summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--BFS/Makefile29
-rw-r--r--BFS/dpu/task.c2
-rw-r--r--BFS/host/app.c60
-rw-r--r--BFS/host/mram-management.h4
-rw-r--r--BFS/include/common.h (renamed from BFS/support/common.h)0
-rw-r--r--BFS/include/dfatool_host.ah27
-rw-r--r--BFS/include/graph.h (renamed from BFS/support/graph.h)0
-rw-r--r--BFS/include/params.h (renamed from BFS/support/params.h)0
-rw-r--r--BFS/include/timer.h8
-rw-r--r--BFS/include/utils.h (renamed from BFS/support/utils.h)0
-rw-r--r--BFS/support/timer.h31
-rw-r--r--include/dfatool_host_dpu.ah20
-rw-r--r--include/timer_base.h11
13 files changed, 124 insertions, 68 deletions
diff --git a/BFS/Makefile b/BFS/Makefile
index d43202f..598de50 100644
--- a/BFS/Makefile
+++ b/BFS/Makefile
@@ -4,17 +4,34 @@ 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=${NR_DPUS} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD}
+aspectc ?= 0
+aspectc_timing ?= 0
+dfatool_timing ?= 1
+
+HOST_CC := ${CC}
+
+COMMON_FLAGS := -Wall -Wextra -g -Iinclude
+HOST_FLAGS := ${COMMON_FLAGS} -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -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 ${UPMEM_HOME}/bin/clang++ -p . --Xcompiler
+else
+ HOST_FLAGS += -std=c11
+endif
+
QUIET = @
-ifdef verbose
+ifeq (${verbose}, 1)
QUIET =
endif
@@ -27,7 +44,9 @@ bin/dpu_code: ${DPU_SOURCES} ${COMMON_INCLUDES} bin
${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
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
clean:
${QUIET}rm -rf bin
diff --git a/BFS/dpu/task.c b/BFS/dpu/task.c
index 44ec214..5275047 100644
--- a/BFS/dpu/task.c
+++ b/BFS/dpu/task.c
@@ -12,7 +12,7 @@
#include <perfcounter.h>
#include "dpu-utils.h"
-#include "../support/common.h"
+#include "common.h"
BARRIER_INIT(my_barrier, NR_TASKLETS);
diff --git a/BFS/host/app.c b/BFS/host/app.c
index 9ba7ffb..763ff7a 100644
--- a/BFS/host/app.c
+++ b/BFS/host/app.c
@@ -3,9 +3,24 @@
* BFS Host Application Source File
*
*/
+#if ASPECTC
+extern "C" {
+#endif
+
#include <dpu.h>
#include <dpu_log.h>
+#ifndef ENERGY
+#define ENERGY 0
+#endif
+#if ENERGY
+#include <dpu_probe.h>
+#endif
+
+#if ASPECTC
+}
+#endif
+
#include <assert.h>
#include <getopt.h>
#include <stdio.h>
@@ -14,18 +29,11 @@
#include <unistd.h>
#include "mram-management.h"
-#include "../support/common.h"
-#include "../support/graph.h"
-#include "../support/params.h"
-#include "../support/timer.h"
-#include "../support/utils.h"
-
-#ifndef ENERGY
-#define ENERGY 0
-#endif
-#if ENERGY
-#include <dpu_probe.h>
-#endif
+#include "common.h"
+#include "graph.h"
+#include "params.h"
+#include "timer.h"
+#include "utils.h"
#define DPU_BINARY "./bin/dpu_code"
@@ -44,10 +52,6 @@ int main(int argc, char **argv)
double tenergy = 0;
#endif
- 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
struct dpu_set_t dpu_set, dpu;
uint32_t numDPUs, numRanks;
@@ -59,7 +63,7 @@ int main(int argc, char **argv)
#if WITH_ALLOC_OVERHEAD
stopTimer(&timer, 0);
#else
- timer.time[0] = 0;
+ zeroTimer(&timer, 0);
#endif
#if WITH_LOAD_OVERHEAD
@@ -69,7 +73,7 @@ int main(int argc, char **argv)
#if WITH_LOAD_OVERHEAD
stopTimer(&timer, 0);
#else
- timer.time[1] = 0;
+ zeroTimer(&timer, 1);
#endif
DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &numDPUs));
@@ -86,10 +90,10 @@ int main(int argc, char **argv)
uint32_t numNodes = csrGraph.numNodes;
uint32_t *nodePtrs = csrGraph.nodePtrs;
uint32_t *neighborIdxs = csrGraph.neighborIdxs;
- uint32_t *nodeLevel = calloc(numNodes, sizeof(uint32_t)); // Node's BFS level (initially all 0 meaning not reachable)
