diff options
Diffstat (limited to 'Microbenchmarks')
67 files changed, 4849 insertions, 0 deletions
diff --git a/Microbenchmarks/Arithmetic-Throughput/Makefile b/Microbenchmarks/Arithmetic-Throughput/Makefile new file mode 100644 index 0000000..137e7b5 --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/Makefile @@ -0,0 +1,46 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 10 +NR_DPUS ?= 1 +OP ?= ADD +TYPE ?= INT32 + +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},${OP},${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) + +.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${OP} -D${TYPE} +DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${OP} -D${TYPE} + +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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/Arithmetic-Throughput/dpu/task.c b/Microbenchmarks/Arithmetic-Throughput/dpu/task.c new file mode 100644 index 0000000..2f538ff --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/dpu/task.c @@ -0,0 +1,100 @@ +/* +* Execution of arithmetic operations with multiple tasklets +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Arithmetic operation +static void update(T *bufferA, T scalar) { + //#pragma unroll + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + // WRAM READ + T temp = bufferA[i]; +#ifdef ADD + temp += scalar; // ADD +#elif SUB + temp -= scalar; // SUB +#elif MUL + temp *= scalar; // MUL +#elif DIV + temp /= scalar; // DIV +#endif + // WRAM WRITE + bufferA[i] = temp; + } +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + // Barrier + barrier_wait(&my_barrier); + perfcounter_cycles cycles; + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + T scalar = (T)input_size_dpu; // Simply use this number as a scalar + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER + + // Update + update(cache_A, scalar); + + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); + + // Write cache to current MRAM block + mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + } + + return 0; +} diff --git a/Microbenchmarks/Arithmetic-Throughput/host/app.c b/Microbenchmarks/Arithmetic-Throughput/host/app.c new file mode 100644 index 0000000..5933d1d --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/host/app.c @@ -0,0 +1,224 @@ +/** +* app.c +* Arithmetic Throughput Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; +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()); + } +} + +// Compute output in the host +static void update_host(T* C, T* A, unsigned int nr_elements) { + for (unsigned int i = 0; i < nr_elements; i++) { +#if ADD + C[i] = A[i] + (nr_elements / NR_DPUS); +#elif SUB + C[i] = A[i] - (nr_elements / NR_DPUS); +#elif MUL + C[i] = A[i] * (nr_elements / NR_DPUS); +#elif DIV + C[i] = A[i] / (nr_elements / NR_DPUS); +#endif + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + T *bufferA = A; + T *bufferB = B; + C2 = malloc(input_size * sizeof(T)); + + // Create an input file with arbitrary data + read_input(A, B, input_size); + + // 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); + update_host(C2, A, 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); + // Input arguments + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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 + + } +#ifdef ADD + printf("ADD\n"); +#elif SUB + printf("SUB\n"); +#elif MUL + printf("MUL\n"); +#elif DIV + printf("DIV\n"); +#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++) { + if(C2[i] != bufferB[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, C2[i], bufferB[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); + free(B); + free(C2); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/Arithmetic-Throughput/run.sh b/Microbenchmarks/Arithmetic-Throughput/run.sh new file mode 100755 index 0000000..9200e54 --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/run.sh @@ -0,0 +1,17 @@ +#!/bin/bash + +for i in ADD SUB MUL DIV +do + for j in INT32 FLOAT UINT32 INT64 DOUBLE UINT64 + do + for k in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 + do + NR_DPUS=1 NR_TASKLETS=$k BL=10 OP=$i TYPE=$j make all + wait + ./bin/host_code -w 0 -e 1 -i 1048576 > profile/${i}_${j}_tl${k}.txt + wait + make clean + wait + done + done +done diff --git a/Microbenchmarks/Arithmetic-Throughput/support/common.h b/Microbenchmarks/Arithmetic-Throughput/support/common.h new file mode 100755 index 0000000..0d1cc8a --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/support/common.h @@ -0,0 +1,48 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#ifdef UINT32 +#define T uint32_t +#elif UINT64 +#define T uint64_t +#elif INT32 +#define T int32_t +#elif INT64 +#define T int64_t +#elif FLOAT +#define T float +#elif DOUBLE +#define T double +#endif + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/Arithmetic-Throughput/support/cyclecount.h b/Microbenchmarks/Arithmetic-Throughput/support/cyclecount.h new file mode 100644 index 0000000..4ef377e --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/support/cyclecount.h @@ -0,0 +1,20 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} + diff --git a/Microbenchmarks/Arithmetic-Throughput/support/params.h b/Microbenchmarks/Arithmetic-Throughput/support/params.h new file mode 100644 index 0000000..4618411 --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/support/params.h @@ -0,0 +1,56 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + 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) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/Arithmetic-Throughput/support/timer.h b/Microbenchmarks/Arithmetic-Throughput/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/Arithmetic-Throughput/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/CPU-DPU/Makefile b/Microbenchmarks/CPU-DPU/Makefile new file mode 100644 index 0000000..09d613a --- /dev/null +++ b/Microbenchmarks/CPU-DPU/Makefile @@ -0,0 +1,45 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 8 +NR_DPUS ?= 1 +TRANSFER ?= PUSH + +define conf_filename + ${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_TRANSFER_$(4).conf +endef +CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${TRANSFER}) + +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 ${BUILDDIR}) + +COMMON_FLAGS := -Wall -Wextra -g -I${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${TRANSFER} +DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TRANSFER} + +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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/CPU-DPU/dpu/task.c b/Microbenchmarks/CPU-DPU/dpu/task.c new file mode 100644 index 0000000..cb68b4c --- /dev/null +++ b/Microbenchmarks/CPU-DPU/dpu/task.c @@ -0,0 +1,36 @@ +/* +* Empty kernel with multiple tasklets +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { +#if PRINT + unsigned int tasklet_id = me(); + printf("tasklet_id = %u\n", tasklet_id); +#endif + return 0; +} diff --git a/Microbenchmarks/CPU-DPU/host/app.c b/Microbenchmarks/CPU-DPU/host/app.c new file mode 100644 index 0000000..6b8ab53 --- /dev/null +++ b/Microbenchmarks/CPU-DPU/host/app.c @@ -0,0 +1,188 @@ +/** +* app.c +* CPU-DPU Communication Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; +static