\n\n\n\n
2025\u5e746\u670830\u65e5<\/p>\n\n\n\n
Robert Mueller-Albrecht<\/strong><\/p>\n\n\n\n \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u30fb\u30d7\u30ed\u30c0\u30af\u30c8\u30fb\u30de\u30fc\u30b1\u30c6\u30a3\u30f3\u30b0\u30fb\u30de\u30cd\u30fc\u30b8\u30e3\u30fc<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb \u30b3\u30fc\u30dd\u30ec\u30fc\u30b7\u30e7\u30f3<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae oneAPI \u30d9\u30fc\u30b9\u30fb\u30c4\u30fc\u30eb\u30ad\u30c3\u30c8<\/a>\u3001\u30a4\u30f3\u30c6\u30eb\u00ae HPC \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8<\/a>\u3001\u30a4\u30f3\u30c6\u30eb\u00ae AI \u30c4\u30fc\u30eb<\/a> (\u82f1\u8a9e) \u306e\u524d\u56de\u306e\u30ea\u30ea\u30fc\u30b9\u304b\u3089 3 \u30ab\u6708\u304c\u7d4c\u904e\u3057\u3001\u305d\u308c\u3089\u306e\u30b3\u30f3\u30dd\u30fc\u30cd\u30f3\u30c8\u306b\u591a\u304f\u306e\u6a5f\u80fd\u62e1\u5f35\u304c\u8ffd\u52a0\u3055\u308c\u307e\u3057\u305f\u3002\u4eca\u56de\u306e\u4e3b\u306a\u7126\u70b9\u306f\u3001HPC \u3068 AI \u306b\u304a\u3051\u308b\u9ad8\u5ea6\u306a\u30de\u30eb\u30c1\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u4e26\u5217\u51e6\u7406\u306e\u305f\u3081\u306e\u3001\u3088\u308a\u5b8c\u5168\u3067\u4f7f\u3044\u3084\u3059\u3044\u6a5f\u80fd\u3068\u76f8\u4e92\u904b\u7528\u6027\u306e\u5305\u62ec\u7684\u306a\u30bb\u30c3\u30c8\u3092\u5b9f\u73fe\u3057\u3001AI\u3001\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u3001\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u5411\u3051\u306b\u6700\u9069\u5316\u3055\u308c\u305f\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3068\u751f\u7523\u6027\u3092\u63d0\u4f9b\u3059\u308b\u3053\u3068\u3067\u3057\u305f\u3002<\/p>\n\n\n\n \u3053\u306e\u6700\u65b0\u30ea\u30ea\u30fc\u30b9\u306e\u7279\u7b46\u3059\u3079\u304d\u30a4\u30ce\u30d9\u30fc\u30b7\u30e7\u30f3\u306e\u4e00\u90e8\u3092\u6b21\u306b\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae HPC \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae \u30c7\u30fc\u30bf\u30bb\u30f3\u30bf\u30fc GPU \u306e\u30c7\u30d0\u30a4\u30b9\u30b3\u30fc\u30c9\u3067\u30a4\u30f3\u30c6\u30eb\u00ae Fortran \u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u306e\u5f37\u5316\u3055\u308c\u305f LLVM \u30e1\u30e2\u30ea\u30fc\u30fb\u30b5\u30cb\u30bf\u30a4\u30b6\u30fc<\/a>\u3092\u30b5\u30dd\u30fc\u30c8\u3057\u307e\u3057\u305f\u3002\u512a\u308c\u305f\u751f\u7523\u6027\u3068\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3001\u5805\u7262\u306a\u30e1\u30e2\u30ea\u30fc\u306e\u4f7f\u7528\u304c\u4fdd\u8a3c\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n \u6700\u65b0\u306e C++ 26\u3001OpenMP 6.0\u3001Fortran 2023\u3001MPI 4.1 \u6a5f\u80fd\u306e\u5b9f\u88c5\u3001\u968e\u5c64\u578b\u5206\u6563\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u306b\u304a\u3051\u308b\u4e26\u5217\u51e6\u7406\u3067\u306e MPI\u3001SYCL\u3001OpenMP \u306e\u30b7\u30fc\u30e0\u30ec\u30b9\u306a\u5171\u5b58\u306a\u3069\u3001\u9032\u5316\u3092\u7d9a\u3051\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n Fortran 2023 \u3068 OpenMP 6.0 \u306b\u3064\u3044\u3066<\/p>\n\n\n\n \u4e3b\u306a\u6a5f\u80fd\u6539\u5584\u306e\u6982\u8981\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30c4\u30fc\u30eb\u6700\u65b0\u60c5\u5831<\/a>\u306e\u4ee5\u4e0b\u306e\u30c8\u30d4\u30c3\u30af\u3067\u7d39\u4ecb\u3055\u308c\u3066\u3044\u307e\u3059\u3002 \u3082\u3061\u308d\u3093\u3001\u958b\u767a\u30c4\u30fc\u30eb\u306e\u3059\u3079\u3066\u306e\u30b3\u30f3\u30dd\u30fc\u30cd\u30f3\u30c8\u306f\u3001\u6700\u65b0\u304a\u3088\u3073\u6b21\u4e16\u4ee3\u306e\u30a4\u30f3\u30c6\u30eb\u00ae Xeon\u00ae \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc\u3001\u30a4\u30f3\u30c6\u30eb\u00ae Core™ \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc\u3001\u30a4\u30f3\u30c6\u30eb\u00ae Arc™ \u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u306e\u30ab\u30d0\u30ec\u30c3\u30b8\u3092\u5f15\u304d\u7d9a\u304d\u62e1\u5927\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306b\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae VTune™ \u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u5206\u6790\u6a5f\u80fd\u3078\u306e\u65b0\u3057\u3044\u30b7\u30f3\u30b0\u30eb\u30bd\u30b1\u30c3\u30c8\u306e\u30a4\u30f3\u30c6\u30eb\u00ae Xeon\u00ae 6 SoC (\u958b\u767a\u30b3\u30fc\u30c9\u540d