def get_pool2d_desc(ksize, stride, pad, mode):
"""Create a 2d pooling descriptor."""
desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(desc, libcudnn.cudnnPoolingMode[mode],
ksize[0], ksize[1], pad[0], pad[1],
stride[0], stride[1])
return Auto(desc, libcudnn.cudnnDestroyPoolingDescriptor)
def get_pool2d_desc(ksize, stride, pad, mode):
"""Create a 2d pooling descriptor."""
desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(
desc, libcudnn.cudnnPoolingMode[mode], ksize[0], ksize[1],
pad[0], pad[1], stride[0], stride[1])
return Auto(desc, libcudnn.cudnnDestroyPoolingDescriptor)
def configure(self, input):
in_images = input.shape[0]
in_channels = input.shape[1]
in_height = input.shape[2]
in_width = input.shape[3]
assert(in_width >= self.kW)
assert(in_height >= self.kH)
out_width = int((math.floor(1.0 * in_width - self.kW + 2*self.padW) / self.dW) + 1)
out_height = int((math.floor(1.0 * in_height - self.kH + 2*self.padH) / self.dH) + 1)
self.output = GPUTensor( (in_images, in_channels, out_height, out_width), input.dtype )
if self.pool_desc:
libcudnn.cudnnDestroyPoolingDescriptor(self.pool_desc)
if self.in_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.in_desc)
if self.out_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.out_desc)
self.in_desc = input.get_cudnn_tensor_desc()
self.out_desc = self.output.get_cudnn_tensor_desc()
self.pool_desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(self.pool_desc,
libcudnn.cudnnPoolingMode["CUDNN_POOLING_MAX"],
# libcudnn.cudnnNanPropagation["CUDNN_NOT_PROPAGATE_NAN"],
self.kH, self.kW, self.padH, self.padW, self.dH, self.dW)
def configure(self, input):
in_images = input.shape[0]
in_channels = input.shape[1]
in_height = input.shape[2]
in_width = input.shape[3]
assert (in_width >= self.kW)
assert (in_height >= self.kH)
out_width = int((math.floor(1.0 * in_width - self.kW + 2 * self.padW) /
self.dW) + 1)
out_height = int(
(math.floor(1.0 * in_height - self.kH + 2 * self.padH) / self.dH) +
1)
self.output = GPUTensor(
(in_images, in_channels, out_height, out_width), input.dtype)
if self.pool_desc:
libcudnn.cudnnDestroyPoolingDescriptor(self.pool_desc)
if self.in_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.in_desc)
if self.out_desc:
libcudnn.cudnnDestroyTensorDescriptor(self.out_desc)
self.in_desc = input.get_cudnn_tensor_desc()
self.out_desc = self.output.get_cudnn_tensor_desc()
self.pool_desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(
self.pool_desc,
libcudnn.cudnnPoolingMode["CUDNN_POOLING_MAX"],
# libcudnn.cudnnNanPropagation["CUDNN_NOT_PROPAGATE_NAN"],
self.kH,
self.kW,
self.padH,
self.padW,
self.dH,
self.dW)
kH = 2
dW = 1
dH = 1
padW = 0
padH = 0
in_width = x.shape[3]
in_height = x.shape[2]
out_width = int((math.floor(1.0 * in_width - kW + 2 * padW) / dW) + 1)
out_height = int((math.floor(1.0 * in_height - kH + 2 * padH) / dH) + 1)
print("Ot:", out_width, out_height)
y = GPUTensor((1, 1, out_height, out_width))
y_desc = y.get_cudnn_tensor_desc()
print(y_desc)
pool_desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(
pool_desc, libcudnn.cudnnPoolingMode["CUDNN_POOLING_MAX"], kH, kW, padH,
padW, dH, dW)
alpha = 1.0
beta = 0.0
libcudnn.cudnnPoolingForward(cudnn_context, pool_desc, alpha, x_desc.ptr,
x.get_gpu_voidp(), beta, y_desc.ptr,
y.get_gpu_voidp())
print("SHAPE:", y_desc.shape)
print(y.get())
kH = 2
dW = 1
dH = 1
padW = 0
padH = 0
in_width = x.shape[3]
in_height = x.shape[2]
out_width = int((math.floor(1.0 * in_width - kW + 2*padW) / dW) + 1)
out_height = int((math.floor(1.0 * in_height - kH + 2*padH) / dH) + 1)
print("Ot:", out_width, out_height)
y = GPUTensor((1,1,out_height, out_width))
y_desc = y.get_cudnn_tensor_desc()
print(y_desc)
pool_desc = libcudnn.cudnnCreatePoolingDescriptor()
libcudnn.cudnnSetPooling2dDescriptor(pool_desc,
libcudnn.cudnnPoolingMode["CUDNN_POOLING_MAX"],
kH, kW, padH, padW, dH, dW)
alpha = 1.0
beta = 0.0
libcudnn.cudnnPoolingForward(cudnn_context, pool_desc, alpha,
x_desc.ptr, x.get_gpu_voidp(), beta, y_desc.ptr, y.get_gpu_voidp())
print("SHAPE:", y_desc.shape)
print(y.get())