int main(int argc, char **argv)
{
signal(SIGSEGV, SigHandler);
int w, i, verbose, single, region, tested, top;
int s_start, d_start, bytes, xor, alignment_test;
gf_t gf, gf_def;
time_t t0;
gf_internal_t *h;
gf_general_t *a, *b, *c, *d, *ai, *bi;
uint8_t a8, b8, c8, *mult4, *div4, *mult8, *div8;
uint16_t a16, b16, c16, d16, *log16, *alog16;
char as[50], bs[50], cs[50], ds[50], ais[50], bis[50];
uint32_t mask;
char *ra, *rb, *rc, *rd, *target;
int align;
if (argc < 4) usage(NULL);
if (sscanf(argv[1], "%d", &w) == 0){
usage("Bad w\n");
}
if (sscanf(argv[3], "%ld", &t0) == 0) usage("Bad seed\n");
if (t0 == -1) t0 = time(0);
MOA_Seed(t0);
if (w > 32 && w != 64 && w != 128) usage("Bad w");
if (create_gf_from_argv(&gf, w, argc, argv, 4) == 0) usage(BM);
printf("Size (bytes): %d\n", gf_size(&gf));
for (i = 0; i < strlen(argv[2]); i++) {
if (strchr("ASRV", argv[2][i]) == NULL) usage("Bad test\n");
}
h = (gf_internal_t *) gf.scratch;
a = (gf_general_t *) malloc(sizeof(gf_general_t));
b = (gf_general_t *) malloc(sizeof(gf_general_t));
c = (gf_general_t *) malloc(sizeof(gf_general_t));
d = (gf_general_t *) malloc(sizeof(gf_general_t));
ai = (gf_general_t *) malloc(sizeof(gf_general_t));
bi = (gf_general_t *) malloc(sizeof(gf_general_t));
//15 bytes extra to make sure it's 16byte aligned
ra = (char *) malloc(sizeof(char)*REGION_SIZE+15);
rb = (char *) malloc(sizeof(char)*REGION_SIZE+15);
rc = (char *) malloc(sizeof(char)*REGION_SIZE+15);
rd = (char *) malloc(sizeof(char)*REGION_SIZE+15);
//this still assumes 8 byte aligned pointer from malloc
//(which is usual on 32-bit machines)
ra += (uint64_t)ra & 0xf;
rb += (uint64_t)rb & 0xf;
rc += (uint64_t)rc & 0xf;
rd += (uint64_t)rd & 0xf;
if (w <= 32) {
mask = 0;
for (i = 0; i < w; i++) mask |= (1 << i);
}
verbose = (strchr(argv[2], 'V') != NULL);
single = (strchr(argv[2], 'S') != NULL || strchr(argv[2], 'A') != NULL);
region = (strchr(argv[2], 'R') != NULL || strchr(argv[2], 'A') != NULL);
if (!gf_init_hard(&gf_def, w, GF_MULT_DEFAULT, GF_REGION_DEFAULT, GF_DIVIDE_DEFAULT,
(h->mult_type != GF_MULT_COMPOSITE) ? h->prim_poly : 0, 0, 0, NULL, NULL))
problem("No default for this value of w");
if (w == 4) {
mult4 = gf_w4_get_mult_table(&gf);
div4 = gf_w4_get_div_table(&gf);
}
if (w == 8) {
mult8 = gf_w8_get_mult_table(&gf);
div8 = gf_w8_get_div_table(&gf);
}
if (w == 16) {
log16 = gf_w16_get_log_table(&gf);
alog16 = gf_w16_get_mult_alog_table(&gf);
}
if (verbose) printf("Seed: %ld\n", t0);
if (single) {
if (gf.multiply.w32 == NULL) problem("No multiplication operation defined.");
if (verbose) { printf("Testing single multiplications/divisions.\n"); fflush(stdout); }
if (w <= 10) {
top = (1 << w)*(1 << w);
} else {
top = 1024*1024;
}
for (i = 0; i < top; i++) {
if (w <= 10) {
a->w32 = i % (1 << w);
b->w32 = (i >> w);
//Allen: the following conditions were being run 10 times each. That didn't seem like nearly enough to
//me for these special cases, so I converted to doing this mod stuff to easily make the number of times
//run both larger and proportional to the total size of the run.
} else {
switch (i % 32)
{
case 0:
gf_general_set_zero(a, w);
gf_general_set_random(b, w, 1);
break;
case 1:
gf_general_set_random(a, w, 1);
gf_general_set_zero(b, w);
break;
case 2:
gf_general_set_one(a, w);
gf_general_set_random(b, w, 1);
break;
case 3:
gf_general_set_random(a, w, 1);
gf_general_set_one(b, w);
break;
default:
gf_general_set_random(a, w, 1);
gf_general_set_random(b, w, 1);
}
}
//Allen: the following special cases for w=64 are based on the code below for w=128.
