xla
cac652c3 - Boxed cpu fallback (#2945)

Commit
4 years ago
Boxed cpu fallback (#2945) * use in-tree public codegen API. * use in-tree public codegen API. * stop pinning to feature branch * remove torch_pin file before merging. preemptively update yaml to incorporate #2891 * use in-tree public codegen API. * stop pinning to feature branch * [WIP] remove bridge:: api from codegen. Start auto-gen'ing out/inplace kernels when possible * updated torch pin * fix missing ops in yaml * remove torch_pin file before merging. preemptively update yaml to incorporate #2891 * make codegen backend-agnostic, except for CPU fallbacks * remove logging * rename aten_xla_type.h -> XLANativeFunctions.h * rename aten_xla_type.h -> XLANativeFunctions.h * fix rename, delete aten_xla_type.h * update torch_pin after the file rename * remove logging info from yaml * make codegen backend-agnostic, except for CPU fallbacks * rename aten_xla_type.h -> XLANativeFunctions.h * fix rename, delete aten_xla_type.h * use codegen'd inplace kernels, and delete manually written inplace kernels * fix test change * lint * make codegen backend-agnostic, except for CPU fallbacks * rename aten_xla_type.h -> XLANativeFunctions.h * use codegen'd inplace kernels, and delete manually written inplace kernels * [WIP] remove bridge:: api from codegen. Start auto-gen'ing out/inplace kernels when possible * use a boxed CPU fallback kernel * remove AtenXlaType from straggler ops * fix failing tests * formatting + test fix * fix overload naming for fallback, made calling it require a bit less boilerplate * fix name overloads for boxed fallback * reference the boxed fallback in aten_cpu_fallback.h instead of aten_xla_type_default.h * fix accidental diff * revert accidental yaml change * remove accidentally commited file * update boxed fallback API usage * remove accidentally committed file * lint * fix issue with calling XLA_COUNTER macro from boxed fallback * lint * better comment * minor pr feedback * add boxed fallback call to 2 new ops * remove to_cpu() from yaml, _to_cpu is the right one * remove torch pin
Author
Parents
Loading