summaryrefslogtreecommitdiff
path: root/TRNS
diff options
context:
space:
mode:
Diffstat (limited to 'TRNS')
-rw-r--r--TRNS/Makefile31
-rw-r--r--TRNS/baselines/cpu/Makefile28
-rw-r--r--TRNS/baselines/cpu/main.cpp16
-rwxr-xr-xTRNS/baselines/cpu/run-perf.sh6
-rwxr-xr-xTRNS/benchmark-scripts/ccmcc25-sim.sh53
-rwxr-xr-xTRNS/benchmark-scripts/ccmcc25.sh32
-rwxr-xr-xTRNS/dimes-hetsim-hbm.sh7
-rwxr-xr-xTRNS/dimes-hetsim-nmc.sh4
-rw-r--r--TRNS/dpu/task.c2
-rw-r--r--TRNS/host/app.c62
-rw-r--r--[-rwxr-xr-x]TRNS/include/common.h (renamed from TRNS/support/common.h)12
-rw-r--r--TRNS/include/dfatool_host.ah36
-rw-r--r--TRNS/include/params.h (renamed from TRNS/support/params.h)4
-rw-r--r--TRNS/include/timer.h5
-rwxr-xr-xTRNS/run-fgbs24a.sh29
-rwxr-xr-xTRNS/run-paper-strong-full.sh30
-rwxr-xr-xTRNS/run-paper-strong-rank.sh29
-rwxr-xr-xTRNS/run-paper-weak.sh28
-rwxr-xr-xTRNS/run-rank.sh32
-rwxr-xr-xTRNS/run.sh32
-rwxr-xr-xTRNS/support/timer.h66
21 files changed, 242 insertions, 302 deletions
diff --git a/TRNS/Makefile b/TRNS/Makefile
index fd3f493..302fcd6 100644
--- a/TRNS/Makefile
+++ b/TRNS/Makefile
@@ -1,18 +1,32 @@
NR_DPUS ?= 1
NR_TASKLETS ?= 16
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=${NR_DPUS} -DENERGY=${ENERGY} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD}
+aspectc ?= 0
+aspectc_timing ?= *
+dfatool_timing ?= 1
+
+HOST_CC := ${CC}
+
+COMMON_FLAGS := -Wall -Wextra -g -Iinclude
+HOST_FLAGS := ${COMMON_FLAGS} -O3 -march=native `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DENERGY=${ENERGY} -DDFATOOL_TIMING=${dfatool_timing} -DASPECTC=${aspectc}
DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS}
+ifeq (${aspectc_timing}, 1)
+ ASPECTC_HOST_FLAGS += -ainclude/dfatool_host_dpu.ah -ainclude/dfatool_host.ah
+endif
+
+ASPECTC_HOST_FLAGS ?= -a0
+
+ifeq (${aspectc}, 1)
+ HOST_CC = ag++ -r repo.acp -v 0 ${ASPECTC_HOST_FLAGS} --c_compiler ${UPMEM_HOME}/bin/clang++ -p . --Xcompiler
+else
+ HOST_FLAGS += -std=c11
+endif
+
QUIET = @
ifdef verbose
@@ -24,8 +38,11 @@ all: bin/host_code bin/dpu_code
bin:
${QUIET}mkdir -p bin
+# cp/rm are needed to work around AspectC++ not liking symlinks
bin/host_code: ${HOST_SOURCES} ${COMMON_INCLUDES} bin
- ${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+ ${QUIET}cp ../include/dfatool_host_dpu.ah include
+ ${QUIET}${HOST_CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+ ${QUIET}rm -f include/dfatool_host_dpu.ah
