search

intel / torch-xpu-ops

Apache License 2.0
25 stars 18 forks source link

need tf32 support in convolution #941

Open daisyden opened 1 week ago

daisyden commented 1 week ago

🚀 The feature, motivation and pitch

ipex convolution switched to stock pytorch, because we don't have tf32 support this case got 46% regresion compared to ipex 2.3

RVP resnet50_tf32_train_plain_nhwc High 256 FAIL 1650.43 1672.41 902.86 -46%

Here is a comparison of verbose:

convolution jit:ir forward_training src_f32::blocked:acdb::f0 wei_f32::blocked:acdb::f0 bia_undef::undef::: dst_f32::blocked:acdb::f0 attr-scratchpad:user attr-fpmath:tf32 alg:convolution_direct mb16_ic128oc128_ih28oh28kh3sh1dh0ph1_iw28ow28kw3sw1dw0pw1 0.087158
convolution jit:ir forward_training src_f32::blocked:acdb::f0 wei_f32::blocked:acdb::f0 bia_undef::undef::: dst_f32::blocked:acdb::f0 attr-scratchpad:user alg:convolution_direct mb16_ic128oc128_ih28oh28kh3sh1dh0ph1_iw28ow28kw3sw1dw0pw1 0.196045

Alternatives

No response

Additional context

No response

daisyden commented 1 week ago

see https://jira.devtools.intel.com/browse/PYTORCHDGQ-5209