[VTA] Does VTA currently work with 1-bit weights?

Questions
1

Hey everyone,
I am currently playing with VTA on the ZedBoard and it is so far working very well.

Now, I have a problem with 1-bit weights. I use the default PYNQ configuration with LOG_ACC_BUFF_SIZE changed to 15 and LOG_WGT_WIDTH to 0.
With this configuration, I get the following error:
include/tvm/expr.h:62: Check failed: data_bits % 8 == 0U (1 vs. 0) Need to load/store by multiple of bytes
when I run the matrix_multiply.py from the VTA tutorial.

I do not understand the error message because to my understanding, the VTA DMA Unit should be capable of loading 1-bit weights. It always loads multiple weights in one go (vta.cc:112) [which should be a multiple of 8-bit in my example]. Is that right?
I have no idea how I need to change the matrix_multiply code to accomplish the 1-bit weight dma loading.
Thank you very much for any hints.

2

I’m no expert in the matter, but it seems that the problem isn’t really VTA (based on the path of the file that us throwing the error.

I took a look at the source code which I paste here:

// Get number of bytes considering vector type.
inline int GetVectorBytes(Type dtype) {
  int data_bits = dtype.bits() * dtype.lanes();
  // allow bool to exist
  if (dtype == Bool()) return 1;
  CHECK_EQ(data_bits % 8, 0U)
      << "Need to load/store by multiple of bytes"; //this is line 62
  return data_bits / 8;
}

So out of the top of my head, it seems that you have a problem with your definition of dtype. If its 1bit, you need to define it as Bool and at least the if (dtype == Bool()) return 1; should return a correct value.
Don’t know if there will be problems down the line since I haven’t seen a 1bit VTA implementation working.