xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/common/arg_min_max.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1/*
2 * Copyright (c) 2019-2021 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#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE_OUTPUT)
27
28#define VEC_TYPE_IN VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
29#define VEC_TYPE_OUT VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
30#define VEC_SELECT_IN SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
31#define VEC_SIGNED_INT_IN SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
32
33#if defined(FLOAT_DATA_TYPE)
34#define ISGREATER(x, y) (VEC_SELECT_IN) isgreater(x, y)
35#define ISLESS(x, y) (VEC_SELECT_IN) isless(x, y)
36#else // !FLOAT_DATA_TYPE
37#if defined(WIDTH)
38#define ISGREATER(x, y) (x > y) ? 1 : 0
39#define ISLESS(x, y) (x < y) ? 1 : 0
40#else // !defined(WIDTH)
41#define ISGREATER(x, y) select((VEC_SIGNED_INT_IN)0, (VEC_SIGNED_INT_IN)-1, (VEC_SIGNED_INT_IN)(x > y))
42#define ISLESS(x, y) select((VEC_SIGNED_INT_IN)0, (VEC_SIGNED_INT_IN)-1, (VEC_SIGNED_INT_IN)(x < y))
43#endif // defined(WIDTH)
44#endif // defined(FLOAT_DATA_TYPE)
45
46#if defined(ARG_MAX)
47#define CONDITION_TO_USE(x, y) ISGREATER(x, y)
48#elif defined(ARG_MIN)
49#define CONDITION_TO_USE(x, y) ISLESS(x, y)
50#else // !(defined(ARG_MAX) || defined(ARG_MIN))
51#error "Unsupported reduction operation!"
52#endif // defined(ARG_MAX)
53
54#if defined(WIDTH)
55#if defined(ARG_MIN)
56#if defined(PREV_OUTPUT)
57/** Find index minimum value of a vector
58 *
59 * @param[in] input Pointer to the first value.
60 *
61 * @return index of the vector.
62 */
63inline DATA_TYPE_OUTPUT arg_idx_min_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx)
64{
65    int end_elem = (x_idx + 1) * 16;
66    if(end_elem > WIDTH)
67    {
68        end_elem = WIDTH - x_idx * 16;
69    }
70    DATA_TYPE_OUTPUT res = prev_res[0];
71    for(int x_v = 1; x_v < end_elem; ++x_v)
72    {
73        res = select(res, prev_res[x_v], *(input + prev_res[x_v]) < * (input + res));
74    }
75    return res;
76}
77#else // !defined(PREV_OUTPUT)
78/** Find index minimum value of a vector
79 *
80 * @param[in] input Pointer to the first value.
81 *
82 * @return index of the vector.
83 */
84inline DATA_TYPE_OUTPUT arg_idx_min(__global const DATA_TYPE *input, const int x_idx)
85{
86#if WIDTH < 16
87    DATA_TYPE_OUTPUT res = 0;
88    for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v)
89    {
90        res = select(res, x_v, *(input + x_v) < * (input + res));
91    }
92    return res;
93#else  // WIDTH >= 16
94    int       x_elem   = x_idx * 16;
95    const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH);
96    x_elem -= x_goback;
97
98    VEC_DATA_TYPE(DATA_TYPE, 16)
99    in = vload16(0, input - x_goback);
100    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)
101    res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
102
103    SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 8)
104    idx_sel       = (in.s01234567 <= in.s89abcdef);
105    in.s01234567  = select(in.s89abcdef, in.s01234567, idx_sel);
106    res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8));
107
108    idx_sel.s0123 = (in.s0123 < in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 4)));
109    in.s0123      = select(in.s4567, in.s0123, idx_sel.s0123);
110    res.s0123     = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4));
111
112    idx_sel.s01 = (in.s01 < in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 2)));
113    in.s01      = select(in.s23, in.s01, idx_sel.s01);
114    res.s01     = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
115
116    idx_sel.s0 = (in.s0 < in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), SIGNED_INT_DATA_TYPE(DATA_TYPE)));
117    res.s0     = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
118
119    return res.s0 + x_elem;
120#endif // WIDTH < 16
121}
122#endif // defined(PREV_OUTPUT)
123#endif // defined(ARG_MIN)
124#if defined(ARG_MAX)
125#if defined(PREV_OUTPUT)
126/** Find index maximum value of a vector
127 *
128 * @param[in] input Pointer to the first value.
