == DataLayout::kNCHW) {
      // get launch 3D config
      int nc = n * c;
      platform::GpuLaunchConfig config_3d =
          GetGpuLaunchConfig3D(ctx.cuda_device_context(), nc, out_h, out_w);
      KeNearestNeighborInterpNCHWFw<
          T><<<config_3d.block_per_grid, config_3d.thread_per_block, 0,
               ctx.cuda_device_context().stream()>>>(
          input_data, in_h, in_w, output_data, out_h, out_w, nc, ratio_h,
          ratio_w, align_corners);
    } else {
      int64_t cw = c * out_w;
      auto interp_divmods = FastDivModForInterpolate(c, out_chw, cw);
      KeNearestNeighborInterpFw<
          T><<<config.block_per_grid, config.thread_per_block, 0,
               ctx.cuda_device_context().stream()>>>(
          input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n,
          out_chw, c, ratio_h, ratio_w, align_corners, interp_divmods);
    }
X
xiaoting 已提交
1408
  } else if ("bilinear" == interp_method) {
F
feng_shuai 已提交
1409 1410 1411 1412 1413 1414
    dim3 thread_num = config.thread_per_block;
#ifdef WITH_NV_JETSON
    if (config.compute_capability == 53 || config.compute_capability == 62) {
      thread_num = 512;
    }
#endif
1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433
    const T align_type_value = (align_mode == 0 && !align_corners) ? 0.5f : 0;
    if (data_layout == DataLayout::kNCHW) {
      // get launch 3D config
      int nc = n * c;
      platform::GpuLaunchConfig config_3d =
          GetGpuLaunchConfig3D(ctx.cuda_device_context(), nc, out_h, out_w);
      KeBilinearInterpNCHWFw<
          T><<<config_3d.block_per_grid, config_3d.thread_per_block, 0,
               ctx.cuda_device_context().stream()>>>(
          input_data, in_h, in_w, output_data, out_h, out_w, nc, ratio_h,
          ratio_w, align_type_value);
    } else {
      int64_t cw = c * out_w;
      auto interp_divmods = FastDivModForInterpolate(c, out_chw, cw);
      KeBilinearInterpFw<T><<<config.block_per_grid, thread_num, 0,
                              ctx.cuda_device_context().stream()>>>(
          input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n,
          out_chw, c, ratio_h, ratio_w, align_type_value, interp_divmods);
    }
X
xiaoting 已提交
1434
  } else if ("bicubic" == interp_method) {
1435 1436 1437 1438 1439 1440
#ifdef __HIPCC__
    constexpr int thread_per_block = 256;
#else
    constexpr int thread_per_block = 512;
#endif
    KeBicubicInterpFw<T><<<config.block_per_grid, thread_per_block, 0,
1441
                           ctx.cuda_device_context().stream()>>>(
X
xiaoting 已提交
1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465
        input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n,
        out_chw, c, ratio_h, ratio_w, align_corners, data_layout);
  }
}

template <typename T>
static void Interpolate3DCUDAFwd(const framework::ExecutionContext& ctx,
                                 const Tensor& input, Tensor* output) {
  auto* input_data = input.data<T>();

  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_d = ctx.Attr<int>("out_d");
  int out_h = ctx.Attr<int>("out_h");
  int out_w = ctx.Attr<int>("out_w");

  auto list_new_shape_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
1466 1467 1468
  float scale_w = -1;
  float scale_d = -1;
  float scale_h = -1;
X
xiaoting 已提交
1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488
  if (list_new_shape_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_shape_tensor);
    out_d = new_size[0];
    out_h = new_size[1];
    out_w = new_size[2];
  } else {
    auto scale_tensor = ctx.Input<Tensor>("Scale");
    auto scale = ctx.Attr<std::vector<float>>("scale");
    if (scale_tensor != nullptr) {
      auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
      if (scale_data.size() > 1) {
        scale_d = scale_data[0];
        scale_h = scale_data[1];
        scale_w = scale_data[2];
      } else {
        scale_d = scale_data[0];
        scale_h = scale_data[0];
        scale_w = scale_data[0];
      }
K
Kqnonrime 已提交
1489 1490 1491 1492 1493 1494 1495