T* C; +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] = A[i]; + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // Allocate DPUs and load binary + DPU_ASSERT(dpu_alloc(NR_DPUS, "nrThreadPerPool=8", &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); + + unsigned int i = 0; + unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + C = malloc(input_size * sizeof(T)); + C2 = malloc(input_size * sizeof(T)); + T *bufferA = A; + T *bufferC = C; + + // Create an input file with arbitrary data + read_input(A, B, input_size); + + // 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++) { + + printf("Load input data\n"); + // Input arguments + const unsigned int input_size_dpu = input_size / nr_of_dpus; + // Copy input arrays + if(rep >= p.n_warmup) + start(&timer, 1, rep - p.n_warmup); + i = 0; +#ifdef SERIAL + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } +#elif BROADCAST + DPU_ASSERT(dpu_broadcast_to(dpu_set, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT)); +#else + DPU_FOREACH(dpu_set, dpu, i) { + DPU_ASSERT(dpu_prepare_xfer(dpu, bufferA + input_size_dpu * i)); + } + DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT)); +#endif + 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); + //DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#endif + + printf("Retrieve results\n"); + if(rep >= p.n_warmup) + start(&timer, 3, rep - p.n_warmup); + i = 0; +#ifdef SERIAL + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferC + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } +#else + DPU_FOREACH(dpu_set, dpu, i) { + DPU_ASSERT(dpu_prepare_xfer(dpu, bufferC + input_size_dpu * i)); + } + DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT)); +#endif + if(rep >= p.n_warmup) + stop(&timer, 3); + + } + + // Print timing results + printf("CPU-DPU "); + print(&timer, 1, p.n_reps); + double time_load = timer.time[1] / (1000 * p.n_reps); + printf("CPU-DPU Bandwidth (GB/s): %f\n", (input_size * 8)/(time_load*1e6)); + printf("DPU Kernel "); + print(&timer, 2, p.n_reps); + printf("\n"); + printf("DPU-CPU "); + print(&timer, 3, p.n_reps); + double time_retrieve = timer.time[3] / (1000 * p.n_reps); + printf("DPU-CPU Bandwidth (GB/s): %f\n", (input_size * 8)/(time_retrieve*1e6)); + + // Check output + bool status = true; +#ifdef BROADCAST + for (i = 0; i < input_size/nr_of_dpus; i++) { + if(B[i] != bufferC[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, B[i], bufferA[i]); +#endif + } + } +#else + for (i = 0; i < input_size; i++) { + if(B[i] != bufferC[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, B[i], bufferA[i]); +#endif + } + } +#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); + free(B); + free(C); + free(C2); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/CPU-DPU/run.sh b/Microbenchmarks/CPU-DPU/run.sh new file mode 100755 index 0000000..09cdd7a --- /dev/null +++ b/Microbenchmarks/CPU-DPU/run.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +for i in 1 2 4 8 16 32 64 +do + for j in 1 + do + for k in SERIAL PUSH BROADCAST + do + for l in 1 4 16 64 256 1024 4096 16384 65536 262144 1048576 4194304 + do + NR_DPUS=$i NR_TASKLETS=$j BL=10 TRANSFER=$k make all + wait + ./bin/host_code -w 5 -e 20 -i ${l} >& profile/${i}_tl${j}_TR${k}_i${l}.txt + wait + make clean + wait + done + done + done +done diff --git a/Microbenchmarks/CPU-DPU/support/common.h b/Microbenchmarks/CPU-DPU/support/common.h new file mode 100755 index 0000000..3c3c6ae --- /dev/null +++ b/Microbenchmarks/CPU-DPU/support/common.h @@ -0,0 +1,31 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#define T uint64_t + +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/CPU-DPU/support/params.h b/Microbenchmarks/CPU-DPU/support/params.h new file mode 100644 index 0000000..4618411 --- /dev/null +++ b/Microbenchmarks/CPU-DPU/support/params.h @@ -0,0 +1,56 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + 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) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/CPU-DPU/support/timer.h b/Microbenchmarks/CPU-DPU/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/CPU-DPU/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/MRAM-Latency/Makefile b/Microbenchmarks/MRAM-Latency/Makefile new file mode 100644 index 0000000..b8dc7e2 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/Makefile @@ -0,0 +1,46 @@ +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) + +.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${OP} -D${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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/MRAM-Latency/dpu/copy.c b/Microbenchmarks/MRAM-Latency/dpu/copy.c new file mode 100644 index 0000000..5cfd141 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/dpu/copy.c @@ -0,0 +1,84 @@ +/* +* MRAM-WRAM R/W Latency with multiple tasklets +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + + 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 + +#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 + } + + return 0; +} diff --git a/Microbenchmarks/MRAM-Latency/host/app.c b/Microbenchmarks/MRAM-Latency/host/app.c new file mode 100644 index 0000000..c968749 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/host/app.c @@ -0,0 +1,208 @@ +/** +* app.c +* MRAM Latency Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; +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()); + } +} + +// Compute output in the host +static void stream_host(T* C, T* A, unsigned int nr_elements) { + for (unsigned int i = 0; i < nr_elements; i++) { + C[i] = A[i]; + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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!"); + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + T *bufferA = A; + T *bufferB = B; + C2 = malloc(input_size * sizeof(T)); + + // Create an input file with arbitrary data + read_input(A, B, input_size); + + // 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); + stream_host(C2, A, 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); + // Input arguments + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { + if(C2[i] != bufferB[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, C2[i], bufferB[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); + free(B); + free(C2); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/MRAM-Latency/run.sh b/Microbenchmarks/MRAM-Latency/run.sh new file mode 100755 index 0000000..2ae57bf --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/run.sh @@ -0,0 +1,21 @@ +#!