Hi Linhaifeng,
On Tue, May 12, 2015 at 1:13 AM, Linhaifeng <haifeng.lin at huawei.com> wrote:
> Hi, Ravi Kerur
>
> On 2015/5/9 5:19, Ravi Kerur wrote:
> > Preliminary results on Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz, Ubuntu
> > 14.04 x86_64 shows comparisons using AVX/SSE instructions taking 1/3rd
> > CPU ticks for 16, 32, 48 and 64 bytes comparison. In addition,
>
> I had write a program to test rte_memcmp and I have a question about the
> result.
> Why cost same CPU ticks for 128 256 512 1024 1500 bytes? Is there any
> problem in
> my test?
>
>
If you can wait until Thursday I will probably send v3 patch which will
have full memcmp support.
In your program try with volatile pointer and see if it helps.
>
> [root at localhost test]# gcc avx_test.c -O3 -I
> /data/linhf/v2r2c00/open-source/dpdk/dpdk-2.0.0/x86_64-native-linuxapp-gcc/include/
> -mavx2 -DRTE_MACHINE_CPUFLAG_AVX2
> [root at localhost test]# ./a.out 0
> each test run 100000000 times
> copy 16 bytes costs average 7(rte_memcmp) 10(memcmp) ticks
> copy 32 bytes costs average 9(rte_memcmp) 11(memcmp) ticks
> copy 64 bytes costs average 6(rte_memcmp) 13(memcmp) ticks
> copy 128 bytes costs average 11(rte_memcmp) 14(memcmp) ticks
> copy 256 bytes costs average 9(rte_memcmp) 14(memcmp) ticks
> copy 512 bytes costs average 9(rte_memcmp) 14(memcmp) ticks
> copy 1024 bytes costs average 9(rte_memcmp) 14(memcmp) ticks
> copy 1500 bytes costs average 11(rte_memcmp) 14(memcmp) ticks
> [root at localhost test]# ./a.out 1
> each test run 100000000 times
> copy 16 bytes costs average 2(rte_memcpy) 10(memcpy) ticks
> copy 32 bytes costs average 2(rte_memcpy) 10(memcpy) ticks
> copy 64 bytes costs average 3(rte_memcpy) 10(memcpy) ticks
> copy 128 bytes costs average 7(rte_memcpy) 12(memcpy) ticks
> copy 256 bytes costs average 9(rte_memcpy) 23(memcpy) ticks
> copy 512 bytes costs average 14(rte_memcpy) 34(memcpy) ticks
> copy 1024 bytes costs average 37(rte_memcpy) 61(memcpy) ticks
> copy 1500 bytes costs average 62(rte_memcpy) 87(memcpy) ticks
>
>
> Here is my program:
>
> #include <stdio.h>
> #include <rte_cycles.h>
> #include <smmintrin.h>
> #include <rte_memcpy.h>
> #include <rte_memcmp.h>
>
> #define TIMES 100000000L
>
> void test_memcpy(size_t n)
> {
> uint64_t start, end, i, start2, end2;
> uint8_t *src, *dst;
>
> src = (uint8_t*)malloc(n * sizeof(uint8_t));
> dst = (uint8_t*)malloc(n * sizeof(uint8_t));
>
> start = rte_rdtsc();
> for (i = 0; i < TIMES; i++) {
> rte_memcpy(dst, src, n);
> }
> end = rte_rdtsc();
>
> start2 = rte_rdtsc();
> for (i = 0; i < TIMES; i++) {
> memcpy(dst, src, n);
> }
> end2 = rte_rdtsc();
>
>
> free(src);
> free(dst);
>
> printf("copy %u bytes costs average %llu(rte_memcpy) %llu(memcpy)
> ticks\n", n, (end - start)/TIMES, (end2 - start2)/TIMES);
> }
>
> int test_memcmp(size_t n)
> {
> uint64_t start, end, i, start2, end2, j;
> uint8_t *src, *dst;
> int *ret;
>
> src = (uint8_t*)malloc(n * sizeof(uint8_t));
> dst = (uint8_t*)malloc(n * sizeof(uint8_t));
> ret = (int*)malloc(TIMES * sizeof(int));
>
> start = rte_rdtsc();
> for (i = 0; i < TIMES; i++) {
> ret[i] = rte_memcmp(dst, src, n);
> }
> end = rte_rdtsc();
>
> start2 = rte_rdtsc();
> for (i = 0; i < TIMES; i++) {
> ret[i] = memcmp(dst, src, n);
> }
> end2 = rte_rdtsc();
>
> // avoid gcc to optimize memcmp
> for (i = 0; i < TIMES; i++) {
> t += ret[i];
> }
>
> free(src);
> free(dst);
>
> printf("copy %u bytes costs average %llu(rte_memcmp) %llu(memcmp)
> ticks\n", n, (end - start)/TIMES, (end2 - start2)/TIMES);
> return t;
> }
>
>
>
>
> int main(int narg, char** args)
> {
> printf("each test run %llu times\n", TIMES);
>
> if (narg < 2) {
> printf("usage:./avx_test 0/1 1:test memcpy 0:test
> memcmp\n");
> return -1;
> }
>
> if (atoi(args[1])) {
> test_memcpy(16);
> test_memcpy(32);
> test_memcpy(64);
> test_memcpy(128);
> test_memcpy(256);
> test_memcpy(512);
> test_memcpy(1024);
> test_memcpy(1500);
> } else {
> test_memcmp(16);
> test_memcmp(32);
> test_memcmp(64);
> test_memcmp(128);
> test_memcmp(256);
> test_memcmp(512);
> test_memcmp(1024);
> test_memcmp(1500);
> }
> }
>
>
>
>
>
>
>