bin/dpu_code: ${DPU_SOURCES} ${COMMON_INCLUDES} bin
${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
diff --git a/TRNS/baselines/cpu/Makefile b/TRNS/baselines/cpu/Makefile
index 236f7bb..2f28738 100644
--- a/TRNS/baselines/cpu/Makefile
+++ b/TRNS/baselines/cpu/Makefile
@@ -32,16 +32,30 @@
# THE SOFTWARE.
#
-NUMA ?= 0
-NUMA_MEMCPY ?= 0
-FLAGS =
+benchmark ?= 1
+debug ?= 0
+native ?= 1
+nop_sync ?= 0
+numa ?= 0
+numa_memcpy ?= 0
-ifeq (${NUMA}, 1)
- FLAGS += -lnuma
+CFLAGS =
+LDFLAGS =
+
+ifeq (${debug}, 1)
+ CFLAGS += -g
+endif
+
+ifeq (${native}, 1)
+ CFLAGS += -march=native
+endif
+
+ifeq (${numa}, 1)
+ LDFLAGS += -lnuma
endif
CXX=g++
-CXX_FLAGS=-std=c++11 -Wall -Wextra -pedantic -DNUMA=${NUMA} -DNUMA_MEMCPY=${NUMA_MEMCPY}
+CXX_FLAGS=-std=c++11 -Wall -Wextra -pedantic -DNUMA=${numa} -DNUMA_MEMCPY=${numa_memcpy} -DNOP_SYNC=${nop_sync} -DWITH_BENCHMARK=${benchmark}
LIB=-L/usr/lib/ -lm -pthread
@@ -52,7 +66,7 @@ EXE=trns
all: trns
trns: ${SRC}
- $(CXX) -O2 $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE) $(FLAGS)
+ $(CXX) -O3 $(CXX_FLAGS) ${CFLAGS} $(SRC) $(LIB) -o $(EXE) ${LDFLAGS}
trns_O0: ${SRC}
$(CXX) $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)_O0
diff --git a/TRNS/baselines/cpu/main.cpp b/TRNS/baselines/cpu/main.cpp
index c8cccaf..b4cd149 100644
--- a/TRNS/baselines/cpu/main.cpp
+++ b/TRNS/baselines/cpu/main.cpp
@@ -36,9 +36,18 @@
#include "support/setup.h"
#include "kernel.h"
#include "support/common.h"
-#include "support/timer.h"
#include "support/verify.h"
+#if WITH_BENCHMARK
+#include "support/timer.h"
+#else
+#include <string>
+struct Timer {
+ inline void start(std::string name) {(void)name;}
+ inline void stop(std::string name) {(void)name;}
+};
+#endif
+
#include <unistd.h>
#include <thread>
#include <string.h>
@@ -362,6 +371,7 @@ int main(int argc, char **argv) {
timer.stop("free");
#endif
+#if WITH_BENCHMARK
if (rep >= p.n_warmup) {
#if NUMA_MEMCPY
printf("[::] TRNS-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%d"
@@ -396,10 +406,8 @@ int main(int argc, char **argv) {
timer.get("Step 1") + timer.get("Step 2") + timer.get("Step 3"));
#endif // NUMA_MEMCPY
}
+#endif // WITH_BENCHMARK
}
- //timer.print("Step 1", p.n_reps);
- //timer.print("Step 2", p.n_reps);
- //timer.print("Step 3", p.n_reps);
// Verify answer
//verify(h_local, h_in_backup, M_ * m, N_ * n, 1);
diff --git a/TRNS/baselines/cpu/run-perf.sh b/TRNS/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..f16a3b1
--- /dev/null
+++ b/TRNS/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B numa=1
+
+perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 1 -a 4 -c 4
+perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 4 -a 4 -c 4
diff --git a/TRNS/benchmark-scripts/ccmcc25-sim.sh b/TRNS/benchmark-scripts/ccmcc25-sim.sh
new file mode 100755
index 0000000..a7aa79c
--- /dev/null
+++ b/TRNS/benchmark-scripts/ccmcc25-sim.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname)
+
+# Args: -m m -n n -o M_ -p N_
+#
+# Input: (M_ * m) × (N_ * n) matrix
+# Output: (N_* n) × (M_ * m) matrix
+# Step 1: transpose (M_ * m) × N_ matrix that consists of tiles of size n
+# CPU version: explicit
+# DPU version: implicit (M_ * m write operations of #DPUs * n elements to DPUs)
+# Step 2: transpose m × n matrix; this happens N_ * M_ times.
+# DPU version: Each tasklet transposes a single m × n matrix / tile.
+# (16 × 8 tile takes up 1 KiB WRAM)
+# Step 3: Transpose M_ × n matrix that consists of tiles of size m.
