{"id":2457,"date":"2026-08-06T22:46:45","date_gmt":"2026-08-06T14:46:45","guid":{"rendered":"https:\/\/ferina.top\/?p=2457"},"modified":"2026-08-06T22:46:47","modified_gmt":"2026-08-06T14:46:47","slug":"nsight-graphics","status":"publish","type":"post","link":"https:\/\/ferina.top\/index.php\/2026\/08\/06\/nsight-graphics\/","title":{"rendered":"Nsight Graphics"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u4e4b\u524d\u5206\u6790GPU\u7cfb\u7edf\u7684\u65f6\u5019\uff0c\u7528\u7684\u591a\u662fNsight System\u3002\u4f46\u76ee\u524d\u53d1\u73b0Nsight Graphics\u53ef\u80fd\u66f4\u9002\u5408\u5206\u6790\u4e00\u6bb5\u65f6\u95f4\u5185\u7684\u6e32\u67d3\u7ec6\u8282\uff0c\u4ee5\u53ca\u5bf9\u5e94\u7684GPU\u884c\u4e3a\u3002\u82e6\u4e8e\u770b\u4e0d\u61c2\u754c\u9762\uff0c\u4e8e\u662f\u51b3\u5b9a\u628a\u4e00\u4e9b\u5b66\u4e60\u7684\u8bb0\u5f55\u6574\u7406\u4e0b\u6765\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">1. Self Instruction Mix <\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u57fa\u672c\u6267\u884c\u5355\u5143<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/image.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"937\" height=\"834\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/image.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-2458\"  sizes=\"auto, (max-width: 937px) 100vw, 937px\" \/><\/div><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9996\u5148\uff0c\u4ece\u8ba4\u8bc6\u57fa\u672c\u7684\u6267\u884c\u5355\u5143\u5f00\u59cb\uff0c\u4ee5\u4e0b\u662f\u4e00\u89c8\u8868\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7f29\u5199<\/th><th>\u542b\u4e49<\/th><th>\u5b9e\u9645\u4f5c\u7528<\/th><\/tr><\/thead><tbody><tr><td>FMA<\/td><td>Fused Multiply-Add\/Accumulate<\/td><td>\u4e3b\u8981\u6267\u884c FP32 \u52a0\u3001\u4e58\u3001\u878d\u5408\u4e58\u52a0\uff0c\u4e5f\u6267\u884c\u90e8\u5206\u6574\u6570\u4e58\u6cd5\/\u4e58\u52a0\u3002<\/td><\/tr><tr><td>ALU<\/td><td>Arithmetic Logic Unit<\/td><td>\u666e\u901a\u6574\u6570\u8fd0\u7b97\u3001\u6bd4\u8f83\u3001\u903b\u8f91\u3001\u4f4d\u64cd\u4f5c\u3001\u6570\u636e\u79fb\u52a8\uff0c\u4ee5\u53ca\u90e8\u5206\u7b80\u5355 FP32 \u64cd\u4f5c\u3002<\/td><\/tr><tr><td>LDCU<\/td><td>\u5e38\u91cf\u52a0\u8f7d\u5230 Uniform Register \u7684\u6307\u4ee4\/\u901a\u8def\u5206\u7c7b<\/td><td>\u4ece Constant Memory \u5411 warp-uniform \u5bc4\u5b58\u5668\u52a0\u8f7d\u6807\u91cf\u503c\u3002\u5b98\u65b9\u5bf9\u5b83\u7684\u5b9a\u4e49\u662f\u201cLoad a Value from Constant Memory into a Uniform Register\u201d\u3002<a href=\"https:\/\/docs.nvidia.com\/cuda\/archive\/13.2.0\/cuda-binary-utilities\/index.html\">CUDA \u6307\u4ee4\u8bf4\u660e<\/a><\/td><\/tr><tr><td>UDP<\/td><td>Uniform Data Path<\/td><td>\u7edf\u4e00\/\u6807\u91cf\u6570\u636e\u901a\u8def\u3002\u5904\u7406\u6574\u4e2a warp \u90fd\u76f8\u540c\u7684\u503c\uff0c\u907f\u514d\u6bcf\u4e2a lane \u91cd\u590d\u505a\u540c\u4e00\u8ba1\u7b97\u3002\u65b0\u7248\u6307\u6807\u4e2d\u4e5f\u5e38\u79f0 <code>UNIFORM<\/code>\u3002<\/td><\/tr><tr><td>ADU<\/td><td>Address Divergence Unit<\/td><td>\u5730\u5740\u5206\u6b67\u5904\u7406\u3001\u90e8\u5206\u8df3\u8f6c\/\u5206\u652f\u3001\u5e38\u91cf\u52a0\u8f7d\u3001CTA\/block barrier \u7b49\u63a7\u5236\u548c\u5bfb\u5740\u5de5\u4f5c\u3002<\/td><\/tr><tr><td>SCH<\/td><td>Scheduler \/ \u8c03\u5ea6\u63a7\u5236\u8def\u5f84<\/td><td>\u8c03\u5ea6\u548c\u63a7\u5236\u7c7b\u5fae\u64cd\u4f5c\u3001NOP\uff0c\u4ee5\u53ca\u90e8\u5206\u5149\u8ffd\u63a7\u5236\u6307\u4ee4\u3002\u5b83\u4e0d\u7b49\u4e8e RT Core \u672c\u4f53\u3002<\/td><\/tr><tr><td>CBU<\/td><td>Convergence Barrier Unit<\/td><td>warp \u5206\u652f\u6c47\u5408\u3001\u63a7\u5236\u6d41\u3001barrier\u3001branch\/call\/return\/exit \u7b49\u3002<\/td><\/tr><tr><td>LSU<\/td><td>Load\/Store Unit<\/td><td>\u53d1\u51fa Global\u3001Local\u3001Shared Memory \u7684 load\/store\/atomic\uff0c\u4ee5\u53ca\u90e8\u5206\u7279\u6b8a\u5bc4\u5b58\u5668\u3001warp primitive \u548c barrier \u6307\u4ee4\u3002<\/td><\/tr><tr><td>XU<\/td><td>Transcendental and Data Type Conversion Unit<\/td><td>\u7279\u6b8a\u51fd\u6570\u548c\u7c7b\u578b\u8f6c\u6362\uff0c\u5982 <code>sin\/cos\/rcp\/rsqrt<\/code>\u3001float\u2194int\uff1b\u7ecf\u5e38\u4e5f\u88ab\u79f0\u4e3a SFU \u8def\u5f84\u3002<\/td><\/tr><tr><td>TEX<\/td><td>Texture Unit<\/td><td>\u7eb9\u7406\u5bfb\u5740\u3001LOD\u3001\u8fc7\u6ee4\u3001\u683c\u5f0f\u8f6c\u6362\u3001Texture\/Surface \u8bbf\u95ee\u3002<\/td><\/tr><tr><td>RTCORE<\/td><td>Ray Tracing Core<\/td><td>BVH \u904d\u5386\u3001\u5149\u7ebf\u4e0e\u5305\u56f4\u76d2\/\u4e09\u89d2\u5f62\u6c42\u4ea4\u7b49\u786c\u4ef6\u5149\u8ffd\u5de5\u4f5c\uff1b\u4e0d\u662f\u6574\u4e2a ray tracing shader \u7684\u5168\u90e8\u6267\u884c\u3002<\/td><\/tr><tr><td>IPA<\/td><td>Pixel Shader \u5c5e\u6027\u63d2\u503c\/\u8bbf\u95ee\u8def\u5f84<\/td><td>\u8bfb\u53d6\u6216\u63d2\u503c Pixel