// Find the best block size parameters for the Dslash and DslashXpay kernels
void DiracClover::Tune(cudaColorSpinorField &out, const cudaColorSpinorField &in,
const cudaColorSpinorField &x) {
DiracWilson::Tune(out, in, x);
setDslashTuning(QUDA_TUNE_YES);
{ // Tune clover application
TuneDiracClover cloverTune(*this, out, in);
cloverTune.Benchmark(tuneClover);
}
setDslashTuning(QUDA_TUNE_NO);
}
static int dslashTest()
{
int accuracy_level = 0;
init();
int attempts = 1;
for (int i=0; i<attempts; i++) {
if (tune) { // warm-up run
printfQuda("Tuning...\n");
setDslashTuning(QUDA_TUNE_YES, QUDA_VERBOSE);
dslashCUDA(1);
}
printfQuda("Executing %d kernel loops...", loops);
double secs = dslashCUDA(loops);
#ifdef DSLASH_PROFILING
printDslashProfile();
#endif
if (!transfer) *spinorOut = *cudaSpinorOut;
printfQuda("\n%fms per loop\n", 1000*secs);
staggeredDslashRef();
unsigned long long flops = dirac->Flops();
int link_floats = 8*gaugeParam.reconstruct+8*18;
int spinor_floats = 8*6*2 + 6;
int link_float_size = prec;
int spinor_float_size = 0;
link_floats = test_type ? (2*link_floats) : link_floats;
spinor_floats = test_type ? (2*spinor_floats) : spinor_floats;
int bytes_for_one_site = link_floats * link_float_size + spinor_floats * spinor_float_size;
if (prec == QUDA_HALF_PRECISION) bytes_for_one_site += (8*2 + 1)*4;
printfQuda("GFLOPS = %f\n", 1.0e-9*flops/secs);
printfQuda("GB/s = %f\n\n", 1.0*Vh*bytes_for_one_site/((secs/loops)*1e+9));
if (!transfer) {
double spinor_ref_norm2 = norm2(*spinorRef);
double cuda_spinor_out_norm2 = norm2(*cudaSpinorOut);
double spinor_out_norm2 = norm2(*spinorOut);
printfQuda("Results: CPU=%f, CUDA=%f, CPU-CUDA=%f\n", spinor_ref_norm2, cuda_spinor_out_norm2,
spinor_out_norm2);
} else {
double spinor_ref_norm2 = norm2(*spinorRef);
double spinor_out_norm2 = norm2(*spinorOut);
printfQuda("Result: CPU=%f , CPU-CUDA=%f", spinor_ref_norm2, spinor_out_norm2);
}
accuracy_level = cpuColorSpinorField::Compare(*spinorRef, *spinorOut);
}
end();
return accuracy_level;
}
// Find the best block size parameters for the Dslash and DslashXpay kernels
void DiracDomainWall::Tune(cudaColorSpinorField &out, const cudaColorSpinorField &in,
const cudaColorSpinorField &x) {
setDslashTuning(QUDA_TUNE_YES);
{ // Tune Dslash
TuneDiracDomainWallDslash dslashTune(*this, out, in);
dslashTune.Benchmark(tuneDslash[0]);
for (int i=0; i<4; i++)
if (commDimPartitioned(i))
dslashTune.Benchmark(tuneDslash[i+1]);
}
{ // Tune DslashXpay
TuneDiracDomainWallDslashXpay dslashXpayTune(*this, out, in, x);
dslashXpayTune.Benchmark(tuneDslashXpay[0]);
for (int i=0; i<4; i++)
if (commDimPartitioned(i))
dslashXpayTune.Benchmark(tuneDslashXpay[i+1]);
}
setDslashTuning(QUDA_TUNE_NO);
}
int main(int argc, char **argv)
{
init();
float spinorGiB = (float)Vh*Ls*spinorSiteSize*sizeof(inv_param.cpu_prec) / (1 << 30);
printf("\nSpinor mem: %.3f GiB\n", spinorGiB);
printf("Gauge mem: %.3f GiB\n", gauge_param.gaugeGiB);
int attempts = 1;
dslashRef();
for (int i=0; i<attempts; i++) {
if (tune) { // warm-up run
printfQuda("Tuning...\n");
setDslashTuning(QUDA_TUNE_YES, QUDA_VERBOSE);
dslashCUDA(1);
}
double secs = dslashCUDA();
if (!transfer) *spinorOut = *cudaSpinorOut;
// print timing information
printf("%fms per loop\n", 1000*secs);
unsigned long long flops = 0;
if (!transfer) flops = dirac->Flops();
int spinor_floats = test_type ? 2*(9*24+24)+24 : 9*24+24;
if (inv_param.cuda_prec == QUDA_HALF_PRECISION)
