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
|
/*
* 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.
*/
#ifndef ARM_COMPUTE_SVEASYMM_H
#define ARM_COMPUTE_SVEASYMM_H
#if defined(ARM_COMPUTE_ENABLE_SVE2)
#include "src/core/NEON/SVEMath.h"
#include <arm_sve.h>
namespace arm_compute
{
/** Perform a multiply-accumulate on all components of a QASYMM8 vector
*
* vd*vs + vo
*
* @param[in] pg Predicate value.
* @param[in] vd Input vector value in QASYMM8 format
* @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes.
* @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes.
*
* @return A vector in QASYMM8 format, saturated to fit
*/
svuint8_t svmla_qasymm8_z(svbool_t pg, svuint8_t vd, svfloat32_t vs, svfloat32_t vo);
/** Perform a multiply-accumulate on all components of a QASYMM8_SIGNED vector
*
* vd*vs + vo
*
* @param[in] pg Predicate value.
* @param[in] vd Input vector value in QASYMM8_SIGNED format
* @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes.
* @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes.
*
* @return A vector in QASYMM8_SIGNED format, saturated to fit
*/
svint8_t svmla_qasymm8_signed_z(svbool_t pg, svint8_t vd, svfloat32_t vs, svfloat32_t vo);
/** Dequantize following an asymmetric quantization scheme a sve vector.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be dequantized.
* @param[in] scale Quantization scaling factor.
* @param[in] offset Zero quantization offset.
*
* @return Dequantized values in an sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svuint8_t &qv, float scale, int32_t offset)
{
const auto voffset = svdup_n_s32(offset);
const auto vscale = svdup_n_f32(scale);
const svfloat32x4_t vdequantized_input = svcreate4_f32(
svmul_f32_z(pg,
svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(qv))), voffset)),
vscale),
svmul_f32_z(pg,
svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(qv))), voffset)),
vscale),
svmul_f32_z(pg,
svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(qv))), voffset)),
vscale),
svmul_f32_z(pg,
svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(qv))), voffset)),
vscale));
return vdequantized_input;
}
/** Dequantize an sve vector
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be dequantized.
* @param[in] qi Quantization information to be used in the computation.
*
* @return Dequantized values in an sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svuint8_t &qv, const UniformQuantizationInfo &qi)
{
return svdequantize_z(pg, qv, qi.scale, qi.offset);
}
/** Dequantize an sve vector stored as signed asymmetric.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be dequantized.
* @param[in] scale Quantization scaling factor.
* @param[in] offset Zero quantization offset.
*
* @return Dequantized values in a sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, float scale, int32_t offset)
{
const auto voffset = svdup_n_s32(offset);
const auto vscale = svdup_n_f32(scale);
const svfloat32x4_t vdequantized_input = svcreate4_f32(
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(qv)), voffset)), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(qv)), voffset)), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(qv)), voffset)), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(qv)), voffset)), vscale));
return vdequantized_input;
}
/** Dequantize an sve vector.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be dequantized.
* @param[in] qi Quantization information to be used in the computation.
*
* @return Dequantized values in an sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, const UniformQuantizationInfo &qi)
{
return svdequantize_z(pg, qv, qi.scale, qi.offset);
}
/** Dequantize following symmetric quantization scheme on an sve vector.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be dequantized.
* @param[in] vscale Vector containing quantization scaling factors.
*
* @return Dequantized values in a sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, const svfloat32x4_t vscale)
{
const svfloat32x4_t vdequantized_input =
svcreate4_f32(svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlb_s16(qv))), svget4_f32(vscale, 0)),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlb_s16(qv))), svget4_f32(vscale, 1)),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlt_s16(qv))), svget4_f32(vscale, 2)),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlt_s16(qv))), svget4_f32(vscale, 3)));
return vdequantized_input;
}
/** Dequantize following a symmetric quantization scheme an sve vector.
*
* @param[in] qv Input values to be dequantized.
* @param[in] scale Quantization scaling factor.
*
* @return Dequantized values in a sve vector
*/
inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, float scale)
{
const auto vscale = svdup_n_f32(scale);
const svfloat32x4_t vdequantized_input =
svcreate4_f32(svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlb_s16(qv))), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlb_s16(qv))), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlt_s16(qv))), vscale),
svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlt_s16(qv))), vscale));
return vdequantized_input;
}
/** Quantize an sve vector holding floating point values.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be quantized.
* @param[in] qi Quantization information to be used in the computation.
*
* @return An sve vector holding the quantized values
*/
inline svuint8_t svquantize_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi)
{
const float scale = qi.scale;
const int offset = qi.offset;
const auto voffset = svdup_n_f32(offset);
const auto vinvscale = svdup_n_f32(1.f / scale);
const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale));
const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale));
const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale));
const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale));
const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
return svqxtnt_u16(svqxtnb_u16(pa), pb);
}
/** Signed quantize an sve vector holding floating point values.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be quantized.
* @param[in] qi Quantization information to be used in the computation.
*
* @return An sve vector holding the quantized values
*/
inline svint8_t svquantize_signed_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi)
{
const float scale = qi.scale;
const int offset = qi.offset;
const auto voffset = svdup_n_f32(offset);
const auto vinvscale = svdup_n_f32(1.f / scale);
const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale));
const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale));
const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale));
const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale));
const auto pa = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1);
const auto pb = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3);
return svqxtnt_s16(svqxtnb_s16(pa), pb);
}
/** Quantize to QASYMM16 an sve vector holding 16 floating point values.
*
* @param[in] pg Predicate value.
* @param[in] qv Input values to be quantized.
* @param[in] qi Quantization information to be used in the computation.
*
* @return An sve vector holding the quantized values
*/
inline svuint16x2_t svquantize_qasymm16_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi)
{
const float scale = qi.scale;
const int offset = qi.offset;
const auto voffset = svdup_n_f32(offset);
const auto vinvscale = svdup_n_f32(1.f / scale);
const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale));
const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale));
const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale));
const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale));
const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
return svcreate2_u16(pa, pb);
}
} // namespace arm_compute
#include "src/core/NEON/SVEAsymm.inl"
#endif /* defined(ARM_COMPUTE_ENABLE_SVE2) */
#endif // ARM_COMPUTE_NEASYMM_H
|
