AndrewZhaoLuo opened a new issue #8340:
URL: https://github.com/apache/tvm/issues/8340


   Some operations like conv2d and matmul support heterogenous output 
datatypes. 
   
   E.g. the operation could be done with float16 operands but accumulate the 
result in float32. 
   
   However, some schedules for these ops don't actually support mixed precision 
types. For example `python/tvm/topi/cuda/conv2d_winograd.py` has a type error 
if the output dtype is not the same as the input dtypes.
   
   This issue is to audit operations which support output datatypes (e.g. have 
the `out_dtype` attribute) and make sure corresponding schedules support the 
feature. For every schedule this will involve adding a unittest.
   
   List of ops to examine
   - [ ] "nn.conv1d",
   - [ ] "nn.conv2d",
   - [ ] "nn.conv3d",
   - [ ] "nn.conv1d_transpose",
   - [ ] "nn.conv2d_transpose",
   - [ ] "nn.conv3d_transpose",
   - [ ] "nn.dense",
   - [ ] "nn.batch_matmul"


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to