xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/nhwc/im2col.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1/*
2 * Copyright (c) 2018-2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
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:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
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
22 * SOFTWARE.
23 */
24#include "helpers.h"
25
26#define VECTOR_N VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
27#define COND_N SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
28
29#if defined(IM2COL_3X3) || defined(IM2COL_9X9)
30/** Store a 1x9 row or a 3x3 block in a boundary-aware manner to avoid paddings in the channel dimension
31 *  @name IM2COL1X9_NHWC_STORE
32 *
33 *  @note To use this macro for a 3x3 block, @p ROW has to be 0
34 *
35 * @param[in] VECTOR_SIZE          The non-boundary vector width of @p DATA. Supported: 1(scalar), 2, 3, 4, 8, 16
36 * @param[in] BOUNDARY_VECTOR_SIZE The boundary vector width of @p DATA. Supported: 1-16, but has to be <= @p size
37 * @param[in] DATA_TYPE            Data type of @p DATA
38 * @param[in] SRC_DEPTH            Input channel size / depth
39 * @param[in] DATA                 Value variable base name
40 * @param[in] ROW                  The row number to store. Supported: 0-8
41 * @param[in] OUTPUT_PTR           Output pointer
42 * @{
43 */
44#if defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
45#define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR)         \
46    const bool at_channel_boundary = get_global_id(0) == 0;                                                          \
47    if(at_channel_boundary)                                                                                          \
48    {                                                                                                                \
49        IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
50    }                                                                                                                \
51    else                                                                                                             \
52    {                                                                                                                \
53        IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR)                    \
54    }
55#else // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
56#define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
57    IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR)
58#endif // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
59
60#define IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
61    VSTORE(VECTOR_SIZE)                                                                           \
62    (DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH);                 \
63    VSTORE(VECTOR_SIZE)                                                                           \
64    (DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH);                 \
65    VSTORE(VECTOR_SIZE)                                                                           \
66    (DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH);                 \
67    VSTORE(VECTOR_SIZE)                                                                           \
68    (DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH);                 \
69    VSTORE(VECTOR_SIZE)                                                                           \
70    (DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH);                 \
71    VSTORE(VECTOR_SIZE)                                                                           \
72    (DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH);                 \
73    VSTORE(VECTOR_SIZE)                                                                           \
74    (DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH);                 \
75    VSTORE(VECTOR_SIZE)                                                                           \
76    (DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH);                 \
77    VSTORE(VECTOR_SIZE)                                                                           \
78    (DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH);
79
80#define IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
81    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
82    (DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH);                                    \
83    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
84    (DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH);                                    \
85    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
86    (DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH);                                    \
87    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
88    (DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH);                                    \
89    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
90    (DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH);                                    \
91    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
92    (DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH);                                    \
93    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
94    (DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH);                                    \
95    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
96    (DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH);                                    \
97    VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)                                                                \
98    (DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH);
99/** @}*/
100#endif // defined(IM2COL_3X3) || defined(IM2COL_9X9)
101
102#if defined(IM2COL_3X3)
103/** This kernel performs im2col when the kernel size is 3x3 and the data layout is NHWC
104 *
105 * @note This kernel computes VECTOR_SIZE elements
106 * @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
107 * @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
108 * @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
109 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
110 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
111 * @note The kernel depth must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=3
112 * @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
113 * @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.
114 *
115 * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
116 * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
117 * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
118 * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
119 * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
120 * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
121 * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
122 * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
123 * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
124 * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
125 * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
126 * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
127 * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
128 * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
129 * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes).
130 * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes).
