tree d75d816e7751236473295a1cb23868dba1e0a1c4
parent ee167be15f4c532d5e834c53db1938db3fd5a3fe
author Nicolas Vasilache <nicolasvasilache@users.noreply.github.com> 1671014677 +0100
committer GitHub <noreply@github.com> 1671014677 +0100
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJjmakVCRBK7hj4Ov3rIwAA4ecIADhx8WRdip0F7MXvcsd2w3d4
 PrlZAZE2hh3qGvii+6k3K583nfzh+oe6Ow5BwhxpUw+Db46PadZHbNh+2QPUxN7x
 flAd4tnszxrxqMAKOOfVJMUV3X9VMTyZw2+ojx7bRyrH+Lx91+WpWSxLRlk1tZpv
 JYkgh2k+77IPWRLaeb2dn2mvGFnn5YX3u1torxx7pEW3oAtmMYbEb+lNYgvb8IUQ
 F/r5+uH6wGrrzrvm2MaDhpATIflXA7npXL9jQtvhd3I/Xl3YMAw9bI9TecC58Mbs
 qsWYlwnvS8wYBE8Glf6V7Ze+nc3WaNaYoyrzjgwWZbpsFYCtVVB5qlptLkGvHzI=
 =cYcH
 -----END PGP SIGNATURE-----
 

Emit better reduction schedule from JIT (#11548)

Emit the reduction schedule splitting the reduction directly to
"scf.for" rather than to a generic over inputs with expanded shape. This
allows us to ensure the generation of vector loads and stores, and to
target the vector size appropriate to the input. Also plug in the
matcher into the schedule instead of the fragile reliance on the order
of operations.
    
This is a rebase of #11498 on main.

Co-authored-by: Oleksandr "Alex" Zinenko <zinenko@google.com>