1 // Copyright (c) 2019 Intel Corporation
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
7 // http://www.apache.org/licenses/LICENSE-2.0
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.
16 #include "include/include_all.cl"
19 #define INIT_VAL CHAR_MIN
27 inline int FUNC(apply_pooling)(int tmp, int in)
36 KERNEL(pooling_gpu_b_fs_yx_fsv4)(
37 const __global UNIT_TYPE* input,
38 __global UNIT_TYPE* output)
40 const uint x = (uint)get_global_id(0);
41 const uint y = (uint)get_global_id(1);
42 const uint bf = (uint)get_global_id(2);
43 const uint f = (bf * 4) % INPUT0_FEATURE_NUM;
44 const uint b = (bf * 4) / INPUT0_FEATURE_NUM;
46 const int offset_x = (int)x*STRIDE_SIZE_X - PADDING_SIZE_X;
47 const int offset_y = (int)y*STRIDE_SIZE_Y - PADDING_SIZE_Y;
49 int result[4] = { INIT_VAL, INIT_VAL, INIT_VAL, INIT_VAL };
52 if (offset_x + POOL_SIZE_X < 0 || offset_x >= INPUT0_SIZE_X ||
53 offset_y + POOL_SIZE_Y < 0 || offset_y >= INPUT0_SIZE_Y)
58 #ifdef DYNAMIC_KERNEL_DIVIDER
59 uint num_elementes = 0;
62 const uint batch_and_feature_offset = GET_DATA_B_FS_YX_FSV4_INDEX(INPUT0, b, f, 0, 0);
63 for(uint j = 0; j < POOL_SIZE_Y; j++)
65 int input_offset_y = offset_y + j;
66 bool zero_y = input_offset_y >= INPUT0_SIZE_Y || input_offset_y < 0;
69 for(uint i = 0; i < POOL_SIZE_X; i++)
71 int input_offset_x = offset_x + i;
72 bool zero = input_offset_x >= INPUT0_SIZE_X || input_offset_x < 0;
75 const uint input_idx = batch_and_feature_offset + input_offset_y*IN_Y_PITCH + input_offset_x*IN_X_PITCH;
77 int int_data = *((const __global int*)(input + input_idx));
78 char4 ch4_data = as_char4(int_data);
79 result[0] = FUNC_CALL(apply_pooling)(result[0], (int)ch4_data[0]);
80 result[1] = FUNC_CALL(apply_pooling)(result[1], (int)ch4_data[1]);
81 result[2] = FUNC_CALL(apply_pooling)(result[2], (int)ch4_data[2]);
82 result[3] = FUNC_CALL(apply_pooling)(result[3], (int)ch4_data[3]);
84 #ifdef DYNAMIC_KERNEL_DIVIDER
91 #ifdef DYNAMIC_WITH_PADDING_KERNEL_DIVIDER
92 const int hend = min(offset_y + POOL_SIZE_Y, INPUT0_SIZE_Y + PADDING_SIZE_Y);
93 const int wend = min(offset_x + POOL_SIZE_X, INPUT0_SIZE_X + PADDING_SIZE_X);
94 const uint num_elementes = (hend - offset_y) * (wend - offset_x);
96 #else // !CHECK_BOUNDRY
97 uint input_idx = GET_DATA_B_FS_YX_FSV4_INDEX(INPUT0, b, f, offset_y, offset_x);
99 for(uint j = 0; j < POOL_SIZE_Y; j++)
101 for(uint i = 0; i < POOL_SIZE_X; i++)
103 int int_data = *((const __global int*)(input + input_idx));
104 char4 ch4_data = as_char4(int_data);
105 result[0] = FUNC_CALL(apply_pooling)(result[0], (int)ch4_data[0]);
106 result[1] = FUNC_CALL(apply_pooling)(result[1], (int)ch4_data[1]);
107 result[2] = FUNC_CALL(apply_pooling)(result[2], (int)ch4_data[2]);
108 result[3] = FUNC_CALL(apply_pooling)(result[3], (int)ch4_data[3]);
110 input_idx += IN_X_PITCH;
112 input_idx += (IN_Y_PITCH - POOL_SIZE_X*IN_X_PITCH);
115 #if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
116 const uint num_elementes = POOL_SIZE_X*POOL_SIZE_Y;
120 #if defined AVG_POOLING
121 #if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
122 for(uint i = 0; i < 4; i++)
124 result[i] = convert_int(round(((float)result[i] / max(num_elementes, (uint)1)));
127 for(uint i = 0; i < 4; i++)
129 result[i] = convert_int(round((float)result[i] / (int)(POOL_SIZE_Y * POOL_SIZE_X)));
135 for(uint op = 0; op < 4; op++)
137 char_res[op] = ACTIVATION(convert_char(result[op]), NL_M ,NL_N);
139 const uint output_pos = GET_DATA_B_FS_YX_FSV4_INDEX(OUTPUT, b, f, y, x);
140 *((__global int*)(output + output_pos)) = as_int(char_res);