def execute(self, repeat=1, unbind=True):
for r in range(repeat):
if self.bsum_zero:
drv.memset_d32_async(*self.bsum_zero)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:], shared_size=self.shared)
if unbind:
self.bsum_zero = None
for kernel_params in self.kernels:
kernel_params[3:11] = (None,) * 8
def execute(self, repeat=1, unbind=True):
for r in range(repeat):
if self.bsum_zero:
drv.memset_d32_async(*self.bsum_zero)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:], shared_size=self.shared)
if unbind:
self.bsum_zero = None
for kernel_params in self.kernels:
kernel_params[3:11] = (None,) * 8
def execute(self, repeat=1, unbind=True):
for r in range(repeat):
if not self.determ:
drv.memset_d32_async(*self.zero_args)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:])
if self.convert_args:
_fp_convert(*self.convert_args)
if unbind:
self.zero_args = self.convert_args = None
for kernel_params in self.kernels:
kernel_params[3:8] = (None,) * 5
def execute(self, repeat=1, unbind=True):
for r in range(repeat):
if not self.determ:
drv.memset_d32_async(*self.zero_args)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:])
if self.convert_args:
_fp_convert(*self.convert_args)
if unbind:
self.zero_args = self.convert_args = None
for kernel_params in self.kernels:
kernel_params[3:8] = (None, ) * 5
def execute(self, repeat=1, unbind=True):
shuffle_kernel = _get_transpose_kernel(self.dtype_str)
kernel = kernel_specs.get_kernel(self.kernel[0])
for r in range(repeat):
# let atomic adds accumulate on top
if not self.beta:
drv.memset_d8_async(*self.zero_args)
shuffle_kernel.prepared_async_call(*self.shuffle_args)
kernel.prepared_async_call(*self.kernel[1:])
if unbind:
self.zero_args = None
self.shuffle_args[2:5] = (None,) * 3
self.kernel[3:8] = (None,) * 5
def execute(self, repeat=1, unbind=True):
shuffle_kernel = _get_transpose_kernel(self.dtype_str)
kernel = kernel_specs.get_kernel(self.kernel[0])
for r in range(repeat):
# let atomic adds accumulate on top
if not self.beta:
drv.memset_d8_async(*self.zero_args)
shuffle_kernel.prepared_async_call(*self.shuffle_args)
kernel.prepared_async_call(*self.kernel[1:])
if unbind:
self.zero_args = None
self.shuffle_args[2:5] = (None, ) * 3
self.kernel[3:8] = (None, ) * 5
def execute(self, repeat=1, unbind=True):
shuffle_kernel = _get_shuffle_kernel(self.dtype_str)
for r in range(repeat):
if self.bsum_zero:
drv.memset_d32_async(*self.bsum_zero)
shuffle_kernel.prepared_async_call(*self.shuffle_args)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:], shared_size=self.shared)
if unbind:
self.bsum_zero = None
self.shuffle_args[2:5] = (None,) * 3
for kernel_params in self.kernels:
kernel_params[3:11] = (None,) * 8
def execute(self, repeat=1, unbind=True):
shuffle_kernel = _get_shuffle_kernel(self.dtype_str)
for r in range(repeat):
if self.bsum_zero:
drv.memset_d32_async(*self.bsum_zero)
shuffle_kernel.prepared_async_call(*self.shuffle_args)
for kernel_params in self.kernels:
kernel = kernel_specs.get_kernel(kernel_params[0])
kernel.prepared_async_call(*kernel_params[1:], shared_size=self.shared)
if unbind:
self.bsum_zero = None
self.shuffle_args[2:5] = (None,) * 3
for kernel_params in self.kernels:
kernel_params[3:11] = (None,) * 8