summaryrefslogtreecommitdiff
path: root/TRNS/host
diff options
context:
space:
mode:
Diffstat (limited to 'TRNS/host')
-rw-r--r--TRNS/host/app.c62
1 files changed, 38 insertions, 24 deletions
diff --git a/TRNS/host/app.c b/TRNS/host/app.c
index 452b894..c178a19 100644
--- a/TRNS/host/app.c
+++ b/TRNS/host/app.c
@@ -7,16 +7,32 @@
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
+
+#if ASPECTC
+extern "C" {
+#endif
+
#include <dpu.h>
#include <dpu_log.h>
+#include <dpu_management.h>
+#include <dpu_target_macros.h>
+
+#if ENERGY
+#include <dpu_probe.h>
+#endif
+
+#if ASPECTC
+}
+#endif
+
#include <unistd.h>
#include <getopt.h>
#include <assert.h>
#include <math.h>
-#include "../support/common.h"
-#include "../support/timer.h"
-#include "../support/params.h"
+#include "common.h"
+#include "timer.h"
+#include "params.h"
#define XSTR(x) STR(x)
#define STR(x) #x
@@ -26,18 +42,13 @@
#define DPU_BINARY "./bin/dpu_code"
#endif
-#if ENERGY
-#include <dpu_probe.h>
-#endif
-
-#include <dpu_management.h>
-#include <dpu_target_macros.h>
-
// Pointer declaration
static T* A_host;
static T* A_backup;
static T* A_result;
+unsigned int kernel = 0;
+
// Create input arrays
static void read_input(T* A, unsigned int nr_elements) {
srand(0);
@@ -84,10 +95,10 @@ int main(int argc, char **argv) {
N_ = p.exp == 0 ? N_ * NR_DPUS : N_;
// Input/output allocation
- A_host = malloc(M_ * m * N_ * n * sizeof(T));
- A_backup = malloc(M_ * m * N_ * n * sizeof(T));
- A_result = malloc(M_ * m * N_ * n * sizeof(T));
- T* done_host = malloc(M_ * n); // Host array to reset done array of step 3
+ A_host = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ A_backup = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ A_result = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ T* done_host = (T*)malloc(M_ * n); // Host array to reset done array of step 3
memset(done_host, 0, M_ * n);
// Create an input file with arbitrary data
@@ -131,6 +142,7 @@ int main(int argc, char **argv) {
DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
stop(&timer, 2);
DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
+ DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
} else if (first_round){
start(&timer, 1, 0);
DPU_ASSERT(dpu_alloc(active_dpus, NULL, &dpu_set));
@@ -174,8 +186,8 @@ int main(int argc, char **argv) {
start(&timer, 5, !first_round);
}
- unsigned int kernel = 0;
- dpu_arguments_t input_arguments = {m, n, M_, kernel};
+ kernel = 0;
+ dpu_arguments_t input_arguments = {m, n, M_, (enum kernels)kernel};
// transfer control instructions to DPUs (run first program part)
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments));
@@ -215,7 +227,7 @@ int main(int argc, char **argv) {
start(&timer, 7, !first_round);
}
kernel = 1;
- dpu_arguments_t input_arguments2 = {m, n, M_, kernel};
+ dpu_arguments_t input_arguments2 = {m, n, M_, (enum kernels)kernel};
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments2));
}
@@ -299,7 +311,9 @@ int main(int argc, char **argv) {
}
if (status) {
printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+#if DFATOOL_TIMING
unsigned long input_size = M_ * m * N_ * n;
+#endif
if (rep >= p.n_warmup) {
/*
* timer 0: CPU version
@@ -313,35 +327,35 @@ int main(int argc, char **argv) {
* timer 8: run DPU program (second kernel)
* timer 9: read transposed matrix
*/
- printf("[::] TRNS-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%lu numa_node_rank=%d ",
+ dfatool_printf("[::] TRNS-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%lu numa_node_rank=%d ",
NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), input_size, numa_node_rank);
- printf("| latency_cpu_us=%f latency_realloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f",
+ dfatool_printf("| latency_cpu_us=%f latency_realloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f",
timer.time[0], // CPU
timer.time[1], // free + alloc
timer.time[2], // load
timer.time[3] + timer.time[4] + timer.time[5] + timer.time[7], // write
timer.time[6] + timer.time[8], // kernel
timer.time[9]); // read
- printf(" latency_write1_us=%f latency_write2_us=%f latency_write3_us=%f latency_write4_us=%f latency_kernel1_us=%f latency_kernel2_us=%f",
+ dfatool_printf(" latency_write1_us=%f latency_write2_us=%f latency_write3_us=%f latency_write4_us=%f latency_kernel1_us=%f latency_kernel2_us=%f",
timer.time[3],
timer.time[4],
timer.time[5],
timer.time[7],
timer.time[6],
timer.time[8]);
- printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
+ dfatool_printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
input_size * sizeof(T) / timer.time[0],
input_size * sizeof(T) / (timer.time[6] + timer.time[8]),
input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
+ dfatool_printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
input_size * sizeof(T) / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size * sizeof(T) / (timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
+ dfatool_printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
input_size / timer.time[0],
input_size / (timer.time[6] + timer.time[8]),
input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n",
+ dfatool_printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n",
input_size / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size / (timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
generated by cgit v1.2.3 (git 2.39.1) at 2025-07-01 06:13:27 +0000