Shader \u8f93\u5165\u5c5e\u6027\u3002\u516c\u5f00\u6587\u6863\u4e00\u822c\u76f4\u63a5\u4f7f\u7528 IPA\uff0c\u6ca1\u6709\u7a33\u5b9a\u516c\u5f00\u7684\u5168\u79f0\u3002\u4e0d\u8981\u4e0e\u7f16\u8bd1\u5668\u4e2d\u7684 Interprocedural Analysis \u6df7\u6dc6\u3002<\/td><\/tr><tr><td>REDUX<\/td><td>Reduction \u8def\u5f84\/\u6307\u4ee4<\/td><td>\u5bf9 warp \u4e2d\u591a\u4e2a lane \u7684\u503c\u6267\u884c\u5f52\u7ea6\uff0c\u5e76\u4ea7\u751f uniform \u7ed3\u679c\u3002<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a5\u4e0b\u6765\u5c55\u5f00\u4e00\u4e9b\u7ec6\u8282\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">FMA &amp; ALU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1a\u4e3a\u4ec0\u4e48FMA\u8981\u628a\u4e58\u6cd5\u548c\u52a0\u6cd5\u7edf\u5408\u5728\u4e00\u8d77\uff1f\u4e0d\u80fd\u5148\u4e58\u518d\u52a0\u5417\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A\uff1a\u5148\u4e58\u518d\u52a0\uff0c\u4e2d\u95f4\u7ed3\u679c\u548c\u6700\u540e\u7ed3\u679c\u4e00\u5171\u8981rounded 2\u6b21\uff1bFMA \u53ea rounded \u4e00\u6b21\uff0c\u6570\u503c\u7cbe\u5ea6\u66f4\u9ad8<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u5728\u7b97FLOPS\uff08Floating point operations per second\uff09\u65f6\uff0cFMA\u7b97\u4e24\u6b21\u6d6e\u70b9\u8fd0\u7b97<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d6e\u70b9\u4e58\u6cd5\u52a0\u6cd5\u5b9e\u9645\u4e0a\u590d\u7528\u4e86FMA\u7684\u7ed3\u6784\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3a x * y + 0\u8fd9\u6837\u7684\u5f62\u5f0f\uff08\u5f53\u7136\u4e0d\u4e00\u5b9a\u771f\u662f\u8fd9\u6837\u7f16\u7801\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1aFMA \u548c ALU\u7684\u804c\u8d23\u8303\u56f4\uff1f\u4e3a\u4ec0\u4e48\u4e00\u5b9a\u8981\u5206\u79bb\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\uff1a<strong>\u4e58\u6cd5\u6bd4\u52a0\u6cd5\/\u903b\u8f91\/\u6bd4\u8f83\u8d35\u7684\u591a<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-1 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/FMA-1024x389.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"389\" data-id=\"2465\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/FMA-1024x389.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-2465\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n<\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/ALU-1-1024x360.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"360\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/ALU-1-1024x360.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-2467\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u8fd9\u4e24\u5f20\u5b9e\u9645Nsight\u622a\u4e0b\u6765\u7684\u56fe\u53ef\u4ee5\u5927\u81f4\u770b\u51fa\u4e24\u4e2apipe\u7684\u529f\u80fd\uff08\u4ee5\u53ca\u5728\u4e00\u5e27\u6e32\u67d3\u4e2d\u7684\u5927\u6982\u6307\u4ee4\u6570\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bb0\u5f55\u4e00\u4e9b\u4e2a\u4eba\u7684\u7591\u70b9\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1aLogic Operation \u548c FP32 Logic\u7684\u533a\u522b\uff1f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Logic Op\uff1a\u6309\u4f4d\u4e0e\uff0c\u6216\uff0c\u975e\uff0c\u5f02\u6216\uff0cbool\u8fd0\u7b97\u7b49\uff0c<strong>\u6bcf\u4e00\u4f4d<\/strong>\u72ec\u7acb\u53c2\u52a0\u8fd0\u7b97\uff0c\u4e0d\u5728\u610f\u8bed\u4e49<\/li>\n\n\n\n<li>FP32 Logic\uff1amin\/max\u3001\u9009\u62e9\uff08float r2 = condition ? a : b;\uff09\u3001saturate\uff08clamp\uff09\uff0c\u9700\u8981\u8003\u8651\u8bed\u4e49<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1aCompare \u548c Logic\u7684\u533a\u522b\uff1f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Compare\uff1a\u5c0f\u4e8e\u5927\u4e8e\u7b49\u4e8e\uff0c\u8f93\u51fabool\u503c\uff0c\u4ee3\u8868<strong>\u4e24\u4e2a\u6570\u7684\u76f8\u5bf9\u5173\u7cfb<\/strong>\uff0c\u56e0\u6b64\u6ca1\u6709\u7b2c\u4e09\u4e2a\u6570<\/li>\n\n\n\n<li>Logic Op\uff1a\u4e0e\u6216\u975e\uff0c\u4e0d\u5173\u5fc3\u8f93\u51fa\u4ec0\u4e48\uff0c\u53ef\u4ee5\u8f93\u51fabool\u503c\uff08\u5982&amp;\u4e0d\u8f93\u51fabool\u503c\uff0c\u800c&amp;&amp;\u8f93\u51fabool\u503c\uff09\uff0c\u5173\u5fc3\u7ed3\u679c\u6570<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1a\u4ec0\u4e48\u662fData Movement\uff1f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8d4b\u503c\uff08a = b\uff09\uff0c\u4e09\u5143\u9009\u62e9\uff08\uff09\uff0cswizzle\uff0c\u6253\u5305<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1aFMA int\u4e58\u52a0\u548c ALU int\u8ba1\u7b97\u7684\u8fb9\u754c\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A\uff1a\u6240\u6709\u9700\u8981\u4e58\u6cd5\u7684int \u8ba1\u7b97\u7528FMA\uff1b\u5176\u4ed6\u7684\u6240\u6709\u90fd\u7528ALU <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1aData Movement\u662f\u4ec0\u4e48\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A\uff1a\u7c7b\u4f3c\u6307\u4ee4\u96c6\u91cc\u7684MOV\uff0c\u8868\u793a\u5bc4\u5b58\u5668 