WRAM WiP

3 files changed, 78 insertions, 41 deletions

diff --git a/Microbenchmarks/WRAM/Makefile b/Microbenchmarks/WRAM/Makefile
index f824e41..24817fb 100644
--- a/Microbenchmarks/WRAM/Makefile
+++ b/Microbenchmarks/WRAM/Makefile
@@ -1,6 +1,4 @@
 DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_TASKLETS ?= 16
 BL ?= 10
 NR_DPUS ?= 1
@@ -8,40 +6,34 @@ OP ?= streaming
 MEM ?= WRAM
 TYPE ?= INT64
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_$(4)_$(5).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${OP},${MEM})
-
-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)
 
 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${OP} -D${MEM} -D${TYPE}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${OP} -D${MEM} -D${TYPE}
 
-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/task.c ${COMMON_INCLUDES} bin
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ dpu/task.c
 
 clean:
-	$(RM) -r $(BUILDDIR)
+	${QUIET}rm -rf bin
 
 test: all
-	./${HOST_TARGET}
+	${QUIET}bin/host_code
+
+.PHONY: all clean test
diff --git a/Microbenchmarks/WRAM/host/app.c b/Microbenchmarks/WRAM/host/app.c
index a0131da..65193fc 100644
--- a/Microbenchmarks/WRAM/host/app.c
+++ b/Microbenchmarks/WRAM/host/app.c
@@ -22,6 +22,9 @@
 #define DPU_BINARY "./bin/dpu_code"
 #endif
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 // Pointer declaration
 static unsigned int* A;
 static T* B;
@@ -49,6 +52,14 @@ static void read_input(unsigned int* A, T* B, unsigned int nr_elements) {
     }
 }
 
+#ifdef streaming
+char transfer_mode[] = "streaming";
+#elif strided
+char transfer_mode[] = "strided";
+#else
+char transfer_mode[] = "random";
+#endif
+
 // Compute output in the host
 static void copy_host(T* C, T* B, unsigned int* A, unsigned int nr_elements) {
     unsigned int wram_size = BLOCK_SIZE >> DIV;
@@ -67,11 +78,13 @@ int main(int argc, char **argv) {
 
     struct dpu_set_t dpu_set, dpu;
     uint32_t nr_of_dpus;
+    uint32_t nr_of_ranks;
     
     // Allocate DPUs and load binary
     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));
+    DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
     printf("Allocated %d DPU(s)\n", nr_of_dpus);
 
     unsigned int i = 0;
@@ -165,6 +178,10 @@ int main(int argc, char **argv) {
                 dpu_results_t result;
                 result.cycles = 0;
                 DPU_ASSERT(dpu_copy_from(dpu, "DPU_RESULTS", each_tasklet * sizeof(dpu_results_t), &result, sizeof(dpu_results_t)));
+                printf("[::] DMA UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%u e_mode=%s block_size_B=%d"
+                    " | dpu_cycles=%lu\n",
+                    nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), input_size_dpu, transfer_mode, BLOCK_SIZE,
+                    result.cycles);
                 if (result.cycles > results[i].cycles)
                     results[i].cycles = result.cycles;
             }
@@ -195,16 +212,6 @@ int main(int argc, char **argv) {
     }
     printf("DPU cycles  = %g cc\n", cc / p.n_reps);
 
-    // Print timing results
-    printf("CPU ");
-    print(&timer, 0, p.n_reps);
-    printf("CPU-DPU ");
-    print(&timer, 1, p.n_reps);
-    printf("DPU Kernel ");
-    print(&timer, 2, p.n_reps);
-    printf("DPU-CPU ");
-    print(&timer, 3, p.n_reps);
-
     // Check output
     bool status = true;
     for (i = 0; i < input_size; i++) {
@@ -215,11 +222,6 @@ int main(int argc, char **argv) {
 #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);
diff --git a/Microbenchmarks/WRAM/run-transfer.sh b/Microbenchmarks/WRAM/run-transfer.sh
new file mode 100755
index 0000000..7c4a179
--- /dev/null
+++ b/Microbenchmarks/WRAM/run-transfer.sh
@@ -0,0 +1,43 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks WRAM microbenchmark (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for ndpu in 1 4 8 16; do
+	for ntask in 1 2 4 8 12 16 20; do
+		for bl in 4 5 6 7 8 9 10 11; do
+			for op in streaming strided random; do
+				if make -B NR_DPUS=$ndpu NR_TASKLETS=$ntask BL=$bl OP=$op; then
+					bin/host_code -w 0 -e 10 || true
+				fi
+			done
+		done
+	done
+done
+echo "Completed at $(date)"
+) | tee "log-$(hostname)-wram-only.txt"
+
+(
+
+echo "prim-benchmarks WRAM microbenchmark (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for ndpu in 1 4 8 16; do
+	for ntask in 1 2 4 8 12 16 20; do
+		for bl in 4 5 6 7 8 9 10 11; do
+			for op in streaming strided random; do
+				if make -B NR_DPUS=$ndpu NR_TASKLETS=$ntask BL=$bl OP=$op MEM=MRAM; then
+					bin/host_code -w 0 -e 10 || true
+				fi
+			done
+		done
+	done
+done
+echo "Completed at $(date)"
+) | tee "log-$(hostname)-wram-mram.txt"