NVIDIA GeForce RTX 5070ti<\/a> (8960cores)<\/td><\/tr>\n<\/table><\/p>\n2) \u5b9f\u9a13\u5185\u5bb9<\/h2>\n
4\u30d0\u30a4\u30c8\u6d6e\u52d5\u5c0f\u6570\u70b9\u6570\u306e\u914d\u5217\u306e\u8db3\u3057\u7b97\u3092\u3084\u3089\u305b\u308b\u3002
\n\u5f62\u5f0f\u8a00\u8a9e\u7684\u306b\u66f8\u3051\u3070\u3053\u3093\u306a\u611f\u3058\u3060\u3002<\/p>\n
\r\nfloat A[1_000_000_000], B[1_000_000_000], C[1_000_000_000];\r\nC[i] = A[i] + B[i]; \/\/ for i = 0 to 999_999_999\r\n<\/pre>\n\u3082\u3057\u304f\u306f\u3001\u3053\u3093\u306a\u611f\u3058\u3060\u3002<\/p>\n
\r\n\u2200 i \u2208 [0, N): C[i] := A[i] + B[i]\r\n<\/pre>\n\u914d\u5217\u306e\u8981\u7d20\u6570\u306f 109<\/sup>\u500b\u3001\u3059\u306a\u308f\u3061 10\u5104\u500b<\/span> \u3060\u3002<\/p>\n\u300c4\u30d0\u30a4\u30c8\u6d6e\u52d5\u5c0f\u6570\u70b9\u6570\u306e\u8a08\u7b97\u3092\u624b\u8a08\u7b97\u3067 10\u5104\u56de\u3084\u3063\u3066\u304f\u3060\u3055\u3044\u300210\u5104\u5186\u3042\u3052\u307e\u3059\u3002\u300d<\/span>
\n\u3068\u8a00\u308f\u308c\u3066\u3082\u7d76\u5bfe\u306b\u65ad\u308b\u3002<\/p>\n\u4eba\u529b\u3067 5\u79d2\u306b\u4e00\u3064\u3001
\n\u5bdd\u305a\u306b\u8a08\u7b97\u3057\u305f\u3068\u3057\u3066 158\u5e74\u304b\u304b\u308b\u30fb\u30fb\u30fb<\/span><\/p>\n3) \u5b9f\u9a13\u7528\u30d7\u30ed\u30b0\u30e9\u30e0<\/h2>\n
CPU\u7528\u3001GPU\u7528\u306e\u4e8c\u7a2e\u985e\u306e\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u7528\u610f\u3057\u305f\u3002
\nCPU\u7248\u306e single-core or multi-core \u306f\u3001\u30b3\u30f3\u30d1\u30a4\u30eb\u30b9\u30a4\u30c3\u30c1\u6307\u5b9a\u3067\u5206\u3051\u308b\u3002<\/p>\n
(1) CPU\u7528\u5b9f\u9a13\u30b3\u30fc\u30c9<\/h3>\n
\u6ce8\u610f\uff1a<\/span>
\n\u3053\u306e\u30d7\u30ed\u30b0\u30e9\u30e0\u3067\u306f\u3001sizeof(float) x 10\u5104\u500b x 3\u30bb\u30c3\u30c8 = 12GB<\/span> \u306e\u30e1\u30e2\u30ea\u3092 allocate\u3057\u3066\u3044\u308b\u3002
\n\u30e1\u30e2\u30ea\u642d\u8f09\u91cf\u306e\u5c0f\u3055\u3044\u30b7\u30b9\u30c6\u30e0\u3067\u5b9f\u884c\u3059\u308b\u5834\u5408\u306f\u3001\u500b\u6570\u3092\u6e1b\u3089\u3059\u5fc5\u8981\u304c\u3042\u308b\u3002<\/p>\n\r\n#include <stdio.h>\r\n#include <stdlib.h>\r\n#include <time.h>\r\n\r\n#define N 1000000000\r\n\r\n#ifdef USE_OPENMP\r\n#include <omp.h>\r\n#endif\r\n\r\n\/*\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/*\/\r\nint main() {\r\n size_t size = N * sizeof(float);\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* \u30c7\u30fc\u30bf\u4f5c\u6210 *\/\r\n float *A = malloc(size);\r\n float *B = malloc(size);\r\n float *C = malloc(size);\r\n for (int i = 0; i < N; i++) {\r\n A[i] = 1.0f;\r\n B[i] = 2.0f;\r\n }\r\n \/*--------------------------------------------------------------------------*\/\r\n struct timespec start, end;\r\n clock_gettime(CLOCK_MONOTONIC, &start); \/* \u6642\u9593\u8a08\u6e2c\u958b\u59cb *\/\r\n \/****************************************************************************>>> \u8a08\u6e2c\u533a\u9593 *\/\r\n \/* \u6f14\u7b97\u5b9f\u884c *\/\r\n #pragma omp parallel for\r\n for (int i = 0; i < N; i++) {\r\n C[i] = A[i] + B[i];\r\n }\r\n \/****************************************************************************<<< \u8a08\u6e2c\u533a\u9593 *\/\r\n clock_gettime(CLOCK_MONOTONIC, &end); \/* \u6642\u9593\u8a08\u6e2c\u7d42\u4e86 *\/\r\n long sec_diff = end.tv_sec - start.tv_sec;\r\n long nsec_diff = end.tv_nsec - start.tv_nsec;\r\n double elapsed_ms = sec_diff * 1000.0 + nsec_diff \/ 1e6;\r\n printf("time: %.3f ms\\n", elapsed_ms);\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* \u30e1\u30e2\u30ea\u89e3\u653e *\/\r\n free(A); free(B); free(C);\r\n \/*--------------------------------------------------------------------------*\/\r\n return 0;\r\n}\r\n<\/pre>\nCPU\u7528\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3001Single thread\u7528\u30d7\u30ed\u30b0\u30e9\u30e0 cpu_1<\/span> \u3092\u751f\u6210\u3059\u308b\u3002<\/p>\n\r\n$ gcc cpu.c -o cpu_1<\/span>\r\n<\/pre>\nCPU\u7528\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3001Multi thread\u7528\u30d7\u30ed\u30b0\u30e9\u30e0 cpu_N<\/span> \u3092\u751f\u6210\u3059\u308b\u3002
\n#pragma omp parallel for \u3068\u66f8\u3044\u305f\u3068\u3053\u308d\u304c\u30de\u30eb\u30c1\u30b9\u30ec\u30c3\u30c9\u5b9f\u884c\u30b3\u30fc\u30c9\u306b\u5c55\u958b\u3055\u308c\u308b\u3002<\/p>\n\r\n$ gcc cpu.c -o cpu_N -fopenmp<\/span>\r\n<\/pre>\n(2) GPU\u7528\u5b9f\u9a13\u30b3\u30fc\u30c9<\/h3>\n
CPU\u7248\u3068\u6bd4\u8f03\u3059\u308b\u3068\u3001\u4ee5\u4e0b\u306e\u51e6\u7406\u304c\u5897\u3048\u305f\u5206\u3060\u3051\u30b3\u30fc\u30c9\u91cf\u304c\u591a\u3044\u3002
\n\u30fbGPU\u30e1\u30e2\u30ea\u78ba\u4fdd
\n\u30fbCPU \u2192 GPU\u30c7\u30fc\u30bf\u8ee2\u9001\uff08INPUT\uff09
\n\u30fbGPU\u7528\u6f14\u7b97\u30b3\u30fc\u30c9 (CUDA)
\n\u30fbGPU \u2192 CPU\u30c7\u30fc\u30bf\u8ee2\u9001\uff08OUTPUT\uff09<\/p>\n
\u307e\u305a\u306f\u3001\u4e0a\u8a18\u306e\u30aa\u30fc\u30d0\u30fc\u30d8\u30c3\u30c9\u90e8\u5206\u3092\u7121\u8996\u3057\u3001
\n\u7d14\u7c8b\u306b\u30d9\u30af\u30c8\u30eb\u8a08\u7b97\u3060\u3051\u3092\u5bfe\u8c61\u306b\u901f\u5ea6\u8a08\u6e2c\u3057\u3066\u307f\u308b\u3002<\/span><\/p>\n\u6ce8\u610f\uff1a<\/span>
\n\u3053\u306e\u30d7\u30ed\u30b0\u30e9\u30e0\u3067\u306f\u3001\u30db\u30b9\u30c8(CPU)\u5074\u3001\u30c7\u30d0\u30a4\u30b9(GPU)\u5074\u3067\u305d\u308c\u305e\u308c\u306b
