Re: [PATCH v4 0/3] RISC-V: ifunced memcpy using new kernel hwprobe interface

[Palmer Dabbelt <palmer@rivosinc.com>]

On Thu, 06 Jul 2023 12:29:43 PDT (-0700), Evan Green wrote:
>
> This series illustrates the use of a recently accepted Linux syscall that
> enumerates architectural information about the RISC-V cores the system
> is running on. In this series we expose a small wrapper function around
> the syscall. An ifunc selector for memcpy queries it to see if unaligned
> access is "fast" on this hardware. If it is, it selects a newly provided
> implementation of memcpy that doesn't work hard at aligning the src and
> destination buffers.
>
> I opted to spin the whole series, though it's perfectly safe to take
> just the first two patches for the hwprobe interface and abandon the
> third patch as a separate issue.

Thanks.  Given that it has a meaningful performance increase on the 
T-Head hardware it seems reasonable to take it for the next release.  I 
don't remember if I've looked super closely at the implementation, I'll 
do so before testing and merging it -- certainly not this week, though, 
as the merge window will probably eat all my spare cycles.

The only issue on my end is the assembly memcpy routine, which we were 
generally trying to avoid.  +Jeff, as IIUC the Ventana folks were 
interested in memcpy on fast-misaligned systems.  Do you guys happen to 
have one lying around for the C implementation?  It'd be nice to see if 
we're getting any real performance benefit from the assembly.

