1 // Copyright (c) 2018 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 __attribute__((intel_reqd_sub_group_size(8)))
37 KERNEL(pooling_gpu_fs_bs_yx_bsv4_fsv32)(
38 const __global UNIT_TYPE* input,
39 __global UNIT_TYPE* output)
41 const uint x = (uint)get_global_id(0);
42 const uint y = (uint)get_global_id(1);
43 const uint bf = (uint)get_global_id(2);
44 // we process 4 features per workitem that's why we need to divide it
45 const uint aligned32_features = ((INPUT0_FEATURE_NUM + 31) / 32) * 32;
46 const uint f = (get_global_id(2) * 4) % aligned32_features;
47 const uint b = 4 * ((get_global_id(2) * 4) / aligned32_features);
49 if (x >= OUTPUT_SIZE_X)
54 const int offset_x = (int)x*STRIDE_SIZE_X - PADDING_SIZE_X;
55 const int offset_y = (int)y*STRIDE_SIZE_Y - PADDING_SIZE_Y;
57 int4 result[4] = { INIT_VAL };
60 if (offset_x + POOL_SIZE_X < 0 || offset_x >= INPUT0_SIZE_X ||
61 offset_y + POOL_SIZE_Y < 0 || offset_y >= INPUT0_SIZE_Y)
66 #ifdef DYNAMIC_KERNEL_DIVIDER
67 uint num_elementes = 0;
70 const uint batch_and_feature_offset = GET_DATA_FS_BS_YX_BSV4_FSV32_INDEX(INPUT0, b, f, 0, 0);
71 for(uint j = 0; j < POOL_SIZE_Y; j++)
73 int input_offset_y = offset_y + j;
74 bool zero_y = input_offset_y >= INPUT0_SIZE_Y || input_offset_y < 0;
77 for(uint i = 0; i < POOL_SIZE_X; i++)
79 int input_offset_x = offset_x + i;
80 bool zero = input_offset_x >= INPUT0_SIZE_X || input_offset_x < 0;
83 const uint input_idx = batch_and_feature_offset + input_offset_y*IN_Y_PITCH + input_offset_x*IN_X_PITCH;
85 int4 int_data = as_int4(intel_sub_group_block_read4((const __global uint*)(input + input_idx)));
86 for(uint b = 0; b < 4; b++)
88 char4 input_data = as_char4(int_data[b]);
89 result[b][0] = FUNC_CALL(apply_pooling)(result[b][0], (int)input_data[0]);
90 result[b][1] = FUNC_CALL(apply_pooling)(result[b][1], (int)input_data[1]);
91 result[b][2] = FUNC_CALL(apply_pooling)(result[b][2], (int)input_data[2]);
92 result[b][3] = FUNC_CALL(apply_pooling)(result[b][3], (int)input_data[3]);
96 #ifdef DYNAMIC_KERNEL_DIVIDER
103 #ifdef DYNAMIC_WITH_PADDING_KERNEL_DIVIDER
104 const int hend = min(offset_y + POOL_SIZE_Y, INPUT0_SIZE_Y + PADDING_SIZE_Y);
105 const int wend = min(offset_x + POOL_SIZE_X, INPUT0_SIZE_X + PADDING_SIZE_X);
106 const uint num_elementes = (hend - offset_y) * (wend - offset_x);
109 uint input_idx = GET_DATA_FS_BS_YX_BSV4_FSV32_INDEX(INPUT0, b, f, offset_y, offset_x);
111 for(uint j = 0; j < POOL_SIZE_Y; j++)
113 for(uint i = 0; i < POOL_SIZE_X; i++)
115 int4 int_data = as_int4(intel_sub_group_block_read4((const __global uint*)(input + input_idx)));
116 for(uint b = 0; b < 4; b++)
118 char4 input_data = as_char4(int_data[b]);
119 result[b][0] = FUNC_CALL(apply_pooling)(result[b][0], (int)input_data[0]);
120 result[b][1] = FUNC_CALL(apply_pooling)(result[b][1], (int)input_data[1]);
121 result[b][2] = FUNC_CALL(apply_pooling)(result[b][2], (int)input_data[2]);
122 result[b][3] = FUNC_CALL(apply_pooling)(result[b][3], (int)input_data[3]);
125 input_idx += IN_X_PITCH;
127 input_idx += (IN_Y_PITCH - POOL_SIZE_X*IN_X_PITCH);
130 #if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
131 const uint num_elementes = POOL_SIZE_X*POOL_SIZE_Y;
135 #if defined AVG_POOLING
136 #if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
137 for(uint b = 0; b < 4; b++)
139 for(uint i = 0; i < 4; i++)
141 result[b][i] = convert_int(round(((float)result[b][i] / max(num_elementes, (uint)1)));
145 for(uint b = 0; b < 4; b++)
147 for(uint i = 0; i < 4; i++)
149 result[b][i] = convert_int(round((float)result[b][i] / (int)(POOL_SIZE_Y * POOL_SIZE_X)));
156 for(uint b = 0; b < 4; b++)
158 char4 char_res = as_char4(char_result[b]);
159 for(uint op = 0; op < 4; op++)
161 char_res[op] = ACTIVATION(convert_char(result[b][op]), NL_M ,NL_N);
163 char_result[b] = as_int(char_res);
165 const uint output_pos = GET_DATA_FS_BS_YX_BSV4_FSV32_INDEX(OUTPUT, b, f, y, x);
166 intel_sub_group_block_write4((__global uint*)(output + output_pos), as_uint4(char_result));