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_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)
# 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)
# Get pointers to GPU memory
X_data = ctypes.c_void_p(int(X.gpudata))
cuO = ng.empty(dimO[::-1], dtype=np.float32)
cuI[:] = 2 * (.5 - ng.rand())
cuF[:] = 2 * (.5 - ng.rand())
cuE[:] = 2 * (.5 - ng.rand())
#print drv.mem_get_info()
I_data = ctypes.c_void_p(int(cuI.gpudata))
F_data = ctypes.c_void_p(int(cuF.gpudata))
O_data = ctypes.c_void_p(int(cuO.gpudata))
E_data = ctypes.c_void_p(int(cuE.gpudata))
B_data = ctypes.c_void_p(int(cuB.gpudata))
U_data = ctypes.c_void_p(int(cuU.gpudata))
libcudnn.cudnnSetConvolution2dDescriptor(C_desc, pad_h, pad_w, str_h, str_w, 1, 1, conv_mode)
libcudnn.cudnnSetTensor4dDescriptor(I_desc, NCHW_fmt, cu_dtype, N, C, H, W)
libcudnn.cudnnSetTensor4dDescriptor(B_desc, NCHW_fmt, cu_dtype, N, C, H, W)
libcudnn.cudnnSetTensor4dDescriptor(O_desc, NCHW_fmt, cu_dtype, N, K, P, Q)
libcudnn.cudnnSetTensor4dDescriptor(E_desc, NCHW_fmt, cu_dtype, N, K, P, Q)
libcudnn.cudnnSetFilter4dDescriptor(F_desc, cu_dtype, K, C, R, S)
libcudnn.cudnnSetFilter4dDescriptor(U_desc, cu_dtype, K, C, R, S)
algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(cudnn, I_desc, F_desc, C_desc, O_desc, fwd_pref, 0)
ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(cudnn, I_desc, F_desc, C_desc, O_desc, algo)
#print algo.value, ws_size.value
ws_ptr = drv.mem_alloc(ws_size.value) if ws_size.value > 0 else 0
ws_data = ctypes.c_void_p(int(ws_ptr))
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))
cuU = ng.empty(dimF[::-1], dtype=np.float32)
cuO = ng.empty(dimO[::-1], dtype=np.float32)
cuI[:] = 2 * (.5 - ng.rand())
cuF[:] = 2 * (.5 - ng.rand())
cuE[:] = 2 * (.5 - ng.rand())
#print drv.mem_get_info()
I_data = ctypes.c_void_p(int(cuI.gpudata))
F_data = ctypes.c_void_p(int(cuF.gpudata))
O_data = ctypes.c_void_p(int(cuO.gpudata))
E_data = ctypes.c_void_p(int(cuE.gpudata))
B_data = ctypes.c_void_p(int(cuB.gpudata))
U_data = ctypes.c_void_p(int(cuU.gpudata))
libcudnn.cudnnSetConvolution2dDescriptor(C_desc, pad_h, pad_w, str_h,
str_w, 1, 1, conv_mode)
libcudnn.cudnnSetTensor4dDescriptor(I_desc, NCHW_fmt, cu_dtype, N, C, H, W)
libcudnn.cudnnSetTensor4dDescriptor(B_desc, NCHW_fmt, cu_dtype, N, C, H, W)
libcudnn.cudnnSetTensor4dDescriptor(O_desc, NCHW_fmt, cu_dtype, N, K, P, Q)
libcudnn.cudnnSetTensor4dDescriptor(E_desc, NCHW_fmt, cu_dtype, N, K, P, Q)
libcudnn.cudnnSetFilter4dDescriptor(F_desc, cu_dtype, K, C, R, S)
libcudnn.cudnnSetFilter4dDescriptor(U_desc, cu_dtype, K, C, R, S)
algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(
cudnn, I_desc, F_desc, C_desc, O_desc, fwd_pref, 0)
ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(
cudnn, I_desc, F_desc, C_desc, O_desc, algo)
#print algo.value, ws_size.value
ws_ptr = drv.mem_alloc(ws_size.value) if ws_size.value > 0 else 0