port UNI to dfatool

9 files changed, 160 insertions, 65 deletions

diff --git a/UNI/Makefile b/UNI/Makefile
index 65c7962..f9cdf0b 100644
--- a/UNI/Makefile
+++ b/UNI/Makefile
@@ -1,45 +1,37 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_TASKLETS ?= 16
 BL ?= 10
 NR_DPUS ?= 1
 ENERGY ?= 0
 
-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}/host_code
-DPU_TARGET := ${BUILDDIR}/dpu_code
-
 COMMON_INCLUDES := support
-HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
-DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
-
-.PHONY: all clean test
-
-__dirs := $(shell mkdir -p ${BUILDDIR})
+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=${NR_DPUS} -DBL=${BL} -DENERGY=${ENERGY} 
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} 
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+QUIET = @
 
-${CONF}:
-	$(RM) $(call conf_filename,*,*)
-	touch ${CONF}
+ifdef verbose
+	QUIET =
+endif
 
-${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	$(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+all: bin/host_code bin/dpu_code
 
-${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+bin:
+	${QUIET} mkdir -p bin
+
+bin/host_code: ${HOST_SOURCES} ${COMMON_INCLUDES} bin
+	${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+
+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}
+	bin/host_code
+
+.PHONY: all clean test
diff --git a/UNI/baselines/cpu/Makefile b/UNI/baselines/cpu/Makefile
index 10e4f9b..ec3f403 100644
--- a/UNI/baselines/cpu/Makefile
+++ b/UNI/baselines/cpu/Makefile
@@ -16,13 +16,15 @@ uni_O2: app_baseline.c
 run: uni
 	./uni -i 1258291200 -t 4
 
+# -e 20 was not part of the upstream instructions
+
 .PHONY: run_O0
 run_O0: uni_O0
-	./uni_O0 -i 1258291200 -t 4
+	./uni_O0 -i 1258291200 -t 4 -e 20
 
 .PHONY: run_O2
 run_O2: uni_O2
-	./uni_O2 -i 1258291200 -t 4
+	./uni_O2 -i 1258291200 -t 4 -e 20
 
 .PHONY: clean
 clean:
diff --git a/UNI/baselines/cpu/app_baseline.c b/UNI/baselines/cpu/app_baseline.c
index 39f225a..f4f303f 100644
--- a/UNI/baselines/cpu/app_baseline.c
+++ b/UNI/baselines/cpu/app_baseline.c
@@ -32,7 +32,7 @@ static T *create_test_file(unsigned int nr_elements) {
     B = (T*) malloc(nr_elements * sizeof(T));
     C = (T*) malloc(nr_elements * sizeof(T));
 
-    printf("nr_elements\t%u\t", nr_elements);
+    //printf("nr_elements\t%u\t", nr_elements);
     for (int i = 0; i < nr_elements; i++) {
         //A[i] = (unsigned int) (rand());
         //A[i] = i+1;
@@ -141,16 +141,12 @@ int main(int argc, char **argv) {
         nr_threads++;
 
         if (rep >= p.n_warmup) {
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_MBps=%f\n",
-                nr_threads, XSTR(T), file_size,
-                file_size * 2 * sizeof(T) / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_MOpps=%f\n",
+            printf("[::] UNI CPU | n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_MBps=%f",
                 nr_threads, XSTR(T), file_size,
+                file_size * sizeof(T) / timer.time[0]);
+            printf(" throughput_MOpps=%f",
                 file_size / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d | ",
-                nr_threads, XSTR(T), file_size);
             printall(&timer, 0);
         }
     }
diff --git a/UNI/baselines/cpu/run-opti.sh b/UNI/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..0a3a4a3
--- /dev/null
+++ b/UNI/baselines/cpu/run-opti.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+make run_O0 | sed 's/CPU/CPU O0/' | tee "${HOST}-O0.txt"
+
+make run_O2 | sed 's/CPU/CPU O2/' | tee "${HOST}-O2.txt"
diff --git a/UNI/baselines/cpu/run.sh b/UNI/baselines/cpu/run.sh
index 3c99ad2..37a6a9c 100755
--- a/UNI/baselines/cpu/run.sh
+++ b/UNI/baselines/cpu/run.sh
@@ -2,16 +2,23 @@
 
 set -e
 
+HOST="$(hostname)"
+
+echo $HOST
+
+(
+
 echo "prim-benchmarks UNI CPU (dfatool edition)"
 echo "Started at $(date)"
 echo "Revision $(git describe --always)"
 
