{"id":7253,"date":"2019-04-24T21:10:49","date_gmt":"2019-04-24T12:10:49","guid":{"rendered":"http:\/\/umezawa.dyndns.info\/wordpress\/?p=7253"},"modified":"2019-04-29T02:05:16","modified_gmt":"2019-04-28T17:05:16","slug":"utf-8%e3%81%ae%e3%82%b3%e3%83%bc%e3%83%89%e3%83%9d%e3%82%a4%e3%83%b3%e3%83%88%e3%81%af%e3%81%a9%e3%81%86%e3%82%84%e3%81%a3%e3%81%a6avx-512%e3%81%a7%e9%ab%98%e9%80%9f%e3%81%ab%e6%95%b0%e3%81%88","status":"publish","type":"post","link":"http:\/\/umezawa.dyndns.info\/wordpress\/?p=7253","title":{"rendered":"UTF-8\u306e\u30b3\u30fc\u30c9\u30dd\u30a4\u30f3\u30c8\u306f\u3069\u3046\u3084\u3063\u3066AVX-512\u3067\u9ad8\u901f\u306b\u6570\u3048\u308b\u304b"},"content":{"rendered":"<p>\u524d\u306e\u8a18\u4e8b\u3067\u3042\u308b<a href=\"http:\/\/umezawa.dyndns.info\/wordpress\/?p=7236\">UTF-8\u306e\u30b3\u30fc\u30c9\u30dd\u30a4\u30f3\u30c8\u306f\u3069\u3046\u3084\u3063\u3066\u3082\u3063\u3068\u9ad8\u901f\u306b\u6570\u3048\u308b\u304b<\/a>\u3067 AVX2 \u3067\u306e\u9ad8\u901f\u5316\u3092\u3084\u308a\u307e\u3057\u305f\u304c\u3001\u4eca\u3067\u306f\u3055\u3089\u306b\u30d9\u30af\u30bf\u306e\u9577\u3044 AVX-512 \u3068\u3044\u3046\u3082\u306e\u304c\u3042\u308b\u306e\u3067\u3001\u3053\u308c\u3067\u3082\u3084\u3063\u3066\u307f\u3088\u3046\u304b\u3068\u601d\u3044\u307e\u3059\u3002\u306a\u304a\u3001\u60f3\u5b9a\u3059\u308b\u30d7\u30ed\u30bb\u30c3\u30b5\u306f Skylake-X \u3067\u3059\u3002<\/p>\n<p><!--more--><\/p>\n<p>\u307e\u305a\u3001zmm \u30ec\u30b8\u30b9\u30bf\u5411\u3051\u306e\u6c34\u5e73\u52a0\u7b97\u95a2\u6570\u3092\u5148\u306b\u66f8\u3044\u3066\u304a\u304d\u307e\u3059\u3002<\/p>\n<blockquote>\n<pre>\r\ninline int32_t avx512_horizontal_sum_epi8(__m512i x)\r\n{\r\n    __m512i sumhi = _mm512_unpackhi_epi8(x, _mm512_setzero_si512());\r\n    __m512i sumlo = _mm512_unpacklo_epi8(x, _mm512_setzero_si512());\r\n    __m512i sum16x32 = _mm512_add_epi16(sumhi, sumlo);\r\n    __m256i sum16x16 = _mm256_add_epi16(_mm512_castsi512_si256(sum16x32), _mm512_extracti64x4_epi64(sum16x32, 1));\r\n    __m128i sum16x8 = _mm_add_epi16(_mm256_castsi256_si128(sum16x16), _mm256_extracti128_si256(sum16x16, 1));\r\n    __m128i sum16x4 = _mm_add_epi16(sum16x8, _mm_srli_si128(sum16x8, 8));\r\n    uint64_t tmp = _mm_cvtsi128_si64(sum16x4);\r\n    tmp += (tmp &gt;&gt; 32);\r\n    tmp += (tmp &gt;&gt; 16);\r\n    return tmp & 0xffff;\r\n}\r\n<\/pre>\n<\/blockquote>\n<h3>\u524d\u7f6e\u304d<\/h3>\n<p>AVX-512 \u304c AVX2 \u307e\u3067\u3068\u7570\u306a\u308b\u70b9\u3068\u3057\u3066\u3001\u30d9\u30af\u30bf\u6bd4\u8f03\u547d\u4ee4\u306e\u7d50\u679c\u3084\u30d6\u30ec\u30f3\u30c9\u547d\u4ee4\u306e\u5236\u5fa1\u30aa\u30da\u30e9\u30f3\u30c9\u306a\u3069\u306e\u3001\u5b9f\u8cea\u7684\u306b bool \u306e\u30d9\u30af\u30bf\u3067\u3042\u308b\u3088\u3046\u306a\u5024\u3092\u4fdd\u6301\u3059\u308b\u305f\u3081\u306b\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u3068\u3044\u3046\u3082\u306e\u3092\u4f7f\u3046\u3088\u3046\u306b\u306a\u3063\u305f\u3001\u3068\u3044\u3046\u70b9\u304c\u6319\u3052\u3089\u308c\u307e\u3059\u3002<\/p>\n<p>\u524d\u306e\u8a18\u4e8b\u3067\u306f\u6700\u7d42\u7684\u306b