1 // Copyright (c) 2016-2017 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"
17 #include "include/sub_group.cl"
19 KERNEL(convolution_gpu_yxfb_yxio_b16)(
20 const __global float* input,
21 __global UNIT_TYPE* output,
22 const __global float* filter,
24 const __global float* bias,
28 // get_global_size(0) -> Number of work items needed to compute all features and all batches for single output spatial position
29 // (single (x, y) point in output).
30 // get_global_size(1) -> Output size in X-dimension.
31 // get_global_size(2) -> Output size in Y-dimension.
32 // get_global_id(0) -> Id of work item computing single spatial point of output indicated by get_global_id(1), get_global_id(2).
33 // get_global_id(1) -> Current x-position in output.
34 // get_global_id(2) -> Current y-position in output.
36 // WORK_ITEMS_PER_SINGLE_BATCHES_ELEMENTS -> Number of work items needed to compute entire one batch for at least one feature and one spatial point.
37 // (this number in current implementation computes also OFM_PER_WORK_ITEM output features at the same time).
38 // FILTER_ARRAY_NUM -> Number of filters groups (split size).
40 const uint out_x = get_global_id(1);
41 const uint out_y = get_global_id(2);
43 const uint output_f_size = OUTPUT_PAD_BEFORE_FEATURE_NUM + OUTPUT_FEATURE_NUM + OUTPUT_PAD_AFTER_FEATURE_NUM;
44 const uint output_x_size = OUTPUT_PAD_BEFORE_SIZE_X + OUTPUT_SIZE_X + OUTPUT_PAD_AFTER_SIZE_X;
45 const uint linear_id_xy = OUTPUT_PAD_BEFORE_SIZE_X + out_x + output_x_size * (out_y + OUTPUT_PAD_BEFORE_SIZE_Y);
46 uint global_id = (((uint)get_global_id(0) / WORK_ITEMS_PER_SINGLE_BATCHES_ELEMENTS) + (linear_id_xy * FILTER_ARRAY_NUM + split_idx) * (output_f_size / OFM_PER_WORK_ITEM)) * WORK_ITEMS_PER_SINGLE_BATCHES_ELEMENTS;
48 const uint sub_group_id = get_local_id(0);
50 #if defined(USE_BLOCK_READ_2) || defined(USE_BLOCK_READ_1)
51 const uint chunk_size = sizeof(uint)/sizeof(UNIT_TYPE);
53 const uint chunk_size = 1;
56 const uint out_batch_id = chunk_size * sub_group_id + LOCAL_WORK_GROUP_SIZE * BATCHES_PER_WORK_ITEM * ((uint)get_group_id(0) % LOCAL_WORK_GROUPS_PER_SINGLE_BATCHES_ELEMENTS);
58 const uint out_id = (global_id / WORK_ITEMS_PER_SINGLE_BATCHES_ELEMENTS) * OFM_PER_WORK_ITEM * OUTPUT_FEATURE_PITCH + OUTPUT_PAD_BEFORE_FEATURE_NUM * OUTPUT_FEATURE_PITCH + OUTPUT_PAD_BEFORE_BATCH_NUM + out_batch_id;
60 const uint ofm_offset = ((global_id * OFM_PER_WORK_ITEM) / WORK_ITEMS_PER_SINGLE_BATCHES_ELEMENTS) % output_f_size;
62 // Each component of vector element contains computation for separate output feature.
63 float8 _data[BATCHES_PER_WORK_ITEM];
64 for(uint i = 0; i < BATCHES_PER_WORK_ITEM; i++)
66 _data[i] = UNIT_VAL_ZERO;
69 const int x = (int)out_x * STRIDE_SIZE_X - PADDING_SIZE_X;
70 const int y = (int)out_y * STRIDE_SIZE_Y - PADDING_SIZE_Y;
72 for (uint i = 0; i < FILTER_SIZE_Y; i++)
74 const int input_offset_y = y + i * DILATION_SIZE_Y;
75 const bool zero_y = input_offset_y >= INPUT0_SIZE_Y || input_offset_y < 0;
79 for (uint j = 0; j < FILTER_SIZE_X; j++)
81 const int input_offset_x = x + j * DILATION_SIZE_X;
82 const bool zero = input_offset_x >= INPUT0_SIZE_X || input_offset_x < 0;
86 uint input_idx = input_offset_x*INPUT0_X_PITCH + input_offset_y*INPUT0_Y_PITCH;
87 input_idx += INPUT0_OFFSET + split_idx * FILTER_IFM_NUM * INPUT0_FEATURE_PITCH;
88 input_idx += out_batch_id;
90 //sub_group_id used as offset to make each workitem load different filter, and then shuffle it
91 uint filter_idx = ofm_offset + sub_group_id + i*FILTER_Y_PITCH + j*FILTER_X_PITCH;
93 for (uint h = 0; h < FILTER_IFM_NUM; h++)
95 #ifdef USE_BLOCK_READ_2
96 float2 _input = as_float2(intel_sub_group_block_read2((const __global uint*)input + input_idx));
97 float8 filter_transp = TRANSPOSE_BLOCK_8(filter[filter_idx]);
98 _data[0] = fma(_input.s0, filter_transp, _data[0]);
99 _data[1] = fma(_input.s1, filter_transp, _data[1]);
100 input_idx += INPUT0_FEATURE_PITCH;
102 float8 filter_transp = TRANSPOSE_BLOCK_8(filter[filter_idx]);
103 for(uint s = 0; s < BATCHES_PER_WORK_ITEM; s++)
105 _data[s] = fma(input[input_idx], filter_transp, _data[s]);
106 input_idx += LOCAL_WORK_GROUP_SIZE;
108 input_idx += INPUT0_FEATURE_PITCH - BATCHES_PER_WORK_ITEM * LOCAL_WORK_GROUP_SIZE;
110 filter_idx += FILTER_IFM_PITCH;
118 float bias_val = bias[ofm_offset + sub_group_id];
119 for(uint s = 0; s < BATCHES_PER_WORK_ITEM; s++)
121 ADD_BIAS_8(_data[s], bias_val);
124 for(uint s = 0; s < BATCHES_PER_WORK_ITEM; s++)
126 _data[s] = ACTIVATION(_data[s], NL_M, NL_N);
129 for(uint s = 0; s < BATCHES_PER_WORK_ITEM; s++)
131 uint _out_id = OUTPUT_VIEW_OFFSET + out_id + s * LOCAL_WORK_GROUP_SIZE;
132 output[_out_id] = _data[s].s0; _out_id += OUTPUT_FEATURE_PITCH;
133 output[_out_id] = _data[s].s1; _out_id += OUTPUT_FEATURE_PITCH;
134 output[_out_id] = _data[s].s2; _out_id += OUTPUT_FEATURE_PITCH;
135 output[_out_id] = _data[s].s3; _out_id += OUTPUT_FEATURE_PITCH;
136 output[_out_id] = _data[s].s4; _out_id += OUTPUT_FEATURE_PITCH;
137 output[_out_id] = _data[s].s5; _out_id += OUTPUT_FEATURE_PITCH;
138 output[_out_id] = _data[s].s6; _out_id += OUTPUT_FEATURE_PITCH;
139 output[_out_id] = _data[s].s7; _out_id += OUTPUT_FEATURE_PITCH;