matmul produce wrong (empty) results on ROCm builds #294
Description
Using the master branch, the matmul function gives uninitialized(?) results. It is relatively easy to reproduce with following script
import numpy
import torch
from torch_sparse.matmul import matmul
from torch_sparse.tensor import SparseTensor
def main():
src = torch.from_numpy(numpy.ones((3,3)).astype("f4")).to("cuda:0")
other = torch.from_numpy(numpy.ones((3,3)).astype("f4")).to("cuda:0")
src = SparseTensor.from_dense(src)
out = matmul(src, other)
print(out)
main()
out will be all 0, or nan with reduce == sum, or maximum number initialized with min.
I am actually not sure if it is a bug rooted in pytorch_sparse itself or ROCm, as unit tests are all passed in other related libs (e.g. pytorch_scatter). This is especially interesting as if kernel codes are compiled with -O2, out is filled with 0, if compiled with -O1, out is filled with nan. If compiled with -O0, the program will simply hang due to some error reported from amdgpu.
Any debugging instruction would be helpful.