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
|
/*
* Copyright (c) 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.
*/
#ifndef SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H
#define SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H
#if defined(ENABLE_SVE)
#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
{
template <typename ScalarType>
void sve_logits_1d_max(const ITensor *in, ITensor *out, const Window &window);
template <typename ScalarType>
void sve_softmax_logits_1d_float(const ITensor *in, const ITensor *max, void *const tmp,
ITensor *out, const float beta, bool is_log, const Window &window);
#if defined(__ARM_FEATURE_SVE2)
template <typename ScalarType>
void sve_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
ITensor *out, float beta, bool is_log, const Window &window)
{
const int start_x = in->info()->valid_region().anchor.x();
const int input_width = in->info()->valid_region().shape.x();
const float scale_beta = -beta * in->info()->quantization_info().uniform().scale;
const auto scale_beta_vec = svdup_n_f32(scale_beta);
Iterator in_it(in, window);
Iterator max_it(max, window);
Iterator out_it(out, window);
const auto all_true_pg = wrapper::svptrue<ScalarType>();
using SVEType = typename wrapper::traits::sve_vector<ScalarType>::type;
const int inc_1 = static_cast<int>(svcntw());
const int inc_2 = static_cast<int>(2 * svcntw());
const int inc_3 = static_cast<int>(3 * svcntw());
execute_window_loop(window, [&](const Coordinates &)
{
/* Get pointers */
const auto in_ptr = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
const auto tmp_ptr = reinterpret_cast<float *>(tmp);
float sum{};
/* Compute exponentials and sum */
{
/* Get max value */
const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
const auto vec_max = wrapper::svdup_n(max_val);
/* Init sum to zero */
auto vec_sum_0 = svdup_n_f32(0.f);
auto vec_sum_1 = svdup_n_f32(0.f);
auto vec_sum_2 = svdup_n_f32(0.f);
auto vec_sum_3 = svdup_n_f32(0.f);
/* Loop over row and compute exponentials and sum */
int x = 0;
svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
svbool_t pg_0 = svunpklo(svunpklo(pg));
svbool_t pg_1 = svunpkhi(svunpklo(pg));
svbool_t pg_2 = svunpklo(svunpkhi(pg));
svbool_t pg_3 = svunpkhi(svunpkhi(pg));
do
{
auto vec_elements = svld1(pg, in_ptr + x);
vec_elements = svsub_z(pg, vec_max, vec_elements);
auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements)));
auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements)));
auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements)));
auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements)));
if(is_log)
{
vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
}
else
{
vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
vec_sum_0 = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
vec_sum_1 = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
vec_sum_2 = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
vec_sum_3 = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
}
svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
x += wrapper::svcnt<ScalarType>();
pg = wrapper::svwhilelt<ScalarType>(x, input_width);
pg_0 = svunpklo(svunpklo(pg));
pg_1 = svunpkhi(svunpklo(pg));
pg_2 = svunpklo(svunpkhi(pg));
pg_3 = svunpkhi(svunpkhi(pg));
}
while(svptest_any(all_true_pg, pg));
/* Reduce sum */
const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1), svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
sum = svaddv_f32(all_true_pg, vec_sum);
/* Run remaining elements */
x = 0;
if(is_log)
{
sum = std::log(sum);
}
else
{
sum = 256.f / sum;
}
}
/* Normalize exponentials */
{
constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
/* Loop over row and compute softmax */
int x = 0;
svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
svbool_t pg_0 = svunpklo(svunpklo(pg));
svbool_t pg_1 = svunpkhi(svunpklo(pg));
svbool_t pg_2 = svunpklo(svunpkhi(pg));
svbool_t pg_3 = svunpkhi(svunpkhi(pg));
do
{
auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
svfloat32_t res_0{};
svfloat32_t res_1{};
svfloat32_t res_2{};
svfloat32_t res_3{};
if(is_log)
{
res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
}
else
{
res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
if(is_qasymm8_signed)
{
const auto offset_vec = svdup_n_f32(128.f);
res_0 = svsub_z(pg_0, vec_in_0, offset_vec);
res_1 = svsub_z(pg_1, vec_in_1, offset_vec);
res_2 = svsub_z(pg_2, vec_in_2, offset_vec);
res_3 = svsub_z(pg_3, vec_in_3, offset_vec);
}
}
// Store value
const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
svst1(pg, out_ptr + x, out);
x += wrapper::svcnt<ScalarType>();
pg = wrapper::svwhilelt<ScalarType>(x, input_width);
pg_0 = svunpklo(svunpklo(pg));
pg_1 = svunpkhi(svunpklo(pg));
pg_2 = svunpklo(svunpkhi(pg));
pg_3 = svunpkhi(svunpkhi(pg));
}
while(svptest_any(all_true_pg, pg));
}
},
in_it, max_it, out_it);
}
#endif /* defined(__ARM_FEATURE_SVE2) */
} // namespace cpu
} // namespace arm_compute
#endif /* defined(ENABLE_SVE) */
#endif /* SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H */
|