/bin/bash + +for i in 3 4 5 6 7 8 9 10 11 +do + for j in 1 + do + NR_DPUS=1 NR_TASKLETS=$j BL=$i OP=READ make all + wait + ./bin/host_code -w 0 -e 1 -i 2097152 > profile/read_tl${j}_bl${i}.txt + wait + make clean + wait + + NR_DPUS=1 NR_TASKLETS=$j BL=$i OP=WRITE make all + wait + ./bin/host_code -w 0 -e 1 -i 2097152 > profile/write_tl${j}_bl${i}.txt + wait + make clean + wait + done +done diff --git a/Microbenchmarks/MRAM-Latency/support/common.h b/Microbenchmarks/MRAM-Latency/support/common.h new file mode 100755 index 0000000..8e2e59b --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/support/common.h @@ -0,0 +1,36 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#define T uint64_t + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/MRAM-Latency/support/cyclecount.h b/Microbenchmarks/MRAM-Latency/support/cyclecount.h new file mode 100644 index 0000000..c4247b5 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/support/cyclecount.h @@ -0,0 +1,19 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} diff --git a/Microbenchmarks/MRAM-Latency/support/params.h b/Microbenchmarks/MRAM-Latency/support/params.h new file mode 100644 index 0000000..4618411 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/support/params.h @@ -0,0 +1,56 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + 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) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/MRAM-Latency/support/timer.h b/Microbenchmarks/MRAM-Latency/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/MRAM-Latency/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/Operational-Intensity/Makefile b/Microbenchmarks/Operational-Intensity/Makefile new file mode 100644 index 0000000..0fcbd8c --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/Makefile @@ -0,0 +1,46 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 10 +NR_DPUS ?= 1 +TYPE ?= INT32 +OP ?= ADD + +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},${TYPE},${OP}) + +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 ${BUILDDIR}) + +COMMON_FLAGS := -Wall -Wextra -g -I${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${TYPE} -D${OP} +DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} -D${OP} + +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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/Operational-Intensity/dpu/task.c b/Microbenchmarks/Operational-Intensity/dpu/task.c new file mode 100644 index 0000000..ea40fb8 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/dpu/task.c @@ -0,0 +1,99 @@ +/* +* Arithmetic Throughtput versus Operational Intensity +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Update +static void update(T* bufferA, T scalar, uint32_t rep, uint32_t str) { + for (unsigned int r = 0; r < rep; r++){ + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i+=str){ +#ifdef ADD + bufferA[i] += scalar; // ADD +#elif SUB + bufferA[i] -= scalar; // SUB +#elif MUL + bufferA[i] *= scalar; // MUL +#elif DIV + bufferA[i] /= scalar; // DIV +#endif + } + } +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + // Barrier + barrier_wait(&my_barrier); + perfcounter_cycles cycles; + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + uint32_t rep = DPU_INPUT_ARGUMENTS.repetitions; + uint32_t str = DPU_INPUT_ARGUMENTS.stride; + T scalar = DPU_INPUT_ARGUMENTS.scalar; // Simply use this number as a scalar + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + + // Update + update(cache_A, scalar, rep, str); + + // Write cache to current MRAM block + mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + + } + + result->cycles = timer_stop(&cycles); // STOP TIMER + barrier_wait(&my_barrier); + + return 0; +} diff --git a/Microbenchmarks/Operational-Intensity/host/app.c b/Microbenchmarks/Operational-Intensity/host/app.c new file mode 100644 index 0000000..1fc5f46 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/host/app.c @@ -0,0 +1,221 @@ +/** +* app.c +* Operational Intensity Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; + +// Create input arrays +static void read_input(T* A, 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()); + } +} + +// Compute output in the host +static void update_host(T* A, T scalar, unsigned int nr_elements, unsigned int rep, unsigned int str) { + for (unsigned int j = 0; j < nr_elements; j+=(BLOCK_SIZE / sizeof(T))) { + for (unsigned int r = 0; r < rep; r++){ + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i+=str){ +#ifdef ADD + A[j + i] += scalar; // ADD +#elif SUB + A[j + i] -= scalar; // SUB +#elif MUL + A[j + i] *= scalar; // MUL +#elif DIV + A[j + i] /= scalar; // DIV +#endif + } + } + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + T *bufferA = A; + + // Create an input file with arbitrary data + read_input(A, input_size); + memcpy(B, A, input_size * sizeof(T)); + + // Timer declaration + Timer timer; + + printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL); + + unsigned int repetitions = p.repetitions >= 1.0 ? (unsigned int)p.repetitions : 1; + unsigned int stride = p.repetitions >= 1.0 ? 1 : (unsigned int)(1 / p.repetitions); + + printf("p.repetitions\t%f\trepetitions\t%u\tstride\t%u\n", p.repetitions, repetitions, stride); + + // 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); + update_host(B, (T)p.n_reps /*use as a scalar*/, input_size, repetitions, stride); + 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 + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + T scalar = (T)p.n_reps; // Just a scalar + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel, repetitions, stride, scalar}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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), bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + +#if PERF + results[i].cycles = 0; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { + if(B[i] != bufferA[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, B[i], bufferA[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); + free(B); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/Operational-Intensity/run.sh b/Microbenchmarks/Operational-Intensity/run.sh new file mode 100755 index 0000000..a540154 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/run.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +for i in ADD SUB MUL DIV +do + for j in CHAR SHORT INT32 FLOAT INT64 DOUBLE + do + for k in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 + do + for l in 0.001953125 0.00390625 0.0078125 0.015625 0.03125 0.0625 0.125 0.25 0.5 1 2 4 8 16 32 64 128 256 512 + do + NR_DPUS=1 NR_TASKLETS=$k BL=10 OP=$i TYPE=$j make all + wait + ./bin/host_code -w 0 -e 1 -i 1048576 -p ${l} >& profile/${i}_${j}_tl${k}_p${l}.txt + wait + make clean + wait + done + done + done +done diff --git a/Microbenchmarks/Operational-Intensity/support/common.h b/Microbenchmarks/Operational-Intensity/support/common.h new file mode 100755 index 0000000..aac3eed --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/support/common.h @@ -0,0 +1,55 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#ifdef UINT32 +#define T uint32_t +#elif UINT64 +#define T uint64_t +#elif INT32 +#define T int32_t +#elif INT64 +#define T int64_t +#elif FLOAT +#define T float +#elif DOUBLE +#define T double +#elif CHAR +#define T char +#elif SHORT +#define T short +#endif + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; + uint32_t repetitions; + uint32_t stride; + T scalar; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/Operational-Intensity/support/cyclecount.h b/Microbenchmarks/Operational-Intensity/support/cyclecount.h new file mode 100644 index 0000000..c4247b5 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/support/cyclecount.h @@ -0,0 +1,19 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} diff --git