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
|
/* { dg-do compile } */
/* { dg-require-effective-target ilp32 } */
/* { dg-options "-O3 -msse2 -fdump-rtl-csa" } */
/* { dg-skip-if "no stdint" { vxworks_kernel } } */
#include <emmintrin.h>
#include <stdint.h>
typedef __SIZE_TYPE__ size_t;
typedef float vFloat __attribute__ ((__vector_size__ (16)));
typedef double vDouble __attribute__ ((__vector_size__ (16)));
typedef struct buf
{
void *data;
unsigned long h;
unsigned long w;
size_t bytes;
} buf;
typedef struct job
{
struct Job *next;
void * info;
long (*func)(struct Job *job);
long error;
} job;
typedef struct fj
{
job hd;
buf src;
buf dest;
float g;
unsigned int flags;
} fj;
static const double r[256], t[256];
long bar (const buf *src, const buf *dest, float g, unsigned int flags)
{
float *d0 = (float*) src->data;
float *d1 = (float*) dest->data;
uintptr_t w = dest->w;
uintptr_t idx;
vFloat p0;
static const vFloat m0;
static const vDouble p[3], m, b;
float *sr = d0;
float *dr = d1;
for( idx = 0; idx + 8 <= w; idx += 8 )
{
vFloat f0 = _mm_loadu_ps (sr);
vFloat f1 = _mm_loadu_ps (sr + 4);
sr += 8;
vFloat fa0 = _mm_andnot_ps (m0, f0);
vFloat fa1 = _mm_andnot_ps (m0, f1);
vDouble v0 = _mm_cvtps_pd (fa0);
vDouble v1 = _mm_cvtps_pd (_mm_movehl_ps (fa0, fa0));
vDouble v2 = _mm_cvtps_pd (fa1);
vDouble v3 = _mm_cvtps_pd (_mm_movehl_ps (fa1, fa1));
vDouble vi0, vi1, vi2, vi3;
__m128i b0, b1, b2, b3;
b0 = _mm_packs_epi32 (_mm_packs_epi32 (b0, b1), _mm_packs_epi32 (b2, b3));
b1 = _mm_srli_epi64 (b0, 32);
unsigned int i0 = _mm_cvtsi128_si32 (b0);
unsigned int i2 = _mm_cvtsi128_si32 (b1);
v0 -= _mm_loadh_pd (_mm_load_sd (r + (i0 & 0xff)), r + (i0 >> 16));
v1 -= _mm_loadh_pd (_mm_load_sd (r + (i2 & 0xff)), r + (i2 >> 16));
b0 = _mm_unpackhi_epi64 (b0, b0);
b1 = _mm_unpackhi_epi64 (b1, b1);
unsigned int i4 = _mm_cvtsi128_si32 (b0);
unsigned int i6 = _mm_cvtsi128_si32 (b1);
v2 -= _mm_loadh_pd (_mm_load_sd (r + (i4 & 0xff)), r + (i4 >> 16));
v3 -= _mm_loadh_pd (_mm_load_sd (r + (i6 & 0xff)), r + (i6 >> 16));
v0 = p[0] + (p[1] + p[2] * v0) * v0;
v1 = p[0] + (p[1] + p[2] * v1) * v1;
v2 = p[0] + (p[1] + p[2] * v2) * v2;
v3 = p[0] + (p[1] + p[2] * v3) * v3;
vi0 = (vDouble) _mm_slli_epi64 ((__m128i)((vi0 + b) + m), 52);
vi1 = (vDouble) _mm_slli_epi64 ((__m128i)((vi1 + b) + m), 52);
vi2 = (vDouble) _mm_slli_epi64 ((__m128i)((vi2 + b) + m), 52);
vi3 = (vDouble) _mm_slli_epi64 ((__m128i)((vi3 + b) + m), 52);
vi0 *= _mm_loadh_pd (_mm_load_sd (t + (i0 & 0xff)), t + (i0 >> 16));
vi1 *= _mm_loadh_pd (_mm_load_sd (t + (i2 & 0xff)), t + (i2 >> 16));
vi2 *= _mm_loadh_pd (_mm_load_sd (t + (i4 & 0xff)), t + (i4 >> 16));
vi3 *= _mm_loadh_pd (_mm_load_sd (t + (i6 & 0xff)), t + (i6 >> 16));
v0 *= vi0;
v1 *= vi1;
v2 *= vi2;
v3 *= vi3;
vFloat r0 = _mm_movelh_ps (_mm_cvtpd_ps( v0 ), _mm_cvtpd_ps (v1));
vFloat r1 = _mm_movelh_ps (_mm_cvtpd_ps( v2 ), _mm_cvtpd_ps (v3));
vFloat z0 = _mm_cmpeq_ps (f0, _mm_setzero_ps());
vFloat z1 = _mm_cmpeq_ps (f1, _mm_setzero_ps());
r0 = _mm_andnot_ps (z0, r0);
r1 = _mm_andnot_ps (z1, r1);
z0 = _mm_and_ps (z0, p0);
z1 = _mm_and_ps (z1, p0);
r0 = _mm_or_ps (r0, z0);
r1 = _mm_or_ps (r1, z1);
_mm_storeu_ps (dr, r0);
_mm_storeu_ps (dr + 4, r1);
dr += 8;
}
return 0;
}
long foo (job *j )
{
fj *jd = (fj*) j;
return bar (&jd->src, &jd->dest, jd->g, jd->flags);
}
/* { dg-final { scan-rtl-dump-not "deleted 1 dead insns" "csa" } } */
/* { dg-final { cleanup-rtl-dump "csa" } } */
|