+#
+# Note for DPU version: if M_ > #DPUs, steps 1 through 3 are repeated.
+# Number of repetitions == ceil(M_ / #DPUS)
+# For Hetsim benchmarks, we set M_ == #DPUs to simplify the task graph (no repetitions that depend on the number of available DPUs).
+# Just in case, there is also a configuration with M_ == 2048 independent of #DPUs
+#
+# input size: uint64(DPU)/double(CPU) * M_ * m * N_ * n
+# output size: uint64(DPU)/double(CPU) * M_ * m * N_ * n -- on DPU only; CPU version operates in-place
+# Upstream DPU version uses int64_t, -p 2048 -o 12288 -x 1 [implicit -m 16 -n 8]
+# Upstream CPU version uses double, -p 2556 -o 4096 -m 16 -n 8 and fails with -o 12288 (allocation error)
+#
+# -p 2048 -o 2048 -m 16 -n 8 -> matrix size: 4 GiB
+# -p [64 .. 2304] -o 2048 -m 16 -n 8 -> matrix size: 128 MiB .. 4.5 GiB
+
+run_benchmark_nmc() {
+ local "$@"
+ set -e
+ make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} \
+ aspectc=1 aspectc_timing=1 dfatool_timing=0
+ bin/host_code -w 0 -e 2 -p ${cols} -o ${rows} -m ${tile_rows} -n ${tile_cols}
+}
+
+export -f run_benchmark_nmc
+
+fn=log/$(hostname)/ccmcc25-sdk${sdk}-sim
+
+source ~/lib/local/upmem/upmem-2025.1.0-Linux-x86_64/upmem_env.sh simulator
+
+echo "prim-benchmarks TRNS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt
+
+parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 cols={cols} rows={rows} tile_cols={tile_cols} tile_rows={tile_rows} \
+ ::: nr_dpus 1 4 16 32 64 \
+ ::: rows 64 128 256 512 \
+ ::: cols 64 128 256 512 \
+ ::: tile_rows 16 \
+ ::: tile_cols 8 \
+>> ${fn}.txt
diff --git a/TRNS/benchmark-scripts/ccmcc25.sh b/TRNS/benchmark-scripts/ccmcc25.sh
new file mode 100755
index 0000000..7c66306
--- /dev/null
+++ b/TRNS/benchmark-scripts/ccmcc25.sh
@@ -0,0 +1,32 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname)
+
+run_benchmark_nmc() {
+ local "$@"
+ set -e
+ sudo limit_ranks_to_numa_node ${numa_rank}
+ make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} dfatool_timing=0 aspectc=1 aspectc_timing=1
+ bin/host_code -w 0 -e 4 -p ${cols} -o ${rows} -m ${tile_rows} -n ${tile_cols}
+}
+
+export -f run_benchmark_nmc
+
+for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do
+
+ fn=log/$(hostname)/ccmcc25-sdk${sdk}
+
+ source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh
+
+ echo "prim-benchmarks TRNS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt
+
+ parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank=any cols={cols} rows={rows} tile_cols={tile_cols} tile_rows={tile_rows} \
+ ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 \
+ ::: rows 1024 2048 4096 \
+ ::: cols 64 128 256 512 768 1024 1536 2048 2304 \
+ ::: tile_rows 16 \
+ ::: tile_cols 8 \
+ >> ${fn}.txt
+
+done
diff --git a/TRNS/dimes-hetsim-hbm.sh b/TRNS/dimes-hetsim-hbm.sh
index e2efaee..cc5dc68 100755
--- a/TRNS/dimes-hetsim-hbm.sh
+++ b/TRNS/dimes-hetsim-hbm.sh
@@ -32,7 +32,7 @@ fn=log/$(hostname)/dimes-hetsim-hbm
(
-make -B NUMA=1 NUMA_MEMCPY=1
+make -B numa=1 numa_memcpy=1
echo "CPU single-node operation with setup cost, memcpy node == input node, cpu node == output node (1/3)" >&2
@@ -43,10 +43,9 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
::: ram_in $(seq 0 15) \
:::+ cpu_memcpy $(seq 0 7) $(seq 0 7) \
::: ram_local $(seq 0 15) \
- :::+ cpu $(seq 0 7) $(seq 0 7) \
- ::: input_size 167772160
+ :::+ cpu $(seq 0 7) $(seq 0 7)
-make -B NUMA=1
+make -B numa=1
echo "CPU single-node operation (2/3)" >&2
diff --git a/TRNS/dimes-hetsim-nmc.sh b/TRNS/dimes-hetsim-nmc.sh