to \u5bc4\u5b58\u5668\uff0c\u6216\u8005\u7acb\u5373\u6570 to \u5bc4\u5b58\u5668<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">LSU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">LSU\u4ee3\u8868\u7740\u663e\u5b58\u548c\u5bc4\u5b58\u5668\u7684\u4ea4\u4e92\uff0c\u4e3b\u8981\u5206Global Shared Local\u4e09\u7ea7<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e00\u4e2a\u8bef\u533a\uff1a\u6211\u4eec\u90fd\u77e5\u9053GPU\u6709\u4e09\u7ea7\u7f13\u5b58\uff0cL2\uff08\u4e00\u5757GPU\u7684\u6240\u6709SM\u5171\u4eab\uff09\uff0cL1\uff08\u4e00\u4e2aSM\u7684\u6240\u6709subcore\u5171\u4eab\uff09\uff0cL0\uff08\u4e00\u4e2asubcore\u4e0a\u7684\u6240\u6709FP 32\/INT 32\u548c\u4e00\u4e2aTensor core\u5171\u4eab\uff09\uff0c\u4f46\u4ed6\u4eec\u548cLSU\u8fd9\u91cc\u7684\u4e09\u7ea7\u4e0d\u662f\u4e00\u4e00\u5bf9\u5e94\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LSU\u8fd9\u91cc\u5bf9\u5e94\u7684\uff0c\u66f4\u50cf\u662fshader\u53ef\u89c1\u7684\u865a\u62df\u5730\u5740\u7a7a\u95f4\uff0c\u5176\u4e0e\u5b9e\u9645\u7269\u7406\u7a7a\u95f4\u7684\u5bf9\u5e94\u662f\u4e0d\u786e\u5b9a\u7684\u3002Cache\u4ec5\u4ec5\u662f\u52a0\u901f\u7ed3\u6784\uff0c\u6bd4\u5982\u5b9e\u9645\u4e0a\u662f\u4e0d\u80fd\u76f4\u63a5\u5206\u914dL2 Cache\u7684\uff0c\u53ea\u80fd\u5728Load VM\u7684\u65f6\u5019\u52a0\u8f7d\u5230L2 Cache<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Global Memory\u662f\u6700\u5e38\u89c1\u7684memory\uff0c\u57fa\u672c\u6240\u6709\u6ca1\u6709\u7279\u6b8a\u8bbe\u7f6e\u7684\u663e\u5b58\u64cd\u4f5c\uff0c\u90fd\u662f\u5bf9Global memory\u7684\u64cd\u4f5c\uff0c\u4e00\u822c\u6765\u8bf4\uff0c\u5bf9\u4e8eGlobal Memory\u7684instruction\u548cSample\u5747\u8fdc\u9ad8\u4e8eShared \u548c Local<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Shared Memory\u5f88\u7279\u6b8a\uff0c\u867d\u7136\u4e4b\u524d\u8bf4\u8fc7Cache\u662f\u4e0d\u80fd\u5206\u914d\u7684\uff0c\u4f46Shared Memory\u786e\u5b9e\u548cL1 Cache\u5171\u4eab\u540c\u6837\u7684\u7269\u7406\u7a7a\u95f4\uff0c\u5c5e\u4e8e\u6240\u8c13\u7684 On-chip \u7247\u4e0a\u5185\u5b58\uff0cLoad\/Store \u90fd\u76f8\u5f53\u7684\u5feb\u3002\u56e0\u6b64\u5e38\u7528\u4e8e\u7279\u6b8a\u7684shader\u4f18\u5316\uff0c\u6bd4\u5982\u4e00\u4e2awarp\u5171\u4eab\u7684\u6570\u636e\u53ef\u4ee5\u653e\u5728shared memory\u4e2d\uff08TODO\uff1acheck this\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Local Memory \u662f\u5bc4\u5b58\u5668\u8d44\u6e90\u8017\u5c3d\u540e\u7684\u5907\u9009\u65b9\u6848\uff1a\u7ebf\u7a0b\u7684\u79c1\u6709\u6570\u636e\u672c\u6765\u5e94\u8be5\u5168\u90e8\u653e\u5728\u5bc4\u5b58\u5668\u4e2d\uff0c\u4f46\u5982\u679c\u5bc4\u5b58\u5668\u8d44\u6e90\u8017\u5c3d\uff0c\u5c31\u4f1a\u88ab\u653e\u5728\u663e\u5b58\u4e2d\u3002\u4e00\u822c\u7528\u6765\u89c2\u5bdf\u5bc4\u5b58\u5668\u8d44\u6e90\u662f\u5426\u6bd4\u8f83\u6781\u9650\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GPU\u67b6\u6784\uff1a\u4ee5Blackwell\u67b6\u6784\u4e3a\u4f8b\uff0c\u4e00\u4e2asubcore\u542b\u670916\u4e2aFP32\u4e13\u7528lane\uff0c\u4ee5\u53ca16\u4e2a\u53ef\u9009FP32\/INT32 lane\uff0c\u5ba3\u4f20\u7684\u65f6\u5019\u79f0\u4e3a32\u4e2a cuda core\u3002\u4e00\u4e2aSM\u67094\u4e2asubcore\uff0c\u56e0\u6b64\u6709128\u4e2acuda core\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u822c\u6765\u8bf4\uff0c\u5bf9\u4e8eGlobal Memory\u7684\u8c03\u7528\u6b21\u6570\u548cSample\u5747\u8fdc\u9ad8\u4e8eShared \u548c Local<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/NVIDIA-RTX-Blackwell-SM-Figure5-page11-791x1024.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"791\" height=\"1024\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/NVIDIA-RTX-Blackwell-SM-Figure5-page11-791x1024.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-2477\"  sizes=\"auto, (max-width: 791px) 100vw, 791px\" \/><\/div><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">LDCU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Load a Value from Constant Memory into a Uniform Register<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Q\uff1a\u4ec0\u4e48\u662fConstant Memory\uff1fConstant Memory \u548c Shared Memory \u7684\u533a\u522b\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e24\u79cd Memory \u7406\u8bba\u4e0a\u90fd\u662f\u6240\u8c13\u7247\u4e0a\u7f13\u5b58\uff08on-chip\u5c31\u662fSM\u5185\u90e8\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>shared memory<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>L1 cache\u7684\u4e00\u90e8\u5206\uff0c\u662f\u771f\u6b63\u7684\u7247\u4e0a\u7f13\u5b58<\/li>\n\n\n\n<li>\u4ec5\u4ec5\u5728\u4e00\u4e2ablock\u4e2d\u7684\u6240\u6709\u7ebf\u7a0b\u5171\u4eab <\/li>\n\n\n\n<li>\u53ef\u8bfb\u53ef\u5199<\/li>\n\n\n\n<li>Shared memory\u662fSM\u7684\u7a00\u7f3a\u8d44\u6e90\uff0c\u4f7f\u7528\u8fc7\u591a\u4f1a\u9650\u5236SM\u4e0a\u540c\u65f6\u9a7b\u7559\u7684\u7ebf\u7a0b\u5757\u6570\u91cf\uff08Occupancy\uff09<\/li>\n\n\n\n<li>\u4ee5RTX