port HST-L to dfatool

6 files changed, 116 insertions, 40 deletions

diff --git a/HST-L/Makefile b/HST-L/Makefile
index fb1ce5a..1888b0a 100644
--- a/HST-L/Makefile
+++ b/HST-L/Makefile
@@ -1,46 +1,38 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
-NR_TASKLETS ?= 16
-BL ?= 8
 NR_DPUS ?= 1
+NR_TASKLETS ?= 16
 NR_HISTO ?= 1
+BL ?= 10
 ENERGY ?= 0
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_NR_DPUS_$(4).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${NR_DPUS})
-
-HOST_TARGET := ${BUILDDIR}/host_code
-DPU_TARGET := ${BUILDDIR}/dpu_code
-
 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
+COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES} -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -DNR_HISTO=${NR_HISTO}
+HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DENERGY=${ENERGY}
+DPU_FLAGS := ${COMMON_FLAGS} -O2
 
-__dirs := $(shell mkdir -p ${BUILDDIR})
+QUIET = @
 
-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} -DENERGY=${ENERGY}
-DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -DNR_HISTO=${NR_HISTO} 
+ifdef verbose
+	QUIET =
+endif
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+all: bin/host_code bin/dpu_code
 
-${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/host_code: ${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/dpu_code: ${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}
+	${QUIET}bin/host_code
+
+.PHONY: all clean test
diff --git a/HST-L/host/app.c b/HST-L/host/app.c
index 09ab578..b9c07f9 100644
--- a/HST-L/host/app.c
+++ b/HST-L/host/app.c
@@ -23,6 +23,9 @@
 #define DPU_BINARY "./bin/dpu_code"
 #endif
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 #if ENERGY
 #include <dpu_probe.h>
 #endif
@@ -40,7 +43,7 @@ static void read_input(T* A, const Params p) {
 
     // Open input file
     unsigned short temp;
-    sprintf(dctFileName, p.file_name);
+    sprintf(dctFileName, "%s", p.file_name);
     if((File = fopen(dctFileName, "rb")) != NULL) {
         for(unsigned int y = 0; y < p.input_size; y++) {
             fread(&temp, sizeof(unsigned short), 1, File);
@@ -138,14 +141,14 @@ int main(int argc, char **argv) {
 
         // Compute output on CPU (performance comparison and verification purposes)
         if(rep >= p.n_warmup)
-            start(&timer, 0, rep - p.n_warmup);
+            start(&timer, 0, 0);
         histogram_host(histo_host, A, p.bins, p.input_size, 1, nr_of_dpus);
         if(rep >= p.n_warmup)
             stop(&timer, 0);
 
         printf("Load input data\n");
         if(rep >= p.n_warmup)
-            start(&timer, 1, rep - p.n_warmup);
+            start(&timer, 1, 0);
         // Input arguments
         unsigned int kernel = 0;
         i = 0;
@@ -177,11 +180,12 @@ int main(int argc, char **argv) {
         printf("Run program on DPU(s) \n");
         // Run DPU kernel
         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
         }
+
         DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
         if(rep >= p.n_warmup) {
             stop(&timer, 2);
@@ -205,22 +209,36 @@ int main(int argc, char **argv) {
         printf("Retrieve results\n");
         i = 0;
         if(rep >= p.n_warmup)
-            start(&timer, 3, rep - p.n_warmup);
+            start(&timer, 3, 0);
         // PARALLEL RETRIEVE TRANSFER
         DPU_FOREACH(dpu_set, dpu, i) {
             DPU_ASSERT(dpu_prepare_xfer(dpu, histo + p.bins * i));
         }
         DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu_8bytes * sizeof(T), p.bins * sizeof(unsigned int), DPU_XFER_DEFAULT));
-		
+
         // Final histogram merging
         for(i = 1; i < nr_of_dpus; i++){
             for(unsigned int j = 0; j < p.bins; j++){
                 histo[j] += histo[j + i * p.bins];
-            }			
-        }		
+            }
+        }
         if(rep >= p.n_warmup)
             stop(&timer, 3);
 
+        if (rep >= p.n_warmup) {
+            printf("[::] HST-L NMC | n_dpus=%d n_tasklets=%d e_type=%s n_elements=%u n_bins=%d "
+                "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f",
+                nr_of_dpus, NR_TASKLETS, XSTR(T), input_size, p.bins,
+                input_size * sizeof(T) / timer.time[0],
+                input_size * sizeof(T) / timer.time[2],
+                input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
+            printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f",
+                input_size / timer.time[0],
+                input_size / timer.time[2],
+                input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
+            printall(&timer, 3);
+        }
+
     }
 
     // Print timing results
@@ -237,8 +255,7 @@ int main(int argc, char **argv) {
     double energy;
     DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
     printf("DPU Energy (J): %f\t", energy);
-    #endif	
-
+    #endif
 
     // Check output
     bool status = true;
diff --git a/HST-L/run-paper-strong-full.sh b/HST-L/run-paper-strong-full.sh
new file mode 100755
index 0000000..0108d40
--- /dev/null
+++ b/HST-L/run-paper-strong-full.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks HST-S strong-full (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# >2048 are not part of upstream
+for nr_dpus in 2543 2304 256 512 1024 2048; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -b 256 -x 2 || true
+		fi
+	done
+done
+) | tee log-paper-strong-full.txt
diff --git a/HST-L/run-paper-strong-rank.sh b/HST-L/run-paper-strong-rank.sh
new file mode 100755
index 0000000..f2f80b1
--- /dev/null
+++ b/HST-L/run-paper-strong-rank.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks HST-S strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# >64 are not part of upstream config space
+for nr_dpus in 128 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -b 256 -x 1 || true
+		fi
+	done
+done
+) | tee log-paper-strong-rank.txt
diff --git a/HST-L/run-paper-weak.sh b/HST-L/run-paper-weak.sh
new file mode 100755
index 0000000..3ddd801
--- /dev/null
+++ b/HST-L/run-paper-weak.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks HST-S weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# upstream does not include >64
+for nr_dpus in 256 512 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -b 256 -x 0 || true
+		fi
+	done
+done
+) | tee log-paper-weak.txt
diff --git a/HST-L/support/timer.h b/HST-L/support/timer.h
index eedc385..5c00213 100755
--- a/HST-L/support/timer.h
+++ b/HST-L/support/timer.h
@@ -57,3 +57,10 @@ void stop(Timer *timer, int i) {
 }

 

 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");

+}