{"id":7440,"date":"2026-07-22T12:32:57","date_gmt":"2026-07-22T04:32:57","guid":{"rendered":"https:\/\/i007.cc\/wordpress\/?p=7440"},"modified":"2026-07-22T12:32:57","modified_gmt":"2026-07-22T04:32:57","slug":"simd-%e5%90%91%e9%87%8f%e5%8c%96%e4%bc%98%e5%8c%96","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/7440","title":{"rendered":"SIMD \u5411\u91cf\u5316\u4f18\u5316"},"content":{"rendered":"

\ud83d\udd37 \u4e00\u3001\u4ec0\u4e48\u662f SIMD\uff1f<\/h3>\n

SIMD = Single Instruction, Multiple Data<\/strong>\uff08\u5355\u6307\u4ee4\u591a\u6570\u636e\uff09<\/p>\n

\u666e\u901a\u6807\u91cf\u8ba1\u7b97\uff1a\u4e00\u6761\u6307\u4ee4\u5904\u7406\u4e00\u4e2a\u6570\u636e<\/p>\n

\n
\n
<\/div>\n<\/div>\n
\n
add r1, r2  \u2192  1\u6b21\u52a0\u6cd5\uff0c\u5904\u74061\u5bf9\u6570\u5b57<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
SIMD\uff1a\u4e00\u6761\u6307\u4ee4\u540c\u65f6\u5904\u7406\u591a\u4e2a\u6570\u636e<\/p>\n
\n
\n
<\/div>\n<\/div>\n
\n
vaddps ymm1, ymm2  \u2192  1\u6b21\u6307\u4ee4\uff0c\u540c\u65f6\u5904\u74068\u5bf9float\uff08AVX2\uff0c256\u4f4d\u5bc4\u5b58\u5668\uff09<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\n
<\/div>\n<\/div>\n
\n
\u6807\u91cf\u8ba1\u7b97\uff1a\r\n[a1] + [b1] = [c1]   \u2190 \u7b2c1\u6b21\u6307\u4ee4\r\n[a2] + [b2] = [c2]   \u2190 \u7b2c2\u6b21\u6307\u4ee4\r\n[a3] + [b3] = [c3]   \u2190 \u7b2c3\u6b21\u6307\u4ee4\r\n[a4] + [b4] = [c4]   \u2190 \u7b2c4\u6b21\u6307\u4ee4\r\n\uff084\u6761\u6307\u4ee4\uff09\r\n\r\nSIMD\u8ba1\u7b97\uff1a\r\n[a1 a2 a3 a4] + [b1 b2 b3 b4] = [c1 c2 c3 c4]\r\n\uff081\u6761\u6307\u4ee4\uff0c\u7406\u8bba\u4e0a\u5feb4\u500d\uff01\uff09<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\ud83d\udd37 \u4e8c\u3001SIMD \u6307\u4ee4\u96c6\u5bb6\u65cf\uff08\u6309\u5bc4\u5b58\u5668\u5bbd\u5ea6\uff09<\/h3>\n
\n\n\n\n\n\n\n\n
\u6307\u4ee4\u96c6<\/th>\n\u5bc4\u5b58\u5668\u5bbd\u5ea6<\/th>\n\u53ef\u5e76\u884c\u5904\u7406\u7684 float \u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n
SSE<\/td>\n128\u4f4d<\/td>\n4\u4e2a<\/td>\n<\/tr>\n
AVX\/AVX2<\/td>\n256\u4f4d<\/td>\n8\u4e2a<\/td>\n<\/tr>\n
AVX-512<\/td>\n512\u4f4d<\/td>\n16\u4e2a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n
\n
\n
<\/div>\n<\/div>\n
bash<\/div>\n
\n