      PADDLE_ENFORCE_EQ(
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
X
xiaoting 已提交
1496
      PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507
          scale_h > 0, true,
          platform::errors::InvalidArgument(
              "The scale_h in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
      PADDLE_ENFORCE_EQ(
          scale_d > 0, true,
          platform::errors::InvalidArgument(
              "The scale_d in input 'Scale' Tensor of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_d));
X
xiaoting 已提交
1508 1509 1510 1511 1512 1513 1514
    } else {
      if (scale.size() > 1) {
        scale_d = scale[0];
        scale_h = scale[1];
        scale_w = scale[2];

        PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531
            scale_w > 0, true,
            platform::errors::InvalidArgument(
                "The scale_w in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_w));
        PADDLE_ENFORCE_EQ(
            scale_h > 0, true,
            platform::errors::InvalidArgument(
                "The scale_h in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_h));
        PADDLE_ENFORCE_EQ(
            scale_d > 0, true,
            platform::errors::InvalidArgument(
                "The scale_d in Attr(scale) of Operator(interpolate) "
                "should be greater than 0, but received value is %d.",
                scale_d));
X
xiaoting 已提交
1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575
      }
    }
    if (scale_d > 0. && scale_h > 0. && scale_w > 0.) {
      out_d = static_cast<int>(in_d * scale_d);
      out_h = static_cast<int>(in_h * scale_h);
      out_w = static_cast<int>(in_w * scale_w);
    }
    auto out_size = ctx.Input<Tensor>("OutSize");
    if (out_size != nullptr) {
      Tensor sizes;
      framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
      auto size_data = sizes.data<int>();
      out_d = size_data[0];
      out_h = size_data[1];
      out_w = size_data[2];
    }
  }
  PADDLE_ENFORCE_GT(out_d, 0, platform::errors::InvalidArgument(
                                  "out_d in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));
  PADDLE_ENFORCE_GT(out_h, 0, platform::errors::InvalidArgument(
                                  "out_h in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));
  PADDLE_ENFORCE_GT(out_w, 0, platform::errors::InvalidArgument(
                                  "out_w in Attr(out_shape) of Op(interpolate) "
                                  "should be greater than 0."));

  framework::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {n, c, out_d, out_h, out_w};
  } else {
    dim_out = {n, out_d, out_h, out_w, c};
  }
  auto output_data = output->mutable_data<T>(dim_out, ctx.GetPlace());

  if (in_d == out_d && in_h == out_h && in_w == out_w) {
    framework::TensorCopy(input, ctx.GetPlace(), output);
    return;
  }

  float ratio_d = 0.f;
  float ratio_h = 0.f;
  float ratio_w = 0.f;
  if (out_d > 1) {
1576 1577 1578
    float new_scale_d = 0.f;
    new_scale_d = (scale_d > 0) ? static_cast<float>(1. / scale_d)
                                : static_cast<float>(in_d) / out_d;
X
xiaoting 已提交
1579
    ratio_d = (align_corners) ? static_cast<float>(in_d - 1) / (out_d - 1)
1580
                              : static_cast<float>(new_scale_d);
X
xiaoting 已提交
1581 1582
  }
  if (out_h > 1) {
1583 1584 1585
    float new_scale_h = 0.f;
    new_scale_h = (scale_h > 0) ? static_cast<float>(1. / scale_h)
                                : static_cast<float>(in_h) / out_h;
X
xiaoting 已提交
1586
    ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
1587
                              : static_cast<float>(new_scale_h);
X
xiaoting 已提交
1588 1589
  }
  if (out_w > 1) {
1590 1591 1592
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
1593
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
1594
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
1595 1596
  }

X
xiaoting 已提交
1597 1598 1599 1600
  int64_t in_dhw = in_d * in_h * in_w;
  int64_t out_dhw = out_d * out_h * out_w;
  int64_t in_cdhw = c * in_dhw;
  int64_t out_cdhw = c * out_dhw;
X
xiaoting 已提交
1601

