def build_checker():
if hasattr(build_checker, "dll"):
return build_checker.dll
checker_code_path = os.path.join(
os.path.dirname(inspect.getfile(daceml.onnx)), "include",
"op_checker.h")
with open(checker_code_path, "r") as f:
checker_code = f.read()
program = codeobject.CodeObject("onnx_op_checker",
checker_code,
"cpp",
targets.cpu.CPUCodeGen,
"ONNXOpChecker",
environments={"ONNXRuntime"})
BUILD_PATH = os.path.join('.dacecache', "onnx_op_checker")
compiler.generate_program_folder(None, [program], BUILD_PATH)
compiler.configure_and_compile(BUILD_PATH)
checker_dll = ctypes.CDLL(
compiler.get_binary_name(BUILD_PATH, "onnx_op_checker"))
build_checker.dll = checker_dll
return checker_dll
def cuda_helper():
helper_code = """
#include <dace/dace.h>
extern "C" {
int host_to_gpu(void* gpu, void* host, size_t size) {
auto result = cudaMemcpy(gpu, host, size, cudaMemcpyHostToDevice);
DACE_CUDA_CHECK(cudaGetLastError());
DACE_CUDA_CHECK(cudaDeviceSynchronize());
return result;
}
}
"""
program = codeobject.CodeObject("cuda_helper", helper_code, "cpp",
targets.cpu.CPUCodeGen, "CudaHelper")
dummy_cuda_target = codeobject.CodeObject("dummy", "", "cu",
targets.cuda.CUDACodeGen,
"CudaDummy")
build_folder = dace.Config.get('default_build_folder')
BUILD_PATH = os.path.join(build_folder, "cuda_helper")
compiler.generate_program_folder(None, [program, dummy_cuda_target],
BUILD_PATH)
compiler.configure_and_compile(BUILD_PATH)
checker_dll = compiled_sdfg.ReloadableDLL(
compiler.get_binary_name(BUILD_PATH, "cuda_helper"), "cuda_helper")
class CudaHelper:
def __init__(self):
self.dll = checker_dll
checker_dll.load()
self._host_to_gpu = checker_dll.get_symbol("host_to_gpu")
self._host_to_gpu.restype = ctypes.c_int
def __del__(self):
self.dll.unload()
def host_to_gpu(self, gpu_ptr: int, numpy_array: np.ndarray):
size = ctypes.sizeof(
dtypes._FFI_CTYPES[numpy_array.dtype.type]) * numpy_array.size
result = ctypes.c_int(
self._host_to_gpu(
ctypes.c_void_p(gpu_ptr),
ctypes.c_void_p(
numpy_array.__array_interface__["data"][0]),
ctypes.c_size_t(size)))
if result.value != 0:
raise ValueError("host_to_gpu returned nonzero result!")
return CudaHelper()