/** * app.c * UNI Host Application Source File * */ #include #include #include #include #include #include #include #include #include #include "../support/common.h" #include "../support/timer.h" #include "../support/params.h" // Define the DPU Binary path as DPU_BINARY here #ifndef DPU_BINARY #define DPU_BINARY "./bin/dpu_code" #endif #define XSTR(x) STR(x) #define STR(x) #x #if ENERGY #include #endif // Pointer declaration static T* A; static T* C; static T* C2; // Create input arrays static void read_input(T* A, unsigned int nr_elements, unsigned int nr_elements_round) { //srand(0); printf("nr_elements\t%u\t", nr_elements); for (unsigned int i = 0; i < nr_elements; i++) { //A[i] = (T) (rand()); A[i] = i%2==0?i:i+1; } for (unsigned int i = nr_elements; i < nr_elements_round; i++) { A[i] = A[nr_elements - 1]; } } // Compute output in the host static unsigned int unique_host(T* C, T* A, unsigned int nr_elements) { unsigned int pos = 0; C[pos] = A[pos]; pos++; for(unsigned int i = 1; i < nr_elements; i++) { if(A[i] != A[i-1]) { C[pos] = A[i]; pos++; } } return pos; } // Main of the Host Application int main(int argc, char **argv) { struct Params p = input_params(argc, argv); struct dpu_set_t dpu_set, dpu; uint32_t nr_of_dpus; #if ENERGY struct dpu_probe_t probe; DPU_ASSERT(dpu_probe_init("energy_probe", &probe)); #endif // Timer declaration Timer timer; // Allocate DPUs and load binary #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); timer.time[0] = 0; // alloc #endif #if !WITH_LOAD_OVERHEAD DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); assert(nr_of_dpus == NR_DPUS); timer.time[1] = 0; // load #endif #if !WITH_FREE_OVERHEAD timer.time[6] = 0; // free #endif unsigned int i = 0; uint32_t accum = 0; uint32_t total_count = 0; const unsigned int input_size = p.exp == 0 ? p.input_size * NR_DPUS : p.input_size; // Total input size (weak or strong scaling) const unsigned int input_size_dpu_ = divceil(input_size, NR_DPUS); // Input size per DPU (max.) const unsigned int input_size_dpu_round = (input_size_dpu_ % (NR_TASKLETS * REGS) != 0) ? roundup(input_size_dpu_, (NR_TASKLETS * REGS)) : input_size_dpu_; // Input size per DPU (max.), 8-byte aligned // Input/output allocation A = malloc(input_size_dpu_round * NR_DPUS * sizeof(T)); C = malloc(input_size_dpu_round * NR_DPUS * sizeof(T)); C2 = malloc(input_size_dpu_round * NR_DPUS * sizeof(T)); T *bufferA = A; T *bufferC = C2; // Create an input file with arbitrary data read_input(A, input_size, input_size_dpu_round * NR_DPUS); // Loop over main kernel for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) { #if WITH_ALLOC_OVERHEAD if(rep >= p.n_warmup) { start(&timer, 0, 0); } DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); if(rep >= p.n_warmup) { stop(&timer, 0); } #endif #if WITH_LOAD_OVERHEAD if(rep >= p.n_warmup) { start(&timer, 1, 0); } DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); if(rep >= p.n_warmup) { stop(&timer, 1); } DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); assert(nr_of_dpus == NR_DPUS); #endif // Compute output on CPU (performance comparison and verification purposes) if(rep >= p.n_warmup) { start(&timer, 2, 0); } total_count = unique_host(C, A, input_size); if(rep >= p.n_warmup) { stop(&timer, 2); } if(rep >= p.n_warmup) { start(&timer, 3, 0); } // Input arguments const unsigned int input_size_dpu = input_size_dpu_round; unsigned int kernel = 0; dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel}; // Copy input arrays i = 0; DPU_FOREACH(dpu_set, dpu, i) { DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments)); } DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments), DPU_XFER_DEFAULT)); DPU_FOREACH(dpu_set, dpu, i) { DPU_ASSERT(dpu_prepare_xfer(dpu, bufferA + input_size_dpu * i)); } DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT)); if(rep >= p.n_warmup) { stop(&timer, 3); } // Run DPU kernel if(rep >= p.n_warmup) { start(&timer, 4, 0); #if ENERGY DPU_ASSERT(dpu_probe_start(&probe)); #endif } DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); if(rep >= p.n_warmup) { stop(&timer, 4); #if ENERGY DPU_ASSERT(dpu_probe_stop(&probe)); #endif } #if PRINT { unsigned int each_dpu = 0; printf("Display DPU Logs\n"); DPU_FOREACH (dpu_set, dpu) { printf("DPU#%d:\n", each_dpu); DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); each_dpu++; } } #endif dpu_results_t results[nr_of_dpus]; uint32_t* results_scan = malloc(nr_of_dpus * sizeof(uint32_t)); uint32_t* offset = calloc(nr_of_dpus, sizeof(uint32_t)); uint32_t* offset_scan = calloc(nr_of_dpus, sizeof(uint32_t)); i = 0; accum = 0; if(rep >= p.n_warmup) { start(&timer, 5, 0); } // PARALLEL RETRIEVE TRANSFER dpu_results_t* results_retrieve[nr_of_dpus]; DPU_FOREACH(dpu_set, dpu, i) { results_retrieve[i] = (dpu_results_t*)malloc(NR_TASKLETS * sizeof(dpu_results_t)); DPU_ASSERT(dpu_prepare_xfer(dpu, results_retrieve[i])); } DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, "DPU_RESULTS", 0, NR_TASKLETS * sizeof(dpu_results_t), DPU_XFER_DEFAULT)); DPU_FOREACH(dpu_set, dpu, i) { // Retrieve tasklet timings for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { // First output element of this DPU if(each_tasklet == 0){ results[i].first = results_retrieve[i][each_tasklet].first; } // Last output element of this DPU and count if(each_tasklet == NR_TASKLETS - 1){ results[i].t_count = results_retrieve[i][each_tasklet].t_count; results[i].last = results_retrieve[i][each_tasklet].last; } } // Check if first(i) == last(i-1) -- offset if(i != 0){ if(results[i].first == results[i - 1].last) offset[i] = 1; // Sequential scan - offset offset_scan[i] += offset[i]; } // Sequential scan uint32_t temp = results[i].t_count - offset[i]; results_scan[i] = accum; accum += temp; #if PRINT printf("i=%d -- %u, %u, %u -- %u\n", i, results_scan[i], accum, temp, offset_scan[i]); #endif free(results_retrieve[i]); } if(rep >= p.n_warmup) { stop(&timer, 5); } i = 0; if(rep >= p.n_warmup) { start(&timer, 6, 0); } DPU_FOREACH (dpu_set, dpu) { // Copy output array DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferC + results_scan[i] - offset_scan[i], results[i].t_count * sizeof(T))); i++; } if(rep >= p.n_warmup) { stop(&timer, 6); } #if WITH_ALLOC_OVERHEAD #if WITH_FREE_OVERHEAD if(rep >= p.n_warmup) { start(&timer, 7, 0); } #endif DPU_ASSERT(dpu_free(dpu_set)); #if WITH_FREE_OVERHEAD if(rep >= p.n_warmup) { stop(&timer, 7); } #endif #endif // Free memory free(results_scan); free(offset); free(offset_scan); // Check output bool status = true; if(accum != total_count) status = false; #if PRINT printf("accum %u, total_count %u\n", accum, total_count); #endif for (i = 0; i < accum; i++) { if(C[i] != bufferC[i]){ status = false; #if PRINT printf("%d: %lu -- %lu\n", i, C[i], bufferC[i]); #endif } } if (status) { printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n"); if (rep >= p.n_warmup) { printf("[::] UNI UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d", NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size); printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ", WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD); printf("| latency_alloc_us=%f latency_load_us=%f latency_cpu_us=%f latency_write_us=%f latency_kernel_us=%f latency_sync_us=%f latency_read_us=%f latency_free_us=%f", timer.time[0], timer.time[1], timer.time[2], timer.time[3], timer.time[4], timer.time[5], timer.time[6], timer.time[7]); printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f", input_size * 3 * sizeof(T) / timer.time[2], input_size * 3 * sizeof(T) / (timer.time[4]), input_size * 3 * sizeof(T) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7])); printf(" throughput_upmem_wxsr_MBps=%f throughput_upmem_lwxsr_MBps=%f throughput_upmem_alwxsr_MBps=%f", input_size * 3 * sizeof(T) / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]), input_size * 3 * sizeof(T) / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]), input_size * 3 * sizeof(T) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6])); printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f", input_size / timer.time[2], input_size / (timer.time[4]), input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7])); printf(" throughput_upmem_wxsr_MOpps=%f throughput_upmem_lwxsr_MOpps=%f throughput_upmem_alwxsr_MOpps=%f\n", input_size / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]), input_size / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]), input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6])); printall(&timer, 4); } } else { printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n"); } } // Print timing results /* printf("CPU "); print(&timer, 0, p.n_reps); printf("CPU-DPU "); print(&timer, 1, p.n_reps); printf("DPU Kernel "); print(&timer, 2, p.n_reps); printf("Inter-DPU "); print(&timer, 3, p.n_reps); printf("DPU-CPU "); print(&timer, 4, p.n_reps); */ #if ENERGY double energy; DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy)); printf("DPU Energy (J): %f\t", energy); #endif // Deallocation free(A); free(C); free(C2); #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_free(dpu_set)); #endif return 0; }