- uint64_t *visited = calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
- uint64_t *currentFrontier = calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
- uint64_t *nextFrontier = calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
+ uint32_t *nodeLevel = (uint32_t*)calloc(numNodes, sizeof(uint32_t)); // Node's BFS level (initially all 0 meaning not reachable)
+ uint64_t *visited = (uint64_t*)calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
+ uint64_t *currentFrontier = (uint64_t*)calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
+ uint64_t *nextFrontier = (uint64_t*)calloc(numNodes / 64, sizeof(uint64_t)); // Bit vector with one bit per node
setBit(nextFrontier[0], 0); // Initialize frontier to first node
uint32_t level = 1;
@@ -325,7 +329,7 @@ int main(int argc, char **argv)
// Calculating result on CPU
PRINT_INFO(p.verbosity >= 1, "Calculating result on CPU");
- uint32_t *nodeLevelReference = calloc(numNodes, sizeof(uint32_t)); // Node's BFS level (initially all 0 meaning not reachable)
+ uint32_t *nodeLevelReference = (uint32_t*) calloc(numNodes, sizeof(uint32_t)); // Node's BFS level (initially all 0 meaning not reachable)
memset(nextFrontier, 0, numNodes / 64 * sizeof(uint64_t));
setBit(nextFrontier[0], 0); // Initialize frontier to first node
nextFrontierEmpty = 0;
@@ -395,7 +399,7 @@ int main(int argc, char **argv)
#if WITH_FREE_OVERHEAD
stopTimer(&timer, 7);
#else
- timer.time[7] = 0;
+ zeroTimer(&timer, 7);
#endif
// Verify the result
@@ -412,9 +416,9 @@ int main(int argc, char **argv)
}
if (status) {
- printf
+ dfatool_printf
("[::] BFS-UMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%d "
- "| throughput_pim_MBps=%f throughput_MBps=%f", numDPUs,
+ "| throughput_pim_MBps=%f throughput_MBps=%f", numDPUs, numRanks,
NR_TASKLETS, "uint32_t", numNodes,
numNodes * sizeof(uint32_t) / (timer.time[2] +
timer.time[3]),
@@ -423,12 +427,12 @@ int main(int argc, char **argv)
timer.time[2] +
timer.time[3] +
timer.time[4]));
- printf(" throughput_pim_MOpps=%f throughput_MOpps=%f",
+ dfatool_printf(" throughput_pim_MOpps=%f throughput_MOpps=%f",
numNodes / (timer.time[2] + timer.time[3]),
numNodes / (timer.time[0] + timer.time[1] +
timer.time[2] + timer.time[3] +
timer.time[4]));
- printf
+ dfatool_printf
(" latency_alloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_sync_us=%f latency_read_us=%f latency_cpu_us=%f latency_free_us=%f\n",
timer.time[0], timer.time[1], timer.time[2], timer.time[3],
timer.time[4], timer.time[5], timer.time[6],
diff --git a/BFS/host/mram-management.h b/BFS/host/mram-management.h
index f2ee031..61616da 100644
--- a/BFS/host/mram-management.h
+++ b/BFS/host/mram-management.h
@@ -2,8 +2,8 @@
#ifndef _MRAM_MANAGEMENT_H_
#define _MRAM_MANAGEMENT_H_
-#include "../support/common.h"
-#include "../support/utils.h"
+#include "common.h"
+#include "utils.h"
#define DPU_CAPACITY (64 << 20) // A DPU's capacity is 64 MiB
diff --git a/BFS/support/common.h b/BFS/include/common.h
index 5f2aa0d..5f2aa0d 100644
--- a/BFS/support/common.h
+++ b/BFS/include/common.h
diff --git a/BFS/include/dfatool_host.ah b/BFS/include/dfatool_host.ah
new file mode 100644
index 0000000..592e6ec
--- /dev/null
+++ b/BFS/include/dfatool_host.ah
@@ -0,0 +1,27 @@
+#pragma once
+
+#include <sys/time.h>
+#include "dfatool_host_dpu.ah"
+
+aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
+
+ virtual int getKernel() { return 1; }
+
+ DfatoolHostTiming() {
+ element_size = sizeof(uint32_t);
+ }
+
