def _fp_convert(src_data, src_type, dest_tensor, reduce_shape):
if reduce_shape:
kernel = _get_reduce_kernel(dest_tensor.dtype.str[1:])
blocks = _ceil_div(reduce_shape[1], 32)
kernel.prepared_async_call((blocks, 1, 1), (32, 1, 1),
dest_tensor.backend.stream,
dest_tensor.gpudata,
src_data,
reduce_shape[1],
reduce_shape[0]*reduce_shape[1])
else:
from neon.backends.nervanagpu import GPUTensor
from neon.backends.float_ew import _get_compound_kernel, _get_fast_ew_dims
# quick wrapper to convert raw fp32 scratch data to a destination tensor
shape, strides = _get_fast_ew_dims(dest_tensor.size)
kernel_args = [0,
dest_tensor.gpudata, strides[0], strides[1],
src_data, strides[0], strides[1],
shape[1]]
kernel = _get_compound_kernel((
(GPUTensor, 0, dest_tensor.dtype.str[1:], 0, False),
(GPUTensor, 1, src_type, 0, False),
('assign', 0, False, 32)),
dest_tensor.backend.compute_capability)
kernel.prepared_async_call((shape[0], 1, 1),
(32, 1, 1),
dest_tensor.backend.stream,
*kernel_args)
def _fp_convert(src_data, src_type, dest_tensor, reduce_shape):
if reduce_shape:
kernel = _get_reduce_kernel(dest_tensor.dtype.str[1:])
blocks = _ceil_div(reduce_shape[1], 32)
kernel.prepared_async_call(
(blocks, 1, 1), (32, 1, 1), dest_tensor.backend.stream,
dest_tensor.gpudata, src_data, reduce_shape[1],
reduce_shape[0] * reduce_shape[1])
else:
from neon.backends.nervanagpu import GPUTensor
from neon.backends.float_ew import _get_compound_kernel, _get_fast_ew_dims
# quick wrapper to convert raw fp32 scratch data to a destination tensor
shape, strides = _get_fast_ew_dims(dest_tensor.size)
kernel_args = [
0, dest_tensor.gpudata, strides[0], strides[1], src_data,
strides[0], strides[1], shape[1]
]
kernel = _get_compound_kernel(
((GPUTensor, 0, dest_tensor.dtype.str[1:], 0, False),
(GPUTensor, 1, src_type, 0, False), ('assign', 0, False, 32)),
dest_tensor.backend.compute_capability)
kernel.prepared_async_call((shape[0], 1, 1), (32, 1, 1),
dest_tensor.backend.stream, *kernel_args)