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.
15 #include "include/include_all.cl"
17 #define LOCAL_SIZE INPUT0_BATCH_NUM
19 __attribute__((reqd_work_group_size(LOCAL_SIZE, 1, 1)))
21 __global INPUT0_TYPE* input,
22 __global OUTPUT_TYPE* output,
23 __global FILTER_TYPE* weights,
25 __global BIAS_TYPE* biases,
28 __global INPUT0_TYPE* scale_in
30 , __global INPUT0_TYPE* scale_bias
33 , __global INPUT0_TYPE* inv_var,
34 __global INPUT0_TYPE* conv_output,
35 __global INPUT0_TYPE* bn_output
39 const uint f = get_global_id(1);
40 const uint b = get_global_id(0);
42 UNIT_TYPE conv_out = UNIT_VAL_ZERO;
44 const uint in_split_offset = split_idx * INPUT0_FEATURE_PITCH * FILTER_IFM_NUM;
46 const uint filter_offset = f*FILTER_OFM_PITCH;
47 const uint input_offset = b*INPUT0_BATCH_PITCH + INPUT0_OFFSET + in_split_offset;
49 for (uint y = 0; y < OUTPUT_SIZE_Y; ++y)
51 const int input_y = y * STRIDE_SIZE_Y - PADDING_SIZE_Y;
52 for (uint x = 0; x < OUTPUT_SIZE_X; ++x)
54 const int input_x = x * STRIDE_SIZE_X - PADDING_SIZE_X;
55 for (uint k = 0; k < FILTER_IFM_NUM; ++k)
57 for (uint j = 0; j < FILTER_SIZE_Y ; ++j)
59 const int input_offset_y = input_y + j * DILATION_SIZE_Y;
60 const bool zero_y = input_offset_y >= INPUT0_SIZE_Y || input_offset_y < 0;
64 for (uint i = 0; i < FILTER_SIZE_X ; ++i)
66 const int input_offset_x = input_x + i * DILATION_SIZE_X;
67 const bool zero_x = input_offset_x >= INPUT0_SIZE_X || input_offset_x < 0;
71 uint input_idx = input_offset + (uint)input_offset_x*INPUT0_X_PITCH + (uint)input_offset_y*INPUT0_Y_PITCH + k*INPUT0_FEATURE_PITCH;
72 uint filter_idx = filter_offset + k*FILTER_IFM_PITCH + j*FILTER_Y_PITCH + i*FILTER_X_PITCH;
73 conv_out += input[input_idx] * weights[filter_idx];
80 conv_out += (UNIT_TYPE)biases[f];
83 const uint out_split_offset = split_idx * OUTPUT_FEATURE_PITCH * OUTPUT_FEATURE_NUM;
84 const uint dst_index = GET_DATA_INDEX(OUTPUT, b, f, y, x) + out_split_offset;
86 conv_output[dst_index] = conv_out;
88 output[dst_index] = conv_out;
95 barrier(CLK_LOCAL_MEM_FENCE);
97 __local ACCUMULATOR_TYPE sum[LOCAL_SIZE];
99 const uint local_idx = b;
103 uint input_idx = GET_DATA_INDEX(OUTPUT, local_idx, f, 0, 0);
104 for (uint y = 0; y < OUTPUT_SIZE_Y; y++)
106 for (uint x = 0; x < OUTPUT_SIZE_X; x++)
108 #ifdef FUSED_TRAINING
109 UNIT_TYPE in = conv_output[input_idx];
111 UNIT_TYPE in = output[input_idx];
113 sum[local_idx] += in;
114 input_idx += OUTPUT_X_PITCH;
116 input_idx += OUTPUT_Y_PITCH - OUTPUT_SIZE_X * OUTPUT_X_PITCH;
119 barrier(CLK_LOCAL_MEM_FENCE);
121 for(uint offset = LOCAL_SIZE / 2; offset > 0; offset /= 2)
123 if (local_idx < offset)
125 sum[local_idx] += sum[local_idx + offset];
127 barrier(CLK_LOCAL_MEM_FENCE);
130 UNIT_TYPE mean = sum[0] / (OUTPUT_BATCH_NUM * OUTPUT_SIZE_X * OUTPUT_SIZE_Y);
134 input_idx = GET_DATA_INDEX(OUTPUT, local_idx, f, 0, 0);
135 for (uint y = 0; y < OUTPUT_SIZE_Y; y++)
137 for (uint x = 0; x < OUTPUT_SIZE_X; x++)
139 #ifdef FUSED_TRAINING
140 UNIT_TYPE in = conv_output[input_idx] - mean;
142 UNIT_TYPE in = output[input_idx] - mean;
144 sum[local_idx] += in * in;
145 input_idx += OUTPUT_X_PITCH;
147 input_idx += OUTPUT_Y_PITCH - OUTPUT_SIZE_X * OUTPUT_X_PITCH;
150 barrier(CLK_LOCAL_MEM_FENCE);
152 for(uint offset = LOCAL_SIZE / 2; offset > 0; offset /= 2)
154 if (local_idx < offset)
156 sum[local_idx] += sum[local_idx + offset];
158 barrier(CLK_LOCAL_MEM_FENCE);
161 float variance = sum[0] / (OUTPUT_BATCH_NUM * OUTPUT_SIZE_X * OUTPUT_SIZE_Y);
163 float inv_variance = (float)(1.0 / sqrt(variance + EPSILON));
165 #ifdef FUSED_TRAINING
167 inv_var[f] = inv_variance;
170 uint out_idx = GET_DATA_INDEX(OUTPUT, local_idx, f, 0, 0);
171 for (uint y = 0; y < OUTPUT_SIZE_Y; y++)
173 for (uint x = 0; x < OUTPUT_SIZE_X; x++)
175 #ifdef FUSED_TRAINING
176 UNIT_TYPE out_val = inv_variance * (conv_output[out_idx] - mean);
177 bn_output[out_idx] = out_val;
178 #ifdef SCALE_BIAS_TERM
179 output[out_idx] = ACTIVATION(out_val * scale_in[f] + scale_bias[f], NL_M, NL_N);
181 output[out_idx] = ACTIVATION(out_val * scale_in[f], NL_M, NL_N);
184 #ifdef SCALE_BIAS_TERM
185 output[out_idx] = ACTIVATION(inv_variance * (output[out_idx] - mean) * scale_in[f] + scale_bias[f], NL_M, NL_N);
187 output[out_idx] = ACTIVATION(inv_variance * (output[out_idx] - mean) * scale_in[f], NL_M, NL_N);
190 out_idx += OUTPUT_X_PITCH;
192 out_idx += OUTPUT_Y_PITCH - OUTPUT_SIZE_X * OUTPUT_X_PITCH;