Menu
▾
▴
[linuxsh-dev] Slow 32 bytes long "mov" delta - Hitachi SH4(7750), SolutionEngine(MS7750SE01E)
|
From: Vladimir D. <vdo...@wa...> - 2001-12-06 21:57:16
|
Hi,
Can anyone tell me what is going on with a performance.
Isn't it the processor issue?
If I copy data from one place to another after each megabyte (delta
between src and dst)
there is some kind of "black hole" 32 bytes long in which processor
works 15 times
slower.
It happens after 1M, 2M, 3M, have not tested more.
0..................1M|32bytes.............2M|32bytes.............3M|32bytes
|--------------------|xxx-------------------|xxx-------------------|xxx
N = 1.2.3
Delta between src and dst 0M <= DELTA < 1M -fast
Delta between src and dst 1M*N <= DELTA <= 1M*N+31 -15 times
slower
Delta between src and dst 1M*N+31 < DELTA < 1M*N -fast
It doesn't matter is memory area PAGE_SIZE aligned and not.
I even tested with mmap'ed physical memory allocated at boot time with
"alloc_bootmem_pages" to be sure that it is continuous and not-swappable
- the same result ...
Here is a simplified demo code, output, and disassembly.
Thanks,
Vova
Output:
u8copy1mm1 3145728 chars = 0.174000 sec.
u8copy1m 3145728 chars = 2.392500 sec.
u8copy1mp31 3145728 chars = 2.391000 sec.
u8copy1mp32 3145728 chars = 0.174000 sec.
Code:
//1024*1024*3
#define THREEMEG 3145728
//1M
#define ONEMEGMINUS1 1048575
#define ONEMEG 1048576
#define ONEMEGPLUS31 1048607
#define ONEMEGPLUS32 1048608
void u8copy1mm1(void)
{
for(ii = 0; ii < THREEMEG; ii++)
mem8_1[0]=mem8_1[ONEMEGMINUS1];
}
void u8copy1m(void)
{
for(ii = 0; ii < THREEMEG; ii++)
mem8_1[0]=mem8_1[ONEMEG];
}
void u8copy1mp31(void)
{
for(ii = 0; ii < THREEMEG; ii++)
mem8_1[0]=mem8_1[ONEMEGPLUS31];
}
void u8copy1mp32(void)
{
for(ii = 0; ii < THREEMEG; ii++)
mem8_1[0]=mem8_1[ONEMEGPLUS32];
}
start = clock();
for( k = 0; k < NUMBEROFLOOPS; k++ )
u8copy1mm1();
end = clock();
cpu_time_used = (((double)(end-start))/CLOCKS_PER_SEC)/k;
printf("u8copy1mm1 %i chars = %f sec.\n",THREEMEG,cpu_time_used);
start = clock();
for( k = 0; k < NUMBEROFLOOPS; k++ )
u8copy1m();
end = clock();
cpu_time_used = (((double)(end-start))/CLOCKS_PER_SEC)/k;
printf("u8copy1m %i chars = %f sec.\n",THREEMEG,cpu_time_used);
start = clock();
for( k = 0; k < NUMBEROFLOOPS; k++ )
u8copy1mp31();
end = clock();
cpu_time_used = (((double)(end-start))/CLOCKS_PER_SEC)/k;
printf("u8copy1mp31 %i chars = %f sec.\n",THREEMEG,cpu_time_used);
start = clock();
for( k = 0; k < NUMBEROFLOOPS; k++ )
u8copy1mp32();
end = clock();
cpu_time_used = (((double)(end-start))/CLOCKS_PER_SEC)/k;
printf("u8copy1mp32 %i chars = %f sec.\n",THREEMEG,cpu_time_used);
Output:
mem1 = 0x29711000
THREEMEG = 3145728 NUMBEROFLOOPS = 10
u8copy1mm1 3145728 chars = 0.174000 sec.
u8copy1m 3145728 chars = 2.392500 sec.
u8copy1mp31 3145728 chars = 2.391000 sec.
u8copy1mp32 3145728 chars = 0.174000 sec.
