tree 12e503276357c92e35a185b852e714c71170737a
parent af08a7c817dbfafbd6872670e2c792d4732319f7
author Yuwei Sun <49397791+YuWei-CH@users.noreply.github.com> 1782980196 -0700
committer GitHub <noreply@github.com> 1782980196 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqRh5kCRC1aQ7uu5UhlAAAlJoQAFKGw7nX+PbIh1T6hefE3oTa
 bHkD6yFVmp8bI+9osknj6XmCDZGq50PqvyxV3mU6lNi6TAKGLBvny4XuU6XRQrb9
 bt0zbMWEHrBNl8jtW8C+gvpaHtJBB7qeLUa/pJwT28poArfXkXz3HPPE/YHRvVqg
 zfUInZIIn0fK2uD/wCezr2xGCuX15FvxxfL2WESibNPfn9011U30+0JTLuT82DgF
 z4ToJslrXSZ1ilG1j7U/V8cDgplMVZGtVq/45vIX+DOAE2OT0FldzYa2qAZJJQuG
 9u2VYYjvPb1wCU45KXhYN/p4RM969vIxLcJGAiXSqZ1ohSmvzs7AA37tAXYWp5dU
 yobeNDYyWHzBcN+vbNH7I13G49AShMK/o9yfj2gu+F/sRPUOIUG6s039A/VoKaAL
 yqV2Pr1uOqyhsncHI0qm9VuqkBWyhK9f5zNEjaG/P8y39+x+qVVe94PnHMXcCD4V
 9BSoBDeng80klXWKCT0056PIqospo/gWESadzoAIH6lyCoWLm7pypf3maFoYfhN/
 YVJ27Rq9Ntqj1wOJ50uSICxnxL9tA0oo/aJy7+J2JL1ewgeN6U20R9jRpVr6ul4/
 JPy6b0H2ewPKnsi2ZtNZbNJ4mnIgnNgbxtLgf1xqCSeKktphp4pN7jp7u1DxLsRA
 mWgquounzu+gwNUAZPaH
 =6yHJ
 -----END PGP SIGNATURE-----
 

[StableHLO] Guard convolution widen operand fusion (#24424)

The motivation for this PR is to keep StableHLO preprocessing valid for
mixed-precision models that contain bf16/f32 convolutions. IREE
currently folds widening converts into dense ops to reduce explicit
casts, but doing this for stablehlo.convolution can create invalid
operand type combinations that StableHLO type inference rejects. This PR
avoid folding widening converts into stablehlo.convolution when doing so
would make the lhs and rhs element types invalid for StableHLO type
inference.

---------

Signed-off-by: Hao Ren <rhao8608@gmail.com>
Co-authored-by: Hao Ren <rhao8608@gmail.com>