index b5f6f13..80987e7 100755
--- a/TRNS/dimes-hetsim-nmc.sh
+++ b/TRNS/dimes-hetsim-nmc.sh
@@ -73,7 +73,7 @@ parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
) >> ${fn}.txt
cd baselines/cpu
-make -B NUMA=1
+make -B numa=1
(
@@ -97,7 +97,7 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
) >> ${fn}.txt
-make -B NUMA=1 NUMA_MEMCPY=1
+make -B numa=1 numa_memcpy=1
(
diff --git a/TRNS/dpu/task.c b/TRNS/dpu/task.c
index 0f5e4be..9c0e0a8 100644
--- a/TRNS/dpu/task.c
+++ b/TRNS/dpu/task.c
@@ -12,7 +12,7 @@
#include <mutex.h>
#include <barrier.h>
-#include "../support/common.h"
+#include "common.h"
__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
diff --git a/TRNS/host/app.c b/TRNS/host/app.c
index 452b894..c178a19 100644
--- a/TRNS/host/app.c
+++ b/TRNS/host/app.c
@@ -7,16 +7,32 @@
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
+
+#if ASPECTC
+extern "C" {
+#endif
+
#include <dpu.h>
#include <dpu_log.h>
+#include <dpu_management.h>
+#include <dpu_target_macros.h>
+
+#if ENERGY
+#include <dpu_probe.h>
+#endif
+
+#if ASPECTC
+}
+#endif
+
#include <unistd.h>
#include <getopt.h>
#include <assert.h>
#include <math.h>
-#include "../support/common.h"
-#include "../support/timer.h"
-#include "../support/params.h"
+#include "common.h"
+#include "timer.h"
+#include "params.h"
#define XSTR(x) STR(x)
#define STR(x) #x
@@ -26,18 +42,13 @@
#define DPU_BINARY "./bin/dpu_code"
#endif
-#if ENERGY
-#include <dpu_probe.h>
-#endif
-
-#include <dpu_management.h>
-#include <dpu_target_macros.h>
-
// Pointer declaration
static T* A_host;
static T* A_backup;
static T* A_result;
+unsigned int kernel = 0;
+
// Create input arrays
static void read_input(T* A, unsigned int nr_elements) {
srand(0);
@@ -84,10 +95,10 @@ int main(int argc, char **argv) {
N_ = p.exp == 0 ? N_ * NR_DPUS : N_;
// Input/output allocation
- A_host = malloc(M_ * m * N_ * n * sizeof(T));
- A_backup = malloc(M_ * m * N_ * n * sizeof(T));
- A_result = malloc(M_ * m * N_ * n * sizeof(T));
- T* done_host = malloc(M_ * n); // Host array to reset done array of step 3
+ A_host = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ A_backup = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ A_result = (T*)malloc(M_ * m * N_ * n * sizeof(T));
+ T* done_host = (T*)malloc(M_ * n); // Host array to reset done array of step 3
memset(done_host, 0, M_ * n);
// Create an input file with arbitrary data
@@ -131,6 +142,7 @@ int main(int argc, char **argv) {
DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
stop(&timer, 2);
DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
+ DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
} else if (first_round){
start(&timer, 1, 0);
DPU_ASSERT(dpu_alloc(active_dpus, NULL, &dpu_set));
@@ -174,8 +186,8 @@ int main(int argc, char **argv) {
start(&timer, 5, !first_round);
}
- unsigned int kernel = 0;
- dpu_arguments_t input_arguments = {m, n, M_, kernel};
+ kernel = 0;
+ dpu_arguments_t input_arguments = {m, n, M_, (enum kernels)kernel};
// transfer control instructions to DPUs (run first program part)
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments));
@@ -215,7 +227,7 @@ int main(int argc, char **argv) {
start(&timer, 7, !first_round);
}
kernel = 1;
- dpu_arguments_t input_arguments2 = {m, n, M_, kernel};
+ dpu_arguments_t input_arguments2 = {m, n, M_, (enum kernels)kernel};
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments2));
}
@@ -299,7 +311,9 @@ int main(int argc, char **argv) {