PRO 6000\u4e3a\u4f8b\uff0c\u4e00\u4e2aSM\u4e0a\u9762\u6709 256 KB\u7684L1 Cache\/Shared memory\u5171\u7528\u533a\u57df\uff0c\u5728\u4e0d\u505a\u7279\u6b8a\u8c03\u6574\u7684\u60c5\u51b5\u4e0b\uff0cL1 Cache \u548c Shared memory\u662f\u5404128 KB\uff0c\u4f46\u53ef\u4ee5\u901a\u8fc7\u51fd\u6570\u6765\u8bbe\u7f6e\u504f\u597dL1\u6216\u8005Shared<\/li>\n\n\n\n<li><strong><code>cudaDeviceSetCacheConfig(cudaFuncCache cacheConfig)<\/code><\/strong> \u5168\u5c40\u504f\u597d<\/li>\n\n\n\n<li><strong><code>cudaFuncSetCacheConfig(const void* func, cudaFuncCache cacheConfig)<\/code><\/strong> \u5185\u6838\u5c40\u90e8\u504f\u597d<\/li>\n\n\n\n<li>L1 Cache \u9002\u7528\u6570\u636e\u8bbf\u95ee\u4e0d\u89c4\u5219\u7684\u4efb\u52a1\uff0cShared Memory\u9002\u5408\u6570\u636e\u8bbf\u95ee\u89c4\u5219\uff0c\u7ebf\u7a0b\u95f4\u901a\u4fe1\u7b49\u4efb\u52a1<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-code\"><code>RWStructuredBuffer&lt;int> outputBuffer;\ngroupshared int sharedVar; \/\/ \u58f0\u660e\u4e00\u4e2a\u7ebf\u7a0b\u7ec4\u5171\u4eab\u7684int\n\n&#91;numthreads(64, 1, 1)]\nvoid CSMain(uint3 globalIdx : SV_DispatchThreadID,\n            uint3 localIdx : SV_GroupThreadID,\n            uint3 groupIdx : SV_GroupID) {\n    if (localIdx.x == 0)\n        sharedVar = groupIdx.x; \/\/ \u7ebf\u7a0b0\u5199\u5165\n\n    GroupMemoryBarrierWithGroupSync(); \/\/ \u540c\u6b65\u7ec4\u5185\u6240\u6709\u7ebf\u7a0b\n\n    outputBuffer&#91;globalIdx.x] = sharedVar; \/\/ \u6240\u6709\u7ebf\u7a0b\u8bfb\u53d6\n}<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>constant memory<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>constant memory  \u4f4d\u4e8e\u663e\u5b58\u4e2d\uff0c\u4f46\u6709\u4e13\u7528\u7684Constant Cache\u52a0\u901f\u8bbf\u95ee\u3002<\/li>\n\n\n\n<li>Constant Cache\u4e0eL1 Cache\u540c\u7ea7\uff0c\u4f46\u662f\u72ec\u7acb\u7684\u7f13\u5b58\u5355\u5143<\/li>\n\n\n\n<li>\u6bcf\u4e2aSM\u7684Constant Cache 8KB\uff0c\u903b\u8f91\u4e0a\u603b\u7684Constant Memory\u662f64KB<\/li>\n\n\n\n<li>\u53ea\u8bfb<\/li>\n\n\n\n<li>Push Constant \u548c  UBOS\uff08GLSL\uff09\/ cbuffer\uff08HLSL\uff09<\/li>\n\n\n\n<li>\u7ecf\u9a8c\u6765\u8bb2\uff0c\u5927\u90e8\u5206\u9700\u8981\u8bbf\u95ee\u7684\u5e38\u91cf\uff0c\u5305\u62ecMVP\u77e9\u9635\uff0c\u706f\u5149\u53c2\u6570\uff0c\u6e32\u67d3\u8bbe\u7f6e\uff0c\u9aa8\u9abc\u77e9\u9635\u90fd\u653e\u5728constant buffer\u91cc\u9762\uff0c\u56e0\u6b64\u5176\u8bbf\u95ee\u76f8\u5f53\u9891\u7e41\u3002\u5b9e\u9645\u4e0a\uff0c\u8be5\u5e27\u4e2d\u5176\u6307\u4ee4\u6570\u4ec5\u6b64\u4e8e\u51e0\u9879\u9ad8\u9891\u8ba1\u7b97\u3002 <\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-code\"><code>cbuffer matrixBuffer : register(b0) { \/\/ \u7ed1\u5b9a\u5230\u5bc4\u5b58\u5668\u69fd b0\n    float4x4 worldMatrix;\n    float4x4 viewMatrix;\n    float4x4 projectionMatrix;\n};<\/code><\/pre>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">UDP<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Uniform Data Path\uff0c\u5904\u7406\u6574\u4e2aWarp\u90fd\u76f8\u540c\u7684\u503c\u65f6\uff0c\u5e94\u7528\u7684\u4e00\u79cd\u4f18\u5316\uff0c\u901a\u5e38Sample\/Instruction\u6bd4\u8f83\u9ad8\uff0c\u8bf4\u660e\u8fd9\u91cc\u5355\u6761\u6307\u4ee4\u7684\u6d88\u8017\u66f4\u9ad8\u3002\u5f53\u7136\u8fd9\u91cc\u8fd8\u9700\u8981\u66f4\u591a\u7684\u89c2\u5bdf<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">ADU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Address Divergence Unit\uff0c\u5730\u5740\u5206\u6b67\u5904\u7406\u5355\u5143\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5904\u7406\u5206\u652f\u548c\u8df3\u8f6c<\/strong>\uff1a\u5904\u7406<code>if-else<\/code>\u3001<code>switch<\/code>\u3001\u5faa\u73af\u7b49\u5bfc\u81f4\u7684\u7ebf\u7a0b\u675f\u5185\u6267\u884c\u8def\u5f84\u5206\u6b67\uff08Warp Divergence\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5904\u7406\u7d22\u5f15\u5e38\u91cf\u52a0\u8f7d<\/strong>\uff1a\u8d1f\u8d23\u5904\u7406\u90a3\u4e9b\u7d22\u5f15\u503c\u5728warp\u5185\u5404\u7ebf\u7a0b\u95f4\u4e0d\u4e00\u81f4\u7684\u5e38\u91cf\u5185\u5b58\u8bbf\u95ee\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5904\u7406\u5c4f\u969c\u4e0e\u540c\u6b65<\/strong>\uff1a\u53c2\u4e0e\u5904\u7406<code>__syncthreads()<\/code>\u7b49\u7ebf\u7a0b\u5757\u5185\u540c\u6b65\u64cd\u4f5c\u7684\u63a7\u5236\u903b\u8f91<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u80fd\u53ef\u4ee5\u770b\u51fa\u5206\u652f\u7d22\u5f15\u5bfc\u81f4\u7684\u4f4e\u6548\u95ee\u9898\uff0c\u4f46\u5927\u90e8\u5206\u60c5\u51b5\u4e0b\u4e0d\u662f\u4e3b\u8981\u74f6\u9888\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">SCH<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Scheduler \u8c03\u5ea6\u5668\u3002\u8bb0\u5f55\u7684\u662f GPU \u7684<strong>\u8c03\u5ea6\u5668\u548c\u63a7\u5236\u903b\u8f91\u5355\u5143<\/strong>\u5904\u4e8e\u6d3b\u8dc3\u72b6\u6001\u7684\u65f6\u95f4\u5360\u6bd4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u7a7a\u8f6c\u7b49\u5f85\u4e5f\u4f1a\u88ab\u8bb0\u5f55\u5728SCH\u7684Sample\u91cc\u9762\uff0c\u6240\u4ee5\u5176\u5f02\u5e38\u53ef\u80fd\u53cd\u6620\u4e86\u8c03\u5ea6\u5668\u7b49\u5f85\u95ee\u9898\u3002\u53ef\u80fd\u6765\u81ea\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>SM\u4e0a\u6d3b\u8dc3\u7684 warp \u6570\u91cf\u5f88\u5c11\u65f6\uff0c\u5f53\u4e00\u4e9b warp \u6267\u884c Load \u65f6\uff08LSU\/LDCU\u7684\u5de5\u4f5c\uff09\uff0cSCH\u672c\u6765\u5e94\u8be5\u5207\u53e6\u4e00\u6279warp\u4e0a\u6765\u5de5\u4f5c\uff0c\u4f46\u7531\u4e8ewarp\u4e0d\u8db3\u65e0\u6cd5\u5207\u6362\uff0c\u56e0\u6b64\u4f1a\u63d2\u5165 NOP \u5bfc\u81f4\u7a7a\u8f6c\u3002<\/li>\n\n\n\n<li>__syncthreads()\uff0c\u4f7f\u540c\u4e00\u4e2ablock\u5185\u90e8\u7684warp\u7b49\u5f85\u5230\u6700\u540e\u4e00\u4e2a\u6267\u884c\u5b8c\u6bd5\uff0c\u5f15\u5165\u540c\u6b65\u5f00\u9500\uff0c\u8fd9\u91cc\u4e5f\u7b97\u662f\u7a7a\u8f6c<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">CBU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">Convergence Barrier Unit\uff0c\u5173\u6ce8<strong>\u7ebf\u7a0b\u675f\u5206\u6b67<\/strong>\u95ee\u9898\u672c\u8eab\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cADU\u66f4\u5173\u6ce8\u7684\u662f\u4e00\u4e9b\u884c\u4e3a\u9020\u6210\u7684\uff0c<strong>\u975e\u5408\u5e76\u7684\u5730\u5740\u8bbf\u95ee\u6a21\u5f0f<\/strong>\u95ee\u9898\u3002\u53ea\u662f\u5bfc\u81f4\u4e24\u8005\u7684\u8fd9\u4e9b\u884c\u4e3a\u76f8\u8fd1\uff0c\u56e0\u6b64\u8fd9\u4e24\u4e2a\u6307\u6807\u4e5f\u5b58\u5728\u4e00\u5b9a\u5173\u8054\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Branch<\/strong>\uff1a<strong>\u5206<\/strong>\uff0c\u5373\u4ee3\u8868\u5206\u652f\u6307\u4ee4\uff08<code>if-else<\/code>\u3001<code>switch\uff09<\/code>\u672c\u8eab\u5e26\u6765\u7684\u5f00\u9500<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>SIMT<\/strong>\uff1a<strong>\u5408<\/strong>\uff0c\u5373\u4ee3\u8868__syncthreads() \u8fd9\u7c7b\u540c\u6b65\u7b49\u5f85\u5f00\u9500<\/li>\n\n\n\n<li><strong>EXIT<\/strong>\uff1a\u7ec8\uff0c\u5373\u4ee3\u8868<strong>Kernel Exit<\/strong>\u8fd9\u4e2a\u7ed3\u5c3e\u6bb5\u7684\u5f00\u9500<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">XU<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u7279\u6b8a\u51fd\u6570\u4e0e\u7c7b\u578b\u8f6c\u6362\u5355\u5143\uff0c\u7c7b\u578b\u8f6c\u6362\u4e00\u822c\u5728Operation\u5217\u91cc\u4f1a\u76f4\u63a5\u5199\u660e<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">TEX<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u9996\u5148\uff0c\u5728Shader\u4e2d\u6211\u4eec\u63a5\u89e6\u5230\u7684\u56fe\u50cf\u5bf9\u8c61\uff0c\u53ef\u4ee5\u88ab\u5206\u4e3aSurface\u548cTexture<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Texture\uff0c\u5373\u4f20\u7edf\u610f\u4e49\u4e0a\u7684\u7eb9\u7406\uff0c\u53ea\u8bfb\uff0c\u4e14\u53ea\u80fd\u7528\u4e8e<strong>sample<\/strong>\uff08\u5728Nsight graphics\u91cc\u9762\u53eb<strong>Fetch<\/strong>\uff09\uff0c\u56e0\u6b64\u901a\u5e38\u4f34\u968f\u7740\u786c\u4ef6\u6ee4\u6ce2\uff08\u5982\u53cc\u7ebf\u6027\u3001\u4e09\u7ebf\u6027\u63d2\u503c\uff09<\/li>\n\n\n\n<li>Surface\uff0c\u50cf\u662fps\u4e2d\u5199\u5165render target\uff0c\u6216\u8005cs\u4e2d\u5199\u5165UAV\uff0c\u50cf\u7d20\u4e4b\u95f4\u901a\u5e38\u662f\u72ec\u7acb\u7684\uff0c\u4e14\u5927\u90e8\u5206\u60c5\u51b5\u4e0b\u662f\u6570\u636e\u7684\u8f7d\u4f53\u3002\u6709Fetch\uff0cStore\u548cAtomic\u4e09\u79cd\u64cd\u4f5c\u3002Atomic\u4e00\u822c\u901a\u8fc7<code>InterlockedAdd<\/code>,\u00a0<code>InterlockedCompareExchange<\/code>\u8fd9\u4e9b\u51fd\u6570\u5b9e\u73b0\uff0c\u4fdd\u8bc1UAV\u8d44\u6e90\u5199\u5165\u7684\u539f\u5b50\u6027\uff0c\u4e0d\u4f1a\u56e0\u4e3a\u7ade\u4e89\u4ea7\u751f\u9519\u8bef\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">RTCORE<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u987e\u540d\u601d\u4e49\uff0c\u6307\u4f1a\u8c03\u7528Ray tracing core\u7684\u884c\u4e3a\uff0c\u7c7b\u4f3ctrace ray\u8fd9\u79cd\u51fd\u6570\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u6307\u6807\u9605\u8bfb<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6240\u6709\u8fd9\u4e2a\u680f\u76ee\u7684\u6570\u636e\u4e2d\uff0c\u90fd\u542b\u6709 Sample \u548c Instruction\u4e24\u9879\uff0c\u4e3a\u4e86\u771f\u6b63\u7406\u89e3\u6570\u636e\u7684\u542b\u4e49\uff0c\u9996\u5148\u9700\u8981\u5bf9\u8fd9\u4e24\u9879\u6307\u6807\u505a\u4e00\u4e0b\u4e25\u8c28\u7684\u5b9a\u4e49\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. Sample <\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u5047\u8bbe\uff0cnsight