X
xiaoting 已提交
1602
  auto pixelNum = n * out_cdhw;
X
xiaoting 已提交
1603 1604

  platform::GpuLaunchConfig config =
1605
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
1606 1607

  if ("trilinear" == interp_method) {
1608
    KeTrilinearInterpFw<T><<<config.block_per_grid, config.thread_per_block, 0,
X
xiaoting 已提交
1609 1610 1611 1612
                             ctx.cuda_device_context().stream()>>>(
        input_data, in_d, in_h, in_w, n, in_cdhw, output_data, out_d, out_h,
        out_w, n, out_cdhw, c, ratio_d, ratio_h, ratio_w, align_corners,
        align_mode, data_layout);
1613 1614 1615 1616 1617 1618 1619
  } else if ("nearest" == interp_method) {
    KeNearestNeighbor3DInterpFw<
        T><<<config.block_per_grid, config.thread_per_block, 0,
             ctx.cuda_device_context().stream()>>>(
        input_data, in_d, in_h, in_w, n, in_cdhw, output_data, out_d, out_h,
        out_w, n, out_cdhw, c, ratio_d, ratio_h, ratio_w, align_corners,
        data_layout);
X
xiaoting 已提交
1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
  }
}

template <typename T>
static void Interpolate1DCUDABwd(const framework::ExecutionContext& ctx,
                                 Tensor* input_grad, const Tensor output_grad) {
  auto* input = ctx.Input<Tensor>("X");
  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input->dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_w = ctx.Attr<int>("out_w");
  float scale_w = -1;
  auto scale_tensor = ctx.Input<Tensor>("Scale");
  auto scale = ctx.Attr<std::vector<float>>("scale");
  if (scale_tensor != nullptr) {
    auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
    scale_w = scale_data[0];
K
Kqnonrime 已提交
1643 1644 1645 1646 1647 1648
    PADDLE_ENFORCE_EQ(
        scale_w > 0, true,
        platform::errors::InvalidArgument(
            "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_w));
X
xiaoting 已提交
1649 1650 1651 1652
  } else {
    if (scale.size() > 0) {
      scale_w = scale[0];

K
Kqnonrime 已提交
1653 1654 1655 1656 1657 1658
      PADDLE_ENFORCE_EQ(
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
X
xiaoting 已提交
1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688
    }
  }
  if (scale_w > 0.) {
    out_w = static_cast<int>(in_w * scale_w);
  }

  auto out_size = ctx.Input<Tensor>("OutSize");
  if (out_size != nullptr) {
    Tensor sizes;
    framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
    auto size_data = sizes.data<int>();
    out_w = size_data[0];
  }
  auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
  if (list_new_size_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_size_tensor);
    out_w = new_size[0];
  }

  auto* output_grad_data = output_grad.data<T>();
  framework::DDim dim_grad;
  if (data_layout == DataLayout::kNCHW) {
    dim_grad = {n, c, in_w};
  } else {
    dim_grad = {n, in_w, c};
  }
  input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
  auto* input_grad_data = input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
  auto& device_ctx = ctx.template device_context<platform::CUDADeviceContext>();
1689
  pten::funcs::SetConstant<platform::CUDADeviceContext, T> zero;
X
xiaoting 已提交
1690 1691 1692 1693 1694 1695 1696 1697 1698
  zero(device_ctx, input_grad, static_cast<T>(0.0));

  if (in_w == out_w) {
    framework::TensorCopy(output_grad, ctx.GetPlace(), input_grad);
    return;
  }

  float ratio_w = 0.f;
  if (out_w > 1) {
1699 1700 1701
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
1702
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
1703
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
1704
  }
X
xiaoting 已提交
1705 1706 1707
  int64_t in_cw = c * in_w;
  int64_t out_cw = c * out_w;
  auto pixelNum = n * out_cw;
X
xiaoting 已提交
1708 1709

  platform::GpuLaunchConfig config =
1710
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
1711 1712

  if ("linear" == interp_method) {
1713
    KeLinearInterpBw<T><<<config.block_per_grid, config.thread_per_block, 0,
X
xiaoting 已提交
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747
                          ctx.cuda_device_context().stream()>>>(
        input_grad_data, in_w, in_cw, output_grad_data, out_w, n, out_cw, c,
        ratio_w, align_corners, align_mode, data_layout);
  }
}