129 *
130 * @return index of the vector.
131 */
132inline DATA_TYPE_OUTPUT arg_idx_max_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx)
133{
134    int end_elem = (x_idx + 1) * 16;
135    if(end_elem > WIDTH)
136    {
137        end_elem = WIDTH - x_idx * 16;
138    }
139    DATA_TYPE_OUTPUT res = prev_res[0];
140    for(int x_v = 1; x_v < end_elem; ++x_v)
141    {
142        res = select(res, prev_res[x_v], *(input + prev_res[x_v]) > *(input + res));
143    }
144    return res;
145}
146#else // !defined(PREV_OUTPUT)
147/** Find index maximum value of a vector
148 *
149 * @param[in] input Pointer to the first value.
150 *
151 * @return index of the vector.
152 */
153inline DATA_TYPE_OUTPUT arg_idx_max(__global const DATA_TYPE *input, const int x_idx)
154{
155#if WIDTH < 16
156    DATA_TYPE_OUTPUT res = 0;
157    for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v)
158    {
159        res = select(res, x_v, *(input + x_v) > *(input + res));
160    }
161    return res;
162#else  // WIDTH >= 16
163    int       x_elem   = x_idx * 16;
164    const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH);
165    x_elem -= x_goback;
166
167    VEC_DATA_TYPE(DATA_TYPE, 16)
168    in = vload16(0, input - x_goback);
169    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)
170    res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
171
172    SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 8)
173    idx_sel       = (in.s01234567 >= in.s89abcdef);
174    in.s01234567  = select(in.s89abcdef, in.s01234567, idx_sel);
175    res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8));
176
177    idx_sel.s0123 = (in.s0123 > in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 4)));
178    in.s0123      = select(in.s4567, in.s0123, idx_sel.s0123);
179    res.s0123     = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4));
180
181    idx_sel.s01 = (in.s01 > in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 2)));
182    in.s01      = select(in.s23, in.s01, idx_sel.s01);
183    res.s01     = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
184
185    idx_sel.s0 = (in.s0 > in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), SIGNED_INT_DATA_TYPE(DATA_TYPE)));
186    res.s0     = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
187
188    return res.s0 + x_elem;
189#endif // WIDTH < 16
190}
191#endif // defined(PREV_OUTPUT)
192#endif // defined(ARG_MAX)
193
194/** This kernel performs parallel reduction given an operation on x-axis.
