Hi Li,
Thank for details, we may keep current method.
Regards,
Min
At 2025-05-20 17:42:58, "Li Zhang" <[email protected]> wrote:
Hi Chen,
Thanks for the comment.
LDP+STP is recommended in optimization guide for the memory copy loops.
Older compilers sometimes struggle to generate optimal code from the
vld1q_<x>_x2 intrinsics.
Using 2 vld1q_<x> is most likely to get most compilers to generate something
optimal (LDP + STP).
Regards,
Li
From: chen <[email protected]>
Date: Tuesday, 2025. May 20. at 5:18
To: Development for x265 <[email protected]>
Cc: nd <[email protected]>, Li Zhang <[email protected]>
Subject: Re:[x265] [PATCH 0/8] AArch64: Clean up and optimize block copy
primitives
Hi Li,
Thank for the improve patches.
It looks good to me, just a little comment below
In the most function,
+ int16x8_t a0 = vld1q_s16(src + w + 0); + int16x8_t a1 = vld1q_s16(src + w +
8);
How about performance compare to vld1q_s16_x2 ?
Regards,
Chen
At 2025-05-20 00:41:39, "Li Zhang" <[email protected]> wrote:
>Hello,
>
>This patch series optimizes and implements several AArch64 block copy
>primitives using Neon intrinsics. It also cleans up and removes the Neon
>and SVE assembly implementations that are either slower or offer no
>performance benefit.
>
>Many thanks,
>Li
>
>Li Zhang (8):
> AArch64: Optimize blockcopy_pp_neon intrinsics implementation
> AArch64: Optimize blockcopy_ps Neon intrinsics implementation
> AArch64: Implement blockcopy_ss primitives using Neon intrinsics
> AArch64: Implement blockcopy_sp primitives using Neon intrinsics
> AArch64: Optimize cpy1Dto2D_shl Neon intrinsics implementation
> AArch64: Optimize cpy2Dto1D_shl Neon intrinsics implementation
> AArch64: Implement cpy2Dto1D_shr using Neon intrinsics
> AArch64: Implement cpy1Dto2D_shr using Neon intrinsics
>
> source/common/CMakeLists.txt | 2 +-
> source/common/aarch64/asm-primitives.cpp | 180 ---
> source/common/aarch64/blockcopy8-common.S | 54 -
> source/common/aarch64/blockcopy8-sve.S | 1346 ---------------------
> source/common/aarch64/blockcopy8.S | 1049 ----------------
> source/common/aarch64/pixel-prim.cpp | 358 +++++-
> 6 files changed, 305 insertions(+), 2684 deletions(-)
> delete mode 100644 source/common/aarch64/blockcopy8-common.S
>
>--
>2.39.5 (Apple Git-154)
>
>_______________________________________________
>x265-devel mailing list
>[email protected]
>https://mailman.videolan.org/listinfo/x265-devel
_______________________________________________
x265-devel mailing list
[email protected]
https://mailman.videolan.org/listinfo/x265-devel
