2 * Copyright (c) 2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #if defined(FIXED_POINT_POSITION)
27 #include "fixed_point.h"
28 #endif // FIXED_POINT_POSITION
30 #if defined(DATA_TYPE) && defined(ELEMENT_SIZE)
31 #if !defined(FIXED_POINT_POSITION)
34 #define COND_DATA_TYPE char
35 #elif ELEMENT_SIZE == 2
36 #define COND_DATA_TYPE short
37 #elif ELEMENT_SIZE == 4
38 #define COND_DATA_TYPE int
40 #error "Element size not support"
41 #endif // ELEMENT_SIZE
43 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_Y) && defined(KERNEL_DEPTH)
44 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM when the kernel size is 1x1 and the stride_x = 1
46 * @note This kernel computes 4 elements
47 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
48 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
49 * @note The kernel depth must be passed at compile time using -DKERNEL_DEPTH: e.g. -DKERNEL_DEPTH=3
50 * @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
51 * @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.
53 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
54 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
55 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
56 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
57 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
58 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
59 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
60 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
61 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
62 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
63 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
64 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
65 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
66 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
67 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
68 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
70 __kernel void im2col1x1_stridex1_dchw(
71 TENSOR3D_DECLARATION(src),
72 IMAGE_DECLARATION(dst),
76 const uint xc = get_global_id(0) * 4; // x coordinate in the convolved tensor
77 const uint yc = get_global_id(1); // y coordinate in the convolved tensor
78 const uint ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
79 const uint batch = get_global_id(2) / KERNEL_DEPTH; // batch size
82 // The strategy clamps at "xc" as it will be a valid value for sure
83 uint4 xc_clamped = xc + (uint4)(0, 1, 2, 3);
85 // Check which values are valid
86 const VEC_DATA_TYPE(COND_DATA_TYPE, 4) cond0 = CONVERT((xc_clamped < SRC_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, 4));
88 xc_clamped = select((uint4)xc, xc_clamped, convert_int4(cond0));
90 // Calculate input indices
92 const uint yi = yc * STRIDE_Y;
94 // Calculate output indices
96 const uint4 yo = xc_clamped + yc * CONVOLVED_WIDTH; // Index of the convolution
98 // Get input and output address
99 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * src_stride_x + yi * src_stride_y + ch * src_stride_z + batch * src_stride_w;
101 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + batch * dst_stride_w;
103 VEC_DATA_TYPE(DATA_TYPE, 4)
104 data = vload4(0, (__global DATA_TYPE *)input_ptr);
106 // If out-of-bound, overwrite with the first element
107 data = select((VEC_DATA_TYPE(DATA_TYPE, 4))data.s0, data, cond0);
109 *(__global DATA_TYPE *)(output_ptr + yo.s0 * dst_stride_y) = data.s0;
110 *(__global DATA_TYPE *)(output_ptr + yo.s1 * dst_stride_y) = data.s1;
111 *(__global DATA_TYPE *)(output_ptr + yo.s2 * dst_stride_y) = data.s2;
112 *(__global DATA_TYPE *)(output_ptr + yo.s3 * dst_stride_y) = data.s3;
115 if(ch == (KERNEL_DEPTH - 1))
117 *((__global DATA_TYPE *)(output_ptr + yo.s0 * dst_stride_y) + 1) = 1.0f;
118 *((__global DATA_TYPE *)(output_ptr + yo.s1 * dst_stride_y) + 1) = 1.0f;
119 *((__global DATA_TYPE *)(output_ptr + yo.s2 * dst_stride_y) + 1) = 1.0f;
120 *((__global DATA_TYPE *)(output_ptr + yo.s3 * dst_stride_y) + 1) = 1.0f;
124 #endif // defined(CONVOLVED_WIDTH) && defined(STRIDE_Y) && defined(KERNEL_DEPTH)
126 #if defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE)
127 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM when the kernel size is 3x3
129 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
130 * @note The width and height of the input tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT: e.g. -DSRC_WIDTH=128 and -DSRC_HEIGHT=128
131 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
132 * @note The kernel depth must be passed at compile time using -DKERNEL_DEPTH: e.g. -DKERNEL_DEPTH=3
133 * @note The pad_left, pad_right, pad_top and pad_bottom must be passed at compile time using -DPAD_LEFT, -DPAD_RIGHT, -DPAD_TOP and -DPAD_BOTTOM: e.g. -DPAD_LEFT=1, -DPAD_RIGHT=2, -DPAD_TOP=3 and -DPAD_BOTTOM=2
134 * @note The zero value to store in case we load values out-of-bounds must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0.0
135 * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1
136 * @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.
