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
|
//
// Copyright 2011-2012 Ettus Research LLC
// Copyright 2018 Ettus Research, a National Instruments Company
// Copyright 2019 Ettus Research, a National Instruments Brand
//
// SPDX-License-Identifier: GPL-3.0-or-later
//
#include "convert_common.hpp"
#include <uhd/utils/byteswap.hpp>
#include <emmintrin.h>
using namespace uhd::convert;
DECLARE_CONVERTER(fc32, 1, sc16_item32_le, 1, PRIORITY_SIMD)
{
const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]);
item32_t* output = reinterpret_cast<item32_t*>(outputs[0]);
const __m128 scalar = _mm_set_ps1(float(scale_factor));
// this macro converts values faster by using SSE intrinsics to convert 4 values at a time
#define convert_fc32_1_to_item32_1_nswap_guts(_al_) \
for (; i + 3 < nsamps; i += 4) { \
/* load from input */ \
__m128 tmplo = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \
__m128 tmphi = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 2)); \
\
/* convert and scale */ \
__m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \
__m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \
\
/* pack + swap 16-bit pairs */ \
__m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \
tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
\
/* store to output */ \
_mm_storeu_si128(reinterpret_cast<__m128i*>(output + i), tmpi); \
}
size_t i = 0;
// need to dispatch according to alignment for fastest conversion
switch (size_t(input) & 0xf) {
case 0x0:
// the data is 16-byte aligned, so do the fast processing of the bulk of the
// samples
convert_fc32_1_to_item32_1_nswap_guts(_) break;
case 0x8:
// the first sample is 8-byte aligned - process it to align the remainder of
// the samples to 16-bytes
xx_to_item32_sc16<uhd::htowx>(input, output, 1, scale_factor);
i++;
// do faster processing of the bulk of the samples now that we are 16-byte
// aligned
convert_fc32_1_to_item32_1_nswap_guts(_) break;
default:
// we are not 8 or 16-byte aligned, so do fast processing with the unaligned
// load
convert_fc32_1_to_item32_1_nswap_guts(u_)
}
// convert any remaining samples
xx_to_item32_sc16<uhd::htowx>(input + i, output + i, nsamps - i, scale_factor);
}
DECLARE_CONVERTER(fc32, 1, sc16_item32_be, 1, PRIORITY_SIMD)
{
const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]);
item32_t* output = reinterpret_cast<item32_t*>(outputs[0]);
const __m128 scalar = _mm_set_ps1(float(scale_factor));
// this macro converts values faster by using SSE intrinsics to convert 4 values at a time
#define convert_fc32_1_to_item32_1_bswap_guts(_al_) \
for (; i + 3 < nsamps; i += 4) { \
/* load from input */ \
__m128 tmplo = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \
__m128 tmphi = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 2)); \
\
/* convert and scale */ \
__m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \
__m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \
\
/* pack + byteswap -> byteswap 16 bit words */ \
__m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \
tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); \
\
/* store to output */ \
_mm_storeu_si128(reinterpret_cast<__m128i*>(output + i), tmpi); \
}
size_t i = 0;
// need to dispatch according to alignment for fastest conversion
switch (size_t(input) & 0xf) {
case 0x0:
// the data is 16-byte aligned, so do the fast processing of the bulk of the
// samples
convert_fc32_1_to_item32_1_bswap_guts(_) break;
case 0x8:
// the first value is 8-byte aligned - process it and prepare the bulk of the
// data for fast conversion
xx_to_item32_sc16<uhd::htonx>(input, output, 1, scale_factor);
i++;
// do faster processing of the remaining samples now that we are 16-byte
// aligned
convert_fc32_1_to_item32_1_bswap_guts(_) break;
default:
// we are not 8 or 16-byte aligned, so do fast processing with the unaligned
// load
convert_fc32_1_to_item32_1_bswap_guts(u_)
}
// convert any remaining samples
xx_to_item32_sc16<uhd::htonx>(input + i, output + i, nsamps - i, scale_factor);
}
DECLARE_CONVERTER(fc32, 1, sc16_chdr, 1, PRIORITY_SIMD)
{
const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]);
sc16_t* output = reinterpret_cast<sc16_t*>(outputs[0]);
const __m128 scalar = _mm_set_ps1(float(scale_factor));
// this macro converts values faster by using SSE intrinsics to convert 4 values at a time
#define convert_fc32_1_to_item32_1_guts(_al_) \
for (; i + 3 < nsamps; i += 4) { \
/* load from input */ \
__m128 tmplo = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \
__m128 tmphi = \
_mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 2)); \
\
/* convert and scale */ \
__m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \
__m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \
\
/* pack from 32 bit integers to 16 bit */ \
__m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \
\
/* store to output */ \
_mm_storeu_si128(reinterpret_cast<__m128i*>(output + i), tmpi); \
}
size_t i = 0;
// need to dispatch according to alignment for fastest conversion
switch (size_t(input) & 0xf) {
case 0x0:
// the data is 16-byte aligned, so do the fast processing of the bulk of the
// samples
convert_fc32_1_to_item32_1_guts(_) break;
case 0x8:
// the first sample is 8-byte aligned - process it to align the remainder of
// the samples to 16-bytes
xx_to_chdr_sc16(input, output, 1, scale_factor);
i++;
// do faster processing of the bulk of the samples now that we are 16-byte
// aligned
convert_fc32_1_to_item32_1_guts(_) break;
default:
// we are not 8 or 16-byte aligned, so do fast processing with the unaligned
// load
convert_fc32_1_to_item32_1_guts(u_)
}
// convert any remaining samples
xx_to_chdr_sc16(input + i, output + i, nsamps - i, scale_factor);
}
|