131 */
132__kernel void im2col3x3_nhwc(
133    TENSOR3D_DECLARATION(src),
134    IMAGE_DECLARATION(dst),
135    uint src_stride_w,
136    uint dst_stride_w)
137{
138    // input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
139    const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
140    const int ch           = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
141    const int yo           = get_global_id(1);
142    const int batch        = get_global_id(2); // batch size
143
144    // Calculate input indices
145    const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
146    const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
147
148    // Get input and output address
149    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w;
150    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
151
152    int  yi_coord = 0;
153    int3 offset   = 0;
154
155    // Clamp xi
156    int3 xi_offset = ((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT);
157#if PAD_LEFT != 0 || PAD_RIGHT != 0
158#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
159    xi_offset = CLAMP(xi_offset, (int3)0, (int3)(SRC_WIDTH - 1));
160#endif // PAD_LEFT != 0 || PAD_RIGHT != 0
161    // Multiply by src_stride_y as the width (X) dimension here is the second (y) dimension in src NHWC tensor
162    xi_offset *= (int3)src_stride_y;
163
164    // Out-of-bound condition for X
165    int3 x_cond = (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) < (int3)0) || (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) >= (int3)SRC_WIDTH);
166
167    // yi == 0
168    // Clamp yi
169    // yi_coord is casted to unsigned int in order to use just a min() operation
170    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
171    // This is a trick so that the values loaded in the padding areas are always from the last row (SRC_HEIGHT - 1),
172    // because of the negative yi_coord wrap-around, but it gets overwritten by PAD_VALUE immediately as the wrap-around
173    // also causes y_cond (y padding condition) to be satisfied
174    yi_coord = yi - (int)PAD_TOP;
175
176    // Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
177#if PAD_TOP != 0 || PAD_BOTTOM != 0
178    yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
179#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
180
181    // Compute offset
182    offset = xi_offset + (yi_coord * (int)src_stride_z);
183
184    // Load input values
185    VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
186    VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
187    VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
188
189#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
190    // Replace invalid values with PAD_VALUE
191    int y_cond = (int)((uint)(yi - (int)PAD_TOP) >= (uint)(SRC_HEIGHT));
192    values0    = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
193    values1    = select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
194    values2    = select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
195#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
196
197    // yi == 1
198    // Clamp yi_coord (it can be negative if PAD_TOP > 1)
199    yi_coord = yi - (int)PAD_TOP + 1 * DILATION_Y;
200
201    // Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
202#if PAD_TOP != 0 || PAD_BOTTOM != 0
203    yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
204#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
205
206    // Compute offset
207    offset = xi_offset + (yi_coord * (int)src_stride_z);
208
209    // Load input values
210    VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
211    VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
212    VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
213
214#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
215    // Replace invalid values with zeros
216    y_cond  = (int)((uint)(yi - (int)PAD_TOP + 1 * DILATION_Y) >= (uint)(SRC_HEIGHT));
217    values3 = select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
218    values4 = select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
219    values5 = select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
220#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
221
222    // yi == 2
223    // Clamp yi_coord
224    yi_coord = yi - (int)PAD_TOP + 2 * DILATION_Y;
225
226    // Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
227#if PAD_TOP != 0 || PAD_BOTTOM != 0
228    yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
229#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
230
231    // Compute offset
232    offset = xi_offset + (yi_coord * (int)src_stride_z);
233
234    // Load input values
235    VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
236    VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
237    VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
238
239#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
240    // Replace invalid values with PAD_VALUE
241    y_cond  = (int)((uint)(yi - (int)PAD_TOP + 2 * DILATION_Y) >= (uint)(SRC_HEIGHT));
242    values6 = select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
243    values7 = select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
244    values8 = select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
245#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
246
247    // Store in a boundary-aware way to avoid padding
248    IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, 0, output_ptr)
249
250#ifdef HAS_BIAS
251    // We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
252    // added at the end of the channel, while the boundary vec is at the beginning of the channel.