VPCMPGTB \u547d\u4ee4\u306e\u7d50\u679c\u3092 VPSUBB \u306b\u76f4\u63a5\u6e21\u3059\u3068\u3044\u3046\u624b\u6cd5\u306b\u7740\u5730\u3057\u307e\u3057\u305f\u304c\u3001\u3053\u308c\u306f VPCMPGTB \u306e\u7d50\u679c\u304c ymm \u30ec\u30b8\u30b9\u30bf\u306b\u5165\u308b\u304b\u3089\u3067\u304d\u305f\u3053\u3068\u3067\u3059\u3002 AVX-512 \u3067\u306f\u7d50\u679c\u304c\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u306b\u5165\u308b\u305b\u3044\u3067\u3053\u306e\u624b\u6cd5\u306f\u76f4\u63a5\u4f7f\u3048\u306a\u3044\u305f\u3081\u3001\u5225\u306e\u65b9\u6cd5\u3092\u8003\u3048\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<h3>POPCNT \u306b\u5e30\u7740\u3059\u308b<\/h3>\n<p>AVX-512 \u306e\u6bd4\u8f03\u547d\u4ee4\u306e\u7d50\u679c\u3068\u3057\u3066\u5f97\u3089\u308c\u308b\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u306f\u524d\u8ff0\u306e\u3068\u304a\u308a bool \u306e\u30d9\u30af\u30bf\u3067\u3059\u304c\u3001\u6700\u7d42\u7684\u306b\u3084\u308a\u305f\u3044\u3053\u3068\u306f true \u306e\u6570\u3092\u6570\u3048\u308b\u3053\u3068\u3067\u3059\u3002\u3053\u308c\u306f\u3064\u307e\u308a\u7acb\u3063\u3066\u3044\u308b\u30d3\u30c3\u30c8\u3092\u6570\u3048\u308b\u3053\u3068\u306b\u4ed6\u306a\u3089\u306a\u3044\u306e\u3067\u3001\u6c4e\u7528\u30ec\u30b8\u30b9\u30bf\u306b\u30b3\u30d4\u30fc\u3057\u3066 POPCNT \u3059\u308b\u3001\u3068\u3044\u3046\u306e\u304c\u81ea\u7136\u306a\u767a\u60f3\u3067\u3057\u3087\u3046\u3002\u3068\u3044\u3046\u308f\u3051\u3067\u3053\u3093\u306a\u30b3\u30fc\u30c9\u304c\u51fa\u6765\u4e0a\u304c\u308a\u307e\u3059\u3002<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_cmpgt_popcnt(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz; i += 64) {\r\n        __m512i s = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i));\r\n        __mmask64 m = _mm512_cmpgt_epi8_mask(s, _mm512_set1_epi8(-0x41));\r\n        result += _mm_popcnt_u64(m);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<p>\u30eb\u30fc\u30d7\u304c\u4e00\u91cd\u3067\u6e08\u3080\u306e\u3067\u898b\u305f\u76ee\u304c\u30b7\u30f3\u30d7\u30eb\u306a\u306e\u304c\u5229\u70b9\u3067\u3059\u3002<\/p>\n<h3>VPMOVM2B \u3092\u4f7f\u3046<\/h3>\n<p>\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u306e\u5024\u3092 AVX2 \u98a8\u306e\u30d9\u30af\u30bf\u306e\u5024\u306b\u5909\u63db\u3059\u308b VPMOVM2B \u3068\u3044\u3046\u547d\u4ee4\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u308c\u3092\u4f7f\u3046\u30681\u547d\u4ee4\u5897\u3048\u308b\u3082\u306e\u306e AVX2 \u306e\u6642\u3068\u5168\u304f\u540c\u3058\u624b\u6cd5\u304c\u4f7f\u3048\u307e\u3059\u3002<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_cmpgt_movm2b_sub(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz;) {\r\n        __m512i