| 1 |
//===----- CGCUDARuntime.h - Interface to CUDA Runtimes ---------*- C++ -*-===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
// |
| 9 |
// This provides an abstract class for CUDA code generation. Concrete |
| 10 |
// subclasses of this implement code generation for specific CUDA |
| 11 |
// runtime libraries. |
| 12 |
// |
| 13 |
//===----------------------------------------------------------------------===// |
| 14 |
|
| 15 |
#ifndef LLVM_CLANG_LIB_CODEGEN_CGCUDARUNTIME_H |
| 16 |
#define LLVM_CLANG_LIB_CODEGEN_CGCUDARUNTIME_H |
| 17 |
|
| 18 |
#include "clang/AST/GlobalDecl.h" |
| 19 |
#include "llvm/ADT/StringRef.h" |
| 20 |
#include "llvm/IR/GlobalValue.h" |
| 21 |
|
| 22 |
namespace llvm { |
| 23 |
class Function; |
| 24 |
class GlobalVariable; |
| 25 |
} |
| 26 |
|
| 27 |
namespace clang { |
| 28 |
|
| 29 |
class CUDAKernelCallExpr; |
| 30 |
class NamedDecl; |
| 31 |
class VarDecl; |
| 32 |
|
| 33 |
namespace CodeGen { |
| 34 |
|
| 35 |
class CodeGenFunction; |
| 36 |
class CodeGenModule; |
| 37 |
class FunctionArgList; |
| 38 |
class ReturnValueSlot; |
| 39 |
class RValue; |
| 40 |
|
| 41 |
class CGCUDARuntime { |
| 42 |
protected: |
| 43 |
CodeGenModule &CGM; |
| 44 |
|
| 45 |
public: |
| 46 |
// Global variable properties that must be passed to CUDA runtime. |
| 47 |
class DeviceVarFlags { |
| 48 |
public: |
| 49 |
enum DeviceVarKind { |
| 50 |
Variable, // Variable |
| 51 |
Surface, // Builtin surface |
| 52 |
Texture, // Builtin texture |
| 53 |
}; |
| 54 |
|
| 55 |
/// The kind flag for an offloading entry. |
| 56 |
enum OffloadEntryKindFlag : uint32_t { |
| 57 |
/// Mark the entry as a global entry. This indicates the presense of a |
| 58 |
/// kernel if the size field is zero and a variable otherwise. |
| 59 |
OffloadGlobalEntry = 0x0, |
| 60 |
/// Mark the entry as a managed global variable. |
| 61 |
OffloadGlobalManagedEntry = 0x1, |
| 62 |
/// Mark the entry as a surface variable. |
| 63 |
OffloadGlobalSurfaceEntry = 0x2, |
| 64 |
/// Mark the entry as a texture variable. |
| 65 |
OffloadGlobalTextureEntry = 0x3, |
| 66 |
}; |
| 67 |
|
| 68 |
private: |
| 69 |
unsigned Kind : 2; |
| 70 |
unsigned Extern : 1; |
| 71 |
unsigned Constant : 1; // Constant variable. |
| 72 |
unsigned Managed : 1; // Managed variable. |
| 73 |
unsigned Normalized : 1; // Normalized texture. |
| 74 |
int SurfTexType; // Type of surface/texutre. |
| 75 |
|
| 76 |
public: |
| 77 |
DeviceVarFlags(DeviceVarKind K, bool E, bool C, bool M, bool N, int T) |
| 78 |
: Kind(K), Extern(E), Constant(C), Managed(M), Normalized(N), |
| 79 |
SurfTexType(T) {} |
| 80 |
|
| 81 |
DeviceVarKind getKind() const { return static_cast<DeviceVarKind>(Kind); } |
| 82 |
bool isExtern() const { return Extern; } |
| 83 |
bool isConstant() const { return Constant; } |
| 84 |
bool isManaged() const { return Managed; } |
| 85 |
bool isNormalized() const { return Normalized; } |
| 86 |
int getSurfTexType() const { return SurfTexType; } |
| 87 |
}; |
| 88 |
|
| 89 |
CGCUDARuntime(CodeGenModule &CGM) : CGM(CGM) {} |
| 90 |
virtual ~CGCUDARuntime(); |
| 91 |
|
| 92 |
virtual RValue EmitCUDAKernelCallExpr(CodeGenFunction &CGF, |
| 93 |
const CUDAKernelCallExpr *E, |
| 94 |
ReturnValueSlot ReturnValue); |
| 95 |
|
| 96 |
/// Emits a kernel launch stub. |
| 97 |
virtual void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) = 0; |
| 98 |
|
| 99 |
/// Check whether a variable is a device variable and register it if true. |
| 100 |
virtual void handleVarRegistration(const VarDecl *VD, |
| 101 |
llvm::GlobalVariable &Var) = 0; |
| 102 |
|
| 103 |
/// Finalize generated LLVM module. Returns a module constructor function |
| 104 |
/// to be added or a null pointer. |
| 105 |
virtual llvm::Function *finalizeModule() = 0; |
| 106 |
|
| 107 |
/// Returns function or variable name on device side even if the current |
| 108 |
/// compilation is for host. |
| 109 |
virtual std::string getDeviceSideName(const NamedDecl *ND) = 0; |
| 110 |
|
| 111 |
/// Get kernel handle by stub function. |
| 112 |
virtual llvm::GlobalValue *getKernelHandle(llvm::Function *Stub, |
| 113 |
GlobalDecl GD) = 0; |
| 114 |
|
| 115 |
/// Get kernel stub by kernel handle. |
| 116 |
virtual llvm::Function *getKernelStub(llvm::GlobalValue *Handle) = 0; |
| 117 |
|
| 118 |
/// Adjust linkage of shadow variables in host compilation. |
| 119 |
virtual void |
| 120 |
internalizeDeviceSideVar(const VarDecl *D, |
| 121 |
llvm::GlobalValue::LinkageTypes &Linkage) = 0; |
| 122 |
}; |
| 123 |
|
| 124 |
/// Creates an instance of a CUDA runtime class. |
| 125 |
CGCUDARuntime *CreateNVCUDARuntime(CodeGenModule &CGM); |
| 126 |
|
| 127 |
} |
| 128 |
} |
| 129 |
|
| 130 |
#endif |