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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
|
/*
* Copyright (c) 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.
*/
layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
#include "helpers.h"
#ifdef DATA_TYPE_FP16
BUFFER_DECLARATION(src, 1, uint, readonly);
BUFFER_DECLARATION(dst, 2, uint, restrict);
#else // DATA_TYPE_FP16
BUFFER_DECLARATION(src, 1, float, readonly);
BUFFER_DECLARATION(dst, 2, float, restrict);
#endif // DATA_TYPE_FP16
layout(std140) uniform shader_params
{
#ifdef IM2COL_GENERIC
TENSOR3D_PARAM_DECLARATION(src);
IMAGE_PARAM_DECLARATION(dst);
uint filter_depth;
uint src_stride_w;
uint dst_stride_w;
#endif // IM2COL_GENERIC
#ifdef IM2COL_REDUCED
TENSOR3D_PARAM_DECLARATION(src);
VECTOR_PARAM_DECLARATION(dst);
uint width;
uint height;
#endif // IM2COL_REDUCED
#ifdef COL2IM
IMAGE_PARAM_DECLARATION(src);
TENSOR3D_PARAM_DECLARATION(dst);
uint width;
#endif // COL2IM
};
#ifdef DATA_TYPE_FP16
precision mediump float;
#ifdef IM2COL_REDUCED
/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
*
* @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
* @note In case biases will be added in late stage, "#define HAS_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
* @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 Y 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. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] width The width of the input tensor
* @param[in] height The height of the input tensor
*/
void main(void)
{
uvec3 pos = uvec3(gl_GlobalInvocationID.xyz);
uvec3 size = uvec3(gl_WorkGroupSize.xyz);
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
Tensor3D src_nostep = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(src);
Vector dst = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(dst);
uint image_size = width * height;
uint element_count = src_step_x / src_stride_x;
uint tmp_out_offset = dst.current_offset + ((pos.x * element_count + pos.y * width + pos.z * image_size) * dst.stride_x);
uint width_fp16 = ((width + uint(1)) >> uint(1));
uint tmp;
// odd width
if(width % uint(2) != uint(0))
{
// even row
if((pos.y + pos.z * height) % uint(2) == uint(0))
{
LOAD1(tmp, src, src.current_offset >> uint(2));
STORE1(dst, tmp_out_offset >> uint(2), tmp);
}
else
{
// special op
uint tmpleft = uint(0);
uint tmpright = uint(0);
LOAD1(tmpright, src, src.current_offset >> uint(2)); // right half
if(pos.x == uint(0))
{
LOAD1(tmpleft, src, tensor3D_offset_fp16(src_nostep, int(width), int(pos.y) - 1, int(pos.z)) >> uint(2)); // left half
tmpright = (tmpleft & uint(0xffff)) + (tmpright << uint(16));
}
else
{
LOAD1(tmpleft, src, tensor3D_offset_fp16(src_nostep, (int(pos.x) - 1) * int(element_count), int(pos.y), int(pos.z)) >> uint(2)); // left half
tmpright = ((tmpleft >> uint(16)) + (tmpright << uint(16)));
}
STORE1(dst, tmp_out_offset >> uint(2), tmpright);
}
}
else
{
LOAD1(tmp, src, src.current_offset >> uint(2));
STORE1(dst, tmp_out_offset >> uint(2), tmp);
}
#ifdef HAS_BIAS
// If it is the last thread in the 3 dimensional workgroup
if(pos.x == (size.x - 1) && pos.y == (size.y - 1) && pos.z == (size.z - 1))
{
tmp_out_offset += dst.stride_x;
// FIXME: need odd/even detection for tmp_out_offset?
mediump vec2 bias_vec = vec2(1.0f, 1.0f);
uint bias_u = packHalf2x16(bias_vec);
STORE1(dst, tmp_out_offset >> uint(2), bias_u);
}
#endif // HAS_BIAS
}
#endif // IM2COL_REDUCED
#elif defined(DATA_TYPE_FP32)
#ifdef IM2COL_GENERIC
/** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM.