//These w=64 cases are based on Dr. Plank's suggestion because some of the methods for w=64
//involve splitting it in two. I think they're less likely to give errors than the 128-bit case
//though, because the 128 bit case is always split in two.
//As with w=128, I'm arbitrarily deciding to do this sort of thing with a quarter of the cases
if (w == 64) {
switch (i % 32)
{
case 0: if (!gf_general_is_one(a, w)) a->w64 &= RMASK; break;
case 1: if (!gf_general_is_one(a, w)) a->w64 &= LMASK; break;
case 2: if (!gf_general_is_one(a, w)) a->w64 &= RMASK; if (!gf_general_is_one(b, w)) b->w64 &= RMASK; break;
case 3: if (!gf_general_is_one(a, w)) a->w64 &= RMASK; if (!gf_general_is_one(b, w)) b->w64 &= LMASK; break;
case 4: if (!gf_general_is_one(a, w)) a->w64 &= LMASK; if (!gf_general_is_one(b, w)) b->w64 &= RMASK; break;
case 5: if (!gf_general_is_one(a, w)) a->w64 &= LMASK; if (!gf_general_is_one(b, w)) b->w64 &= LMASK; break;
case 6: if (!gf_general_is_one(b, w)) b->w64 &= RMASK; break;
case 7: if (!gf_general_is_one(b, w)) b->w64 &= LMASK; break;
}
}
//Allen: for w=128, we have important special cases where one half or the other of the number is all
//zeros. The probability of hitting such a number randomly is 1^-64, so if we don't force these cases
//we'll probably never hit them. This could be implemented more efficiently by changing the set-random
//function for w=128, but I think this is easier to follow.
//I'm arbitrarily deciding to do this sort of thing with a quarter of the cases
if (w == 128) {
switch (i % 32)
{
case 0: if (!gf_general_is_one(a, w)) a->w128[0] = 0; break;
case 1: if (!gf_general_is_one(a, w)) a->w128[1] = 0; break;
case 2: if (!gf_general_is_one(a, w)) a->w128[0] = 0; if (!gf_general_is_one(b, w)) b->w128[0] = 0; break;
case 3: if (!gf_general_is_one(a, w)) a->w128[0] = 0; if (!gf_general_is_one(b, w)) b->w128[1] = 0; break;
case 4: if (!gf_general_is_one(a, w)) a->w128[1] = 0; if (!gf_general_is_one(b, w)) b->w128[0] = 0; break;
case 5: if (!gf_general_is_one(a, w)) a->w128[1] = 0; if (!gf_general_is_one(b, w)) b->w128[1] = 0; break;
case 6: if (!gf_general_is_one(b, w)) b->w128[0] = 0; break;
case 7: if (!gf_general_is_one(b, w)) b->w128[1] = 0; break;
}
}
tested = 0;
gf_general_multiply(&gf, a, b, c);
/* If w is 4, 8 or 16, then there are inline multiplication/division methods.
Test them here. */
if (w == 4 && mult4 != NULL) {
a8 = a->w32;
b8 = b->w32;
c8 = GF_W4_INLINE_MULTDIV(mult4, a8, b8);
if (c8 != c->w32) {
printf("Error in inline multiplication. %d * %d. Inline = %d. Default = %d.\n",
a8, b8, c8, c->w32);
exit(1);
}
}
if (w == 8 && mult8 != NULL) {
a8 = a->w32;
b8 = b->w32;
c8 = GF_W8_INLINE_MULTDIV(mult8, a8, b8);
if (c8 != c->w32) {
printf("Error in inline multiplication. %d * %d. Inline = %d. Default = %d.\n",
a8, b8, c8, c->w32);
exit(1);
}
}
if (w == 16 && log16 != NULL) {
a16 = a->w32;
b16 = b->w32;
c16 = GF_W16_INLINE_MULT(log16, alog16, a16, b16);
if (c16 != c->w32) {
printf("Error in inline multiplication. %d * %d. Inline = %d. Default = %d.\n",
a16, b16, c16, c->w32);
printf("%d %d\n", log16[a16], log16[b16]);
top = log16[a16] + log16[b16];
printf("%d %d\n", top, alog16[top]);
exit(1);
}
}
/* If this is not composite, then first test against the default: */
if (h->mult_type != GF_MULT_COMPOSITE) {
tested = 1;
gf_general_multiply(&gf_def, a, b, d);
if (!gf_general_are_equal(c, d, w)) {
