Add necessary lowerings and decompositions to run adaptive_avg_pool2d #1121
Conversation
As per title. [ghstack-poisoned]
…_avg_pool2d" As per title. [ghstack-poisoned]
|
Done! |
…_avg_pool2d" As per title. [ghstack-poisoned]
|
It would be good to have a test for |
torchinductor/decomposition.py
Outdated
| aten.expand_as, | ||
| aten.flip, | ||
| aten.floor_divide.default, | ||
| aten.fmod.Tensor, |
There was a problem hiding this comment.
even though nominally we have fmod decomposition, attempting to use it leads to
def forward(self, arg0_1, arg1_1):
fmod = torch.ops.prims.fmod.default(arg0_1, arg1_1); arg0_1 = arg1_1 = None
return (fmod,)
torchinductor.graph: [WARNING] Creating implicit fallback for:
target: prims.fmod.default
args[0]: TensorBox(StorageBox(
InputBuffer(name='arg0_1', layout=FixedLayout('cuda', torch.int32, size=[s0], stride=[1]))
))
args[1]: TensorBox(StorageBox(
InputBuffer(name='arg1_1', layout=FixedLayout('cuda', torch.int32, size=[s0], stride=[1]))
))
torchinductor.ir: [WARNING] Using FallbackKernel: torch.ops.prims.fmod.default
(because decomposition just calls directly into prim)
so again we'd either have to write a real decomposition on the pytorch side, or lowering on the inductor side to support fmod.
Libdevice has fmod, there's also std::fmod so we should just add lowering
…_avg_pool2d" As per title. [ghstack-poisoned]
|
Done. I also added some tests. For signbit, there was no equivalent triton function, so I added an implementation. The implementation is not 100% correct, as it would return |
…_avg_pool2d" As per title. [ghstack-poisoned]
…_avg_pool2d" As per title. [ghstack-poisoned]
…_avg_pool2d" As per title. [ghstack-poisoned]
|
The error seems unrelated |
…_avg_pool2d" As per title. [ghstack-poisoned]
|
Ok, this PR had the same issue as described here #1120 (comment). Please tell me if I need to do anything on my side to fix this. |
5 participants
Stack from ghstack:
As per title.