sum = _mm512_setzero_si512();\r\n        size_t j = 0;\r\n        size_t limit = std::min&lt;size_t&gt;(255 * 64, sz - i);\r\n        for (; j &lt; limit; j += 64) {\r\n            __m512i s = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j));\r\n            __mmask64 m = _mm512_cmpgt_epi8_mask(s, _mm512_set1_epi8(-0x41));\r\n            sum = _mm512_sub_epi8(sum, _mm512_movm_epi8(m));\r\n        }\r\n        i += j;\r\n        result += avx512_horizontal_sum_epi8(sum);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<h3>VPMOVDQU8 \u3092\u4f7f\u3046<\/h3>\n<p>\u524d\u63b2\u306e VPMOVM2B \u547d\u4ee4\u3067\u3059\u304c\u3001\u307f\u306a\u3055\u3093\u304a\u4e16\u8a71\u306b\u306a\u3063\u3066\u3044\u308b <a href=\"https:\/\/www.agner.org\/\">Agner Fog<\/a> \u3055\u3093\u304c\u51fa\u3057\u3066\u3044\u308b<a href=\"https:\/\/www.agner.org\/optimize\/\">\u6700\u9069\u5316\u60c5\u5831\u30ea\u30bd\u30fc\u30b9<\/a>\u306e\u547d\u4ee4\u8868 (4. Instruction tables) \u3092\u898b\u308b\u3068\u3001\u3044\u304b\u306b\u3082\u5358\u7d14\u306a\u547d\u4ee4\u306a\u306e\u306b\u306a\u3093\u3068\u30ec\u30a4\u30c6\u30f3\u30b7\u304c3\u3082\u3042\u308a\u307e\u3059\u3002\u4e00\u65b9\u3001VPMOVDQU8 \u3068\u3044\u3046\u30de\u30b9\u30af\u4ed8\u304d mov \u3067\u3082\u540c\u3058\u3053\u3068\u304c\u3067\u304d\u3066\u3001\u3053\u3061\u3089\u306f\u666e\u901a\u306b\u30ec\u30a4\u30c6\u30f3\u30b7\u304c1\u3067\u3059\uff08\u6b63\u78ba\u306b\u306f\u547d\u4ee4\u8868\u306b\u306f 8\/16 \u306e\u547d\u4ee4\u306b\u3064\u3044\u3066\u306f\u306f\u63b2\u8f09\u3055\u308c\u3066\u3044\u306a\u3044\u306e\u3067\u3059\u304c\u3001 32\/64 \u306e\u547d\u4ee4\u306f\u63b2\u8f09\u3055\u308c\u3066\u3044\u3066\u305d\u308c\u3068\u7570\u306a\u308b\u7406\u7531\u306f\u306a\u3055\u305d\u3046\u306a\u306e\u3067\u3053\u3046\u66f8\u3044\u3066\u3044\u307e\u3059\uff09\u3002\u3068\u3044\u3046\u308f\u3051\u3067 VPMOVDQU8 \u3092\u4f7f\u3046\u3068 VPMOVM2B \u3088\u308a\u30ec\u30a4\u30c6\u30f3\u30b7\u304c\u6539\u5584\u3055\u308c\u305d\u3046\u306b\u601d\u308f\u308c\u307e\u3059\u3002<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_cmpgt_movdqu8_add(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz;) {\r\n        __m512i sum = _mm512_setzero_si512();\r\n        size_t j = 0;\r\n        size_t limit = std::min&lt;size_t&gt;(255 * 64, sz - i);\r\n        for (; j &lt; limit; j += 64) {\r\n            __m512i s = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j));\r\n            __mmask64 m = _mm512_cmpgt_epi8_mask(s, _mm512_set1_epi8(-0x41));\r\n            sum = _mm512_add_epi8(sum, _mm512_maskz_mov_epi8(m, _mm512_set1_epi8(1)));\r\n        }\r\n        i += j;\r\n        result += avx512_horizontal_sum_epi8(sum);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<h3>\u30de\u30b9\u30af\u4ed8\u304d\u52a0\u7b97\u3092\u4f7f\u3046<\/h3>\n<p>\u4e0a\u306e2\u3064\u3067\u306f\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u3092\u30d9\u30af\u30bf\u30ec\u30b8\u30b9\u30bf\u306b\u5909\u63db\u3057\u3066\u304b\u3089\u52a0\u7b97\u3057\u3066\u307e\u3057\u305f\u304c\u3001\u52a0\u7b97\u306b\u5bfe\u3057\u3066\u3082\u30de\u30b9\u30af\u304c\u4f7f\u3048\u308b\u306e\u3060\u304b\u3089\u7d20\u76f4\u306b\u30de\u30b9\u30af\u4ed8\u304d\u52a0\u7b97\u3092\u3057\u3066\u3057\u307e\u3048\u3070\u8a71\u304c\u5358\u7d14\u3067\u3059\u3002<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_cmpgt_maskadd(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz;) {\r\n        __m512i sum = _mm512_setzero_si512();\r\n        size_t j = 0;\r\n        size_t limit = std::min&lt;size_t&gt;(255 * 64, sz - i);\r\n        for (; j &lt; limit; j += 64) {\r\n            __m512i s = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j));\r\n            __mmask64 m = _mm512_cmpgt_epi8_mask(s, _mm512_set1_epi8(-0x41));\r\n            sum = _mm512_mask_add_epi8(sum, m, sum, _mm512_set1_epi8(1));\r\n        }\r\n        i += j;\r\n        result += avx512_horizontal_sum_epi8(sum);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<p>\u305f\u3060\u3057\u3001\u3053\u306e\u624b\u6cd5\u3092\u53d6\u308b\u5834\u5408\u3001\u52a0\u7b97\u306e\u3068\u3053\u308d\u306b\u6ce8\u610f\u304c\u5fc5\u8981\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u3002\u30de\u30b9\u30af\u306a\u3057\u547d\u4ee4\u3068\u30de\u30b9\u30af\u4ed8\u304d\u547d\u4ee4\u3068\u3067\u30ec\u30a4\u30c6\u30f3\u30b7\u304c\u7570\u306a\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b\u304b\u3089\u3067\u3059\u3002\u5b9f\u969b\u3001\u30b5\u30a4\u30dc\u30a6\u30ba\u30fb\u30e9\u30dc\u306e\u5149\u6210\u6ecb\u751f\u3055\u3093\u306f <a href=\"https:\/\/www.slideshare.net\/herumi\/avx512\/13\">AVX-512 \u8a73\u89e3\u306e\u767a\u8868\u306b\u304a\u3044\u3066\u30de\u30b9\u30af\u3092\u4f7f\u3046\u3068\u300c\u5c11\u3057\u4f4e\u901f\u300d\u3068\u8a00\u3063\u3066\u3044\u307e\u3059<\/a>\u3002<\/p>\n<p>\u3068\u306f\u3044\u3048\u8ffd\u52a0\u306e\u30ec\u30a4\u30c6\u30f3\u30b7\u304c\u8907\u6570\u30af\u30ed\u30c3\u30af\u304b\u304b\u308b\u3068\u3082\u601d\u3048\u306a\u3044\u306e\u3067\u3001\u30ec\u30a4\u30c6\u30f3\u30b7\u304c1\u304b\u30892\u306b\u306a\u308b\u3082\u306e\u3068\u3057\u3066\u305d\u308c\u3092\u56de\u907f\u3059\u308b\u3088\u3046\u306b\u624b\u52d5\u30672\u500d\u306b\u30a2\u30f3\u30ed\u30fc\u30ea\u30f3\u30b0\u3092\u304b\u3051\u3066\u307f\u307e\u3057\u3087\u3046\uff08\u3068\u308a\u3042\u3048\u305a\u3053\u306e\u90e8\u5206\u306b\u306f\u30b3\u30f3\u30d1\u30a4\u30e9\u306b\u3088\u308b\u30a2\u30f3\u30ed\u30fc\u30ea\u30f3\u30b0\u306f\u671f\u5f85\u3057\u306a\u3044\u3082\u306e\u3068\u3057\u307e\u3059\uff09<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_cmpgt_maskadd128(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz;) {\r\n        __m512i sum0 = _mm512_setzero_si512();\r\n        __m512i sum1 = _mm512_setzero_si512();\r\n        size_t j = 0;\r\n        size_t