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
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
|
/*
* Copyright (c) 2022-2023 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.
*/
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensorPack.h"
#include "arm_compute/core/Window.h"
#include "src/core/helpers/ScaleHelpers.h"
#include "src/core/NEON/NEMath.h"
#include "src/core/NEON/wrapper/wrapper.h"
#include "src/core/utils/ScaleUtils.h"
#include "src/cpu/kernels/scale/neon/list.h"
#include "support/Rounding.h"
#include <arm_neon.h>
#include <cmath>
#include <cstddef>
namespace arm_compute
{
namespace
{
void fp16_neon_scale_nearest(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
float sampling_offset,
bool align_corners,
const Window &window)
{
const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
const size_t in_stride_wc = in_stride_w * in_stride_c;
const size_t in_dim_h = src->info()->dimension(2);
// Compute the ratio between source height and destination height
const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
const auto window_start_x = static_cast<int32_t>(window.x().start());
const auto window_end_x = static_cast<int32_t>(window.x().end());
const int window_step_x = 8;
Window win(window);
win.set(Window::DimX, Window::Dimension(0, 1, 1));
Iterator out(dst, win);
const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
execute_window_loop(
win,
[&](const Coordinates &id)
{
const int32_t offset =
*reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
const auto in_hi = static_cast<int>(
align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr)
: std::floor((id.z() + sampling_offset) * hr));
const int offset_row = in_hi * in_stride_wc;
int32_t x = window_start_x;
const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
for (; x <= window_end_x - window_step_x; x += window_step_x)
{
wrapper::vstore(reinterpret_cast<float16_t *>(out.ptr()) + x,
wrapper::vloadq(in_ptr + offset + offset_row + x));
}
for (; x < window_end_x; ++x)
{
*(reinterpret_cast<float16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
}
},
out);
}
void fp16_neon_scale_bilinear(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
const ITensor *dx,
const ITensor *dy,
BorderMode border_mode,
PixelValue constant_border_value,
float sampling_offset,
bool align_corners,
const Window &window)
{
// Compute the ratio between source height and destination height
const auto hr =
scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
Iterator out(dst, window);
const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
const int in_dim_w = src->info()->dimension(1);
const int in_dim_h = src->info()->dimension(2);
const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
// Don't increment in Y and Z direction for the input tensor
// A pointer to the start of this plane is needed as base for the precomputed offsets
Window win_in(window);
win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
Iterator in(src, win_in);
if (border_mode == BorderMode::CONSTANT)
{
using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type;
const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>());
execute_window_loop(
window,
[&](const Coordinates &id)
{
const auto offset =
*reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
const float16_t *in_ptr =
reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
const auto a00 =
(0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h)
? *(in_ptr + in_stride_c)
: const_border_value;
const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1)
? *(in_ptr + in_stride_wc)
: const_border_value;
const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1)
? *(in_ptr + in_stride_c + in_stride_wc)
: const_border_value;
*reinterpret_cast<float16_t *>(out.ptr()) =
static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
},
in, out);
}
else if (border_mode == BorderMode::REPLICATE)
{
execute_window_loop(
window,
[&](const Coordinates &id)
{
const auto offset =
*reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c +
clamped_h * in_stride_wc);
const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c +
clamped_h * in_stride_wc);
const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c +
clamped_h1 * in_stride_wc);
const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c +
clamped_h1 * in_stride_wc);
*reinterpret_cast<float16_t *>(out.ptr()) =
static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
},
in, out);
}
else
{
ARM_COMPUTE_ERROR("Not implemented");
}
}
} // namespace
namespace cpu
{
#ifdef ENABLE_NCHW_KERNELS
void fp16_bilinear_neon_scale_nchw(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
const ITensor *dx,
const ITensor *dy,
InterpolationPolicy policy,
BorderMode border_mode,
PixelValue constant_border_value,
float sampling_offset,
bool align_corners,
const Window &window)
{
ARM_COMPUTE_UNUSED(policy);
arm_compute::cpu::scale_bilinear_nchw<float16_t>(src, dst, dx, dy, offsets, border_mode, constant_border_value,
sampling_offset, align_corners, window);
}
void fp16_nearest_neon_scale_nchw(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
const ITensor *dx,
const ITensor *dy,
InterpolationPolicy policy,
BorderMode border_mode,
PixelValue constant_border_value,
float sampling_offset,
bool align_corners,
const Window &window)
{
ARM_COMPUTE_UNUSED(policy);
ARM_COMPUTE_UNUSED(border_mode);
arm_compute::cpu::scale_nearest_nchw<float16_t>(src, dst, dx, dy, offsets, constant_border_value, sampling_offset,
align_corners, window);
}
#endif // ENABLE_NCHW_KERNELS
void fp16_neon_scale(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
const ITensor *dx,
const ITensor *dy,
InterpolationPolicy policy,
BorderMode border_mode,
PixelValue constant_border_value,
float sampling_offset,
bool align_corners,
const Window &window)
{
if (policy == InterpolationPolicy::BILINEAR)
{
fp16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset,
align_corners, window);
}
else if (policy == InterpolationPolicy::NEAREST_NEIGHBOR)
{
fp16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
}
}
void fp16_common_neon_scale(const ITensor *src,
ITensor *dst,
const ITensor *offsets,
const ITensor *dx,
const ITensor *dy,
InterpolationPolicy policy,
BorderMode border_mode,
PixelValue constant_border_value,
float sampling_offset,
bool align_corners,
const Window &window)
{
arm_compute::cpu::common_neon_scale<float16_t>(src, dst, offsets, dx, dy, policy, border_mode,
constant_border_value, sampling_offset, align_corners, window);
}
} // namespace cpu
} // namespace arm_compute
#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
|