a/Microbenchmarks/Operational-Intensity/support/params.h b/Microbenchmarks/Operational-Intensity/support/params.h new file mode 100644 index 0000000..485dae5 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/support/params.h @@ -0,0 +1,60 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + float repetitions; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\n -h help" + "\n -w <W> # of untimed warmup iterations (default=2)" + "\n -e <E> # of timed repetition iterations (default=5)" + "\n -x <X> Weak (0) or strong (1) scaling (default=0)" + "\n" + "\nBenchmark-specific options:" + "\n -i <I> input size (default=8K elements)" + "\n -p <P> # of compute repetitions (default=2)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + p.input_size = 8 << 10; + p.repetitions = 1.0; + p.n_warmup = 2; + p.n_reps = 5; + p.exp = 0; + + int opt; + while((opt = getopt(argc, argv, "hi:p:w:e:")) >= 0) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + case 'i': p.input_size = atoi(optarg); break; + case 'p': p.repetitions = atof(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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/Operational-Intensity/support/timer.h b/Microbenchmarks/Operational-Intensity/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/Operational-Intensity/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/Random-GUPS/Makefile b/Microbenchmarks/Random-GUPS/Makefile new file mode 100644 index 0000000..d517fa7 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/Makefile @@ -0,0 +1,44 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 8 +NR_DPUS ?= 1 + +define conf_filename + ${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3).conf +endef +CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL}) + +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 ${BUILDDIR}) + +COMMON_FLAGS := -Wall -Wextra -g -I${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} +DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} + +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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/Random-GUPS/dpu/gups.c b/Microbenchmarks/Random-GUPS/dpu/gups.c new file mode 100644 index 0000000..dadde35 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/dpu/gups.c @@ -0,0 +1,95 @@ +/* +* Random Access (GUPS) with multiple tasklets +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + + +T ran[128]; // Current random numbers + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap +#if PERF + perfcounter_config(COUNT_CYCLES, true); +#endif + } + // Barrier + barrier_wait(&my_barrier); +#if PERF + perfcounter_cycles cycles; + timer_start(&cycles); // START TIMER + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->cycles = 0; +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + // Number of updates to table (suggested: 16x number of table entries) + int NUPDATE = 16 * input_size_dpu; + + for(int j = tasklet_id; j < 128; j += NR_TASKLETS){ + ran[j] = HPCC_starts((NUPDATE/128) * j); + } + // Barrier + barrier_wait(&my_barrier); + + // Address of the current processing block in MRAM + uint32_t mram_base_addr_A = (uint32_t)DPU_MRAM_HEAP_POINTER; + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(sizeof(T)); + + for (int i = 0; i < NUPDATE/128; i++){ + for (int j = tasklet_id; j < 128; j += NR_TASKLETS){ + + ran[j] = (ran[j] << 1) ^ ((S) ran[j] < 0 ? POLY : 0); + + // Table[ran[j] & (TableSize-1)] ^= ran[j]; is computed as follows (3 steps) + // 1. Load cache + mram_read((__mram_ptr void const*)(mram_base_addr_A + (ran[j] & (input_size_dpu - 1)) * sizeof(T)), cache_A, sizeof(T)); + + // 2. Update + //*cache_A ^= ran[j]; + *cache_A = ran[j] & (input_size_dpu - 1); + + // 3. Write cache + mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_A + (ran[j] & (input_size_dpu - 1)) * sizeof(T)), sizeof(T)); + } + } + +#if PERF + result->cycles += timer_stop(&cycles); // STOP TIMER +#endif + + return 0; +} diff --git a/Microbenchmarks/Random-GUPS/host/app.c b/Microbenchmarks/Random-GUPS/host/app.c new file mode 100644 index 0000000..74499be --- /dev/null +++ b/Microbenchmarks/Random-GUPS/host/app.c @@ -0,0 +1,221 @@ +/** +* app.c +* Random Access (GUPS) Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; + +// 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] = A[i]; + } +} + +// Compute output in the host +static void gups_host(T* B, unsigned int nr_elements) { + + // Number of updates to table (suggested: 16x number of table entries) + unsigned int NUPDATE = 16 * (nr_elements / NR_DPUS); + + for (unsigned int d = 0; d < NR_DPUS; d++){ + + T ran[128]; // Current random numbers + for(unsigned int j = 0; j < 128; j++){ + ran[j] = HPCC_starts((NUPDATE/128) * j); + } + + for (unsigned int i = 0; i < NUPDATE/128; i++){ + for (unsigned int j = 0; j < NR_TASKLETS; j++){ + for (unsigned int k = j; k < 128; k += NR_TASKLETS){ + ran[k] = (ran[k] << 1) ^ ((S) ran[k] < 0 ? POLY : 0); + //B[((nr_elements / NR_DPUS) * d) + (ran[k] & ((nr_elements / NR_DPUS) - 1))] ^= ran[k]; + B[((nr_elements / NR_DPUS) * d) + (ran[k] & ((nr_elements / NR_DPUS) - 1))] = ran[k] & ((nr_elements / NR_DPUS) - 1); + } + } + } + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + T *bufferA = A; + + // Create an input file with arbitrary data + read_input(A, B, input_size); + + // 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); + gups_host(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); + // Input arguments + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); + +#if PERF + results[i].cycles = 0; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { + if(B[i] != bufferA[i]){ + status = false; +#if PRINT + printf("%d: %lu -- %lu\n", i, B[i], bufferA[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); + free(B); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/Random-GUPS/run.sh b/Microbenchmarks/Random-GUPS/run.sh new file mode 100755 index 0000000..8c33392 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/run.sh @@ -0,0 +1,14 @@ +#!/bin/bash + +for i in 1 +do + for j in 1 2 4 8 12 16 + do + NR_DPUS=$i NR_TASKLETS=$j BL=10 make all + wait + ./bin/host_code -w 0 -e 1 -i 2097152 >& profile/gups_${i}_tl${j}.txt + wait + make clean + wait + done +done diff --git a/Microbenchmarks/Random-GUPS/support/common.h b/Microbenchmarks/Random-GUPS/support/common.h new file mode 100755 index 0000000..f4eb611 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/support/common.h @@ -0,0 +1,76 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#define T uint64_t +#define S int64_t + +// HPCC_starts function from HPCC benchmark (https://github.com/icl-utk-edu/hpcc/blob/main/RandomAccess/utility.c) +#define POLY 7ULL +#define PERIOD 1317624576693539401LL + +T HPCC_starts(S n) { + int i, j; + T m2[64]; + T temp, ran; + + while (n < 0) n += PERIOD; + while (n > PERIOD) n -= PERIOD; + if (n == 0) return 0x1; + + temp = 0x1; + for (i=0; i<64; i++) { + m2[i] = temp; + temp = (temp << 1) ^ ((S) temp < 0 ? POLY : 0); + temp = (temp << 1) ^ ((S) temp < 0 ? POLY : 0); + } + + for (i=62; i>=0; i--) + if ((n >> i) & 1) + break; + + ran = 0x2; + while (i > 0) { + temp = 0; + for (j=0; j<64; j++) + if ((ran >> j) & 1) + temp ^= m2[j]; + ran = temp; + i -= 1; + if ((n >> i) & 1) + ran = (ran << 1) ^ ((S) ran < 0 ? POLY : 0); + } + + return ran; +} + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/Random-GUPS/support/cyclecount.h b/Microbenchmarks/Random-GUPS/support/cyclecount.h new file mode 100644 index 0000000..4ef377e --- /dev/null +++ b/Microbenchmarks/Random-GUPS/support/cyclecount.h @@ -0,0 +1,20 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} + diff --git a/Microbenchmarks/Random-GUPS/support/params.h b/Microbenchmarks/Random-GUPS/support/params.h new file mode 100644 index 0000000..4618411 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/support/params.h @@ -0,0 +1,56 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + 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) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/Random-GUPS/support/timer.h b/Microbenchmarks/Random-GUPS/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/Random-GUPS/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/STREAM/Makefile b/Microbenchmarks/STREAM/Makefile new file mode 100644 index 0000000..c669857 --- /dev/null +++ b/Microbenchmarks/STREAM/Makefile @@ -0,0 +1,46 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 10 +NR_DPUS ?= 1 +OP ?= copy +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 ${DPU_DIR}/${OP}.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${OP} -D${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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/STREAM/dpu/add.c b/Microbenchmarks/STREAM/dpu/add.c new file mode 100644 index 0000000..faf6504 --- /dev/null +++ b/Microbenchmarks/STREAM/dpu/add.c @@ -0,0 +1,103 @@ +/* +* STREAM Add +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Add +static void add_dpu(T *bufferC, T *bufferA, T *bufferB) { + + #pragma unroll + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferC[i] = bufferA[i] + bufferB[i]; + } + +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + uint32_t mram_base_addr_C = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + 2 * input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE); + +#ifdef WRAM + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER +#endif + + // Add + add_dpu(cache_B, cache_A, cache_B); + +#ifdef WRAM + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); +#endif + + // Write cache to current MRAM block + mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_C + byte_index), BLOCK_SIZE); + + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/STREAM/dpu/copy.c b/Microbenchmarks/STREAM/dpu/copy.c new file mode 100644 index 0000000..ff4df42 --- /dev/null +++ b/Microbenchmarks/STREAM/dpu/copy.c @@ -0,0 +1,75 @@ +/* +* STREAM Copy +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + + // Write cache to current MRAM block + mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/STREAM/dpu/copyw.c b/Microbenchmarks/STREAM/dpu/copyw.c new file mode 100644 index 0000000..eff7a3b --- /dev/null +++ b/Microbenchmarks/STREAM/dpu/copyw.c @@ -0,0 +1,101 @@ +/* +* STREAM Copy (WRAM) +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Copy +static void copyw_dpu(T *bufferB, T *bufferA) { + + #pragma unroll + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferB[i] = bufferA[i]; + } + +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + +#ifdef WRAM + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER +#endif + + // Copy + copyw_dpu(cache_B, cache_A); + +#ifdef WRAM + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); +#endif + + // Write cache to current MRAM block + mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/STREAM/dpu/scale.c b/Microbenchmarks/STREAM/dpu/scale.c new file mode 100644 index 0000000..4247aac --- /dev/null +++ b/Microbenchmarks/STREAM/dpu/scale.c @@ -0,0 +1,103 @@ +/* +* STREAM Scale +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Scale +static void scale_dpu(T *bufferB, T *bufferA, T scalar) { + + #pragma unroll + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferB[i] = scalar * bufferA[i]; + } + +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + T scalar = (T)input_size_dpu; // Simply use this number as a scalar + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + +#ifdef WRAM + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER +#endif + + // Scale + scale_dpu(cache_B, cache_A, scalar); + +#ifdef WRAM + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); +#endif + + // Write cache to current MRAM block + mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/STREAM/dpu/triad.c b/Microbenchmarks/STREAM/dpu/triad.c new file mode 100644 index 0000000..e81cb59 --- /dev/null +++ b/Microbenchmarks/STREAM/dpu/triad.c @@ -0,0 +1,105 @@ +/* +* STREAM Triad +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Triad +static void triad_dpu(T *bufferC, T *bufferA, T *bufferB, T scalar) { + + #pragma unroll + for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){ + bufferC[i] = bufferA[i] + scalar * bufferB[i]; + } + +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + + T scalar = (T)input_size_dpu; // Simply use this number as a scalar + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->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)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T)); + uint32_t mram_base_addr_C = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + 2 * input_size_dpu * sizeof(T)); + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE); + +#ifdef WRAM + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER +#endif + + // Triad + triad_dpu(cache_B, cache_A, cache_B, scalar); + +#ifdef WRAM + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); +#endif + + // Write cache to current MRAM block + mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_C + byte_index), BLOCK_SIZE); + + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/STREAM/host/app.c b/Microbenchmarks/STREAM/host/app.c new file mode 100644 index 0000000..4571e96 --- /dev/null +++ b/Microbenchmarks/STREAM/host/app.c @@ -0,0 +1,248 @@ +/** +* app.c +* STREAM Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static T* A; +static T* B; +#if defined(add) || defined(triad) +static T* C; +#endif +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()); + } +} + +// Compute output in the host +#if defined(add) || defined(triad) +static void stream_host(T* C, T* B, T* A, unsigned int nr_elements) { +#else +static void stream_host(T* C, T* A, unsigned int nr_elements) { +#endif + for (unsigned int i = 0; i < nr_elements; i++) { +#ifdef scale + C[i] = (nr_elements / NR_DPUS) * A[i]; +#elif add + C[i] = A[i] + B[i]; +#elif triad + C[i] = A[i] + (nr_elements / NR_DPUS) * B[i]; +#else // copy + C[i] = A[i]; +#endif + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + T *bufferA = A; + T *bufferB = B; +#if defined(add) || defined(triad) + C = malloc(input_size * sizeof(T)); + T *bufferC = C; +#endif + C2 = malloc(input_size * sizeof(T)); + + // Create an input file with arbitrary data + read_input(A, B, input_size); + + // 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); +#if defined(add) || defined(triad) + stream_host(C2, B, A, input_size); +#else + stream_host(C2, A, input_size); +#endif + 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 + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); +#if defined(add) || defined(triad) + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T))); +#endif + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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 +#if defined(add) || defined(triad) + DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, 2 * input_size_dpu * sizeof(T), bufferC + input_size_dpu * i, input_size_dpu * sizeof(T))); +#else + 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))); +#endif + +#if PERF + results[i].cycles = 0; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { +#if defined(add) || defined(triad) + if(C2[i] != bufferC[i]){ +#else + if(C2[i] != bufferB[i]){ +#endif + status = false; +#if PRINT +#if defined(add) || defined(triad) + printf("%d: %u -- %u\n", i, C2[i], bufferC[i]); +#else + printf("%d: %u -- %u\n", i, C2[i], bufferB[i]); +#endif +#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); + free(B); +#if defined(add) || defined(triad) + free(C); +#endif + free(C2); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/STREAM/run.sh b/Microbenchmarks/STREAM/run.sh new file mode 100755 index 0000000..141b01c --- /dev/null +++ b/Microbenchmarks/STREAM/run.sh @@ -0,0 +1,35 @@ +#!/bin/bash + +# 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 + +# 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 + done +done diff --git a/Microbenchmarks/STREAM/support/common.h b/Microbenchmarks/STREAM/support/common.h new file mode 100755 index 0000000..8e2e59b --- /dev/null +++ b/Microbenchmarks/STREAM/support/common.h @@ -0,0 +1,36 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#define T uint64_t + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/STREAM/support/cyclecount.h b/Microbenchmarks/STREAM/support/cyclecount.h new file mode 100644 index 0000000..4ef377e --- /dev/null +++ b/Microbenchmarks/STREAM/support/cyclecount.h @@ -0,0 +1,20 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} + diff --git a/Microbenchmarks/STREAM/support/params.h b/Microbenchmarks/STREAM/support/params.h new file mode 100644 index 0000000..4618411 --- /dev/null +++ b/Microbenchmarks/STREAM/support/params.h @@ -0,0 +1,56 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + 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) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/STREAM/support/timer.h b/Microbenchmarks/STREAM/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/STREAM/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/STRIDED/Makefile b/Microbenchmarks/STRIDED/Makefile new file mode 100644 index 0000000..c5f3071 --- /dev/null +++ b/Microbenchmarks/STRIDED/Makefile @@ -0,0 +1,45 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 8 +NR_DPUS ?= 1 +OP ?= COARSECOARSE + +define conf_filename + ${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_$(4).conf +endef +CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${OP}) + +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 ${BUILDDIR}) + +COMMON_FLAGS := -Wall -Wextra -g -I${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} +DPU_FLAGS := ${COMMON_FLAGS} -O2 -flto -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${OP} + +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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/STRIDED/dpu/strided.c b/Microbenchmarks/STRIDED/dpu/strided.c new file mode 100644 index 0000000..273b57c --- /dev/null +++ b/Microbenchmarks/STRIDED/dpu/strided.c @@ -0,0 +1,103 @@ +/* +* Strided access with multiple tasklets +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->cycles = 0; + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T); + uint32_t s = DPU_INPUT_ARGUMENTS.stride; + + // Address of the current processing block in MRAM + uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id * (input_size_dpu * sizeof(T) / NR_TASKLETS))); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id * (input_size_dpu * sizeof(T) / NR_TASKLETS)) + input_size_dpu * sizeof(T)); + +#ifdef COARSECOARSE + // BLOCK SIZE + uint32_t B_SIZE = BLOCK_SIZE / sizeof(T); + uint32_t ADDR = (input_size_dpu/NR_TASKLETS) * tasklet_id; + uint32_t j = 0; + + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(BLOCK_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T) / NR_TASKLETS; byte_index += BLOCK_SIZE){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE); + mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE); + + // Copy + if(((ADDR + j * B_SIZE) & (s - 1)) == 0){ + + for(unsigned int i = 0; i < B_SIZE; i += s){ + cache_B[i] = cache_A[i]; + } + + } + + // Write cache to current MRAM block + mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE); + j++; + } +#else // FINEFINE + // Initialize a local cache to store the MRAM block + T *cache_A = (T *) mem_alloc(sizeof(T)); + uint32_t stride = (uint32_t)(s * sizeof(T)); + + for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T) / NR_TASKLETS; byte_index += stride){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, sizeof(T)); + + // Write cache to current MRAM block + mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_B + byte_index), sizeof(T)); + } +#endif + + result->cycles = timer_stop(&cycles); // STOP TIMER + + return 0; +} diff --git a/Microbenchmarks/STRIDED/host/app.c b/Microbenchmarks/STRIDED/host/app.c new file mode 100644 index 0000000..e9d119a --- /dev/null +++ b/Microbenchmarks/STRIDED/host/app.c @@ -0,0 +1,211 @@ +/** +* app.c +* Strided Access Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Create input arrays +static void read_input(T* A, T* B, T* C, 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] = A[i]; + C[i] = A[i]; + } +} + +// Compute output in the host +static void stride_host(T* B, T* A, unsigned int nr_elements, int s) { + for (unsigned int i = 0; i < nr_elements; i+=s){ + B[i] = A[i]; + } +} + +// Pointer declaration +static T* A; +static T* B; +static T* C; + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(T)); + B = malloc(input_size * sizeof(T)); + C = malloc(input_size * sizeof(T)); + T *bufferA = A; + T *bufferB = B; + + // Create an input file with arbitrary data + read_input(A, B, C, input_size); + + // 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); + stride_host(C, A, input_size, p.stride); + 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 + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), p.stride, kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T))); +#ifdef COARSECOARSE + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T))); +#endif + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { + if(C[i] != bufferB[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, C[i], bufferB[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); + free(B); + free(C); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/STRIDED/run.sh b/Microbenchmarks/STRIDED/run.sh new file mode 100755 index 0000000..153e633 --- /dev/null +++ b/Microbenchmarks/STRIDED/run.