template <typename T>
static void Interpolate2DCUDABwd(const framework::ExecutionContext& ctx,
                                 Tensor* input_grad, const Tensor output_grad) {
  auto* input = ctx.Input<Tensor>("X");
  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input->dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_h = ctx.Attr<int>("out_h");
  int out_w = ctx.Attr<int>("out_w");
  float scale_h = -1;
  float scale_w = -1;
  auto scale_tensor = ctx.Input<Tensor>("Scale");
  auto scale = ctx.Attr<std::vector<float>>("scale");
  if (scale_tensor != nullptr) {
    auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
    if (scale_data.size() > 1) {
      scale_h = scale_data[0];
      scale_w = scale_data[1];
    } else {
      scale_h = scale_data[0];
      scale_w = scale_data[0];
    }
K
Kqnonrime 已提交
1748 1749 1750 1751 1752 1753 1754

    PADDLE_ENFORCE_EQ(
        scale_w > 0, true,
        platform::errors::InvalidArgument(
            "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_w));
X
xiaoting 已提交
1755
    PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1756 1757 1758 1759 1760
        scale_h > 0, true,
        platform::errors::InvalidArgument(
            "The scale_h in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_h));
X
xiaoting 已提交
1761 1762 1763 1764 1765 1766
  } else {
    if (scale.size() > 1) {
      scale_w = scale[1];
      scale_h = scale[0];

      PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      PADDLE_ENFORCE_EQ(
          scale_h > 0, true,
          platform::errors::InvalidArgument(
              "The scale_h in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
X
xiaoting 已提交
1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810
    }
  }
  if (scale_w > 0. && scale_h > 0.) {
    out_h = static_cast<int>(in_h * scale_h);
    out_w = static_cast<int>(in_w * scale_w);
  }

  auto out_size = ctx.Input<Tensor>("OutSize");
  if (out_size != nullptr) {
    Tensor sizes;
    framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
    auto size_data = sizes.data<int>();
    out_h = size_data[0];
    out_w = size_data[1];
  }
  auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
  if (list_new_size_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_size_tensor);
    out_h = new_size[0];
    out_w = new_size[1];
  }

  auto* output_grad_data = output_grad.data<T>();
  framework::DDim dim_grad;
  if (data_layout == DataLayout::kNCHW) {
    dim_grad = {n, c, in_h, in_w};
  } else {
    dim_grad = {n, in_h, in_w, c};
  }
  input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
  auto* input_grad_data = input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
  auto& device_ctx = ctx.template device_context<platform::CUDADeviceContext>();
1811
  pten::funcs::SetConstant<platform::CUDADeviceContext, T> zero;
X
xiaoting 已提交
1812 1813 1814 1815 1816 1817 1818 1819 1820 1821
  zero(device_ctx, input_grad, static_cast<T>(0.0));

  if (in_h == out_h && in_w == out_w) {
    framework::TensorCopy(output_grad, ctx.GetPlace(), input_grad);
    return;
  }

  float ratio_h = 0.f;
  float ratio_w = 0.f;
  if (out_h > 1) {
1822 1823 1824
    float new_scale_h = 0.f;
    new_scale_h = (scale_h > 0) ? static_cast<float>(1. / scale_h)
                                : static_cast<float>(in_h) / out_h;
X
xiaoting 已提交
1825
    ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
1826
                              : static_cast<float>(new_scale_h);
X
xiaoting 已提交
1827 1828
  }
  if (out_w > 1) {
1829 1830 1831
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
1832
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
1833
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
1834 1835
  }

X
xiaoting 已提交
1836 1837 1838 1839 1840
  int64_t in_hw = in_h * in_w;
  int64_t out_hw = out_h * out_w;
  int64_t in_chw = c * in_hw;
  int64_t out_chw = c * out_hw;
  auto pixelNum = n * out_chw;
X
xiaoting 已提交
1841 1842

  platform::GpuLaunchConfig config =
1843
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
1844 1845