# \u67e5\u770b\u4f60\u7684CPU\u652f\u6301\u54ea\u4e9b\u6307\u4ee4\u96c6\r\ncat \/proc\/cpuinfo | grep flags | head -1\r\n# \u770b\u5230 avx2\u3001sse4_2 \u7b49\u5b57\u6837<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\ud83d\udd37 \u4e09\u3001\u6700\u7b80\u5355\u7684\u65b9\u5f0f\uff1a\u8ba9\u7f16\u8bd1\u5668\u81ea\u52a8\u5411\u91cf\u5316<\/h3>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u7f16\u8bd1\u65f6\u52a0\u4e0a\u4f18\u5316\u9009\u9879\r\n\/\/ g++ -O3 -march=native myprogram.cpp\r\n\r\nvoid addArrays(float* a, float* b, float* result, int n) {\r\n    for (int i = 0; i < n; i++) {\r\n        result[i] = a[i] + b[i];   \/\/ \u7f16\u8bd1\u5668\u53ef\u80fd\u81ea\u52a8\u8f6c\u6210 SIMD \u6307\u4ee4\r\n    }\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
-O3<\/code> \u4f1a\u5f00\u542f\u81ea\u52a8\u5411\u91cf\u5316\uff0c-march=native<\/code> \u8ba9\u7f16\u8bd1\u5668\u4f7f\u7528\u4f60CPU\u652f\u6301\u7684\u6700\u9ad8\u6307\u4ee4\u96c6\u3002<\/p>\n
\u9a8c\u8bc1\u7f16\u8bd1\u5668\u662f\u5426\u771f\u7684\u5411\u91cf\u5316\u4e86<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
bash<\/div>\n
\n
g++ -O3 -march=native -S myprogram.cpp -o output.asm\r\ngrep -i \"vaddps\\|vmulps\" output.asm   # \u770b\u5230\u8fd9\u4e9b\u8bf4\u660e\u7528\u4e0a\u4e86AVX\u6307\u4ee4<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\u6216\u8005\u7528 Compiler Explorer<\/a> \u5728\u7ebf\u67e5\u770b\u6c47\u7f16\u3002<\/p>\n
\n
\ud83d\udd37 \u56db\u3001\u4e3a\u4ec0\u4e48\u7f16\u8bd1\u5668\u6709\u65f6”\u7f62\u5de5”\u4e0d\u5411\u91cf\u5316\uff1f<\/h3>\n
\u9677\u96311\uff1a\u6570\u636e\u4f9d\u8d56\uff08\u524d\u4e00\u6b21\u7ed3\u679c\u5f71\u54cd\u4e0b\u4e00\u6b21\uff09<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u274c \u65e0\u6cd5\u5411\u91cf\u5316\uff1asum \u4f9d\u8d56\u4e0a\u4e00\u6b21\u8fed\u4ee3\u7684\u7ed3\u679c\r\nfloat sum = 0;\r\nfor (int i = 0; i < n; i++) {\r\n    sum += a[i];   \/\/ \u6bcf\u6b21\u90fd\u4f9d\u8d56\u4e0a\u4e00\u6b21\u7684 sum\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\u89e3\u51b3\uff1a\u7528\u591a\u4e2a\u72ec\u7acb\u7684\u7d2f\u52a0\u5668\uff0c\u6700\u540e\u5408\u5e76\uff08\u6253\u7834\u4f9d\u8d56\u94fe\uff09\uff1a<\/p>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u2705 4\u8def\u5c55\u5f00\uff0c4\u4e2a\u72ec\u7acb\u7d2f\u52a0\u5668\uff0c\u53ef\u4ee5\u5e76\u884c\r\nfloat sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;\r\nint i = 0;\r\nfor (; i + 4 <= n; i += 4) {\r\n    sum0 += a[i];\r\n    sum1 += a[i+1];\r\n    sum2 += a[i+2];\r\n    sum3 += a[i+3];\r\n}\r\nfloat sum = sum0 + sum1 + sum2 + sum3;\r\nfor (; i < n; i++) sum += a[i];  \/\/ \u5904\u7406\u5269\u4f59\u7684<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\u9677\u96312\uff1a\u6307\u9488\u522b\u540d\uff08\u7f16\u8bd1\u5668\u4e0d\u786e\u5b9a\u5185\u5b58\u662f\u5426\u91cd\u53e0\uff09<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u274c \u7f16\u8bd1\u5668\u4e0d\u77e5\u9053 a\u3001b\u3001result \u662f\u5426\u6307\u5411\u91cd\u53e0\u5185\u5b58\uff0c\u4e0d\u6562\u968f\u610f\u91cd\u6392\u8ba1\u7b97\u987a\u5e8f\r\nvoid addArrays(float* a, float* b, float* result, int n) {\r\n    for (int i = 0; i < n; i++) {\r\n        result[i] = a[i] + b[i];\r\n    }\r\n}\r\n\r\n\/\/ \u2705 restrict \u5173\u952e\u5b57\u627f\u8bfa\"\u8fd9\u4e9b\u6307\u9488\u4e0d\u91cd\u53e0\"\uff0c\u7f16\u8bd1\u5668\u624d\u6562\u5927\u80c6\u5411\u91cf\u5316\r\nvoid addArrays(float* __restrict a, float* __restrict b, \r\n                float* __restrict result, int n) {\r\n    for (int i = 0; i < n; i++) {\r\n        result[i] = a[i] + b[i];\r\n    }\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\u9677\u96313\uff1a\u5faa\u73af\u91cc\u6709\u5206\u652f\u8df3\u8f6c<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u274c \u96be\u4ee5\u5411\u91cf\u5316\uff1a\u6bcf\u4e2a\u5143\u7d20\u8d70\u7684\u5206\u652f\u53ef\u80fd\u4e0d\u540c\r\nfor (int i = 0; i < n; i++) {\r\n    if (a[i] > 0) result[i] = a[i] * 2;\r\n    else result[i] = 0;\r\n}\r\n\r\n\/\/ \u2705 \u6539\u5199\u6210\u65e0\u5206\u652f\u5f62\u5f0f\uff08\u7528\u4e58\u6cd5\u4ee3\u66ffif\uff09\uff0c\u66f4\u5bb9\u6613\u5411\u91cf\u5316\r\nfor (int i = 0; i < n; i++) {\r\n    result[i] = (a[i] > 0) * a[i] * 2;   \/\/ bool \u81ea\u52a8\u8f6c 0\/1\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\ud83d\udd37 \u4e94\u3001\u624b\u52a8 SIMD\uff1aIntrinsics\uff08\u76f4\u63a5\u63a7\u5236\u5e95\u5c42\u6307\u4ee4\uff09<\/h3>\n
\u5f53\u7f16\u8bd1\u5668\u81ea\u52a8\u5411\u91cf\u5316\u4e0d\u591f\u7406\u60f3\u65f6\uff0c\u53ef\u4ee5\u624b\u5199 SIMD \u6307\u4ee4<\/strong>\u3002<\/p>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
#include <immintrin.h>   \/\/ AVX intrinsics\r\n\r\nvoid addArraysSIMD(float* a, float* b, float* result, int n) {\r\n    int i = 0;\r\n    \r\n    \/\/ \u6bcf\u6b21\u5904\u74068\u4e2afloat\uff08AVX2\uff0c256\u4f4d = 8\u00d732\u4f4d\uff09\r\n    for (; i + 8 <= n; i += 8) {\r\n        __m256 va = _mm256_loadu_ps(&a[i]);       \/\/ \u52a0\u8f7d8\u4e2afloat\u8fdb\u5bc4\u5b58\u5668\r\n        __m256 vb = _mm256_loadu_ps(&b[i]);\r\n        __m256 vresult = _mm256_add_ps(va, vb);   \/\/ \u4e00\u6761\u6307\u4ee4\u5b8c\u62108\u6b21\u52a0\u6cd5\r\n        _mm256_storeu_ps(&result[i], vresult);    \/\/ \u5b58\u56de\u5185\u5b58\r\n    }\r\n    \r\n    \/\/ \u5904\u7406\u5269\u4f59\u4e0d\u8db38\u4e2a\u7684\u5143\u7d20\uff08\u6807\u91cf\u65b9\u5f0f\u8865\u5c3e\uff09\r\n    for (; i < n; i++) {\r\n        result[i] = a[i] + b[i];\r\n    }\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\u6027\u80fd\u5bf9\u6bd4\u5b9e\u6d4b\uff08\u5178\u578b\u91cf\u7ea7\uff09<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u5904\u74061000\u4e07\u4e2afloat\u76f8\u52a0\r\n\r\n\/\/ \u6807\u91cf\u7248\u672c\r\nfor (int i = 0; i < n; i++) result[i] = a[i] + b[i];\r\n\/\/ \u8017\u65f6\uff1a~8ms\r\n\r\n\/\/ AVX2\u624b\u52a8\u5411\u91cf\u5316\r\naddArraysSIMD(a, b, result, n);\r\n\/\/ \u8017\u65f6\uff1a~2ms\uff08\u7ea64\u500d\u63d0\u5347\uff0c\u53d7\u5185\u5b58\u5e26\u5bbd\u9650\u5236\uff0c\u672a\u8fbe\u5230\u7406\u8bba8\u500d\uff09<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\u6ce8\u610f<\/strong>\uff1a\u5b9e\u9645\u52a0\u901f\u6bd4\u5f80\u5f80\u8fbe\u4e0d\u5230\u7406\u8bba\u503c\uff088\u500d\uff09\uff0c\u56e0\u4e3a\u5185\u5b58\u5e26\u5bbd\u3001\u6570\u636e\u52a0\u8f7d\/\u5b58\u50a8\u5f00\u9500\u4e5f\u662f\u74f6\u9888\uff0c\u4e0d\u662f\u5355\u7eaf\u7684\u8ba1\u7b97\u53d8\u5feb\u4e86\u5c31\u4e07\u4e8b\u5927\u5409\u3002<\/p>\n<\/blockquote>\n
\n
\ud83d\udd37 \u516d\u3001\u5e38\u7528 SIMD Intrinsics \u901f\u67e5<\/h3>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u52a0\u8f7d\/\u5b58\u50a8\r\n__m256 v = _mm256_loadu_ps(ptr);        \/\/ \u52a0\u8f7d8\u4e2afloat\uff08\u65e0\u5bf9\u9f50\u8981\u6c42\uff09\r\n__m256 v = _mm256_load_ps(ptr);         \/\/ \u52a0\u8f7d8\u4e2afloat\uff08\u8981\u6c4232\u5b57\u8282\u5bf9\u9f50\uff0c\u66f4\u5feb\uff09\r\n_mm256_storeu_ps(ptr, v);               \/\/ \u5b58\u50a8\r\n\r\n\/\/ \u7b97\u672f\u8fd0\u7b97\r\n_mm256_add_ps(a, b);    \/\/ \u52a0\u6cd5\r\n_mm256_sub_ps(a, b);    \/\/ \u51cf\u6cd5\r\n_mm256_mul_ps(a, b);    \/\/ \u4e58\u6cd5\r\n_mm256_div_ps(a, b);    \/\/ \u9664\u6cd5\r\n\r\n\/\/ \u6bd4\u8f83\u4e0e\u9009\u62e9\r\n_mm256_cmp_ps(a, b, _CMP_GT_OS);        \/\/ \u6bd4\u8f83\uff0c\u751f\u6210\u63a9\u7801\r\n_mm256_blendv_ps(a, b, mask);           \/\/ \u6839\u636e\u63a9\u7801\u9009\u62e9\r\n\r\n\/\/ \u6c34\u5e73\u6c42\u548c\uff08\u628a8\u4e2a\u6570\u52a0\u62101\u4e2a\uff0c\u5e38\u7528\u4e8ereduce\uff09\r\n__m256 v = ...;\r\n__m128 hi = _mm256_extractf128_ps(v, 1);\r\n__m128 lo = _mm256_castps256_ps128(v);\r\n__m128 sum = _mm_add_ps(hi, lo);\r\n\/\/ ...\u7ee7\u7eed\u6c34\u5e73\u76f8\u52a0\u76f4\u5230\u52691\u4e2a\u6570<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\u624b\u5199 Intrinsics \u4ee3\u7801\u5197\u957f\u3001\u53ef\u8bfb\u6027\u5dee\u3001\u8fd8\u8981\u9488\u5bf9\u4e0d\u540cCPU\u6307\u4ee4\u96c6\u5199\u591a\u4efd\u4ee3\u7801\u2014\u2014\u8fd9\u6b63\u662f\u4e3a\u4ec0\u4e48\u5b9e\u8df5\u4e2d\u66f4\u63a8\u8350\u7528\u4e13\u95e8\u7684\u5e93<\/strong>\u3002<\/p>\n<\/blockquote>\n