limit = std::min&lt;size_t&gt;(255 * 128, sz - i);\r\n        for (; j &lt; limit; j += 128) {\r\n            __m512i s0 = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j));\r\n            __m512i s1 = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j + 64));\r\n            __mmask64 m0 = _mm512_cmpgt_epi8_mask(s0, _mm512_set1_epi8(-0x41));\r\n            __mmask64 m1 = _mm512_cmpgt_epi8_mask(s1, _mm512_set1_epi8(-0x41));\r\n            sum0 = _mm512_mask_add_epi8(sum0, m0, sum0, _mm512_set1_epi8(1));\r\n            sum1 = _mm512_mask_add_epi8(sum1, m1, sum1, _mm512_set1_epi8(1));\r\n        }\r\n        i += j;\r\n        result += avx512_horizontal_sum_epi8(sum0);\r\n        result += avx512_horizontal_sum_epi8(sum1);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<h3>VPSHUFB \u306b\u56de\u5e30\u3059\u308b<\/h3>\n<p>\u4eca\u307e\u3067\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u3092\u3069\u3046\u3084\u3063\u3066\u30d9\u30af\u30bf\u30ec\u30b8\u30b9\u30bf\u306b\u5909\u63db\u3059\u308b\u304b\u3092\u8003\u3048\u3066\u3044\u307e\u3057\u305f\u304c\u3001\u5927\u5143\u306e\u8a18\u4e8b\u306b\u66f8\u3044\u3066\u3042\u308b\u624b\u6cd5\u3067\u3042\u308b VPSHUFB \u3092\u4f7f\u3063\u305f\u65b9\u6cd5\u306a\u3089\u7d50\u679c\u306f\u6700\u521d\u304b\u3089\u30d9\u30af\u30bf\u30ec\u30b8\u30b9\u30bf\u3067\u5f97\u3089\u308c\u308b\u306e\u3067\u3001\u5358\u7d14\u306b\u8db3\u3059\u3060\u3051\u3067\u3059\u3002\u3068\u308a\u3042\u3048\u305a\u3053\u308c\u3082\u5b9f\u88c5\u3057\u3066\u304a\u304d\u307e\u3057\u3087\u3046\u3002<\/p>\n<p>\u4f59\u8ac7\u3067\u3059\u304c\u3001intrinsic \u3068\u3057\u3066\u306f \\_mm_setr_epi8 \u3084 \\_mm128_setr_epi8 \u306f\u3042\u3063\u3066\u3082\u4f55\u6545\u304b \\_mm512_setr_epi8 \u306f\u306a\u304f\u3066 \\_mm512_set_epi8 \u3057\u304b\u306a\u304f\u3001\u9762\u98df\u3089\u3044\u307e\u3057\u305f\u3002<a href=\"https:\/\/software.intel.com\/sites\/landingpage\/IntrinsicsGuide\/#\">Intel \u306e intrinsic guide<\/a> \u306b\u3082\u306a\u3044\u3093\u3067\u3059\u304c\u3001\u306a\u3093\u3067\u3067\u3059\u304b\u306d\u2026\uff1f<\/p>\n<blockquote>\n<pre>\r\nsize_t avx512_pshufb_add(const char *p, size_t sz)\r\n{\r\n    size_t result = 0;\r\n    for (size_t i = 0; i &lt; sz;) {\r\n        __m512i sum = _mm512_setzero_si512();\r\n        size_t j = 0;\r\n        size_t limit = std::min&lt;size_t&gt;(255 * 64, sz - i);\r\n        for (; j &lt; limit; j += 64) {\r\n            const __m512i table = _mm512_set_epi8(\r\n                1, 1, 1, 1,             \/\/ 0xF .. 0xC\r\n                0, 0, 0, 0,             \/\/ 0xB .. 0x8\r\n                1, 1, 1, 1, 1, 1, 1, 1, \/\/ 0x7 ..\r\n                1, 1, 1, 1,             \/\/ 0xF .. 0xC\r\n                0, 0, 0, 0,             \/\/ 0xB .. 0x8\r\n                1, 1, 1, 1, 1, 1, 1, 1, \/\/ 0x7 ..\r\n                1, 1, 1, 1,             \/\/ 0xF .. 0xC\r\n                0, 0, 0, 0,             \/\/ 0xB .. 