}
if (status) {
printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+#if DFATOOL_TIMING
unsigned long input_size = M_ * m * N_ * n;
+#endif
if (rep >= p.n_warmup) {
/*
* timer 0: CPU version
@@ -313,35 +327,35 @@ int main(int argc, char **argv) {
* timer 8: run DPU program (second kernel)
* timer 9: read transposed matrix
*/
- printf("[::] TRNS-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%lu numa_node_rank=%d ",
+ dfatool_printf("[::] TRNS-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%lu numa_node_rank=%d ",
NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), input_size, numa_node_rank);
- printf("| latency_cpu_us=%f latency_realloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f",
+ dfatool_printf("| latency_cpu_us=%f latency_realloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f",
timer.time[0], // CPU
timer.time[1], // free + alloc
timer.time[2], // load
timer.time[3] + timer.time[4] + timer.time[5] + timer.time[7], // write
timer.time[6] + timer.time[8], // kernel
timer.time[9]); // read
- printf(" latency_write1_us=%f latency_write2_us=%f latency_write3_us=%f latency_write4_us=%f latency_kernel1_us=%f latency_kernel2_us=%f",
+ dfatool_printf(" latency_write1_us=%f latency_write2_us=%f latency_write3_us=%f latency_write4_us=%f latency_kernel1_us=%f latency_kernel2_us=%f",
timer.time[3],
timer.time[4],
timer.time[5],
timer.time[7],
timer.time[6],
timer.time[8]);
- printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
+ dfatool_printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
input_size * sizeof(T) / timer.time[0],
input_size * sizeof(T) / (timer.time[6] + timer.time[8]),
input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
+ dfatool_printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
input_size * sizeof(T) / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size * sizeof(T) / (timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
+ dfatool_printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
input_size / timer.time[0],
input_size / (timer.time[6] + timer.time[8]),
input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
- printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n",
+ dfatool_printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n",
input_size / (timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size / (timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]),
input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6] + timer.time[7] + timer.time[8] + timer.time[9]));
diff --git a/TRNS/support/common.h b/TRNS/include/common.h
index 2ba56c5..6a94c62 100755..100644
--- a/TRNS/support/common.h
+++ b/TRNS/include/common.h
@@ -14,16 +14,18 @@
// Data type
#define T int64_t
+enum kernels {
+ kernel1 = 0,
+ kernel2 = 1,
+ nr_kernels = 2,
+};
+
// Structures used by both the host and the dpu to communicate information
typedef struct {
uint32_t m;
uint32_t n;
uint32_t M_;
- enum kernels {
- kernel1 = 0,
- kernel2 = 1,
- nr_kernels = 2,
- } kernel;
+ enum kernels kernel;
} dpu_arguments_t;
#ifndef ENERGY
diff --git a/TRNS/include/dfatool_host.ah b/TRNS/include/dfatool_host.ah
new file mode 100644
index 0000000..72978cc
--- /dev/null
+++ b/TRNS/include/dfatool_host.ah
@@ -0,0 +1,36 @@
+#pragma once
+
+#include <sys/time.h>
+#include "dfatool_host_dpu.ah"
+
+aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
+
+ unsigned int n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile;