\nsizeof(float) x 10\u5104\u500b x 3\u30bb\u30c3\u30c8 = 12GB<\/span> \u306e\u30e1\u30e2\u30ea\u3092 allocate\u3057\u3066\u3044\u308b\u3002
\n\u30e1\u30e2\u30ea\u642d\u8f09\u91cf\u306e\u5c0f\u3055\u3044\u30b7\u30b9\u30c6\u30e0\u3067\u5b9f\u884c\u3059\u308b\u5834\u5408\u306f\u3001\u500b\u6570\u3092\u6e1b\u3089\u3059\u5fc5\u8981\u304c\u3042\u308b\u3002<\/p>\n\r\n#include <stdio.h>\r\n#include <cuda_runtime.h>\r\n\r\n#define N 1000000000\r\n#define THREADS_PER_BLOCK 1024\r\n\r\n\/*\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/*\/\r\n__global__ void vector_add(float *A, float *B, float *C, int n) {\r\n int i = blockIdx.x * blockDim.x + threadIdx.x;\r\n if (i < n) {\r\n C[i] = A[i] + B[i];\r\n }\r\n}\r\n\r\n\/*\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/*\/\r\nint main() {\r\n size_t size = N * sizeof(float);\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* \u30c7\u30fc\u30bf\u4f5c\u6210 *\/\r\n float *A = (float*)malloc(size);\r\n float *B = (float*)malloc(size);\r\n float *C = (float*)malloc(size);\r\n for (int i = 0; i < N; i++) {\r\n A[i] = 1.0f;\r\n B[i] = 2.0f;\r\n }\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* CPU \u2192 GPU\u30c7\u30fc\u30bf\u8ee2\u9001 *\/\r\n float *d_A, *d_B, *d_C;\r\n cudaMalloc(&d_A, size);\r\n cudaMalloc(&d_B, size);\r\n cudaMalloc(&d_C, size);\r\n cudaMemcpy(d_A, A, size, cudaMemcpyHostToDevice);\r\n cudaMemcpy(d_B, B, size, cudaMemcpyHostToDevice);\r\n \/*--------------------------------------------------------------------------*\/\r\n struct timespec start, end;\r\n clock_gettime(CLOCK_MONOTONIC, &start); \/* \u6642\u9593\u8a08\u6e2c\u958b\u59cb *\/\r\n \/****************************************************************************>>> \u8a08\u6e2c\u533a\u9593 *\/\r\n \/* \u6f14\u7b97\u5b9f\u884c *\/\r\n int blocks = (N + THREADS_PER_BLOCK - 1) \/ THREADS_PER_BLOCK;\r\n vector_add <<< blocks, THREADS_PER_BLOCK >>> (d_A, d_B, d_C, N);\r\n cudaDeviceSynchronize(); \/* GPU\u5074\u306e\u51e6\u7406\u5b8c\u4e86\u3068\u540c\u671f *\/\r\n \/****************************************************************************<<< \u8a08\u6e2c\u533a\u9593 *\/\r\n clock_gettime(CLOCK_MONOTONIC, &end); \/* \u6642\u9593\u8a08\u6e2c\u7d42\u4e86 *\/\r\n long sec_diff = end.tv_sec - start.tv_sec;\r\n long nsec_diff = end.tv_nsec - start.tv_nsec;\r\n double elapsed_ms = sec_diff * 1000.0 + nsec_diff \/ 1e6;\r\n printf("time: %.3f ms\\n", elapsed_ms);\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* GPU \u2192 CPU\u30c7\u30fc\u30bf\u8ee2\u9001 *\/\r\n cudaMemcpy(C, d_C, size, cudaMemcpyDeviceToHost);\r\n \/*--------------------------------------------------------------------------*\/\r\n \/* \u30e1\u30e2\u30ea\u89e3\u653e *\/\r\n cudaFree(d_A); cudaFree(d_B); cudaFree(d_C);\r\n free(A); free(B); free(C);\r\n \/*--------------------------------------------------------------------------*\/\r\n return 0;\r\n}\r\n<\/pre>\nGPU\u7528\u30bd\u30fc\u30b9\u30d5\u30a1\u30a4\u30eb\u306f nvcc\uff08NVIDIA Cuda Compiler\uff09<\/span> \u3067\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3001\u5b9f\u884c\u30d7\u30ed\u30b0\u30e9\u30e0 gpu<\/span> \u3092\u751f\u6210\u3059\u308b\u3002<\/p>\n\r\n$ nvcc gpu.cu -o gpu<\/span>\r\n<\/pre>\n4) \u5b9f\u884c\u7d50\u679c<\/h2>\n
\u524d\u8ff0\u306e 3\u7a2e\u985e\u306e\u30d7\u30ed\u30b0\u30e9\u30e0\u306e\u5b9f\u884c\u7d50\u679c\u306f\u4ee5\u4e0b\u306e\u901a\u308a\u3002<\/p>\n
\r\n$ .\/cpu_1<\/span>\r\ntime: 1536.077 ms\r\n\r\n$ .\/cpu_N<\/span>\r\ntime: 206.356 ms\r\n\r\n$ .\/gpu<\/span>\r\ntime: 15.783 ms\r\n<\/pre>\n\u500b\u4eba\u7684\u306a\u671f\u5f85\u5024\u3088\u308a\u306f\u3060\u3044\u3076\u4f4e\u304b\u3063\u305f\u304c GPU\u304c\u9ad8\u901f\u3060\u3002<\/span><\/p>\nGPU\u306f\u3001CPU\uff08single-core\uff09\u306e 97\u500d<\/span>
\nGPU\u306f\u3001CPU\uff0820 multi-cores\uff09\u306e 13\u500d<\/span><\/p>\n5) \u30c7\u30fc\u30bf\u8ee2\u9001\u306e\u30aa\u30fc\u30d0\u30fc\u30d8\u30c3\u30c9\u6642\u9593\u3092\u542b\u3081\u3066\u8a08\u6e2c\u3057\u3066\u307f\u308b\u3002<\/h2>\n
\u4e0a\u306e\u65b9\u3067\u3001
\n\u307e\u305a\u306f\u3001\u4e0a\u8a18\u306e\u30aa\u30fc\u30d0\u30fc\u30d8\u30c3\u30c9\u90e8\u5206\u3092\u7121\u8996\u3057\u3001\u7d14\u7c8b\u306b\u30d9\u30af\u30c8\u30eb\u8a08\u7b97\u3060\u3051\u3092\u5bfe\u8c61\u306b\u901f\u5ea6\u8a08\u6e2c\u3057\u3066\u307f\u308b\u3002<\/span>
\n\u3068\u66f8\u3044\u305f\u304c\u3001\u5b9f\u904b\u7528\u3059\u308b\u5834\u5408\u306b\u306f\u3001\u3053\u306e\u30aa\u30fc\u30d0\u30fc\u30d8\u30c3\u30c9\u90e8\u5206\u3092\u7121\u8996\u3059\u308b\u308f\u3051\u306b\u306f\u3044\u304b\u306a\u3044\u3002<\/p>\n\u51e6\u7406\u6642\u9593\u306e\u8a08\u6e2c\u7bc4\u56f2\u3092\u30e1\u30e2\u30ea\u8ee2\u9001\u3092\u542b\u3080\u7bc4\u56f2\u306b\u5909\u66f4\u3057\u3066\u307f\u305f\u3002<\/span>
\n\u3053\u306e\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u5b9f\u884c\u3059\u308b\u3068\u30fb\u30fb\u30fb<\/p>\n\r\n$ .\/gpu<\/span>\r\ntime: 2020.624 ms\r\n<\/pre>\n\u3093\uff1f\uff1f\uff1f<\/span><\/p>\nCPU\uff08Single thread\uff09\u3088\u308a\u3082\u9045\u304f\u306a\u3063\u305f\u305e\u30fb\u30fb\u30fb<\/p>\n
