| line |
stmt |
bran |
cond |
sub |
pod |
time |
code |
|
1
|
|
|
|
|
|
|
#define PERL_NO_GET_CONTEXT |
|
2
|
|
|
|
|
|
|
#include "EXTERN.h" |
|
3
|
|
|
|
|
|
|
#include "perl.h" |
|
4
|
|
|
|
|
|
|
#include "XSUB.h" |
|
5
|
|
|
|
|
|
|
|
|
6
|
|
|
|
|
|
|
#include |
|
7
|
|
|
|
|
|
|
|
|
8
|
|
|
|
|
|
|
#if defined(__AVX2__) || defined(__SSE4_2__) |
|
9
|
|
|
|
|
|
|
#include |
|
10
|
|
|
|
|
|
|
#endif |
|
11
|
|
|
|
|
|
|
|
|
12
|
|
|
|
|
|
|
/* Fetch the raw int64_t buffer behind a packed Perl string. rw forces |
|
13
|
|
|
|
|
|
|
the SV out of copy-on-write before we write into its buffer. */ |
|
14
|
|
|
|
|
|
|
static int64_t * |
|
15
|
7246
|
|
|
|
|
|
aeps_buf(pTHX_ SV *sv, IV window, int rw) |
|
16
|
|
|
|
|
|
|
{ |
|
17
|
|
|
|
|
|
|
STRLEN len; |
|
18
|
7246
|
100
|
|
|
|
|
char *p = rw ? SvPV_force(sv, len) : SvPV(sv, len); |
|
19
|
7246
|
100
|
|
|
|
|
if ((IV) len < window * (IV) sizeof(int64_t)) |
|
20
|
4
|
|
|
|
|
|
croak("Algorithm::EventsPerSecond: ring buffer smaller than window"); |
|
21
|
7242
|
|
|
|
|
|
return (int64_t *) p; |
|
22
|
|
|
|
|
|
|
} |
|
23
|
|
|
|
|
|
|
|
|
24
|
|
|
|
|
|
|
static int64_t |
|
25
|
2022
|
|
|
|
|
|
aeps_count(int64_t *buckets, int64_t *stamps, IV window, IV oldest) |
|
26
|
|
|
|
|
|
|
{ |
|
27
|
2022
|
|
|
|
|
|
int64_t sum = 0; |
|
28
|
2022
|
|
|
|
|
|
IV i = 0; |
|
29
|
|
|
|
|
|
|
|
|
30
|
|
|
|
|
|
|
#if defined(__AVX2__) |
|
31
|
|
|
|
|
|
|
/* 4 buckets per iteration: stamps >= oldest becomes an all-ones |
|
32
|
|
|
|
|
|
|
lane mask (>= via > oldest-1), which gates the bucket values |
|
33
|
|
|
|
|
|
|
feeding a vector accumulator. */ |
|
34
|
|
|
|
|
|
|
{ |
|
35
|
|
|
|
|
|
|
__m256i vsum = _mm256_setzero_si256(); |
|
36
|
|
|
|
|
|
|
__m256i cutoff = _mm256_set1_epi64x((int64_t) oldest - 1); |
|
37
|
|
|
|
|
|
|
int64_t lanes[4]; |
|
38
|
|
|
|
|
|
|
for (; i + 4 <= window; i += 4) { |
|
39
|
|
|
|
|
|
|
__m256i s = _mm256_loadu_si256((__m256i const *) (stamps + i)); |
|
40
|
|
|
|
|
|
|
__m256i b = _mm256_loadu_si256((__m256i const *) (buckets + i)); |
|
41
|
|
|
|
|
|
|
__m256i keep = _mm256_cmpgt_epi64(s, cutoff); |
|
42
|
|
|
|
|
|
|
vsum = _mm256_add_epi64(vsum, _mm256_and_si256(b, keep)); |
|
43
|
|
|
|
|
|
|
} |
|
44
|
|
|
|
|
|
|
_mm256_storeu_si256((__m256i *) lanes, vsum); |
|
45
|
|
|
|
|
|
|
sum = lanes[0] + lanes[1] + lanes[2] + lanes[3]; |
|
46
|
|
|
|
|
|
|
} |
|
47
|
|
|
|
|
|
|
#elif defined(__SSE4_2__) |
|
48
|
|
|
|
|
|
|
/* Same masked sum, 2 buckets per iteration. */ |
|
49
|
|
|
|
|
|
|
{ |
|
50
|
|
|
|
|
|
|
__m128i vsum = _mm_setzero_si128(); |
|
51
|
|
|
|
|
|
|
__m128i cutoff = _mm_set1_epi64x((int64_t) oldest - 1); |
|
52
|
|
|
|
|
|
|