+ advice call("% input_params(...)"): after() {
+ printf("[>>] BFS | n_dpus=%u\n", NR_DPUS);
+ }
+
+ advice call("% coo2csr(...)") : after() {
+ struct CSRGraph *g = tjp->result();
+ input_size = g->numNodes;
+ printf("[--] BFS | n_dpus=%u n_elements=%lu\n", NR_DPUS, input_size);
+ }
+
+ advice execution("% main(...)") : after() {
+ printf("[<<] BFS | n_dpus=%u n_elements=%lu\n", NR_DPUS, input_size);
+ }
+};
diff --git a/BFS/support/graph.h b/BFS/include/graph.h
index 2a19f67..2a19f67 100644
--- a/BFS/support/graph.h
+++ b/BFS/include/graph.h
diff --git a/BFS/support/params.h b/BFS/include/params.h
index f9169bc..f9169bc 100644
--- a/BFS/support/params.h
+++ b/BFS/include/params.h
diff --git a/BFS/include/timer.h b/BFS/include/timer.h
new file mode 100644
index 0000000..e85490f
--- /dev/null
+++ b/BFS/include/timer.h
@@ -0,0 +1,8 @@
+#pragma once
+
+#define N_TIMERS 8
+#define startTimer start
+#define stopTimer stop
+#define zeroTimer zero
+#include "../../include/timer_base.h"
+#undef N_TIMERS
diff --git a/BFS/support/utils.h b/BFS/include/utils.h
index ccd8fbd..ccd8fbd 100644
--- a/BFS/support/utils.h
+++ b/BFS/include/utils.h
diff --git a/BFS/support/timer.h b/BFS/support/timer.h
deleted file mode 100644
index 63b5567..0000000
--- a/BFS/support/timer.h
+++ /dev/null
@@ -1,31 +0,0 @@
-
-#ifndef _TIMER_H_
-#define _TIMER_H_
-
-#include <stdio.h>
-#include <sys/time.h>
-
-typedef struct Timer {
- struct timeval startTime[8];
- struct timeval stopTime[8];
- double time[8];
-} Timer;
-
-static void startTimer(Timer *timer, int i, int rep)
-{
- if (rep == 0) {
- timer->time[i] = 0.0;
- }
- gettimeofday(&timer->startTime[i], NULL);
-}
-
-static void stopTimer(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);
-}
-
-#endif
diff --git a/include/dfatool_host_dpu.ah b/include/dfatool_host_dpu.ah
index 0e03d31..010c18c 100644
--- a/include/dfatool_host_dpu.ah
+++ b/include/dfatool_host_dpu.ah
@@ -12,6 +12,8 @@ aspect DfatoolHostDPUTiming {
double const M_to_Mi = 1.048576; /* 2^20 / 1e6 */
+ virtual int getKernel() = 0;
+
advice call("% dpu_get_nr_dpus(...)") : after() {
n_dpus = **(tjp->arg<1>());
}
@@ -79,7 +81,7 @@ aspect DfatoolHostDPUTiming {
tjp->filename(),
tjp->line(),
n_dpus, n_ranks,
- kernel + 1,
+ getKernel(),
input_size,
input_size / n_dpus,
latency_us,
@@ -104,6 +106,22 @@ aspect DfatoolHostDPUTiming {
);
}
+ advice call("% dpu_copy_from(...)") : around() {
+ size_t payload_size = *(tjp->arg<4>());
+ gettimeofday(&starttime, NULL);
+ tjp->proceed();
+ gettimeofday(&stoptime, NULL);
+ double time_us = (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec);
+ printf("[::] dpu_copy_from @ %s:%d | n_dpus=%u n_ranks=%u payload_B=%lu | latency_us=%f throughput_MiBps=%f\n",
+ tjp->filename(),
+ tjp->line(),
+ n_dpus, n_ranks,
+ payload_size,
+ time_us,
+ payload_size / (time_us * M_to_Mi)
+ );
+ }
+
advice call("% dpu_push_xfer(...)") : around() {
size_t payload_size = *(tjp->arg<4>());
gettimeofday(&starttime, NULL);
diff --git a/include/timer_base.h b/include/timer_base.h
index fed8bec..dcc73b7 100644
--- a/include/timer_base.h
+++ b/include/timer_base.h
@@ -31,6 +31,11 @@ void stop(Timer *timer, int i)
(timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);
}
+void zero(Timer *timer, int i)
+{
+ timer->time[0] = 0;
+}
+
#else
#define dfatool_printf(fmt, ...) do {} while (0)
@@ -50,4 +55,10 @@ void stop(Timer *timer, int i)
(void)i;
}
+void zero(Timer *timer, int i)
+{
+ (void)timer;
+ (void)i;
+}
+
#endif
generated by cgit v1.2.3 (git 2.39.1) at 2025-05-22 20:34:54 +0000