138 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
139 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
140 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
141 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
142 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
143 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
144 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
145 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
146 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
147 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
148 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
149 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
150 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
151 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
152 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
153 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
155 __kernel void im2col3x3_dchw(
156 TENSOR3D_DECLARATION(src),
157 IMAGE_DECLARATION(dst),
161 const int xc = get_global_id(0); // x coordinate in the convolved tensor
162 const int yc = get_global_id(1); // y coordinate in the convolved tensor
163 const int ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
164 const int batch = get_global_id(2) / KERNEL_DEPTH; // batch size
166 // Calculate input indices
167 const int xi = xc * STRIDE_X - PAD_LEFT;
168 const int yi = yc * STRIDE_Y - PAD_TOP;
170 // Calculate output indices
171 const int xo = ch * 9; // 3x3
172 const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
174 // Get input and output address
175 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * (int)src_stride_x + yi * (int)src_stride_y + ch * src_stride_z + batch * src_stride_w;
177 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
179 VEC_DATA_TYPE(DATA_TYPE, 3)
180 row0 = vload3(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
181 VEC_DATA_TYPE(DATA_TYPE, 3)
182 row1 = vload3(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
183 VEC_DATA_TYPE(DATA_TYPE, 3)
184 row2 = vload3(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
186 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
187 // Put 0 if the value is out-of-bound
188 int3 x = (int3)xi + (int3)(0, 1, 2);
189 int3 y = (int3)yi + (int3)(0, 1, 2);
191 VEC_DATA_TYPE(COND_DATA_TYPE, 3)
192 cond0 = CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s0 >= 0 && y.s0 < SRC_HEIGHT)), VEC_DATA_TYPE(COND_DATA_TYPE, 3));
193 VEC_DATA_TYPE(COND_DATA_TYPE, 3)
194 cond1 = CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s1 >= 0 && y.s1 < SRC_HEIGHT)), VEC_DATA_TYPE(COND_DATA_TYPE, 3));
195 VEC_DATA_TYPE(COND_DATA_TYPE, 3)
196 cond2 = CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s2 >= 0 && y.s2 < SRC_HEIGHT)), VEC_DATA_TYPE(COND_DATA_TYPE, 3));
198 row0 = select((VEC_DATA_TYPE(DATA_TYPE, 3))PAD_VALUE, row0, cond0);
199 row1 = select((VEC_DATA_TYPE(DATA_TYPE, 3))PAD_VALUE, row1, cond1);
200 row2 = select((VEC_DATA_TYPE(DATA_TYPE, 3))PAD_VALUE, row2, cond2);
201 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
203 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row0.s012, row1.s012, row2.s01), 0, (__global DATA_TYPE *)output_ptr);
204 *((__global DATA_TYPE *)output_ptr + 8) = row2.s2;
207 if(ch == (KERNEL_DEPTH - 1))
209 *((__global DATA_TYPE *)output_ptr + 9) = 1.0f;
214 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM when the kernel size is 5x5
216 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
217 * @note The width and height of the input tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT: e.g. -DSRC_WIDTH=128 and -DSRC_HEIGHT=128
218 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
219 * @note The kernel depth must be passed at compile time using -DKERNEL_DEPTH: e.g. -DKERNEL_DEPTH=3
220 * @note The pad_left, pad_right, pad_top and pad_bottom must be passed at compile time using -DPAD_LEFT, -DPAD_RIGHT, -DPAD_TOP and -DPAD_BOTTOM: e.g. -DPAD_LEFT=1, -DPAD_RIGHT=2, -DPAD_TOP=3 and -DPAD_BOTTOM=2
221 * @note The zero value to store in case we load values out-of-bounds must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0.0
222 * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1
223 * @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.