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +for i in COARSECOARSE FINEFINE +do + for j in 1 + do + for k in 1 2 4 8 16 + do + for l in 1 2 4 8 16 32 64 128 256 512 1024 2048 4096 + do + NR_DPUS=$j NR_TASKLETS=$k BL=10 OP=$i make all + wait + ./bin/host_code -w 0 -e 1 -i 2097152 -s ${l} >& profile/${i}_${j}_tl${k}_s${l}.txt + wait + make clean + wait + done + done + done +done diff --git a/Microbenchmarks/STRIDED/support/common.h b/Microbenchmarks/STRIDED/support/common.h new file mode 100755 index 0000000..c71f33a --- /dev/null +++ b/Microbenchmarks/STRIDED/support/common.h @@ -0,0 +1,37 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + uint32_t stride; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#define T uint64_t + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" +#endif diff --git a/Microbenchmarks/STRIDED/support/cyclecount.h b/Microbenchmarks/STRIDED/support/cyclecount.h new file mode 100644 index 0000000..4ef377e --- /dev/null +++ b/Microbenchmarks/STRIDED/support/cyclecount.h @@ -0,0 +1,20 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} + diff --git a/Microbenchmarks/STRIDED/support/params.h b/Microbenchmarks/STRIDED/support/params.h new file mode 100644 index 0000000..e49cc36 --- /dev/null +++ b/Microbenchmarks/STRIDED/support/params.h @@ -0,0 +1,60 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + unsigned int stride; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n -s <S> stride (default=2)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + p.input_size = 8 << 10; + p.stride = 2; + p.n_warmup = 1; + p.n_reps = 3; + p.exp = 0; + + int opt; + while((opt = getopt(argc, argv, "hi:s:w:e:x:")) >= 0) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + case 'i': p.input_size = atoi(optarg); break; + case 's': p.stride = 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(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/Microbenchmarks/STRIDED/support/timer.h b/Microbenchmarks/STRIDED/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/STRIDED/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
diff --git a/Microbenchmarks/WRAM/Makefile b/Microbenchmarks/WRAM/Makefile new file mode 100644 index 0000000..f824e41 --- /dev/null +++ b/Microbenchmarks/WRAM/Makefile @@ -0,0 +1,47 @@ +DPU_DIR := dpu +HOST_DIR := host +BUILDDIR ?= bin +NR_TASKLETS ?= 16 +BL ?= 10 +NR_DPUS ?= 1 +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 ${BUILDDIR}) + +COMMON_FLAGS := -Wall -Wextra -g -I${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} ${DPU_TARGET} + +${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} + +clean: + $(RM) -r $(BUILDDIR) + +test: all + ./${HOST_TARGET} diff --git a/Microbenchmarks/WRAM/dpu/task.c b/Microbenchmarks/WRAM/dpu/task.c new file mode 100644 index 0000000..9ded3da --- /dev/null +++ b/Microbenchmarks/WRAM/dpu/task.c @@ -0,0 +1,111 @@ +/* +* WRAM Access +* +*/ +#include <stdint.h> +#include <stdio.h> +#include <defs.h> +#include <mram.h> +#include <alloc.h> +#include <perfcounter.h> +#include <barrier.h> + +#include "../support/common.h" +#include "../support/cyclecount.h" + +__host dpu_arguments_t DPU_INPUT_ARGUMENTS; +__host dpu_results_t DPU_RESULTS[NR_TASKLETS]; + +// Copy +static void copy_pattern_dpu(T *bufferC, T *bufferB, uint32_t *bufferA) { + + #pragma unroll + for (unsigned int i = 0; i < (BLOCK_SIZE >> DIV); i++){ + + uint32_t address = bufferA[i]; + bufferC[address] = bufferB[address]; + + } + +} + +// Barrier +BARRIER_INIT(my_barrier, NR_TASKLETS); + +extern int main_kernel1(void); + +int (*kernels[nr_kernels])(void) = {main_kernel1}; + +int main(void) { + // Kernel + return kernels[DPU_INPUT_ARGUMENTS.kernel](); +} + +// main_kernel1 +int main_kernel1() { + unsigned int tasklet_id = me(); +#if PRINT + printf("tasklet_id = %u\n", tasklet_id); +#endif + if (tasklet_id == 0){ // Initialize once the cycle counter + mem_reset(); // Reset the heap + + perfcounter_config(COUNT_CYCLES, true); + } + perfcounter_cycles cycles; + // Barrier + barrier_wait(&my_barrier); +#ifndef WRAM + timer_start(&cycles); // START TIMER +#endif + + uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size; + + dpu_results_t *result = &DPU_RESULTS[tasklet_id]; + result->cycles = 0; + + const uint32_t A_SIZE = (BLOCK_SIZE >> DIV) << 2; + // Address of the current processing block in MRAM + uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id * A_SIZE)); + uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(uint32_t)); + uint32_t mram_base_addr_C = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * (sizeof(uint32_t) + sizeof(T))); + + // Initialize a local cache to store the MRAM block + uint32_t *cache_A = (uint32_t *) mem_alloc(A_SIZE); + T *cache_B = (T *) mem_alloc(BLOCK_SIZE); + T *cache_C = (T *) mem_alloc(BLOCK_SIZE); + + uint32_t A_byte_index = 0; + for(unsigned int byte_index = 0; byte_index < (input_size_dpu << DIV); byte_index += BLOCK_SIZE * NR_TASKLETS){ + + // Load cache with current MRAM block + mram_read((__mram_ptr void const*)(mram_base_addr_A + A_byte_index), cache_A, A_SIZE); + mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE); + mram_read((__mram_ptr void const*)(mram_base_addr_C + byte_index), cache_C, BLOCK_SIZE); // Clean cache_C + +#ifdef WRAM + // Barrier + barrier_wait(&my_barrier); + timer_start(&cycles); // START TIMER +#endif + + // Copy + copy_pattern_dpu(cache_C, cache_B, cache_A); + +#ifdef WRAM + result->cycles += timer_stop(&cycles); // STOP TIMER + // Barrier + barrier_wait(&my_barrier); +#endif + + // Write cache to current MRAM block + mram_write(cache_C, (__mram_ptr void*)(mram_base_addr_C + byte_index), BLOCK_SIZE); + + A_byte_index += A_SIZE * NR_TASKLETS; + } + +#ifndef WRAM + result->cycles = timer_stop(&cycles); // STOP TIMER +#endif + return 0; +} diff --git a/Microbenchmarks/WRAM/host/app.c b/Microbenchmarks/WRAM/host/app.c new file mode 100644 index 0000000..a0131da --- /dev/null +++ b/Microbenchmarks/WRAM/host/app.c @@ -0,0 +1,232 @@ +/** +* app.c +* WRAM Access Host Application Source File +* +*/ +#include <stdio.h> +#include <stdlib.h> +#include <stdbool.h> +#include <string.h> +#include <dpu.h> +#include <dpu_log.h> +#include <unistd.h> +#include <getopt.h> +#include <assert.h> + +#include "../support/common.h" +#include "../support/timer.h" +#include "../support/params.h" + +// Define the DPU Binary path as DPU_BINARY here +#ifndef DPU_BINARY +#define DPU_BINARY "./bin/dpu_code" +#endif + +// Pointer declaration +static unsigned int* A; +static T* B; +static T* C; +static T* C2; + +// Create input arrays +#ifdef strided +static void read_input(unsigned int* A, T* B, unsigned int nr_elements, unsigned int stride) { +#else +static void read_input(unsigned int* A, T* B, unsigned int nr_elements) { +#endif + srand(0); + printf("nr_elements\t%u\t", nr_elements); + for (unsigned int i = 0; i < nr_elements; i++) { +#ifdef streaming + A[i] = i % (BLOCK_SIZE >> DIV); +#elif strided + A[i] = ((i>0 ? A[i-1]:0) + stride) % (BLOCK_SIZE >> DIV); +#else + A[i] = ((unsigned int)rand()) % (BLOCK_SIZE >> DIV); +#endif + B[i] = (T)(rand()); + C[i] = 0; + } +} + +// 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; + for (unsigned int i = 0; i < nr_elements / wram_size; i++) { + for (unsigned int j = 0; j < wram_size; j++) { + unsigned int address = A[i * wram_size + j]; + C[i * wram_size + address] = B[i * wram_size + address]; + } + } +} + +// Main of the Host Application +int main(int argc, char **argv) { + + struct Params p = input_params(argc, argv); + + struct dpu_set_t dpu_set, dpu; + uint32_t nr_of_dpus; + + // 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); + + 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; + + // Input/output allocation + A = malloc(input_size * sizeof(unsigned int)); + unsigned int *bufferA = A; + B = malloc(input_size * sizeof(T)); + T *bufferB = B; + C = malloc(input_size * sizeof(T)); + T *bufferC = C; + C2 = malloc(input_size * sizeof(T)); + + // Create an input file with arbitrary data +#ifdef strided + read_input(A, B, input_size, p.stride); +#else + read_input(A, B, input_size); +#endif + + // 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); + copy_host(C2, B, A, 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); + // Input arguments + const unsigned int input_size_dpu = input_size / nr_of_dpus; + unsigned int kernel = 0; + dpu_arguments_t input_arguments = {input_size_dpu, kernel}; + DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments))); + // Copy input arrays + i = 0; + DPU_FOREACH (dpu_set, dpu) { + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(unsigned int))); + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(unsigned int), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T))); + DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * (sizeof(unsigned int) + sizeof(T)), bufferC + input_size_dpu * i, input_size_dpu * sizeof(T))); + i++; + } + 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); + DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); + if(rep >= p.n_warmup) + stop(&timer, 2); + +#if PRINT + { + unsigned int each_dpu = 0; + printf("Display DPU Logs\n"); + DPU_FOREACH (dpu_set, dpu) { + printf("DPU#%d:\n", each_dpu); + DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout)); + each_dpu++; + } + } +#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(unsigned int) + sizeof(T)), bufferC + input_size_dpu * i, input_size_dpu * sizeof(T))); + +#if PERF + results[i].cycles = 0; + // Retrieve tasklet timings + for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { + 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))); + if (result.cycles > results[i].cycles) + results[i].cycles = result.cycles; + } +#endif + i++; + } + 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; + for (i = 0; i < input_size; i++) { + if(C2[i] != bufferC[i]){ + status = false; +#if PRINT + printf("%d: %u -- %u\n", i, C2[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); + free(B); + free(C); + free(C2); + DPU_ASSERT(dpu_free(dpu_set)); + + return status ? 0 : -1; +} diff --git a/Microbenchmarks/WRAM/run.sh b/Microbenchmarks/WRAM/run.sh new file mode 100755 index 0000000..12fca6e --- /dev/null +++ b/Microbenchmarks/WRAM/run.sh @@ -0,0 +1,37 @@ +#!/bin/bash + +# WRAM +for i in streaming random +do + for j in 1 + do + for k in 1 2 4 8 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}_s1_WRAM.txt + wait + make clean + wait + done + done +done + +for i in strided +do + for j in 1 + do + for k in 1 2 4 8 16 + do + for l in 1 2 4 8 16 32 64 + 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 -s ${l} >& profile/${i}_${j}_tl${k}_s${l}_WRAM.txt + wait + make clean + wait + done + done + done +done diff --git a/Microbenchmarks/WRAM/support/common.h b/Microbenchmarks/WRAM/support/common.h new file mode 100755 index 0000000..42914ac --- /dev/null +++ b/Microbenchmarks/WRAM/support/common.h @@ -0,0 +1,43 @@ +#ifndef _COMMON_H_ +#define _COMMON_H_ + +// Structures used by both the host and the dpu to communicate information +typedef struct { + uint32_t size; + enum kernels { + kernel1 = 0, + nr_kernels = 1, + } kernel; +} dpu_arguments_t; + +typedef struct { + uint64_t cycles; +} dpu_results_t; + +// Transfer size between MRAM and WRAM +#ifdef BL +#define BLOCK_SIZE_LOG2 BL +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#else +#define BLOCK_SIZE_LOG2 8 +#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2) +#define BL BLOCK_SIZE_LOG2 +#endif + +// Data type +#ifdef INT32 +#define T int32_t +#define DIV 2 // Shift right to divide by sizeof(T) +#else +#define T int64_t +#define DIV 3 // Shift right to divide by sizeof(T) +#endif + +#define PERF 1 // Use perfcounters? +#define PRINT 0 + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_RESET "\x1b[0m" + +#endif diff --git a/Microbenchmarks/WRAM/support/cyclecount.h b/Microbenchmarks/WRAM/support/cyclecount.h new file mode 100644 index 0000000..c4247b5 --- /dev/null +++ b/Microbenchmarks/WRAM/support/cyclecount.h @@ -0,0 +1,19 @@ +#include <perfcounter.h> + +// Timer +typedef struct perfcounter_cycles{ + perfcounter_t start; + perfcounter_t end; + perfcounter_t end2; + +}perfcounter_cycles; + +void timer_start(perfcounter_cycles *cycles){ + cycles->start = perfcounter_get(); // START TIMER +} + +uint64_t timer_stop(perfcounter_cycles *cycles){ + cycles->end = perfcounter_get(); // STOP TIMER + cycles->end2 = perfcounter_get(); // STOP TIMER + return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4); +} diff --git a/Microbenchmarks/WRAM/support/params.h b/Microbenchmarks/WRAM/support/params.h new file mode 100644 index 0000000..bbc2a19 --- /dev/null +++ b/Microbenchmarks/WRAM/support/params.h @@ -0,0 +1,61 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +typedef struct Params { + unsigned int input_size; + unsigned int stride; + int n_warmup; + int n_reps; + int exp; +}Params; + +static void usage() { + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\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=8K elements)" + "\n -s <S> stride (default=2)" + "\n"); +} + +struct Params input_params(int argc, char **argv) { + struct Params p; + p.input_size = 8 << 10; + p.stride = 2; + p.n_warmup = 1; + p.n_reps = 3; + p.exp = 0; + + int opt; + while((opt = getopt(argc, argv, "hi:w:e:x:s:")) >= 0) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + 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; + case 's': p.stride = atoi(optarg); break; + default: + fprintf(stderr, "\nUnrecognized option!\n"); + usage(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + assert((NR_TASKLETS & (NR_TASKLETS - 1)) == 0 && "Use a power-of-two number of tasklets!"); + + return p; +} +#endif diff --git a/Microbenchmarks/WRAM/support/timer.h b/Microbenchmarks/WRAM/support/timer.h new file mode 100755 index 0000000..eedc385 --- /dev/null +++ b/Microbenchmarks/WRAM/support/timer.h @@ -0,0 +1,59 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+
+typedef struct Timer{
+
+ struct timeval startTime[4];
+ struct timeval stopTime[4];
+ double time[4];
+
+}Timer;
+
+void start(Timer *timer, int i, int rep) {
+ if(rep == 0) {
+ timer->time[i] = 0.0;
+ }
+ gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i) {
+ gettimeofday(&timer->stopTime[i], NULL);
+ timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
+ (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)); }
|