\n
\ud83d\udd37 \u4e03\u3001\u66f4\u53cb\u597d\u7684\u65b9\u5f0f\uff1aSIMD \u5c01\u88c5\u5e93<\/h3>\n
1. std::experimental::simd<\/code>\uff08C++\u6807\u51c6\u63d0\u6848\uff0c\u90e8\u5206\u7f16\u8bd1\u5668\u652f\u6301\uff09<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
#include <experimental\/simd>\r\nnamespace stdx = std::experimental;\r\n\r\nvoid addArrays(float* a, float* b, float* result, int n) {\r\n    using simd_t = stdx::native_simd<float>;\r\n    int width = simd_t::size();   \/\/ \u81ea\u52a8\u83b7\u53d6\u5f53\u524dCPU\u7684SIMD\u5bbd\u5ea6\r\n    \r\n    int i = 0;\r\n    for (; i + width <= n; i += width) {\r\n        simd_t va(a + i, stdx::element_aligned);\r\n        simd_t vb(b + i, stdx::element_aligned);\r\n        simd_t vresult = va + vb;   \/\/ \u50cf\u666e\u901a\u8fd0\u7b97\u7b26\u4e00\u6837\u5199\u4ee3\u7801\uff01\r\n        vresult.copy_to(result + i, stdx::element_aligned);\r\n    }\r\n    for (; i < n; i++) result[i] = a[i] + b[i];\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
2. \u7b2c\u4e09\u65b9\u5e93\uff1axsimd<\/code>\u3001highway<\/code>\uff08Google\uff09<\/h4>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
#include <xsimd\/xsimd.hpp>\r\n\r\nvoid addArrays(float* a, float* b, float* result, int n) {\r\n    using batch = xsimd::batch<float>;\r\n    std::size_t inc = batch::size;\r\n    \r\n    std::size_t i = 0;\r\n    for (; i < n - n % inc; i += inc) {\r\n        batch va = batch::load_unaligned(&a[i]);\r\n        batch vb = batch::load_unaligned(&b[i]);\r\n        batch vres = va + vb;\r\n        vres.store_unaligned(&result[i]);\r\n    }\r\n    for (; i < n; i++) result[i] = a[i] + b[i];\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\u4f18\u52bf<\/strong>\uff1a\u5199\u4e00\u4efd\u4ee3\u7801\uff0c\u5e93\u81ea\u52a8\u9002\u914d SSE\/AVX2\/AVX-512\/ARM NEON \u7b49\u4e0d\u540c\u786c\u4ef6\uff0c\u4e0d\u7528\u624b\u5199\u591a\u4efd Intrinsics\u3002<\/p>\n<\/blockquote>\n
\n
\ud83d\udd37 \u516b\u3001SIMD \u53cb\u597d\u7684\u6570\u636e\u5e03\u5c40\uff08\u547c\u5e94\u4e0a\u4e00\u8282 SoA\uff09<\/h3>\n
\n
\n
<\/div>\n<\/div>\n
cpp<\/div>\n
\n
