daiduo/TurboPFor-Integer-Compression
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

IcBench: Sample Benchmark App

Browse Source
This commit is contained in:
x
2018-01-13 16:02:24 +01:00
parent a135a7b125
commit 257ec4e17f
1 changed files with 89 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

126
icapp.c
View File

@ -65,55 +65,93 @@ void libmemcpy(unsigned char *dst, unsigned char *src, int len) {
memcpy_ptr(dst, src, len);
}
#define ID_MEMCPY 14
unsigned bench(unsigned char *in, unsigned n, unsigned char *out, unsigned char *cpy, int id) {
unsigned l;
#define ID_MEMCPY 16
unsigned bench32(unsigned char *in, unsigned n, unsigned char *out, unsigned char *cpy, int id) {
unsigned l,m=(n+3)/4;
memrcpy(cpy,in,n);
switch(id) {
case 1: TMBENCH("\np4nenc16 ",l=p4nenc16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec16( out, n/2, cpy) ,n); break;
case 2: TMBENCH("\np4nenc16v ",l=p4nenc128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec128v16( out, n/2, cpy) ,n); break;
case 1: TMBENCH("\np4nenc32 ",l=p4nenc32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec32( out, m, cpy) ,n); break;
case 2: TMBENCH("\np4nenc32v ",l=p4nenc128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec128v32( out, m, cpy) ,n); break;
case 3: TMBENCH("\np4nzenc16 ",l=p4nzenc16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec16( out, n/2, cpy) ,n); break;
case 4: TMBENCH("\np4nzenc16v ",l=p4nzenc128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec128v16( out, n/2, cpy) ,n); break;
case 3: TMBENCH("\np4nzenc32 ",l=p4nzenc32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec32( out, m, cpy) ,n); break;
case 4: TMBENCH("\np4nzenc32v ",l=p4nzenc128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec128v32( out, m, cpy) ,n); break;
case 5: TMBENCH("\np4ndenc16 ",l=p4ndenc16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec16( out, n/2, cpy) ,n); break;
case 6: TMBENCH("\np4ndenc16v ",l=p4ndenc128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec128v16( out, n/2, cpy) ,n); break;
case 5: TMBENCH("\np4ndenc32 ",l=p4ndenc32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec32( out, m, cpy) ,n); break;
case 6: TMBENCH("\np4ndenc32v ",l=p4ndenc128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec128v32( out, m, cpy) ,n); break;
case 7: TMBENCH("\nvbzenc16 ",l=vbzenc16( in, n/2, out,0)-out,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",vbzdec16( out, n/2, cpy,0) ,n); break;
case 7: TMBENCH("\nvbzenc32 ",l=vbzenc32( in, m, out,0)-out,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",vbzdec32( out, m, cpy,0) ,n); break;
case 8: TMBENCH("\nbitnpack16 ",l=bitnpack16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack16( out, n/2, cpy) ,n); break;
case 9: TMBENCH("\nbitnpack16v ",l=bitnpack128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack128v16( out, n/2, cpy) ,n); break;
case 8: TMBENCH("\nbitnpack32 ",l=bitnpack32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack32( out, m, cpy) ,n); break;
case 9: TMBENCH("\nbitnpack32v ",l=bitnpack128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack128v32( out, m, cpy) ,n); break;
case 10: TMBENCH("\nbitnzpack16 ",l=bitnzpack16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack16( out, n/2, cpy) ,n); break;
case 11: TMBENCH("\nbitnzpack16v",l=bitnzpack128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack128v16(out, n/2, cpy) ,n); break;
case 10: TMBENCH("\nbitnzpack32 ",l=bitnzpack32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack32( out, m, cpy) ,n); break;
