Re: Cost of handling misaligned access

On Mon, 3 Feb 2025 21:40:21 +0000, BGB wrote:

On 2/3/2025 1:41 PM, Thomas Koenig wrote:

EricP <ThatWouldBeTelling@thevillage.com> schrieb:
>

That is fine for code that is being actively maintained and backward
data structure compatibility is not required (like those inside a
kernel).
>
However for x86 there was a few billion lines of legacy code that likely
assumed 2-byte alignment, or followed the fp64 aligned to 32-bits
advice,
and a C language that mandates structs be laid out in memory exactly as
specified (no automatic struct optimization). Also I seem to recall some
amount of squawking about SIMD when it required naturally aligned
buffers.
As SIMD no longer requires alignment, presumably code no longer does so.

>
Looking at Intel's optimization manual, they state in
"15.6 DATA ALIGNMENT FOR INTEL® AVX"
>
"Assembly/Compiler Coding Rule 65. (H impact, M generality) Align
data to 32-byte boundary when possible. Prefer store alignment
over load alignment."
>
and further down, about AVX-512,
>
"18.23.1 Align Data to 64 Bytes"
>
"Aligning data to vector length is recommended. For best results,
when using Intel AVX-512 instructions, align data to 64 bytes.
>
When doing a 64-byte Intel AVX-512 unaligned load/store, every
load/store is a cache-line split, since the cache-line is 64
bytes. This is double the cache line split rate of Intel AVX2
code that uses 32-byte registers. A high cache-line split rate in
memory-intensive code can cause poor performance."
>
This sounds reasonable, and good advice if you want to go
down SIMD lane.
>

>
This is, ironically, a place where SIMD via ganged registers has an
advantage over SIMD via large monolithic registers.

Iroincally^2 is that vVM allows each implementation to decide on how
many and how wide the SIMD register are. LBI/O might have 8 128-bit
flip-flops, while GBOoO might have 32 512-bit flip-flops. All running
the same binary and all running that same binary as fast as any binary
that that machine could run; in addition, HW looks at the loop index
and possibly predication) to create masks on the lanes of execution.
SW ASCII should describe the calculations to be performed,
The compiler should produce a vVM loop for those calculations
in loops.
Any implementation should run that vVM loop as fast as it can.
See, no change to ISA and you still get 98^ of SIMD you know and love
and the number and width of the registers is an implementation variable!