inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/convolution_gpu_yxfb_ref.cl

   1 // Copyright (c) 2016-2017 Intel Corporation
   2 //
   3 // Licensed under the Apache License, Version 2.0 (the "License");
   4 // you may not use this file except in compliance with the License.
   5 // You may obtain a copy of the License at
   6 //
   7 //      http://www.apache.org/licenses/LICENSE-2.0
   8 //
   9 // Unless required by applicable law or agreed to in writing, software
  10 // distributed under the License is distributed on an "AS IS" BASIS,
  11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12 // See the License for the specific language governing permissions and
  13 // limitations under the License.
  14
  15 #include "include/include_all.cl"
  16
  17 KERNEL(convolution_gpu_yxfb_ref)(
  18     const __global UNIT_TYPE* input,
  19     __global UNIT_TYPE* output,
  20     const __global UNIT_TYPE* filter,
  21 #if BIAS_TERM
  22     const __global UNIT_TYPE* bias,
  23 #endif
  24     uint split_idx)
  25 {
  26     UNIT_TYPE result = UNIT_VAL_ZERO;
  27
  28     const uint batch_offset = (uint)get_global_id(0) % INPUT0_BATCH_NUM;
  29     const uint ofm_offset   = (uint)get_global_id(0) / INPUT0_BATCH_NUM;
  30     const uint out_x        = (uint)get_global_id(1);
  31     const uint out_y        = (uint)get_global_id(2);
  32
  33     const int x = (int)out_x * STRIDE_SIZE_X - PADDING_SIZE_X;
  34     const int y = (int)out_y * STRIDE_SIZE_Y - PADDING_SIZE_Y;
  35
  36 #if DEPTHWISE_SEPARABLE_OPT
  37     const uint in_split_offset = (ofm_offset / FILTER_OFM_NUM) * INPUT0_FEATURE_PITCH * FILTER_IFM_NUM;
  38 #else
  39     const uint in_split_offset = split_idx * INPUT0_FEATURE_PITCH * FILTER_IFM_NUM;
  40 #endif
  41     const uint input_offset = INPUT0_OFFSET + batch_offset*INPUT0_BATCH_PITCH + in_split_offset;
  42 #if GROUPED && !DEPTHWISE_SEPARABLE_OPT
  43     const uint filter_offset = split_idx * FILTER_LENGTH;
  44 #else
  45     const uint filter_offset = 0;
  46 #endif
  47
  48     for (uint i = 0; i < FILTER_SIZE_Y; i++)
  49     {
  50         const int input_offset_y = y + i * DILATION_SIZE_Y;
  51         const bool zero_y = input_offset_y >= INPUT0_SIZE_Y || input_offset_y < 0;
  52
  53         if(!zero_y)
  54         {
  55             for (uint j = 0; j < FILTER_SIZE_X; j++)
  56             {
  57                 const int input_offset_x = x + j * DILATION_SIZE_X;
  58                 const bool zero = input_offset_x >= INPUT0_SIZE_X || input_offset_x < 0;
  59
  60                 if(!zero)
  61                 {
  62                     uint input_idx = input_offset + (uint)input_offset_x*INPUT0_X_PITCH + (uint)input_offset_y*INPUT0_Y_PITCH;
  63                     uint filter_idx = filter_offset + ofm_offset*FILTER_OFM_PITCH + i*FILTER_Y_PITCH + j*FILTER_X_PITCH;
  64
  65                     for (uint h = 0; h < FILTER_IFM_NUM; h++)
  66                     {
  67                         result = fma(input[input_idx], filter[filter_idx], result);
  68                         filter_idx += FILTER_IFM_PITCH;
  69                         input_idx += INPUT0_FEATURE_PITCH;
  70                     }
  71                 }
  72             }
  73         }
  74     }
  75 #if BIAS_TERM
  76 #if GROUPED && !DEPTHWISE_SEPARABLE_OPT
  77     const uint bias_offset = split_idx * BIAS_LENGTH;
  78 #else
  79     const uint bias_offset = 0;
  80 #endif
  81     result += bias[ofm_offset + bias_offset];
  82 #endif
  83     const uint out_split_offset = split_idx * OUTPUT_FEATURE_PITCH * FILTER_OFM_NUM;
  84     const uint dst_index = batch_offset*OUTPUT_BATCH_PITCH + ofm_offset*OUTPUT_FEATURE_PITCH + out_y*OUTPUT_Y_PITCH + out_x*OUTPUT_X_PITCH + OUTPUT_OFFSET + out_split_offset;
  85     output[dst_index] = ACTIVATION(result, NL_M, NL_N);
  86 }