195 *
196 * @note In case the results of previous stages are passed the flag PREV_OUTPUT has to be passed using -DPREV_OUTPUT
197 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
198 * @note The data type of the output must be passed at compile time using -DDATA_TYPE_OUTPUT: e.g. -DDATA_TYPE_OUTPUT=uint
199 * @note The arg_max flag must be passed at compile time using -DARG_MAX if we want to compute the ArgMax
200 * @note The arg_min flag must be passed at compile time using -DARG_MIN if we want to compute the ArgMin
201 *
202 * @param[in] src_ptr                                   Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/F32
203 * @param[in] src_stride_x                              Stride of the source tensor in X dimension (in bytes)
204 * @param[in] src_step_x                                src_stride_x * number of elements along X processed per workitem(in bytes)
205 * @param[in] src_stride_y                              Stride of the source tensor in Y dimension (in bytes)
206 * @param[in] src_step_y                                src_stride_y * number of elements along Y processed per workitem(in bytes)
207 * @param[in] src_offset_first_element_in_bytes         The offset of the first element in the source tensor
208 * @param[in] prev_res_ptr                              (Optional) Pointer to previous results tensor. Supported data types: U32/S32
209 * @param[in] prev_res_stride_x                         (Optional) Stride of the output tensor in X dimension (in bytes)
210 * @param[in] prev_res_step_x                           (Optional) prev_res_stride_x * number of elements along X processed per workitem(in bytes)
211 * @param[in] prev_res_stride_y                         (Optional) Stride of the output tensor in Y dimension (in bytes)
212 * @param[in] prev_res_step_y                           (Optional) prev_res_stride_y * number of elements along Y processed per workitem(in bytes)
213 * @param[in] prev_res_offset_first_element_in_bytes    (Optional) The offset of the first element in the previous results tensor
214 * @param[in] partial_res_ptr                           The local buffer to hold partial result values. Supported data types: U32/S32
215 * @param[in] partial_res_stride_x                      Stride of the output tensor in X dimension (in bytes)
216 * @param[in] partial_res_step_x                        partial_res_stride_x * number of elements along X processed per workitem(in bytes)
217 * @param[in] partial_res_stride_y                      Stride of the output tensor in Y dimension (in bytes)
218 * @param[in] partial_res_step_y                        partial_res_stride_y * number of elements along Y processed per workitem(in bytes)
219 * @param[in] partial_res_offset_first_element_in_bytes The offset of the first element in the source tensor
220 * @param[in] local_results                             Local buffer for storing the partial result
221 */
222__kernel void arg_min_max_x(
223    IMAGE_DECLARATION(src),
224#if defined(PREV_OUTPUT)
225    IMAGE_DECLARATION(prev_res),
226#endif // defined(PREV_OUTPUT)
227    IMAGE_DECLARATION(partial_res),
228    __local DATA_TYPE_OUTPUT *local_results)
229{
230#if defined(PREV_OUTPUT)
231    Image src      = CONVERT_TO_IMAGE_STRUCT_NO_STEP(src);
232    Image prev_res = CONVERT_TO_IMAGE_STRUCT(prev_res);
233#else  // !defined(PREV_OUTPUT)
234    Image src                      = CONVERT_TO_IMAGE_STRUCT(src);
235#endif // defined(PREV_OUTPUT)
236    Image partial_res = CONVERT_TO_IMAGE_STRUCT(partial_res);
237
238    unsigned int lsize = get_local_size(0);
239    unsigned int lid   = get_local_id(0);
240
241    const uint     x_idx                 = get_global_id(0);
242    const uint     y_idx                 = get_global_id(1);
243    const __global DATA_TYPE *src_in_row = (const __global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + y_idx * src_step_y);
244
245    for(unsigned int y = 0; y < get_local_size(1); ++y)
246    {
247#if defined(ARG_MAX)
248#if defined(PREV_OUTPUT)
249        local_results[lid] = arg_idx_max_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx);
250#else  // !defined(PREV_OUTPUT)
251        local_results[lid] = arg_idx_max((__global DATA_TYPE *)offset(&src, 0, y), x_idx);
252#endif // defined(PREV_OUTPUT)
253#else  // defined(ARG_MIN)
254#if defined(PREV_OUTPUT)
255        local_results[lid]         = arg_idx_min_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx);
256#else  // !defined(PREV_OUTPUT)
257        local_results[lid] = arg_idx_min((__global DATA_TYPE *)offset(&src, 0, y), x_idx);
258#endif // defined(PREV_OUTPUT)
259#endif // defined(ARG_MAX) || defined(ARG_MIN)
260
261        barrier(CLK_LOCAL_MEM_FENCE);
262
263        // Looking for the next highest power of 2 (maximum value of lsize is 8)
264        unsigned int middle = lsize - 1;
265        middle |= middle >> 1;
266        middle |= middle >> 2;
267        middle += 1;
268        // Perform parallel reduction
269        for(unsigned int i = middle; i > 0; i >>= 1)
270        {
271            if(lid < i && lid + i < lsize)
272            {
273                DATA_TYPE tmp0 = *(src_in_row + local_results[lid]);
274                DATA_TYPE tmp1 = *(src_in_row + local_results[lid + i]);
275#if defined(ARG_MAX)
276                local_results[lid] = select(
277                                         local_results[lid],
278                                         local_results[lid + i],
279                                         ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 < tmp1));
280#else  // defined(ARG_MIN)
281                local_results[lid] = select(
282                                         local_results[lid],
283                                         local_results[lid + i],
284                                         ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 > tmp1));
285#endif // defined(ARG_MAX) || defined(ARG_MIN)
286            }
287            barrier(CLK_LOCAL_MEM_FENCE);
288        }
289
290        if(lid == 0)
291        {
292            ((__global DATA_TYPE_OUTPUT *)offset(&partial_res, get_group_id(0), y))[0] = local_results[0];
293        }
294    }
295}
296#endif // defined(WIDTH)
297
298#if defined(HEIGHT)
299/** This kernel performs reduction on y-axis.
