summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--GEMV/Makefile44
-rw-r--r--GEMV/host/app.c81
-rwxr-xr-xGEMV/run.sh26
3 files changed, 94 insertions, 57 deletions
diff --git a/GEMV/Makefile b/GEMV/Makefile
index fbba9b8..25174d3 100644
--- a/GEMV/Makefile
+++ b/GEMV/Makefile
@@ -1,44 +1,36 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
+NR_DPUS ?= 1
NR_TASKLETS ?= 16
BL ?= 10
-NR_DPUS ?= 1
-
-define conf_filename
- ${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL})
-
-HOST_TARGET := ${BUILDDIR}/gemv_host
-DPU_TARGET := ${BUILDDIR}/gemv_dpu
COMMON_INCLUDES := support
-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=${NR_DPUS} -DBL=${BL}
DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL}
-all: ${HOST_TARGET} ${DPU_TARGET}
+QUIET = @
+
+ifdef verbose
+ QUIET =
+endif
+
+all: bin/gemv_host bin/gemv_dpu
-${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}
+bin/gemv_host: ${HOST_SOURCES} ${COMMON_INCLUDES} bin
+ ${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
-${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
- dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+bin/gemv_dpu: ${DPU_SOURCES} ${COMMON_INCLUDES} bin
+ ${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
clean:
- $(RM) -r $(BUILDDIR)
+ ${QUIET}rm -rf bin
test: all
- ./${HOST_TARGET} -m 1024 -n 1024
+ bin/gemv_host -m 1024 -n 1024
diff --git a/GEMV/host/app.c b/GEMV/host/app.c
index de2e847..daf4427 100644
--- a/GEMV/host/app.c
+++ b/GEMV/host/app.c
@@ -18,6 +18,9 @@
#include <dpu_probe.h>
#endif
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
#include "../support/common.h"
#include "../support/timer.h"
#include "../support/params.h"
@@ -140,15 +143,12 @@ int main(int argc, char **argv) {
Timer timer;
// Compute output on CPU (performance comparison and verification purposes)
- start(&timer, 0, 0);
- gemv_host(C, A, B, m_size, n_size);
- stop(&timer, 0);
for (unsigned int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
-
-
-
+ start(&timer, 0, 0);
+ gemv_host(C, A, B, m_size, n_size);
+ stop(&timer, 0);
if (rep >= p.n_warmup)
- start(&timer, 1, rep - p.n_warmup);
+ start(&timer, 1, 0);
// Input arguments
i = 0;
DPU_FOREACH(dpu_set, dpu, i) {
@@ -177,7 +177,7 @@ int main(int argc, char **argv) {
// 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
@@ -202,7 +202,7 @@ int main(int argc, char **argv) {
// Retrieve results
C_dpu = malloc(max_rows_per_dpu * nr_of_dpus * sizeof(T));
if (rep >= p.n_warmup)
- start(&timer, 3, rep - p.n_warmup);
+ start(&timer, 3, 0);
i = 0;
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, C_dpu + i * max_rows_per_dpu));
@@ -210,6 +210,46 @@ int main(int argc, char **argv) {
DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, DPU_MRAM_HEAP_POINTER_NAME, max_rows_per_dpu * n_size_pad * sizeof(T) + n_size_pad * sizeof(T), max_rows_per_dpu * sizeof(T), DPU_XFER_DEFAULT));
if(rep >= p.n_warmup)
stop(&timer, 3);
+
+
+ // Check output
+ bool status = true;
+ unsigned int n,j;
+ i = 0;
+ for (n = 0; n < nr_of_dpus; n++) {
+ for (j = 0; j < dpu_info[n].rows_per_dpu; j++) {
+ if(C[i] != C_dpu[n * max_rows_per_dpu + j]) {
+ status = false;
+#if PRINT
+ // printf("%d: %d -- %d\n", i, C[i], C_dpu[n * max_rows_per_dpu + j]);
+#endif
+ }
+ i++;
+ }
+ }
+ 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 n_elements=%d "
+ "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f\n",
+ nr_of_dpus, NR_TASKLETS, XSTR(T), n_size * m_size,
+ n_size * m_size * sizeof(T) / timer.time[0],
+ n_size * m_size * sizeof(T) / timer.time[2],
+ n_size * m_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
+ printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d "
+ "| throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
+ nr_of_dpus, NR_TASKLETS, XSTR(T), n_size * m_size,
+ n_size * m_size / timer.time[0],
+ n_size * m_size / timer.time[2],
+ n_size * m_size / (timer.time[1] + timer.time[2] + timer.time[3]));
+ printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d |",
+ nr_of_dpus, NR_TASKLETS, XSTR(T), n_size * m_size);
+ printall(&timer, 3);
+ }
+ } else {
+ printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+ }
+
}
#if ENERGY
double acc_energy, avg_energy, acc_time, avg_time;
@@ -233,27 +273,6 @@ int main(int argc, char **argv) {
printf("Energy (J): %f J\t", avg_energy);
#endif
- // Check output
- bool status = true;
- unsigned int n,j;
- i = 0;
- for (n = 0; n < nr_of_dpus; n++) {
- for (j = 0; j < dpu_info[n].rows_per_dpu; j++) {
- if(C[i] != C_dpu[n * max_rows_per_dpu + j]) {
- status = false;
-#if PRINT
- // printf("%d: %d -- %d\n", i, C[i], C_dpu[n * max_rows_per_dpu + j]);
-#endif
- }
- i++;
- }
- }
- 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(B);
@@ -265,5 +284,5 @@ int main(int argc, char **argv) {
DPU_ASSERT(dpu_probe_deinit(&probe));
#endif
- return status ? 0 : -1;
+ return 0;
}
diff --git a/GEMV/run.sh b/GEMV/run.sh
new file mode 100755
index 0000000..2f61ef2
--- /dev/null
+++ b/GEMV/run.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -e
+
+# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
+# -w: number of un-timed warmup iterations
+# -e: number of timed iterations
+# -m: number of rows
+# -n: number of cols
+
+echo "prim-benchmarks VA (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for n in 512 1024 2048 4096; do
+ for m in 512 1024 2048 4096; do
+ for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do
+ for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do
+ echo
+ if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+ timeout --foreground -k 1m 30m bin/gemv_host -w 0 -e 100 -m $m -n $n || true
+ fi
+ done
+ done
+ done
+done
generated by cgit v1.2.3 (git 2.39.1) at 2025-02-16 19:05:34 +0000