-for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+for nr_threads in 88 64 44 1 2 4 6 8 12 16 20 24 32; do
 	for i in 19660800 1258291200 629145600 314572800 157286400 78643200 39321600; do
 		for dt in int8_t int16_t int32_t int64_t float double; do
 			if make -B TYPE=${dt}; then
-				timeout -k 1m 30m ./uni -i ${i} -w 0 -e 100 -t ${nr_threads} || true
+				timeout --foreground -k 1m 60m ./uni -i ${i} -w 0 -e 100 -t ${nr_threads} || true
 			fi
 		done
 	done
 done
+) | tee "${HOST}-explore.txt"
diff --git a/UNI/host/app.c b/UNI/host/app.c
index 1b91bba..e068d6b 100644
--- a/UNI/host/app.c
+++ b/UNI/host/app.c
@@ -22,6 +22,9 @@
 #define DPU_BINARY "./bin/dpu_code"
 #endif
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 #if ENERGY
 #include <dpu_probe.h>
 #endif
@@ -106,14 +109,14 @@ 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);
         total_count = unique_host(C, A, input_size);
         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
         const unsigned int input_size_dpu = input_size_dpu_round;
         unsigned int kernel = 0;
@@ -134,7 +137,7 @@ 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
@@ -168,7 +171,7 @@ int main(int argc, char **argv) {
         accum = 0;
 
         if(rep >= p.n_warmup)
-            start(&timer, 3, rep - p.n_warmup);
+            start(&timer, 3, 0);
         // PARALLEL RETRIEVE TRANSFER
         dpu_results_t* results_retrieve[nr_of_dpus];
 
@@ -212,7 +215,7 @@ int main(int argc, char **argv) {
 
         i = 0;
         if(rep >= p.n_warmup)
-            start(&timer, 4, rep - p.n_warmup);
+            start(&timer, 4, 0);
         DPU_FOREACH (dpu_set, dpu) {
             // Copy output array
             DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferC + results_scan[i] - offset_scan[i], results[i].t_count * sizeof(T)));
@@ -227,6 +230,39 @@ int main(int argc, char **argv) {
         free(offset);
         free(offset_scan);
 
+        // Check output
+        bool status = true;
+        if(accum != total_count) status = false;
+#if PRINT
+        printf("accum %u, total_count %u\n", accum, total_count);
+#endif
+        for (i = 0; i < accum; i++) {
+            if(C[i] != bufferC[i]){ 
+                status = false;
+#if PRINT
+                printf("%d: %lu -- %lu\n", i, C[i], bufferC[i]);
+#endif
+            }
+        }
+        if (status) {
+            printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+            if (rep >= p.n_warmup) {
+                printf("[::] UNI NMC | n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
+                    "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f",
+                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
+                    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] + timer.time[4]));
+                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] + timer.time[4]));
+                printall(&timer, 4);
+            }
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
+
     }
 
     // Print timing results
@@ -247,25 +283,6 @@ int main(int argc, char **argv) {
     printf("DPU Energy (J): %f\t", energy);
 #endif	
 
-    // Check output
-    bool status = true;
-    if(accum != total_count) status = false;
-#if PRINT
-    printf("accum %u, total_count %u\n", accum, total_count);
-#endif
-    for (i = 0; i < accum; i++) {
-        if(C[i] != bufferC[i]){ 
-            status = false;
-#if PRINT
-            printf("%d: %lu -- %lu\n", i, C[i], bufferC[i]);
-#endif
-        }
-    }
-    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);
@@ -273,5 +290,5 @@ int main(int argc, char **argv) {
     free(C2);
     DPU_ASSERT(dpu_free(dpu_set));
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/UNI/run-paper-strong-full.sh b/UNI/run-paper-strong-full.sh
new file mode 100755
index 0000000..9e5ac97
--- /dev/null
+++ b/UNI/run-paper-strong-full.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 UNI strong-full (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 256 512 1024 2048; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 251658240 -x 1 || true
+		fi
+	done
+done | tee log-paper-strong-full.txt
diff --git a/UNI/run-paper-strong-rank.sh b/UNI/run-paper-strong-rank.sh
new file mode 100755
index 0000000..6cffd65
--- /dev/null
+++ b/UNI/run-paper-strong-rank.sh
@@ -0,0 +1,26 @@
+#!/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 UNI strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream config space
+for nr_dpus in 512 256 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 3932160 -x 1 || true
+		fi
+	done
+done
+) | tee log-paper-strong-rank.txt
diff --git a/UNI/run-paper-weak.sh b/UNI/run-paper-weak.sh
new file mode 100755
index 0000000..e9b5e04
--- /dev/null
+++ b/UNI/run-paper-weak.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 UNI 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} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 3932160 -x 0 || true
+		fi
+	done
+done | tee log-paper-weak.txt