\/\/ \u274c AoS\uff1aSIMD\u96be\u4ee5\u52a0\u8f7d\uff0c\u56e0\u4e3ax\/y\/z\u4ea4\u9519\u6392\u5217\uff0c\u4e0d\u8fde\u7eed\r\nstruct Point { float x, y, z; };\r\nstd::vector<Point> points;\r\n\r\n\/\/ \u2705 SoA\uff1ax\u5750\u6807\u8fde\u7eed\u5b58\u50a8\uff0c\u53ef\u4ee5\u76f4\u63a5SIMD\u6279\u91cf\u52a0\u8f7d\r\nstruct PointsSoA {\r\n    std::vector<float> x, y, z;\r\n};\r\n\r\nvoid scalePoints(PointsSoA& p, float factor) {\r\n    addArraysSIMD(p.x.data(), ..., factor, p.x.size());  \/\/ \u76f4\u63a5SIMD\u5904\u7406\u6574\u4e2ax\u6570\u7ec4\r\n}<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u4e0a\u8282\u8bb2\u7684 SoA \u548c\u8fd9\u8282\u7684 SIMD \u662f”\u5929\u4f5c\u4e4b\u5408”<\/strong>\u2014\u2014SIMD \u8981\u6c42\u6570\u636e\u8fde\u7eed\u4e14\u540c\u8d28\uff0cSoA \u6070\u597d\u63d0\u4f9b\u4e86\u8fd9\u79cd\u5e03\u5c40\u3002<\/p>\n<\/blockquote>\n
\n
\ud83d\udd37 \u4e5d\u3001\u4f55\u65f6\u4e0d\u503c\u5f97\u624b\u52a8 SIMD\uff1f<\/h3>\n
\n
\n
<\/div>\n<\/div>\n
\n
\u2705 \u503c\u5f97\u624b\u52a8\u4f18\u5316\u7684\u573a\u666f\uff1a\r\n- \u5927\u89c4\u6a21\u6570\u503c\u8ba1\u7b97\uff08\u79d1\u5b66\u8ba1\u7b97\u3001\u56fe\u5f62\u5b66\u3001\u97f3\u89c6\u9891\u5904\u7406\uff09\r\n- \u5df2\u786e\u8ba4\u662f\u6027\u80fd\u74f6\u9888\uff08\u7528 perf \u9a8c\u8bc1\u8fc7\uff09\r\n- \u6570\u636e\u89c4\u6a21\u5927\uff08\u51e0\u5343\u4e2a\u5143\u7d20\u4ee5\u4e0a\u624d\u80fd\u644a\u8584\u5f00\u9500\uff09\r\n\r\n\u274c \u4e0d\u503c\u5f97\u7684\u573a\u666f\uff1a\r\n- \u6570\u636e\u91cf\u5f88\u5c0f\uff08SIMD\u7684\"\u542f\u52a8\u5f00\u9500\"\u53ef\u80fd\u8d85\u8fc7\u6536\u76ca\uff09\r\n- \u903b\u8f91\u590d\u6742\u3001\u5206\u652f\u591a\uff08\u5411\u91cf\u5316\u6536\u76ca\u6709\u9650\uff0c\u4ee3\u7801\u53ef\u8bfb\u6027\u5927\u5e45\u4e0b\u964d\uff09\r\n- \u8fd8\u6ca1\u786e\u8ba4\u662f\u74f6\u9888\uff08\u8fc7\u65e9\u4f18\u5316\u662f\u4e07\u6076\u4e4b\u6e90\uff09<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\ud83d\udd37 \u5b8c\u6574\u4f18\u5316\u5c42\u6b21\u56de\u987e<\/h3>\n
\n
\n
<\/div>\n<\/div>\n
\n
\u7b2c1\u5c42\uff1a\u9009\u5bf9\u7b97\u6cd5\u548c\u6570\u636e\u7ed3\u6784\uff08O(n log n) vs O(n\u00b2)\uff09\u2192 \u6536\u76ca\u6700\u5927\r\n\u7b2c2\u5c42\uff1aCache-friendly \u5e03\u5c40\uff08SoA\u3001\u8fde\u7eed\u5185\u5b58\uff09       \u2192 \u901a\u5e382-4\u500d\u63d0\u5347\r\n\u7b2c3\u5c42\uff1aSIMD \u5411\u91cf\u5316                                \u2192 \u901a\u5e382-8\u500d\u63d0\u5347\uff08\u53e0\u52a0\u5728\u7b2c2\u5c42\u4e4b\u4e0a\uff09\r\n\u7b2c4\u5c42\uff1a\u624b\u5199\u6c47\u7f16\/\u6781\u81f4\u6307\u4ee4\u8c03\u4f18                       \u2192 \u8fb9\u9645\u6536\u76ca\u9012\u51cf\uff0c\u7ef4\u62a4\u6210\u672c\u5267\u589e<\/pre>\n