  if ("nearest" == interp_method) {
1846 1847 1848
    KeNearestNeighborInterpBw<
        T><<<config.block_per_grid, config.thread_per_block, 0,
             ctx.cuda_device_context().stream()>>>(
X
xiaoting 已提交
1849 1850 1851
        input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, out_w,
        n, out_chw, c, ratio_h, ratio_w, align_corners, data_layout);
  } else if ("bilinear" == interp_method) {
1852 1853 1854
    const T align_type_value = (align_mode == 0 && !align_corners) ? 0.5f : 0;
    bool is_nchw = (data_layout == DataLayout::kNCHW) ? true : false;
    bool optimize_flag = false;
1855
#ifndef __HIPCC__
1856 1857 1858
    optimize_flag = (in_h < (out_h >> 6) && in_w < (out_w >> 6))
                        ? true
                        : ((in_h == 1 && in_w == 1) ? true : false);
1859
#endif
1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872

    if (optimize_flag & is_nchw) {
      KeBilinearInterpBwShareMemory<
          T><<<config.block_per_grid, config.thread_per_block, 0,
               ctx.cuda_device_context().stream()>>>(
          input_grad_data, in_h, in_w, output_grad_data, out_h, out_w, n, c,
          ratio_h, ratio_w, align_type_value, is_nchw);
    } else {
      KeBilinearInterpBw<T><<<config.block_per_grid, config.thread_per_block, 0,
                              ctx.cuda_device_context().stream()>>>(
          input_grad_data, in_h, in_w, output_grad_data, out_h, out_w, n, c,
          ratio_h, ratio_w, align_type_value, is_nchw);
    }
X
xiaoting 已提交
1873
  } else if ("bicubic" == interp_method) {
1874 1875 1876 1877 1878 1879
#ifdef __HIPCC__
    constexpr int thread_per_block = 256;
#else
    constexpr int thread_per_block = 512;
#endif
    KeBicubicInterpBw<T><<<config.block_per_grid, thread_per_block, 0,
1880
                           ctx.cuda_device_context().stream()>>>(
X
xiaoting 已提交
1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919
        input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, out_w,
        n, out_chw, c, ratio_h, ratio_w, align_corners, data_layout);
  }
}

template <typename T>
static void Interpolate3DCUDABwd(const framework::ExecutionContext& ctx,
                                 Tensor* input_grad,
                                 const Tensor& output_grad) {
  auto* input = ctx.Input<Tensor>("X");
  const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
  const DataLayout data_layout = framework::StringToDataLayout(data_layout_str);
  int n, c, in_d, in_h, in_w;
  ExtractNCDWH(input->dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);

  auto interp_method = ctx.Attr<std::string>("interp_method");
  bool align_corners = ctx.Attr<bool>("align_corners");
  int align_mode = ctx.Attr<int>("align_mode");

  int out_d = ctx.Attr<int>("out_d");
  int out_h = ctx.Attr<int>("out_h");
  int out_w = ctx.Attr<int>("out_w");
  float scale_d = -1;
  float scale_h = -1;
  float scale_w = -1;
  auto scale_tensor = ctx.Input<Tensor>("Scale");
  auto scale = ctx.Attr<std::vector<float>>("scale");
  if (scale_tensor != nullptr) {
    auto scale_data = get_new_data_from_tensor<float>(scale_tensor);
    if (scale_data.size() > 1) {
      scale_d = scale_data[0];
      scale_h = scale_data[1];
      scale_w = scale_data[2];
    } else {
      scale_d = scale_data[0];
      scale_h = scale_data[0];
      scale_w = scale_data[0];
    }
    PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936
        scale_w > 0, true,
        platform::errors::InvalidArgument(
            "The scale_w in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_w));
    PADDLE_ENFORCE_EQ(
        scale_h > 0, true,
        platform::errors::InvalidArgument(
            "The scale_h in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_h));
    PADDLE_ENFORCE_EQ(
        scale_d > 0, true,
        platform::errors::InvalidArgument(
            "The scale_d in input 'Scale' Tensor of Operator(interpolate) "
            "should be greater than 0, but received value is %d.",
            scale_d));
X
xiaoting 已提交
1937 1938 1939 1940 1941 1942 1943
  } else {
    if (scale.size() > 1) {
      scale_d = scale[0];
      scale_h = scale[1];
      scale_w = scale[2];

