dlsyscourse / hw3 Goto Github PK

Q5 has following line that runs local tests: !python3 -m pytest -v -k "matmul and cpu".

However this doesn't run test_matmul_tiled test.
To run this test one must add another notebook cell with command !python3 -m pytest -v -k "matmul_tiled"

I guess it's better to update existing line with the following !python3 -m pytest -v -k "matmul and cpu or matmul_tiled"

Steps to reproduce:

!python3 -m pytest -v -k "(permute or reshape or broadcast or getitem) and cpu and not compact"

Traceback:
../../../opt/anaconda3/envs/dlsys/lib/python3.8/importlib/init.py:127: in import_module
return _bootstrap._gcd_import(name[level:], package, level)
tests/test_ndarray.py:6: in
import needle as ndl
python/needle/init.py:1: in
from . import ops
python/needle/ops.py:5: in
from .autograd import NDArray
python/needle/autograd.py:12: in
from .backend_selection import Device, array_api, NDArray, default_device
python/needle/backend_selection.py:9: in
from . import backend_ndarray as array_api
python/needle/backend_ndarray/init.py:1: in
from .ndarray import *
python/needle/backend_ndarray/ndarray.py:6: in
from . import ndarray_backend_cpu
E ImportError: cannot import name 'ndarray_backend_cpu' from partially initialized module 'needle.backend_ndarray' (most likely due to a circular import) (/Users/talurj/Documents/dlsys/hw3/python/needle/backend_ndarray/init.py)

In hw3/python/needle/backend_ndarray/ndarray.py, Row 299, there is code like this , "start = self.shape[dim]".

In my opinion, it should be "start = self.shape[dim] + start"

For people who are not super familiar with pybind or cmake, if you want to generate the debug information for the shared library, you will need to comment the line pybind11_strip(ndarray_backend_cpu) and then compile it with regular cmake debug flag. Otherwise, the debug info will get stripped and you could not use debugging tools like (gdb, valgrind, etc.).

In src/ ndarray_bakcend_cuda.cu file, The comment of the parameter 'strides' and 'offset' with the function CompactKernel and Compact is not consistency.

CompactKernel:

*   strides: vector of strides of out array
*   offset: offset of out array

Compact:

* strides: strides of the *a* array (not out, which has compact strides)
*  offset: offset of the *a* array (not out, which has zero offset, being compact)

I think the parameters strides and offsets all belong to the input array "a".

This should be (a.shape[1] * tile, tile, a.shape[1], 1),)