225 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
226 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
227 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
228 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
229 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
230 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
231 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
232 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
233 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
234 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
235 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
236 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
237 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
238 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
239 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
240 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
242 __kernel void im2col5x5_dchw(
243 TENSOR3D_DECLARATION(src),
244 IMAGE_DECLARATION(dst),
248 const int xc = get_global_id(0); // x coordinate in the convolved tensor
249 const int yc = get_global_id(1); // y coordinate in the convolved tensor
250 const int ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
251 const int batch = get_global_id(2) / KERNEL_DEPTH; // batch size
253 // Calculate input indices
254 const int xi = xc * STRIDE_X - PAD_LEFT;
255 const int yi = yc * STRIDE_Y - PAD_TOP;
257 // Calculate output indices
258 const int xo = ch * 25; // 5x5
259 const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
261 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
262 // Put 0 if the value is out-of-bound
263 int4 x0 = (int4)xi + (int4)(0, 1, 2, 3);
264 int4 y0 = (int4)yi + (int4)(0, 1, 2, 3);
268 // Check if we could have out-of-bounds elements in the x direction
269 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
270 x0_condition = CONVERT((x0 >= (int4)0 && x0 < (int4)SRC_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, 4));
271 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
272 y0_condition = CONVERT((y0 >= (int4)0 && y0 < (int4)SRC_HEIGHT), VEC_DATA_TYPE(COND_DATA_TYPE, 4));
273 COND_DATA_TYPE x1_condition = (COND_DATA_TYPE)(x1 >= 0 && x1 < SRC_WIDTH);
274 COND_DATA_TYPE y1_condition = (COND_DATA_TYPE)(y1 >= 0 && y1 < SRC_HEIGHT);
275 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
277 // Get input and output address
278 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * (int)src_stride_x + yi * (int)src_stride_y + ch * src_stride_z + batch * src_stride_w;
280 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
283 VEC_DATA_TYPE(DATA_TYPE, 4)
284 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
286 row01 = *((__global DATA_TYPE *)input_ptr + 4);
288 input_ptr += src_stride_y;
290 VEC_DATA_TYPE(DATA_TYPE, 4)
291 row10 = vload4(0, (__global DATA_TYPE *)input_ptr);
293 row11 = *((__global DATA_TYPE *)input_ptr + 4);
295 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
296 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
297 cond00 = x0_condition && (VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s0;
298 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
299 cond10 = x0_condition && (VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s1;
300 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y0_condition.s0);
301 COND_DATA_TYPE cond11 = (COND_DATA_TYPE)(x1_condition && y0_condition.s1);
303 // Replace with 0 if the value is not valid
304 row00 = select((VEC_DATA_TYPE(DATA_TYPE, 4))PAD_VALUE, row00, cond00);
305 row10 = select((VEC_DATA_TYPE(DATA_TYPE, 4))PAD_VALUE, row10, cond10);
306 row01 = select((DATA_TYPE)PAD_VALUE, row01, cond01);
307 row11 = select((DATA_TYPE)PAD_VALUE, row11, cond11);
308 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
310 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s0123, row01,
312 0, (__global DATA_TYPE *)output_ptr);
313 vstore2((VEC_DATA_TYPE(DATA_TYPE, 2))(row10.s3, row11), 0, (__global DATA_TYPE *)output_ptr + 8);
315 input_ptr += src_stride_y;
316 output_ptr += 10 * dst_stride_x;
320 VEC_DATA_TYPE(DATA_TYPE, 4)
321 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
323 row01 = *((__global DATA_TYPE *)input_ptr + 4);
325 input_ptr += src_stride_y;
327 VEC_DATA_TYPE(DATA_TYPE, 4)
328 row10 = vload4(0, (__global DATA_TYPE *)input_ptr);
330 row11 = *((__global DATA_TYPE *)input_ptr + 4);
332 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
333 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
334 cond00 = x0_condition && (VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s2;
335 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
336 cond10 = x0_condition && (VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s3;
337 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y0_condition.s2);
338 COND_DATA_TYPE cond11 = (COND_DATA_TYPE)(x1_condition && y0_condition.s3);
340 // Replace with 0 if the value is not valid
341 row00 = select((VEC_DATA_TYPE(DATA_TYPE, 4))PAD_VALUE, row00, cond00);
342 row10 = select((VEC_DATA_TYPE(DATA_TYPE, 4))PAD_VALUE, row10, cond10);
343 row01 = select((DATA_TYPE)PAD_VALUE, row01, cond01);
344 row11 = select((DATA_TYPE)PAD_VALUE, row11, cond11);
345 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
347 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s0123, row01,
349 0, (__global DATA_TYPE *)output_ptr);
350 vstore2((VEC_DATA_TYPE(DATA_TYPE, 2))(row10.s3, row11), 0, (__global DATA_TYPE *)output_ptr + 8);
352 input_ptr += src_stride_y;
353 output_ptr += 10 * dst_stride_x;
357 VEC_DATA_TYPE(DATA_TYPE, 4)
358 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
