def configure(self, input):
# print("Convolution::configure: input shape =", input.shape)
in_images = input.shape[0]
in_channels = input.shape[1]
in_height = input.shape[2]
in_width = input.shape[3]
assert (in_channels == self.num_filter_channels)
out_width = int((1.0 * in_width + 2 * self.padW - self.kW) / self.dW +
1)
out_height = int((1.0 * in_height + 2 * self.padH - self.kH) /
self.dH + 1)
self.output = GPUTensor(
(in_images, self.num_filter_maps, out_height, out_width),
input.dtype)
# print("ONCV:", input.dtype, self.output.dtype)
# print("Convolution::configure: output shape =", self.output.shape)
# initialize cudnn descriptors
if self.in_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.in_desc.ptr)
if self.out_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.out_desc.ptr)
self.in_desc = input.get_cudnn_tensor_desc()
# Get output dimensions (first two values are n_input and filters_out)
_, _, out_height2, out_width2 = libcudnn.cudnnGetConvolution2dForwardOutputDim(
self.conv_desc, self.in_desc.ptr, self.filt_desc)
assert (out_width == out_width2)
assert (out_height == out_height2)
self.out_desc = self.output.get_cudnn_tensor_desc()
# find best convolution algorithm
self.algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(
context.cudnn, self.in_desc.ptr, self.filt_desc, self.conv_desc,
self.out_desc.ptr, self.convolution_fwd_pref, 0)
print("Convolution::configure: algo=%s" % str(self.algo.value))
self.ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(
context.cudnn, self.in_desc.ptr, self.filt_desc, self.conv_desc,
self.out_desc.ptr, self.algo)
self.ws_ptr = drv.mem_alloc(
self.ws_size.value) if self.ws_size.value > 0 else 0
print("Convolution::configure: workspace size=%d" % self.ws_size.value)
def configure(self, input):
# print("Convolution::configure: input shape =", input.shape)
in_images = input.shape[0]
in_channels = input.shape[1]
in_height = input.shape[2]
in_width = input.shape[3]
assert(in_channels == self.num_filter_channels)
out_width = int((1.0 * in_width + 2*self.padW - self.kW) / self.dW + 1);
out_height = int((1.0 * in_height + 2*self.padH - self.kH) / self.dH + 1);
self.output = GPUTensor((in_images, self.num_filter_maps, out_height, out_width),
input.dtype)
# print("ONCV:", input.dtype, self.output.dtype)
# print("Convolution::configure: output shape =", self.output.shape)
# initialize cudnn descriptors
if self.in_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.in_desc.ptr)
if self.out_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.out_desc.ptr)
self.in_desc = input.get_cudnn_tensor_desc()
# Get output dimensions (first two values are n_input and filters_out)
_, _, out_height2, out_width2 = libcudnn.cudnnGetConvolution2dForwardOutputDim(
self.conv_desc, self.in_desc.ptr, self.filt_desc)
assert(out_width == out_width2)
assert(out_height == out_height2)
self.out_desc = self.output.get_cudnn_tensor_desc()
# find best convolution algorithm
self.algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(context.cudnn, self.in_desc.ptr,
self.filt_desc, self.conv_desc, self.out_desc.ptr, self.convolution_fwd_pref, 0)
print("Convolution::configure: algo=%s" % str(self.algo.value))
self.ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(context.cudnn,
self.in_desc.ptr, self.filt_desc, self.conv_desc, self.out_desc.ptr, self.algo)
self.ws_ptr = drv.mem_alloc(self.ws_size.value) if self.ws_size.value > 0 else 0
print("Convolution::configure: workspace size=%d" % self.ws_size.value)
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
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))
start_bench()
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)
end_bench("fprop")
ws_ptr = None
# Clean up
libcudnn.cudnnDestroyTensorDescriptor(X_desc)
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))
start_bench()
for r in (range(repeat)):
libcudnn.cudnnConvolutionForward(cudnn, alpha, I_desc, I_data, F_desc, F_data, C_desc, algo, ws_data, ws_size.value, beta, O_desc, O_data)
end_bench("fprop")
ws_ptr = None
start_bench()
for r in (range(repeat)):
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))
start_bench()
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)
end_bench("fprop")
ws_ptr = None
# Clean up
libcudnn.cudnnDestroyTensorDescriptor(X_desc)
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))
start_bench()
for r in (range(repeat)):
libcudnn.cudnnConvolutionForward(cudnn, alpha, I_desc, I_data, F_desc,
F_data, C_desc, algo, ws_data,
ws_size.value, beta, O_desc, O_data)
end_bench("fprop")
ws_ptr = None