1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
|
/*
* Copyright (c) 2016, 2017 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
#if defined(FIXED_POINT_POSITION)
#include "fixed_point.h"
#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE_PROMOTED, 8)
#define MUL_OP(a, b) MUL_SAT_OP_EXPAND(CONVERT((a), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), CONVERT((b), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), DATA_TYPE_PROMOTED, 8, FIXED_POINT_POSITION)
// There is no need to have a larger intermediate type for qs32 because all the arguments are already promoted
MULQ_SAT_IMPL(qs32x8, qs32x8)
#else /* FIXED_POINT_POSITION */
#undef CONVERT_SAT
#define ADD_OP(a, b) ((a) + (b))
#define MUL_OP(a, b) ((a) * (b))
#define CONVERT_SAT(a, b) ((a))
#endif /* FIXED_POINT_POSITION */
#if STRIDE_X == 3
#define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size
#define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size)
#elif STRIDE_X == 2
#define INPUT_PIXEL(data_size) extract_input_stride2
#elif STRIDE_X == 1
#define INPUT_PIXEL(data_size) extract_input_stride1
#else /* STRIDE_X not equals 1, 2 or 3 */
#error "Only support strides 1, 2 and 3"
#endif /* STRIDE_X == 3 */
/** Extracts a 1D horizontal vector from the input tensor with stride as 1.
*
* @param[in] input_pixel Pointer to the first pixel.
*
* @return extracted input pixels.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_pixel)
{
return vload8(0, input_pixel);
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 2.
*
* @param[in] input_pixel Pointer to the first pixel.
*
* @return extracted input pixels.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_pixel)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
temp = vload16(0, input_pixel);
return temp.s02468ace;
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 32-bit data size.
*
* @param[in] input_pixel Pointer to the first pixel.
*
* @return extracted input pixels.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_TYPE *input_pixel)
{
VEC_DATA_TYPE(DATA_TYPE, 4)
temp1 = vload4(0, input_pixel);
VEC_DATA_TYPE(DATA_TYPE, 4)
temp2 = vload4(0, input_pixel + 6);
VEC_DATA_TYPE(DATA_TYPE, 4)
temp3 = vload4(0, input_pixel + 12);
VEC_DATA_TYPE(DATA_TYPE, 4)
temp4 = vload4(0, input_pixel + 18);
return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s03, temp2.s03, temp3.s03, temp4.s03);
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 16-bit data size.
*
* @param[in] input_pixel Pointer to the first pixel.
*
* @return extracted input pixels.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_TYPE *input_pixel)
{
VEC_DATA_TYPE(DATA_TYPE, 8)
temp1 = vload8(0, input_pixel);
VEC_DATA_TYPE(DATA_TYPE, 8)
temp2 = vload8(0, input_pixel + 8);
VEC_DATA_TYPE(DATA_TYPE, 8)
temp3 = vload8(0, input_pixel + 16);
return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s036, temp2.s147, temp3.s25);
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size.
*
* @param[in] input_pixel Pointer to the first pixel.
*
* @return extracted input pixels.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_TYPE *input_pixel)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
temp1 = vload16(0, input_pixel);
VEC_DATA_TYPE(DATA_TYPE, 16)
temp2 = vload16(0, input_pixel + 12);
return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369);
}
/** This kernel performs a direct convolution to convolve the low three dimensions.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
* @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
* @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p weights_ptr
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
* @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor
* @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension
*/
#if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
__kernel void direct_convolution1x1(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
#ifdef HAS_BIAS
VECTOR_DECLARATION(biases),
#endif /* defined(HAS_BIAS) */
unsigned int weights_stride_w)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
#ifdef HAS_BIAS
Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
#endif /* defined(HAS_BIAS) */
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
pixels = 0;
const uint z_index = get_global_id(2);
weights.ptr += z_index * weights_stride_w;
for(int d = 0; d < WEIGHTS_DEPTH; ++d)
{
DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr;
VEC_DATA_TYPE(DATA_TYPE, 8)
input_pixel = INPUT_PIXEL(DATA_SIZE)((__global DATA_TYPE *)src.ptr);
pixels = ADD_OP(pixels, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, input_pixel));
src.ptr += src_stride_z;
weights.ptr += weights_stride_z;
}
#ifdef HAS_BIAS
pixels = ADD_OP(pixels, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index))));
#endif /* defined(HAS_BIAS) */
vstore8(CONVERT_SAT(pixels, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
}
#endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
|