def compile(self):
if self.ptx is None:
program = _NVRTCProgram(kernel.encode(),
"recurrent_forget_mult.cu".encode())
GPUForgetMult.ptx = program.compile()
if torch.cuda.current_device() not in GPUForgetMult.configured_gpus:
m = function.Module()
m.load(bytes(self.ptx.encode()))
self.forget_mult = m.get_function("recurrent_forget_mult")
self.bwd_forget_mult = m.get_function("bwd_recurrent_forget_mult")
Stream = namedtuple("Stream", ["ptr"])
self.stream = Stream(ptr=torch.cuda.current_stream().cuda_stream)
GPUForgetMult.configured_gpus[torch.cuda.current_device()] = (
self.forget_mult,
self.bwd_forget_mult,
self.stream,
)
(
self.forget_mult,
self.bwd_forget_mult,
self.stream,
) = GPUForgetMult.configured_gpus[torch.cuda.current_device()]
def compile(self):
if self.ptx is None:
program = _NVRTCProgram(kernel.encode(), 'recurrent_forget_mult.cu'.encode())
GPUForgetMult.ptx = program.compile()
if torch.cuda.current_device() not in GPUForgetMult.configured_gpus:
m = function.Module()
m.load(bytes(self.ptx.encode()))
self.forget_mult = m.get_function('recurrent_forget_mult')
self.bwd_forget_mult = m.get_function('bwd_recurrent_forget_mult')
Stream = namedtuple('Stream', ['ptr'])
self.stream = Stream(ptr=torch.cuda.current_stream().cuda_stream)
GPUForgetMult.configured_gpus[torch.cuda.current_device()] = (self.forget_mult, self.bwd_forget_mult, self.stream)
self.forget_mult, self.bwd_forget_mult, self.stream = GPUForgetMult.configured_gpus[torch.cuda.current_device()]