VA: handle alloc/load/free overhead as parameters; vary them for fgbs24a

2 files changed, 11 insertions, 18 deletions

diff --git a/VA/host/app.c b/VA/host/app.c
index 065af72..93d6e57 100644
--- a/VA/host/app.c
+++ b/VA/host/app.c
@@ -76,8 +76,6 @@ int main(int argc, char **argv) {
     DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
 #endif
 
-    printf("WITH_ALLOC_OVERHEAD=%d WITH_LOAD_OVERHEAD=%d WITH_FREE_OVERHEAD=%d\n", WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD);
-
     // Timer declaration
     Timer timer;
 
@@ -263,8 +261,10 @@ int main(int argc, char **argv) {
         if (status) {
             printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
             if (rep >= p.n_warmup) {
-                printf("[::] VA UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d n_elements_per_dpu=%d b_sdk_singlethreaded=%d ",
-                    nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size, input_size / NR_DPUS, SDK_SINGLETHREADED);
+                printf("[::] VA UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d n_elements_per_dpu=%d",
+                    nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size, input_size / NR_DPUS);
+                printf(" b_sdk_singlethreaded=%d b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ",
+                    SDK_SINGLETHREADED, 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_read_us=%f latency_free_us=%f",
                     timer.time[0],
                     timer.time[1],
@@ -297,18 +297,6 @@ int main(int argc, char **argv) {
     printf("throughput_*_MOpps == n_elements / (+ latency_*_us ...)\n");
     printf("throughput_*_MBps == 3 * sizeof(e_type) * throughput_*_MOpps \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("DPU-CPU ");
-    print(&timer, 3, p.n_reps);
-    */
-
 #if ENERGY
     double energy;
     DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
diff --git a/VA/run-fgbs24a.sh b/VA/run-fgbs24a.sh
index f8941fb..0aa6999 100755
--- a/VA/run-fgbs24a.sh
+++ b/VA/run-fgbs24a.sh
@@ -4,6 +4,8 @@ set -e
 
 mkdir -p $(hostname)
 
+ts=$(date +%Y%m%d)
+
 # BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
 # T: data type
 # -w: number of un-timed warmup iterations
@@ -16,13 +18,16 @@ echo "prim-benchmarks VA (dfatool fgbs24a edition)"
 echo "Started at $(date)"
 echo "Revision $(git describe --always)"
 
-for nr_dpus in 2543 2304 2048; do
+for nr_dpus in 2304 2048 2543; do
 	for nr_tasklets in 16; do
 		echo
 		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
 			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 167772160 -x 1 || true
 		fi
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 167772160 -x 1 || true
+		fi
 	done
 done
 echo "Completed at $(date)"
-) | tee "$(hostname)/fgbs24a.txt"
+) | tee "$(hostname)/${ts}-fgbs24a.txt"