[dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on i40e

[Rong, Leyi]

> -----Original Message-----
> From: David Marchand <david.marchand at redhat.com>
> Sent: Monday, January 25, 2021 10:35 PM
> To: Kadam, Pallavi <pallavi.kadam at intel.com>
> Cc: Yigit, Ferruh <ferruh.yigit at intel.com>; Ali Alnubani <alialnu at nvidia.com>;
> NBU-Contact-Thomas Monjalon <thomas at monjalon.net>; Richardson, Bruce
> <bruce.richardson at intel.com>; Tal Shnaiderman <talshn at nvidia.com>; Odi Assli
> <odia at nvidia.com>; Rong, Leyi <leyi.rong at intel.com>; Zhang, Qi Z
> <qi.z.zhang at intel.com>; Lu, Wenzhuo <wenzhuo.lu at intel.com>; Xing, Beilei
> <beilei.xing at intel.com>; Menon, Ranjit <ranjit.menon at intel.com>;
> dev at dpdk.org; Raslan Darawsheh <rasland at nvidia.com>
> Subject: Re: [dpdk-dev] [PATCH v3 0/3] AVX512 vPMD on i40e
> 
> On Thu, Jan 21, 2021 at 6:02 AM Kadam, Pallavi <pallavi.kadam at intel.com>
> wrote:
> > On 1/20/2021 11:21 PM, Ferruh Yigit wrote:
> >
> > And for the mingw, I have same result with Ali, I can reproduce with (Fedora
> MinGW 9.2.1-6.fc32).
> >
> > But different from the CI, I am getting the error [1] for all following files:
> > rte_random.c
> > i40e_rxtx_vec_sse.c
> > i40e_rxtx_vec_avx512.c
> > i40e_rxtx_vec_avx2.c
> > rte_ethdev.c
> >
> > [1] Error: invalid register for .seh_savexmm
> >
> >
> > There is a stackoverflow entry for it:
> > https://stackoverflow.com/questions/43152633/invalid-register-for-seh-
> > savexmm-in-cygwin
> >
> > If I use '-fno-asynchronous-unwind-tables' as suggested there, the build works
> fine.
> >
> > So the problem may not be just 'i40e_rxtx_vec_avx512.c'.
> >
> >
> > If I change the machine type from 'native' to 'corei7' [2], the build error
> reduced to only 'i40e_rxtx_vec_avx512.c', so the problem seems happens when
> avx512 is supported by CPU, in this case compiler seems has a defect.
> > And since for 'i40e_rxtx_vec_avx512.c' the '-march=skylake-avx512' explicitly
> set can cause the problem seen in all machines.
> >
> > [2]
> >  diff --git a/config/x86/cross-mingw b/config/x86/cross-mingw  index
> > 4c15a7fa2e..7cee238add 100644
> >  --- a/config/x86/cross-mingw
> >  +++ b/config/x86/cross-mingw
> >  @@ -9,5 +9,5 @@ pkgconfig = 'x86_64-w64-mingw32-pkg-config'
> >   [host_machine]
> >   system = 'windows'
> >   cpu_family = 'x86_64'
> >  -cpu = 'native'
> >  +cpu = 'corei7'
> >   endian = 'little'
> >
> > @Ranjit, @Pallavi,
> > Are you building using mingw, and if so are you observing same problem?
> >
> 
> Thanks Ferruh.
> 
> 
> > We usually build using Clang. However, we verify with mingw as well before
> submitting the patch.
> > As mentioned in the patch [1] comments, we replaced #include
> > x86intrin.h with <rte_vect.h> in the file i40e_rxtx_vec_avx2.c And this helped
> fixing an error related to conflicting types for '__m_prefethw' with Clang on
> Windows.
> > I was able to build this patch using Clang as well as mingw.
> > [1] http://patches.dpdk.org/patch/84770/
> >
> > I verified patch fix submitted by Rong, Leyi, it builds successfully with Clang.
> > However, I am getting same error "Error: invalid register for .seh_savexmm"
> using mingw for 'i40e_rxtx_vec_avx512.c' file.
> 
> The patch I sent https://patchwork.dpdk.org/patch/86999/ is not enough.
> I have neither time nor environment to find a fix.
> 
> For now I simply stopped checking mingw builds.
> 
> Will it get fixed?
> 
> 
> --
> David Marchand

Hi,

Send patch https://patchwork.dpdk.org/patch/87349/ to disable avx512 on windows as the workaround.
Pallavi also tested with this patch, and passed the mingw and clang build. Will try to see if has better way to fix.