graphics\u6bcf\u4e00\u79d2\u90fd\u5bf9GPU\u8fdb\u884c\u5927\u91cf\u7684sample\uff0c\u968f\u673a\u7684sample\u5728\u6570\u91cf\u591f\u591a\u65f6\u53cd\u6620\u6982\u7387\u3002\u56e0\u6b64\uff0csample\u6307\u6807\u53cd\u6620\u4e86<strong>\u4e00\u6bb5\u65f6\u95f4\u5185\uff0cGPU\u5728\u8fdb\u884c\u67d0\u9879\u884c\u4e3a\u7684\u6982\u7387<\/strong>\u3002\u5728\u5b9e\u9645\u4f7f\u7528\u65f6\uff0c\u53ef\u4ee5\u88ab\u8fd1\u4f3c\u7684\u89c6\u4e3a\uff0c\u67d0\u9879\u884c\u4e3a\u5728\u4e00\u4e2a\u533a\u95f4\u5185\u7684<strong>\u65f6\u5ef6\u60c5\u51b5<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5b9e\u9645\u6d4b\u91cf\u4e2d\uff0c\u89c2\u5bdf\u5230\u603bSample\u6570\u5e76\u4e0d\u662f\u65f6\u95f4\u5747\u5300\u7684\u3002\u6bd4\u5982UE render\u4e2d\uff0c\u5728TSR\uff08\u540e\u5904\u7406\u9636\u6bb5\u7684\u65f6\u5e8f\u8d85\u5206\uff09\u4e2d\uff0c\u5176\u5728322\u5fae\u79d2\u5185\uff0c\u603b\u91c7\u6837\u6b21\u6570\u8fbe\u5230\u4e8614536\uff0cLumen Screen Probe Gather \u548cDirect Lighting\u4e5f\u5f88\u9891\u7e41\u3002\u4f46\u5728Nanite draw geometry\u6bb5\u5c31\u8981\u4f4e\u7684\u591a\u3002\u4ece\u89c2\u5bdf\u4e0a\u770b\uff0c\u4f3c\u4e4e\u4e0eSM Warp Occupancy\u76f8\u5173\u3002\u731c\u6d4b\u662f\uff0c\u867d\u7136\u5355\u4f4d\u65f6\u95f4\u7684sample\u6570\u4e00\u5b9a\uff0c\u4f46\u53ea\u6709sample\u5230\u6d3b\u8dc3\u6838\u5fc3\u7684\u624d\u4f1a\u88ab\u4fdd\u7559\u5230\u62a5\u544a\u4e2d\u3002\u56e0\u6b64workload\u6bd4\u8f83\u9ad8\uff0csample\u624d\u6bd4\u8f83\u5bc6\u96c6\u3002\u5f53\u7136\u8fd9\u91cc\u76ee\u524d\u8fd8\u6ca1\u6709\u6570\u503c\u5316\u7684\u8bc1\u636e\u652f\u6491\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. Instruction<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4Sample\u53cd\u6620\u4e86\u65f6\u5ef6\uff0c\u90a3\u4e48Instruction\u5c31\u5bf9\u5e94\u4e86\u9891\u7387\uff0c\u5373\u5728\u5355\u4f4d\u65f6\u95f4\u5185\u8fd9\u4e2a\u884c\u4e3a\u53d1\u751f\u4e86\u591a\u5c11\u6b21\u3002\u914d\u5408\u65f6\u5ef6\uff0c\u5047\u5982\u4e00\u4e2a\u884c\u4e3a\u7684Sample\u5f88\u9ad8\uff0c\u4f46Instruction\u5f88\u4f4e\uff0c\u8bf4\u660e\u5176\u53ef\u80fd\u662f\u5355\u6b21\u6d88\u8017\u5f88\u9ad8\uff0c\u6216\u8005\u6d41\u7a0b\u4e2d\u5b58\u5728\u5927\u91cfstall\u7684\u60c5\u51b5\uff0c\u662f\u9700\u8981\u989d\u5916\u8b66\u60d5\u7684\u60c5\u51b5\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. Sample \u7ec6\u8282<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728Sample\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5f85\u5176\u88ab\u5212\u5206\u4e3a\u4ee5\u4e0b\u8fd9\u4e9b\u7ec6\u8282\u7edf\u8ba1\u9879\u3002\u8fd9\u4e9b\u6570\u636e\u662f<strong> Warp Stall Reasons<\/strong>\u3002\u5728\u67d0\u4e00\u4e2a\u65f6\u523b\uff0c\u4e00\u4e2awarp\u603b\u662f\u5728\u7b49\u5f85\u4ec0\u4e48\uff1a\u53ef\u80fd\u662f\u5728\u7b49\u6570\u636e\uff0c\u7b49\u6267\u884c\uff0c\u7b49\u6307\u4ee4\uff0c\u6216\u8005\u7b49\u540c\u6b65\u3002\u56e0\u6b64\uff0c\u628a\u8fd9\u4e9b\u6570\u636e\u5168\u90e8\u5217\u51fa\uff0c\u6709\u52a9\u4e8e\u660e\u786e\u67d0\u4e00\u9879\u5177\u4f53\u7684\u6307\u4ee4\u7684\u74f6\u9888\u5728\u54ea\u91cc\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/image-1.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"315\" height=\"267\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2026\/08\/image-1.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-2507\"  sizes=\"auto, (max-width: 315px) 100vw, 315px\" \/><\/div><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Data Dependency<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9996\u5148\u662f\u6700\u5e38\u89c1\u7684stall reason\uff0c\u5373<strong>\u7b49\u6570\u636e<\/strong>\uff0c\u5176\u88ab\u5206\u4e3a\u4e0b\u9762\u4e24\u79cd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Long Scoreboard (\u957f\u8bb0\u5206\u677f)<\/strong>\uff1a\u7ebf\u7a0b\u675f\u6b63\u5728\u7b49\u5f85\u6765\u81ea<strong>\u7247\u5916<\/strong>\u5185\u5b58\uff08\u5982\u5168\u5c40\u5185\u5b58\u3001\u7eb9\u7406\u5185\u5b58\uff09\u7684<strong>\u957f\u5ef6\u8fdf<\/strong>\u6570\u636e<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Short Scoreboard (\u77ed\u8bb0\u5206\u677f)<\/strong>\uff1a\u7ebf\u7a0b\u675f\u6b63\u5728\u7b49\u5f85\u6765\u81ea<strong>\u7247\u4e0a<\/strong>\u5b58\u50a8\uff08\u5982\u5171\u4eab\u5185\u5b58\uff09\u7684<strong>\u77ed\u5ef6\u8fdf<\/strong>\u6570\u636e<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u80fd\u4f1a\u6709\u8fd9\u6837\u7684\u7591\u95ee\uff1a\u5bf9\u4e8eFMA\u7684FP32\u6765\u8bf4\uff0c\u5176\u672c\u8eab\u4ee3\u8868\u7740\u5bf9\u4e24\u4e2a\u5bc4\u5b58\u5668\u7684\u64cd\u4f5c\uff0c\u90a3\u4e48\u4e3a\u4ec0\u4e48\u5176Sample\u4e2d\uff0cLong\/Short Scoreboard\u4ecd\u7136\u5360\u636e\u4e86\u5927\u90e8\u5206\uff1f\u7406\u8bba\u4e0a\u4e0d\u662f\u5e94\u8be5\u5728LSU\/LDCU\u8fd9\u79cd\u663e\u5b58\u64cd\u4f5c\u4e2d\u624d\u6bd4\u8f83\u660e\u663e\u5417\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9645\u4e0a\uff0c\u8fd9\u662f\u56e0\u4e3a FMA \u4f9d\u8d56\u7684\u5bc4\u5b58\u5668\uff0c\u5728launch\u8fd9\u4e00\u6761\u6307\u4ee4\u7684\u65f6\u5019\uff0c\u4e0d\u4e00\u5b9a\u51c6\u5907\u597d\u4e86\u3002\u4e3e\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>float val = tex2D(myTex, uv).x;  \/\/ \u5168\u5c40\u7eb9\u7406\u8bbf\u95ee\uff0c\u5ef6\u8fdf~200-300 cycles\nfloat result = val * 2.0 + 0.5;  \/\/ FMA\u4f9d\u8d56val<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9636\u6bb5\u4e00\uff1a\u53d1\u5c04 TEX \u6307\u4ee4<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8c03\u5ea6\u5668\u9009\u4e2d\u8be5 Warp\uff0c\u68c0\u67e5 TEX \u6307\u4ee4\u7684\u6e90\u5bc4\u5b58\u5668\uff08\u7eb9\u7406\u5750\u6807\uff09\u5df2\u5c31\u7eea\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8c03\u5ea6\u5668\u5c06 TEX \u6307\u4ee4<strong>\u53d1\u5c04<\/strong>\u7ed9 TEX \u5355\u5143\uff08TMU\uff09\u3002\u6b64\u65f6\uff0c\u8fd9\u6761\u6307\u4ee4\u5c31\u7ed3\u675f\u4e86\uff0c\u5b83\u4e0d\u518d\u5360\u7528\u8c03\u5ea6\u5668\u7684\u53d1\u5c04\u5e26\u5bbd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TEX \u5355\u5143\u53d1\u9001\u5185\u5b58\u8bf7\u6c42\u5230 L1\/L2\/DRAM\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9636\u6bb5\u4e8c\uff1aWarp \u8fdb\u5165\u201c\u8bb0\u5206\u677f\u7b49\u5f85\u201d\u72b6\u6001\uff08Long Scoreboard\uff09<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u5728\uff0c\u6307\u4ee4\u6307\u9488\uff08PC\uff09\u6307\u5411\u4e0b\u4e00\u6761\u6307\u4ee4\uff0c\u5373\u4f9d\u8d56\u4e8e TEX \u7ed3\u679c\u7684 FMA\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8c03\u5ea6\u5668\u4f1a\u68c0\u67e5\u8fd9\u6761 FMA \u6307\u4ee4\u7684\u6e90\u5bc4\u5b58\u5668\uff08\u7528\u4e8e\u5b58\u653e\u7eb9\u7406\u8fd4\u56de\u503c\u7684\u76ee\u6807\u5bc4\u5b58\u5668\uff09\u3002\u8bb0\u5206\u677f\u663e\u793a\u8be5\u5bc4\u5b58\u5668\u5904\u4e8e\u00a0<strong>\u201c\u672a\u5c31\u7eea\uff08Pending\uff09\u201d<\/strong>\u00a0\u72b6\u6001\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5173\u952e\u70b9<\/strong>\uff1a\u7531\u4e8e\u6e90\u64cd\u4f5c\u6570\u672a\u5c31\u7eea\uff0c\u8c03\u5ea6\u5668<strong>\u4e0d\u4f1a<\/strong>\u53d1\u5c04\u8fd9\u6761 FMA\u3002\u5b83\u4e0d\u4f1a\u5c06 FMA \u63a8\u8fdb FP \u6d41\u6c34\u7ebf\u3002\u5b83\u53ea\u662f\u8ba9\u8fd9\u4e2a Warp \u8fdb\u5165\u00a0<strong>Stall\uff08\u505c\u987f\uff09<\/strong>\u00a0\u72b6\u6001\uff0c\u5e76\u8bb0\u5f55\u4e0b\u00a0<strong>Long Scoreboard\u00a0<\/strong>\u505c\u987f\u539f\u56e0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PS\uff1a\u6b64\u65f6\u5b9e\u9645\u4e0aTEX\u5355\u5143\u81ea\u5df1\u4e5f\u5728\u8bb0\u5f55 <strong>Long Scoreboard<\/strong>\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0cTEX\u5728\u7b49\u8bbf\u5b58\uff0cFMA\u5728\u7b49TEX\u5199\u5165\u5bc4\u5b58\u5668\uff0c\u5b9e\u9645\u4e0a\u8fd9\u4e24\u4e2a <strong>Long Scoreboard<\/strong> \u5728\u7edd\u5bf9\u65f6\u95f4\u8f74\u4e0a\u7684\u91cd\u5408\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9636\u6bb5\u4e09\uff1a\u5ef6\u8fdf\u9690\u85cf<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6839\u636eGPU\u7684\u8bbe\u8ba1\uff0c\u6b64\u65f6\u8c03\u5ea6\u5668\u4f1a\u8c03\u53e6\u5916\u4e00\u4e2a <strong>Ready<\/strong> \u7684 Warp \u4e0a\u6765\uff0c\u53d1\u5c04\u5b83\u7684\u7b97\u672f\/\u8bbf\u5b58\u6307\u4ee4<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9636\u6bb5\u56db\uff1a\u6570\u636e\u8fd4\u56de &amp; \u53d1\u5c04 FMA<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ecf\u8fc7\u51e0\u767e\u4e2a\u5468\u671f\u540e\uff0c\u7eb9\u7406\u6570\u636e\u4ece\u663e\u5b58\u8fd4\u56de\uff0c\u5199\u5165\u76ee\u6807\u5bc4\u5b58\u5668\u3002\u8bb0\u5206\u677f\u5c06\u8be5\u5bc4\u5b58\u5668\u6807\u8bb0\u4e3a <strong>Ready<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4e0b\u4e00\u4e2a\u8c03\u5ea6\u5468\u671f\uff0c\u8c03\u5ea6\u5668\u518d\u6b21\u68c0\u67e5\u8fd9\u4e2a Warp \u7684 FMA \u6307\u4ee4\uff0c\u53d1\u73b0\u6e90\u5bc4\u5b58\u5668\u5c31\u7eea\uff0c\u4e8e\u662f<strong>\u73b0\u5728\u624d\u771f\u6b63\u5c06 FMA \u6307\u4ee4\u53d1\u5c04<\/strong>\u7ed9 FP \u6d41\u6c34\u7ebf\u6267\u884c\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Pipeline Busy<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7c7b\u505c\u987f\u6e90\u4e8e\u6267\u884c\u6307\u4ee4\u6240\u9700\u7684\u786c\u4ef6\u5355\u5143\u6b63\u5fd9\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Math Pipe Throttle (\u6570\u5b66\u6d41\u6c34\u7ebf\u8282\u6d41)<\/strong>\uff1a\u6307\u4ee4\u6240\u9700\u7684\u8ba1\u7b97\u6d41\u6c34\u7ebf\uff08\u5982<code>ALU<\/code>,\u00a0<code>FMA<\/code>,\u00a0<code>Tensor Core<\/code>\uff09\u6b63\u5fd9<a href=\"https:\/\/docs.nvidia.com\/nsight-compute\/pdf\/ProfilingGuide.pdf#9#6\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\u3002\u8fd9\u901a\u5e38\u610f\u5473\u7740\u8ba1\u7b97\u4efb\u52a1\u672c\u8eab\u8fc7\u91cd\u3002\u8fd9\u4e2a\u6307\u6807\u53ea\u5728\u8ba1\u7b97\u76f8\u5173\u7684 