{"author":"qazal","author_email":"77887910+Qazalin@users.noreply.github.com","author_time":1702674872,"commit_time":1702674872,"committer":"GitHub","committer_email":"noreply@github.com","hash":"f3ed96a92953ead19636e32441362989a6322a50","message":"Use the reduceop dtype to define the acc in linearizer (#2625)\n\n* upcast the other way\r\n\r\n* Revert \"upcast the other way\"\r\n\r\nThis reverts commit 355692ba796d9026c0ad0c92829fe6ddaf81d09e.\r\n\r\n* remove uop cast, this should have never been there\r\n\r\n* add regression test\r\n\r\n* now fuzz it\r\n\r\ncorrect test\r\n\r\n* the accumulator is always the output type\r\n\r\nlint\r\n\r\n* fuzz all reduce ops\r\n\r\n* MULACC upcast_dtype could be half too\r\n\r\nopencl supports it https://man.opencl.org/mad.html\r\n\r\n* cast to the same dtype is a noop\r\n\r\n* internal casting support for MULACC\r\n\r\n* fuzz test mulacc internal casting\r\n\r\n* get_reduce_dtype\r\n\r\nhandle vectorized acc\r\n\r\nupdate get_reduce_acc calls with the correct dtype\r\n\r\nupdate tests\r\n\r\n* pending _complete_ implementation of a function that gets the dtype based on self.reduceop\r\n\r\n+more failing tests\r\n\r\n* get_reduce_dtype try 2\r\n\r\nadd TODO\r\n\r\n* get_lazyop_info already does it\r\n\r\n* cleanup\r\n\r\n* bring back internal casting support for mulacc\r\n\r\n* use the scalar version of the acc dtype\r\n\r\n* conceptual diff cleanup\r\n\r\n* one extra line to a cleaner linearizer\r\n\r\n* correct test assumptions - these should promote?\r\n\r\n* rm mulacc cast, the cast of vins happens with the acc dtype promotion\r\n\r\nlinearizer hacks\r\n\r\n* Revert \"rm mulacc cast, the cast of vins happens with the acc dtype promotion\"\r\n\r\nThis reverts commit afdd5407331d366c9421f30aa2626a4a359cd8e1.\r\n\r\nRevert \"correct test assumptions - these should promote?\"\r\n\r\nThis reverts commit 49ae2206ed8f5c70a727ba40af03ebbc512f316b.\r\n\r\n* skip tests blocked by MULACC->lazyop cleanup\r\n\r\n* final changes to add back internal casting for MULACC and update skip test logic, upcast works but downcast does not\r\n\r\n* only test the linearizer abstraction layer\r\n\r\nwe wanna ensure that linearizer matches whatever lazy is returning\r\n\r\n* remove unused hypothesis module\r\n\r\n* remove mulacc related changes, those will move to the lazy pr\r\n\r\n* remove midcast test\r\n\r\n* move to helpers\r\n\r\n* Revert \"remove midcast test\"\r\n\r\nThis reverts commit 86e74d796023e71122a09b9355bef493da6ae086.\r\n\r\nadd TODO with skip\r\n\r\n---------\r\n\r\nCo-authored-by: chenyu <chenyu@fastmail.com>","parents":["765f8b05e5d428e0c607f1448ed4a1f4ecf959f5"],"tree_hash":"1b37a968ea655830f63b2596d8ea976c98da267b"}