360 row01 = *((__global DATA_TYPE *)input_ptr + 4);
362 input_ptr += src_stride_y;
364 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
365 VEC_DATA_TYPE(COND_DATA_TYPE, 4)
366 cond00 = x0_condition && (VEC_DATA_TYPE(COND_DATA_TYPE, 4))y1_condition;
367 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y1_condition);
369 // Replace with 0 if the value is not valid
370 row00 = select((VEC_DATA_TYPE(DATA_TYPE, 4))PAD_VALUE, row00, cond00);
371 row01 = select((DATA_TYPE)PAD_VALUE, row01, cond01);
372 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
374 vstore4(row00, 0, (__global DATA_TYPE *)output_ptr);
375 *((__global DATA_TYPE *)output_ptr + 4) = row01;
377 output_ptr += 5 * dst_stride_x;
381 if(ch == (KERNEL_DEPTH - 1))
383 *((__global DATA_TYPE *)output_ptr) = 1.0f;
387 #endif // defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE)
389 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_DEPTH)
390 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM when the kernel size is 11x11
392 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
393 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
394 * @note The kernel depth must be passed at compile time using -DKERNEL_DEPTH: e.g. -DKERNEL_DEPTH=3
395 * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1
396 * @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.
398 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
399 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
400 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
401 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
402 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
403 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
404 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
405 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
406 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
407 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
408 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
409 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
410 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
411 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
412 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
413 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
415 __kernel void im2col11x11_padx0_pady0_dchw(
416 TENSOR3D_DECLARATION(src),
417 IMAGE_DECLARATION(dst),
421 const int xc = get_global_id(0); // x coordinate in the convolved tensor
422 const int yc = get_global_id(1); // y coordinate in the convolved tensor
423 const int ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
424 const int batch = get_global_id(2) / KERNEL_DEPTH; // batch size
426 // Calculate input indices
427 const int xi = xc * STRIDE_X;
428 const int yi = yc * STRIDE_Y;
430 // Calculate output indices
431 const int xo = ch * 121; // 11x11
432 const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
434 // Get input and output address
435 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * src_stride_x + yi * src_stride_y + ch * src_stride_z + batch * src_stride_w;
437 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
439 VEC_DATA_TYPE(DATA_TYPE, 8)
440 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
441 VEC_DATA_TYPE(DATA_TYPE, 3)
442 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
444 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
445 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
447 input_ptr += src_stride_y;
448 output_ptr += 11 * src_stride_x;
452 VEC_DATA_TYPE(DATA_TYPE, 8)
453 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
454 VEC_DATA_TYPE(DATA_TYPE, 3)
455 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
457 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
458 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
460 input_ptr += src_stride_y;
461 output_ptr += 11 * src_stride_x;
465 VEC_DATA_TYPE(DATA_TYPE, 8)
466 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
467 VEC_DATA_TYPE(DATA_TYPE, 3)
468 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
470 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
471 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
473 input_ptr += src_stride_y;
474 output_ptr += 11 * src_stride_x;
478 VEC_DATA_TYPE(DATA_TYPE, 8)
479 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
480 VEC_DATA_TYPE(DATA_TYPE, 3)
481 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
483 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
484 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
486 input_ptr += src_stride_y;
487 output_ptr += 11 * src_stride_x;
491 VEC_DATA_TYPE(DATA_TYPE, 8)
492 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
493 VEC_DATA_TYPE(DATA_TYPE, 3)
494 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
496 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
497 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
499 input_ptr += src_stride_y;
500 output_ptr += 11 * src_stride_x;
504 VEC_DATA_TYPE(DATA_TYPE, 8)
505 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
506 VEC_DATA_TYPE(DATA_TYPE, 3)
507 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
509 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
510 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
512 input_ptr += src_stride_y;
513 output_ptr += 11 * src_stride_x;
517 VEC_DATA_TYPE(DATA_TYPE, 8)
518 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
519 VEC_DATA_TYPE(DATA_TYPE, 3)
520 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
522 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
523 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
525 input_ptr += src_stride_y;
526 output_ptr += 11 * src_stride_x;
