[dpdk-dev] [PATCH v3 1/6] eal/arm: add 64-bit armv8 version of rte_memcpy.h

[Jerin Jacob]

On Mon, Nov 02, 2015 at 12:22:47PM +0000, Hunt, David wrote:
> On 02/11/2015 04:57, Jerin Jacob wrote:
> >On Fri, Oct 30, 2015 at 01:49:14PM +0000, David Hunt wrote:
> >>Signed-off-by: David Hunt <david.hunt at intel.com>
> --snip--
> >>+#ifndef _RTE_MEMCPY_ARM_64_H_
> >>+#define _RTE_MEMCPY_ARM_64_H_
> >>+
> >>+#include <stdint.h>
> >>+#include <string.h>
> >>+
> >>+#ifdef __cplusplus
> >>+extern "C" {
> >>+#endif
> >>+
> >>+#include "generic/rte_memcpy.h"
> >>+
> >>+#ifdef __ARM_NEON_FP
> >
> >SIMD is not optional in armv8 spec.So every armv8 machine will have
> >SIMD instruction unlike armv7.More over LDP/STP instruction is
> >not part of SIMD.So this check is not required or it can
> >be replaced with a check that select memcpy from either libc or this specific
> >implementation
> 
> Jerin,
>    I've just benchmarked the libc version against the hand-coded version of
> the memcpy routines, and the libc wins in most cases. This code was just an
> initial attempt at optimising the memccpy's, so I feel that with the current
> benchmark results, it would better just to remove the assembly versions, and
> use the libc version for the initial release on ARMv8.
> Then, in the future, the ARMv8 experts are free to submit an optimised
> version as a patch in the future. Does that sound reasonable to you?

Make sense. Based on my understanding, other blocks are also not optimized 
for arm64.
So better to revert back to CONFIG_RTE_FORCE_INTRINSICS and
libc for initial version.

BTW: I just tested ./arm64-armv8a-linuxapp-gcc/app/test and
"byteorder_autotest" is broken. I think existing arm64 code is not optimized
beyond CONFIG_RTE_FORCE_INTRINSICS. So better to use verified
CONFIG_RTE_FORCE_INTRINSICS scheme.

if you guys are OK with arm and arm64 as two different platform then
I can summit the complete working patch for arm64.(as in my current source
code "arm64" is a different platform(lib/librte_eal/common/include/arch/arm64/)


> Rgds,
> Dave.
> 
> 
> --snip--
> 
> 
>