port BS NMC to dfatool

3 files changed, 89 insertions, 45 deletions

diff --git a/BS/Makefile b/BS/Makefile
index 129ade1..99dd0f8 100644
--- a/BS/Makefile
+++ b/BS/Makefile
@@ -1,44 +1,39 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
-NR_TASKLETS ?= 16
 NR_DPUS ?= 1
-PROBLEM_SIZE ?= 2
-
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS})
+NR_TASKLETS ?= 16
+PROBLEM_SIZE ?= 8192 # original 2, wtf
 
 COMMON_INCLUDES := support
-HOST_TARGET := ${BUILDDIR}/bs_host
-DPU_TARGET := ${BUILDDIR}/bs_dpu
+HOST_TARGET := bin/bs_host
+DPU_TARGET := bin/bs_dpu
 
-HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
-DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
+HOST_SOURCES := $(wildcard host/*.c)
+DPU_SOURCES := $(wildcard dpu/*.c)
 
 .PHONY: all clean test
 
-__dirs := $(shell mkdir -p ${BUILDDIR})
-
 COMMON_FLAGS := -Wall -Wextra  -g -I${COMMON_INCLUDES}
 HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DPROBLEM_SIZE=${PROBLEM_SIZE}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS}
 
+QUIET = @
+
+ifdef verbose
+	QUIET =
+endif
+
 all: ${HOST_TARGET} ${DPU_TARGET}
 
-${CONF}:
-	$(RM) $(call conf_filename,*,*)
-	touch ${CONF}
+bin:
+	${QUIET}mkdir -p bin
 
 ${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	$(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+	${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
 
 ${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
 
 clean:
-	$(RM) -r $(BUILDDIR)
+	${QUIET}rm -rf bin
 
 test: all
-	./${HOST_TARGET} -i 262144
+	${HOST_TARGET} -i 262144
diff --git a/BS/host/app.c b/BS/host/app.c
index 3929877..85587ac 100644
--- a/BS/host/app.c
+++ b/BS/host/app.c
@@ -18,6 +18,9 @@
 #include <dpu_probe.h>
 #endif
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 #include "params.h"
 #include "timer.h"
 
@@ -48,6 +51,7 @@ int64_t binarySearch(DTYPE * input, DTYPE * querys, DTYPE input_size, uint64_t n
 		while (l <= r) {
 			DTYPE m = l + (r - l) / 2;
 
+			// XXX shouldn't this short-circuit?
 			// Check if x is present at mid
 			if (input[m] == querys[q])
 			result = m;
@@ -101,11 +105,6 @@ int main(int argc, char **argv) {
 	// Create an input file with arbitrary data
 	create_test_file(input, querys, input_size, num_querys);
 
-	// Compute host solution
-	start(&timer, 0, 0);
-	result_host = binarySearch(input, querys, input_size - 1, num_querys);
-	stop(&timer, 0);
-
 	// Create kernel arguments
 	uint64_t slice_per_dpu          = num_querys / nr_of_dpus;
 	dpu_arguments_t input_arguments = {input_size, slice_per_dpu, 0};
@@ -114,8 +113,14 @@ int main(int argc, char **argv) {
 		// Perform input transfers
 		uint64_t i = 0;
 
-		if (rep >= p.n_warmup)
-		start(&timer, 1, rep - p.n_warmup);
+		// Compute host solution
+		start(&timer, 0, 0);
+		result_host = binarySearch(input, querys, input_size - 1, num_querys);
+		stop(&timer, 0);
+
+		if (rep >= p.n_warmup) {
+			start(&timer, 1, 0);
+		}
 
 		DPU_FOREACH(dpu_set, dpu, i)
 		{
@@ -142,13 +147,14 @@ int main(int argc, char **argv) {
 
 		DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, input_size * sizeof(DTYPE), slice_per_dpu * sizeof(DTYPE), DPU_XFER_DEFAULT));
 
-		if (rep >= p.n_warmup)
-		stop(&timer, 1);
+		if (rep >= p.n_warmup) {
+			stop(&timer, 1);
+		}
 
 		// Run kernel on DPUs
 		if (rep >= p.n_warmup)
 		{
-			start(&timer, 2, rep - p.n_warmup);
+			start(&timer, 2, 0);
 			#if ENERGY
 			DPU_ASSERT(dpu_probe_start(&probe));
 			#endif
@@ -176,8 +182,9 @@ int main(int argc, char **argv) {
 		#endif
 
 		// Retrieve results
-		if (rep >= p.n_warmup)
-		start(&timer, 3, rep - p.n_warmup);
+		if (rep >= p.n_warmup) {
+			start(&timer, 3, 0);
+		}
 		dpu_results_t* results_retrieve[nr_of_dpus];
 		i = 0;
 		DPU_FOREACH(dpu_set, dpu, i)
@@ -199,10 +206,36 @@ int main(int argc, char **argv) {
 			}
 			free(results_retrieve[i]);
 		}
-		if(rep >= p.n_warmup)
-		stop(&timer, 3);
+		if(rep >= p.n_warmup) {
+			stop(&timer, 3);
+		}
+
+		int status = (result_dpu == result_host);
+		if (status) {
+			printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] results are equal\n");
+			if (rep >= p.n_warmup) {
+				printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%lu "
+					"| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f\n",
+					nr_of_dpus, NR_TASKLETS, XSTR(DTYPE), input_size,
+					num_querys * sizeof(DTYPE) / timer.time[0],
+					num_querys * sizeof(DTYPE) / timer.time[2],
+					num_querys * sizeof(DTYPE) / (timer.time[1] + timer.time[2] + timer.time[3]));
+				printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%lu "
+					"| throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
+					nr_of_dpus, NR_TASKLETS, XSTR(DTYPE), input_size,
+					num_querys / timer.time[0],
+					num_querys / timer.time[2],
+					num_querys / (timer.time[1] + timer.time[2] + timer.time[3]));
+				printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%lu |",
+					nr_of_dpus, NR_TASKLETS, XSTR(DTYPE), input_size);
+				printall(&timer, 3);
+			}
+		} else {
+			printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] results differ!\n");
+		}
 	}
 	// Print timing results
+	/*
 	printf("CPU Version Time (ms): ");
 	print(&timer, 0, p.n_reps);
 	printf("CPU-DPU Time (ms): ");
@@ -211,6 +244,7 @@ int main(int argc, char **argv) {
 	print(&timer, 2, p.n_reps);
 	printf("DPU-CPU Time (ms): ");
 	print(&timer, 3, p.n_reps);
+	*/
 
 	#if ENERGY
 	double energy;
@@ -218,15 +252,8 @@ int main(int argc, char **argv) {
 	printf("DPU Energy (J): %f\t", energy * num_iterations);
 	#endif
 
-	int status = (result_dpu == result_host);
-	if (status) {
-		printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] results are equal\n");
-	} else {
-		printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] results differ!\n");
-	}
-
 	free(input);
 	DPU_ASSERT(dpu_free(dpu_set));
 
-	return status ? 0 : 1;
+	return 0;
 }
diff --git a/BS/run.sh b/BS/run.sh
new file mode 100755
index 0000000..cf32e4f
--- /dev/null
+++ b/BS/run.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 (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do
+	for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 || true
+		fi
+	done
+done