RED: update dfatool integration; port nmc version; add eval scripts

8 files changed, 142 insertions, 59 deletions

diff --git a/RED/Makefile b/RED/Makefile
index ffdeecc..fc8298d 100644
--- a/RED/Makefile
+++ b/RED/Makefile
@@ -1,6 +1,3 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_TASKLETS ?= 16
 BL ?= 10
 NR_DPUS ?= 1
@@ -10,40 +7,35 @@ TYPE ?= INT64
 ENERGY ?= 0
 PERF ?= 0
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_VERSION_$(4)_SYNC_$(5)_TYPE_$(6).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${VERSION},${SYNC},${TYPE})
-
-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} -D${VERSION} -D${SYNC} -D${TYPE} -DENERGY=${ENERGY} -DPERF=${PERF}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${VERSION} -D${SYNC} -D${TYPE} -DPERF=${PERF}
 
-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} -w 0 -e 1 -i 6553600 -x 1
+	${QUIET}bin/host_code -w 0 -e 1 -i 6553600 -x 1
+
+.PHONY: all clean test
diff --git a/RED/baselines/cpu/Makefile b/RED/baselines/cpu/Makefile
index e4d935c..4350185 100644
--- a/RED/baselines/cpu/Makefile
+++ b/RED/baselines/cpu/Makefile
@@ -1,5 +1,12 @@
-all:
+all: red
+
+red: app_baseline.cpp
 	g++ -O2 app_baseline.cpp -fopenmp -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_OMP -lgomp -lm -o red -D${TYPE}
 
+run: red
+	./red -i 1048576000 -t 4
+
 clean:
 	rm red
+
+.PHONY: all clean run
diff --git a/RED/baselines/cpu/app_baseline.cpp b/RED/baselines/cpu/app_baseline.cpp
index db4b060..10534c6 100644
--- a/RED/baselines/cpu/app_baseline.cpp
+++ b/RED/baselines/cpu/app_baseline.cpp
@@ -194,18 +194,14 @@ int main(int argc, char **argv) {
         if (status) {
             printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
             if(rep >= p.n_warmup) {
-                printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                    "| throughput_cpu_ref_MBps=%f throughput_cpu_thrust_MBps=%f\n",
+                printf("[::] RED CPU | n_threads=%d e_type=%s n_elements=%u "
+                    "| throughput_seq_MBps=%f throughput_MBps=%f",
                     nr_threads, XSTR(T), input_size,
                     input_size * sizeof(T) / timer.time[0],
                     input_size * sizeof(T) / timer.time[1]);
-                printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                    "| throughput_cpu_ref_MOpps=%f throughput_cpu_thrust_MOpps=%f\n",
-                    nr_threads, XSTR(T), input_size,
+                printf(" throughput_seq_MOpps=%f throughput_MOpps=%f",
                     input_size / timer.time[0],
                     input_size / timer.time[1]);
-                printf("[::] n_threads=%d e_type=%s n_elements=%d | ",
-                    nr_threads, XSTR(T), input_size);
                 printall(&timer, 1);
             }
         } else {
diff --git a/RED/baselines/cpu/run-opti.sh b/RED/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..bd380c7
--- /dev/null
+++ b/RED/baselines/cpu/run-opti.sh
@@ -0,0 +1,9 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make -B TYPE=UINT64 verbose=1
+
+./red -i 1048576000 -t 4 -w 0 -e 100 | sed 's/CPU/CPU Baseline/' | tee "${HOST}-baseline.txt"
diff --git a/RED/baselines/cpu/run.sh b/RED/baselines/cpu/run.sh
index 1707edf..3a0fb09 100755
--- a/RED/baselines/cpu/run.sh
+++ b/RED/baselines/cpu/run.sh
@@ -2,16 +2,26 @@
 