\u4ee5\u4e0b\u306e\u500b\u6240\u306b +2005[ms]<\/span> \u306e\u6642\u9593\u304c\u304b\u304b\u3063\u305f\u3068\u3044\u3046\u3053\u3068\u3060\u3002<\/p>\n\r\n \/* CPU \u2192 GPU\u30c7\u30fc\u30bf\u8ee2\u9001 *\/\r\n float *d_A, *d_B, *d_C;\r\n cudaMalloc(&d_A, size);\r\n cudaMalloc(&d_B, size);\r\n cudaMalloc(&d_C, size);\r\n cudaMemcpy(d_A, A, size, cudaMemcpyHostToDevice);\r\n cudaMemcpy(d_B, B, size, cudaMemcpyHostToDevice);\r\n\uff1a\r\n \/* GPU \u2192 CPU\u30c7\u30fc\u30bf\u8ee2\u9001 *\/\r\n cudaMemcpy(C, d_C, size, cudaMemcpyDeviceToHost);\r\n<\/pre>\n\u30e1\u30a4\u30f3\u306e\u30d9\u30af\u30c8\u30eb\u8a08\u7b97\u81ea\u4f53\u306f\u305f\u3063\u305f\u306e 15.783[ms]<\/span> \u306a\u306e\u306b\u30fb\u30fb\u30fb
\n\u304a\u305d\u308b\u3079\u3057\u30e1\u30e2\u30ea\u8ee2\u9001\u30b3\u30b9\u30c8\u30fb\u30fb\u30fb<\/p>\n\u30103\u3011\u6240\u611f<\/h1>\n
CUDA\u3067\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u5b9f\u88c5\u3059\u308b\u969b\u306b\u306f\u3001
\nCPU(Host)-GPU(Device)\u9593\u306e\u30c7\u30fc\u30bf\u8ee2\u9001\u30b3\u30b9\u30c8\u306b\u6ce8\u610f\u3057\u307e\u3057\u3087\u3046\u3002
\nCPU(Host)-GPU(Device)\u9593\u306e\u30c7\u30fc\u30bf\u8ee2\u9001\u306e\u91cf\u3068\u983b\u5ea6\u3092\u6e1b\u3089\u305b\u308b\u3088\u3046\u306b\u3001\u8a2d\u8a08\u3092\u6700\u9069\u5316\u3057\u307e\u3057\u3087\u3046\u3002<\/span>
\n\u3068\u3044\u3046\u3053\u3068\u3060\u306a\u3002<\/p>\n\u7d9a\u304d\u306f\u3053\u3061\u3089\u30fb\u30fb\u30fb
\n(49) Python vs C\u8a00\u8a9e\u3067CUDA\u30d9\u30af\u30c8\u30eb\u8a08\u7b97\u901f\u5ea6\u3092\u6bd4\u8f03\u3059\u308b\u3002<\/a><\/p>\n
\n","protected":false},"excerpt":{"rendered":"Single processor, Multi processor\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002 Single-core, Multi-core\u3067\u306e\u6bd4\u8f03\u3067\u3059\u3002 \u30101\u3011\u3084\u308a\u305f\u3044\u3053\u3068 \u904e\u53bb\u8a18\u4e8b (46)\u3010Othello AI\u3011\u30aa\u30bb\u30edAI\u306b\u2026 \u7d9a\u304d\u3092\u8aad\u3080 »<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8,30,6,29],"tags":[],"class_list":["post-1721","post","type-post","status-publish","format-standard","hentry","category-cuda","category-openmp","category-6","category-29"],"views":769,"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/1721","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/comments?post=1721"}],"version-history":[{"count":58,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/1721\/revisions"}],"predecessor-version":[{"id":1989,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/1721\/revisions\/1989"}],"wp:attachment":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/media?parent=1721"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/categories?post=1721"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/tags?post=1721"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}