diff options
| author | Daniel Friesel <daniel.friesel@uos.de> | 2023-06-02 15:27:22 +0200 |
|---|---|---|
| committer | Daniel Friesel <daniel.friesel@uos.de> | 2023-06-02 15:27:22 +0200 |
| commit | 7e0597a2ddb2d07ff7b72a6fa47245774e04f964 (patch) | |
| tree | 51dbd792f11428c5f61283fb5de3fc7d4f9606f7 /HST-S/host/app.c | |
| parent | f438d54780af4ae90638f5d57acdd5e98d6596e9 (diff) | |
port HST-S to dfatool
Diffstat (limited to 'HST-S/host/app.c')
| -rw-r--r-- | HST-S/host/app.c | 27 |
1 files changed, 22 insertions, 5 deletions
diff --git a/HST-S/host/app.c b/HST-S/host/app.c index e50fd62..5dafa3a 100644 --- a/HST-S/host/app.c +++ b/HST-S/host/app.c @@ -23,6 +23,9 @@ #define DPU_BINARY "./bin/dpu_code" #endif +#define XSTR(x) STR(x) +#define STR(x) #x + #if ENERGY #include <dpu_probe.h> #endif @@ -40,7 +43,7 @@ static void read_input(T* A, const Params p) { // Open input file unsigned short temp; - sprintf(dctFileName, p.file_name); + sprintf(dctFileName, "%s", p.file_name); if((File = fopen(dctFileName, "rb")) != NULL) { for(unsigned int y = 0; y < p.input_size; y++) { fread(&temp, sizeof(unsigned short), 1, File); @@ -138,14 +141,14 @@ int main(int argc, char **argv) { // Compute output on CPU (performance comparison and verification purposes) if(rep >= p.n_warmup) - start(&timer, 0, rep - p.n_warmup); + start(&timer, 0, 0); histogram_host(histo_host, A, p.bins, p.input_size, 1, nr_of_dpus); if(rep >= p.n_warmup) stop(&timer, 0); printf("Load input data\n"); if(rep >= p.n_warmup) - start(&timer, 1, rep - p.n_warmup); + start(&timer, 1, 0); // Input arguments unsigned int kernel = 0; i = 0; @@ -177,7 +180,7 @@ int main(int argc, char **argv) { printf("Run program on DPU(s) \n"); // Run DPU kernel if(rep >= p.n_warmup) { - start(&timer, 2, rep - p.n_warmup); + start(&timer, 2, 0); #if ENERGY DPU_ASSERT(dpu_probe_start(&probe)); #endif @@ -206,7 +209,7 @@ int main(int argc, char **argv) { printf("Retrieve results\n"); i = 0; if(rep >= p.n_warmup) - start(&timer, 3, rep - p.n_warmup); + start(&timer, 3, 0); // PARALLEL RETRIEVE TRANSFER DPU_FOREACH(dpu_set, dpu, i) { DPU_ASSERT(dpu_prepare_xfer(dpu, histo + p.bins * i)); @@ -222,6 +225,20 @@ int main(int argc, char **argv) { if(rep >= p.n_warmup) stop(&timer, 3); + if (rep >= p.n_warmup) { + printf("[::] HST-S NMC | n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d n_bins=%d " + "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f", + nr_of_dpus, NR_TASKLETS, XSTR(T), input_size, p.bins, + input_size * sizeof(T) / timer.time[0], + input_size * sizeof(T) / timer.time[2], + input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3])); + printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f", + input_size / timer.time[0], + input_size / timer.time[2], + input_size / (timer.time[1] + timer.time[2] + timer.time[3])); + printall(&timer, 3); + } + } // Print timing results |