pytorch

526659db - whitelist ops we can build shapes for (#49125)

whitelist ops we can build shapes for (#49125) Summary: Whitelist ops we can build shapes for. Otherwise, `buildShapeExpressions` assumes that `aten::unsqueeze` is just a regular op. ``` [DUMP tensorexpr_fuser.cpp:329] buildShapeExpressions for [DUMP tensorexpr_fuser.cpp:329] graph(%1 : float, [DUMP tensorexpr_fuser.cpp:329] %3 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0), [DUMP tensorexpr_fuser.cpp:329] %8 : float, [DUMP tensorexpr_fuser.cpp:329] %10 : Float(50, strides=[1], requires_grad=0, device=cuda:0)): [DUMP tensorexpr_fuser.cpp:329] %11 : int = prim::Constant[value=1]() [DUMP tensorexpr_fuser.cpp:329] %12 : Float(50, 1, strides=[1, 1], requires_grad=0, device=cuda:0) = aten::unsqueeze(%10, %11) [DUMP tensorexpr_fuser.cpp:329] %9 : Float(50, 1, strides=[1, 1], requires_grad=0, device=cuda:0) = aten::mul(%12, %8) [DUMP tensorexpr_fuser.cpp:329] %6 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0) = aten::add(%3, %9, %11) [DUMP tensorexpr_fuser.cpp:329] %2 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0) = aten::div(%6, %1) [DUMP tensorexpr_fuser.cpp:329] return (%2, %6, %9) [DEBUG tensorexpr_fuser.cpp:347] Adding a mapping for %3 %162 : int[] = aten::size(%27) [DEBUG tensorexpr_fuser.cpp:347] Adding a mapping for %10 %163 : int[] = aten::size(%23) [DEBUG tensorexpr_fuser.cpp:402] Building sizes for %12 : Float(50, 1, strides=[1, 1], requires_grad=0, device=cuda:0) = aten::unsqueeze(%10, %11) [DEBUG tensorexpr_fuser.cpp:405] Getting aten::size for %10 [DEBUG tensorexpr_fuser.cpp:402] Building sizes for %9 : Float(50, 1, strides=[1, 1], requires_grad=0, device=cuda:0) = aten::mul(%12, %8) [DEBUG tensorexpr_fuser.cpp:405] Getting aten::size for %12 [DEBUG tensorexpr_fuser.cpp:402] Building sizes for %6 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0) = aten::add(%3, %9, %11) [DEBUG tensorexpr_fuser.cpp:405] Getting aten::size for %3 [DEBUG tensorexpr_fuser.cpp:405] Getting aten::size for %9 [DEBUG tensorexpr_fuser.cpp:402] Building sizes for %2 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0) = aten::div(%6, %1) [DEBUG tensorexpr_fuser.cpp:405] Getting aten::size for %6 [DEBUG tensorexpr_fuser.cpp:907] Inserting a typecheck guard for a node%156 : Float(50, 28, strides=[28, 1], requires_grad=0, device=cuda:0) = prim::TensorExprGroup[Subgraph=<Graph>](%3, %27, %16, %23) [DUMP tensorexpr_fuser.cpp:463] After guarding fusion groups: ``` Pull Request resolved: https://github.com/pytorch/pytorch/pull/49125 Reviewed By: albanD Differential Revision: D25926997 Pulled By: Krovatkin fbshipit-source-id: f8041bbfc12be16c329754c6d16911d12aa352ef