gf_general_val_to_s(a, w, as, 1);
gf_general_val_to_s(b, w, bs, 1);
gf_general_val_to_s(c, w, cs, 1);
gf_general_val_to_s(d, w, ds, 1);
printf("Error in single multiplication (all numbers in hex):\n\n");
printf(" gf.multiply(gf, %s, %s) = %s\n", as, bs, cs);
printf(" The default gf multiplier returned %s\n", ds);
exit(1);
}
}
/* Now, we also need to double-check by other means, in case the default is wanky,
and when we're performing composite operations. Start with 0 and 1, where we know
what the result should be. */
if (gf_general_is_zero(a, w) || gf_general_is_zero(b, w) ||
gf_general_is_one(a, w) || gf_general_is_one(b, w)) {
tested = 1;
if (((gf_general_is_zero(a, w) || gf_general_is_zero(b, w)) && !gf_general_is_zero(c, w)) ||
(gf_general_is_one(a, w) && !gf_general_are_equal(b, c, w)) ||
(gf_general_is_one(b, w) && !gf_general_are_equal(a, c, w))) {
gf_general_val_to_s(a, w, as, 1);
gf_general_val_to_s(b, w, bs, 1);
gf_general_val_to_s(c, w, cs, 1);
printf("Error in single multiplication (all numbers in hex):\n\n");
printf(" gf.multiply(gf, %s, %s) = %s, which is clearly wrong.\n", as, bs, cs);
;
exit(1);
}
}
/* Dumb check to make sure that it's not returning numbers that are too big: */
if (w < 32 && (c->w32 & mask) != c->w32) {
gf_general_val_to_s(a, w, as, 1);
gf_general_val_to_s(b, w, bs, 1);
gf_general_val_to_s(c, w, cs, 1);
printf("Error in single multiplication (all numbers in hex):\n\n");
printf(" gf.multiply.w32(gf, %s, %s) = %s, which is too big.\n", as, bs, cs);
exit(1);
}
/* Finally, let's check to see that multiplication and division work together */
if (!gf_general_is_zero(a, w)) {
gf_general_divide(&gf, c, a, d);
if (!gf_general_are_equal(b, d, w)) {
gf_general_val_to_s(a, w, as, 1);
gf_general_val_to_s(b, w, bs, 1);
gf_general_val_to_s(c, w, cs, 1);
gf_general_val_to_s(d, w, ds, 1);
printf("Error in single multiplication/division (all numbers in hex):\n\n");
printf(" gf.multiply(gf, %s, %s) = %s, but gf.divide(gf, %s, %s) = %s\n", as, bs, cs, cs, as, ds);
exit(1);
}
}
}
void gf_general_set_up_single_timing_test(int w, void *ra, void *rb, int size)
{
void *top;
gf_general_t g;
uint8_t *r8, *r8a;
uint16_t *r16;
uint32_t *r32;
uint64_t *r64;
int i;
top = rb+size;
/* If w is 8, 16, 32, 64 or 128, fill the regions with random bytes.
However, don't allow for zeros in rb, because that will screw up
division.
When w is 4, you fill the regions with random 4-bit words in each byte.
Otherwise, treat every four bytes as an uint32_t
and fill it with a random value mod (1 << w).
*/
if (w == 8 || w == 16 || w == 32 || w == 64 || w == 128) {
MOA_Fill_Random_Region (ra, size);
while (rb < top) {
gf_general_set_random(&g, w, 0);
switch (w) {
case 8:
r8 = (uint8_t *) rb;
*r8 = g.w32;
break;
case 16:
r16 = (uint16_t *) rb;
*r16 = g.w32;
break;
case 32:
r32 = (uint32_t *) rb;
*r32 = g.w32;
break;
case 64:
r64 = (uint64_t *) rb;
*r64 = g.w64;
break;
case 128:
r64 = (uint64_t *) rb;
r64[0] = g.w128[0];
r64[1] = g.w128[1];
break;
}
rb += (w/8);
}
} else if (w == 4) {
r8a = (uint8_t *) ra;
r8 = (uint8_t *) rb;
while (r8 < (uint8_t *) top) {
gf_general_set_random(&g, w, 1);
*r8a = g.w32;
gf_general_set_random(&g, w, 0);
*r8 = g.w32;
r8a++;
r8++;
}
} else {
r32 = (uint32_t *) ra;
for (i = 0; i < size/4; i++) r32[i] = MOA_Random_W(w, 1);
r32 = (uint32_t *) rb;
for (i = 0; i < size/4; i++) r32[i] = MOA_Random_W(w, 0);
}
}
int main(int argc, char **argv)
