VA: new dfatool format; -O and reproduction benchmarks

6 files changed, 65 insertions, 15 deletions

diff --git a/VA/baselines/cpu/Makefile b/VA/baselines/cpu/Makefile
index 3171e19..fb7c383 100644
--- a/VA/baselines/cpu/Makefile
+++ b/VA/baselines/cpu/Makefile
@@ -16,13 +16,15 @@ va_O2: app_baseline.c
 run: va
 	./va -t 4
 
+# -w/-e are not part of upstream config
+
 .PHONY: run_O0
 run_O0: va_O0
-	./va_O0 -t 4
+	./va_O0 -t 4 -w 1 -e 100
 
 .PHONY: run_O2
 run_O2: va_O2
-	./va_O2 -t 4
+	./va_O2 -t 4 -w 1 -e 100
 
 .PHONY: clean
 clean:
diff --git a/VA/baselines/cpu/app_baseline.c b/VA/baselines/cpu/app_baseline.c
index 3fdc10b..1ac6c14 100644
--- a/VA/baselines/cpu/app_baseline.c
+++ b/VA/baselines/cpu/app_baseline.c
@@ -137,16 +137,13 @@ int main(int argc, char **argv) {
         nr_threads++;
 
         if (rep >= p.n_warmup) {
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_MBps=%f\n",
+            printf("[::] VA CPU | n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_MBps=%f",
                 nr_threads, XSTR(T), file_size,
                 file_size * 3 * sizeof(T) / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_MOpps=%f\n",
+            printf(" throughput_MOpps=%f",
                 nr_threads, XSTR(T), file_size,
                 file_size / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d |",
-                nr_threads, XSTR(T), file_size);
             printall(&timer, 0);
         }
     }
diff --git a/VA/baselines/cpu/run-opti.sh b/VA/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..0a3a4a3
--- /dev/null
+++ b/VA/baselines/cpu/run-opti.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+make run_O0 | sed 's/CPU/CPU O0/' | tee "${HOST}-O0.txt"
+
+make run_O2 | sed 's/CPU/CPU O2/' | tee "${HOST}-O2.txt"
diff --git a/VA/host/app.c b/VA/host/app.c
index bfa6806..7758d65 100644
--- a/VA/host/app.c
+++ b/VA/host/app.c
@@ -187,20 +187,16 @@ 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("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
-                    "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f\n",
+                printf("[::] VA NMC | n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
+                    "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f",
                     nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
                     input_size * 3 * sizeof(T) / timer.time[0],
                     input_size * 3 * sizeof(T) / (timer.time[1]),
                     input_size * 3 * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
-                printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
-                    "| throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
-                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
+                printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
                     input_size / timer.time[0],
                     input_size / (timer.time[1]),
                     input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
-                printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d | ",
-                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size);
                 printall(&timer, 3);
             }
         } else {
diff --git a/VA/run-paper-strong-rank.sh b/VA/run-paper-strong-rank.sh
new file mode 100755
index 0000000..671eafc
--- /dev/null
+++ b/VA/run-paper-strong-rank.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
+# T: data type
+# -w: number of un-timed warmup iterations
+# -e: number of timed iterations
+# -i; ignored, always uses 262144 elements
+
+echo "prim-benchmarks VA strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 2621440 -x 1 || true
+		fi
+	done
+done | tee log-paper-strong-rank.txt
diff --git a/VA/run-paper-weak.sh b/VA/run-paper-weak.sh
new file mode 100755
index 0000000..99eed79
--- /dev/null
+++ b/VA/run-paper-weak.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
+# T: data type
+# -w: number of un-timed warmup iterations
+# -e: number of timed iterations
+# -i; ignored, always uses 262144 elements
+
+echo "prim-benchmarks VA weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 2621440 -x 0 || true
+		fi
+	done
+done | tee log-paper-weak.txt