Pipe \u4e2d\u51fa\u73b0<\/li>\n\n\n\n<li><strong>MIO Throttle (MIO \u8282\u6d41)<\/strong>\uff1a\u6307\u4ee4\u6240\u9700\u7684<strong>MIO<\/strong>\uff08Memory Input\/Output\uff09\u5355\u5143\u961f\u5217\u5df2\u6ee1<a href=\"https:\/\/docs.nvidia.com\/nsight-compute\/pdf\/ProfilingGuide.pdf#9#6\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\u3002MIO\u8d1f\u8d23\u5904\u7406\u5171\u4eab\u5185\u5b58\u3001\u7279\u6b8a\u6570\u5b66\u6307\u4ee4\u548c\u52a8\u6001\u5206\u652f\u7b49\u64cd\u4f5c<a href=\"https:\/\/66ring.github.io\/2024\/02\/28\/universe\/gpu\/nsight_by_study_case\/#SOL-Speed-of-Light\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><a href=\"https:\/\/www.techqa.club\/v\/q\/when-does-mio-throttle-stall-happen-66233462\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\u3002\u8fd9\u4e2a\u6307\u6807\u53ea\u5728LSU\uff0cLDCU\u8fd9\u4e9b Pipe \u4e2d\u51fa\u73b0<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Scheduling &amp; Fetch<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7c7b\u505c\u987f\u4e0e\u6307\u4ee4\u7684\u8c03\u5ea6\u548c\u83b7\u53d6\u6709\u5173\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Dispatch Stall (\u8c03\u5ea6\u505c\u987f)<\/strong>\uff1a\u56e0\u8c03\u5ea6\u8d44\u6e90\u4e0d\u53ef\u7528\u800c\u65e0\u6cd5\u53d1\u5c04\u6307\u4ee4\u3002\u53ef\u80fd\u662f\u7531\u4e8e\u6307\u4ee4\u7f13\u5b58\u672a\u547d\u4e2d\u3001\u6d41\u6c34\u7ebf\u51b2\u5237\u7b49\u539f\u56e0\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>No Instructions (\u65e0\u6307\u4ee4)<\/strong>\uff1a\u7ebf\u7a0b\u675f<strong>\u6ca1\u6709\u8981\u6267\u884c\u7684\u6307\u4ee4<\/strong>\u3002\u53ef\u80fd\u56e0\u4e3a\u6307\u4ee4\u7f13\u5b58\u672a\u547d\u4e2d<a href=\"https:\/\/66ring.github.io\/2024\/02\/28\/universe\/gpu\/nsight_by_study_case\/#SOL-Speed-of-Light\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><a href=\"https:\/\/docs.nvidia.com\/nsight-compute\/pdf\/ProfilingGuide.pdf#9#6\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\uff0c\u6216\u5185\u6838\u672c\u8eab\u5de5\u4f5c\u91cf\u8fc7\u5c0f<a href=\"https:\/\/docs.nvidia.com\/nsight-compute\/pdf\/ProfilingGuide.pdf#9#6\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Not Selected (\u672a\u88ab\u9009\u4e2d)<\/strong>\uff1a\u7ebf\u7a0b\u675f\u672c\u8eab<strong>\u53ef\u4ee5\u6267\u884c<\/strong>\uff08\u6570\u636e\u5df2\u5c31\u7eea\uff09\uff0c\u4f46\u8c03\u5ea6\u5668\u9009\u62e9\u4e86<strong>\u5176\u4ed6<\/strong>\u7ebf\u7a0b\u675f<a href=\"https:\/\/66ring.github.io\/2024\/02\/28\/universe\/gpu\/nsight_by_study_case\/#SOL-Speed-of-Light\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><a href=\"https:\/\/docs.nvidia.com\/nsight-compute\/pdf\/ProfilingGuide.pdf#9#6\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a>\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Branch Resolving (\u5206\u652f\u89e3\u6790)<\/strong>\uff1a\u7ebf\u7a0b\u675f\u6b63\u5728\u5904\u7406\u4e00\u4e2a<strong>\u5206\u652f\u6307\u4ee4<\/strong>\uff0c\u5bfc\u81f4\u505c\u987f\u3002\u901a\u5e38\u4e0e\u52a8\u6001\u5206\u652f\u6216\u7ebf\u7a0b\u675f\u5185\u5206\u6b67\uff08Divergence\uff09\u6709\u5173\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Others<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Wait\uff1a\u4e00\u822c\u6307\u540c\u6b65\u7b49\u5f85\u72b6\u6001\uff0c\u5982__syncthreads()<\/li>\n\n\n\n<li>Selected\uff1a\u6b63\u5e38\u7684\u5de5\u4f5c\u72b6\u6001\uff0c\u4e0d\u662f\u505c\u987f\u539f\u56e0<\/li>\n\n\n\n<li>Drain\uff1a\u6307\u6d41\u6c34\u7ebf\u6216\u961f\u5217\u6b63\u5728\u88ab\u201c\u6392\u7a7a\u201d\u3002\u4f8b\u5982\uff0c\u5728\u8c03\u5ea6\u4e2d\u95f4\u53d1\u751f\u7684\u505c\u987f\uff0c\u53ef\u80fd\u4e0e\u5b50\u901a\u9053\u5207\u6362\u6709\u5173<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u4e4b\u524d\u5206\u6790GPU\u7cfb\u7edf\u7684\u65f6\u5019\uff0c\u7528\u7684\u591a\u662fNsight System\u3002\u4f46\u76ee\u524d\u53d1\u73b0Nsight Graphics\u53ef\u80fd\u66f4\u9002 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[29,32],"class_list":["post-2457","post","type-post","status-publish","format-standard","hentry","category-8","tag-gpu","tag-32"],"_links":{"self":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2457","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/comments?post=2457"}],"version-history":[{"count":57,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2457\/revisions"}],"predecessor-version":[{"id":2520,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2457\/revisions\/2520"}],"wp:attachment":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/media?parent=2457"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/categories?post=2457"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/tags?post=2457"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}