1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
|
/*
* Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers_asymm.h"
#include "warp_helpers_quantized.h"
/** Transforms four 2D coordinates. This is used to map the output coordinates to the input coordinates.
*
* @param[in] coord 2D coordinates to transform.
* @param[in] scale input/output scale ratio
*
* @return a float8 containing 4 2D transformed values in the input image.
*/
inline const float8 transform_bilinear_quantized(const float2 coord, const float2 scale)
{
const float4 in_x_coords = (float4)(coord.s0, 1 + coord.s0, 2 + coord.s0, 3 + coord.s0);
#ifdef SAMPLING_POLICY_TOP_LEFT
const float4 new_x = in_x_coords * (float4)(scale.s0);
const float4 new_y = (float4)(coord.s1 * scale.s1);
return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3);
#elif SAMPLING_POLICY_CENTER
const float4 new_x = (in_x_coords + ((float4)(0.5f))) * (float4)(scale.s0) - (float4)(0.5f);
const float4 new_y = (float4)((coord.s1 + 0.5f) * scale.s1 - 0.5f);
return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3);
#else /* SAMPLING_POLICY */
#error("Unsupported sampling policy");
#endif /* SAMPLING_POLICY */
}
/** Performs an affine transformation on an image interpolating with the BILINEAR method.
*
* @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT
* @note Scale value for QASYMM8 data type to used is passed as -DSCALE=<VALUE> e.g. -DSCALE=0.5
* @note Offset value for QASYMM8 data type to used is passed as -DOFFSET=<VALUE> e.g. -DOFFSET=1
*
* @param[in] in_ptr Pointer to the source image. Supported data types: QASYMM8.
* @param[in] in_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] in_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] out_ptr Pointer to the destination image. Supported data types: U8, S16. (Must be the same as the input)
* @param[in] out_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image
* @param[in] input_width Input image width
* @param[in] input_height Input image height
* @param[in] scale_x The scale factor along x dimension
* @param[in] scale_y The scale factor along y dimension
*/
__kernel void scale_bilinear_quantized_nchw(
IMAGE_DECLARATION(in),
IMAGE_DECLARATION(out),
const float input_width,
const float input_height,
const float scale_x,
const float scale_y)
{
Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in);
Image out = CONVERT_TO_IMAGE_STRUCT(out);
const float2 r = (float2)(scale_x, scale_y);
const float8 tc = transform_bilinear_quantized(get_current_coords_quantized(), r);
vstore4(bilinear_interpolate_with_border_quantized(&in, tc, input_width, input_height, BORDER_SIZE, SCALE, OFFSET), 0, (__global DATA_TYPE *)out.ptr);
}
|