Granite Rapids-D) (\u30b3\u30f3\u30d1\u30af\u30c8\u306a\u30d5\u30a9\u30fc\u30e0\u30d5\u30a1\u30af\u30bf\u30fc\u306e\u30a4\u30f3\u30c6\u30eb\u00ae Xeon\u00ae \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc\u30fb\u30d1\u30c3\u30b1\u30fc\u30b8\u3067\u3001\u9ad8\u5e2f\u57df\u5e45\u306e\u7523\u696d\u7528\u30b2\u30fc\u30c8\u30a6\u30a7\u30a4\u3084\u6a5f\u5185\u822a\u7a7a\u7ba1\u5236\u30b7\u30b9\u30c6\u30e0\u306a\u3069\u3067\u5e83\u304f\u4f7f\u7528\u3055\u308c\u3066\u3044\u307e\u3059) \u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u8a8d\u8b58\u306e\u8ffd\u52a0\u3082\u542b\u307e\u308c\u307e\u3059\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae \u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u306b\u5c0e\u5165\u3055\u308c\u305f\u6a5f\u80fd\u306b\u306f\u3001SYCL \u3068 OpenMP \u306e\u63d0\u6848\u3055\u308c\u305f\u62e1\u5f35\u306e\u521d\u671f\u5b9f\u88c5\u304c\u542b\u307e\u308c\u307e\u3059\u3002<\/p>\n\n\n\n \u8a66\u9a13\u7684 SYCL \u540d\u524d\u7a7a\u9593 sycl::ext::oneAPI::experimental::matrix \u306f\u3001\u30b8\u30e7\u30a4\u30f3\u30c8\u884c\u5217\u6f14\u7b97\u3067\u4f7f\u7528\u3055\u308c\u308b\u65b0\u3057\u3044\u30c7\u30fc\u30bf\u578b\u3092\u5b9a\u7fa9\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u4e00\u822c\u7684\u306a\u30e1\u30e2\u30ea\u30fc\u30ed\u30fc\u30c9\u3001\u30b9\u30c8\u30a2\u3001\u30dd\u30a4\u30f3\u30bf\u30fc\u6f14\u7b97\u306f\u3001SYCL \u30ef\u30fc\u30af\u30b0\u30eb\u30fc\u30d7\u5168\u4f53\u306b\u9069\u7528\u3067\u304d\u3001\u30c7\u30fc\u30bf\u3092\u30d7\u30ea\u30d5\u30a7\u30c3\u30c1\u3057\u3066\u52b9\u7387\u7684\u306b\u5b9f\u884c\u3067\u304d\u307e\u3059\u3002\u30dd\u30a4\u30f3\u30bf\u30fc\u306f\u30aa\u30fc\u30d0\u30fc\u30ed\u30fc\u30c9\u53ef\u80fd\u3067\u3059\u3002<\/p>\n\n\n\n \u884c\u5217\u306e\u4e57\u7b97\u3068\u52a0\u7b97 (D=A*B+C) \u306f\u3001\u8981\u7d20\u3054\u3068\u306e\u6f14\u7b97\u3001\u6d3b\u6027\u5316\u95a2\u6570\u3001(\u9006) \u91cf\u5b50\u5316\u3001\u30c7\u30fc\u30bf\u578b\u5909\u63db\u3092\u542b\u3080\u884c\u5217\u306e\u30b3\u30d4\u30fc\u306a\u3069\u3001\u57fa\u672c\u7684\u306a\u64cd\u4f5c\u3092\u3059\u3079\u3066\u30b5\u30dd\u30fc\u30c8\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n SYCL \u884c\u5217\u3068\u30c6\u30f3\u30bd\u30eb\u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u3067\u8a08\u7b97\u30aa\u30d6\u30b8\u30a7\u30af\u30c8\u306e\u30d7\u30ed\u30d1\u30c6\u30a3\u30fc\u3092\u52b9\u7387\u7684\u306b\u8868\u73fe\u3057\u3066\u64cd\u4f5c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u3001\u30b7\u30df\u30e5\u30ec\u30fc\u30b7\u30e7\u30f3\u3001\u9ad8\u5ea6\u306a\u6570\u5024\u5fae\u5206\u65b9\u7a0b\u5f0f\u3001\u30d3\u30b8\u30e5\u30a2\u30eb\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u3001CGI VFX\u3001\u91d1\u878d\u30ea\u30b9\u30af\u5206\u6790\u3001\u30b2\u30fc\u30e0\u3001AI \u306a\u3069\u306e\u5e83\u7bc4\u306a\u8a08\u7b97\u3092\u9ad8\u901f\u5316\u3059\u308b\u30d3\u30b8\u30e7\u30f3\u3068\u6b63\u5473\u306e\u52b9\u679c\u304c\u5f97\u3089\u308c\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u8a08\u7b97\u306f\u3059\u3079\u3066\u3001SPIR-V \u5354\u8abf\u884c\u5217\u62e1\u5f35\u3068\u4e92\u63db\u6027\u306e\u3042\u308b\u884c\u5217\u304a\u3088\u3073\u30c6\u30f3\u30bd\u30eb\u6f14\u7b97\u7528\u306e\u7d71\u5408\u30a4\u30f3\u30bf\u30fc\u30d5\u30a7\u30a4\u30b9\u306b\u3088\u308a\u884c\u308f\u308c\u307e\u3059\u3002<\/p>\n\n\n\n \u3053\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u306b\u3088\u308a\u3001\u7d71\u5408\u30b8\u30e7\u30a4\u30f3\u30c8\u884c\u5217\u304c\u672c\u8cea\u7684\u306b\u30b9\u30b1\u30fc\u30e9\u30d6\u30eb\u306b\u306a\u308a\u3001\u7570\u306a\u308b\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u3084\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u9593\u3067\u79fb\u690d\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n \u7d71\u5408\u30b8\u30e7\u30a4\u30f3\u30c8\u884c\u5217\u306e SYCL \u62e1\u5f35\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u8a18\u4e8b\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002 OpenMP \u306f\u3001\u79fb\u690d\u6027\u3068\u30b9\u30b1\u30fc\u30e9\u30d3\u30ea\u30c6\u30a3\u30fc\u306b\u512a\u308c\u305f\u30e2\u30c7\u30eb\u3092\u63d0\u4f9b\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u3066\u3001\u30e9\u30c3\u30d7\u30c8\u30c3\u30d7\u304b\u3089\u30b9\u30fc\u30d1\u30fc\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fc\u307e\u3067\u3001\u3055\u307e\u3056\u307e\u306a\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u7528\u306e\u4e26\u5217\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3092\u958b\u767a\u3059\u308b\u305f\u3081\u306e\u67d4\u8edf\u306a\u30a4\u30f3\u30bf\u30fc\u30d5\u30a7\u30a4\u30b9\u3092\u30d7\u30ed\u30b0\u30e9\u30de\u30fc\u306b\u63d0\u4f9b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u4e3b\u306a\u6a5f\u80fd\u306f\u3001\u4ee5\u4e0b\u306e\u3068\u304a\u308a\u3067\u3059\u3002<\/p>\n\n\n\n \u8907\u6570\u306e\u30bd\u30b1\u30c3\u30c8\u3001\u8907\u6570\u306e\u30aa\u30d5\u30ed\u30fc\u30c9\u30fb\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u30fb\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u30fb\u30e2\u30c7\u30eb\u3001\u305d\u3057\u3066\u7570\u306a\u308b NUMA \u30c9\u30e1\u30a4\u30f3\u304b\u3089\u306a\u308b\u74b0\u5883\u3092\u8003\u3048\u3066\u307f\u307e\u3057\u3087\u3046\u3002FPGA\u3001\u30ab\u30b9\u30bf\u30e0\u30fb\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\u30fc\u3001\u30c1\u30c3\u30d7\u30ec\u30c3\u30c8\u3001GPU \u306e\u4eba\u6c17\u304c\u9ad8\u307e\u308a\u3001\u7af6\u4e89\u304c\u6fc0\u5316\u3059\u308b\u306b\u3064\u308c\u3066\u3001\u30e1\u30e2\u30ea\u30fc\u30fb\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u306f\u3055\u3089\u306b\u8907\u96d1\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n \u6211\u3005\u306e\u30b3\u30f3\u30dd\u30fc\u30b6\u30d3\u30ea\u30c6\u30a3\u30fc\u30fb\u30a2\u30d7\u30ed\u30fc\u30c1\u306f\u3001\u6700\u5c0f\u9650\u306e\u5909\u66f4\u3067 SYCL \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3001OpenMP\u3001MPI \u306e\u5171\u5b58\u3092\u53ef\u80fd\u306b\u3057\u307e\u3059\u3002\u3082\u3061\u308d\u3093\u3001\u307b\u304b\u306e\u591a\u5c64\u578b\u306e\u30cf\u30a4\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u30fb\u30b9\u30bf\u30c3\u30af\u3068\u540c\u69d8\u306b\u3001\u8907\u96d1\u306b\u306a\u308b\u3068\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306e\u8981\u4ef6\u3082\u5897\u3048\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u56f3 1 \u306b\u793a\u3059\u3088\u3046\u306b\u3001\u305d\u306e\u6982\u5ff5\u306f\u975e\u5e38\u306b\u5358\u7d14\u3067\u3059\u3002<\/p>\n\n\n\n \u56f3 1:<\/strong> OpenMP \u30b3\u30f3\u30dd\u30fc\u30b6\u30d3\u30ea\u30c6\u30a3\u30fc\u306e\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0<\/p>\n\n\n\n OpenMP \u30b3\u30f3\u30dd\u30fc\u30b6\u30d3\u30ea\u30c6\u30a3\u30fc\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u8a18\u4e8b\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002 \u4eee\u60f3\u95a2\u6570\u306f\u3001\u8a00\u8a9e\u306e\u8a95\u751f\u4ee5\u6765\u3001\u73fe\u4ee3\u306e C++ \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u306b\u304a\u3051\u308b\u4e0d\u53ef\u6b20\u306a\u8981\u7d20\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002\u30cf\u30a4\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u79d1\u5b66\u8a08\u7b97\u3067\u5e83\u304f\u4f7f\u7528\u3055\u308c\u3066\u3044\u308b Kokkos C++ \u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u79fb\u690d\u6027\u30a4\u30f3\u30bf\u30fc\u30d5\u30a7\u30a4\u30b9<\/a> (\u82f1\u8a9e) \u306a\u3069\u306e\u4e00\u822c\u7684\u306a\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u62bd\u8c61\u5316\u306f\u3001\u4eee\u60f3\u95a2\u6570\u306b\u4f9d\u5b58\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n \u305f\u3060\u3057\u3001SYCL \u4ed5\u69d8\u306e 1 \u5b9a\u7fa9\u30eb\u30fc\u30eb (ODR) \u306b\u3088\u308b\u4eee\u60f3\u7d99\u627f\u306e\u4f7f\u7528\u306b\u306f\u82e5\u5e72\u306e\u5236\u9650\u304c\u3042\u308a\u307e\u3059\u3002\u4eee\u60f3\u30e1\u30f3\u30d0\u30fc\u95a2\u6570\u306f\u30c7\u30d0\u30a4\u30b9\u95a2\u6570\u5185\u3067\u547c\u3073\u51fa\u305b\u306a\u3044\u305f\u3081\u3001\u4e00\u90e8\u306e\u30b3\u30fc\u30c9\u30d9\u30fc\u30b9\u3092 SYCL \u306b\u79fb\u690d\u3059\u308b\u969b\u306b\u8907\u96d1\u306a\u554f\u984c\u304c\u767a\u751f\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u3053\u306e\u554f\u984c\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u3001\u6211\u3005\u306f oneAPI \u30ab\u30fc\u30cd\u30eb\u30fb\u30d7\u30ed\u30d1\u30c6\u30a3\u30fc\u62e1\u5f35\u6a5f\u80fd (indirectly_callable \u30d7\u30ed\u30d1\u30c6\u30a3\u30fc) \u3092\u5c0e\u5165\u3057\u3066\u3001\u4eee\u60f3\u95a2\u6570\u3092\u3059\u3079\u3066\u306e\u5236\u9650\u3092\u542b\u3080\u7279\u5225\u306a\u30c7\u30d0\u30a4\u30b9\u95a2\u6570\u3068\u3057\u3066\u5ba3\u8a00\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n \u56f3 2<\/strong>: oneAPI \u30ab\u30fc\u30cd\u30eb\u30fb\u30d7\u30ed\u30d1\u30c6\u30a3\u30fc\u62e1\u5f35\u6a5f\u80fd\u3092\u5229\u7528\u3057\u305f\u57fa\u672c\u7684\u306a\u4eee\u60f3\u30af\u30e9\u30b9\u7d99\u627f\u306e\u30bb\u30c3\u30c8\u30a2\u30c3\u30d7<\/p>\n\n\n\n SYCL \u3067\u306e\u4eee\u60f3\u95a2\u6570\u306e\u4f7f\u7528\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u8a18\u4e8b\u304a\u3088\u3073\u30d7\u30ec\u30bc\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306e\u9332\u753b<\/a> (\u82f1\u8a9e) \u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002 \u30a4\u30f3\u30c6\u30eb\u304c\u8868\u660e\u3057\u305f\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u30fb\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u30fc\u306b\u5bfe\u3059\u308b\u30b5\u30dd\u30fc\u30c8\u306b\u6cbf\u3063\u3066\u3001\u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u958b\u767a\u8005\u306f\u3001OpenMP\u3001SYCL\u3001MPI \u306a\u3069\u306e\u4e26\u5217\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u624b\u6cd5\u3001SPIR-V \u306a\u3069\u306e\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u5b9a\u7fa9\u3001LLVM \u306a\u3069\u306e\u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u30fb\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3001TensorFlow \u3084 PyTorch \u306a\u3069\u306e AI \u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306a\u3069\u3001\u3053\u308c\u3089\u306e\u8b70\u8ad6\u3084\u305d\u306e\u4ed6\u306e\u6700\u5148\u7aef\u306e\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306b\u95a2\u3059\u308b\u8b70\u8ad6\u306b\u7a4d\u6975\u7684\u306b\u8ca2\u732e\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n \u753b\u50cf\u51e6\u7406\u3068\u30d3\u30b8\u30e5\u30a2\u30eb\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u306f\u3001\u591a\u304f\u306e\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u3001\u30b2\u30fc\u30e0\u3001CGI VFX \u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306e\u91cd\u8981\u306a\u8981\u7d20\u3067\u3059\u3002<\/p>\n\n\n\n \u305d\u306e\u305f\u3081\u3001\u30a4\u30f3\u30c6\u30eb\u00ae oneAPI DPC++\/C++ \u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u306e Vulkan \u304a\u3088\u3073 DirectX \u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9 API \u3068\u306e SYCL \u76f8\u4e92\u904b\u7528\u6027\u304c\u5f37\u5316\u3055\u308c\u3001\u30af\u30ed\u30b9\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u306e\u4e92\u63db\u6027\u3001\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3001\u52b9\u7387\u6027\u304c\u5411\u4e0a\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u3089\u306e\u5f37\u5316\u304c\u30a4\u30f3\u30c6\u30eb\u00ae DPC++ \u4e92\u63db\u6027\u30c4\u30fc\u30eb\u3067\u5b8c\u5168\u306b\u8a8d\u8b58\u3055\u308c\u308b\u3088\u3046\u306b\u306a\u3063\u305f\u3053\u3068\u3067\u3001CUDA GPU \u30aa\u30d5\u30ed\u30fc\u30c9\u30b3\u30fc\u30c9\u3092\u30aa\u30fc\u30d7\u30f3\u6a19\u6e96\u306e\u30af\u30ed\u30b9\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u304a\u3088\u3073\u30de\u30eb\u30c1\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc SYCL \u30b3\u30fc\u30c9\u306b\u79fb\u884c\u3059\u308b\u30d7\u30ed\u30bb\u30b9\u306b\u3082\u53cd\u6620\u3055\u308c\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002\u30a4\u30f3\u30c6\u30eb\u00ae DPC++ \u4e92\u63db\u6027\u30c4\u30fc\u30eb\u3067\u3001\u4e00\u822c\u7684\u306a AI \u304a\u3088\u3073\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u30fb\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3067\u4f7f\u7528\u3055\u308c\u308b 357 \u306e API \u3092\u81ea\u52d5\u7684\u306b\u79fb\u884c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n\n\n\n \u30d0\u30a4\u30f3\u30c9\u30ec\u30b9\u30fb\u30a4\u30e1\u30fc\u30b8\u306f\u3001SYCL 2020 \u6a19\u6e96\u306e\u753b\u50cf\u30b5\u30dd\u30fc\u30c8\u304c\u62e1\u5f35\u3055\u308c\u3001\u30c7\u30d0\u30a4\u30b9\u3078\u306e\u753b\u50cf\u306e\u4fdd\u5b58\u65b9\u6cd5\u3068\u3001\u4f7f\u7528\u3059\u308b\u30e1\u30e2\u30ea\u30fc\u30fb\u30a2\u30af\u30bb\u30b9\u30fb\u30e2\u30c7\u30eb (\u7d71\u5408\u5171\u6709\u30e1\u30e2\u30ea\u30fc (USM)\u3001\u30d0\u30c3\u30d5\u30a1\u30fc\u30a2\u30af\u30bb\u30b5\u30fc\u3001\u30c7\u30d0\u30a4\u30b9\u6700\u9069\u5316\u30e1\u30e2\u30ea\u30fc\u30ec\u30a4\u30a2\u30a6\u30c8\u3001\u30a4\u30f3\u30dd\u30fc\u30c8\u30e1\u30e2\u30ea\u30fc\u306a\u3069) \u3092\u5236\u5fa1\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002\u30a2\u30af\u30bb\u30b5\u30fc\u30d9\u30fc\u30b9\u306e\u30e1\u30e2\u30ea\u30fc\u30fb\u30a2\u30af\u30bb\u30b9\u30fb\u30ea\u30af\u30a8\u30b9\u30c8\u3092\u8d85\u3048\u3066\u3001\u30e9\u30f3\u30bf\u30a4\u30e0\u30fb\u30a4\u30e1\u30fc\u30b8\u30fb\u30ec\u30f3\u30c0\u30ea\u30f3\u30b0\u3001\u30df\u30c3\u30d7\u30de\u30c3\u30d7\u3001\u30ad\u30e5\u30fc\u30d6\u30de\u30c3\u30d7\u3001\u30a4\u30e1\u30fc\u30b8\u914d\u5217\u306a\u3069\u3092\u30b5\u30dd\u30fc\u30c8\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u67d4\u8edf\u306a\u88dc\u52a9\u30b3\u30d4\u30fc\u6a5f\u80fd\u306b\u3088\u308a\u753b\u50cf\u30c7\u30fc\u30bf\u3092\u30b3\u30d4\u30fc\u3057\u3066\u518d\u89e3\u91c8\u3057\u3001\u5168\u4f53\u50cf\u3092\u5b8c\u6210\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n \u3053\u308c\u306b\u3088\u308a\u3001CPU \u30db\u30b9\u30c8\u30d7\u30ed\u30bb\u30b9\u3068\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\u30fc\u30fb\u30c7\u30d0\u30a4\u30b9\u9593\u3067\u8ee2\u9001\u3055\u308c\u308b\u30c7\u30fc\u30bf\u306e\u91cf\u3092\u6700\u5c0f\u9650\u306b\u6291\u3048\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n \u5b9f\u884c\u30d5\u30ed\u30fc\u306f\u56f3 3 \u304b\u3089\u3001\u3088\u308a\u30b7\u30f3\u30d7\u30eb\u3067\u30ea\u30bd\u30fc\u30b9\u52b9\u7387\u304c\u5927\u5e45\u306b\u5411\u4e0a\u3057\u305f\u56f3 4 \u306b\u4ee3\u308f\u308a\u307e\u3059\u3002<\/p>\n\n\n\n \u56f3 3<\/strong>: \u76f8\u4e92\u904b\u7528\u6027\u306e\u306a\u3044\u30db\u30b9\u30c8\u3068\u30c7\u30d0\u30a4\u30b9\u9593\u306e\u30c7\u30fc\u30bf\u8ee2\u9001<\/p>\n\n\n\n \u56f3 4<\/strong>: \u76f8\u4e92\u904b\u7528\u6027\u306e\u3042\u308b\u30db\u30b9\u30c8\u3068\u30c7\u30d0\u30a4\u30b9\u9593\u306e\u30c7\u30fc\u30bf\u8ee2\u9001<\/p>\n\n\n\n SYCL \u753b\u50cf\u51e6\u7406\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306b\u5916\u90e8\u30e1\u30e2\u30ea\u30fc\u3092\u30a4\u30f3\u30dd\u30fc\u30c8\u3059\u308b\u305f\u3081\u306e API \u3092\u5c0e\u5165\u3059\u308b\u3053\u3068\u306b\u3088\u308a\u3001SYCL \u3068 Vulkan \u307e\u305f\u306f DirectX 12 \u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u5185\u3067\u540c\u3058\u30e1\u30e2\u30ea\u30fc\u5272\u308a\u5f53\u3066\u3092\u518d\u5229\u7528\u3057\u3066\u3001\u4f59\u5206\u306a\u30b3\u30d4\u30fc\u30b9\u30c6\u30c3\u30d7\u3092 2 \u3064\u6e1b\u3089\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n SYCL \u3068 DirectX \u304a\u3088\u3073 Vukan \u306e\u76f8\u4e92\u904b\u7528\u6027\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u8a18\u4e8b\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002 \u30b9\u30b1\u30fc\u30e9\u30d6\u30eb\u306a\u30de\u30eb\u30c1\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u4e26\u5217\u51e6\u7406\u3068\u306f\u3001\u4e00\u822c\u7684\u306a\u8a00\u8a9e\u3084\u8a00\u8a9e\u62e1\u5f35\u3067\u9ad8\u5ea6\u306a\u4e26\u5217\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u306e\u6982\u5ff5\u3092\u30b5\u30dd\u30fc\u30c8\u3059\u308b\u3053\u3068\u3060\u3051\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u8a00\u8a9e\u62e1\u5f35\u3068\u57fa\u76e4\u3068\u306a\u308b\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u6a5f\u80fd\u306e\u4e21\u65b9\u304b\u3089\u76f4\u63a5\u30e1\u30ea\u30c3\u30c8\u3092\u53d7\u3051\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3001\u6700\u9069\u5316\u3055\u308c\u305f\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u30fb\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u306e\u8c4a\u5bcc\u306a\u30a8\u30b3\u30b7\u30b9\u30c6\u30e0\u3082\u91cd\u8981\u3067\u3059\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae \u30a4\u30f3\u30c6\u30b0\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u30fb\u30d7\u30ea\u30df\u30c6\u30a3\u30d6 (\u30a4\u30f3\u30c6\u30eb\u00ae IPP)<\/a> \u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae IPP \u306e\u30ef\u30fc\u30d7\u30fb\u30d1\u30fc\u30b9\u30da\u30af\u30c6\u30a3\u30d6\u95a2\u6570\u3068\u6700\u8fd1\u508d\u88dc\u9593\u307e\u305f\u306f OpenCV \u3092\u4f7f\u7528\u3057\u3066\u300132 \u30d3\u30c3\u30c8\u6d6e\u52d5\u5c0f\u6570\u70b9\u753b\u50cf\u306e\u64cd\u4f5c\u3068\u30a2\u30e9\u30a4\u30e1\u30f3\u30c8\u3092\u9ad8\u901f\u5316\u3059\u308b\u3053\u3068\u306b\u3088\u308a\u3001\u30a4\u30f3\u30c6\u30eb\u00ae AVX-512 \u306b\u5bfe\u5fdc\u3057\u305f\u30a4\u30f3\u30c6\u30eb\u00ae CPU \u3067\u306e\u9ad8\u7cbe\u5ea6\u753b\u50cf\u51e6\u7406\u3092\u9ad8\u901f\u5316\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u753b\u50cf\u51e6\u7406\u306b\u52a0\u3048\u3066\u3001\u4fe1\u53f7\u51e6\u7406\u3067\u3082\u3001\u30a4\u30f3\u30c6\u30eb\u00ae AVX2 \u304a\u3088\u3073\u30a4\u30f3\u30c6\u30eb\u00ae AVX-512 \u306b\u5bfe\u5fdc\u3057\u305f CPU \u3067\u306e 16 \u30d3\u30c3\u30c8\u30fb\u30c7\u30fc\u30bf\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304c\u5927\u5e45\u306b\u5411\u4e0a\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae oneAPI \u30de\u30b9\u30fb\u30ab\u30fc\u30cd\u30eb\u30fb\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc (\u30a4\u30f3\u30c6\u30eb\u00ae oneMKL)<\/a> \u306f\u3001\u9ad8\u901f\u30d5\u30fc\u30ea\u30a8\u5909\u63db (FFT) \u306e\u30af\u30ed\u30b9\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u4e26\u5217\u5b9f\u884c\u3092\u65b0\u305f\u306a\u30ec\u30d9\u30eb\u306b\u5f15\u304d\u4e0a\u3052\u3001\u65b0\u3057\u3044 SYCL DFT API \u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u8907\u6570\u306e GPU \u3092\u7d44\u307f\u5408\u308f\u305b\u305f 2D \u304a\u3088\u3073 3D \u975e\u30d0\u30c3\u30c1 FFT \u51e6\u7406\u3092\u5bb9\u6613\u306b\u5b9f\u73fe\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u5206\u6563\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u30fb\u30af\u30e9\u30b9\u30bf\u30fc\u3067 CPU \u306b\u8907\u6570\u306e GPU \u3092\u63a5\u7d9a\u3059\u308b\u306e\u306f\u3001HPC \u306b\u304a\u3051\u308b\u4e00\u822c\u7684\u306a\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u69cb\u6210\u3067\u3059 (\u524d\u8ff0\u306e SYCL \u304a\u3088\u3073 MPI \u3068\u306e OpenMP \u30b3\u30f3\u30dd\u30fc\u30b6\u30d3\u30ea\u30c6\u30a3\u30fc\u3092\u53c2\u7167)\u3002<\/p>\n\n\n\n OpenMP 6.0 \u306e\u65b0\u3057\u3044 stripe \u30eb\u30fc\u30d7\u5909\u63db\u69cb\u9020\u3068\u3001nowait \u7bc0\u306e\u30aa\u30d7\u30b7\u30e7\u30f3\u306e\u30d6\u30fc\u30eb\u5f15\u6570\u306b\u3088\u308a\u3001\u30a4\u30f3\u30c6\u30eb\u00ae \u30b3\u30f3\u30d1\u30a4\u30e9\u30fc\u3067\u975e\u540c\u671f\u307e\u305f\u306f\u540c\u671f\u30aa\u30d5\u30ed\u30fc\u30c9\u3092\u6761\u4ef6\u4ed8\u304d\u3067\u9078\u629e\u53ef\u80fd\u306b\u306a\u308a\u3001\u30aa\u30d5\u30ed\u30fc\u30c9\u51e6\u7406\u3092\u67d4\u8edf\u306b\u5236\u5fa1\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc<\/a>\u3067\u4ee5\u4e0b\u306e\u6a5f\u80fd\u3092\u5229\u7528\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n\n\n\n \u3053\u308c\u3089\u306b\u3088\u308a\u3001CPU \u306e\u307f\u3092\u4f7f\u7528\u3059\u308b\u30e9\u30c3\u30d7\u30c8\u30c3\u30d7\u304b\u3089\u3001HPC \u304a\u3088\u3073 AI \u5411\u3051\u306e\u30a8\u30af\u30b5\u30b9\u30b1\u30fc\u30eb\u30fb\u30b9\u30fc\u30d1\u30fc\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fc\u306b\u81f3\u308b\u307e\u3067\u3001\u30b9\u30b1\u30fc\u30e9\u30d6\u30eb\u306a\u4e26\u5217\u51e6\u7406\u304c\u5b9f\u73fe\u3057\u307e\u3059\u3002<\/p>\n\n\n\n AI \u306b\u95a2\u3057\u3066\u306f\u3001PyTorch \u304a\u3088\u3073\u30a4\u30f3\u30c6\u30eb\u00ae oneAPI \u30c7\u30a3\u30fc\u30d7\u30fb\u30cb\u30e5\u30fc\u30e9\u30eb\u30fb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u30fb\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc (\u30a4\u30f3\u30c6\u30eb\u00ae oneDNN)<\/a> (\u82f1\u8a9e) \u306b\u5bfe\u3059\u308b\u30a4\u30f3\u30c6\u30eb\u306e\u8ca2\u732e\u3068\u3057\u3066\u3055\u3089\u306a\u308b\u6700\u9069\u5316\u3092\u8ffd\u52a0\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae Core™ Ultra \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc (\u30b7\u30ea\u30fc\u30ba 2)\u3001P-cores \u642d\u8f09\u30a4\u30f3\u30c6\u30eb\u00ae Xeon\u00ae 6 \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc\u3001\u30a4\u30f3\u30c6\u30eb\u00ae Arc™ GPU \u5411\u3051\u306e\u30a4\u30f3\u30c6\u30eb\u00ae oneDNN \u306e\u6700\u65b0\u306e\u6700\u9069\u5316\u306f PyTorch 2.7 \u30ea\u30ea\u30fc\u30b9\u306b\u30a2\u30c3\u30d7\u30b9\u30c8\u30ea\u30fc\u30e0\u3055\u308c\u3001BERT\u3001Llama\u3001GPT \u306a\u3069\u306e\u4e00\u822c\u7684\u306a