rt_mod_cuda.cc 3.31 KB
Newer Older
1
#include "codegen_cuda.h"
2
#include "runtime/cuda/cuda_module.h"
3
4
5
6

namespace tvm {
namespace codegen {

7
8
static std::unordered_map<std::string, runtime::FunctionInfo>
ExtractFuncInfo(const IRModule &mod) {
9
10
11
  std::unordered_map<std::string, runtime::FunctionInfo> fmap;

  for (auto kv : mod->functions) {
12
13
    ICHECK(kv.second->IsInstance<tir::PrimFuncNode>())
        << "Can only lower IR Module with PrimFuncs";
14
15
16
17
18
19
20
21
22
23
24
25
26
27
    auto f = Downcast<tir::PrimFunc>(kv.second);

    runtime::FunctionInfo info;
    for (size_t i = 0; i < f->params.size(); ++i) {
      if (f->params[i]->dtype.is_handle()) {
        auto ptr = f->params[i]->type_annotation.as<PointerTypeNode>();
        if (ptr && ptr->storage_scope == "grid_constant") {
          info.arg_types.push_back(DataType(kTVMGridConstant, 64, 1));
          continue;
        }
      }
      info.arg_types.push_back(f->params[i].dtype());
    }
    if (auto opt = f->GetAttr<Array<String>>(tir::attr::kKernelLaunchParams)) {
28
      for (const auto &tag : opt.value()) {
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
        info.launch_param_tags.push_back(tag);
      }
    }
    auto global_symbol = f->GetAttr<String>(tvm::attr::kGlobalSymbol);
    fmap[static_cast<std::string>(global_symbol.value())] = info;
  }
  return fmap;
}

runtime::Module BuildTileLangCUDA(IRModule mod, Target target) {
  using tvm::runtime::Registry;
  bool output_ssa = false;
  CodeGenTileLangCUDA cg;
  cg.Init(output_ssa);

  for (auto kv : mod->functions) {
45
46
    ICHECK(kv.second->IsInstance<PrimFuncNode>())
        << "CodeGenTileLangCUDA: Can only take PrimFunc";
47
    auto gvar = Downcast<GlobalVar>(kv.first);
48
49
50
    auto f = Downcast<PrimFunc>(kv.second);
    auto calling_conv = f->GetAttr<Integer>(tvm::attr::kCallingConv);
    ICHECK(calling_conv == CallingConv::kDeviceKernelLaunch);
51
    cg.AddFunction(gvar, f);
52
53
54
  }

  std::string code = cg.Finish();
55
  if (const auto *f = Registry::Get("tilelang_callback_cuda_postproc")) {
56
57
58
59
    code = (*f)(code, target).operator std::string();
  }
  std::string fmt = "ptx";
  std::string ptx;
60
  if (const auto *f = Registry::Get("tilelang_callback_cuda_compile")) {
61
    ptx = (*f)(code, target).operator std::string();
62
63
    if (ptx[0] != '/')
      fmt = "cubin";
64
65
66
67
68
69
  } else {
    ICHECK(0);
  }
  return runtime::CUDAModuleCreate(ptx, fmt, ExtractFuncInfo(mod), code);
}

70
runtime::Module BuildTileLangCUDAWithoutCompile(IRModule mod, Target target) {
71
72
73
74
75
76
  using tvm::runtime::Registry;
  bool output_ssa = false;
  CodeGenTileLangCUDA cg;
  cg.Init(output_ssa);

  for (auto kv : mod->functions) {
77
78
    ICHECK(kv.second->IsInstance<PrimFuncNode>())
        << "CodeGenTileLangCUDA: Can only take PrimFunc";
79
    auto gvar = Downcast<GlobalVar>(kv.first);
80
81
82
    auto f = Downcast<PrimFunc>(kv.second);
    auto calling_conv = f->GetAttr<Integer>(tvm::attr::kCallingConv);
    ICHECK(calling_conv == CallingConv::kDeviceKernelLaunch);
83
    cg.AddFunction(gvar, f);
84
85
86
  }

  std::string code = cg.Finish();
87
  if (const auto *f = Registry::Get("tilelang_callback_cuda_postproc")) {
88
89
    code = (*f)(code, target).operator std::string();
  }
90
  return runtime::CUDAModuleCreate("ptx", "ptx", ExtractFuncInfo(mod), code);
91
92
}

93
94
TVM_REGISTER_GLOBAL("target.build.tilelang_cuda")
    .set_body_typed(BuildTileLangCUDA);
95
96
TVM_REGISTER_GLOBAL("target.build.tilelang_cuda_without_compile")
    .set_body_typed(BuildTileLangCUDAWithoutCompile);
97

98
99
} // namespace codegen
} // namespace tvm