 <\/p>\n<\/div>\n<\/div>\n
\n
\u9ec4\u91d1\u6cd5\u5219<\/strong>\uff1a\u5148\u7528 perf<\/code> \u627e\u5230\u771f\u6b63\u7684\u74f6\u9888\uff0c\u786e\u8ba4\u662f”\u8ba1\u7b97\u5bc6\u96c6\u578b\u5faa\u73af”\uff0c\u518d\u8003\u8651 SIMD\uff1b\u5426\u5219\u5f88\u53ef\u80fd\u4f18\u5316\u4e86\u4e00\u6bb5\u6839\u672c\u4e0d\u5f71\u54cd\u6574\u4f53\u6027\u80fd\u7684\u4ee3\u7801\u3002<\/p>\n<\/blockquote>\n
\n
\u4e00\u53e5\u8bdd\u603b\u7ed3<\/h3>\n
\n
SIMD \u7684\u672c\u8d28\u662f”\u7528\u4e00\u6761\u6307\u4ee4\u505a\u591a\u4efd\u76f8\u540c\u7684\u5de5\u4f5c”<\/strong>\uff0c\u524d\u63d0\u662f\u6570\u636e\u8fde\u7eed\u3001\u64cd\u4f5c\u4e00\u81f4\u3001\u6ca1\u6709\u5206\u652f\u4f9d\u8d56\u3002\u5b83\u548c cache-friendly \u8bbe\u8ba1\uff08\u5c24\u5176\u662f SoA\uff09\u662f\u5929\u7136\u642d\u6863\u2014\u2014\u597d\u7684\u5185\u5b58\u5e03\u5c40\u662f SIMD \u80fd\u53d1\u6325\u5a01\u529b\u7684\u57fa\u7840\uff0c\u4e8c\u8005\u7ed3\u5408\u624d\u80fd\u771f\u6b63\u69a8\u5e72\u73b0\u4ee3 CPU \u7684\u7b97\u529b\u3002<\/p>\n<\/blockquote>\n
\u8981\u4e0d\u8981\u770b\u770b\u4e00\u4e2a\u5b8c\u6574\u7684\u5b9e\u6218\u6848\u4f8b\u2014\u2014\u7528 Google Benchmark \u5b9e\u6d4b\u5bf9\u6bd4<\/strong>“\u6807\u91cf vs \u81ea\u52a8\u5411\u91cf\u5316 vs \u624b\u52a8SIMD”\u4e09\u79cd\u5b9e\u73b0\u7684\u6027\u80fd\u5dee\u5f02\uff1f\u6211\u53ef\u4ee5\u5e2e\u4f60\u5199\u51fa\u53ef\u76f4\u63a5\u7f16\u8bd1\u8fd0\u884c\u7684\u4ee3\u7801\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\ud83d\udd37 \u4e00\u3001\u4ec0\u4e48\u662f SIMD\uff1f SIMD =<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[24],"tags":[],"class_list":["post-7440","post","type-post","status-publish","format-standard","hentry","category-value_docs"],"_links":{"self":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/7440","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/comments?post=7440"}],"version-history":[{"count":1,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/7440\/revisions"}],"predecessor-version":[{"id":7441,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/7440\/revisions\/7441"}],"wp:attachment":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/media?parent=7440"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/categories?post=7440"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/tags?post=7440"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}