> Performance numbers were compared using a small test program [1], run on
> a D1 Nezha board, which supports fast unaligned access. "Fast" here
> means copying unaligned words is faster than copying byte-wise, but
> still slower than copying aligned words. Here's the speed of various
> memcpy()s with the generic implementation:
>
> memcpy size 1 count 1000000 offset 0 took 109564 us
> memcpy size 3 count 1000000 offset 0 took 138425 us
> memcpy size 4 count 1000000 offset 0 took 148374 us
> memcpy size 7 count 1000000 offset 0 took 178433 us
> memcpy size 8 count 1000000 offset 0 took 188430 us
> memcpy size f count 1000000 offset 0 took 266118 us
> memcpy size f count 1000000 offset 1 took 265940 us
> memcpy size f count 1000000 offset 3 took 265934 us
> memcpy size f count 1000000 offset 7 took 266215 us
> memcpy size f count 1000000 offset 8 took 265954 us
> memcpy size f count 1000000 offset 9 took 265886 us
> memcpy size 10 count 1000000 offset 0 took 195308 us
> memcpy size 11 count 1000000 offset 0 took 205161 us
> memcpy size 17 count 1000000 offset 0 took 274376 us
> memcpy size 18 count 1000000 offset 0 took 199188 us
> memcpy size 19 count 1000000 offset 0 took 209258 us
> memcpy size 1f count 1000000 offset 0 took 278263 us
> memcpy size 20 count 1000000 offset 0 took 207364 us
> memcpy size 21 count 1000000 offset 0 took 217143 us
> memcpy size 3f count 1000000 offset 0 took 300023 us
> memcpy size 40 count 1000000 offset 0 took 231063 us
> memcpy size 41 count 1000000 offset 0 took 241259 us
> memcpy size 7c count 100000 offset 0 took 32807 us
> memcpy size 7f count 100000 offset 0 took 36274 us
> memcpy size ff count 100000 offset 0 took 47818 us
> memcpy size ff count 100000 offset 0 took 47932 us
> memcpy size 100 count 100000 offset 0 took 40468 us
> memcpy size 200 count 100000 offset 0 took 64245 us
> memcpy size 27f count 100000 offset 0 took 82549 us
> memcpy size 400 count 100000 offset 0 took 111254 us
> memcpy size 407 count 100000 offset 0 took 119364 us
> memcpy size 800 count 100000 offset 0 took 203899 us
> memcpy size 87f count 100000 offset 0 took 222465 us
> memcpy size 87f count 100000 offset 3 took 222289 us
> memcpy size 1000 count 100000 offset 0 took 388846 us
> memcpy size 1000 count 100000 offset 1 took 468827 us
> memcpy size 1000 count 100000 offset 3 took 397098 us
> memcpy size 1000 count 100000 offset 4 took 397379 us
> memcpy size 1000 count 100000 offset 5 took 397368 us
> memcpy size 1000 count 100000 offset 7 took 396867 us
> memcpy size 1000 count 100000 offset 8 took 389227 us
> memcpy size 1000 count 100000 offset 9 took 395949 us
> memcpy size 3000 count 50000 offset 0 took 674837 us
> memcpy size 3000 count 50000 offset 1 took 676944 us
> memcpy size 3000 count 50000 offset 3 took 679709 us
> memcpy size 3000 count 50000 offset 4 took 680829 us
> memcpy size 3000 count 50000 offset 5 took 678024 us
> memcpy size 3000 count 50000 offset 7 took 681097 us
> memcpy size 3000 count 50000 offset 8 took 670004 us
> memcpy size 3000 count 50000 offset 9 took 674553 us
>
> Here is that same test run with the assembly memcpy() in this series:
> memcpy size 1 count 1000000 offset 0 took 92703 us
> memcpy size 3 count 1000000 offset 0 took 112527 us
> memcpy size 4 count 1000000 offset 0 took 120481 us
> memcpy size 7 count 1000000 offset 0 took 149558 us
> memcpy size 8 count 1000000 offset 0 took 90617 us
> memcpy size f count 1000000 offset 0 took 174373 us
> memcpy size f count 1000000 offset 1 took 178615 us
> memcpy size f count 1000000 offset 3 took 178845 us
> memcpy size f count 1000000 offset 7 took 178636 us
> memcpy size f count 1000000 offset 8 took 174442 us
> memcpy size f count 1000000 offset 9 took 178660 us
> memcpy size 10 count 1000000 offset 0 took 99845 us
> memcpy size 11 count 1000000 offset 0 took 112522 us
> memcpy size 17 count 1000000 offset 0 took 179735 us
> memcpy size 18 count 1000000 offset 0 took 110870 us
> memcpy size 19 count 1000000 offset 0 took 121472 us
> memcpy size 1f count 1000000 offset 0 took 188231 us
> memcpy size 20 count 1000000 offset 0 took 119571 us
> memcpy size 21 count 1000000 offset 0 took 132429 us
> memcpy size 3f count 1000000 offset 0 took 227021 us
