Frédéric Bastien

[XLA:GPU] HorizontalLoopFusion now generates a cleaner graph

2

This ask less works by other optimization pass. This is a safe subset of another PR. In case of revert, it will revert less changes. The main changes are: -...

I’m making this issue here in DLPack, as I do not think of a better place for this. This issue is between many software and as this is the goal...

secure_dump should predump in the same folder as the destination

2

https://github.com/mila-udem/blocks/blob/master/blocks/serialization.py#L226 This do a dump to a temp file and then move it to the destination. This is safer in case of a crash during the dump. But it cause...

[XLA] When dumping the html files, also dump a version without the fusion details.

7

It allows to not dump the fusion internal. This WAR the issue that "too big" HTML files aren't rendered by the browser. We can manually look at the txt hlo...

JIT+Shardmap+donate_argnums doesn't work, while pmap+donate_argnums works.

4

If I use: ``` jax.pmap(f_pmap, in_axes=0, out_axes=0, axis_name='x', donate_argnums=0) ``` It works. But when this code is ported to jax.array + shardmap to support multi-node, this run, but warn that...

[DRAFT] Custom gpu ops custom partitioning

1

This in progress PR modify the docs/Custom_Operation_for_GPUs.py tutorial to use custom_partitioning instead of xmap. Don't review now, there is still much work done. - [x] finish the forward code. -...

If we add 3 tensor and 2 of them are c contiguous but not the last, it seam wasteful to compute the index 2 times for each c contiguous array.

We could use PyOpenCL w http://documen.tician.de/pyopencl/array.html#complex-numbers In addition, I've added a rudimentary facility for translating Fortran kernels to OpenCL, see here: https://github.com/inducer/pyopencl/tree/master/contrib/fortran-to-opencl