TConv3DBpFilterTiling Structure

The TConv3DBpFilterTiling structure contains the Conv3dBackpropFilter operator specifications and parameters related to the tiling algorithm. It is passed to the Conv3dBackpropFilter kernel for data tiling, data transfer, and computation. For details about the parameters of the TConv3DBpFilterTiling structure, see Table 1.

Call the GetTiling API to obtain the TConv3DBpFilterTiling structure. For details, see Conv3DBackpropFilter Tiling Instructions. Currently, the parameters in the TConv3DBpFilterTiling structure cannot be customized.

Table 1 TConv3DBpFilterTiling structure description

Parameter

Description

batch

Batch size of the input GradOutput. The unit is element.

cin

Channel of the input. The unit is element.

cout

Channel of the input GradOutput. The unit is element.

cin1G

Reserved parameter, which is imperceptible to users.

cout1G

Reserved parameter, which is imperceptible to users.

dout

Depth of the input GradOutput. The unit is element.

ho

Height of the input GradOutput. The unit is element.

wo

Width of the input GradOutput. The unit is element.

di

Depth of the input. The unit is element.

hi

Height of the input. The unit is element.

wi

Width of the input. The unit is element.

dk

Depth of the output weight. The unit is element.

hk

Height of the output weight. The unit is element.

wk

Width of the output weight. The unit is element.

group

Reserved parameter, which is imperceptible to users.

strideD

Depth of the stride in backpropagation convolution computation. The unit is element.

strideH

Height of the stride in backpropagation convolution computation. The unit is element.

strideW

Width of the stride in backpropagation convolution computation. The unit is element.

padFront

Forward direction of the depth dimension of the padding in backpropagation convolution computation. The unit is element.

padBack

Backward direction of the depth dimension of the padding in backpropagation convolution computation. The unit is element.

padUp

Upward direction of the height dimension of the padding in backpropagation convolution computation. The unit is element.

padDown

Downward direction of the height dimension of the padding in backpropagation convolution computation. The unit is element.

padLeft

Leftward direction of the width dimension of the padding in backpropagation convolution computation. The unit is element.

padRight

Rightward direction of the width dimension of the padding in backpropagation convolution computation. The unit is element.

dilationD

Depth of the dilation in backpropagation convolution computation. The unit is element.

dilationH

Height of the dilation in backpropagation convolution computation. The unit is element.

dilationW

Width of the dilation in backpropagation convolution computation. The unit is element.

channelSize

Size of C0 in the current input data type. Currently, this parameter can only be set to 16.

al0Pbuffer

The value 1 indicates that DoubleBuffer is disabled, and the value 2 indicates that DoubleBuffer is enabled.

bl0Pbuffer

The value 1 indicates that DoubleBuffer is disabled, and the value 2 indicates that DoubleBuffer is enabled.

cl0Pbuffer

The value 1 indicates that DoubleBuffer is disabled, and the value 2 indicates that DoubleBuffer is enabled.

al1Pbuffer

The value 1 indicates that DoubleBuffer is disabled, and the value 2 indicates that DoubleBuffer is enabled.

bl1Pbuffer

The value 1 indicates that DoubleBuffer is disabled, and the value 2 indicates that DoubleBuffer is enabled.

baseM

Size of the M direction on L0. The unit is element.

baseK

Size of the K direction on L0. The unit is element.

baseN

Size of the N direction on L0. The unit is element.

m0

Size of the minimum fractal in the M direction on L0.

k0

Size of the minimum fractal in the K direction on L0.

n0

Size of the minimum fractal in the N direction on L0.

stepM

Multiple of baseM in the buffer M direction of the matrix cached in L1.

stepN

Multiple of baseN in the buffer N direction of the matrix cached in L1.

stepKa

Multiple of baseK in the buffer K direction of the matrix cached in L1.

stepKb

Multiple of baseK in the buffer K direction of the matrix cached in L1.

iterateOrder

Reserved parameter, which is imperceptible to users.

bl1Bound

Maximum amount of data loaded to the GradOutput matrix in L1.

hf32Flag

Reserved parameter, which is imperceptible to users.

singleCoreDK

Reserved parameter, which is imperceptible to users.

singleCoreGroup

Reserved parameter, which is imperceptible to users.

singleCoreCout

Size of the cout data to be computed in the M direction of a single core, in elements.

singleCoreHo

Size of the ho data to be computed in the K direction of a single core, in elements.

singleCoreBatch

Size of batch on a single core. The unit is element.

singleCoreCin

Size of the cin data to be computed in the N direction of a single core. The unit is element.

totalL1Size

L1 size. The unit is element.

singleCoreM

Size of M on a single core. The unit is element.

singleCoreN

Size of N on a single core. The unit is element.

singleCoreK

Size of K on a single core. The unit is element.