300 *
301 * @note The input data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
302 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
303 * @note The data type of the output must be passed at compile time using -DDATA_TYPE_OUTPUT: e.g. -DDATA_TYPE_OUTPUT=uint
304 * @note The height size must be passed at compile time using -DHEIGHT e.g. -DHEIGHT=128
305 *
306 * @param[in] input_ptr                            Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/F32
307 * @param[in] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
308 * @param[in] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
309 * @param[in] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
310 * @param[in] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
311 * @param[in] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
312 * @param[in] output_ptr                           The local buffer to hold sumed values. Supported data types: U32/S32
313 * @param[in] output_stride_x                      Stride of the output tensor in X dimension (in bytes)
314 * @param[in] output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
315 * @param[in] output_stride_y                      Stride of the output tensor in Y dimension (in bytes)
316 * @param[in] output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
317 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor
318 */
319__kernel void arg_min_max_y(
320    IMAGE_DECLARATION(input),
321    IMAGE_DECLARATION(output))
322{
323    const int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
324
325    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y;
326    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE_OUTPUT) + get_global_id(1) * output_stride_y;
327
328    VEC_TYPE_IN res = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr), VEC_TYPE_IN);
329
330    VEC_TYPE_OUT indx0 = 0;
331    for(DATA_TYPE_OUTPUT y = 1; y < HEIGHT; ++y)
332    {
333        VEC_TYPE_IN in = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + y * input_stride_y)), VEC_TYPE_IN);
334
335        VEC_TYPE_OUT cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_TYPE_OUT);
336        indx0                  = select(indx0, (VEC_TYPE_OUT)y, cond_conv);
337        res                    = select(res, in, CONDITION_TO_USE(in, res));
338    }
339
340    // Store result
341    STORE_VECTOR_SELECT(indx, DATA_TYPE_OUTPUT, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
342}
343#endif // defined(HEIGHT)
344
345#if defined(DEPTH) && !defined(BATCH)
346/** This kernel performs reduction on z-axis.
