Performance Summary

Performance data was obtained on the J7AEP EVM. EVM warm cycle obtained by profiling the kernel's compute code execution after a cold run of the same code. Please refer to the kernel's documentation for more information about the parameters in the tables shown below.

---

CNN kernels

This section contains tables that depict the expected performance numbers for the CNN kernels.

Row-wise Convolution (MMALIB_CNN_convolveBias_row_ixX_ixX_oxX)

Bit Width	Stride	Fr	Fc	Ni	No	inWidth	validColsPerRowIn	validRowsIn	ValidColsIn	inChOffset	Bias	quantMethod	Pad left	Pad Right	Pad Top	Pad Bottom	EVM Cycles
8	1	3	3	256	64	64	NA	NA	768	1024	Yes	Yes	1	1	0	0	25261
8	1	5	5	64	64	64	NA	NA	1280	2048	Yes	Yes	2	2	0	0	27441
8	1	1	1	64	128	256	NA	NA	1024	1088	Yes	Yes	0	0	0	0	2657
8	2	5	5	12	64	128	128	18	NA	3072	Yes	Yes	2	2	2	2	3516

Depth-wise Convolution (MMALIB_CNN_convolve_col_smallNo_ixX_ixX_oxX)

Bit Width	Dilation Y	Top Pad	Bottom Pad	Left Pad	Right Pad	Pad Fill Value	Lc	Lr	Num. In Channels	Num Out Channels	Num. Groups Per Kernel	EVM Warm Cycles
8	1	1	1	1	1	0	256	256	1	1	4	4823
8	1	2	2	2	2	0	256	256	1	1	4	4745
8	1	1	1	1	1	-87	512	128	1	1	4	6884
16	1	2	2	2	2	-17	256	128	1	1	4	11160

Depth-wise Convolution with High Precision (MMALIB_CNN_convolve_col_smallNo_highPrecision)

Bit Width	kernelWidth	kernelHeight	strideX	strideY	dilationX	Dilation Y	Top Pad	Bottom Pad	Left Pad	Right Pad	Pad Fill Value	Lc	Lr	Num. In Channels	Num Out Channels	Num. Groups Per Kernel	EVM Warm Cycles
8	3	3	1	1	1	1	1	1	1	1	0	256	256	1	1	2	2529
8	5	5	1	1	1	1	2	2	2	2	0	256	256	1	1	2	2517
8	3	3	2	2	1	1	1	1	1	1	-56	512	128	1	1	2	3584
16	5	5	2	2	1	1	2	2	2	2	70	256	128	1	1	2	5728

Fully-connected Layer (MMALIB_CNN_fullyConnectedBias_ixX_ixX_oxX)

Bit Width	Batch Size	Ni	No	Stride Y	Stride X	Stride H	EVM Warm Cycles
8	64	1280	128	64	1600	81920	3214
16	64	1280	32	64	2624	81920	3110

LINALG kernels

This section contains tables that depict expected performance numbers for the linear algebra kernels.

Matrix-matrix Multiplication

Bit Width	m	k	n	Stride C	Stride A	Stride B	EVM Warm Cycles
8	255	255	255	320	320	320	4521
8	2048	64	64	64	64	64	2306

Matrix-matrix Multiply and Accumulate

 Bit Width	m	k	n	EVM Warm Cycles
8	65	257	257	5190
32	65	17	257	9484

Pointwise Matrix-matrix Multiply

 Bit Width	m	n	EVM Warm Cycles
8	1025	65	2228
16	513	33	1193
32	255	63	1211

Matrix Transpose

Bit Width	M	N	EVM Warm Cycles
32	56	56	412
8	256	256	1470

DSP kernels

Shown below are the expected performance numbers for the FIR filter kernel.

FIR Filter

Bit Width	Data Size	Batch Size	Filter Size	EVM Warm Cycles
32	2048	1	512	3840
32	4096	4	16	2231

FFT kernels

Shown below are the expected performance numbers for the FFT kernel.

Bit Width	FFT Size	Batch Size	EVM Warm Cycles
16	256	16	1060
16	4096	1	1668