1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
|
#pragma once
#include <c10/macros/Export.h>
#include <torch/csrc/jit/ir/ir.h>
#include <torch/csrc/jit/runtime/profiling_record.h>
#include <torch/csrc/jit/codegen/cuda/fusion.h>
/*
* This file handles Parsing PyTorch jit ir;
*
* It is used in two places:
* 1. When partitioning PyTorch jit ir to create prim::CudaFusionGroup, each
* node is queried by `isNodeParsible` to determine whether the node could
* be handled by the fuser (whether a given PyTorch jit operator should be
* merged);
* 2. lowering PyTorch jit ir to CUDA codegen ir.
* creates a `Fusion` by traversing a PyTorch jit graph.
*
* TODO: we could consider exposing API to allow custom registration of parsing
* rules for a given PyTorch jit operator.
*/
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
constexpr int kPwThreadX = 128;
constexpr int kFcdReductionThreadX = 128;
constexpr int kNonFcdReductionThreadX = 32;
constexpr int kNonFcdReductionThreadY = 32;
TORCH_CUDA_CU_API bool hasReductionNode(const Block* block);
TORCH_CUDA_CU_API bool isReductionToSizeNode(const Node* node);
TORCH_CUDA_CU_API bool isReductionNode(const Node* node);
TORCH_CUDA_CU_API bool hasNormalizationNode(const Block* block);
TORCH_CUDA_CU_API bool isNormalizationNode(const Node* node);
TORCH_CUDA_CU_API bool isElementWiseNode(const Node* node);
// returns whether or not a parsing function exists for the given node type.
TORCH_CUDA_CU_API bool isNodeParsible(const Node* node);
TORCH_CUDA_CU_API bool shouldProfileNode(const Node* node);
TORCH_CUDA_CU_API bool skipNodeKind(const std::string& symbol_str, bool flip);
void InsertProfileNodes(ProfilingRecord* pr);
// lowers PyTorch jit graph to `Fusion`.
TORCH_CUDA_CU_API std::unique_ptr<Fusion> parseJitIR(
const std::shared_ptr<Graph>& graph);
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch
|