case 11: TMBENCH("\nbitnzpack32v ",l=bitnzpack128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack128v32(out, m, cpy) ,n); break;
case 12: TMBENCH("\nbitndpack16 ",l=bitndpack16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack16( out, n/2, cpy) ,n); break;
case 13: TMBENCH("\nbitndpack16v",l=bitndpack128v16( in, n/2, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack128v16(out, n/2, cpy) ,n); break;
case ID_MEMCPY:TMBENCH("\nmemcpy ",libmemcpy( in,out,n) ,n); printf("%10u %5.1f%%", n, (double)100.0); return n;
case 12: TMBENCH("\nbitndpack32 ",l=bitndpack32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack32( out, m, cpy) ,n); break;
case 13: TMBENCH("\nbitndpack32v ",l=bitndpack128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack128v32(out, m, cpy) ,n); break;
case 14: TMBENCH("\nbitnd1pack32 ",l=bitnd1pack32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnd1unpack32( out, m, cpy) ,n); break;
case 15: TMBENCH("\nbitnd1pack32v",l=bitnd1pack128v32( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnd1unpack128v32(out, m, cpy) ,n); break;
case ID_MEMCPY: TMBENCH("\nmemcpy ",libmemcpy( in,out,n) ,n); printf("%10u %5.1f%%", n, (double)100.0); return n;
default: return l;
}
memcheck(in,(n/2)*2,cpy);
memcheck(in,n,cpy);
return l;
}
unsigned bench16(unsigned char *in, unsigned n, unsigned char *out, unsigned char *cpy, int id) {
unsigned l,m=(n+1)/2;
memrcpy(cpy,in,n);
switch(id) {
case 1: TMBENCH("\np4nenc16 ",l=p4nenc16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec16( out, m, cpy) ,n); break;
case 2: TMBENCH("\np4nenc16v ",l=p4nenc128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4ndec128v16( out, m, cpy) ,n); break;
case 3: TMBENCH("\np4nzenc16 ",l=p4nzenc16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec16( out, m, cpy) ,n); break;
case 4: TMBENCH("\np4nzenc16v ",l=p4nzenc128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nzdec128v16( out, m, cpy) ,n); break;
case 5: TMBENCH("\np4ndenc16 ",l=p4ndenc16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec16( out, m, cpy) ,n); break;
case 6: TMBENCH("\np4ndenc16v ",l=p4ndenc128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",p4nddec128v16( out, m, cpy) ,n); break;
case 7: TMBENCH("\nvbzenc16 ",l=vbzenc16( in, m, out,0)-out,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",vbzdec16( out, m, cpy,0) ,n); break;
case 8: TMBENCH("\nbitnpack16 ",l=bitnpack16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack16( out, m, cpy) ,n); break;
case 9: TMBENCH("\nbitnpack16v ",l=bitnpack128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnunpack128v16( out, m, cpy) ,n); break;
case 10: TMBENCH("\nbitnzpack16 ",l=bitnzpack16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack16( out, m, cpy) ,n); break;
case 11: TMBENCH("\nbitnzpack16v ",l=bitnzpack128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnzunpack128v16(out, m, cpy) ,n); break;
case 12: TMBENCH("\nbitndpack16 ",l=bitndpack16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack16( out, m, cpy) ,n); break;
case 13: TMBENCH("\nbitndpack16v ",l=bitndpack128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitndunpack128v16(out, m, cpy) ,n); break;
case 14: TMBENCH("\nbitnd1pack16 ",l=bitnd1pack16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnd1unpack16( out, m, cpy) ,n); break;
case 15: TMBENCH("\nbitnd1pack16v",l=bitnd1pack128v16( in, m, out) ,n); printf("%10u %5.1f%%", l, (double)l*100.0/n); TMBENCH("",bitnd1unpack128v16(out, m, cpy) ,n); break;
case ID_MEMCPY: TMBENCH("\nmemcpy ",libmemcpy( in,out,n) ,n); printf("%10u %5.1f%%", n, (double)100.0); return n;
default: return l;
}
memcheck(in,n,cpy);
return l;
}
