port SCAN-SSA to dfatool

7 files changed, 123 insertions, 70 deletions

diff --git a/SCAN-SSA/Makefile b/SCAN-SSA/Makefile
index 4930e70..00af65c 100644
--- a/SCAN-SSA/Makefile
+++ b/SCAN-SSA/Makefile
@@ -1,46 +1,36 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_DPUS ?= 1
 NR_TASKLETS ?= 16
 BL ?= 10
 TYPE ?= INT64
 ENERGY ?= 0
+UNROLL ?= 1
 
-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
+HOST_SOURCES := $(wildcard host/*.c)
+DPU_SOURCES := $(wildcard dpu/*.c)
 
 COMMON_INCLUDES := support
-HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
-DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
-
-.PHONY: all clean test
+COMMON_FLAGS = -Wall -Wextra -O2 -I${COMMON_INCLUDES} -DNR_DPUS=${NR_DPUS} -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} -DUNROLL=${UNROLL}
+HOST_FLAGS = ${COMMON_FLAGS} -std=c11 `dpu-pkg-config --cflags --libs dpu`
+DPU_FLAGS = ${COMMON_FLAGS}
 
-__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} -D${TYPE} -DENERGY=${ENERGY}
-DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} 
+ifdef verbose
+	QUIET =
+endif
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+all: bin/dpu_code bin/host_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/dpu_code: ${DPU_SOURCES}
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} ${DPU_SOURCES} -o $@
 
-${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+bin/host_code: ${HOST_SOURCES}
+	${QUIET}${CC} ${HOST_FLAGS} ${HOST_SOURCES} -o $@
 
 clean:
-	$(RM) -r $(BUILDDIR)
+	${QUIET}${RM} -f bin/dpu_code bin/host_code
 
-test: all
-	./${HOST_TARGET}
+.PHONY: all clean
diff --git a/SCAN-SSA/dpu/task.c b/SCAN-SSA/dpu/task.c
index c6871f7..15411a4 100644
--- a/SCAN-SSA/dpu/task.c
+++ b/SCAN-SSA/dpu/task.c
@@ -23,10 +23,16 @@ T message_partial_count;
 // Scan in each tasklet
 static T scan(T *output, T *input){
     output[0] = input[0];
+#if UNROLL
     #pragma unroll
     for(unsigned int j = 1; j < REGS; j++) {
         output[j] = output[j - 1] + input[j];
     }
+#else
+    for(unsigned int j = 1; j < REGS; j++) {
+        output[j] = output[j - 1] + input[j];
+    }
+#endif
     return output[REGS - 1];
 }
 
@@ -53,10 +59,16 @@ BARRIER_INIT(my_barrier, NR_TASKLETS);
 
 // Add in each tasklet
 static void add(T *output, T p_count){
+#if UNROLL
     #pragma unroll
     for(unsigned int j = 0; j < REGS; j++) {
         output[j] += p_count;
     }
+#else
+    for(unsigned int j = 0; j < REGS; j++) {
+        output[j] += p_count;
+    }
+#endif
 }
 
 extern int main_kernel1(void);
diff --git a/SCAN-SSA/host/app.c b/SCAN-SSA/host/app.c
index 2cc8af7..93650a8 100644
--- a/SCAN-SSA/host/app.c
+++ b/SCAN-SSA/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
@@ -51,6 +54,17 @@ static void scan_host(T* C, T* A, unsigned int nr_elements) {
     }
 }
 
+/*
+ * "SCAN" := [a1, a1, ..., an] -> [a1, a1 + a2, ..., a1 + a2 + ... + an]
+ * From the paper:
+ * SCAN-SSA has three steps:
+ * * (copy data to the DPU) (timer 1)
+ * * compute the partial scan operation locally inside each DPU (timer 2)
+ * * retrieve last element of each local scan from the DPUs, do a local (complete) scan on the last elements, and push them to the next DPU (timer 3)
+ * * finish the scan operation in each DPU (add sum [a1 ... ax] to a(x+1), ..., a(x+m) for DPU size m) (timer 4)
+ * * (retrieve results from DPU) (timer 5)
+ */
+
 // Main of the Host Application
 int main(int argc, char **argv) {
 
@@ -73,7 +87,7 @@ int main(int argc, char **argv) {
     unsigned int i = 0;
     T accum = 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 * nr_of_dpus; // Total input size (weak or strong scaling)
     const unsigned int input_size_dpu_ = divceil(input_size, nr_of_dpus); // Input size per DPU (max.)
     const unsigned int input_size_dpu_round = 
         (input_size_dpu_ % (NR_TASKLETS * REGS) != 0) ? roundup(input_size_dpu_, (NR_TASKLETS * REGS)) : input_size_dpu_; // Input size per DPU (max.), 8-byte aligned
@@ -98,14 +112,14 @@ 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);
         scan_host(C, 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);
         // Input arguments
         const unsigned int input_size_dpu = input_size_dpu_round;
         unsigned int kernel = 0;
@@ -123,10 +137,10 @@ 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
@@ -152,7 +166,7 @@ int main(int argc, char **argv) {
         }
 #endif
 