253    // The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
254    // the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
255    // See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
256    if((ch + VECTOR_SIZE) >= SRC_DEPTH)
257    {
258        *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 9) = 1.0f;
259    }
260#endif // HAS_BIAS
261}
262#endif // defined(IM2COL_3X3)
263
264#if defined(IM2COL_9X9)
265#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
266#define IM2COL1x9(i)                                                                                         \
267    ({                                                                                                       \
268        yi_coord = yi - (int)PAD_TOP + i * DILATION_Y;                                                       \
269        yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));                                              \
270        \
271        offset0 = xi_offset0 + (yi_coord * (int)src_stride_z);                                               \
272        offset1 = xi_offset1 + (yi_coord * (int)src_stride_z);                                               \
273        \
274        VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0));            \
275        VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1));            \
276        VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2));            \
277        VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3));            \
278        VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4));            \
279        VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5));            \
280        VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6));            \
281        VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7));            \
282        VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1));               \
283        \
284        int y_cond = (int)((uint)(yi - (int)PAD_TOP + i * DILATION_Y) >= (uint)(SRC_HEIGHT));                \
285        values0    = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s0))); \
286        values1    = select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s1))); \
287        values2    = select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s2))); \
288        values3    = select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s3))); \
289        values4    = select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s4))); \
290        values5    = select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s5))); \
291        values6    = select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s6))); \
292        values7    = select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s7))); \
293        values8    = select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond1)));    \
294        \
295        IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \
296    })
297#else // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
298#define IM2COL1x9(i)                                                                                         \
299    ({                                                                                                       \
300        yi_coord = yi - (int)PAD_TOP + i * DILATION_Y;                                                       \
301        yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));                                              \
302        \
303        offset0 = xi_offset0 + (yi_coord * (int)src_stride_z);                                               \
304        offset1 = xi_offset1 + (yi_coord * (int)src_stride_z);                                               \
305        \
306        VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0));            \
307        VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1));            \
308        VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2));            \
309        VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3));            \
310        VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4));            \
311        VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5));            \
312        VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6));            \
313        VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7));            \
314        VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1));               \
315        \
316        IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \
317    })
318#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
319
320/** This kernel performs im2col when the kernel size is 9x9 and the data layout is NHWC
321 *
322 * @note This kernel computes VECTOR_SIZE elements
323 * @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
324 * @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
325 * @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
326 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
327 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
328 * @note The kernel depth must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=3
329 * @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
330 * @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.
331 *
332 * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
333 * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
334 * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
335 * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
336 * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
337 * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
338 * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
339 * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
340 * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
341 * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
342 * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
343 * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
344 * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
345 * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
346 * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes).
347 * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes).
348 */
349__kernel void im2col9x9_nhwc(
350    TENSOR3D_DECLARATION(src),
351    IMAGE_DECLARATION(dst),
352    uint src_stride_w,
353    uint dst_stride_w)
354{
355    // input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
356    const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
357    const int ch           = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
358    const int yo           = get_global_id(1);
359    const int batch        = get_global_id(2); // batch size
360
361    // Calculate input indices
362    const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
363    const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
364
365    // Get input and output address
366    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w;
367    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
368
369    int  yi_coord = 0;
370    int8 offset0  = 0;
371    int  offset1  = 0;
372
373    // Clamp xi
374    int8 xi_offset0 = ((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT);
375    int  xi_offset1 = ((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT);
376
377#if PAD_LEFT != 0 || PAD_RIGHT != 0
378#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
379    xi_offset0 = CLAMP(xi_offset0, (int8)0, (int8)(SRC_WIDTH - 1));
380    xi_offset1 = CLAMP(xi_offset1, (int)0, (int)(SRC_WIDTH - 1));
381#endif // PAD_LEFT != 0 || PAD_RIGHT != 0
382    xi_offset0 *= (int8)src_stride_y;
383    xi_offset1 *= (int)src_stride_y;
384
385    // Out-of-bound condition for X
386    int8 x_cond0 = (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) < (int8)0) || (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) >= (int8)SRC_WIDTH);
387    int  x_cond1 = (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) < (int)0) || (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
388
389    IM2COL1x9(0);
390    IM2COL1x9(1);
391    IM2COL1x9(2);
392    IM2COL1x9(3);
393    IM2COL1x9(4);
394    IM2COL1x9(5);
395    IM2COL1x9(6);
396    IM2COL1x9(7);
397    IM2COL1x9(8);
398
399#ifdef HAS_BIAS
400    // We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
401    // added at the end of the channel, while the boundary vec is at the beginning of the channel.