int main(int argc, char* argv[]) {
unsigned trips = 4,cmp=1, b = 1 << 30, esize=4, lz=0, fno,id=0;
unsigned trips = 4,cmp=1, b = 1 << 31, esize=4, lz=0, fno,id=0, usize=4;
unsigned char *in,*out,*cpy;
int c, digit_optind = 0, this_option_optind = optind ? optind : 1, option_index = 0;
static struct option long_options[] = { {"blocsize", 0, 0, 'b'}, {0, 0, 0} };
for(;;) {
if((c = getopt_long(argc, argv, "B:ce:i:I:", long_options, &option_index)) == -1) break;
if((c = getopt_long(argc, argv, "B:ce:i:I:s:", long_options, &option_index)) == -1) break;
switch(c) {
case 0 : printf("Option %s", long_options[option_index].name); if(optarg) printf (" with arg %s", optarg); printf ("\n"); break;
case 'e': id = atoi(optarg); break;
case 'i': tm_rep = atoi(optarg); if(!tm_rep) tm_rep=1,trips=1; break;
case 'I': trips = atoi(optarg); if(!trips) trips=1; break;
case 's': usize = argtoi(optarg,1); break;
case 'B': b = argtoi(optarg,1); break;
case 'c': cmp++; break;
}
}
if(argc - optind < 1) { fprintf(stderr, "16 Bits file not specified\n"); exit(-1); }
if(argc - optind < 1) { fprintf(stderr, "raw file not specified\n"); exit(-1); }
for(fno = optind; fno < argc; fno++) {
char *inname = argv[fno];
@ -129,16 +167,26 @@ int main(int argc, char* argv[]) {
if(!(in = (unsigned char*)malloc(n+1024))) { fprintf(stderr, "malloc error\n"); exit(-1); } cpy = in;
if(!(out = (unsigned char*)malloc(flen*4/3+1024))) { fprintf(stderr, "malloc error\n"); exit(-1); }
if(cmp && !(cpy = (unsigned char*)malloc(n+1024))) { fprintf(stderr, "malloc error\n"); exit(-1); }
n = fread(in, 1, n, fi); printf("File='%s' Length=%u\n", inname, n);
n = fread(in, 1, n, fi); printf("File='%s' Length=%u Integer size=%d\n", inname, n, usize);
fclose(fi);
if(n <= 0) exit(0);
if(fno == optind)
tm_init(trips, 1);
if(!id) {
printf("function E MB/s size ratio D MB/s");
for(i=1; i <= ID_MEMCPY; i++) bench(in,n,out,cpy,i);
} else
bench(in,n,out,cpy,id);
for(i=1; i <= ID_MEMCPY; i++)
switch(usize) {
case 2: bench16(in,n,out,cpy,i); break;
case 4: bench32(in,n,out,cpy,i); break;
default: die("integer size must be 2 or 4\n");
}
} else {
switch(usize) {
case 2: bench16(in,n,out,cpy,id); break;
case 4: bench32(in,n,out,cpy,id); break;
default: die("integer size must be 2 or 4\n");
}
}
printf("\n");
}
}
@ -147,18 +195,22 @@ int main(int argc, char* argv[]) {
./icapp fc_300MHz.dat -I23
File='fc_300MHz.dat' Length=711595784
function E MB/s size ratio D MB/s
p4nenc16 627.18 441887368 62.1% 3284.07
p4nenc16v 684.00 441887368 62.1% 6193.39
p4nzenc16 507.39 332356287 46.7% 1405.11
p4nenc16 639.32 441887368 62.1% 3440.60
p4nenc16v 684.74 441887368 62.1% 6235.07
p4nzenc16 507.39 332356287 46.7% 1796.33
p4nzenc16v 548.99 332356287 46.7% 2651.81
p4ndenc16 450.75 525441761 73.8% 1375.51
p4ndenc16v 459.79 525441761 73.8% 2339.04
vbzenc16 2228.28 390678354 54.9% 2027.44
bitnpack16 4139.66 587990808 82.6% 5244.86
bitnpack16v 6659.51 587990808 82.6% 9207.07
bitnzpack16 1441.67 594311978 83.5% 2442.65
bitnzpack16v 2218.05 594311978 83.5% 4417.44
bitndpack16 2676.71 713562010 100.3% 4937.93
bitndpack16v 2994.07 713562010 100.3% 6393.84
memcpy 13334.75 711595784 100.0%
p4ndenc16v 459.79 525441761 73.8% 2340.81
vbzenc16 2228.28 390678354 54.9% 2028.28
bitnpack16 4139.86 587990808 82.6% 5244.86
bitnpack16v 6725.41 587990808 82.6% 9207.07
bitnzpack16 1439.90 594311978 83.5% 2442.53
bitnzpack16v 2219.25 594311978 83.5% 4417.38
bitndpack16 4637.59 713562010 100.3% 4942.12
bitndpack16v 6753.63 713562010 100.3% 6393.84
bitnd1pack16 2742.09 714375456 100.4% 4952.64
bitnd1pack16v 3600.12 714375456 100.4% 6301.49
memcpy 13429.38 711595784 100.0%
*/