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
|
// Copyright (c) 2021-2024, Intel Corporation
// SPDX-License-Identifier: BSD-3-Clause
#include <benchmark/benchmark.h>
#include <bitset>
#include <cstdint>
#include <stdio.h>
#include "../common.h"
#include "05_packed_load_store_ispc.h"
static Docs
docs("Check packed_load_active/packed_store_active implementation of stdlib functions:\n"
"[int32, int64] x [all_off, 1/16, 1/8, 1/4, 1/2, 3/4, 7/8, 15/16, all_on] versions.\n"
"Observation:\n"
" - it does make sense to test multiple mask pattern, as they behave substantantially differently for "
"different implementations\n"
" - implementations based on LLVM intrinsic is a good one, but not the best, as it don't do popcnt as "
"part of intrinsic\n"
" - performance was not evaluated for different data set sizes\n"
" - 1/16, 1/8 behave like 1/4 for 4 wide targets, similarly 1/16 on 8 wide is 1/8\n"
" - 1/2 even and 1/2 odd have different perfomance, that's not expected.\n"
" - 1/2 even is about 20% regression on AVX2 (LLVM intrinsics vs old manual implementation) that's might be "
"interesting to investigate\n"
"Expectation:\n"
" - No regressions\n");
WARM_UP_RUN();
// Minimum size is maximum target width, i.e. 64.
// Larger buffer is better, but preferably to stay within L1.
#define ARGS Arg(8192)
// #define ARGS RangeMultiplier(2)->Range(64, 64<<15)->Complexity(benchmark::oN)
template <typename T> static void init(T *src, T *dst, int count) {
for (int i = 0; i < count; i++) {
src[i] = static_cast<T>(i + 1);
dst[i] = 0;
}
}
template <typename T>
static void check_packed_load_active_eq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
int width = ispc::width();
int loc = 0;
int checkVal = 0;
for (int i = 0; i < count / width; i++) {
for (int programIndex = 0; programIndex < width; programIndex++) {
if ((programIndex & index) == 0) {
checkVal += src[loc++];
}
}
}
if (checkVal != num) {
printf("Error check_packed_load_active index=%d.\n", index);
}
}
template <typename T>
static void check_packed_load_active_neq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
int width = ispc::width();
int loc = 0;
int checkVal = 0;
for (int i = 0; i < count / width; i++) {
for (int programIndex = 0; programIndex < width; programIndex++) {
if ((programIndex & index) != 0) {
checkVal += src[loc++];
}
}
}
if (checkVal != num) {
printf("Error check_packed_load_active index=%d.\n", index);
}
}
template <typename T>
static void check_packed_store_active_eq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
int width = ispc::width();
int srcLoc = 0;
int dstLoc = 0;
for (int i = 0; i < count / width; i++) {
for (int programIndex = 0; programIndex < width; programIndex++) {
if ((programIndex & index) == 0) {
if (src[srcLoc] != dst[dstLoc++]) {
printf("Error check_packed_store_active loc=%d.\n", index);
}
}
srcLoc++;
}
}
}
template <typename T>
static void check_packed_store_active_neq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
int width = ispc::width();
int srcLoc = 0;
int dstLoc = 0;
for (int i = 0; i < count / width; i++) {
for (int programIndex = 0; programIndex < width; programIndex++) {
if ((programIndex & index) != 0) {
if (src[srcLoc] != dst[dstLoc++]) {
printf("Error check_packed_store_active loc=%d.\n", index);
}
}
srcLoc++;
}
}
}
template <typename T>
static void check_packed_store_active2_eq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
check_packed_store_active_eq(src, dst, index, num, count);
}
template <typename T>
static void check_packed_store_active2_neq(T *src, T *dst, const unsigned int index, unsigned int num, int count) {
check_packed_store_active_neq(src, dst, index, num, count);
}
#define PACKED_LOAD_STORE_COND(T_C, T_ISPC, FUNC, ACTIVE_RATIO) \
static void FUNC##_##T_ISPC##_##ACTIVE_RATIO##_eq(benchmark::State &state) { \
int count = static_cast<int>(state.range(0)); \
T_C *dst = static_cast<T_C *>(aligned_alloc_helper(sizeof(T_C) * count)); \
T_C *src = static_cast<T_C *>(aligned_alloc_helper(sizeof(T_C) * count)); \
init(src, dst, count); \
const unsigned int index = ACTIVE_RATIO; \
unsigned int num = 0; \
\
for (auto _ : state) { \
num = ispc::FUNC##_##T_ISPC##_eq(src, dst, index, count); \
} \
\
check_##FUNC##_eq(src, dst, index, num, count); \
aligned_free_helper(src); \
aligned_free_helper(dst); \
state.SetComplexityN(state.range(0)); \
} \
BENCHMARK(FUNC##_##T_ISPC##_##ACTIVE_RATIO##_eq)->ARGS; \
\
static void FUNC##_##T_ISPC##_##ACTIVE_RATIO##_neq(benchmark::State &state) { \
int count = static_cast<int>(state.range(0)); \
T_C *dst = static_cast<T_C *>(aligned_alloc_helper(sizeof(T_C) * count)); \
T_C *src = static_cast<T_C *>(aligned_alloc_helper(sizeof(T_C) * count)); \
init(src, dst, count); \
const unsigned int index = ACTIVE_RATIO; \
unsigned int num = 0; \
\
for (auto _ : state) { \
num = ispc::FUNC##_##T_ISPC##_neq(src, dst, index, count); \
} \
\
check_##FUNC##_neq(src, dst, index, num, count); \
aligned_free_helper(src); \
aligned_free_helper(dst); \
state.SetComplexityN(state.range(0)); \
} \
BENCHMARK(FUNC##_##T_ISPC##_##ACTIVE_RATIO##_neq)->ARGS;
// The last argument constant means the following:
// 0 is all-on, all-off
// 1 is 1/2 - even, odd
// 3 is 1/4, 3/4
// 7 is 1/8, 7/8
// 15 is 1/16, 15/16
PACKED_LOAD_STORE_COND(int32_t, int32, packed_load_active, 0)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_load_active, 1)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_load_active, 3)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_load_active, 7)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_load_active, 15)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active, 0)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active, 1)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active, 3)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active, 7)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active, 15)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active2, 0)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active2, 1)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active2, 3)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active2, 7)
PACKED_LOAD_STORE_COND(int32_t, int32, packed_store_active2, 15)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_load_active, 0)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_load_active, 1)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_load_active, 3)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_load_active, 7)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_load_active, 15)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active, 0)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active, 1)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active, 3)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active, 7)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active, 15)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active2, 0)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active2, 1)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active2, 3)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active2, 7)
PACKED_LOAD_STORE_COND(int64_t, int64, packed_store_active2, 15)
BENCHMARK_MAIN();
|
