diff options
| -rw-r--r-- | Microbenchmarks/STREAM/Makefile | 49 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/dpu/add.c | 8 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/dpu/copyw.c | 8 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/dpu/scale.c | 8 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/dpu/triad.c | 8 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/host/app.c | 64 | ||||
| -rwxr-xr-x | Microbenchmarks/STREAM/run.sh | 49 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/support/common.h | 2 | ||||
| -rw-r--r-- | Microbenchmarks/STREAM/support/params.h | 5 |
9 files changed, 107 insertions, 94 deletions
diff --git a/Microbenchmarks/STREAM/Makefile b/Microbenchmarks/STREAM/Makefile index c669857..12fd65c 100644 --- a/Microbenchmarks/STREAM/Makefile +++ b/Microbenchmarks/STREAM/Makefile @@ -1,46 +1,31 @@ -DPU_DIR := dpu -HOST_DIR := host -BUILDDIR ?= bin NR_TASKLETS ?= 16 BL ?= 10 NR_DPUS ?= 1 +T ?= uint64_t OP ?= copy MEM ?= MRAM +UNROLL ?= 1 -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}) +DPU_SOURCES = dpu/${OP}.c +HOST_SOURCES = $(wildcard host/*.c) -HOST_TARGET := ${BUILDDIR}/host_code -DPU_TARGET := ${BUILDDIR}/dpu_code +COMMON_INCLUDES = support +COMMON_FLAGS = -Wall -Wextra -O2 -I${COMMON_INCLUDES} -DNR_DPUS=${NR_DPUS} -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -DT=${T} -D${OP} -D${MEM} -DUNROLL=${UNROLL} +HOST_FLAGS = ${COMMON_FLAGS} -std=c11 `dpu-pkg-config --cflags --libs dpu` +DPU_FLAGS = ${COMMON_FLAGS} -flto -COMMON_INCLUDES := support -HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c) -DPU_SOURCES := $(wildcard ${DPU_DIR}/${OP}.c) +all: bin/dpu_code bin/host_code -.PHONY: all clean test +bin: + mkdir -p bin -__dirs := $(shell mkdir -p ${BUILDDIR}) +bin/dpu_code: ${DPU_SOURCES} + dpu-upmem-dpurte-clang ${DPU_FLAGS} ${DPU_SOURCES} -o $@ -COMMON_FLAGS := -Wall -Wextra -g -I${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} - -all: ${HOST_TARGET} ${DPU_TARGET} - -${CONF}: - $(RM) $(call conf_filename,*,*) - touch ${CONF} - -${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF} - $(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS} - -${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF} - dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES} +bin/host_code: ${HOST_SOURCES} + ${CC} ${HOST_FLAGS} ${HOST_SOURCES} -o $@ clean: - $(RM) -r $(BUILDDIR) + ${RM} -f bin/dpu_code bin/host_code -test: all - ./${HOST_TARGET} +.PHONY: all clean diff --git a/Microbenchmarks/STREAM/dpu/add.c b/Microbenchmarks/STREAM/dpu/add.c index faf6504..12854ee 100644 --- a/Microbenchmarks/STREAM/dpu/add.c +++ b/Microbenchmarks/STREAM/dpu/add.c @@ -18,12 +18,16 @@ __host dpu_results_t DPU_RESULTS[NR_TASKLETS]; // Add static void add_dpu(T *bufferC, T *bufferA, T *bufferB) { - +#if UNROLL #pragma unroll for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ bufferC[i] = bufferA[i] + bufferB[i]; } - +#else + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferC[i] = bufferA[i] + bufferB[i]; + } +#endif } // Barrier diff --git a/Microbenchmarks/STREAM/dpu/copyw.c b/Microbenchmarks/STREAM/dpu/copyw.c index eff7a3b..66dd319 100644 --- a/Microbenchmarks/STREAM/dpu/copyw.c +++ b/Microbenchmarks/STREAM/dpu/copyw.c @@ -18,12 +18,16 @@ __host dpu_results_t DPU_RESULTS[NR_TASKLETS]; // Copy static void copyw_dpu(T *bufferB, T *bufferA) { - +#if UNROLL #pragma unroll for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ bufferB[i] = bufferA[i]; } - +#else + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferB[i] = bufferA[i]; + } +#endif } // Barrier diff --git a/Microbenchmarks/STREAM/dpu/scale.c b/Microbenchmarks/STREAM/dpu/scale.c index 4247aac..53a70a2 100644 --- a/Microbenchmarks/STREAM/dpu/scale.c +++ b/Microbenchmarks/STREAM/dpu/scale.c @@ -18,12 +18,16 @@ __host dpu_results_t DPU_RESULTS[NR_TASKLETS]; // Scale static void scale_dpu(T *bufferB, T *bufferA, T scalar) { - +#if UNROLL #pragma unroll for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ bufferB[i] = scalar * bufferA[i]; } - +#else + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferB[i] = scalar * bufferA[i]; + } +#endif } // Barrier diff --git a/Microbenchmarks/STREAM/dpu/triad.c b/Microbenchmarks/STREAM/dpu/triad.c index e81cb59..9e641ea 100644 --- a/Microbenchmarks/STREAM/dpu/triad.c +++ b/Microbenchmarks/STREAM/dpu/triad.c @@ -18,12 +18,16 @@ __host dpu_results_t DPU_RESULTS[NR_TASKLETS]; // Triad static void triad_dpu(T *bufferC, T *bufferA, T *bufferB, T scalar) { - +#if UNROLL #pragma unroll for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ bufferC[i] = bufferA[i] + scalar * bufferB[i]; } - +#else + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferC[i] = bufferA[i] + scalar * bufferB[i]; + } +#endif } // Barrier diff --git a/Microbenchmarks/STREAM/host/app.c b/Microbenchmarks/STREAM/host/app.c