      PADDLE_ENFORCE_EQ(
K
Kqnonrime 已提交
1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960
          scale_w > 0, true,
          platform::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      PADDLE_ENFORCE_EQ(
          scale_h > 0, true,
          platform::errors::InvalidArgument(
              "The scale_h in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
      PADDLE_ENFORCE_EQ(
          scale_d > 0, true,
          platform::errors::InvalidArgument(
              "The scale_d in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_d));
X
xiaoting 已提交
1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
    }
  }
  if (scale_d > 0. && scale_h > 0. && scale_w > 0.) {
    out_d = static_cast<int>(in_d * scale_d);
    out_h = static_cast<int>(in_h * scale_h);
    out_w = static_cast<int>(in_w * scale_w);
  }

  auto out_size = ctx.Input<Tensor>("OutSize");
  if (out_size != nullptr) {
    Tensor sizes;
    framework::TensorCopySync(*out_size, platform::CPUPlace(), &sizes);
    auto size_data = sizes.data<int>();
    out_d = size_data[0];
    out_h = size_data[1];
    out_w = size_data[2];
  }
  auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
  if (list_new_size_tensor.size() > 0) {
    // have size tensor
    auto new_size = get_new_shape(list_new_size_tensor);
    out_d = new_size[0];
    out_h = new_size[1];
    out_w = new_size[2];
  }

  auto* output_grad_data = output_grad.data<T>();
  framework::DDim dim_grad;
  if (data_layout == DataLayout::kNCHW) {
    dim_grad = {n, c, in_d, in_h, in_w};
  } else {
    dim_grad = {n, in_d, in_h, in_w, c};
  }
  auto* input_grad_data = input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
  auto& device_ctx = ctx.template device_context<platform::CUDADeviceContext>();
1996
  pten::funcs::SetConstant<platform::CUDADeviceContext, T> zero;
X
xiaoting 已提交
1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
  zero(device_ctx, input_grad, static_cast<T>(0.0));

  if (in_d == out_d && in_h == out_h && in_w == out_w) {
    framework::TensorCopy(output_grad, ctx.GetPlace(), input_grad);
    return;
  }

  float ratio_d = 0.f;
  float ratio_h = 0.f;
  float ratio_w = 0.f;
  if (out_d > 1) {
2008 2009 2010
    float new_scale_d = 0.f;
    new_scale_d = (scale_d > 0) ? static_cast<float>(1. / scale_d)
                                : static_cast<float>(in_d) / out_d;
X
xiaoting 已提交
2011
    ratio_d = (align_corners) ? static_cast<float>(in_d - 1) / (out_d - 1)
2012
                              : static_cast<float>(new_scale_d);
X
xiaoting 已提交
2013 2014
  }
  if (out_h > 1) {
2015 2016 2017
    float new_scale_h = 0.f;
    new_scale_h = (scale_h > 0) ? static_cast<float>(1. / scale_h)
                                : static_cast<float>(in_h) / out_h;
X
xiaoting 已提交
2018
    ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
2019
                              : static_cast<float>(new_scale_h);
X
xiaoting 已提交
2020 2021
  }
  if (out_w > 1) {
2022 2023 2024
    float new_scale_w = 0.f;
    new_scale_w = (scale_w > 0) ? static_cast<float>(1. / scale_w)
                                : static_cast<float>(in_w) / out_w;
X
xiaoting 已提交
2025
    ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
2026
                              : static_cast<float>(new_scale_w);
X
xiaoting 已提交
2027 2028
  }

X
xiaoting 已提交
2029 2030 2031 2032
  int64_t in_dhw = in_d * in_h * in_w;
  int64_t out_dhw = out_d * out_h * out_w;
  int64_t in_cdhw = c * in_dhw;
  int64_t out_cdhw = c * out_dhw;
X
xiaoting 已提交
2033