> memcpy size 40 count 1000000 offset 0 took 166416 us
> memcpy size 41 count 1000000 offset 0 took 180206 us
> memcpy size 7c count 100000 offset 0 took 28602 us
> memcpy size 7f count 100000 offset 0 took 31676 us
> memcpy size ff count 100000 offset 0 took 39257 us
> memcpy size ff count 100000 offset 0 took 39176 us
> memcpy size 100 count 100000 offset 0 took 21928 us
> memcpy size 200 count 100000 offset 0 took 35814 us
> memcpy size 27f count 100000 offset 0 took 60315 us
> memcpy size 400 count 100000 offset 0 took 63652 us
> memcpy size 407 count 100000 offset 0 took 73160 us
> memcpy size 800 count 100000 offset 0 took 121532 us
> memcpy size 87f count 100000 offset 0 took 147269 us
> memcpy size 87f count 100000 offset 3 took 144744 us
> memcpy size 1000 count 100000 offset 0 took 232057 us
> memcpy size 1000 count 100000 offset 1 took 254319 us
> memcpy size 1000 count 100000 offset 3 took 256973 us
> memcpy size 1000 count 100000 offset 4 took 257655 us
> memcpy size 1000 count 100000 offset 5 took 259456 us
> memcpy size 1000 count 100000 offset 7 took 260849 us
> memcpy size 1000 count 100000 offset 8 took 232347 us
> memcpy size 1000 count 100000 offset 9 took 254330 us
> memcpy size 3000 count 50000 offset 0 took 382376 us
> memcpy size 3000 count 50000 offset 1 took 389872 us
> memcpy size 3000 count 50000 offset 3 took 385310 us
> memcpy size 3000 count 50000 offset 4 took 389748 us
> memcpy size 3000 count 50000 offset 5 took 391707 us
> memcpy size 3000 count 50000 offset 7 took 386778 us
> memcpy size 3000 count 50000 offset 8 took 385691 us
> memcpy size 3000 count 50000 offset 9 took 392030 us
>
> The assembly routine is measurably better.
>
> [1] https://pastebin.com/DRyECNQW
>
>
> Changes in v4:
>  - Remove __USE_GNU (Florian)
>  - __nonnull, __wur, __THROW, and  __fortified_attr_access decorations
>   (Florian)
>  - change long to long int (Florian)
>  - Fix comment formatting (Florian)
>  - Update backup kernel header content copy.
>  - Fix function declaration formatting (Florian)
>  - Changed export versions to 2.38
>  - Fixed comment style (Florian)
>
> Changes in v3:
>  - Update argument types to match v4 kernel interface
>  - Add the "return" to the vsyscall
>  - Fix up vdso arg types to match kernel v4 version
>  - Remove ifdef around INLINE_VSYSCALL (Adhemerval)
>  - Word align dest for large memcpy()s.
>  - Add tags
>  - Remove spurious blank line from sysdeps/riscv/memcpy.c
>
> Changes in v2:
>  - hwprobe.h: Use __has_include and duplicate Linux content to make
>    compilation work when Linux headers are absent (Adhemerval)
>  - hwprobe.h: Put declaration under __USE_GNU (Adhemerval)
>  - Use INLINE_SYSCALL_CALL (Adhemerval)
>  - Update versions
>  - Update UNALIGNED_MASK to match kernel v3 series.
>  - Add vDSO interface
>  - Used _MASK instead of _FAST value itself.
>
> Evan Green (3):
>   riscv: Add Linux hwprobe syscall support
>   riscv: Add hwprobe vdso call support
>   riscv: Add and use alignment-ignorant memcpy
>
>  sysdeps/riscv/memcopy.h                       |  26 ++++
>  sysdeps/riscv/memcpy.c                        |  64 +++++++++
>  sysdeps/riscv/memcpy_noalignment.S            | 121 ++++++++++++++++++
>  sysdeps/unix/sysv/linux/dl-vdso-setup.c       |  10 ++
>  sysdeps/unix/sysv/linux/dl-vdso-setup.h       |   3 +
>  sysdeps/unix/sysv/linux/riscv/Makefile        |   8 +-
>  sysdeps/unix/sysv/linux/riscv/Versions        |   3 +
>  sysdeps/unix/sysv/linux/riscv/hwprobe.c       |  31 +++++
>  .../unix/sysv/linux/riscv/memcpy-generic.c    |  24 ++++
>  .../unix/sysv/linux/riscv/rv32/libc.abilist   |   1 +
>  .../unix/sysv/linux/riscv/rv64/libc.abilist   |   1 +
>  sysdeps/unix/sysv/linux/riscv/sys/hwprobe.h   |  72 +++++++++++
>  sysdeps/unix/sysv/linux/riscv/sysdep.h        |   1 +
>  13 files changed, 363 insertions(+), 2 deletions(-)
>  create mode 100644 sysdeps/riscv/memcopy.h
>  create mode 100644 sysdeps/riscv/memcpy.c
>  create mode 100644 sysdeps/riscv/memcpy_noalignment.S
>  create mode 100644 sysdeps/unix/sysv/linux/riscv/hwprobe.c
>  create mode 100644 sysdeps/unix/sysv/linux/riscv/memcpy-generic.c
>  create mode 100644 sysdeps/unix/sysv/linux/riscv/sys/hwprobe.h