402    // The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
403    // the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
404    // See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
405    if((ch + VECTOR_SIZE) >= SRC_DEPTH)
406    {
407        *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 81) = 1.0f;
408    }
409#endif // HAS_BIAS
410}
411#endif // defined(IM2COL_9X9)
412
413#if defined(IM2COL_GENERIC)
414/** This opencl kernel performs a generic im2col implementation when the data layout is NHWC
415 *
416 * @note This kernel computes VECTOR_SIZE elements
417 * @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
418 * @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
419 * @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
420 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
421 * @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
422 * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
423 * @note The kernel width, height and depth must be passed at compile time using -DKERNEL_WIDTH, -DKERNEL_HEIGHT and -DSRC_DEPTH: e.g. -DKERNEL_WIDTH=3, -DKERNEL_HEIGHT=3 and -DSRC_DEPTH=64
424 * @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
425 * @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
426 * @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
427 * @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1
428 * @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.
429 *
430 * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
431 * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
432 * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
433 * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
434 * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
435 * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
436 * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
437 * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
438 * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
439 * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
440 * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
441 * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
442 * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
443 * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
444 * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes).
445 * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes).
446 */
447__kernel void im2col_generic_nhwc(
448    TENSOR3D_DECLARATION(src),
449    IMAGE_DECLARATION(dst),
450    uint src_stride_w,
451    uint dst_stride_w)
452{
453    // input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
454    const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
455    const int ch           = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
456    const int yo           = get_global_id(1);
457    const int batch        = get_global_id(2); // batch size
458
459    // Calculate input indices
460    const int xi = (yo % CONVOLVED_WIDTH) * STRIDE_X;
461    const int yi = (yo / (int)CONVOLVED_WIDTH) * STRIDE_Y;
462
463    // Get input and output address
464    const int stride_x         = ch * sizeof(DATA_TYPE);
465    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + stride_x + batch * (int)src_stride_w;
466    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + stride_x + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
467
468    int i = 0;
469    for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
470    {
471        // Clamp yi_coord
472        int yi_coord = yi + yk * DILATION_Y - (int)PAD_TOP;
473        yi_coord     = clamp(yi_coord, (int)0, (int)(SRC_HEIGHT - 1));
474
475        // Out-of-bound condition for Y
476        int y_border_condition = ((yi + yk * DILATION_Y - (int)PAD_TOP) < (int)0) || ((yi + yk * DILATION_Y - (int)PAD_TOP) >= (int)SRC_HEIGHT);
477
478        for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
479        {
480            // Clamp xi_coord
481            int xi_coord = (xi + xk * DILATION_X - (int)PAD_LEFT);
482            xi_coord     = clamp(xi_coord, (int)0, (int)(SRC_WIDTH - 1));
483
484            // Out-of-bound condition for X
485            int x_border_condition = ((xi + xk * DILATION_X - (int)PAD_LEFT) < (int)0) || ((xi + xk * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
486
487            int offset = xi_coord * (int)src_stride_y + (yi_coord * (int)src_stride_z);
488
489            VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset));
490
491#if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
492            // Replace with PAD_VALUE if the value is out-of-bound
493            values0 = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)x_border_condition || (COND_N)(y_border_condition)));
494#endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
495
496            // Store in a boundary-aware way to avoid padding
497#if BOUNDARY_VECTOR_SIZE != VECTOR_SIZE
498            const bool at_channel_boundary = get_global_id(0) == 0;
499            if(at_channel_boundary)
500            {
501                VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)
502                (values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
503            }
504            else // at_channel_boundary
505#endif           // BOUNDARY_VECTOR_SIZE != VECTOR_SIZE
506            {
507                VSTORE(VECTOR_SIZE)
508                (values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
509            }
510            i++;
511        }
512    }
513
514#ifdef HAS_BIAS
515    // We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
516    // added at the end of the channel, while the boundary vec is at the beginning of the channel.
517    // The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
518    // the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
519    // See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
520    if((ch + VECTOR_SIZE) >= SRC_DEPTH)
521    {
522        *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * KERNEL_WIDTH * KERNEL_HEIGHT) = 1.0f;
523    }
524#endif // HAS_BIAS
525}
526#endif // defined(IM2COL_GENERIC)