xref: /aosp_15_r20/external/pytorch/torch/csrc/jit/codegen/fuser/compiler.h (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #pragma once
2 
3 #include <ATen/core/stack.h>
4 #include <torch/csrc/Export.h>
5 #include <torch/csrc/jit/codegen/fuser/arg_spec.h>
6 #include <torch/csrc/jit/codegen/fuser/fused_kernel.h>
7 #include <torch/csrc/jit/codegen/fuser/interface.h>
8 #include <torch/csrc/jit/codegen/fuser/kernel_spec.h>
9 #include <torch/csrc/jit/ir/ir.h>
10 
11 #include <cstdint>
12 #include <vector>
13 
14 namespace torch::jit::fuser {
15 
16 // Performs device-independent "upfront" compilation of the given fusion_group,
17 // if it has not been registered already.
18 // Returns a key that can be used to run the fusion later
19 TORCH_API int64_t registerFusion(const Node* fusion_group);
20 
21 // Performs device-specific "runtime" compilation of the given kernel
22 //  with the runtime arguments specified in ArgSpec.
23 //  Outputs are allocated using map_size on the specified device.
24 TORCH_API std::shared_ptr<FusedKernel> compileKernel(
25     const KernelSpec& spec,
26     const ArgSpec& arg_spec,
27     const std::vector<int64_t>& map_size,
28     const at::Device device);
29 
30 TORCH_API size_t nCompiledKernels();
31 
32 TORCH_API int debugFuser();
33 
34 using FusedKernelConstructor = std::function<std::shared_ptr<FusedKernel>(
35     int16_t device,
36     std::string name,
37     std::string code,
38     std::vector<TensorDesc> input_desc,
39     std::vector<TensorDesc> output_desc,
40     std::vector<PartitionDesc> chunk_desc,
41     std::vector<PartitionDesc> concat_desc,
42     bool has_random)>;
43 
44 TORCH_API void registerFusionBackend(
45     at::Device::Type backend_type,
46     FusedKernelConstructor ctor);
47 TORCH_API bool hasFusionBackend(at::Device::Type backend_type);
48 struct TORCH_API RegisterFusionBackend {
RegisterFusionBackendRegisterFusionBackend49   RegisterFusionBackend(
50       at::Device::Type backend_type,
51       FusedKernelConstructor ctor) {
52     registerFusionBackend(backend_type, std::move(ctor));
53   }
54 };
55 
56 } // namespace torch::jit::fuser
57