VA: more detail output

3 files changed, 33 insertions, 7 deletions

diff --git a/VA/Makefile b/VA/Makefile
index e5a856d..b3c6199 100644
--- a/VA/Makefile
+++ b/VA/Makefile
@@ -4,13 +4,15 @@ BL ?= 10
 TYPE ?= INT32
 ENERGY ?= 0
 WITH_ALLOC_OVERHEAD ?= 0
+WITH_LOAD_OVERHEAD ?= 0
+WITH_FREE_OVERHEAD ?= 0
 
 COMMON_INCLUDES := support
 HOST_SOURCES := $(wildcard host/*.c)
 DPU_SOURCES := $(wildcard dpu/*.c)
 
 COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES}
-HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -D${TYPE} -DENERGY=${ENERGY} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD}
+HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -D${TYPE} -DENERGY=${ENERGY} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE}
 
 QUIET = @
diff --git a/VA/host/app.c b/VA/host/app.c
index 2812a6d..98cd782 100644
--- a/VA/host/app.c
+++ b/VA/host/app.c
@@ -64,14 +64,20 @@ int main(int argc, char **argv) {
     DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
 #endif
 
-#if !WITH_ALLOC_OVERHEAD
+    printf("WITH_ALLOC_OVERHEAD=%d WITH_LOAD_OVERHEAD=%d WITH_FREE_OVERHEAD=%d\n", WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD);
+
     // 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[0] = 0; // alloc
     timer.time[1] = 0; // load
+#endif
+#if !WITH_FREE_OVERHEAD
     timer.time[6] = 0; // free
 #endif
 
@@ -109,6 +115,8 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup) {
             stop(&timer, 0);
         }
+#endif
+#if WITH_LOAD_OVERHEAD
         if(rep >= p.n_warmup) {
             start(&timer, 1, 0);
         }
@@ -205,14 +213,18 @@ int main(int argc, char **argv) {
         }
 
 #if WITH_ALLOC_OVERHEAD
+#if WITH_FREE_OVERHEAD
         if(rep >= p.n_warmup) {
             start(&timer, 6, 0);
         }
+#endif
         DPU_ASSERT(dpu_free(dpu_set));
+#if WITH_FREE_OVERHEAD
         if(rep >= p.n_warmup) {
             stop(&timer, 6);
         }
 #endif
+#endif
 
         // Check output
         bool status = true;
@@ -241,15 +253,25 @@ int main(int argc, char **argv) {
                     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]));
-                printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f\n",
+                printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
+                    input_size * 3 * sizeof(T) / (timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size * 3 * sizeof(T) / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size * 3 * sizeof(T) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]));
+                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]));
+                printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f\n",
+                    input_size / (timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]));
             }
         } else {
             printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
         }
     }
+    printf("throughput_*_MOpps == n_elements / (+ latency_*_us ...)\n");
+    printf("throughput_*_MBps == 3 * sizeof(e_type) * throughput_*_MOpps \n");
 
     // Print timing results
     /*
diff --git a/VA/run.sh b/VA/run.sh
index e4cedb2..d99287b 100755
--- a/VA/run.sh
+++ b/VA/run.sh
@@ -16,11 +16,13 @@ echo "Revision $(git describe --always)"
 
 for nr_dpus in 1 4 8 16 32 64 128 256 512 768 1024 1536 2048; do
 	for nr_tasklets in 8 12 16 20 24; do
-		for i in  2048 4096 8192 16384 65536 262144 1048576 2621440 167772160; do
+		# 2621440 run-paper-weak
+		# 167772160 run-paper-strong-full (needs at least 20 DPUs)
+		for i in 2048 4096 8192 16384 65536 262144 1048576 2621440 167772160; do
 			for dt in CHAR SHORT INT32 INT64 FLOAT DOUBLE; do
 				echo
-				if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 TYPE=${dt} WITH_ALLOC_OVERHEAD=1; then
-					timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i ${i} -x 1 || true
+				if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 TYPE=${dt} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
+					timeout --foreground -k 1m 30m bin/host_code -w 0 -e 50 -i ${i} -x 1 || true
 				fi
 			done
 		done