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#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE)
27
28#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
29#define OFFSET_FLT ((float)OFFSET)
30#define SCALE_FLT ((float)SCALE)
31
32/** Apply normalize_planar_yuv layer on tensors with NHWC data layout.
33 *
34 * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
35 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
36 * @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8
37 * @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8
38 * @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
39 *
40 * @param[in]  src_ptr                            Pointer to the first source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
41 * @param[in]  src_stride_x                       Stride of the first source tensor in X dimension (in bytes)
42 * @param[in]  src_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
43 * @param[in]  src_stride_y                       Stride of the first source tensor in Y dimension (in bytes)
44 * @param[in]  src_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
45 * @param[in]  src_stride_z                       Stride of the first source tensor in Z dimension (in bytes)
46 * @param[in]  src_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
47 * @param[in]  src_offset_first_element_in_bytes  The offset of the first element in the first source tensor
48 * @param[out] dst_ptr                            Pointer to the destination tensor. Supported data types: same as @p src_ptr
49 * @param[in]  dst_stride_x                       Stride of the destination tensor in X dimension (in bytes)
50 * @param[in]  dst_step_x                         output_stride_x * number of elements along X processed per workitem(in bytes)
51 * @param[in]  dst_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
52 * @param[in]  dst_step_y                         output_stride_y * number of elements along Y processed per workitem(in bytes)
53 * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
54 * @param[in]  dst_step_z                         output_stride_z * number of elements along Z processed per workitem(in bytes)
55 * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination tensor
56 * @param[in]  mean_ptr                           Pointer to the mean source tensor. Supported data types: same as @p src_ptr
57 * @param[in]  mean_stride_x                      Stride of the mean source tensor in X dimension (in bytes)
58 * @param[in]  mean_step_x                        mean_stride_x * number of elements along X processed per workitem(in bytes)
59 * @param[in]  mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
60 * @param[in]  std_ptr                            Pointer to the std tensor. Supported data types: same as @p src_ptr
61 * @param[in]  std_stride_x                       Stride of the std tensor in X dimension (in bytes)
62 * @param[in]  std_step_x                         std_stride_x * number of elements along X processed per workitem(in bytes)
63 * @param[in]  std_offset_first_element_in_bytes  The offset of the first element in the var source tensor
64 */
65__kernel void normalize_planar_yuv_layer_q8_nhwc(TENSOR3D_DECLARATION(src),
66                                                 TENSOR3D_DECLARATION(dst),
67                                                 VECTOR_DECLARATION(mean),
68                                                 VECTOR_DECLARATION(std))
69{
70    uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0);
71
72    __global uchar *src_addr  = src_ptr + src_offset_first_element_in_bytes + x_offs + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z;
73    __global uchar *dst_addr  = dst_ptr + dst_offset_first_element_in_bytes + x_offs + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z;
74    __global uchar *mean_addr = mean_ptr + mean_offset_first_element_in_bytes + x_offs;
75    __global uchar *std_addr  = std_ptr + std_offset_first_element_in_bytes + x_offs;
76
77    VEC_DATA_TYPE(float, VEC_SIZE)
78    curr_mean_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)mean_addr), VEC_DATA_TYPE(float, VEC_SIZE));
79    curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT;
80
81    VEC_DATA_TYPE(float, VEC_SIZE)
82    curr_std_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)std_addr), VEC_DATA_TYPE(float, VEC_SIZE));
83    curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT;
84
85    VEC_DATA_TYPE(float, VEC_SIZE)
86    data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr), VEC_DATA_TYPE(float, VEC_SIZE));
87    data_flt = round(data_flt - OFFSET_FLT) * (SCALE_FLT);
88
89    // Perform normalization
90    VEC_DATA_TYPE(float, VEC_SIZE)
91    res_flt = (data_flt - curr_mean_flt) / curr_std_flt;
92
93    const TYPE res0 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE);
94    STORE_VECTOR_SELECT(res, DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
95}
96#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE)