23.rv64.zip 一统碎片化的 RISC-V ISA 生态.pdf

《23.rv64.zip 一统碎片化的 RISC-V ISA 生态.pdf》由会员分享，可在线阅读，更多相关《23.rv64.zip 一统碎片化的 RISC-V ISA 生态.pdf（17页珍藏版）》请在三个皮匠报告上搜索。

1、rv64.ziprv64.zip:Unifying:Unifying diverse RISCdiverse RISC-V V ISA ecoISA eco-systemsystem一一统碎片化的RISCRISC-V V ISAISA 生生态YANGYU CHENYANGYU CHEN（陈泱宇）CYYCYYSELF.NAMECYYCYYSELF.NAMEBackgroundBackground Our dream:RVA23U64 However,existing hardware:C920v1C920v120232023RVA20RVA20P550P5502022024 4RVA20+RVA

2、20+ZbaZbaZbbZbbX60X6020242024RVA22+RVA22+VectorVectorZicondZicondNanhuNanhu v2v220242024RVA2RVA20+0+ZbaZbbZbsRVA23U64 Mandatory ExtensionsSoftware distributionSoftware distributionBinary distribution:RV64GCManually dispatch multi-version:glibcffmpegOpenSSLRVA23 distro is not as beautiful as your ima

3、ginationRVA23 distro is not as beautiful as your imaginationRVA23RVA23 compilation can yield worseworse performance than RV64GCRV64GC with current compiler and hardwarecurrent compiler and hardware,especially for VectorVector extensions.RISC-V is not alone.This issue is also present on the aarch64 w

4、ith the SVE2 1 1 https:/blog.cyyself.name/wp-content/uploads/2025/02/Screenshot-2025-02-09-at-19.17.59.pngPerformance degradation with new ExtensionsPerformance degradation with new Extensions00.511.522.5SPECCPU 2006 onSPECCPU 2006 on XuantieXuantie C920v2 with GCC15.0C920v2 with GCC15.0gcgc_zba_zbb

5、gcbv_zicondMaximum 3.34x Maximum 3.34x slowdown!slowdown!5%GEOMEAN 5%GEOMEAN slowdown!slowdown!Opportunities are always thereOpportunities are always thereRemark:Recorded using GCC14.2+O3+RV64GCObservationObservationNew extensions can also limitlimit the performanceThe number of performanceperforman

6、ce-sensitivesensitive functions is smallsmall Solution:Function clone of different extensionsSolution:Function clone of different extensionsPreserve binary Preserve binary compatibilitycompatibility (even with RV64GC)(even with RV64GC)HighHigherer performance on RVA23 hardware via dynamic dispatchpe