int64_t lanes[2]; |
|
53
|
|
|
|
|
|
|
for (; i + 2 <= window; i += 2) { |
|
54
|
|
|
|
|
|
|
__m128i s = _mm_loadu_si128((__m128i const *) (stamps + i)); |
|
55
|
|
|
|
|
|
|
__m128i b = _mm_loadu_si128((__m128i const *) (buckets + i)); |
|
56
|
|
|
|
|
|
|
__m128i keep = _mm_cmpgt_epi64(s, cutoff); |
|
57
|
|
|
|
|
|
|
vsum = _mm_add_epi64(vsum, _mm_and_si128(b, keep)); |
|
58
|
|
|
|
|
|
|
} |
|
59
|
|
|
|
|
|
|
_mm_storeu_si128((__m128i *) lanes, vsum); |
|
60
|
|
|
|
|
|
|
sum = lanes[0] + lanes[1]; |
|
61
|
|
|
|
|
|
|
} |
|
62
|
|
|
|
|
|
|
#endif |
|
63
|
|
|
|
|
|
|
|
|
64
|
|
|
|
|
|
|
/* Tail (and whole loop when no intrinsics were compiled in); |
|
65
|
|
|
|
|
|
|
branchless so -O3 can auto-vectorize it. */ |
|
66
|
284781
|
100
|
|
|
|
|
for (; i < window; i++) |
|
67
|
282759
|
|
|
|
|
|
sum += buckets[i] & -(int64_t) (stamps[i] >= (int64_t) oldest); |
|
68
|
|
|
|
|
|
|
|
|
69
|
2022
|
|
|
|
|
|
return sum; |
|
70
|
|
|
|
|
|
|
} |
|
71
|
|
|
|
|
|
|
|
|
72
|
|
|
|
|
|
|
MODULE = Algorithm::EventsPerSecond::XS PACKAGE = Algorithm::EventsPerSecond::XS |
|
73
|
|
|
|
|
|
|
|
|
74
|
|
|
|
|
|
|
PROTOTYPES: DISABLE |
|
75
|
|
|
|
|
|
|
|
|
76
|
|
|
|
|
|
|
void |
|
77
|
|
|
|
|
|
|
_xs_mark(SV *buckets_sv, SV *stamps_sv, IV window, IV now, IV count) |
|
78
|
|
|
|
|
|
|
CODE: |
|
79
|
|
|
|
|
|
|
{ |
|
80
|
1600
|
|
|
|
|
|
int64_t *buckets = aeps_buf(aTHX_ buckets_sv, window, 1); |
|
81
|
1599
|
|
|
|
|
|
int64_t *stamps = aeps_buf(aTHX_ stamps_sv, window, 1); |
|
82
|
1598
|
|
|
|
|
|
IV i = now % window; |
|
83
|
|
|
|
|
|
|
|
|
84
|
1598
|
100
|
|
|
|
|
if (stamps[i] != (int64_t) now) { |
|
85
|
1590
|
|
|
|
|
|
buckets[i] = 0; |
|
86
|
1590
|
|
|
|
|
|
stamps[i] = (int64_t) now; |
|
87
|
|
|
|
|
|
|
} |
|
88
|
1598
|
|
|
|
|
|
buckets[i] += (int64_t) count; |
|
89
|
|
|
|
|
|
|
} |
|
90
|
|
|
|
|
|
|
|
|
91
|
|
|
|
|
|
|
IV |
|
92
|
|
|
|
|
|
|
_xs_count(SV *buckets_sv, SV *stamps_sv, IV window, IV oldest) |
|
93
|
|
|
|
|
|
|
CODE: |
|
94
|
2024
|
|
|
|
|
|
RETVAL = (IV) aeps_count( |
|
95
|
|
|
|
|
|
|
aeps_buf(aTHX_ buckets_sv, window, 0), |
|
96
|
|
|
|
|
|
|
aeps_buf(aTHX_ stamps_sv, window, 0), |
|
97
|
|
|
|
|
|
|
window, oldest); |
|
98
|
|
|
|
|
|
|
OUTPUT: |
|
99
|
|
|
|
|
|
|
RETVAL |
|
100
|
|
|
|
|
|
|
|
|
101
|
|
|
|
|
|
|
const char * |
|
102
|
|
|
|
|
|
|
_xs_simd() |
|
103
|
|
|
|
|
|
|
CODE: |
|
104
|
|
|
|
|
|
|
#if defined(__AVX2__) |
|
105
|
|
|
|
|
|
|
RETVAL = "AVX2"; |
|
106
|
|
|
|
|
|
|
#elif defined(__SSE4_2__) |
|
107
|
|
|
|
|
|
|
RETVAL = "SSE4.2"; |
|
108
|
|
|
|
|
|
|
#else |
|
109
|
3
|
|
|
|
|
|
RETVAL = "scalar"; |
|
110
|
|
|
|
|
|
|
#endif |
|
111
|
|
|
|
|
|
|
OUTPUT: |
|
112
|
|
|
|
|
|
|
RETVAL |