+ unsigned int element_size;
+
+ virtual int getKernel() { return kernel; }
+
+ DfatoolHostTiming() {
+ element_size = sizeof(T);
+ }
+
+ advice call("% input_params(...)") : after() {
+ Params* p = tjp->result();
+ /*
+ * Input: (M_ * m) × (N_ * n) matrix
+ */
+ n_rows_outer = p->M_;
+ n_rows_tile = p->m;
+ n_cols_outer = p->N_;
+ n_cols_tile = p->n;
+ printf("[>>] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
+ }
+
+ advice call("% trns_host(...)") : after() {
+ printf("[--] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
+ }
+
+ advice execution("% main(...)") : after() {
+ printf("[<<] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
+ }
+};
diff --git a/TRNS/support/params.h b/TRNS/include/params.h
index 6b7e6f2..385490e 100644
--- a/TRNS/support/params.h
+++ b/TRNS/include/params.h
@@ -21,7 +21,7 @@ static void usage() {
"\n -h help"
"\n -w <W> # of untimed warmup iterations (default=1)"
"\n -e <E> # of timed repetition iterations (default=3)"
- "\n -x <X> Weak (0) or strong (1) scaling (default=0)"
+ "\n -x <X> Weak (0) or strong (1) scaling (default=1)"
"\n"
"\nBenchmark-specific options:"
"\n -m <I> m (default=16 elements)"
@@ -39,7 +39,7 @@ struct Params input_params(int argc, char **argv) {
p.n = 8;
p.n_warmup = 1;
p.n_reps = 3;
- p.exp = 0;
+ p.exp = 1;
int opt;
while((opt = getopt(argc, argv, "hw:e:x:m:n:o:p:")) >= 0) {
diff --git a/TRNS/include/timer.h b/TRNS/include/timer.h
new file mode 100644
index 0000000..8d5c3d5
--- /dev/null
+++ b/TRNS/include/timer.h
@@ -0,0 +1,5 @@
+#pragma once
+
+#define N_TIMERS 10
+#include "../../include/timer_base.h"
+#undef N_TIMERS
diff --git a/TRNS/run-fgbs24a.sh b/TRNS/run-fgbs24a.sh
deleted file mode 100755
index 6ba8993..0000000
--- a/TRNS/run-fgbs24a.sh
+++ /dev/null
@@ -1,29 +0,0 @@
-#!/bin/bash
-
-set -e
-
-mkdir -p $(hostname)
-
-# 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 TRNS strong-full (dfatool fgbs24a edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 2304 2048 2543; do
- for nr_tasklets in 16; do
- echo
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
- # upstream uses -p 2048, but then the number of DPUs is always constant...
- timeout --foreground -k 1m 180m bin/host_code -w 0 -e 100 -p $nr_dpus -o 12288 -x 1 || true
- fi
- done
-done
-echo "Completed at $(date)"
-) | tee "$(hostname)/fgbs24a.txt"
diff --git a/TRNS/run-paper-strong-full.sh b/TRNS/run-paper-strong-full.sh
deleted file mode 100755
index 9d3792c..0000000
--- a/TRNS/run-paper-strong-full.sh
+++ /dev/null
@@ -1,30 +0,0 @@
-#!/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 TRNS strong-full (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-# >2048 is not in upstream
-for nr_dpus in 2543 2304 256 512 1024 2048; do
- for nr_tasklets in 1 2 4 8 16; do
- echo
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
- # upstream uses -p 2048, but then the number of DPUs is always constant...
- timeout --foreground -k 1m 90m bin/host_code -w 0 -e 40 -p $nr_dpus -o 12288 -x 1 || true
- fi
- done
-done
-
-echo "Completed at $(date)"
-
-) | tee "log-$(hostname)-prim-strong-full.txt"
diff --git a/TRNS/run-paper-strong-rank.sh b/TRNS/run-paper-strong-rank.sh
deleted file mode 100755
index f5f00cb..0000000
--- a/TRNS/run-paper-strong-rank.sh
+++ /dev/null
@@ -1,29 +0,0 @@
-#!/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 TRNS 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}; then
- # upstream uses -p 64, but then the number of DPUs is always constant...
