def __init__(self, config, name="Convolution"):
super().__init__(config, name)
self.output = None
self.W = self.load_tensor(config, 0)
self.alpha = 1.0
self.beta = 0.0
self.in_desc = None
self.out_desc = None
self.num_filter_maps = self.W.shape[0]
self.num_filter_channels = self.W.shape[1]
self.bias = self.load_tensor(config, 1, shape=(1, self.num_filter_maps, 1, 1))
# assert(self.bias.shape[0] == self.num_filter_maps)
# self.bias = self.bias.reshape((1, self.num_filter_maps, 1, 1))
# print(self.bias.shape)
self.b_desc = self.bias.get_cudnn_tensor_desc()
self.filt_desc = libcudnn.cudnnCreateFilterDescriptor()
print("FILT:", self.W.dtype, gputensor.np_2_cudnn_dtype[self.W.dtype])
print("FILT:", self.W.shape, self.num_filter_maps, self.num_filter_channels, self.kH, self.kW)
libcudnn.cudnnSetFilter4dDescriptor(self.filt_desc,
gputensor.np_2_cudnn_dtype[self.W.dtype], self.num_filter_maps,
self.num_filter_channels, self.kH, self.kW)
# print("B:", self.bias.shape)
# self.bias_desc =
self.conv_desc = libcudnn.cudnnCreateConvolutionDescriptor()
libcudnn.cudnnSetConvolution2dDescriptor(self.conv_desc, self.padH, self.padW,
self.dH, self.dW, 1, 1, self.convolution_mode)
def get_conv2d_desc(pad, stride, mode=_default_conv_mode):
"""Create a 2d convolution descriptor."""
desc = libcudnn.cudnnCreateConvolutionDescriptor()
libcudnn.cudnnSetConvolution2dDescriptor(desc, pad[0], pad[1], stride[0],
stride[1], 1, 1, mode)
return Auto(desc, libcudnn.cudnnDestroyConvolutionDescriptor)
def get_conv2d_desc(pad, stride, mode=_default_conv_mode):
"""Create a 2d convolution descriptor."""
desc = libcudnn.cudnnCreateConvolutionDescriptor()
libcudnn.cudnnSetConvolution2dDescriptor(
desc, pad[0], pad[1], stride[0], stride[1], 1, 1, mode)
return Auto(desc, libcudnn.cudnnDestroyConvolutionDescriptor)
np.random.rand(filters_out, filters_in, height_filter,
width_filter).astype(np.float32))
# Descriptor for input
X_desc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensor4dDescriptor(X_desc, tensor_format, data_type, n_input,
filters_in, height_in, width_in)
# Filter descriptor
filters_desc = libcudnn.cudnnCreateFilterDescriptor()
libcudnn.cudnnSetFilter4dDescriptor(filters_desc, data_type, tensor_format,
filters_out, filters_in, height_filter,
width_filter)
# Convolution descriptor
conv_desc = libcudnn.cudnnCreateConvolutionDescriptor()
libcudnn.cudnnSetConvolution2dDescriptor(conv_desc, pad_h, pad_w,
vertical_stride, horizontal_stride,
upscalex, upscaley, convolution_mode,
data_type)
# Get output dimensions (first two values are n_input and filters_out)
_, _, height_output, width_output = libcudnn.cudnnGetConvolution2dForwardOutputDim(
conv_desc, X_desc, filters_desc)
# Output tensor
Y = gpuarray.empty((n_input, filters_out, height_output, width_output),
np.float32)
Y_desc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensor4dDescriptor(Y_desc, tensor_format, data_type, n_input,
filters_out, height_output, width_output)
def start_bench():
start.record()
def end_bench(op):
end.record()
end.synchronize()
msecs = end.time_since(start) / repeat
gflops = conv.flops / (msecs * 1000000.0)
print "%7.3f msecs %8.3f gflops (%s: %s)" % (msecs, gflops, op, conv)
ng = NervanaGPU(stochastic_round=False, bench=True)
# Create a cuDNN context
cudnn = libcudnn.cudnnCreate()
C_desc = libcudnn.cudnnCreateConvolutionDescriptor()
I_desc = libcudnn.cudnnCreateTensorDescriptor()
O_desc = libcudnn.cudnnCreateTensorDescriptor()
E_desc = libcudnn.cudnnCreateTensorDescriptor()
B_desc = libcudnn.cudnnCreateTensorDescriptor()
F_desc = libcudnn.cudnnCreateFilterDescriptor()
U_desc = libcudnn.cudnnCreateFilterDescriptor()
# Set some options and tensor dimensions
NCHW_fmt = libcudnn.cudnnTensorFormat['CUDNN_TENSOR_NCHW']
cu_dtype = libcudnn.cudnnDataType['CUDNN_DATA_FLOAT']
conv_mode = libcudnn.cudnnConvolutionMode['CUDNN_CROSS_CORRELATION']
fwd_pref = libcudnn.cudnnConvolutionFwdPreference['CUDNN_CONVOLUTION_FWD_NO_WORKSPACE']
# CUDNN_CONVOLUTION_FWD_NO_WORKSPACE
# CUDNN_CONVOLUTION_FWD_PREFER_FASTEST
