diff options
Diffstat (limited to 'VA')
| -rw-r--r-- | VA/Makefile | 22 | ||||
| -rw-r--r-- | VA/host/app.c | 80 | ||||
| -rwxr-xr-x | VA/support/common.h | 14 | ||||
| -rw-r--r-- | VA/support/dfatool_host.ah | 106 | ||||
| -rw-r--r-- | VA/support/params.h | 4 | ||||
| -rwxr-xr-x | VA/support/timer.h | 26 |
6 files changed, 206 insertions, 46 deletions
diff --git a/VA/Makefile b/VA/Makefile index 040dd4a..a79969c 100644 --- a/VA/Makefile +++ b/VA/Makefile @@ -12,10 +12,28 @@ COMMON_INCLUDES := support HOST_SOURCES := $(wildcard host/*.c) DPU_SOURCES := $(wildcard dpu/*.c) +aspectc ?= 0 +aspectc_timing ?= * +dfatool_timing ?= 1 + +HOST_CC := ${CC} + COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES} -HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -D${TYPE} -DENERGY=${ENERGY} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -DWITH_DPUINFO=${WITH_DPUINFO} +HOST_FLAGS := ${COMMON_FLAGS} -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -D${TYPE} -DENERGY=${ENERGY} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -DWITH_DPUINFO=${WITH_DPUINFO} -DDFATOOL_TIMING=${dfatool_timing} -DASPECTC=${aspectc} DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} +ifeq (${aspectc_timing}, 1) + ASPECTC_HOST_FLAGS += -a support/dfatool_host.ah +endif + +ASPECTC_HOST_FLAGS ?= -a0 + +ifeq (${aspectc}, 1) + HOST_CC = ag++ -r repo.acp -v 0 ${ASPECTC_HOST_FLAGS} --c_compiler ${UPMEM_HOME}/bin/clang++ -p . --Xcompiler +else + HOST_FLAGS += -std=c11 +endif + QUIET = @ ifdef verbose @@ -28,7 +46,7 @@ bin: ${QUIET}mkdir -p bin bin/host_code: ${HOST_SOURCES} ${COMMON_INCLUDES} bin - ${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS} + ${QUIET}${HOST_CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS} bin/dpu_code: ${DPU_SOURCES} ${COMMON_INCLUDES} bin ${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES} diff --git a/VA/host/app.c b/VA/host/app.c index 1a2cdfd..2455642 100644 --- a/VA/host/app.c +++ b/VA/host/app.c @@ -7,8 +7,24 @@ #include <stdlib.h> #include <stdbool.h> #include <string.h> + +#if ASPECTC +extern "C" { +#endif + #include <dpu.h> #include <dpu_log.h> +#include <dpu_management.h> +#include <dpu_target_macros.h> + +#if ENERGY +#include <dpu_probe.h> +#endif + +#if ASPECTC +} +#endif + #include <unistd.h> #include <getopt.h> #include <assert.h> @@ -25,33 +41,29 @@ #define XSTR(x) STR(x) #define STR(x) #x -#if ENERGY -#include <dpu_probe.h> -#endif - -#include <dpu_management.h> -#include <dpu_target_macros.h> - // Pointer declaration static T *A; static T *B; static T *C; static T *C2; +struct Params p; +unsigned int kernel; + // Create input arrays -static void read_input(T *A, T *B, unsigned int nr_elements) +static void read_input(T *A, T *B, unsigned long int nr_elements) { srand(0); - for (unsigned int i = 0; i < nr_elements; i++) { + for (unsigned long int i = 0; i < nr_elements; i++) { A[i] = (T) (rand()); B[i] = (T) (rand()); } } // Compute output in the host -static void vector_addition_host(T *C, T *A, T *B, unsigned int nr_elements) +static void vector_addition_host(T *C, T *A, T *B, unsigned long int nr_elements) { - for (unsigned int i = 0; i < nr_elements; i++) { + for (unsigned long int i = 0; i < nr_elements; i++) { C[i] = A[i] + B[i]; } } @@ -60,7 +72,7 @@ static void vector_addition_host(T *C, T *A, T *B, unsigned int nr_elements) int main(int argc, char **argv) { - struct Params p = input_params(argc, argv); + p = input_params(argc, argv); struct dpu_set_t dpu_set, dpu; uint32_t nr_of_dpus; @@ -79,31 +91,37 @@ int main(int argc, char **argv) // Allocate DPUs and load binary #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); +#if DFATOOL_TIMING timer.time[0] = 0; // alloc #endif +#endif #if !WITH_LOAD_OVERHEAD DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks)); assert(nr_of_dpus == NR_DPUS); +#if DFATOOL_TIMING timer.time[1] = 0; // load #endif +#endif #if !WITH_FREE_OVERHEAD +#if DFATOOL_TIMING timer.time[6] = 0; // free #endif +#endif unsigned int i = 0; - const unsigned int input_size = + const unsigned long int input_size = p.exp == 0 ? p.input_size * NR_DPUS : p.input_size; - const unsigned int input_size_8bytes = ((input_size * sizeof(T)) % 8) != 0 ? roundup(input_size, 8) : input_size; // Input size per DPU (max.), 8-byte aligned - const unsigned int input_size_dpu = divceil(input_size, NR_DPUS); // Input size per DPU (max.) - const unsigned int input_size_dpu_8bytes = ((input_size_dpu * sizeof(T)) % 8) != 0 ? roundup(input_size_dpu, 