- timeout --foreground -k 1m 60m bin/host_code -w 0 -e 40 -p $nr_dpus -o 12288 -x 1 || true
- fi
- done
-done
-
-echo "Completed at $(date)"
-
-) | tee "log-$(hostname)-prim-strong-rank.txt"
diff --git a/TRNS/run-paper-weak.sh b/TRNS/run-paper-weak.sh
deleted file mode 100755
index f02d7d6..0000000
--- a/TRNS/run-paper-weak.sh
+++ /dev/null
@@ -1,28 +0,0 @@
-#!/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 TRNS 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}; then
- timeout --foreground -k 1m 60m bin/host_code -w 0 -e 40 -p 1 -o 12288 -x 0 || true
- fi
- done
-done | tee log-paper-weak.txt
-
-echo "Completed at $(date)"
-
-) | tee "log-$(hostname)-prim-weak.txt"
diff --git a/TRNS/run-rank.sh b/TRNS/run-rank.sh
deleted file mode 100755
index 00f6898..0000000
--- a/TRNS/run-rank.sh
+++ /dev/null
@@ -1,32 +0,0 @@
-#!/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: input size (number of elements, not number of bytes!)
-
-(
-
-echo "prim-benchmarks TRNS (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 1 4 8 16 32 48 64; do
- for nr_tasklets in 8 12 16; do
- # 12288 run-paper-weak, run-paper-strong-full
- for i in 12288; do
- echo
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
- # upstream uses -p 2048 in strong-full, but then the number of DPUs is always constant...
- timeout --foreground -k 1m 90m bin/host_code -w 0 -e 40 -p 1 -o 12288 -x 0 || true
- fi
- done
- done
-done
-
-echo "Completed at $(date)"
-
-) | tee "log-$(hostname)-rank.txt"
diff --git a/TRNS/run.sh b/TRNS/run.sh
deleted file mode 100755
index 8d574a9..0000000
--- a/TRNS/run.sh
+++ /dev/null
@@ -1,32 +0,0 @@
-#!/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: input size (number of elements, not number of bytes!)
-
-(
-
-echo "prim-benchmarks TRNS (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 2542 2304 1 4 8 16 32 64 128 256 512 768 1024 1536 2048; do
- for nr_tasklets in 8 12 16; do
- # 12288 run-paper-weak, run-paper-strong-full
- for i in 12288; do
- echo
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
- # upstream uses -p 2048 in strong-full, but then the number of DPUs is always constant...
- timeout --foreground -k 1m 90m bin/host_code -w 0 -e 40 -p $nr_dpus -o 12288 -x 1 || true
- fi
- done
- done
-done
-
-echo "Completed at $(date)"
-
-) | tee "log-$(hostname).txt"
diff --git a/TRNS/support/timer.h b/TRNS/support/timer.h
deleted file mode 100755
index 786c687..0000000
--- a/TRNS/support/timer.h
+++ /dev/null
@@ -1,66 +0,0 @@
-/*
- * Copyright (c) 2016 University of Cordoba and University of Illinois
- * All rights reserved.
- *
- * Developed by: IMPACT Research Group
- * University of Cordoba and University of Illinois
- * http://impact.crhc.illinois.edu/
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * with the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * > Redistributions of source code must retain the above copyright notice,
- * this list of conditions and the following disclaimers.
- * > Redistributions in binary form must reproduce the above copyright
- * notice, this list of conditions and the following disclaimers in the
- * documentation and/or other materials provided with the distribution.
- * > Neither the names of IMPACT Research Group, University of Cordoba,
- * University of Illinois nor the names of its contributors may be used
- * to endorse or promote products derived from this Software without
- * specific prior written permission.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
- * THE SOFTWARE.
- *
- */
-
-#include <sys/time.h>
-
-typedef struct Timer{
-
- struct timeval startTime[10];
- struct timeval stopTime[10];
- double time[10];
-
-}Timer;
-
-void start(Timer *timer, int i, int rep) {
- if(rep == 0) {
- timer->time[i] = 0.0;
- }
- gettimeofday(&timer->startTime[i], NULL);
-}
-
-void stop(Timer *timer, int i) {
- gettimeofday(&timer->stopTime[i], NULL);
- timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
- (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);
-}
-
-void print(Timer *timer, int i, int REP) { printf("Time (ms): %f\t", timer->time[i] / (1000 * REP)); }
-
-void printall(Timer *timer, int maxt) {
- for (int i = 0; i <= maxt; i++) {
- printf(" timer%d_us=%f", i, timer->time[i]);
- }
- printf("\n");
-}
generated by cgit v1.2.3 (git 2.39.1) at 2025-07-01 06:13:30 +0000