spinor_floats += test_type ? 2*(9*2 + 2) + 2 : 9*2 + 2; // relative size of norm is twice a short
int gauge_floats = (test_type ? 2 : 1) * (gauge_param.gauge_fix ? 6 : 8) * gauge_param.reconstruct;
printfQuda("GFLOPS = %f\n", 1.0e-9*flops/secs);
printfQuda("GB/s = %f\n\n",
(float)Vh*Ls*(spinor_floats+gauge_floats)*inv_param.cuda_prec/((secs/loops)*1e+9));
if (!transfer) {
std::cout << "Results: CPU = " << norm2(*spinorRef) << ", CUDA = " << norm2(*cudaSpinorOut) <<
", CPU-CUDA = " << norm2(*spinorOut) << std::endl;
} else {
std::cout << "Result: CPU = " << norm2(*spinorRef) << ", CPU-CUDA = " << norm2(*spinorOut) << std::endl;
}
cpuColorSpinorField::Compare(*spinorRef, *spinorOut);
}
end();
}
// Find the best block size parameters for the Dslash and DslashXpay kernels
void DiracCloverPC::Tune(cudaColorSpinorField &out, const cudaColorSpinorField &in,
const cudaColorSpinorField &x) {
DiracClover::Tune(out, in, x);
setDslashTuning(QUDA_TUNE_YES);
{ // Tune Dslash
TuneDiracCloverDslash dslashTune(*this, out, in);
dslashTune.Benchmark(blockDslash);
#ifdef OVERLAP_COMMS
dslashTune.Benchmark(blockDslashFace);
#endif
}
{ // Tune DslashXpay
TuneDiracCloverDslashXpay dslashXpayTune(*this, out, in, x);
dslashXpayTune.Benchmark(blockDslashXpay);
#ifdef OVERLAP_COMMS
dslashXpayTune.Benchmark(blockDslashXpayFace);
#endif
}
setDslashTuning(QUDA_TUNE_NO);
}
int main(int argc, char **argv)
{
for (int i =1;i < argc; i++){
if(process_command_line_option(argc, argv, &i) == 0){
continue;
}
fprintf(stderr, "ERROR: Invalid option:%s\n", argv[i]);
usage(argv);
}
initCommsQuda(argc, argv, gridsize_from_cmdline, 4);
display_test_info();
init(argc, argv);
float spinorGiB = (float)Vh*spinorSiteSize*inv_param.cuda_prec / (1 << 30);
printfQuda("\nSpinor mem: %.3f GiB\n", spinorGiB);
printfQuda("Gauge mem: %.3f GiB\n", gauge_param.gaugeGiB);
int attempts = 1;
dslashRef();
for (int i=0; i<attempts; i++) {
if (tune) { // warm-up run
printfQuda("Tuning...\n");
setDslashTuning(QUDA_TUNE_YES, QUDA_VERBOSE);
dslashCUDA(1);
}
printfQuda("Executing %d kernel loops...\n", loops);
dirac->Flops();
double secs = dslashCUDA(loops);
printfQuda("done.\n\n");
#ifdef DSLASH_PROFILING
printDslashProfile();
#endif
if (!transfer) *spinorOut = *cudaSpinorOut;
// print timing information
printfQuda("%fms per loop\n", 1000*secs);
unsigned long long flops = 0;
if (!transfer) flops = dirac->Flops();
int spinor_floats = test_type ? 2*(7*24+24)+24 : 7*24+24;
if (inv_param.cuda_prec == QUDA_HALF_PRECISION)
spinor_floats += test_type ? 2*(7*2 + 2) + 2 : 7*2 + 2; // relative size of norm is twice a short
int gauge_floats = (test_type ? 2 : 1) * (gauge_param.gauge_fix ? 6 : 8) * gauge_param.reconstruct;
if (dslash_type == QUDA_CLOVER_WILSON_DSLASH) {
gauge_floats += test_type ? 72*2 : 72;
}
printfQuda("GFLOPS = %f\n", 1.0e-9*flops/secs);
printfQuda("GB/s = %f\n\n",
Vh*(spinor_floats+gauge_floats)*inv_param.cuda_prec/((secs/loops)*1e+9));
if (!transfer) {
double norm2_cpu = norm2(*spinorRef);
double norm2_cuda= norm2(*cudaSpinorOut);
double norm2_cpu_cuda= norm2(*spinorOut);
printfQuda("Results: CPU = %f, CUDA=%f, CPU-CUDA = %f\n", norm2_cpu, norm2_cuda, norm2_cpu_cuda);
} else {
double norm2_cpu = norm2(*spinorRef);
double norm2_cpu_cuda= norm2(*spinorOut);
printfQuda("Result: CPU = %f, CPU-QUDA = %f\n", norm2_cpu, norm2_cpu_cuda);
}
cpuColorSpinorField::Compare(*spinorRef, *spinorOut);
}
end();
endCommsQuda();
}
