FM / tensorexpr

/usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/jit/tensorexpr/

/usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/jit/tensorexpr

Name	Size	Mode	Actions
operators/	-	0755	rm
analysis.h	5888	0644	edit dl rm
block_codegen.h	4211	0644	edit dl rm
bounds_inference.h	2230	0644	edit dl rm
bounds_overlap.h	3329	0644	edit dl rm
codegen.h	6402	0644	edit dl rm
cpp_codegen.h	2278	0644	edit dl rm
cpp_intrinsics.h	719	0644	edit dl rm
cuda_codegen.h	7782	0644	edit dl rm
cuda_random.h	2642	0644	edit dl rm
dim_arg.h	884	0644	edit dl rm
eval.h	9639	0644	edit dl rm
exceptions.h	3253	0644	edit dl rm
expr.h	11588	0644	edit dl rm
external_functions.h	1274	0644	edit dl rm
external_functions_registry.h	2343	0644	edit dl rm
fwd_decls.h	2806	0644	edit dl rm
graph_opt.h	2553	0644	edit dl rm
half_support.h	5038	0644	edit dl rm
hash_provider.h	7930	0644	edit dl rm
intrinsic_symbols.h	420	0644	edit dl rm
ir.h	22622	0644	edit dl rm
ir_cloner.h	2069	0644	edit dl rm
ir_mutator.h	2010	0644	edit dl rm
ir_printer.h	3693	0644	edit dl rm
ir_simplifier.h	15090	0644	edit dl rm
ir_verifier.h	1240	0644	edit dl rm
ir_visitor.h	1825	0644	edit dl rm
kernel.h	9210	0644	edit dl rm
llvm_codegen.h	3180	0644	edit dl rm
llvm_jit.h	1965	0644	edit dl rm
loopnest.h	21599	0644	edit dl rm
mem_dependency_checker.h	13003	0644	edit dl rm
reduction.h	6742	0644	edit dl rm
registerizer.h	12498	0644	edit dl rm
stmt.h	21138	0644	edit dl rm
tensor.h	7640	0644	edit dl rm
tensorexpr_init.h	268	0644	edit dl rm
types.h	3880	0644	edit dl rm
unique_name_manager.h	940	0644	edit dl rm
var_substitutor.h	1753	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/jit/tensorexpr/graph_opt.h (2553B)

#pragma once

#include <torch/csrc/jit/ir/ir.h>

namespace torch {
namespace jit {
namespace tensorexpr {

// Optimize aten::cat ops in the given subgraph.
//
// Moving users of cat to its inputs.
//    Cat ops get lowered into multiple loops, one per input. When the result
//    of cat is used by some other op, it results in a situation where inlining
//    of cat does not happen. This in turn results in intermediate buffers
//    being created for the result of cat, since it is not inlined.
//
//    For example, consider the following graph:
//       graph(%x : Float(10, strides=[1], device=cpu),
//             %y : Float(20, strides=[1], device=cpu)):
//         %dim : int = prim::Constant[value=0]()
//         %xy_list : Tensor[] = prim::ListConstruct(%x, %y)
//         %cat : Float(60, strides=[1], device=cpu) = aten::cat(%xy_list, %dim)
//         %5 : Float(60, strides=[1], device=cpu) = aten::log(%cat)
//         return (%5))IR";
//
//     This will get lowered into:
//         Allocate(aten_cat);
//         for (...)
//           aten_cat[...] = x[...]
//         for (...)
//           aten_cat[...] = y[...]
//         for (...)
//           aten_log[...] = log(aten_cat[...])
//         Free(aten_cat);
//     Note that aten_cat is not inlined into aten_log and it results in
//     an intermediate buffer allocation as well.
//
//     Optimization:
//        We move the ops that use the result of `cat` into its inputs whenever
//     possible.
//
//     The graph above will be transformed to:
//        graph(%x : Float(10, strides=[1], device=cpu),
//              %y : Float(20, strides=[1], device=cpu)):
//          %3 : int = prim::Constant[value=0]()
//          %7 : Float(10, strides=[1], device=cpu) = aten::log(%x)
//          %8 : Float(20, strides=[1], device=cpu) = aten::log(%y)
//          %9 : Tensor[] = prim::ListConstruct(%7, %8)
//          %10 : Float(60, strides=[1], device=cpu) = aten::cat(%9, %3)
//          return (%10)
//
//     This will get lowered into:
//         for (...)
//           aten_cat[...] = log(x[...])
//         for (...)
//           aten_cat[...] = log(y[...])
//     aten_cat is the output buffer here.

bool OptimizeCat(const std::shared_ptr<Graph>& graph);

TORCH_API void annotateInputShapes(
    const std::shared_ptr<Graph>& graph,
    const std::vector<c10::optional<at::Tensor>>& example_inputs);
TORCH_API std::shared_ptr<Graph> removeUnusedSelfArgument(
    const std::shared_ptr<Graph>& graph);

} // namespace tensorexpr
} // namespace jit
} // namespace torch

#pragma once #include namespace torch { namespace jit { namespace tensorexpr { // Optimize aten::cat ops in the given subgraph. // // Moving users of cat to its inputs. // Cat ops get lowered into multiple loops, one per input. When the result // of cat is used by some other op, it results in a situation where inlining // of cat does not happen. This in turn results in intermediate buffers // being created for the result of cat, since it is not inlined. // // For example, consider the following graph: // graph(%x : Float(10, strides=[1], device=cpu), // %y : Float(20, strides=[1], device=cpu)): // %dim : int = prim::Constant[value=0]() // %xy_list : Tensor[] = prim::ListConstruct(%x, %y) // %cat : Float(60, strides=[1], device=cpu) = aten::cat(%xy_list, %dim) // %5 : Float(60, strides=[1], device=cpu) = aten::log(%cat) // return (%5))IR"; // // This will get lowered into: // Allocate(aten_cat); // for (...) // aten_cat[...] = x[...] // for (...) // aten_cat[...] = y[...] // for (...) // aten_log[...] = log(aten_cat[...]) // Free(aten_cat); // Note that aten_cat is not inlined into aten_log and it results in // an intermediate buffer allocation as well. // // Optimization: // We move the ops that use the result of `cat` into its inputs whenever // possible. // // The graph above will be transformed to: // graph(%x : Float(10, strides=[1], device=cpu), // %y : Float(20, strides=[1], device=cpu)): // %3 : int = prim::Constant[value=0]() // %7 : Float(10, strides=[1], device=cpu) = aten::log(%x) // %8 : Float(20, strides=[1], device=cpu) = aten::log(%y) // %9 : Tensor[] = prim::ListConstruct(%7, %8) // %10 : Float(60, strides=[1], device=cpu) = aten::cat(%9, %3) // return (%10) // // This will get lowered into: // for (...) // aten_cat[...] = log(x[...]) // for (...) // aten_cat[...] = log(y[...]) // aten_cat is the output buffer here. bool OptimizeCat(const std::shared_ptr& graph); TORCH_API void annotateInputShapes( const std::shared_ptr& graph, const std::vector>& example_inputs); TORCH_API std::shared_ptr removeUnusedSelfArgument( const std::shared_ptr& graph); } // namespace tensorexpr } // namespace jit } // namespace torch