index 4571e96..6e0ac16 100644 --- a/Microbenchmarks/STREAM/host/app.c +++ b/Microbenchmarks/STREAM/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; @@ -40,6 +43,27 @@ static void read_input(T* A, T* B, unsigned int nr_elements) { } } +static char benchmark_name[] = +#ifdef scale +"SCALE" +#elif add +"ADD" +#elif triad +"TRIAD" +#else +"COPY" +#endif +; + +static char mem_name[] = +#ifdef WRAM +"WRAM" +#endif +#ifdef MRAM +"MRAM" +#endif +; + // Compute output in the host #if defined(add) || defined(triad) static void stream_host(T* C, T* B, T* A, unsigned int nr_elements) { @@ -76,7 +100,7 @@ int main(int argc, char **argv) { 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; + const unsigned int input_size = p.input_size * nr_of_dpus; // Input/output allocation A = malloc(input_size * sizeof(T)); @@ -102,7 +126,7 @@ 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); #if defined(add) || defined(triad) stream_host(C2, B, A, input_size); #else @@ -113,7 +137,7 @@ int main(int argc, char **argv) { 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 / nr_of_dpus; unsigned int kernel = 0; @@ -134,7 +158,7 @@ int main(int argc, char **argv) { 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); DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); if(rep >= p.n_warmup) stop(&timer, 2); @@ -153,7 +177,7 @@ int main(int argc, char **argv) { printf("Retrieve results\n"); if(rep >= p.n_warmup) - start(&timer, 3, rep - p.n_warmup); + start(&timer, 3, 0); dpu_results_t results[nr_of_dpus]; i = 0; DPU_FOREACH (dpu_set, dpu) { @@ -198,19 +222,6 @@ int main(int argc, char **argv) { } #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; for (i = 0; i < input_size; i++) { @@ -231,10 +242,25 @@ int main(int argc, char **argv) { } if (status) { printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n"); + printf("[::] n_dpus=%d n_tasklets=%d e_benchmark=%-6s e_type=%s e_mem=%s b_unroll=%d | throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f \n", nr_of_dpus, NR_TASKLETS, benchmark_name, XSTR(T), mem_name, UNROLL, 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("[::] n_dpus=%d n_tasklets=%d e_benchmark=%-6s e_type=%s e_mem=%s b_unroll=%d | throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f \n", nr_of_dpus, NR_TASKLETS, benchmark_name, XSTR(T), mem_name, UNROLL, input_size / timer.time[0], input_size / timer.time[2], input_size / (timer.time[1] + timer.time[2] + timer.time[3])); } else { printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n"); } + } + 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); + // Deallocation free(A); free(B); @@ -244,5 +270,5 @@ int main(int argc, char **argv) { free(C2); DPU_ASSERT(dpu_free(dpu_set)); - return status ? 0 : -1; + return 0; } diff --git a/Microbenchmarks/STREAM/run.sh b/Microbenchmarks/STREAM/run.sh index 3ed965e..24b08df 100755 --- a/Microbenchmarks/STREAM/run.sh +++ b/Microbenchmarks/STREAM/run.sh @@ -1,38 +1,25 @@ #!/bin/bash -mkdir -p profile set -e -# MRAM -for i in copy copyw add scale triad -do - for j in 1 - do - for k in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 - do - NR_DPUS=$j NR_TASKLETS=$k BL=10 MEM=MRAM OP=$i make all - wait - ./bin/host_code -w 0 -e 1 -i 2097152 >& profile/${i}_${j}_tl${k}_MRAM.txt - wait - make clean - wait - done - done -done +# 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!) +# 2097152 B -> 2M is maximum for 64bit types (due to 16M per DPU) -# WRAM -for i in copyw add scale triad -do - for j in 1 - do - for k in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 - do - NR_DPUS=$j NR_TASKLETS=$k BL=10 MEM=WRAM OP=$i make all - wait - ./bin/host_code -w 0 -e 1 -i 2097152 >& profile/${i}_${j}_tl${k}_WRAM.txt - wait - make clean - wait - done +for mem in MRAM WRAM; do + 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 op in copy copyw add scale triad; do + for dt in uint8_t uint16_t uint32_t uint64_t float double; do + if make -B MEM=${mem} OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=1 \ + || make -B MEM=${mem} OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=0; then + bin/host_code -w 0 -e 20 -i 2097152 + fi + done + done + done done done diff --git a/Microbenchmarks/STREAM/support/common.h b/Microbenchmarks/STREAM/support/common.h index 8e2e59b..7264fbd 100644 --- a/Microbenchmarks/STREAM/support/common.h +++ b/Microbenchmarks/STREAM/support/common.h @@ -25,7 +25,9 @@ typedef struct { #endif // Data type +#ifndef T #define T uint64_t +#endif #define PERF 1 // Use perfcounters? #define PRINT 0 diff --git a/Microbenchmarks/STREAM/support/params.h b/Microbenchmarks/STREAM/support/params.h index 4618411..d28ba65 100644 --- a/Microbenchmarks/STREAM/support/params.h +++ b/Microbenchmarks/STREAM/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 = 8 << 10; 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(); |