static void find_best_gws(int do_benchmark, struct fmt_main *self)
{
int num;
cl_ulong run_time, min_time = CL_ULONG_MAX;
unsigned int SHAspeed, bestSHAspeed = 0, max_gws;
int optimal_gws = local_work_size;
const int sha1perkey = 50004;
unsigned long long int MaxRunTime = 5000000000ULL;
max_gws = get_max_mem_alloc_size(ocl_gpu_id) / (UNICODE_LENGTH * VF);
if (do_benchmark) {
fprintf(stderr, "Calculating best keys per crypt (GWS) for LWS=%zd and max. %llu s duration.\n\n", local_work_size, MaxRunTime / 1000000000UL);
fprintf(stderr, "Raw GPU speed figures including buffer transfers:\n");
}
for (num = local_work_size; max_gws; num *= 2) {
if (!do_benchmark)
advance_cursor();
if (!(run_time = gws_test(num, do_benchmark, self)))
break;
SHAspeed = sha1perkey * (1000000000UL * VF * num / run_time);
if (run_time < min_time)
min_time = run_time;
if (do_benchmark)
fprintf(stderr, "gws %6d%8llu c/s%14u sha1/s%8.3f sec per crypt_all()", num, (1000000000ULL * VF * num / run_time), SHAspeed, (float)run_time / 1000000000.);
if (((float)run_time / (float)min_time) < ((float)SHAspeed / (float)bestSHAspeed)) {
if (do_benchmark)
fprintf(stderr, "!\n");
bestSHAspeed = SHAspeed;
optimal_gws = num;
} else {
if (run_time < MaxRunTime && SHAspeed > (bestSHAspeed * 1.01)) {
if (do_benchmark)
fprintf(stderr, "+\n");
bestSHAspeed = SHAspeed;
optimal_gws = num;
continue;
}
if (do_benchmark)
fprintf(stderr, "\n");
if (run_time >= MaxRunTime)
break;
}
}
global_work_size = optimal_gws;
}
/* --
This function could be used to calculated the best num
of keys per crypt for the given format
-- */
static void find_best_gws(struct fmt_main * self) {
size_t num = 0;
cl_ulong run_time, min_time = CL_ULONG_MAX;
int optimal_gws = local_work_size, step = STEP;
int do_benchmark = 0;
unsigned int SHAspeed, bestSHAspeed = 0;
unsigned long long int max_run_time = 1000000000ULL;
char *tmp_value;
if ((tmp_value = getenv("STEP"))){
step = atoi(tmp_value);
do_benchmark = 1;
}
step = get_multiple(step, local_work_size);
if ((tmp_value = cfg_get_param(SECTION_OPTIONS, SUBSECTION_OPENCL, DUR_CONFIG)))
max_run_time = atoi(tmp_value) * 1000000000UL;
fprintf(stderr, "Calculating best global work size (GWS) for LWS=%zd and max. %llu s duration.\n\n",
local_work_size, max_run_time / 1000000000ULL);
if (do_benchmark)
fprintf(stderr, "Raw speed figures including buffer transfers:\n");
for (num = get_step(num, step, 1); num; num = get_step(num, step, 0)) {
//Check if hardware can handle the size we are going to try now.
if (sizeof(sha256_password) * num * 1.2 > get_max_mem_alloc_size(ocl_gpu_id))
break;
if (! (run_time = gws_test(num, self)))
continue;
if (!do_benchmark)
advance_cursor();
SHAspeed = num / (run_time / 1000000000.);
if (run_time < min_time)
min_time = run_time;
if (do_benchmark) {
fprintf(stderr, "gws: %8zu\t%12lu c/s %8.3f ms per crypt_all()",
num, (long) (num / (run_time / 1000000000.)),
(float) run_time / 1000000.);
if (run_time > max_run_time) {
fprintf(stderr, " - too slow\n");
break;
}
} else {
if (run_time > min_time * 20 || run_time > max_run_time)
break;
}
if (((long) SHAspeed - bestSHAspeed) > 10000) {
if (do_benchmark)
fprintf(stderr, "+");
bestSHAspeed = SHAspeed;
optimal_gws = num;
}
if (do_benchmark)
fprintf(stderr, "\n");
}
fprintf(stderr, "Optimal global work size %d\n", optimal_gws);
fprintf(stderr, "(to avoid this test on next run, put \""
GWS_CONFIG " = %d\" in john.conf, section [" SECTION_OPTIONS
SUBSECTION_OPENCL "])\n", optimal_gws);
global_work_size = optimal_gws;
create_clobj(optimal_gws, self);
}
static void find_best_gws(int do_benchmark)
{
int num;
cl_ulong run_time, min_time = CL_ULONG_MAX;
unsigned int SHAspeed, bestSHAspeed = 0;
int optimal_gws = local_work_size;
const int sha1perkey = (strlen(rar_fmt.params.tests[0].plaintext) * 2 + 8 + 3) * 0x40000 / 64 + 16;
char *conf;
unsigned long long int MaxRunTime = 5000000000ULL;
if ((conf = cfg_get_param(SECTION_OPTIONS, SUBSECTION_OPENCL, "rar_MaxDuration")))
MaxRunTime = atoi(conf) * 1000000000UL;
#ifndef DEBUG
if (do_benchmark)
#endif
{
fprintf(stderr, "Calculating best keys per crypt (GWS) for LWS=%zd and max. %llu s duration.\n\n", local_work_size, MaxRunTime / 1000000000UL);
fprintf(stderr, "Raw GPU speed figures including buffer transfers:\n");
}
for (num = local_work_size; num; num *= 2) {
if (!(run_time = gws_test(num)))
break;
SHAspeed = sha1perkey * (1000000000UL * num * VF / run_time);
if (run_time < min_time)
min_time = run_time;
#ifndef DEBUG
if (do_benchmark)
#endif
fprintf(stderr, "gws %6d\t%4llu c/s%14u sha1/s%8.3f sec per crypt_all()", num, (1000000000ULL * num * VF / run_time), SHAspeed, (float)run_time / 1000000000.);
if (((float)run_time / (float)min_time) < ((float)SHAspeed / (float)bestSHAspeed)) {
#ifndef DEBUG
if (do_benchmark)
#endif
fprintf(stderr, "!\n");
bestSHAspeed = SHAspeed;
optimal_gws = num;
} else {
if (run_time > MaxRunTime) {
#ifndef DEBUG
if (do_benchmark)
#endif
fprintf(stderr, "\n");
break;
}
if (SHAspeed > bestSHAspeed) {
#ifndef DEBUG
if (do_benchmark)
#endif
fprintf(stderr, "+");
bestSHAspeed = SHAspeed;
optimal_gws = num;
}
#ifndef DEBUG
if (do_benchmark)
#endif
fprintf(stderr, "\n");
}
}
if (get_device_type(ocl_gpu_id) != CL_DEVICE_TYPE_CPU) {
fprintf(stderr, "Optimal keys per crypt %d\n",(int)optimal_gws);
fprintf(stderr, "(to avoid this test on next run, put \""
GWS_CONFIG " = %d\" in john.conf, section ["
SECTION_OPTIONS
SUBSECTION_OPENCL "])\n", (int)optimal_gws);
}
*mkpc = VF * (global_work_size = optimal_gws);
}
