Closed curiousdragon closed 1 year ago
Update: I actually ended up finding and fixing a few dimensional bugs, which seemed to be the source of the errors in
conv_dw-rect-baremetal
andconv_trans_input_3120_with_kernel_dilation-rect-baremetal
so these two tests are now passing!
Now, the only rectangular convolution test that is failing is conv_with_pool-rect-baremetal
, and as suggested, I've added an if
statement to exit with an error message and fail if someone tries to run a rectangular convolution with pooling - they would get something along the lines of:
Pooling with rectangular convolutions is currently not supported.
*** FAILED *** (tohost = 1)
I've also added this error message and exit statement to the commented out pooling logic in sp_tiled_conv
and sp_tiled_conv_dw
just in case, since I don't think the gemmini_extended2_config_st
function currently supports differing pooling output dimensions which rectangular convolutions probably require.
Remaining todos (edit: done in #28):
Thanks for the revisions!
Note: Closing this PR since it's based on a branch in a fork, and opening up a new PR (#28) that has an updated branch on this repository specifically for review + merging! If I remember correctly, I think the main reasoning for wanting to have a non-forked development branch was that now with #28, I can make a PR for the gemmini
repository with an updated submodule reference that isn't to a separate fork of this gemmini-rocc-tests
repository.
./build.sh
ingemmini-rocc-tests
, and will end with-rect-baremetal
conv_dw-rect-baremetal
conv_trans_input_3120_with_kernel_dilation-rect-baremetal
conv_with_pool-rect-baremetal
gemmini_extended2_config_st
, called insp_tiled_conv
, since now thepool_out_dim
has been split intopool_out_row_dim
andpool_out_col_dim
, butgemmini_extended2_config_st
seems to be a ROCC instruction so it's unclear to me how exactly to proceed. (Or the bug might be somewhere else, I have yet to investigate further)