Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / e5d76e1574103de405df625e48e5294ea106060c / . / src / core / CL / cl_kernels / normalize_planar_yuv_layer_quantized.cl
blob: d660fffb58cd307cad12a82e1200a768b41b6f9b [file] [log] [blame]
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]1/*
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]2 * Copyright (c) 2018-2021 Arm Limited.
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]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#if defined(NUM_CHANNELS)
33
34/** Apply normalize_planar_yuv layer on tensors with NCHW data layout.
35 *
36 * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
37 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
38 * @note The depth of the input tensor should be given as a preprocessor argument using -DNUM_CHANNELS e.g. -DNUM_CHANNELS=8
39 * @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8
40 * @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8
41 *
Michele Di Giorgiocbbed282019-12-20 13:26:08 +0000[diff] [blame]42 * @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]43 * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes)
44 * @param[in] src_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
45 * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes)
46 * @param[in] src_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
47 * @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes)
48 * @param[in] src_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
49 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor
50 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
51 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
52 * @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
53 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
54 * @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
55 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
56 * @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
57 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
58 * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr
59 * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
60 * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
61 * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
62 * @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr
63 * @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes)
64 * @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes)
65 * @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor
66 */
67__kernel void normalize_planar_yuv_layer_q8_nchw(TENSOR3D_DECLARATION(src),
68 TENSOR3D_DECLARATION(dst),
69 VECTOR_DECLARATION(mean),
70 VECTOR_DECLARATION(std))
71{
72 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
73 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
74 Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
75 Vector std = CONVERT_TO_VECTOR_STRUCT(std);
76
77 const uint current_slice = get_global_id(2) % NUM_CHANNELS;
78
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]79 VEC_DATA_TYPE(float, VEC_SIZE)
80 curr_mean_flt = (VEC_DATA_TYPE(float, VEC_SIZE))(*((__global DATA_TYPE *)(mean.ptr + current_slice * sizeof(DATA_TYPE))));
81 curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]82
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]83 VEC_DATA_TYPE(float, VEC_SIZE)
84 curr_std_flt = (VEC_DATA_TYPE(float, VEC_SIZE))(*((__global DATA_TYPE *)(std.ptr + current_slice * sizeof(DATA_TYPE))));
85 curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]86
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]87 VEC_DATA_TYPE(float, VEC_SIZE)
88 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr), VEC_DATA_TYPE(float, VEC_SIZE));
89 data_flt = round(data_flt - OFFSET_FLT) * SCALE_FLT;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]90
91 // Perform normalization
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]92 VEC_DATA_TYPE(float, VEC_SIZE)
93 res_flt = (data_flt - curr_mean_flt) / curr_std_flt;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]94
95 const TYPE res_u8 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE);
96 VSTORE(VEC_SIZE)
97 (res_u8, 0, (__global DATA_TYPE *)dst.ptr);
98}
99
100#endif // defined(NUM_CHANNELS)
101
102/** Apply normalize_planar_yuv layer on tensors with NHWC data layout.
103 *
104 * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
105 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
106 * @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8
107 * @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]108 * @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
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]109 *
Michele Di Giorgiocbbed282019-12-20 13:26:08 +0000[diff] [blame]110 * @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]111 * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes)
112 * @param[in] src_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
113 * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes)
114 * @param[in] src_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
115 * @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes)
116 * @param[in] src_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
117 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor
118 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
119 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
120 * @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
121 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
122 * @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
123 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
124 * @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
125 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
126 * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr
127 * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
128 * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
129 * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
130 * @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr
131 * @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes)
132 * @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes)
133 * @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor
134 */
135__kernel void normalize_planar_yuv_layer_q8_nhwc(TENSOR3D_DECLARATION(src),
136 TENSOR3D_DECLARATION(dst),
137 VECTOR_DECLARATION(mean),
138 VECTOR_DECLARATION(std))
139{
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]140 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);
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]141
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]142 __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;
143 __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;
144 __global uchar *mean_addr = mean_ptr + mean_offset_first_element_in_bytes + x_offs;
145 __global uchar *std_addr = std_ptr + std_offset_first_element_in_bytes + x_offs;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]146
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]147 VEC_DATA_TYPE(float, VEC_SIZE)
148 curr_mean_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)mean_addr), VEC_DATA_TYPE(float, VEC_SIZE));
149 curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]150
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]151 VEC_DATA_TYPE(float, VEC_SIZE)
152 curr_std_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)std_addr), VEC_DATA_TYPE(float, VEC_SIZE));
153 curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]154
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]155 VEC_DATA_TYPE(float, VEC_SIZE)
156 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr), VEC_DATA_TYPE(float, VEC_SIZE));
157 data_flt = round(data_flt - OFFSET_FLT) * (SCALE_FLT);
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]158
159 // Perform normalization
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]160 VEC_DATA_TYPE(float, VEC_SIZE)
161 res_flt = (data_flt - curr_mean_flt) / curr_std_flt;
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]162
Sheri Zhang4f1650f2021-04-15 12:58:20 +0100[diff] [blame]163 const TYPE res0 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE);
164 STORE_VECTOR_SELECT(res, DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
Michele Di Giorgiod63dfa22018-09-12 10:18:54 +0100[diff] [blame]165}
166#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE)