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
59
60
|
#pragma once
#include <ATen/core/stack.h>
#include <torch/csrc/WindowsTorchApiMacro.h>
#include <torch/csrc/jit/codegen/fuser/arg_spec.h>
#include <torch/csrc/jit/codegen/fuser/fused_kernel.h>
#include <torch/csrc/jit/codegen/fuser/interface.h>
#include <torch/csrc/jit/codegen/fuser/kernel_spec.h>
#include <torch/csrc/jit/ir/ir.h>
#include <cstdint>
#include <vector>
namespace torch {
namespace jit {
namespace fuser {
// Performs device-independent "upfront" compilation of the given fusion_group,
// if it has not been registered already.
// Returns a key that can be used to run the fusion later
TORCH_API int64_t registerFusion(const Node* fusion_group);
// Performs device-specific "runtime" compilation of the given kernel
// with the runtime arguments specified in ArgSpec.
// Outputs are allocated using map_size on the specified device.
TORCH_API std::shared_ptr<FusedKernel> compileKernel(
const KernelSpec& spec,
const ArgSpec& arg_spec,
const std::vector<int64_t>& map_size,
const at::Device device);
TORCH_API size_t nCompiledKernels();
TORCH_API int debugFuser();
using FusedKernelConstructor = std::function<std::shared_ptr<FusedKernel>(
int16_t device,
std::string name,
std::string code,
std::vector<TensorDesc> input_desc,
std::vector<TensorDesc> output_desc,
std::vector<PartitionDesc> chunk_desc,
std::vector<PartitionDesc> concat_desc,
bool has_random)>;
TORCH_API void registerFusionBackend(
at::Device::Type backend_type,
FusedKernelConstructor ctor);
TORCH_API bool hasFusionBackend(at::Device::Type backend_type);
struct TORCH_API RegisterFusionBackend {
RegisterFusionBackend(
at::Device::Type backend_type,
FusedKernelConstructor ctor) {
registerFusionBackend(backend_type, std::move(ctor));
}
};
} // namespace fuser
} // namespace jit
} // namespace torch
|
