port VA to dfatool

4 files changed, 88 insertions, 58 deletions

diff --git a/VA/Makefile b/VA/Makefile
index ab1bd97..d089ab1 100644
--- a/VA/Makefile
+++ b/VA/Makefile
@@ -1,46 +1,38 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
+NR_DPUS ?= 1
 NR_TASKLETS ?= 16
 BL ?= 10
-NR_DPUS ?= 1
 TYPE ?= INT32
 ENERGY ?= 0
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_TYPE_$(4).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${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)
+HOST_SOURCES := $(wildcard host/*.c)
+DPU_SOURCES := $(wildcard dpu/*.c)
 
 .PHONY: all clean test
 
-__dirs := $(shell mkdir -p ${BUILDDIR})
-
 COMMON_FLAGS := -Wall -Wextra -g -I${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${TYPE} -DENERGY=${ENERGY}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE}
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+QUIET = @
+
+ifdef verbose
+	QUIET =
+endif
+
+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}
+	bin/host_code
diff --git a/VA/host/app.c b/VA/host/app.c
index c7c02f3..8fe9e1c 100644
--- a/VA/host/app.c
+++ b/VA/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
@@ -35,7 +38,6 @@ static T* C2;
 // Create input arrays
 static void read_input(T* A, T* B, unsigned int nr_elements) {
     srand(0);
-    printf("nr_elements\t%u\t", nr_elements);
     for (unsigned int i = 0; i < nr_elements; i++) {
         A[i] = (T) (rand());
         B[i] = (T) (rand());
@@ -69,7 +71,7 @@ int main(int argc, char **argv) {
     printf("Allocated %d DPU(s)\n", nr_of_dpus);
     unsigned int i = 0;
 
-    const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size; // Total input size (weak or strong scaling)
+    const unsigned int input_size = p.input_size;
     const unsigned int input_size_8bytes = 
         ((input_size * sizeof(T)) % 8) != 0 ? roundup(input_size, 8) : input_size; // Input size per DPU (max.), 8-byte aligned
     const unsigned int input_size_dpu = divceil(input_size, nr_of_dpus); // Input size per DPU (max.)
@@ -91,21 +93,18 @@ int main(int argc, char **argv) {
     // Timer declaration
     Timer timer;
 
-    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);
         vector_addition_host(C, A, B, 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
         unsigned int kernel = 0;
         dpu_arguments_t input_arguments[NR_DPUS];
@@ -137,10 +136,9 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup)
             stop(&timer, 1);
 
-        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
@@ -165,9 +163,8 @@ int main(int argc, char **argv) {
         }
 #endif
 
-        printf("Retrieve results\n");
         if(rep >= p.n_warmup)
-            start(&timer, 3, rep - p.n_warmup);
+            start(&timer, 3, 0);
         i = 0;
         // PARALLEL RETRIEVE TRANSFER
         DPU_FOREACH(dpu_set, dpu, i) {
@@ -177,9 +174,42 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup)
             stop(&timer, 3);
 
+        // Check output
+        bool status = true;
+        for (i = 0; i < input_size; i++) {
+            if(C[i] != bufferC[i]){ 
+                status = false;
+#if PRINT
+                printf("%d: %u -- %u\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("[::] 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\n",
+                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
+                    input_size * 3 * sizeof(T) / timer.time[0],
+                    input_size * 3 * sizeof(T) / (timer.time[1]),
+                    input_size * 3 * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
+                printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
+                    "| throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f\n",
+                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
+                    input_size / timer.time[0],
+                    input_size / (timer.time[1]),
+                    input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
+                printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d | ",
+                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size);
+                printall(&timer, 3);
+            }
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
     }
 
     // Print timing results
+    /*
     printf("CPU ");
     print(&timer, 0, p.n_reps);
     printf("CPU-DPU ");
@@ -188,6 +218,7 @@ int main(int argc, char **argv) {
     print(&timer, 2, p.n_reps);
     printf("DPU-CPU ");
     print(&timer, 3, p.n_reps);
+    */
 
 #if ENERGY
     double energy;
@@ -195,21 +226,6 @@ int main(int argc, char **argv) {
     printf("DPU Energy (J): %f\t", energy);
 #endif	
 
-    // Check output
-    bool status = true;
-    for (i = 0; i < input_size; i++) {
-        if(C[i] != bufferC[i]){ 
-            status = false;
-#if PRINT
-            printf("%d: %u -- %u\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);
@@ -218,5 +234,5 @@ int main(int argc, char **argv) {
     free(C2);
     DPU_ASSERT(dpu_free(dpu_set));
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/VA/run.sh b/VA/run.sh
new file mode 100755
index 0000000..984d6ef
--- /dev/null
+++ b/VA/run.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: input size (number of elements, not number of bytes!)
+
+echo "prim-benchmarks VA (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do
+	for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do
+		for i in  2048 4096 8192 16384 65536 262144 1048576 2621440; do
+			for dt in CHAR SHORT INT32 INT64 FLOAT DOUBLE; do
+				echo
+				if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 TYPE=${dt}; then
+					bin/host_code -w 0 -e 100 -i ${i} || true
+				fi
+			done
+		done
+	done
+done
diff --git a/VA/support/params.h b/VA/support/params.h
index efb34f3..e797393 100644
--- a/VA/support/params.h
+++ b/VA/support/params.h
@@ -7,7 +7,6 @@ typedef struct Params {
     unsigned int   input_size;
     int   n_warmup;
     int   n_reps;
-    int  exp;
 }Params;
 
 static void usage() {
@@ -18,7 +17,6 @@ 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"
         "\nBenchmark-specific options:"
         "\n    -i <I>    input size (default=2621440 elements)"
@@ -30,10 +28,9 @@ struct Params input_params(int argc, char **argv) {
     p.input_size    = 2621440;
     p.n_warmup      = 1;
     p.n_reps        = 3;
-    p.exp           = 0;
 
     int opt;
-    while((opt = getopt(argc, argv, "hi:w:e:x:")) >= 0) {
+    while((opt = getopt(argc, argv, "hi:w:e:")) >= 0) {
         switch(opt) {
         case 'h':
         usage();
@@ -42,7 +39,6 @@ struct Params input_params(int argc, char **argv) {
         case 'i': p.input_size    = atoi(optarg); break;
         case 'w': p.n_warmup      = atoi(optarg); break;
         case 'e': p.n_reps        = atoi(optarg); break;
-        case 'x': p.exp           = atoi(optarg); break;
         default:
             fprintf(stderr, "\nUnrecognized option!\n");
             usage();