On Thu, Aug 09, 2012 at 04:22:04PM +0100, Jan Beulich wrote:
> >>> On 09.08.12 at 17:03, "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
...
> > ---
> > arch/x86/include/asm/page.h | 2 ++
> > arch/x86/include/asm/string_32.h | 5 +++++
> > arch/x86/include/asm/string_64.h | 5 +++++
> > arch/x86/lib/Makefile | 1 +
> > arch/x86/lib/clear_page_nocache_32.S | 30 ++++++++++++++++++++++++++++++
> > arch/x86/lib/clear_page_nocache_64.S | 29 +++++++++++++++++++++++++++++
>
> Couldn't this more reasonably go into clear_page_{32,64}.S?
We don't have clear_page_32.S.
> >+ xorl %eax,%eax
> >+ movl $4096/64,%ecx
> >+ .p2align 4
> >+.Lloop:
> >+ decl %ecx
> >+#define PUT(x) movnti %eax,x*8(%edi) ; movnti %eax,x*8+4(%edi)
>
> Is doing twice as much unrolling as on 64-bit really worth it?
Moving 64 bytes per cycle is faster on Sandy Bridge, but slower on
Westmere. Any preference? ;)
Westmere:
Performance counter stats for './test_unroll32' (20 runs):
31498.420608 task-clock # 0.998 CPUs utilized ( +- 0.25% )
40 context-switches # 0.001 K/sec ( +- 1.40% )
0 CPU-migrations # 0.000 K/sec ( +-100.00% )
89 page-faults # 0.003 K/sec ( +- 0.13% )
74,728,231,935 cycles # 2.372 GHz ( +- 0.25% ) [83.34%]
53,789,969,009 stalled-cycles-frontend # 71.98% frontend cycles idle ( +- 0.35% ) [83.33%]
41,681,014,054 stalled-cycles-backend # 55.78% backend cycles idle ( +- 0.43% ) [66.67%]
37,992,733,278 instructions # 0.51 insns per cycle
# 1.42 stalled cycles per insn ( +- 0.05% ) [83.33%]
3,561,376,245 branches # 113.065 M/sec ( +- 0.05% ) [83.33%]
27,182,795 branch-misses # 0.76% of all branches ( +- 0.06% ) [83.33%]
31.558545812 seconds time elapsed ( +- 0.25% )
Performance counter stats for './test_unroll64' (20 runs):
31564.753623 task-clock # 0.998 CPUs utilized ( +- 0.19% )
39 context-switches # 0.001 K/sec ( +- 0.40% )
0 CPU-migrations # 0.000 K/sec
90 page-faults # 0.003 K/sec ( +- 0.12% )
74,886,045,192 cycles # 2.372 GHz ( +- 0.19% ) [83.33%]
57,477,323,995 stalled-cycles-frontend # 76.75% frontend cycles idle ( +- 0.26% ) [83.34%]
44,548,142,150 stalled-cycles-backend # 59.49% backend cycles idle ( +- 0.31% ) [66.67%]
32,940,027,099 instructions # 0.44 insns per cycle
# 1.74 stalled cycles per insn ( +- 0.05% ) [83.34%]
1,884,944,093 branches # 59.717 M/sec ( +- 0.05% ) [83.32%]
1,027,135 branch-misses # 0.05% of all branches ( +- 0.56% ) [83.34%]
31.621001407 seconds time elapsed ( +- 0.19% )
Sandy Bridge:
Performance counter stats for './test_unroll32' (20 runs):
8578.382891 task-clock # 0.997 CPUs utilized ( +- 0.08% )
15 context-switches # 0.000 M/sec ( +- 2.97% )
0 CPU-migrations # 0.000 M/sec
84 page-faults # 0.000 M/sec ( +- 0.13% )
