Hi Matthias, Thanks for the links. I will try a build with the Simd256 type and operations enabled, but without the revec phase, and see how things go.
Cheers, Sam On Friday, October 25, 2024 at 4:38:01 PM UTC+1 [email protected] wrote: > Hi Sam, > > As Wasm only supports Simd128, the Simd256 operations are part of an > experimental (work-in-progress) implementation developed by Intel. > The progress is tracked here: https://crbug.com/42202660 > The build flag is controlled here: > https://source.chromium.org/chromium/chromium/src/+/main:v8/BUILD.gn;l=681;drc=9be9b0197a97a1f661c4cbf7b354c21ec3fcbe7f > When the build flag is set, it can be enabled with the flag > --experimental-wasm-revectorize. > > I don't think the current implementation tries to go further than 256 bit > vector widths. > > Cheers, > Matthias > > On Fri, Oct 25, 2024 at 1:23 PM Sam Parker-Haynes <[email protected]> > wrote: > >> Hi, >> >> I'm looking a lot at shuffles, in turboshaft, and thinking about how we >> can generate de-interleaving loads. For AArch64 we have the LD1, LD2, LD3 >> and LD4 instructions which can load in a maximum of 4x128-bit values. >> >> To simplify my life, I'm currently only considering LD2, which will load >> 2x128-bits, so I figure I the existing Simd256 operations would be good for >> this. But these are all currently predicated >> on V8_ENABLE_WASM_SIMD256_REVEC, do they have to be..? >> >> I only want to support wide vectors for loads, using >> Simd256Extract128Lane to produce usable values for arithmetic, etc... but >> are there any parts of the pipeline that make assumptions when 'revec' is >> enabled that will make life hard? >> >> And when trying to support LD3 (3x128-bit) and LD4 (4x128-bit), would >> this approach still scale? >> >> Thanks, >> Sam >> >> -- >> -- >> v8-dev mailing list >> [email protected] >> http://groups.google.com/group/v8-dev >> --- >> You received this message because you are subscribed to the Google Groups >> "v8-dev" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To view this discussion visit >> https://groups.google.com/d/msgid/v8-dev/5a86a10c-4b95-4dcb-b7b1-efce05865b97n%40googlegroups.com >> >> <https://groups.google.com/d/msgid/v8-dev/5a86a10c-4b95-4dcb-b7b1-efce05865b97n%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> > -- -- v8-dev mailing list [email protected] http://groups.google.com/group/v8-dev --- You received this message because you are subscribed to the Google Groups "v8-dev" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/v8-dev/a522601f-4d24-41ff-8f86-812b6a32f36bn%40googlegroups.com.