530 VEC_DATA_TYPE(DATA_TYPE, 8)
531 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
532 VEC_DATA_TYPE(DATA_TYPE, 3)
533 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
535 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
536 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
538 input_ptr += src_stride_y;
539 output_ptr += 11 * src_stride_x;
543 VEC_DATA_TYPE(DATA_TYPE, 8)
544 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
545 VEC_DATA_TYPE(DATA_TYPE, 3)
546 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
548 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
549 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
551 input_ptr += src_stride_y;
552 output_ptr += 11 * src_stride_x;
556 VEC_DATA_TYPE(DATA_TYPE, 8)
557 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
558 VEC_DATA_TYPE(DATA_TYPE, 3)
559 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
561 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
562 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
564 input_ptr += src_stride_y;
565 output_ptr += 11 * src_stride_x;
569 VEC_DATA_TYPE(DATA_TYPE, 8)
570 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
571 VEC_DATA_TYPE(DATA_TYPE, 3)
572 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
574 vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
575 vstore3((VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
577 output_ptr += 11 * src_stride_x;
581 if(ch == (KERNEL_DEPTH - 1))
583 *((__global DATA_TYPE *)output_ptr) = 1.0f;
587 #endif // defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_DEPTH)
588 #endif // !defined(FIXED_POINT_POSITION)
590 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(VECTOR_SIZE) && defined(WIDTH_MOD_VECTOR_SIZE)
591 /** This kernel reshapes the input tensor to a tensor used to perform convolution using GEMM when
592 * the kernel width is greater than 1 (except when the kernel size is 3x3) and pad_x == pad_y == 0.
594 * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float.
595 * @note The vector size must be passed at compile time using -DVECTOR_SIZE e.g. -DVECTOR_SIZE=4.
596 * @note The width modulo vector size must be passed at compile time using -DWIDTH_MOD_VECTOR_SIZE e.g. -DWIDTH_MOD_VECTOR_SIZE=3.
597 * @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.
599 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32
600 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
601 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
602 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
603 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
604 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
605 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
606 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
607 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
608 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
609 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
610 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
611 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
612 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
613 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
614 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
616 __kernel void im2col_generic_padx0_pady0_dchw(
617 TENSOR3D_DECLARATION(src),
618 IMAGE_DECLARATION(dst),
622 const int xc = get_global_id(0); // x coordinate in the convolved tensor
623 const int yc = get_global_id(1); // y coordinate in the convolved tensor
624 const int ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
625 const int batch = get_global_id(2) / KERNEL_DEPTH; // batch size
627 // Calculate input indices
628 const int xi = xc * STRIDE_X;
629 const int yi = yc * STRIDE_Y;
630 // Calculate output indices
631 const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
632 const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
633 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * src_stride_z + batch * src_stride_w;
634 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w)) + xo;
635 // Linearize convolution elements
636 for(int y = yi, y_e = yi + KERNEL_HEIGHT; y < y_e; ++y)
639 for(int x = xi, x_e = xi + KERNEL_WIDTH; x + VECTOR_SIZE <= x_e; x += VECTOR_SIZE, output_ptr += VECTOR_SIZE)
641 VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
642 row = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
644 (row, 0, output_ptr);
647 // Copy the remainder of the row by doing VLOAD(WIDTH_MOD_VECTOR_SIZE) and VSTORE(WIDTH_MOD_VECTOR_SIZE).
648 // Note that x and output_ptr have already been incremented by VECTOR_SIZE by the loop just before exit.
649 #if WIDTH_MOD_VECTOR_SIZE == 1
650 *output_ptr = *((__global DATA_TYPE *)(input_ptr + (last_x + VECTOR_SIZE) * src_stride_x + y * src_stride_y));
651 #elif WIDTH_MOD_VECTOR_SIZE > 1
652 VEC_DATA_TYPE(DATA_TYPE, WIDTH_MOD_VECTOR_SIZE)
653 row = VLOAD(WIDTH_MOD_VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + (last_x + VECTOR_SIZE) * src_stride_x + y * src_stride_y));
654 VSTORE(WIDTH_MOD_VECTOR_SIZE)
655 (row, 0, output_ptr);
656 #endif /* WIDTH_MOD_VECTOR_SIZE */
657 output_ptr += WIDTH_MOD_VECTOR_SIZE;
658 } /* End of loop over KERNEL_HEIGHT */
661 if(ch == (KERNEL_DEPTH - 1))
663 #ifdef FIXED_POINT_POSITION
664 *output_ptr = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
665 #else // FIXED_POINT_POSITION
667 #endif // FIXED_POINT_POSITION
671 #endif //defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(VECTOR_SIZE) && defined(WIDTH_MOD_VECTOR_SIZE)
673 #if defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE)
674 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM.
