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
|
/*
* Copyright (c) 2020-2022 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 "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/utils/misc/Traits.h"
#include "src/core/NEON/SVEMath.h"
#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
#include <arm_sve.h>
namespace arm_compute
{
namespace cpu
{
void add_qasymm8_sve2(
const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
{
ARM_COMPUTE_UNUSED(policy);
// Create input windows
Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
// Clear X Dimension on execution window as we handle manually
Window win = window;
win.set(Window::DimX, Window::Dimension(0, 1, 1));
const auto window_start_x = static_cast<int>(window.x().start());
const auto window_end_x = static_cast<int>(window.x().end());
const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
const auto all_true_pg = svptrue_b8();
const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale);
const auto voffseto = svdup_n_f32(oq_info.offset);
if (is_broadcast_across_x)
{
const bool is_broadcast_input_2 = input2_win.x().step() == 0;
Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
const svfloat32_t vscale1 = is_broadcast_input_2 ? svdup_n_f32(iq1_info.scale) : svdup_n_f32(iq2_info.scale);
const svfloat32_t vscale2 = is_broadcast_input_2 ? svdup_n_f32(iq2_info.scale) : svdup_n_f32(iq1_info.scale);
const svint32_t voffset1 = is_broadcast_input_2 ? svdup_n_s32(iq1_info.offset) : svdup_n_s32(iq2_info.offset);
const svint32_t voffset2 = is_broadcast_input_2 ? svdup_n_s32(iq2_info.offset) : svdup_n_s32(iq1_info.offset);
// Clear X Dimension on execution window as we handle manually
non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
Iterator broadcast_input(broadcast_tensor, broadcast_win);
Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
Iterator output(dst, win);
execute_window_loop(
win,
[&](const Coordinates &)
{
const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
const uint8_t broadcast_value = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
const svuint8_t broadcast_value_vec = svdup_n_u8(broadcast_value);
int x = window_start_x;
svbool_t pg = svwhilelt_b8(x, window_end_x);
const auto bf_0 = svmul_f32_z(
pg,
svcvt_f32_s32_z(
pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(broadcast_value_vec))),
voffset2)),
vscale2);
const auto bf_1 = svmul_f32_z(
pg,
svcvt_f32_s32_z(
pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(broadcast_value_vec))),
voffset2)),
vscale2);
const auto bf_2 = svmul_f32_z(
pg,
svcvt_f32_s32_z(
pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(broadcast_value_vec))),
voffset2)),
vscale2);
const auto bf_3 = svmul_f32_z(
pg,
svcvt_f32_s32_z(
pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(broadcast_value_vec))),
voffset2)),
vscale2);
do
{
const svuint8_t a = svld1_u8(pg, non_broadcast_input_ptr + x);
const auto af_0 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)),
vscale1);
const auto af_1 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)),
vscale1);
const auto af_2 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)),
vscale1);
const auto af_3 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)),
vscale1);
const auto rf_0 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
const auto rf_1 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
const auto rf_2 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
const auto rf_3 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb);
svst1_u8(pg, output_ptr + x, res);
x += svcntb();
pg = svwhilelt_b8(x, window_end_x);
} while (svptest_any(all_true_pg, pg));
},
broadcast_input, non_broadcast_input, output);
}
else
{
// Clear X Dimension on execution window as we handle manually
input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
Iterator input1(src0, input1_win);
Iterator input2(src1, input2_win);
Iterator output(dst, win);
const auto vscale1 = svdup_n_f32(iq1_info.scale);
const auto vscale2 = svdup_n_f32(iq2_info.scale);
const auto voffset1 = svdup_n_s32(iq1_info.offset);
const auto voffset2 = svdup_n_s32(iq2_info.offset);
execute_window_loop(
win,
[&](const Coordinates &)
{
const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
int x = window_start_x;
svbool_t pg = svwhilelt_b8(x, window_end_x);
do
{
const auto a = svld1_u8(pg, input1_ptr + x);
const auto b = svld1_u8(pg, input2_ptr + x);
const auto af_0 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)),
vscale1);
const auto af_1 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)),
vscale1);
const auto af_2 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)),
vscale1);
const auto af_3 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)),
vscale1);
const auto bf_0 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(b))), voffset2)),
vscale2);
const auto bf_1 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(b))), voffset2)),
vscale2);
const auto bf_2 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(b))), voffset2)),
vscale2);
const auto bf_3 = svmul_f32_z(
pg,
svcvt_f32_s32_z(pg,
svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(b))), voffset2)),
vscale2);
const auto rf_0 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
const auto rf_1 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
const auto rf_2 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
const auto rf_3 =
svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb);
svst1_u8(pg, output_ptr + x, res);
x += svcntb();
pg = svwhilelt_b8(x, window_end_x);
} while (svptest_any(all_true_pg, pg));
},
input1, input2, output);
}
}
} // namespace cpu
} // namespace arm_compute
|
