port SEL NMC to dfatool

6 files changed, 114 insertions, 27 deletions

diff --git a/SEL/baselines/cpu/Makefile b/SEL/baselines/cpu/Makefile
index 02d930c..81f6d17 100644
--- a/SEL/baselines/cpu/Makefile
+++ b/SEL/baselines/cpu/Makefile
@@ -16,13 +16,15 @@ sel_O2: app_baseline.c
 run: sel
 	./sel -i 1258291200 -t 4
 
+# upstream code does not include -e 20 and does 3 iterations instead
+
 .PHONY: run_O0
 run_O0: sel_O0
-	./sel_O0 -i 1258291200 -t 4
+	./sel_O0 -i 1258291200 -t 4 -e 20
 
 .PHONY: run_O2
 run_O2: sel_O2
-	./sel_O2 -i 1258291200 -t 4
+	./sel_O2 -i 1258291200 -t 4 -e 20
 
 .PHONY: clean
 clean:
diff --git a/SEL/baselines/cpu/app_baseline.c b/SEL/baselines/cpu/app_baseline.c
index 04a569f..6ee1cae 100644
--- a/SEL/baselines/cpu/app_baseline.c
+++ b/SEL/baselines/cpu/app_baseline.c
@@ -147,16 +147,12 @@ 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("[::] SEL CPU | n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_MBps=%f",
                 nr_threads, XSTR(T), file_size,
                 file_size * 2 * sizeof(T) / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_MOpps=%f\n",
-                nr_threads, XSTR(T), file_size,
+            printf(" throughput_MOpps=%f",
                 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/SEL/host/app.c b/SEL/host/app.c
index ef07cf9..2194c81 100644
--- a/SEL/host/app.c
+++ b/SEL/host/app.c
@@ -22,6 +22,9 @@
 #define DPU_BINARY "./bin/dpu_code"
 #endif
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 #if ENERGY
 #include <dpu_probe.h>
 #endif
@@ -208,6 +211,35 @@ int main(int argc, char **argv) {
 
         // Free memory
         free(results_scan);
+
+        // Check output
+        bool status = true;
+        if(accum != total_count) status = false;
+        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("[::] SEL 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 * 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] + timer.time[4]));
+                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] + timer.time[4]));
+                printall(&timer, 4);
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
     }
 
     // Print timing results
@@ -228,28 +260,11 @@ int main(int argc, char **argv) {
     printf("DPU Energy (J): %f\t", energy);
     #endif	
 
-    // Check output
-    bool status = true;
-    if(accum != total_count) status = false;
-    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");
-    } else {
-        printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
-    }
-
     // Deallocation
     free(A);
     free(C);
     free(C2);
     DPU_ASSERT(dpu_free(dpu_set));
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/SEL/run-paper-strong-full.sh b/SEL/run-paper-strong-full.sh
new file mode 100755
index 0000000..cc1a99d
--- /dev/null
+++ b/SEL/run-paper-strong-full.sh
@@ -0,0 +1,25 @@
+#!/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 UNI strong-full (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 256 512 1024 2048; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 251658240 -x 1 || true
+		fi
+	done
+done
+) | tee log-paper-strong-full.txt
diff --git a/SEL/run-paper-strong-rank.sh b/SEL/run-paper-strong-rank.sh
new file mode 100755
index 0000000..6cffd65
--- /dev/null
+++ b/SEL/run-paper-strong-rank.sh
@@ -0,0 +1,26 @@
+#!/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 UNI strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream config space
+for nr_dpus in 512 256 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 3932160 -x 1 || true
+		fi
+	done
+done
+) | tee log-paper-strong-rank.txt
diff --git a/SEL/run-paper-weak.sh b/SEL/run-paper-weak.sh
new file mode 100755
index 0000000..5e83c5e
--- /dev/null
+++ b/SEL/run-paper-weak.sh
@@ -0,0 +1,23 @@
+#!/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 UNI weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream config space
+for nr_dpus in 512 256 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 3932160 -x 0 || true
+		fi
+	done
+done | tee log-paper-weak.txt