676 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
677 * @note The width and height of the input tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT: e.g. -DSRC_WIDTH=128 and -DSRC_HEIGHT=128
678 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
679 * @note The kernel width, height and depth must be passed at compile time using -DKERNEL_WIDTH, -DKERNEL_HEIGHT and -DKERNEL_DEPTH: e.g. -DKERNEL_WIDTH=3, -DKERNEL_HEIGHT=3 and -DKERNEL_DEPTH=64
680 * @note The pad_left, pad_right, pad_top and pad_bottom must be passed at compile time using -DPAD_LEFT, -DPAD_RIGHT, -DPAD_TOP and -DPAD_BOTTOM: e.g. -DPAD_LEFT=1, -DPAD_RIGHT=2, -DPAD_TOP=3 and -DPAD_BOTTOM=2
681 * @note The zero value to store in case we load values out-of-bounds must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0.0
682 * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1
683 * @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.
685 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
686 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
687 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
688 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
689 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
690 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
691 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
692 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
693 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
694 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
695 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
696 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
697 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
698 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
699 * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
700 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
702 __kernel void im2col_generic_dchw(
703 TENSOR3D_DECLARATION(src),
704 IMAGE_DECLARATION(dst),
708 const int xc = get_global_id(0); // x coordinate in the convolved tensor
709 const int yc = get_global_id(1); // y coordinate in the convolved tensor
710 const int ch = get_global_id(2) % KERNEL_DEPTH; // input feature map
711 const int batch = get_global_id(2) / KERNEL_DEPTH; // batch size
713 // Calculate input indices
714 const int xi = xc * STRIDE_X - PAD_LEFT;
715 const int yi = yc * STRIDE_Y - PAD_TOP;
717 // Calculate output indices
718 const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
719 const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
721 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * src_stride_z + batch * src_stride_w;
722 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w)) + xo;
724 // Linearize convolution elements
725 for(int y = yi, y_e = yi + KERNEL_HEIGHT; y < y_e; ++y)
727 for(int x = xi, x_e = xi + KERNEL_WIDTH; x < x_e; ++x, ++output_ptr)
729 #if PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
730 *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
731 #else // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
732 if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
734 *output_ptr = PAD_VALUE;
738 *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
740 #endif // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
745 if(ch == (KERNEL_DEPTH - 1))
747 #ifdef FIXED_POINT_POSITION
748 *output_ptr = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
749 #else // FIXED_POINT_POSITION
751 #endif // FIXED_POINT_POSITION
755 #endif // defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE)
757 /**This kernel reshapes the input tensor to a tensor used to perform convolution using GEMM when
758 * the kernel width and height are the same of width and height of the input tensor
760 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
761 * @note In case biases will be added in late stage, -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
763 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32
764 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
765 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
766 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
767 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
768 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
769 * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
770 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
771 * @param[out] dst_ptr Pointer to the destination tensor. Same as @p src_ptr
772 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
773 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
774 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
775 * @param[in] width The width of the input tensor
776 * @param[in] height The height of the input tensor
778 __kernel void im2col_reduced_dchw(
779 TENSOR3D_DECLARATION(src),
780 VECTOR_DECLARATION(dst),
781 uint width, uint height)
783 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
785 const uint image_size = width * height;
787 __global uchar *tmp_out_ptr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) + get_global_id(1) * width + get_global_id(2) * image_size) * dst_stride_x;
789 *((__global DATA_TYPE *)tmp_out_ptr) = *((__global DATA_TYPE *)src.ptr);
792 // If it is the last thread in the 3 dimensional workgroup
793 if(get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1))
795 tmp_out_ptr += dst_stride_x;
796 #ifdef FIXED_POINT_POSITION
797 *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
798 #else // FIXED_POINT_POSITION
799 *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)1.0f;
800 #endif // FIXED_POINT_POSITION
804 #endif // defined(DATA_TYPE) && defined(ELEMENT_SIZE)