MRAM latency benchmark: report DMA latency and bandwidth; merge r/w modes

5 files changed, 81 insertions, 84 deletions

diff --git a/Microbenchmarks/MRAM-Latency/Makefile b/Microbenchmarks/MRAM-Latency/Makefile
index b8dc7e2..ef75829 100644
--- a/Microbenchmarks/MRAM-Latency/Makefile
+++ b/Microbenchmarks/MRAM-Latency/Makefile
@@ -1,46 +1,37 @@
 DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_TASKLETS ?= 16
 BL ?= 8
 NR_DPUS ?= 1
-OP ?= READ
 MEM ?= MRAM
 
-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)
+HOST_SOURCES := $(wildcard host/*.c)
 
-.PHONY: all clean test
+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${MEM}
+DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${MEM}
 
-__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${OP} -D${MEM}
-DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${OP} -D${MEM}
+ifdef verbose
+	QUIET =
+endif
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+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/copy.c ${COMMON_INCLUDES} bin
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ dpu/copy.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/MRAM-Latency/dpu/copy.c b/Microbenchmarks/MRAM-Latency/dpu/copy.c
index 5cfd141..de9cca8 100644
--- a/Microbenchmarks/MRAM-Latency/dpu/copy.c
+++ b/Microbenchmarks/MRAM-Latency/dpu/copy.c
@@ -45,7 +45,9 @@ int main_kernel1() {
     uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
 
     dpu_results_t *result = &DPU_RESULTS[tasklet_id];
-    result->cycles = 0;
+    result->count = 0;
+    result->r_cycles = 0;
+    result->w_cycles = 0;
 
     // Address of the current processing block in MRAM
     uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
@@ -57,27 +59,20 @@ int main_kernel1() {
     for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
         __mram_ptr void const* address_A = (__mram_ptr void const*)(mram_base_addr_A + byte_index);
         __mram_ptr void* address_B = (__mram_ptr void*)(mram_base_addr_B + byte_index);
-#ifdef READ
         // Barrier
         timer_start(&cycles); // START TIMER
-#endif
         // Load cache with current MRAM block
         mram_read(address_A, cache_A, BLOCK_SIZE);
-#ifdef READ
         // Barrier
-        result->cycles += timer_stop(&cycles); // STOP TIMER
-#endif
+        result->r_cycles += timer_stop(&cycles); // STOP TIMER
 
-#ifdef WRITE
         // Barrier
         timer_start(&cycles); // START TIMER
-#endif
         // Write cache to current MRAM block
         mram_write(cache_A, address_B, BLOCK_SIZE);
-#ifdef WRITE
         // Barrier
-        result->cycles += timer_stop(&cycles); // STOP TIMER
-#endif
+        result->w_cycles += timer_stop(&cycles); // STOP TIMER
+        result->count += 1;
     }
 
     return 0;
diff --git a/Microbenchmarks/MRAM-Latency/host/app.c b/Microbenchmarks/MRAM-Latency/host/app.c
index c968749..3b5c530 100644
--- a/Microbenchmarks/MRAM-Latency/host/app.c
+++ b/Microbenchmarks/MRAM-Latency/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 T* A;
 static T* B;
@@ -30,7 +33,6 @@ 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());
@@ -50,17 +52,17 @@ int main(int argc, char **argv) {
     struct Params p = input_params(argc, argv);
 
     struct dpu_set_t dpu_set, dpu;
+    uint32_t clocks_per_sec;
     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));
-    printf("Allocated %d DPU(s)\n", nr_of_dpus);
+    DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
 
     unsigned int i = 0;
-    double cc = 0;
-    double cc_min = 0;
     const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size;
     assert(input_size % (nr_of_dpus * NR_TASKLETS) == 0 && "Input size!");
 
@@ -77,8 +79,6 @@ 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++) {
 
@@ -89,7 +89,6 @@ int main(int argc, char **argv) {
         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);
         // Input arguments
@@ -106,7 +105,6 @@ 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);
@@ -126,24 +124,33 @@ int main(int argc, char **argv) {
         }
 #endif
 
-        printf("Retrieve results\n");
         if(rep >= p.n_warmup)
             start(&timer, 3, rep - p.n_warmup);
-        dpu_results_t results[nr_of_dpus];
         i = 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), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T)));
 			