AI \u63a8\u8ad6\u30ef\u30fc\u30af\u30d5\u30ed\u30fc\u3082\u30a4\u30f3\u30c6\u30eb\u00ae Xeon\u00ae 6 \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc\u3067\u30b9\u30fc\u30d1\u30fc\u30c1\u30e3\u30fc\u30b8\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n \u30a4\u30f3\u30c6\u30eb\u00ae VTune™ \u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc\u306e\u65b0\u3057\u3044\u5206\u6790\u30c4\u30fc\u30eb\u306e\u6a5f\u80fd\u3092\u6d3b\u7528\u3057\u3066\u30a4\u30f3\u30c6\u30eb\u306e\u30af\u30e9\u30a4\u30a2\u30f3\u30c8 GPU \u3068 NPU \u306e AI \u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u52b9\u7387\u5316\u304a\u3088\u3073\u6700\u9069\u5316\u3057\u3066\u3001DirectML \u30d9\u30fc\u30b9\u306e AI \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u5411\u4e0a\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n PyTorch \u30d9\u30fc\u30b9\u306e AI \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306e\u9ad8\u901f\u5316\u306b\u5bfe\u3059\u308b\u30a4\u30f3\u30c6\u30eb\u306e\u652f\u63f4\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u30d6\u30ed\u30b0\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002 AI PC \u3067\u5b9f\u884c\u3059\u308b\u5834\u5408\u3067\u3082\u3001\u30b9\u30fc\u30d1\u30fc\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fc\u3067\u5b9f\u884c\u3059\u308b\u5834\u5408\u3067\u3082\u3001\u5305\u62ec\u7684\u3067\u9ad8\u5ea6\u306b\u7d71\u5408\u3055\u308c\u305f\u4e26\u5217\u51e6\u7406\u306b\u3088\u308a\u3001\u3042\u3089\u3086\u308b\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306e\u5b9f\u884c\u3092\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3001\u6700\u9069\u5316\u3001\u52b9\u7387\u5316\u3067\u304d\u307e\u3059\u3002MPI\u3001OpenMP\u3001SYCL\u3001PyTorch\u3001Vulkan\u3001DirectX \u306a\u3069\u306e\u30b7\u30fc\u30e0\u30ec\u30b9\u306a\u76f8\u4e92\u904b\u7528\u6027\u306b\u3088\u308a\u3001AI\u3001HPC\u3001\u753b\u50cf\u51e6\u7406\u306e\u305f\u3081\u306e\u958b\u767a\u30bd\u30ea\u30e5\u30fc\u30b7\u30e7\u30f3\u3092\u63d0\u4f9b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n \u6700\u65b0\u306e\u958b\u767a\u30c4\u30fc\u30eb\u3001\u30d0\u30fc\u30b8\u30e7\u30f3 2025.2 \u3092\u4eca\u3059\u3050\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n \u30cb\u30e5\u30fc\u30b9\u30fb\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8<\/p>\n\n\n\n \u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0<\/p>\n\n\n\n AI \u3068\u30de\u30b7\u30f3\u30e9\u30fc\u30cb\u30f3\u30b0<\/p>\n\n\n\n \u6027\u80fd\u306f\u3001\u4f7f\u7528\u72b6\u6cc1\u3001\u69cb\u6210\u3001\u305d\u306e\u4ed6\u306e\u8981\u56e0\u306b\u3088\u3063\u3066\u7570\u306a\u308a\u307e\u3059\u3002\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001http:\/\/www.intel.com\/PerformanceIndex\/<\/a> (\u82f1\u8a9e) \u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n \u30a8\u30af\u30bb\u30eb\u30bd\u30d5\u30c8\u304c\u63d0\u4f9b\u3059\u308b\u30a4\u30f3\u30c6\u30eb\u00ae \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30c4\u30fc\u30eb\u5411\u3051\u30b5\u30dd\u30fc\u30c8\u30b5\u30fc\u30d3\u30b9\u3067\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30c4\u30fc\u30eb\u306e\u65e7\u30d0\u30fc\u30b8\u30e7\u30f3\u304b\u3089\u65b0\u30d0\u30fc\u30b8\u30e7\u30f3\u3078\u306e\u79fb\u884c\u3001CUDA \u304b\u3089 SYCL \u3078\u306e\u30b3\u30fc\u30c9\u79fb\u884c\u3001\u4ed6\u793e\u88fd GPU \u3068\u306e\u30b3\u30fc\u30c9\u4e92\u63db\u306a\u3069\u3001\u65b0\u3057\u3044\u74b0\u5883\u3067\u3053\u308c\u307e\u3067\u901a\u308a\u696d\u52d9\u3092\u9042\u884c\u3059\u308b\u305f\u3081\u306e\u79fb\u884c\u3092\u652f\u63f4\u3057\u307e\u3059\u3002\u88fd\u54c1\u306e\u79fb\u884c\u306b\u95a2\u3057\u3066\u304a\u60a9\u307f\u3001\u8cea\u554f\u306a\u3069\u304a\u6c17\u8efd\u306b\u304a\u554f\u3044\u5408\u308f\u305b\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n\u65b0\u30ea\u30ea\u30fc\u30b9\u306e\u6982\u8981<\/h2>\n\n\n\n
\n
\n
\u21d2 \u30a4\u30f3\u30c6\u30eb\u00ae \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30c4\u30fc\u30eb\u306e\u30d0\u30fc\u30b8\u30e7\u30f3 2025.2 \u30ea\u30ea\u30fc\u30b9: AI\u3001\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u3001\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30c6\u30c3\u30c9\u30fb\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u5411\u3051\u306b\u6700\u9069\u5316\u3055\u308c\u305f\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3068\u751f\u7523\u6027<\/a><\/p>\n\n\n\n\u30b9\u30b1\u30fc\u30e9\u30d6\u30eb\u306a\u4e26\u5217\u51e6\u7406\u306e\u76f8\u4e92\u904b\u7528\u6027<\/h2>\n\n\n\n