-        printf("Retrieve results\n");
+        //printf("Retrieve results\n");
         dpu_results_t results[nr_of_dpus];
         T* results_scan = malloc(nr_of_dpus * sizeof(T));
         i = 0;
@@ -162,7 +176,7 @@ int main(int argc, char **argv) {
         dpu_results_t* results_retrieve[nr_of_dpus];
 
         if(rep >= p.n_warmup)
-            start(&timer, 3, rep - p.n_warmup);
+            start(&timer, 3, 0);
 
         DPU_FOREACH(dpu_set, dpu, i) {
             results_retrieve[i] = (dpu_results_t*)malloc(NR_TASKLETS * sizeof(dpu_results_t));
@@ -201,10 +215,10 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup)
             stop(&timer, 3);
 
-        printf("Run program on DPU(s) \n");
+        //printf("Run program on DPU(s) \n");
         // Run DPU kernel
         if(rep >= p.n_warmup) {
-            start(&timer, 4, rep - p.n_warmup);
+            start(&timer, 4, 0);
             #if ENERGY
             DPU_ASSERT(dpu_probe_start(&probe));
             #endif
@@ -229,9 +243,9 @@ int main(int argc, char **argv) {
         }
 #endif
 
-        printf("Retrieve results\n");
+        //printf("Retrieve results\n");
         if(rep >= p.n_warmup)
-            start(&timer, 5, rep - p.n_warmup);
+            start(&timer, 5, 0);
         i = 0;
         // PARALLEL RETRIEVE TRANSFER
         DPU_FOREACH(dpu_set, dpu, i) {
@@ -243,8 +257,40 @@ int main(int argc, char **argv) {
 
         // Free memory
         free(results_scan);
+
+        // Check output
+        bool status = true;
+        for (i = 0; i < input_size; 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");
+            printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d b_unroll=%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, UNROLL, input_size,
+                input_size * sizeof(T) / timer.time[0],
+                input_size * sizeof(T) / (timer.time[2] + timer.time[4]),
+                input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
+            printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d b_unroll=%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, UNROLL, input_size,
+                input_size / timer.time[0],
+                input_size / (timer.time[2] + timer.time[4]),
+                input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
+            printf("[::] n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d b_unroll=%d n_elements=%d | ",
+                nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, UNROLL, input_size);
+            printall(&timer);
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
     }
 
+	/*
     // Print timing results
     printf("CPU ");
     print(&timer, 0, p.n_reps);
@@ -258,6 +304,8 @@ int main(int argc, char **argv) {
     print(&timer, 4, p.n_reps);
     printf("DPU-CPU ");
     print(&timer, 5, p.n_reps);
+    printf("\n");
+	*/
 
     #if ENERGY
     double energy;
@@ -266,21 +314,6 @@ int main(int argc, char **argv) {
     #endif	
 
 
-    // Check output
-    bool status = true;
-    for (i = 0; i < input_size; 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);
@@ -288,5 +321,5 @@ int main(int argc, char **argv) {
     free(C2);
     DPU_ASSERT(dpu_free(dpu_set));
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/SCAN-SSA/run.sh b/SCAN-SSA/run.sh
index 8ea8457..10dee20 100755
--- a/SCAN-SSA/run.sh
+++ b/SCAN-SSA/run.sh
@@ -1,11 +1,27 @@
 #!/bin/bash
 
-for i in  2048 4096 8192 16384 65536 262144 1048576 3932160
-do
-		    NR_DPUS=1 NR_TASKLETS=16 BL=10 make all
-			wait
-            ./bin/host_code -w 10 -e 100 -i ${i} > profile/out${i}_tl16_bl10_dpu11
-			wait
-			make clean
-			wait
+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 SCAN-SSA (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do
+	for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do
+		for i in  2048 4096 8192 16384 65536 262144 1048576 3932160; do
+			for dt in UINT32 UINT64 INT32 INT64 FLOAT DOUBLE; do
+				echo
+				if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 TYPE=${dt} UNROLL=1 \
+					|| make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 TYPE=${dt} UNROLL=0; then
+					bin/host_code -w 0 -e 100 -i ${i} || true
+				fi
+			done
+		done
+	done
 done
diff --git a/SCAN-SSA/support/common.h b/SCAN-SSA/support/common.h
index 1c419da..1c419da 100755..100644
--- a/SCAN-SSA/support/common.h
+++ b/SCAN-SSA/support/common.h
diff --git a/SCAN-SSA/support/params.h b/SCAN-SSA/support/params.h
index bb86211..cdc075d 100644
--- a/SCAN-SSA/support/params.h
+++ b/SCAN-SSA/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() {
@@ -30,10 +29,9 @@ struct Params input_params(int argc, char **argv) {
     p.input_size    = 3932160;
     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 +40,6 @@ 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();
diff --git a/SCAN-SSA/support/timer.h b/SCAN-SSA/support/timer.h
index b53d95f..194569f 100755..100644
--- a/SCAN-SSA/support/timer.h
+++ b/SCAN-SSA/support/timer.h
@@ -56,4 +56,9 @@ void stop(Timer *timer, int i) {
                       (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);

 }

 

-void print(Timer *timer, int i, int REP) { printf("Time (ms): %f\t", timer->time[i] / (1000 * REP)); }

+void printall(Timer *timer) {

+    for (int i = 0; i <= 5; i++) {

+        printf(" timer%d_us=%f", i, timer->time[i]);

+    }

+    printf("\n");

+}