8) : input_size_dpu; // Input size per DPU (max.), 8-byte aligned + const unsigned long int input_size_8bytes = ((input_size * sizeof(T)) % 8) != 0 ? roundup(input_size, 8) : input_size; // Input size per DPU (max.), 8-byte aligned + const unsigned long int input_size_dpu = divceil(input_size, NR_DPUS); // Input size per DPU (max.) + const unsigned long int input_size_dpu_8bytes = ((input_size_dpu * sizeof(T)) % 8) != 0 ? roundup(input_size_dpu, 8) : input_size_dpu; // Input size per DPU (max.), 8-byte aligned // Input/output allocation - A = malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); - B = malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); - C = malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); - C2 = malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); + A = (T*)malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); + B = (T*)malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); + C = (T*)malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); + C2 = (T*)malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T)); T *bufferA = A; T *bufferB = B; T *bufferC = C2; @@ -185,21 +203,21 @@ int main(int argc, char **argv) start(&timer, 3, 0); } // Input arguments - unsigned int kernel = 0; + kernel = 0; dpu_arguments_t input_arguments[NR_DPUS]; for (i = 0; i < nr_of_dpus - 1; i++) { input_arguments[i].size = input_size_dpu_8bytes * sizeof(T); input_arguments[i].transfer_size = input_size_dpu_8bytes * sizeof(T); - input_arguments[i].kernel = kernel; + input_arguments[i].kernel = (enum kernels)kernel; } input_arguments[nr_of_dpus - 1].size = (input_size_8bytes - input_size_dpu_8bytes * (NR_DPUS - 1)) * sizeof(T); input_arguments[nr_of_dpus - 1].transfer_size = input_size_dpu_8bytes * sizeof(T); - input_arguments[nr_of_dpus - 1].kernel = kernel; + input_arguments[nr_of_dpus - 1].kernel = (enum kernels)kernel; // Copy input arrays i = 0; @@ -306,22 +324,22 @@ int main(int argc, char **argv) printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n"); if (rep >= p.n_warmup) { - printf - ("[::] VA-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d n_elements_per_dpu=%d", + dfatool_printf + ("[::] VA-UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%lu n_elements_per_dpu=%lu", nr_of_dpus, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size, input_size / NR_DPUS); - printf + dfatool_printf (" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d numa_node_rank=%d ", WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD, numa_node_rank); - printf + dfatool_printf ("| latency_alloc_us=%f latency_load_us=%f latency_cpu_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f", timer.time[0], timer.time[1], timer.time[2], timer.time[3], timer.time[4], timer.time[5], timer.time[6]); - printf + dfatool_printf (" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f", input_size * 3 * sizeof(T) / timer.time[2], input_size * 3 * sizeof(T) / @@ -330,7 +348,7 @@ int main(int argc, char **argv) (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6])); - printf + dfatool_printf (" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f", input_size * 3 * sizeof(T) / (timer.time[3] + timer.time[4] + @@ -342,7 +360,7 @@ int main(int argc, char **argv) (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5])); - printf + dfatool_printf (" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f", input_size / timer.time[2], input_size / (timer.time[4]), @@ -352,7 +370,7 @@ int main(int argc, char **argv) timer.time[4] + timer.time[5] + timer.time[6])); - printf + dfatool_printf (" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n", input_size / (timer.time[3] + timer.time[4] + diff --git a/VA/support/common.h b/VA/support/common.h index cee09e2..6ce6e23 100755 --- a/VA/support/common.h +++ b/VA/support/common.h @@ -1,14 +1,20 @@ #ifndef _COMMON_H_ #define _COMMON_H_ +enum kernels { + kernel1 = 0, + nr_kernels = 1, +}; + // Structures used by both the host and the dpu to communicate information typedef struct { + /* + * Size per DPU cannot exceed 32 bit, as each DPU only has 64 MiB of memory + * (i.e., only needs 26 bit for addressing). + */ uint32_t size; uint32_t transfer_size; - enum kernels { - kernel1 = 0, - nr_kernels = 1, - } kernel; + enum kernels kernel; } dpu_arguments_t; // Transfer size between MRAM and WRAM diff --git a/VA/support/dfatool_host.ah b/VA/support/dfatool_host.ah new file mode 100644 index 0000000..f207ea2 --- /dev/null +++ b/VA/support/dfatool_host.ah @@ -0,0 +1,106 @@ +#pragma once + +#include <sys/time.h> + +aspect DfatoolHostTiming { + struct timeval starttime; + struct timeval stoptime; + uint32_t n_ranks = 0; + uint32_t n_dpus = 0; + + double const M_to_Mi = 1.048576; /* 2^20 / 1e6 */ + + advice call("% input_params(...)") : after() { + Params* p = tjp->result(); + printf("[--] n_dpus=%u n_elements=%lu e_exp=%d \n", NR_DPUS, p->input_size, p->exp); + } + + advice call("% dpu_get_nr_dpus(...)") : after() { + n_dpus = **(tjp->arg<1>()); + } + + advice call("% dpu_get_nr_ranks(...)") : after() { + n_ranks = **(tjp->arg<1>()); + } + + advice call("% dpu_alloc(...)") : around() { + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + n_dpus = *(tjp->arg<0>()); + printf("[::] dpu_alloc | n_dpus=%u | latency_us=%f\n", + n_dpus, + (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec) + ); + } + + advice call("% dpu_alloc_ranks(...)") : around() { + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + n_ranks = *(tjp->arg<0>()); + printf("[::] dpu_alloc_ranks | n_ranks=%u | latency_us=%f\n", + n_ranks, + (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec) + ); + } + + advice call("% dpu_load(...)") : around() { + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + printf("[::] dpu_load | n_dpus=%u n_ranks=%u | latency_us=%f\n", + n_dpus, n_ranks, + (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec) + ); + } + + advice call("% dpu_free(...)") : around() { + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + printf("[::] dpu_free | n_dpus=%u n_ranks=%u | latency_us=%f\n", + n_dpus, n_ranks, + (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec) + ); + } + + advice call("% dpu_launch(...)") : around() { + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + double latency_us = (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec); + unsigned long input_size = p.input_size; + printf("[::] dpu_launch | n_dpus=%u n_ranks=%u e_kernel=kernel%d n_elements=%lu | latency_us=%f throughput_Mrps=%f throughput_MiBps=%f\n", + n_dpus, n_ranks, + kernel + 1, + input_size, + latency_us, + input_size / latency_us, + input_size * sizeof(T) / (latency_us * M_to_Mi) + ); + } + + advice call("% dpu_push_xfer(...)") : around() { + size_t payload_size = *(tjp->arg<4>()); + gettimeofday(&starttime, NULL); + tjp->proceed(); + gettimeofday(&stoptime, NULL); + double time_us = (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec); + if (*(tjp->arg<1>()) == DPU_XFER_TO_DPU) { + printf("[::] dpu_push_to_dpu | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n", + n_dpus, n_ranks, + payload_size * n_dpus, payload_size, + time_us, + payload_size * n_dpus / (time_us * M_to_Mi) + ); + } else if (*(tjp->arg<1>()) == DPU_XFER_FROM_DPU) { + printf("[::] dpu_push_from_dpu | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n", + n_dpus, n_ranks, + payload_size * n_dpus, payload_size, + time_us, + payload_size * n_dpus / (time_us * M_to_Mi) + ); + } + } +}; diff --git a/VA/support/params.h b/VA/support/params.h index 47c10ef..4f41adc 100644 --- a/VA/support/params.h +++ b/VA/support/params.h @@ -4,7 +4,7 @@ #include "common.h" typedef struct Params { - unsigned int input_size; + unsigned long int input_size; int n_warmup; int n_reps; int exp; @@ -41,7 +41,7 @@ struct Params input_params(int argc, char **argv) exit(0); break; case 'i': - p.input_size = atoi(optarg); + p.input_size = atol(optarg); break; case 'w': p.n_warmup = atoi(optarg); diff --git a/VA/support/timer.h b/VA/support/timer.h index df68334..c0208d3 100755 --- a/VA/support/timer.h +++ b/VA/support/timer.h @@ -35,6 +35,8 @@ #include <sys/time.h> +#if DFATOOL_TIMING + typedef struct Timer { struct timeval startTime[7]; @@ -43,6 +45,8 @@ typedef struct Timer { } Timer; +#define dfatool_printf(fmt, ...) do { printf(fmt, __VA_ARGS__); } while (0) + void start(Timer *timer, int i, int rep) { if (rep == 0) { @@ -60,15 +64,23 @@ void stop(Timer *timer, int i) (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec); } -void print(Timer *timer, int i, int REP) +#else + +#define dfatool_printf(fmt, ...) do {} while (0) + +typedef int Timer; + +void start(Timer *timer, int i, int rep) { - printf("Time (ms): %f\t", timer->time[i] / (1000 * REP)); + (void)timer; + (void)i; + (void)rep; } -void printall(Timer *timer, int maxt) +void stop(Timer *timer, int i) { - for (int i = 0; i <= maxt; i++) { - printf(" timer%d_us=%f", i, timer->time[i]); - } - printf("\n"); + (void)timer; + (void)i; } + +#endif |