*
* @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
* @note In case biases will be added to the convolution "#define HAS_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: 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 Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] filter_depth The depth of the used filter
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
void main(void)
{
uint xc = gl_GlobalInvocationID.x; // x coordinate in the convolved tensor
uint yc = gl_GlobalInvocationID.y; // y coordinate in the convolved tensor
uint ch = gl_GlobalInvocationID.z % filter_depth; // input feature map
uint batch = gl_GlobalInvocationID.z / filter_depth; // the batch
// Calculate input indeces
uint xi = xc * uint(STRIDE_X) - uint(PAD_X);
uint yi = yc * uint(STRIDE_Y) - uint(PAD_Y);
uint input_offset = (src_offset_first_element_in_bytes + (ch * src_stride_z) + (batch * src_stride_w)) >> uint(2);
// Calculate output indeces
uint xo = ch * uint(KERNEL_WIDTH) * uint(KERNEL_HEIGHT);
uint yo = xc + yc * uint(CONVOLVED_WIDTH); // Index of the convolution
uint output_offset = (dst_offset_first_element_in_bytes + (yo * dst_stride_y) + (batch * dst_stride_w) + xo) >> uint(2);
// Linearize convolution elements
for(uint y = yi, y_e = yi + uint(KERNEL_HEIGHT); y < y_e; ++y)
{
for(uint x = xi, x_e = xi + uint(KERNEL_WIDTH); x < x_e; ++x)
{
#if PAD_X == 0 && PAD_Y == 0
output_offset = input_offset + ((x * src_stride_x + y * src_stride_y) >> uint(2));
STORE4(dst, output_offset, LOAD4(src, input_offset));
#else // PAD_X == 0 && PAD_Y == 0
if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
{
STORE4(dst, output_offset, 0.0f);
}
else
{
output_offset = input_offset + (x * src_stride_x + y * src_stride_y) >> uint(2));
STORE4(dst, output_offset, LOAD4(src, input_offset));
}
#endif // PAD_X == 0 && PAD_Y == 0
}
}
#ifdef HAS_BIAS
if(ch == (uint(KERNEL_DEPTH) - 1))
{
STORE4(dst, output_offset, 1.0f);
}
#endif // HAS_BIAS
}
#endif // IM2COL_GENERIC
#ifdef IM2COL_REDUCED
/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
*
* @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
* @note In case biases will be added in late stage, "#define HAS_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: 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 Y 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. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] width The width of the input tensor
* @param[in] height The height of the input tensor
*/
void main(void)
{
uvec3 pos = uvec3(gl_GlobalInvocationID.xyz);
uvec3 size = uvec3(gl_WorkGroupSize.xyz);
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Vector dst = CONVERT_TO_VECTOR_STRUCT_NO_STEP(dst);
uint image_size = width * height;
uint tmp_out_offset = dst.current_offset + (((pos.x + pos.y * width + pos.z * image_size) * dst.stride_x) >> 2);
STORE4(dst, tmp_out_offset, LOAD4(src, src.current_offset));
#ifdef HAS_BIAS
// If it is the last thread in the 3 dimensional workgroup
if(pos.x == (size.x - 1) && pos.y == (size.y - 1) && pos.z == (size.z - 1))
{
tmp_out_offset += (dst.stride_x >> uint(2));
STORE4(dst, tmp_out_offset, 1.f);
}
#endif // HAS_BIAS
}
#endif // IM2COL_REDUCED
#ifdef COL2IM
/** This kernel performs a reshaping of the output of the convolution layer.
*
* @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: 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 Y 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[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
*/
void main(void)
{
uvec2 pos = uvec2(gl_GlobalInvocationID.xy);
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
uint idx = pos.x * dst.stride_z + (pos.y / width) * dst.stride_y + (pos.y % width) * dst.stride_x;
uint tmp_out_offset = dst.current_offset + (idx >> 2);
STORE4(dst, tmp_out_offset, LOAD4(src, src.current_offset));
}
#endif // COL2IM
#else // DATA_TYPE_FP16
#error Data type not supported
#endif // DATA_TYPE_FP16
|
