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
|
/* SPDX-License-Identifier: GPL-2.0
*
* Copyright (C) 2018 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
*/
#include <linux/init.h>
#include <linux/module.h>
#include <linux/delay.h>
#include <linux/slab.h>
#include <linux/sort.h>
#include <asm/cpufeature.h>
#include <asm/processor.h>
#include <asm/fpu/api.h>
#include <asm/simd.h>
static unsigned long stamp = 0;
module_param(stamp, ulong, 0);
int dummy;
enum { BLOCKS_PER_CALL = 16 };
static u8 state[128];
static u8 input[64 * BLOCKS_PER_CALL];
#define declare_it(name) \
asmlinkage void blake2s_compress_ ## name(u8 state[128], const u8 *block, const size_t nblocks, const u32 inc); \
static __always_inline u8 name(void) \
{ \
blake2s_compress_ ## name(state, input, BLOCKS_PER_CALL, 0); \
return input[0]; \
}
#define do_it(name) do { \
u32 eax = 0, ebx = 0, ecx = 0, edx = 0; \
for (i = 0; i < WARMUP; ++i) \
ret |= name(); \
asm volatile("cpuid" : "+a" (eax), "=b" (ebx), "=d" (edx), "+c" (ecx)); \
for (i = 0; i <= TRIALS; ++i) { \
trial_times[i] = get_cycles(); \
ret |= name(); \
} \
for (i = 0; i < TRIALS; ++i) \
trial_times[i] = trial_times[i + 1] - trial_times[i]; \
sort(trial_times, TRIALS + 1, sizeof(cycles_t), compare_cycles, NULL); \
median_ ## name = trial_times[TRIALS / 2]; \
} while (0)
#define report_it(name) do { \
pr_err("%lu: %12s: %6llu cycles per block\n", stamp, #name, median_ ## name / BLOCKS_PER_CALL); \
} while (0)
declare_it(avx)
declare_it(avx512_ymm)
declare_it(avx512_zmm)
static int compare_cycles(const void *a, const void *b)
{
return *((cycles_t *)a) - *((cycles_t *)b);
}
static int __init mod_init(void)
{
enum { WARMUP = 6000, TRIALS = 5000, IDLE = 1 * 1000 };
int ret = 0, i;
cycles_t *trial_times;
cycles_t median_avx = 0;
cycles_t median_avx512_ymm = 0;
cycles_t median_avx512_zmm = 0;
unsigned long flags;
DEFINE_SPINLOCK(lock);
trial_times = kcalloc(TRIALS + 1, sizeof(cycles_t), GFP_KERNEL);
if (!trial_times)
return -ENOMEM;
msleep(IDLE);
spin_lock_irqsave(&lock, flags);
kernel_fpu_begin();
do_it(avx);
do_it(avx512_ymm);
do_it(avx512_zmm);
kernel_fpu_end();
spin_unlock_irqrestore(&lock, flags);
report_it(avx);
report_it(avx512_ymm);
report_it(avx512_zmm);
/* Don't let compiler be too clever. */
dummy = ret;
kfree(trial_times);
/* We should never actually agree to insert the module. Choosing
* -0x1000 here is an amazing hack. It causes the kernel to not
* actually load the module, while the standard userspace tools
* don't return an error, because it's too big. */
return -0x1000;
}
module_init(mod_init);
MODULE_LICENSE("GPL v2");
MODULE_DESCRIPTION("kBench9000 Cycle Counter");
MODULE_AUTHOR("Jason A. Donenfeld <Jason@zx2c4.com>");
|
