CodeGen_PTX_Dev.h
#ifndef HALIDE_CODEGEN_PTX_DEV_H
#define HALIDE_CODEGEN_PTX_DEV_H
/** \file
* Defines the code-generator for producing CUDA host code
*/
#include "CodeGen_GPU_Dev.h"
#include "CodeGen_GPU_Host.h"
#include "CodeGen_LLVM.h"
namespace llvm {
class BasicBlock;
}
namespace Halide {
namespace Internal {
/** A code generator that emits GPU code from a given Halide stmt. */
class CodeGen_PTX_Dev : public CodeGen_LLVM, public CodeGen_GPU_Dev {
public:
friend class CodeGen_GPU_Host<CodeGen_X86>;
friend class CodeGen_GPU_Host<CodeGen_ARM>;
/** Create a PTX device code generator. */
CodeGen_PTX_Dev(Target host);
~CodeGen_PTX_Dev();
void add_kernel(Stmt stmt,
const std::string &name,
const std::vector<DeviceArgument> &args);
static void test();
std::vector<char> compile_to_src();
std::string get_current_kernel_name();
void dump();
virtual std::string print_gpu_name(const std::string &name);
std::string api_unique_name() { return "cuda"; }
protected:
using CodeGen_LLVM::visit;
/** (Re)initialize the PTX module. This is separate from compile, since
* a PTX device module will often have many kernels compiled into it for
* a single pipeline. */
/* override */ virtual void init_module();
/** We hold onto the basic block at the start of the device
* function in order to inject allocas */
llvm::BasicBlock *entry_block;
/** Nodes for which we need to override default behavior for the GPU runtime */
// @{
void visit(const For *);
void visit(const Allocate *);
void visit(const Free *);
void visit(const AssertStmt *);
void visit(const Load *);
void visit(const Store *);
// @}
std::string march() const;
std::string mcpu() const;
std::string mattrs() const;
bool use_soft_float_abi() const;
int native_vector_bits() const;
bool promote_indices() const {return false;}
/** Map from simt variable names (e.g. foo.__block_id_x) to the llvm
* ptx intrinsic functions to call to get them. */
std::string simt_intrinsic(const std::string &name);
};
} // namespace Internal
} // namespace Halide
#endif