mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
e429a68478
Summary: Setting `PYTORCH_NVFUSER_ONE_OP_FUSION=1` will take all nodes nvFuser support, instead of waiting for fusion opportunity. Pull Request resolved: https://github.com/pytorch/pytorch/pull/70000 Reviewed By: samdow Differential Revision: D33292195 Pulled By: davidberard98 fbshipit-source-id: 8ed5ce5e82fbb6737e8ab5ce4223b038eaf47756
55 lines
1.7 KiB
C++
55 lines
1.7 KiB
C++
#pragma once
|
|
|
|
#include <torch/csrc/Export.h>
|
|
#include <torch/csrc/jit/ir/ir.h>
|
|
#include <torch/csrc/jit/runtime/profiling_record.h>
|
|
|
|
/*
|
|
* This file contains APIs for cuda fuser;
|
|
*
|
|
* We use an empty static struct to hold the function pointers, which are
|
|
* registered separately. This is to support cpu-only compilation.
|
|
* Registration is done in torch/csrc/jit/codegen/cuda/register_interface.cpp
|
|
*/
|
|
|
|
namespace torch {
|
|
namespace jit {
|
|
namespace fuser {
|
|
namespace cuda {
|
|
|
|
TORCH_API std::atomic<bool>& getCudaFusionGuardMode();
|
|
|
|
C10_EXPORT bool getSingletonFusion();
|
|
C10_EXPORT bool setSingletonFusion(bool value);
|
|
C10_EXPORT bool getHorizontalFusion();
|
|
C10_EXPORT bool setHorizontalFusion(bool value);
|
|
|
|
// dummy struct to allow API registration
|
|
struct CudaFuserInterface {
|
|
void (*fn_compile_n)(Node*) = nullptr;
|
|
void (*fn_run_n_s)(const Node*, Stack&) = nullptr;
|
|
void (*fn_fuse_graph)(std::shared_ptr<Graph>&) = nullptr;
|
|
bool (*fn_can_fuse_n)(const Node*) = nullptr;
|
|
void (*fn_insert_profile_inodes)(ProfilingRecord* pr) = nullptr;
|
|
bool (*fn_profile_n)(const Node*) = nullptr;
|
|
};
|
|
|
|
// Get interface, this is used by registration and user facing API internally
|
|
C10_EXPORT CudaFuserInterface* getFuserInterface();
|
|
|
|
C10_EXPORT void compileFusionGroup(Node* fusion_node);
|
|
C10_EXPORT void runFusionGroup(const Node* fusion_node, Stack& stack);
|
|
C10_EXPORT void fuseGraph(std::shared_ptr<Graph>&);
|
|
C10_EXPORT bool canFuseNode(const Node* node);
|
|
C10_EXPORT void InsertProfileNodesForCUDAFuser(ProfilingRecord* pr);
|
|
C10_EXPORT bool profileNode(const Node* node);
|
|
|
|
C10_EXPORT bool complyWith(
|
|
const at::Tensor& tensor,
|
|
const c10::TensorTypePtr& guard_tensor_type);
|
|
|
|
} // namespace cuda
|
|
} // namespace fuser
|
|
} // namespace jit
|
|
} // namespace torch
|