Disassembly:
00402240 <u8copy1mm1>:
402240: 09 d2 mov.l
402268<u8copy1mm1+0x28>,r2!0x413940
402242: 00 e1 mov #0,r1
402244: 09 d6 mov.l
40226c<u8copy1mm1+0x2c>,r6!0x2fffff
402246: 0a d7 mov.l
402270<u8copy1mm1+0x30>,r7!0x413900
402248: 23 63 mov r2,r3
40224a: 0a d0 mov.l
402274<u8copy1mm1+0x34>,r0!0xfffff
40224c: e6 2f mov.l r14,@-r15
40224e: 12 22 mov.l r1,@r2
402250: f3 6e mov r15,r14
402252: 72 62 mov.l @r7,r2
402254: 2c 04 mov.b @(r0,r2),r4
402256: 40 22 mov.b r4,@r2
402258: 32 61 mov.l @r3,r1
40225a: 01 71 add #1,r1
40225c: 67 31 cmp/gt r6,r1
40225e: f8 8f bf.s 402252 <u8copy1mm1+0x12>
402260: 12 23 mov.l r1,@r3
402262: e3 6f mov r14,r15
402264: 0b 00 rts
402266: f6 6e mov.l @r15+,r14
402268: 40 39 cmp/eq r4,r9
40226a: 41 00 .word 0x0041
40226c: ff ff .word 0xffff
40226e: 2f 00 mac.l @r2+,@r0+
402270: 00 39 cmp/eq r0,r9
402272: 41 00 .word 0x0041
402274: ff ff .word 0xffff
402276: 0f 00 mac.l @r0+,@r0+
402278: 09 00 nop
40227a: 09 00 nop
40227c: 09 00 nop
40227e: 09 00 nop
00402280 <u8copy1m>:
402280: 09 d2 mov.l
4022a8<u8copy1m+0x28>,r2!0x413940
402282: 00 e1 mov #0,r1
402284: 09 d6 mov.l
4022ac<u8copy1m+0x2c>,r6!0x2fffff
402286: 0a d7 mov.l
4022b0<u8copy1m+0x30>,r7!0x413900
402288: 23 63 mov r2,r3
40228a: 0a d0 mov.l
4022b4<u8copy1m+0x34>,r0!0x100000
40228c: e6 2f mov.l r14,@-r15
40228e: 12 22 mov.l r1,@r2
402290: f3 6e mov r15,r14
402292: 72 62 mov.l @r7,r2
402294: 2c 04 mov.b @(r0,r2),r4
402296: 40 22 mov.b r4,@r2
402298: 32 61 mov.l @r3,r1
40229a: 01 71 add #1,r1
40229c: 67 31 cmp/gt r6,r1
40229e: f8 8f bf.s 402292 <u8copy1m+0x12>
4022a0: 12 23 mov.l r1,@r3
4022a2: e3 6f mov r14,r15
4022a4: 0b 00 rts
4022a6: f6 6e mov.l @r15+,r14
4022a8: 40 39 cmp/eq r4,r9
4022aa: 41 00 .word 0x0041
4022ac: ff ff .word 0xffff
4022ae: 2f 00 mac.l @r2+,@r0+
4022b0: 00 39 cmp/eq r0,r9
4022b2: 41 00 .word 0x0041
4022b4: 00 00 .word 0x0000
4022b6: 10 00 .word 0x0010
4022b8: 09 00 nop
4022ba: 09 00 nop
4022bc: 09 00 nop
4022be: 09 00 nop
004022c0 <u8copy1mp31>:
4022c0: 09 d2 mov.l
4022e8<u8copy1mp31+0x28>,r2!0x413940
4022c2: 00 e1 mov #0,r1
4022c4: 09 d6 mov.l
4022ec<u8copy1mp31+0x2c>,r6!0x2fffff
4022c6: 0a d7 mov.l
4022f0<u8copy1mp31+0x30>,r7!0x413900
4022c8: 23 63 mov r2,r3
4022ca: 0a d0 mov.l
4022f4<u8copy1mp31+0x34>,r0!0x10001f
4022cc: e6 2f mov.l r14,@-r15
4022ce: 12 22 mov.l r1,@r2
4022d0: f3 6e mov r15,r14
4022d2: 72 62 mov.l @r7,r2
4022d4: 2c 04 mov.b @(r0,r2),r4
4022d6: 40 22 mov.b r4,@r2
4022d8: 32 61 mov.l @r3,r1
4022da: 01 71 add #1,r1
4022dc: 67 31 cmp/gt r6,r1
4022de: f8 8f bf.s 4022d2 <u8copy1mp31+0x12>
4022e0: 12 23 mov.l r1,@r3
4022e2: e3 6f mov r14,r15
4022e4: 0b 00 rts
4022e6: f6 6e mov.l @r15+,r14
4022e8: 40 39 cmp/eq r4,r9
4022ea: 41 00 .word 0x0041
4022ec: ff ff .word 0xffff
4022ee: 2f 00 mac.l @r2+,@r0+
4022f0: 00 39 cmp/eq r0,r9
4022f2: 41 00 .word 0x0041
4022f4: 1f 00 mac.l @r1+,@r0+
4022f6: 10 00 .word 0x0010
4022f8: 09 00 nop
4022fa: 09 00 nop
4022fc: 09 00 nop
4022fe: 09 00 nop
00402300 <u8copy1mp32>:
402300: 09 d2 mov.l
402328<u8copy1mp32+0x28>,r2!0x413940
402302: 00 e1 mov #0,r1
402304: 09 d6 mov.l
40232c<u8copy1mp32+0x2c>,r6!0x2fffff
402306: 0a d7 mov.l
402330<u8copy1mp32+0x30>,r7!0x413900
402308: 23 63 mov r2,r3
40230a: 0a d0 mov.l
402334<u8copy1mp32+0x34>,r0!0x100020
40230c: e6 2f mov.l r14,@-r15
40230e: 12 22 mov.l r1,@r2
402310: f3 6e mov r15,r14
402312: 72 62 mov.l @r7,r2
402314: 2c 04 mov.b @(r0,r2),r4
402316: 40 22 mov.b r4,@r2
402318: 32 61 mov.l @r3,r1
40231a: 01 71 add #1,r1
40231c: 67 31 cmp/gt r6,r1
40231e: f8 8f bf.s 402312 <u8copy1mp32+0x12>
402320: 12 23 mov.l r1,@r3
402322: e3 6f mov r14,r15
402324: 0b 00 rts
402326: f6 6e mov.l @r15+,r14
402328: 40 39 cmp/eq r4,r9
40232a: 41 00 .word 0x0041
40232c: ff ff .word 0xffff
40232e: 2f 00 mac.l @r2+,@r0+
402330: 00 39 cmp/eq r0,r9
402332: 41 00 .word 0x0041
402334: 20 00 .word 0x0020
402336: 10 00 .word 0x0010
402338: 09 00 nop
40233a: 09 00 nop
40233c: 09 00 nop
40233e: 09 00 nop
|