SYCL \u30b8\u30e7\u30a4\u30f3\u30c8\u884c\u5217\u62e1\u5f35 API<\/h4>\n\n\n\n
using namespace sycl::ext::intel::experimental::matrix;\ntemplate <typename Group, typename T, use Use, size_t Rows,\n size_t Cols, layout Layout = layout::dynamic>\nstruct joint_matrix;\nenum class use { a, b, accumulator};\nenum class layout {row_major, col_major, dynamic};<\/code><\/pre>\n\n\n\nvoid joint_matrix_mad(Group g, joint_matrix<>&D,\n joint_matrix<>&A, joint_matrix<>&B, joint_matrix<>&C);\nvoid joint_matrix_apply(Group g, joint_matrix<>&jm, F&& func);\nvoid joint_matrix_apply(Group g,\n joint_matrix<>& jm0, joint_matrix<>& jm1, F&& func);\nvoid joint_matrix_copy(Group g, joint_matrix<Group, T1, Use1,\n Rows, Cols, Layout1> &dest,\njoint_matrix<Group, T2, Use2, Rows, Cols, Layout2> &src);<\/code><\/pre>\n\n\n\nvoid joint_matrix_fill(Group g, joint_matrix<>&dst, T v);\nvoid joint_matrix_load(Group g, joint_matrix<>&dst,\n multi_ptr<> src, size_t stride, Layout layout);\nvoid joint_matrix_load(Group g, joint_matrix<>&dst,\n multi_ptr<> src, size_t stride);\nvoid joint_matrix_store(Group g, joint_matrix<>src,\n multi_ptr<> dst, unsigned stride, Layout layout);\nvoid joint_matrix_prefetch(Group g, T* ptr, size_t\n stride, layout Layout, Properties properties);<\/code><\/pre>\n\n\n\n
\u21d2\u00a0\u7d71\u5408\u30b8\u30e7\u30a4\u30f3\u30c8\u884c\u5217\u306e SYCL \u62e1\u5f35<\/a><\/p>\n\n\n\nOpenMP \u30b3\u30f3\u30dd\u30fc\u30b6\u30d3\u30ea\u30c6\u30a3\u30fc\u306e\u5f37\u5316<\/h4>\n\n\n\n
\n
![\"\"](\"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-content\/uploads\/sites\/5\/2025\/07\/23_openmp-3-1024x532.png\")
<\/figure>\n\n\n\n
\u21d2 OpenMP \u3068\u79d1\u5b66\u8a08\u7b97\u306e\u5c06\u6765<\/a><\/p>\n\n\n\nSYCL \u306e\u4eee\u60f3\u95a2\u6570<\/h4>\n\n\n\n
![\"\"](\"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-content\/uploads\/sites\/5\/2025\/07\/23_basic-reference-sample-setup-1024x479.png\")
<\/figure>\n\n\n\n
\u21d2 SYCL \u62e1\u5f35\u306e\u91cd\u8981\u306a\u90e8\u5206: C++ \u4eee\u60f3\u95a2\u6570\u3068 SPIR-V \u30d0\u30c3\u30af\u30a8\u30f3\u30c9<\/a> (\u82f1\u8a9e)<\/p>\n\n\n\n\u9ad8\u901f\u753b\u50cf\u51e6\u7406\u3068\u8996\u899a\u5316<\/h2>\n\n\n\n
![\"\"](\"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-content\/uploads\/sites\/5\/2025\/07\/23_data-transfer-without-sycl-interop-1024x409.png\"){kind=spacer}
<\/figure>\n\n\n\n![\"\"](\"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-content\/uploads\/sites\/5\/2025\/07\/23_data-transfer-with-sycl-interop-1024x424.png\"){kind=spacer}
<\/figure>\n\n\n\n
\u21d2 \u30d0\u30a4\u30f3\u30c9\u30ec\u30b9\u30fb\u30a4\u30e1\u30fc\u30b8\u30fb\u30e1\u30e2\u30ea\u30fc\u306e SYCL \u3068 DirectX \u304a\u3088\u3073 Vulkan \u306e\u76f8\u4e92\u904b\u7528\u6027<\/a> (\u82f1\u8a9e)<\/p>\n\n\n\n\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u4e26\u5217\u51e6\u7406\u306b\u3088\u308b\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9<\/h2>\n\n\n\n
\n
CPU \u3068 GPU \u306b\u304a\u3051\u308b AI \u306e\u30b9\u30fc\u30d1\u30fc\u30c1\u30e3\u30fc\u30b8<\/h2>\n\n\n\n
\u21d2 \u30a4\u30f3\u30c6\u30eb\u00ae GPU \u4e0a\u3067 PyTorch 2.7 \u3092\u9ad8\u901f\u5316<\/a> (\u82f1\u8a9e)
\u6b21\u306e\u30d3\u30c7\u30aa\u3082\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002
\u21d2 AOTInductor \u3092\u4f7f\u7528\u3057\u3066\u30a4\u30f3\u30c6\u30eb\u00ae GPU \u306b\u30b3\u30f3\u30d1\u30a4\u30eb\u6e08\u307f PyTorch \u30e2\u30c7\u30eb\u3092\u30c7\u30d7\u30ed\u30a4<\/a> (\u82f1\u8a9e)
\u21d2 \u30a4\u30f3\u30c6\u30eb\u00ae GPU \u3092\u4f7f\u7528\u3057\u305f PyTorch \u30a8\u30af\u30b9\u30dd\u30fc\u30c8\u306e\u91cf\u5b50\u5316<\/a> (\u82f1\u8a9e)<\/p>\n\n\n\n\u4e26\u5217\u30b3\u30fc\u30c9\u3092\u5236\u5fa1\u3059\u308b<\/h2>\n\n\n\n
\n
\u95a2\u9023\u60c5\u5831<\/h2>\n\n\n\n
\n
\n
\n
\u88fd\u54c1\u304a\u3088\u3073\u6027\u80fd\u306b\u95a2\u3059\u308b\u60c5\u5831<\/h4>\n\n\n\n
\u30a4\u30f3\u30c6\u30eb\u00ae \u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u30c4\u30fc\u30eb\u5411\u3051\u30b5\u30dd\u30fc\u30c8\u30b5\u30fc\u30d3\u30b9<\/h3>\n\n\n\n