Polybench Summary

[calebmkim]

Summary

30 Benchmarks:

• 2 don't work for an annoying reason (we don't support arith.select, and we don't support scf.if. arith.select should be pretty easy, I imagine. scf.if is a bit more difficult, because scf.if statement support yield statements)
• 9 don't work for more fundamental reasons. Either a) arith.mlir doesn't have an sqrt operator, or because they multiply by things like 0.33, which will just round to multiply by 0 in integers.

Among the 19 benchmarks that work:

• 9 can be fully static
• 10 can't, and only have "static islands"

(Edit): Summary of the three categories.

1) Don't Work: correlation, cholesky, gramschmidt, deriche, floyd-warshall, nussinov, adi, heat-3d, jacobi-1d, jacobi-2d, fdtd-2d, 2) Work Island: covariance, symm, syr2k, syrk, trmm, durbing, lu, ludcmp, seidel-2d, trisolv, 3) Work Fully Static: gemm, gemver, gesummv, 2mm, 3mm, atax, bicg, doitgen, mvt,

Table

Design	Works?	Fully Static?
correlation	N (sqrt)	n/a
covariance	Y	N
gemm	Y	Y
gemver	Y	Y
gesummv	Y	Y
symm	Y	N
syr2k	Y	N
syrk	Y	N
trmm	Y	N
2mm	Y	Y
3mm	Y	Y
atax	Y	Y
bicg	Y	Y
doitgen	Y	Y
mvt	Y	Y
cholesky	N (sqrt)	n/a
durbin	Y	N
gramschmidt	N (sqrt)	n/a
lu	Y	N
ludcmp	Y	N
trisolv	Y	N
deriche	N (exp)	n/a
floyd-warshall	N(annoything arith.select)	n/a
nussinov	N (annoying things not supported arith.select, scf.if)	n/a
adi	N (divide by 0 b/c we use int)	n/a
fdtd-2d	N (computation not interesting, mult by 0.5)	n/a
heat-3d	N (computation not interesting (multiplies by 0.125 which rounds to 0))	n/a
jacobi-1d	N (computation not interesting, mult by 0.33)	n/a
jacobi-2d	N (computation not interesting, mult by 0.2)	n/a
seidel-2d	Y	N

[rachitnigam]

Awesome, thanks for compiling this list! I think we have a pretty solid story to tell with this set of benchmarks already! Of course, in the future, it would be nice to extend the scf conversion so it can support the other benchmarks as well. I bet the ones that use scf.yield probably also cannot be fully static.

[calebmkim]

Just an update: there were a couple benchmarks that I miscategorized, I've updated the original comment. But this doensn't change anything about the story though.

[rachitnigam]

I've added support for arith.select so we should see if the remaining benchmarks work: https://github.com/llvm/circt/pull/5857