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 __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 benchmark_conv(kw, kh, bsz):
start, end = (drv.Event(), drv.Event())
def start_bench():
start.record()
def end_bench():
end.record()
end.synchronize()
return end.time_since(start)
n_input = bsz
filters_in = 3
filters_out = 64
height_in = 224
width_in = 224
height_filter = kh
width_filter = kw
pad_h = 3
pad_w = 3
vertical_stride = 1
horizontal_stride = 1
upscalex = 1
upscaley = 1
alpha = 1.0
beta = 1.0
# Input tensor
X = gpuarray.to_gpu(np.random.rand(n_input, filters_in, height_in, width_in)
.astype(np.float32))
# Filter tensor
filters = gpuarray.to_gpu(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, 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)
# 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)
# Get pointers to GPU memory
X_data = ctypes.c_void_p(int(X.gpudata))
filters_data = ctypes.c_void_p(int(filters.gpudata))
Y_data = ctypes.c_void_p(int(Y.gpudata))
# Perform convolution
algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(cudnn_context, X_desc,
filters_desc, conv_desc, Y_desc, convolution_fwd_pref, 0)
# print("Cudnn algorithm = %d" % algo.value)
ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(cudnn_context, X_desc, filters_desc, conv_desc, Y_desc, algo)
ws_ptr = drv.mem_alloc(ws_size.value) if ws_size.value > 0 else 0
ws_data = ctypes.c_void_p(int(ws_ptr))
libcudnn.cudnnConvolutionForward(cudnn_context, alpha, X_desc, X_data,
filters_desc, filters_data, conv_desc, algo, ws_data, ws_size.value, beta,
Y_desc, Y_data)
start_bench()
for i in range(10):
libcudnn.cudnnConvolutionForward(cudnn_context, alpha, X_desc, X_data,
filters_desc, filters_data, conv_desc, algo, ws_data, ws_size.value, beta,
Y_desc, Y_data)
ms = end_bench()
ws_ptr = None
libcudnn.cudnnDestroyTensorDescriptor(X_desc)
libcudnn.cudnnDestroyTensorDescriptor(Y_desc)
libcudnn.cudnnDestroyFilterDescriptor(filters_desc)
libcudnn.cudnnDestroyConvolutionDescriptor(conv_desc)
return ms / 10
def get_filter4d_desc(x, mode=_default_conv_mode):
"""Create a 2d convolution filter descriptor."""
k, c, h, w = x.shape
desc = libcudnn.cudnnCreateFilterDescriptor()
libcudnn.cudnnSetFilter4dDescriptor(desc, _dtypes[x.dtype], k, c, h, w)
return Auto(desc, libcudnn.cudnnDestroyFilterDescriptor)
def get_filter4d_desc(x, mode=_default_conv_mode):
"""Create a 2d convolution filter descriptor."""
k, c, h, w = x.shape
desc = libcudnn.cudnnCreateFilterDescriptor()
libcudnn.cudnnSetFilter4dDescriptor(desc, _dtypes[x.dtype], k, c, h, w)
return Auto(desc, libcudnn.cudnnDestroyFilterDescriptor)
libcudnn.cudnnSetTensorNdDescriptor(xdesc, 0, 3,
[inputsize, minibatch, seqlength])
for xdesc in xdescs
]
hxdesc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensorNdDescriptor(hxdesc, 0, 3,
[hiddensize, minibatch, numlayers])
cxdesc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensorNdDescriptor(cxdesc, 0, 3,
[hiddensize, minibatch, numlayers])
paramssize = libcudnn.cudnnGetRNNParamsSize(handle, rnndesc, xdescs)
wdesc = libcudnn.cudnnCreateFilterDescriptor()
libcudnn.cudnnSetFilterNdDescriptor(wdesc, 0, 0, 3, [paramssize, 1, 1])
ydescs = [libcudnn.cudnnCreateTensorDescriptor() for _ in xrange(seqlength)]
[
libcudnn.cudnnSetTensorNdDescriptor(ydesc, 0, 3,
[hiddensize, minibatch, seqlength])
for ydesc in ydescs
]
hydesc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensorNdDescriptor(hydesc, 0, 3,
[hiddensize, minibatch, numlayers])
cydesc = libcudnn.cudnnCreateTensorDescriptor()
libcudnn.cudnnSetTensorNdDescriptor(cydesc, 0, 3,
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
# N C K D H W T R S pad str
for dims in ( ( 64, 3, 64, 1, 224,224, 1, 3, 3, 0,1,1, 1,1,1), # VGG
( 64, 64, 64, 1, 224,224, 1, 3, 3, 0,1,1, 1,1,1),
( 64, 64,128, 1, 112,112, 1, 3, 3, 0,1,1, 1,1,1),
( 64,128,128, 1, 112,112, 1, 3, 3, 0,1,1, 1,1,1),