Open ghost opened 5 years ago
@mmaciag As far as i can see, the lenet.param lenet.bin is NCNN format, FYI: https://github.com/Tencent/ncnn
I have the same CSC error when running on VP.
Previously I thought it might be the problem of KMD atom_size = 32
, but then I realized the CSC assertion is still there with the fix to KMD (so I deleted my previous comment).
The generated flatbuffer is hang on my FPGA due to no response from hardware.
So seems like this compiler still needs some work to clean pipe.
@mmaciag sorry for the delay reply. see comment inline:
- Is caffe2fb output compatible with current UMD _nvdlaruntime from nvdla/sw repository? yes, we just use some header files in this repo.
- Is it compatible with current nvdla/hw master, with default _nvlarge specification file? we just run on VP which configure for full not small, and use the default kmd.ko drm.ko and umd.ko you can download the those default modules from https://github.com/nvdla/sw/tree/master/prebuilt/linux. we haven't yet run on our FPGA, because now our FPGA just support small mode.
- What command line is expected to run flatbuffer file with _nvdlaruntime? ./nvdla_runtime --loadable flatbuffer --image your_image_file --rawdump then you will find the result in out file.
- Is lenet.param your custom format? Is lenet.bin as it is created by official Caffe (binaryproto) or is it also some custom format? those two files is generated by tencent ncnn from official caffe, the following link is the tencent ncnn https://github.com/Tencent/ncnn. we just use the ncnn code to parse the official caffe.
@icubecorp Thank you for response
Ok... so I've just reached first issue :)
I would start from some questions:
I ran the virtual platform with compiled CMOD as nv_large, but unfortunatelly it asserted on some CSC operation.
I think I have already seen that
csc_entries_
in completely different place. It happened when I was running trace_tests. In some of them I have modified the CBUF banks allocation. Here is example of file so you can see what I am talking about. I changed the D_BANK_0 registers. This triggered the CMOD assertion too. Then I realized the weight/feature data could not fit selected banks with 'my' allocation.