{
int w, it, i, size, iterations, xor;
char tests[100];
char test;
char *single_tests = "MDI";
char *region_tests = "G012";
char *tstrings[256];
void *tmethods[256];
gf_t gf;
double timer, elapsed, ds, di, dnum;
int num;
time_t t0;
uint8_t *ra, *rb;
gf_general_t a;
if (argc < 6) usage(NULL);
if (sscanf(argv[1], "%d", &w) == 0){
usage("Bad w[-pp]\n");
}
if (sscanf(argv[3], "%ld", &t0) == 0) usage("Bad seed\n");
if (sscanf(argv[4], "%d", &size) == 0) usage("Bad size\n");
if (sscanf(argv[5], "%d", &iterations) == 0) usage("Bad iterations\n");
if (t0 == -1) t0 = time(0);
MOA_Seed(t0);
ds = size;
di = iterations;
if ((w > 32 && w != 64 && w != 128) || w < 0) usage("Bad w");
if ((size * 8) % w != 0) usage ("Bad size -- must be a multiple of w*8\n");
if (!create_gf_from_argv(&gf, w, argc, argv, 6)) usage(BM);
strcpy(tests, "");
for (i = 0; argv[2][i] != '\0'; i++) {
switch(argv[2][i]) {
case 'A': strcat(tests, single_tests);
strcat(tests, region_tests);
break;
case 'S': strcat(tests, single_tests); break;
case 'R': strcat(tests, region_tests); break;
case 'G': strcat(tests, "G"); break;
case '0': strcat(tests, "0"); break;
case '1': strcat(tests, "1"); break;
case '2': strcat(tests, "2"); break;
case 'M': strcat(tests, "M"); break;
case 'D': strcat(tests, "D"); break;
case 'I': strcat(tests, "I"); break;
default: usage("Bad tests");
}
}
tstrings['M'] = "Multiply";
tstrings['D'] = "Divide";
tstrings['I'] = "Inverse";
tstrings['G'] = "Region-Random";
tstrings['0'] = "Region-By-Zero";
tstrings['1'] = "Region-By-One";
tstrings['2'] = "Region-By-Two";
tmethods['M'] = (void *) gf.multiply.w32;
tmethods['D'] = (void *) gf.divide.w32;
tmethods['I'] = (void *) gf.inverse.w32;
tmethods['G'] = (void *) gf.multiply_region.w32;
tmethods['0'] = (void *) gf.multiply_region.w32;
tmethods['1'] = (void *) gf.multiply_region.w32;
tmethods['2'] = (void *) gf.multiply_region.w32;
printf("Seed: %ld\n", t0);
ra = (uint8_t *) malloc(size);
rb = (uint8_t *) malloc(size);
if (ra == NULL || rb == NULL) { perror("malloc"); exit(1); }
for (i = 0; i < 3; i++) {
test = single_tests[i];
if (strchr(tests, test) != NULL) {
if (tmethods[(int)test] == NULL) {
printf("No %s method.\n", tstrings[(int)test]);
} else {
elapsed = 0;
dnum = 0;
for (it = 0; it < iterations; it++) {
gf_general_set_up_single_timing_test(w, ra, rb, size);
timer_start(&timer);
num = gf_general_do_single_timing_test(&gf, ra, rb, size, test);
dnum += num;
elapsed += timer_split(&timer);
}
printf("%14s: %10.6lf s Mops: %10.3lf %10.3lf Mega-ops/s\n",
tstrings[(int)test], elapsed,
dnum/1024.0/1024.0, dnum/1024.0/1024.0/elapsed);
}
}
}
for (i = 0; i < 4; i++) {
test = region_tests[i];
if (strchr(tests, test) != NULL) {
if (tmethods[(int)test] == NULL) {
printf("No %s method.\n", tstrings[(int)test]);
} else {
if (test == '0') gf_general_set_zero(&a, w);
if (test == '1') gf_general_set_one(&a, w);
if (test == '2') gf_general_set_two(&a, w);
for (xor = 0; xor < 2; xor++) {
elapsed = 0;
for (it = 0; it < iterations; it++) {
if (test == 'G') gf_general_set_random(&a, w, 1);
gf_general_set_up_single_timing_test(8, ra, rb, size);
timer_start(&timer);
gf_general_do_region_multiply(&gf, &a, ra, rb, size, xor);
elapsed += timer_split(&timer);
}
printf("%14s: XOR: %d %10.6lf s MB: %10.3lf %10.3lf MB/s\n",
tstrings[(int)test], xor, elapsed,
ds*di/1024.0/1024.0, ds*di/1024.0/1024.0/elapsed);
}
}
}
}
return 0;
}
