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
|
#include "benchmark_helpers.hpp"
#include "../simd_memory.hpp"
#ifdef __SSE2__
#include <xmmintrin.h>
#include <emmintrin.h>
#endif
const int unroll = 8;
using namespace nova;
void __noinline__ bench_1(float * out, float * in, unsigned int numSamples)
{
copyvec(out, in, numSamples);
}
#ifdef __SSE2__
void __noinline__ bench_2(float * out, float * in, unsigned int numSamples)
{
int loops = numSamples / unroll;
do {
__m128 data = _mm_load_ps(in);
_mm_store_ps(out, data);
__m128 data2 = _mm_load_ps(in+4);
_mm_store_ps(out+4, data2);
in += 8;
out += 8;
} while (--loops);
}
void __noinline__ bench_3(float * out, float * in, unsigned int numSamples)
{
int loops = numSamples / unroll;
do {
__m128 data = _mm_load_ps(in);
__m128 data2 = _mm_load_ps(in+4);
_mm_store_ps(out, data);
_mm_store_ps(out+4, data2);
in += 8;
out += 8;
} while (--loops);
}
void __noinline__ bench_4(float * out, float * in, unsigned int numSamples)
{
int loops = numSamples / unroll;
do {
__m128d data = _mm_load_pd((double*)in);
_mm_store_pd((double*)out, data);
__m128d data2 = _mm_load_pd((double*)in+4);
_mm_store_pd((double*)out+4, data2);
in += 8;
out += 8;
} while (--loops);
}
void __noinline__ bench_5(float * out, float * in, unsigned int numSamples)
{
int loops = numSamples / unroll;
do {
__m128i data = _mm_load_si128((__m128i*)in);
_mm_store_si128((__m128i*)out, data);
__m128i data2 = _mm_load_si128((__m128i*)in+4);
_mm_store_si128((__m128i*)out+4, data2);
in += 8;
out += 8;
} while (--loops);
}
/*
* sse implementation of partially aligned copy operations
* this seems to be more efficient than memcpy on x86, but not on x86_64
*
*/
#if defined(__GNUC__) && defined(NDEBUG)
#define always_inline inline __attribute__((always_inline))
#else
#define always_inline inline
#endif
const int samples_per_loop = 8;
template <int n>
always_inline void copyvec_naa_simd_mp(float *dest, const float *src)
{
__m128 data = _mm_loadu_ps(src);
_mm_store_ps(dest, data);
copyvec_naa_simd_mp<n-4>(dest+4, src+4);
}
template <>
always_inline void copyvec_naa_simd_mp<0>(float *dest, const float *src)
{}
template <int n>
always_inline void copyvec_ana_simd_mp(float *dest, const float *src)
{
__m128 data = _mm_load_ps(src);
_mm_storeu_ps(dest, data);
copyvec_ana_simd_mp<n-4>(dest+4, src+4);
}
template <>
always_inline void copyvec_ana_simd_mp<0>(float *dest, const float *src)
{}
template <int n>
always_inline void copyvec_nana_simd_mp(float *dest, const float *src)
{
__m128 data = _mm_loadu_ps(src);
_mm_storeu_ps(dest, data);
copyvec_nana_simd_mp<n-4>(dest+4, src+4);
}
template <>
always_inline void copyvec_nana_simd_mp<0>(float *dest, const float *src)
{}
inline void copyvec_naa_simd(float * dest, const float * src, uint n)
{
n = n / samples_per_loop;
do
{
copyvec_naa_simd_mp<samples_per_loop>(dest, src);
dest += samples_per_loop;
src += samples_per_loop;
}
while(--n);
}
inline void copyvec_ana_simd(float * dest, const float * src, uint n)
{
n = n / samples_per_loop;
do
{
copyvec_ana_simd_mp<samples_per_loop>(dest, src);
dest += samples_per_loop;
src += samples_per_loop;
}
while(--n);
}
inline void copyvec_nana_simd(float * dest, const float * src, uint n)
{
n = n / samples_per_loop;
do
{
copyvec_nana_simd_mp<samples_per_loop>(dest, src);
dest += samples_per_loop;
src += samples_per_loop;
}
while(--n);
}
void __noinline__ bench_6(float * out, float * in, unsigned int numSamples)
{
copyvec_nana_simd(out, in, numSamples);
}
void __noinline__ bench_7(float * out, float * in, unsigned int numSamples)
{
copyvec_ana_simd(out, in, numSamples);
}
void __noinline__ bench_8(float * out, float * in, unsigned int numSamples)
{
copyvec_naa_simd(out, in, numSamples);
}
#endif
int main(void)
{
using namespace std;
const int iterations = 50000000;
nova::aligned_array<float, 66> __attribute__((aligned(64))) in;
nova::aligned_array<float, 66> __attribute__((aligned(64))) out;
fill_container(in);
fill_container(out);
cout << "memcpy" << endl;
run_bench(boost::bind(bench_1, out.begin(), in.begin(), 64), iterations);
run_bench(boost::bind(bench_1, out.begin()+2, in.begin(), 64), iterations);
run_bench(boost::bind(bench_1, out.begin(), in.begin()+2, 64), iterations);
run_bench(boost::bind(bench_1, out.begin()+2, in.begin()+2, 64), iterations);
cout << endl;
#ifdef __SSE2__
cout << "float/double/int (sse/sse2 instructions)" << endl;
run_bench(boost::bind(bench_2, out.begin(), in.begin(), 64), iterations);
run_bench(boost::bind(bench_3, out.begin(), in.begin(), 64), iterations);
run_bench(boost::bind(bench_4, out.begin(), in.begin(), 64), iterations);
run_bench(boost::bind(bench_5, out.begin(), in.begin(), 64), iterations);
cout << endl;
cout << "nonaligned" << endl;
run_bench(boost::bind(bench_6, out.begin(), in.begin(), 64), iterations);
run_bench(boost::bind(bench_6, out.begin()+2, in.begin()+2, 64), iterations);
cout << endl;
cout << "partially aligned" << endl;
run_bench(boost::bind(bench_7, out.begin()+2, in.begin(), 64), iterations);
run_bench(boost::bind(bench_8, out.begin(), in.begin()+2, 64), iterations);
cout << endl;
#endif
}
|