29,154,476,597 cycles # 3.399 GHz ( +- 0.08% ) [83.33%]
11,851,215,147 stalled-cycles-frontend # 40.65% frontend cycles idle ( +- 0.20% ) [83.33%]
1,530,172,593 stalled-cycles-backend # 5.25% backend cycles idle ( +- 1.44% ) [66.67%]
37,915,778,094 instructions # 1.30 insns per cycle
# 0.31 stalled cycles per insn ( +- 0.00% ) [83.34%]
3,590,533,447 branches # 418.556 M/sec ( +- 0.01% ) [83.35%]
26,500,765 branch-misses # 0.74% of all branches ( +- 0.01% ) [83.34%]
8.604638449 seconds time elapsed ( +- 0.08% )
Performance counter stats for './test_unroll64' (20 runs):
8463.789963 task-clock # 0.997 CPUs utilized ( +- 0.07% )
14 context-switches # 0.000 M/sec ( +- 1.70% )
0 CPU-migrations # 0.000 M/sec ( +-100.00% )
85 page-faults # 0.000 M/sec ( +- 0.12% )
28,763,328,688 cycles # 3.398 GHz ( +- 0.07% ) [83.32%]
13,517,462,952 stalled-cycles-frontend # 47.00% frontend cycles idle ( +- 0.14% ) [83.33%]
1,356,208,859 stalled-cycles-backend # 4.72% backend cycles idle ( +- 1.42% ) [66.68%]
32,885,492,141 instructions # 1.14 insns per cycle
# 0.41 stalled cycles per insn ( +- 0.00% ) [83.34%]
1,912,094,072 branches # 225.915 M/sec ( +- 0.02% ) [83.34%]
305,896 branch-misses # 0.02% of all branches ( +- 1.05% ) [83.33%]
8.488304839 seconds time elapsed ( +- 0.07% )
$ cat test.c
#include <stdio.h>
#include <sys/mman.h>
#define SIZE 1024*1024*1024
void clear_page_nocache_sse2(void *page) __attribute__((regparm(1)));
int main(int argc, char** argv)
{
char *p;
unsigned long i, j;
p = mmap(NULL, SIZE, PROT_WRITE|PROT_READ,
MAP_PRIVATE|MAP_ANONYMOUS|MAP_POPULATE, -1, 0);
for(j = 0; j < 100; j++) {
for(i = 0; i < SIZE; i += 4096) {
clear_page_nocache_sse2(p + i);
}
}
return 0;
}
$ cat clear_page_nocache_unroll32.S
.globl clear_page_nocache_sse2
.align 4,0x90
clear_page_nocache_sse2:
.cfi_startproc
mov %eax,%edx
xorl %eax,%eax
movl $4096/32,%ecx
.p2align 4
.Lloop_sse2:
decl %ecx
#define PUT(x) movnti %eax,x*4(%edx)
PUT(0)
PUT(1)
PUT(2)
PUT(3)
PUT(4)
PUT(5)
PUT(6)
PUT(7)
#undef PUT
lea 32(%edx),%edx
jnz .Lloop_sse2
nop
ret
.cfi_endproc
.type clear_page_nocache_sse2, @function
.size clear_page_nocache_sse2, .-clear_page_nocache_sse2
$ cat clear_page_nocache_unroll64.S
.globl clear_page_nocache_sse2
.align 4,0x90
clear_page_nocache_sse2:
.cfi_startproc
mov %eax,%edx
xorl %eax,%eax
movl $4096/64,%ecx
.p2align 4
.Lloop_sse2:
decl %ecx
#define PUT(x) movnti %eax,x*8(%edx) ; movnti %eax,x*8+4(%edx)
PUT(0)
PUT(1)
PUT(2)
PUT(3)
PUT(4)
PUT(5)
PUT(6)
PUT(7)
#undef PUT
lea 64(%edx),%edx
jnz .Lloop_sse2
nop
ret
.cfi_endproc
.type clear_page_nocache_sse2, @function
.size clear_page_nocache_sse2, .-clear_page_nocache_sse2
--
Kirill A. Shutemov
Attachment:
signature.asc
Description: Digital signature
