cornell-zhang/heterocl

Convert a NumPy array to a HeteroCL array of bitwidth larger than 64

Opened this issue · 1 comments

zhenkun commented 
A_ = hcl.asarray(np.zeros((10,)), dtype=hcl.Int(128))

will result in:

TypeError: data type 'i16' not understood

numpy seems to support large integers:

>>> a = numpy.array([2**64 * 10])
>>> a
array([184467440737095516160], dtype=object)
>>> a.dtype
dtype('O')
seanlatias commented

Following are some pointers:

  1. In this file, it shows how NumPy arrays interact with the HeteroCL/TVM arrays.
    https://github.com/cornell-zhang/heterocl/blob/master/python/heterocl/tvm/_ffi/ndarray.py

  2. More specifically, this function loads NumPy arrays into HeteroCL/TVM.

    heterocl/python/heterocl/tvm/_ffi/ndarray.py

    Line 157 in 83a4c11

    def copyfrom(self, source_array):

  3. And this function loads HeteroCL/TVM array into NumPy arrays.

    heterocl/python/heterocl/tvm/_ffi/ndarray.py

    Line 225 in 83a4c11

    def asnumpy(self):

  4. This is where we handle the array data copy.

    heterocl/tvm/src/runtime/c_runtime_api.cc

    Line 415 in 83a4c11

    int TVMArrayCopyFromBytes(TVMArrayHandle handle, void* data, size_t nbytes) {

So there'd be no problem if we just want to copy the NumPy data into HeteroCL (though we need to know the size to copy first, for sure). The problem is, how do we decode that at the LLVM level.