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
|
#pragma once
#include <c10/macros/Export.h>
#include <torch/csrc/jit/codegen/cuda/dispatch.h>
#include <torch/csrc/jit/codegen/cuda/ir_all_nodes.h>
#include <torch/csrc/jit/codegen/cuda/kernel_ir.h>
#include <vector>
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
//! Collects start and stop offsets of all split root domains. Offsets
//! are zero unless partially split.
class TORCH_CUDA_CU_API PartialSplitMap {
public:
void build(Fusion* fusion);
Val* getStartOffset(IterDomain* root_domain) const;
Val* getStopOffset(IterDomain* root_domain) const;
private:
std::unordered_map<IterDomain*, Val*> start_offset_map_;
std::unordered_map<IterDomain*, Val*> stop_offset_map_;
};
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch
|