347 *
348 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
349 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
350 * @note The depth size must be passed at compile time using -DDEPTH e.g. -DDEPTH=128
351 *
352 * @param[in] input_ptr                            Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/F32
353 * @param[in] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
354 * @param[in] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
355 * @param[in] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
356 * @param[in] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
357 * @param[in] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
358 * @param[in] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
359 * @param[in] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
360 * @param[in] output_ptr                           The local buffer to hold sumed values. Supported data types: U32/S32
361 * @param[in] output_stride_x                      Stride of the output tensor in X dimension (in bytes)
362 * @param[in] output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
363 * @param[in] output_stride_y                      Stride of the output tensor in Y dimension (in bytes)
364 * @param[in] output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
365 * @param[in] output_stride_z                      Stride of the output tensor in Z dimension (in bytes)
366 * @param[in] output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
367 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor
368 */
369__kernel void arg_min_max_z(
370    TENSOR3D_DECLARATION(input),
371    TENSOR3D_DECLARATION(output))
372{
373    const int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
374
375    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
376    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE_OUTPUT) + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
377
378    VEC_TYPE_IN res = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr), VEC_TYPE_IN);
379
380    VEC_TYPE_OUT indx0 = 0;
381    for(DATA_TYPE_OUTPUT z = 1; z < DEPTH; ++z)
382    {
383        VEC_TYPE_IN in = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + z * input_stride_z)), VEC_TYPE_IN);
384
385        VEC_TYPE_OUT cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_TYPE_OUT);
386        indx0                  = select(indx0, (VEC_TYPE_OUT)z, cond_conv);
387        res                    = select(res, in, CONDITION_TO_USE(in, res));
388    }
389
390    // Store result
391    STORE_VECTOR_SELECT(indx, DATA_TYPE_OUTPUT, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
392}
393#endif /* defined(DEPTH)  && !defined(BATCH) */
394
395#if defined(BATCH) && defined(DEPTH)
396/** This kernel performs reduction on w-axis.
397 *
398 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
399 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
400 * @note The batch size must be passed at compile time using -DBATCH e.g. -DBATCH=128
401 * @note The depth size must be passed at compile time using -DBATCH e.g. -DDEPTH=128
402 *
403 * @param[in] input_ptr                            Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/F32
404 * @param[in] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
405 * @param[in] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
406 * @param[in] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
407 * @param[in] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
408 * @param[in] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
409 * @param[in] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
410 * @param[in] input_stride_w                       Stride of the source tensor in W dimension (in bytes)
411 * @param[in] input_step_w                         input_stride_w * number of elements along W processed per workitem(in bytes)
412 * @param[in] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
413 * @param[in] output_ptr                           The local buffer to hold sumed values. Supported data types: U32/S32
414 * @param[in] output_stride_x                      Stride of the output tensor in X dimension (in bytes)
415 * @param[in] output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
416 * @param[in] output_stride_y                      Stride of the output tensor in Y dimension (in bytes)
417 * @param[in] output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
418 * @param[in] output_stride_z                      Stride of the output tensor in Z dimension (in bytes)
419 * @param[in] output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
420 * @param[in] output_stride_w                      Stride of the output tensor in W dimension (in bytes)
421 * @param[in] output_step_w                        output_stride_w * number of elements along W processed per workitem(in bytes)
422 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor
423 */
424__kernel void arg_min_max_w(
425    TENSOR4D_DECLARATION(input),
426    TENSOR4D_DECLARATION(output))
427{
428    const int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
429
430    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y + (get_global_id(2) % DEPTH) * input_stride_z +
431                                  (get_global_id(2) / DEPTH) * input_stride_w;
432    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE_OUTPUT) + get_global_id(1) * output_stride_y + (get_global_id(
433                                      2) % DEPTH) * output_stride_z + (get_global_id(2) / DEPTH) * output_stride_w;
434
435    VEC_TYPE_IN res = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr), VEC_TYPE_IN);
436
437    VEC_TYPE_OUT indx0 = 0;
438    for(DATA_TYPE_OUTPUT w = 1; w < BATCH; ++w)
439    {
440        VEC_TYPE_IN in = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + w * input_stride_w)), VEC_TYPE_IN);
441
442        VEC_TYPE_OUT cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_TYPE_OUT);
443        indx0                  = select(indx0, (VEC_TYPE_OUT)w, cond_conv);
444        res                    = select(res, in, CONDITION_TO_USE(in, res));
445    }
446
447    // Store result
448    STORE_VECTOR_SELECT(indx, DATA_TYPE_OUTPUT, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
449}
450#endif /* defined(BATCH) && defined(DEPTH) */
451#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE_OUTPUT)