連結已複製
7 回應
Thanks for your feedback, even if it'sbad news. It was really cool to have the IACA with AVX support available one full year before the Sandy Bridgelaunch and I was expecting more or less the same thing for AVX2, IIRC it was told at IDF that it's Intel's new way to help developers (or something along this line), too bad it doesn't lastformore thana single tock.
Strictly speaking, IACA doesn't support specific instruction sets, it supports specific Intel microarchitectures. The version of IACA that was released with AVX support prior to Sandy Bridge being launched did not correspond to an actual microarchitecture, it was more like the existing (Nehalem) architecture with some AVX added on. For AVX2 we won't be producing such a hybrid model (it's a significant effort and the results wouldn't exactly correspond to any actual architecture), instead AVX2 will be supported by the release for Haswell support.
The version of IACA that was released with AVX support prior to Sandy Bridge being launched did not correspond to an actual microarchitecture, it was more like the existing (Nehalem) architecture with some AVX added on
well,early versions(I just tested with iacaArchDataAVX.dll & all dated 12-Nov-2009) were a very good start already, they were for example accurately modeling the 2nd load port and the 48B/clock throughput of the L1D$,also the half width VSQRT/VDIV (2 uops for AVX-256) and the port limitation for SHUFFLE,they weremore SNB oriented than NHM arguably since the 2:1 load:store ratio is a drastic change vs. 1:1 previously,another hint itwas adequately targetting SNB from day 1 is thatNehalem support was added only aftwerwards in later versions (starting with v1.1 according to the changes log here: IACA) and that official Sandy Bridge support was never "added". Anyway, if Sandy Bridge support isn't very good as you said when will we get an update ?
