def create_convolution_descriptor(pad,
stride,
dtype,
mode=cudnn.CUDNN_CROSS_CORRELATION):
desc = Descriptor(cudnn.createConvolutionDescriptor(),
cudnn.destroyConvolutionDescriptor)
ndim = len(pad)
if ndim != len(stride):
raise ValueError('pad and stride must be of same length')
if ndim == 2:
cudnn.setConvolution2dDescriptor(desc.value, pad[0], pad[1], stride[0],
stride[1], 1, 1, mode)
else:
c_pad = _to_ctypes_array(pad)
c_stride = _to_ctypes_array(stride)
c_upscale = _to_ctypes_array((1, ) * ndim)
if _cudnn_version >= 3000:
data_type = get_data_type(dtype)
# TODO(takagi) Temporarily use computing precision of FP32 for
# storing precision of FP16.
if dtype == numpy.float16:
data_type = cudnn.CUDNN_DATA_FLOAT
cudnn.setConvolutionNdDescriptor_v3(desc.value, ndim, c_pad.data,
c_stride.data, c_upscale.data,
mode, data_type)
else:
cudnn.setConvolutionNdDescriptor_v2(desc.value, ndim, c_pad.data,
c_stride.data, c_upscale.data,
mode)
return desc
def create_convolution_descriptor(pad, stride, dtype, mode=cudnn.CUDNN_CROSS_CORRELATION):
desc = Descriptor(cudnn.createConvolutionDescriptor(), cudnn.destroyConvolutionDescriptor)
ndim = len(pad)
if ndim != len(stride):
raise ValueError("pad and stride must be of same length")
if ndim == 2:
cudnn.setConvolution2dDescriptor(desc.value, pad[0], pad[1], stride[0], stride[1], 1, 1, mode)
else:
c_pad = _to_ctypes_array(pad)
c_stride = _to_ctypes_array(stride)
c_upscale = _to_ctypes_array((1,) * ndim)
if _cudnn_version >= 3000:
data_type = get_data_type(dtype)
# TODO(takagi) Temporarily use computing precision of FP32 for
# storing precision of FP16.
if dtype == numpy.float16:
data_type = cudnn.CUDNN_DATA_FLOAT
cudnn.setConvolutionNdDescriptor_v3(
desc.value, ndim, c_pad.data, c_stride.data, c_upscale.data, mode, data_type
)
else:
cudnn.setConvolutionNdDescriptor_v2(desc.value, ndim, c_pad.data, c_stride.data, c_upscale.data, mode)
return desc
def create_convolution_descriptor(pad,
stride,
dtype,
mode=cudnn.CUDNN_CROSS_CORRELATION,
dilation=(1, 1),
use_tensor_core=False):
desc = Descriptor(cudnn.createConvolutionDescriptor(),
cudnn.destroyConvolutionDescriptor)
ndim = len(pad)
if ndim != len(stride):
raise ValueError('pad and stride must be of same length')
if ndim == 2:
if _cudnn_version < 6000:
if dilation[0] != 1 or dilation[1] != 1:
raise ValueError('dilation must be one when cudnn < 6.0')
if _cudnn_version >= 5000:
compute_type = get_data_type(dtype)
# TODO(takagi) Temporarily use computing precision of FP32 for
# storing precision of FP16.
if dtype == numpy.float16:
compute_type = cudnn.CUDNN_DATA_FLOAT
cudnn.setConvolution2dDescriptor_v5(desc.value, pad[0], pad[1],
stride[0], stride[1],
dilation[0], dilation[1], mode,
compute_type)
if _cudnn_version >= 7000:
if use_tensor_core:
math_type = cudnn.CUDNN_TENSOR_OP_MATH
cudnn.setConvolutionMathType(desc.value, math_type)
else:
cudnn.setConvolution2dDescriptor_v4(desc.value, pad[0], pad[1],
stride[0], stride[1], 1, 1,
mode)
else:
c_pad = _to_ctypes_array(pad)
c_stride = _to_ctypes_array(stride)
c_dilation = _to_ctypes_array((1, ) * ndim)
compute_type = get_data_type(dtype)
# TODO(takagi) Temporarily use computing precision of FP32 for
# storing precision of FP16.
if dtype == numpy.float16:
compute_type = cudnn.CUDNN_DATA_FLOAT
cudnn.setConvolutionNdDescriptor_v3(desc.value, ndim, c_pad.data,
c_stride.data, c_dilation.data,
mode, compute_type)
return desc