 set -e
 
+HOST="$(hostname)"
+
+echo $HOST
+
+(
+
 echo "prim-benchmarks RED CPU/Thrust (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
+# default: UINT64 -i 1048576000 -t 4
+
+for nr_threads in 88 64 44 1 2 4 6 8 12 16 20 24 32; do
 	for i in 2048 4096 8192 16384 65536 262144 1048576 3932160 15728640 31457280 262144000 1048576000 2097152000; do
-		for dt in UINT32 UINT64 INT32 INT64 FLOAT DOUBLE; do
-			if make -B TYPE=${dt}; then
-				timeout -k 1m 30m ./red -i ${i} -w 0 -e 100 -t ${nr_threads} || true
+		#for dt in UINT32 UINT64 INT32 INT64 FLOAT DOUBLE; do
+		for dt in UINT64; do
+			if make -B TYPE=${dt} verbose=1; then
+				timeout -k 1m 60m ./red -i ${i} -w 0 -e 100 -t ${nr_threads} || true
 			fi
 		done
 	done
 done
+) | tee "${HOST}-explore.txt"
diff --git a/RED/host/app.c b/RED/host/app.c
index c274db0..eefb696 100644
--- a/RED/host/app.c
+++ b/RED/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
@@ -32,7 +35,7 @@ static T* A;
 // Create input arrays
 static void read_input(T* A, unsigned int nr_elements) {
     srand(0);
-    printf("nr_elements\t%u\t", nr_elements);
+    //printf("nr_elements\t%u\t", nr_elements);
     for (unsigned int i = 0; i < nr_elements; i++) {
         A[i] = (T)(rand());
     }
@@ -64,7 +67,7 @@ int main(int argc, char **argv) {
     DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
     DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
     DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
-    printf("Allocated %d DPU(s)\n", nr_of_dpus);
+    //printf("Allocated %d DPU(s)\n", nr_of_dpus);
 
     unsigned int i = 0;
 #if PERF
@@ -91,21 +94,21 @@ int main(int argc, char **argv) {
     // Timer declaration
     Timer timer;
 
-    printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL);
+    //printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL);
 
     // Loop over main kernel
     for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
 
         // 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);
         count_host = reduction_host(A, input_size);
         if(rep >= p.n_warmup)
             stop(&timer, 0);
 
-        printf("Load input data\n");
+        //printf("Load input data\n");
         if(rep >= p.n_warmup)
-            start(&timer, 1, rep - p.n_warmup);
+            start(&timer, 1, 0);
         count = 0;
         // Input arguments
         unsigned int kernel = 0;
@@ -129,10 +132,10 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup)
             stop(&timer, 1);
 
-        printf("Run program on DPU(s) \n");
+        //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
@@ -158,11 +161,11 @@ int main(int argc, char **argv) {
         }
 #endif
 
-        printf("Retrieve results\n");
+        //printf("Retrieve results\n");
         dpu_results_t results[nr_of_dpus];
         T* results_count = malloc(nr_of_dpus * sizeof(T));
         if(rep >= p.n_warmup)
-            start(&timer, 3, rep - p.n_warmup);
+            start(&timer, 3, 0);
         i = 0;
         // PARALLEL RETRIEVE TRANSFER
         dpu_results_t* results_retrieve[nr_of_dpus];
@@ -223,12 +226,36 @@ int main(int argc, char **argv) {
 
         // Free memory
         free(results_count);
+
+        // Check output
+        bool status = true;
+        if(count != count_host) status = false;
+        if (status) {
+            printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+            if (rep >= p.n_warmup) {
+                printf("[::] RED 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]));
+                printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
+                    input_size / timer.time[0],
+                    input_size / (timer.time[2]),
+                    input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
+                printall(&timer, 3);
+            }
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
+
     }
 #if PERF
     printf("DPU cycles  = %g cc\n", cc / p.n_reps);
 #endif
 
     // Print timing results
+/*
     printf("CPU ");
     print(&timer, 0, p.n_reps);
     printf("CPU-DPU ");
@@ -237,6 +264,7 @@ int main(int argc, char **argv) {
     print(&timer, 2, p.n_reps);
     printf("Inter-DPU ");
     print(&timer, 3, p.n_reps);
+*/
 
     #if ENERGY
     double energy;
@@ -244,18 +272,9 @@ int main(int argc, char **argv) {
     printf("DPU Energy (J): %f\t", energy);
     #endif	
 
-    // Check output
-    bool status = true;
-    if(count != count_host) status = false;
-    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);
     DPU_ASSERT(dpu_free(dpu_set));
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/RED/run-paper-strong-rank.sh b/RED/run-paper-strong-rank.sh
new file mode 100755
index 0000000..ed6e60e
--- /dev/null
+++ b/RED/run-paper-strong-rank.sh
@@ -0,0 +1,25 @@
+#!/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
+
+(
+
+echo "prim-benchmarks RED 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 VERSION=SINGLE verbose=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 6553600 -x 1 || true
+		fi
+	done
+done
+) | tee log-paper-strong-rank.txt
diff --git a/RED/run-paper-weak.sh b/RED/run-paper-weak.sh
new file mode 100755
index 0000000..5a66cbc
--- /dev/null
+++ b/RED/run-paper-weak.sh
@@ -0,0 +1,25 @@
+#!/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
+
+(
+
+echo "prim-benchmarks RED weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# upstream does not include 256 and 512 in config space
+for nr_dpus in 512 256 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 VERSION=SINGLE verbose=1; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 6553600 -x 0 || true
+		fi
+	done
+done
+) | tee log-paper-weak.txt