+1, pls add the typedef...
Thanks… Dave
-----Original Message-----
From: Damjan Marion (damarion)
Sent: Saturday, August 19, 2017 9:09 AM
To: Dave Barach (dbarach) <dbar...@cisco.com>
Cc: George Zhao <george.y.z...@huawei.com>; vpp-dev@lists.fd.io;
disc...@lists.fd.io; csit-...@lists.fd.io
Subject: Re: [discuss] Question about VPP support for ARM 64
GCC is able to compile ARM64 code with 256-bit vectors even if target platform
have only 128-bit registers.
I.e. for the u8x32 version of that function it generates:
ARM64:
dpdk_buffer_init_from_template(void*, void*, void*, void*, void*):
ld1 {v0.16b - v1.16b}, [x4], 32
st1 {v0.16b - v1.16b}, [x3], 32
st1 {v0.16b - v1.16b}, [x2], 32
st1 {v0.16b - v1.16b}, [x1], 32
st1 {v0.16b - v1.16b}, [x0], 32
ld1 {v0.16b - v1.16b}, [x4]
st1 {v0.16b - v1.16b}, [x3]
st1 {v0.16b - v1.16b}, [x2]
st1 {v0.16b - v1.16b}, [x1]
st1 {v0.16b - v1.16b}, [x0]
ret
intel x86-64 without AVX2:
dpdk_buffer_init_from_template(void*, void*, void*, void*, void*):
push %rbp
mov %rsp,%rbp
and $0xffffffffffffffe0,%rsp
lea 0x10(%rsp),%rsp
movdqa (%r8),%xmm1
movdqa 0x10(%r8),%xmm0
movdqa %xmm0,0x10(%rcx)
movdqa %xmm1,(%rcx)
movdqa %xmm1,(%rdx)
movdqa %xmm0,0x10(%rdx)
movdqa %xmm1,(%rsi)
movdqa %xmm0,0x10(%rsi)
movdqa %xmm1,(%rdi)
movdqa %xmm0,0x10(%rdi)
movdqa 0x20(%r8),%xmm1
movdqa 0x30(%r8),%xmm0
movdqa %xmm0,0x30(%rcx)
movdqa %xmm1,0x20(%rcx)
movdqa %xmm1,0x20(%rdx)
movdqa %xmm0,0x30(%rdx)
movdqa %xmm1,0x20(%rsi)
movdqa %xmm0,0x30(%rsi)
movdqa %xmm1,0x20(%rdi)
movdqa %xmm0,0x30(%rdi)
leaveq
retq
So i think here it is only about missing typedef….
> On 19 Aug 2017, at 14:51, Dave Barach (dbarach) <dbar...@cisco.com> wrote:
>
> Dear George,
>
> This specific issue isn’t anywhere near as bad as you might think. As given,
> the code confuses 128-bit vectors with 256-bit vectors, and 64-bit vectors
> with 128-bit vectors.
>
> Question: does the hardware involved support 256-bit vectors? Probably not...
> It almost certainly does support 128-bit vectors.
>
> To make progress, use the known-good u8x16 / 128-bit vector code:
>
> static_always_inline void
> dpdk_buffer_init_from_template (void *d0, void *d1, void *d2, void *d3,
> void *s)
> {
> #if defined(CLIB_HAVE_VEC128)
> int i;
> for (i = 0; i < 4; i++)
> {
> *(u8x16 *) (((u8 *) d0) + i * 16) =
> *(u8x16 *) (((u8 *) d1) + i * 16) =
> *(u8x16 *) (((u8 *) d2) + i * 16) =
> *(u8x16 *) (((u8 *) d3) + i * 16) = *(u8x16 *) (((u8 *) s) + i * 16);
> }
> #else
> #error "CLIB_HAVE_VEC128 has to be defined"
> #endif
> }
>
> Responsible parties - they know who they are - will be back from PTO shortly.
> We need to clean up / create CLIB_HAVE_VEC_256 and move the 256-bit vector
> engine code...
>
> You could also try adding “typedef u8 u8x32 _vector_size(32)” but I somehow
> doubt that will produce anything other than a compiler error.
>
> HTH… Dave
>
> From: vpp-dev-boun...@lists.fd.io [mailto:vpp-dev-boun...@lists.fd.io] On
> Behalf Of George Zhao
> Sent: Friday, August 18, 2017 7:32 PM
> To: 'vpp-dev@lists.fd.io' <vpp-dev@lists.fd.io>; 'disc...@lists.fd.io'
> <disc...@lists.fd.io>; 'csit-...@lists.fd.io' <csit-...@lists.fd.io>
> Subject: [vpp-dev] Question about VPP support for ARM 64
>
> We encounter following issues while trying to build VPP over ARM 64. It seems
> right now only ARM32 are supported in the code. I list the steps we tried and
> hope VPP folks can help us work around this issue.
>
> Steps:
> 1. install Ubuntu 16.04 on OD1K
> $>> uname -a
> Linux OD1K 4.4.0-92-generic #115-Ubuntu SMP Thu Aug 10 09:10:33 UTC 2017
> aarch64 aarch64 aarch64 GNU/Linux
>
> 2. git clone VPP 17.04 and build VPP
> ## Error:
> make[2]: Entering directory '/home/huawei/GIT/vpp.1704/dpdk'
> cat: '/sys/bus/pci/devices/0000:00:01.0/uevent': No such file or directory
>
> **Work around to bypass MakeFile:
> ##############################################################################
> # Cavium ThunderX
> ##############################################################################
> #else ifneq (,$(findstring thunder,$(shell cat
> /sys/bus/pci/devices/0000:00:01.0/uevent | grep cavium)))
> else
> export CROSS=""
> DPDK_TARGET ?= arm64-thunderx-linuxapp-$(DPDK_CC)
> DPDK_MACHINE ?= thunderx
> DPDK_TUNE ?= generic
>
> 3. Then, make build and failed following:
> /home/huawei/GIT/vpp.1704/build-data/../src/plugins/dpdk/device/node.c:276:9:
> error: `u8x32' undeclared (first use in this function)
> *(u8x32 *) (((u8 *) d0) + i * 32) =
>
> ** Check vppinfra/vppinfra/vector.h and don’t find u8x32 with “aarch64”
> #if defined (__aarch64__) || defined (__arm__)
> typedef unsigned int u32x4 _vector_size (16);
> typedef u8 u8x16 _vector_size (16);
> typedef u16 u16x8 _vector_size (16);
> typedef u32 u32x4 _vector_size (16);
> typedef u64 u64x2 _vector_size (16);
> #endif
>
> 4. According https://wiki.fd.io/view/VPP/Alternative_builds
> The VPP seems to support arm32 only .
> export PLATFORM=arm32
>
>
> *Questions:
> Did I miss some steps or should include other header files that defines u8x32?
>
>
> Thanks,
> George
>
> _______________________________________________
> discuss mailing list
> disc...@lists.fd.io
> https://lists.fd.io/mailman/listinfo/discuss
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
