[PATCH] D95705: [RISCV] Add initial support for converting fixed vectors to scalable vectors during lowering to use RVV instructions.
Zakk Chen via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Sun Jan 31 23:31:56 PST 2021
khchen added inline comments.
================
Comment at: llvm/lib/Target/RISCV/RISCVISelLowering.cpp:1703
+ // We use the types that will be mapped to LMUL=1 here.
+ // FIXME: Should we have a way to support larger LMULs?
+ switch (VT.getVectorElementType().SimpleTy) {
----------------
arcbbb wrote:
> craig.topper wrote:
> > arcbbb wrote:
> > > For supporting LMUL > 1,
> > > I was thinking calculating LMUL by
> > > ```
> > > unsigned MinVLen = Subtarget.getMinRVVVectorSizeInBits();
> > > unsigned LMul = (VT.getSizeInBits() + MinVLen - 1) / MinVLen;
> > > ```
> > > and expand the `case MVT::i8:` by
> > > ```
> > > case MVT::i8:
> > > if (LMul == 1)
> > > return EVT(MVT::nxv8i8);
> > > else if (LMul == 2)
> > > return EVT(MVT::nxv16i8);
> > > else if (LMul <= 4)
> > > return EVT(MVT::nxv32i8);
> > > else if (LMul <= 8)
> > > return EVT(MVT::nxv64i8);
> > > ```
> > I guess my concern is that large LMULs increase register pressure so depending on the code it might be better to split the operations than use the increased LMUL if it will cause spills.
> I have a use case for [1]
> ```
> blur_x.store_at(blur_y, y).compute_at(blur_y, yi).vectorize(x, 8);
> ```
> User can evaluate various VLS choices by changing the number of element in `.vectorize()` and evaluate which LMul gains most performance.
> not sure if it makes sense.
> [1] https://github.com/halide/Halide/blob/master/apps/blur/halide_blur_generator.cpp
I think maybe the vectorizer or users(Halide) need to aware the register pressure (or performance) when they perform transformation. If codegen want to help on avoiding register pressure problem, maybe backend could have an IR pass to find out the best LMUL for scalabe vector type operation?
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D95705/new/
https://reviews.llvm.org/D95705
More information about the llvm-commits
mailing list