QLinearConv

The convolution operator consumes a quantized input tensor, its scale and zero point, a quantized filter, its scale and zero point, and output's scale and zero point, and computes the quantized output. 2D convolution on H, W dimensions of BCHW using kernels MCKK. Each c-th KxK kernel is applied on C dimension. This is done over M iterations to yield MxHxW per instance. This is done over B iterations to yield B batches. More...

Collaboration diagram for QLinearConv:

Modules
	QLinearConvKernels

Classes
class	QLinearConvGraph< PAD, QLINEARCONV, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Single instance graph that stores weights and biases Max size = 16384 and 4096 bytes respectively. More...
class	QLinearConvStreamGraph< PAD, QLINEARCONV, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Single instance graph that streams weights and biases, significantly slower. More...
class	QLinearConvChunkHGraph< QLINEARCONV, CONCAT, HCHUNK, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Multiinstance graph that stores weights and biases, chunks BCHW by H dimension, maximum 8 chunks. More...
class	QLinearConvChunkHStreamGraph< QLINEARCONV, CONCAT, HCHUNK, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Multiinstance graph that stores weights and biases, chunks BCHW by H dimension, maximum 8 chunks. More...
class	QLinearConvChunkHPktStreamGraph< QLINEARCONV, CONCAT, HCHUNK, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Multiinstance graph that stores weights and biases, chunks BCHW by H dimension, maximum 8 chunks. More...
class	QLinearConvChunkCGraph< QLINEARCONV0, QLINEARCONV1, QLINEARCONV2, CONCAT, CCHUNK, TT, TTPARAM, INP_H, INP_W, INP_W_PAD, OUT_W, OUT_W_PAD, STEP_H, STEP_W, B, C, M, KH, KW, GROUP, H0, H1, W0, W1 >
	Multiinstance graph that stores weights and biases, chunks BCHW by C dimension, maximum 8 chunks. More...

Detailed Description

Template Parameters

QLINEARCONV	Conv2D Kernel
TT	input/output dtype, int8_t or uint8_t only
TTPARAM	weight dtype, int8_t or uint8_t only
INP_H	input height
INP_W	input width to use
INP_W_PAD	input width, padded to vector boundary
OUT_W	output width to use, unable to infer due to width padding
OUT_W_PAD	output width, padded to vector boundary
STEP_H	stride in height dimension
STEP_W	stride in width dimension
B	batch size
C	input channels
M	output channels
KH	kernel height
KW	kernel width
GROUP	split input into groups, CGROUP==0
H0	Pixels added before height (default 0)
H1	Pixels added after height (default 0)
W0	Pixels added before width (default 0)
W1	Pixels added after width (default 0)