0x8\r\n                1, 1, 1, 1, 1, 1, 1, 1, \/\/ 0x7 ..\r\n                1, 1, 1, 1,             \/\/ 0xF .. 0xC\r\n                0, 0, 0, 0,             \/\/ 0xB .. 0x8\r\n                1, 1, 1, 1, 1, 1, 1, 1  \/\/ 0x7 ..\r\n            );\r\n            __m512i s = _mm512_load_si512(reinterpret_cast&lt;const __m512i *&gt;(p + i + j));\r\n            s = _mm512_and_si512(_mm512_srli_epi16(s, 4), _mm512_set1_epi8(0x0F));\r\n            s = _mm512_shuffle_epi8(table, s);\r\n            sum = _mm512_add_epi8(sum, s);\r\n        }\r\n        i += j;\r\n        result += avx512_horizontal_sum_epi8(sum);\r\n    }\r\n    return result;\r\n}\r\n<\/pre>\n<\/blockquote>\n<h3>\u901f\u3044\u306e\u306f\u3069\u308c\u304b<\/h3>\n<p>\u3067\u3001\u7d50\u5c40\u3069\u308c\u304c\u901f\u3044\u306e\u304b\u3067\u3059\u304c\u2026\u5b9f\u6a5f\u304c\u7121\u3044\u306e\u3067\u5206\u304b\u308a\u307e\u305b\u3093\u3002\u304c\u3001\u305d\u308c\u3067\u7d42\u308f\u308b\u306e\u3082\u3069\u3046\u304b\u3068\u601d\u3046\u306e\u3067\u3001\u3072\u3068\u307e\u305a\u30b3\u30f3\u30d1\u30a4\u30e9\u304c\u5410\u3044\u305f\u30d0\u30a4\u30ca\u30ea\u3092\u773a\u3081\u3066\u307f\u307e\u3059\u3002\u30b3\u30f3\u30d1\u30a4\u30eb\u30aa\u30d7\u30b7\u30e7\u30f3\u306f <code>\/O2 \/clang:-march=skylake-avx512 \/clang:-mtune=skylake-avx512<\/code> \u3067\u3059\u3002<\/p>\n<p>\u307e\u305a\u3001avx512_cmpgt_popcnt \u3092\u30eb\u30fc\u30d7\u30a2\u30f3\u30ed\u30fc\u30ea\u30f3\u30b0\u3057\u3066\u304f\u308c\u3066\u3044\u307e\u305b\u3093\u3002 <code>#pragma unroll<\/code> \u3092\u4ed8\u3051\u308b\u3068\u3001\u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u304c\u30a2\u30f3\u30ed\u30fc\u30eb\u3092\u5b9f\u884c\u3067\u304d\u306a\u304b\u3063\u305f\u3068\u8b66\u544a\u3092\u51fa\u3057\u3066\u304d\u307e\u3059\u3002\u3053\u308c\u3092\u4ed6\u306e\u30eb\u30fc\u30c1\u30f3\u3068\u540c\u69d8\u306a\u5f62\u306e\u4e8c\u91cd\u30eb\u30fc\u30d7\u306b\u3059\u308b\u3068\u30a2\u30f3\u30ed\u30fc\u30ea\u30f3\u30b0\u3057\u3066\u304f\u308c\u308b\u306e\u3067\u3059\u304c\u3001\u4e8c\u91cd\u30eb\u30fc\u30d7\u3058\u3083\u306a\u3044\u3068\u30a2\u30f3\u30ed\u30fc\u30ea\u30f3\u30b0\u3067\u304d\u306a\u3044\u3068\u304b\u3058\u3083\u306a\u3044\u3088\u306a\u3041\u2026<\/p>\n<p>\u30de\u30b9\u30af\u30ec\u30b8\u30b9\u30bf\u3092\u5143\u306b\u30d9\u30af\u30bf\u52a0\u7b97\u3059\u308b\u624b\u6cd5\uff08avx512_cmpgt_movm2b_sub\u3001avx512_cmpgt_movdqu8_add\u3001avx512_cmpgt_maskadd\u3001avx512_cmpgt_maskadd128 \u306e4\u3064\uff09\u3092\u898b\u3066\u307f\u308b\u3068\u3001\u5168\u3066\u30de\u30b9\u30af\u3092 VMOVDQU8 \u3067\u30d9\u30af\u30bf\u306b\u5909\u63db\u3057\u3066\u52a0\u7b97\u3059\u308b\u30d0\u30a4\u30ca\u30ea\uff08\u3064\u307e\u308a avx512_cmpgt_movdqu8_add \u3067\u671f\u5f85\u3057\u305f\u306e\u3068\u7b49\u4fa1\u306a\u51e6\u7406\uff09\u306b\u5909\u63db\u3055\u308c\u3066\u3044\u307e\u3057\u305f\u3002VPMOVM2B \u304c VMOVDQU8 \u306b\u5909\u63db\u3055\u308c\u308b\u3068\u3053\u308d\u307e\u3067\u306f\u3042\u308b\u7a0b\u5ea6\u4e88\u6e2c\u306e\u7bc4\u56f2\u5185\u3067\u3057\u305f\u304c\u3001\u30de\u30b9\u30af\u4ed8\u304d\u52a0\u7b97\u306e\u30b3\u30fc\u30c9\u3092\u7f6e\u304d\u63db\u3048\u3066\u304f\u308b\u3068\u306f\u601d\u3044\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u4e0b\u624b\u306a\u8003\u3048\u4f11\u3080\u306b\u4f3c\u305f\u308a\u3068\u3044\u3046\u3053\u3068\u306e\u3088\u3046\u3067\u3059\u3002<\/p>\n<h3>\u307e\u3068\u3081<\/h3>\n<ul>\n<li>AVX-512 \u3067\u306e\u3061\u3087\u3063\u3068\u3057\u305f\u6700\u9069\u5316\u624b\u6cd5\u306b\u3064\u3044\u3066\u66f8\u304d\u307e\u3057\u305f\u304c\u3001\u3042\u307e\u308a\u3046\u307e\u304f\u884c\u3063\u305f\u6c17\u306f\u3057\u307e\u305b\u3093\u3002<\/li>\n<li>Clang \u3059\u3054\u3044\u3063\u3059\u306d\uff0810\u65e5\u3076\u308a2\u56de\u76ee\uff09<\/li>\n<\/ul>\n<h3>Appendix<\/h3>\n<p><a href=\"https:\/\/gist.github.com\/umezawatakeshi\/407b1607b880c5c6a6257d80e61d8353\">\u30bd\u30fc\u30b9\u30b3\u30fc\u30c9\uff08Windows + Visual Studio 2017 + LLVM integration \u5411\u3051\u306a\u306e\u3067 Unix \u5411\u3051\u306b\u306f\u591a\u5c11\u4fee\u6b63\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\uff09<\/a><\/p>\n<p>\u5b9f\u6a5f\u306f\u6301\u3063\u3066\u307e\u305b\u3093\u304c Intel SDE \u3067\u6b63\u3057\u3044\u7d50\u679c\u304c\u5f97\u3089\u308c\u308b\u3053\u3068\u306f\u78ba\u8a8d\u3057\u3066\u3042\u308a\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u306e\u8a18\u4e8b\u3067\u3042\u308bUTF-8\u306e\u30b3\u30fc\u30c9\u30dd\u30a4\u30f3\u30c8\u306f\u3069\u3046\u3084\u3063\u3066\u3082\u3063\u3068\u9ad8\u901f\u306b\u6570\u3048\u308b\u304b\u3067 AVX2 \u3067\u306e\u9ad8\u901f\u5316\u3092\u3084\u308a\u307e\u3057\u305f\u304c\u3001\u4eca\u3067\u306f\u3055\u3089\u306b\u30d9\u30af\u30bf\u306e\u9577\u3044 AVX-512 \u3068\u3044\u3046\u3082\u306e\u304c\u3042\u308b\u306e\u3067\u3001\u3053\u308c\u3067\u3082\u3084\u3063\u3066\u307f\u3088\u3046\u304b\u3068\u601d\u3044\u307e\u3059\u3002\u306a\u304a\u3001\u60f3\u5b9a [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[27],"tags":[],"class_list":["post-7253","post","type-post","status-publish","format-standard","hentry","category-technology"],"_links":{"self":[{"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7253","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7253"}],"version-history":[{"count":6,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7253\/revisions"}],"predecessor-version":[{"id":7260,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7253\/revisions\/7260"}],"wp:attachment":[{"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7253"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7253"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/umezawa.dyndns.info\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7253"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}