kernels

Classes

Functions

k_arm_convolve_s16_group_ch_mult_1

k_arm_convolve_s16_group_ch_mult_1() -> Conv2dKernelSpec

Group s16 conv2d with ch_mult==1: filter C==1 and input_ch==output_ch.

resolve_float_conv_scratch_size

resolve_float_conv_scratch_size(*, cin: int, kh: int, kw: int, elt_size: int, batch: int, in_h: int, in_w: int, out_h: int, out_w: int, stride_height: int, stride_width: int, padding_height: int, padding_width: int, dilation_height: int, dilation_width: int) -> int

Port arm_convolve_{f16,f32}_get_buffer_size route selection.

CONV_2D's float kernel specs and DEPTHWISE_CONV_2D's convert-to-convolution reroute (see depthwise_conv_2d.operator._resolve_float_scratch) both hand a [Cout, kH, kW, Cin] filter and matching input/output dims to this same generic float convolution wrapper, so one port of its route selection sizes both.

Parameters:

• cin

  (int) –

  Input channel count.

• kh

  (int) –

  Filter height.

• kw

  (int) –

  Filter width.

• elt_size

  (int) –

  Size in bytes of one input element.

• batch

  (int) –

  Input batch count.

• in_h

  (int) –

  Input height.

• in_w

  (int) –

  Input width.

• out_h

  (int) –

  Output height.

• out_w

  (int) –

  Output width.

• stride_height

  (int) –

  Convolution stride along height.

• stride_width

  (int) –

  Convolution stride along width.

• padding_height

  (int) –

  Padding along height.

• padding_width

  (int) –

  Padding along width.

• dilation_height

  (int) –

  Dilation along height.

• dilation_width

  (int) –

  Dilation along width.

Returns:

• int ( int ) –

  Bytes the wrapper asks for, zero if it runs without scratch.