 #if PERF
-            results[i].cycles = 0;
+            DPU_ASSERT(dpu_copy_from(dpu, "CLOCKS_PER_SEC", 0, &clocks_per_sec, sizeof(clocks_per_sec)));
             // Retrieve tasklet timings
             for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) {
                 dpu_results_t result;
-                result.cycles = 0;
+                result.count = 0;
+                result.r_cycles = 0;
+                result.w_cycles = 0;
                 DPU_ASSERT(dpu_copy_from(dpu, "DPU_RESULTS", each_tasklet * sizeof(dpu_results_t), &result, sizeof(dpu_results_t)));
-                if (result.cycles > results[i].cycles)
-                    results[i].cycles = result.cycles;
+                printf("[::] DMA UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%u block_size_B=%d"
+                    " | latency_mram_read_us=%f latency_mram_write_us=%f"
+                    " throughput_dpu_mram_read_MBps=%f throughput_dpu_mram_write_MBps=%f"
+                    " throughput_mram_read_MBps=%f throughput_mram_write_MBps=%f\n",
+                    nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), input_size_dpu, BLOCK_SIZE,
+                    ((double)result.r_cycles * 1e6 / clocks_per_sec) / result.count,
+                    ((double)result.w_cycles * 1e6 / clocks_per_sec) / result.count,
+                    input_size * sizeof(T) / ((double)result.r_cycles * 1e6 / clocks_per_sec),
+                    input_size * sizeof(T) / ((double)result.w_cycles * 1e6 / clocks_per_sec),
+                    input_size * sizeof(T) / ((double)result.r_cycles * 1e6 * NR_TASKLETS / clocks_per_sec),
+                    input_size * sizeof(T) / ((double)result.w_cycles * 1e6 * NR_TASKLETS / clocks_per_sec));
             }
 #endif
             i++;
@@ -151,36 +158,7 @@ int main(int argc, char **argv) {
         if(rep >= p.n_warmup)
             stop(&timer, 3);
 
-#if PERF
-        uint64_t max_cycles = 0;
-        uint64_t min_cycles = 0xFFFFFFFFFFFFFFFF;
-        // Print performance results
-        if(rep >= p.n_warmup){
-            i = 0;
-            DPU_FOREACH(dpu_set, dpu) {
-                if(results[i].cycles > max_cycles)
-                    max_cycles = results[i].cycles;
-                if(results[i].cycles < min_cycles)
-                    min_cycles = results[i].cycles;
-                i++;
-            }
-            cc += (double)max_cycles;
-            cc_min += (double)min_cycles;
-        }
-#endif
-
     }
-    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;
@@ -193,7 +171,6 @@ int main(int argc, char **argv) {
         }
     }
     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");
     }
diff --git a/Microbenchmarks/MRAM-Latency/run-transfer.sh b/Microbenchmarks/MRAM-Latency/run-transfer.sh
new file mode 100755
index 0000000..f4b8e08
--- /dev/null
+++ b/Microbenchmarks/MRAM-Latency/run-transfer.sh
@@ -0,0 +1,32 @@
+#!/bin/bash
+
+# The DPU application reads blocks of data from MRAM to WRAM via mram_read
+# and copies blocks of data from WRAM to MRAM via mram_write.
+# Each transfer handles BLOCK_SIZE (BL<<2) bytes via DMA.
+# Each DPU reports its total runtime, i.e., the number of cycles accumulated
+# over all mram_read or mram_write calls.
+# Total data size is 8192 uint64 elements, so 65536 B (64 KiB)
+# Overall throughput is 65536 B / dpu_seconds
+# Per-DPU throughput is 65536 B / (dpu_seconds * nr_tasklets)
+# Latency per mram_* call is dpu_seconds / dpu_count
+
+set -e
+
+(
+
+echo "prim-benchmarks MRAM microbenchmark (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for ndpu in 1 4 8 16 32 48 64; do
+	for ntask in 1 2 4 8 12 16 20; do
+		for bl in 4 5 6 7 8 9 10 11; do
+			if make -B NR_DPUS=$ndpu NR_TASKLETS=$ntask BL=$bl; then
+				bin/host_code -w 0 -e 100 || true
+			fi
+		done
+	done
+done
+echo "Completed at $(date)"
+) | tee "log-$(hostname).txt"
+xz -v -9 -M 800M "log-$(hostname).txt"
diff --git a/Microbenchmarks/MRAM-Latency/support/common.h b/Microbenchmarks/MRAM-Latency/support/common.h
index 8e2e59b..ed2c69b 100755
--- a/Microbenchmarks/MRAM-Latency/support/common.h
+++ b/Microbenchmarks/MRAM-Latency/support/common.h
@@ -11,7 +11,9 @@ typedef struct {
 } dpu_arguments_t;
 
 typedef struct {
-    uint64_t cycles;
+    uint64_t r_cycles;
+    uint64_t w_cycles;
+    uint32_t count;
 } dpu_results_t;
 
 // Transfer size between MRAM and WRAM