X
xiaoting 已提交
2034
  auto pixelNum = n * out_cdhw;
X
xiaoting 已提交
2035 2036

  platform::GpuLaunchConfig config =
2037
      platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), pixelNum);
X
xiaoting 已提交
2038 2039

  if ("trilinear" == interp_method) {
2040
    KeTrilinearInterpBw<T><<<config.block_per_grid, config.thread_per_block, 0,
X
xiaoting 已提交
2041 2042 2043 2044
                             ctx.cuda_device_context().stream()>>>(
        input_grad_data, in_d, in_h, in_w, n, in_cdhw, output_grad_data, out_d,
        out_h, out_w, n, out_cdhw, c, ratio_d, ratio_h, ratio_w, align_corners,
        align_mode, data_layout);
2045 2046 2047 2048 2049 2050 2051
  } else if ("nearest" == interp_method) {
    KeNearestNeighbor3DInterpBw<
        T><<<config.block_per_grid, config.thread_per_block, 0,
             ctx.cuda_device_context().stream()>>>(
        input_grad_data, in_d, in_h, in_w, n, in_cdhw, output_grad_data, out_d,
        out_h, out_w, n, out_cdhw, c, ratio_d, ratio_h, ratio_w, align_corners,
        data_layout);
X
xiaoting 已提交
2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111
  }
}

template <typename T>
class InterpolateOpV2CUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        platform::errors::NotFound("This kernel only runs on GPU device."));
    auto* input = ctx.Input<Tensor>("X");
    auto* output = ctx.Output<Tensor>("Out");

    auto input_dims = input->dims();
    if (input_dims.size() == 3) {  // 1D interpolation
      Interpolate1DCUDAFwd<T>(ctx, *input, output);
    } else if (input_dims.size() == 4) {  // 2D interpolation
      Interpolate2DCUDAFwd<T>(ctx, *input, output);
    } else if (input_dims.size() == 5) {  // 3D interpolation
      Interpolate3DCUDAFwd<T>(ctx, *input, output);
    }
  }
};

template <typename T>
class InterpolateV2GradOpCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        platform::errors::NotFound("This kernel only runs on GPU device."));
    auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));

    auto output_grad_dims = output_grad->dims();
    if (output_grad_dims.size() == 3) {  // 1D interpolation
      Interpolate1DCUDABwd<T>(ctx, input_grad, *output_grad);
    } else if (output_grad_dims.size() == 4) {  // 2D interpolation
      Interpolate2DCUDABwd<T>(ctx, input_grad, *output_grad);
    } else if (output_grad_dims.size() == 5) {  // 3D interpolation
      Interpolate3DCUDABwd<T>(ctx, input_grad, *output_grad);
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(bilinear_interp_v2,
                        ops::InterpolateOpV2CUDAKernel<float>,
                        ops::InterpolateOpV2CUDAKernel<double>,
                        ops::InterpolateOpV2CUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(bilinear_interp_v2_grad,
                        ops::InterpolateV2GradOpCUDAKernel<float>,
                        ops::InterpolateV2GradOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(nearest_interp_v2,
                        ops::InterpolateOpV2CUDAKernel<float>,
                        ops::InterpolateOpV2CUDAKernel<double>,
2112
                        ops::InterpolateOpV2CUDAKernel<int64_t>,
X
xiaoting 已提交
2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136
                        ops::InterpolateOpV2CUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(nearest_interp_v2_grad,
                        ops::InterpolateV2GradOpCUDAKernel<float>,
                        ops::InterpolateV2GradOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(trilinear_interp_v2,
                        ops::InterpolateOpV2CUDAKernel<float>,
                        ops::InterpolateOpV2CUDAKernel<double>,
                        ops::InterpolateOpV2CUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(trilinear_interp_v2_grad,
                        ops::InterpolateV2GradOpCUDAKernel<float>,
                        ops::InterpolateV2GradOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(linear_interp_v2, ops::InterpolateOpV2CUDAKernel<float>,
                        ops::InterpolateOpV2CUDAKernel<double>,
                        ops::InterpolateOpV2CUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(linear_interp_v2_grad,
                        ops::InterpolateV2GradOpCUDAKernel<float>,
                        ops::InterpolateV2GradOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(bicubic_interp_v2,
                        ops::InterpolateOpV2CUDAKernel<float>,
                        ops::InterpolateOpV2CUDAKernel<double>,
                        ops::InterpolateOpV2CUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(bicubic_interp_v2_grad,
                        ops::InterpolateV2GradOpCUDAKernel<float>,
                        ops::InterpolateV2GradOpCUDAKernel<double>);