MMALIB Release Notes

Version: 02.02.00.03

Introduction
Licensing
Getting Started
Documentation
What's New
Upgrade and Compatibility Information
Device Support
Validation Information
Fixed Issues
Known Issues
Technical Support
Package Versioning

Introduction

The MMALIB package consists of the Texas Instruments optimized kernels for CNN, FFT, LINALG algorithms.

Licensing

The licensing information of this library and a complete manifest along with export control information is detailed here [HTML].

Getting Started

The MMALIB User Guide [USER_GUIDE] provides the documentation and references necessary to begin development on TI's platforms.

Documentation

Refer to following documentation for further details:

MMALIB User Guide	Build instructions, API Guide	[USER_GUIDE]
Test Reports	Misra C reports, conformance test reports, TI platform test reports	[TEST_RESULTS]
Software Manifest	Licenses, terms of use	[HTML]

What's New

Here are a few of the new features supported in this release:

[new] Tensor convert for formating data and tensor types

CNN

MMALIB_CNN_convolve_row_ixX_ixX_oxX
MMALIB_CNN_convolve_row_1x1stride2_ixX_ixX_oxX
MMALIB_CNN_convolve_row_3x3stride2_ixX_ixX_oxX
MMALIB_CNN_convolve_row_3x3stride3_ixX_ixX_oxX
MMALIB_CNN_convolve_row_5x5stride2_ixX_ixX_oxX
MMALIB_CNN_convolve_row_7x7stride2_ixX_ixX_oxX
MMALIB_CNN_convolve_row_11x11stride4_ixX_ixX_oxX
MMALIB_CNN_convolve_col_smallNo_ixX_ixX_oxX
MMALIB_CNN_convolve_col_smallNo_3x3stride2_ixX_ixX_oxX
MMALIB_CNN_deconvolve_row_2x2Stride2_ixX_ixX_oxX
MMALIB_CNN_deconvolve_row_4x4Stride2_ixX_ixX_oxX
MMALIB_CNN_deconvolve_row_8x8Stride2_ixX_ixX_oxX
MMALIB_CNN_pixelShuffleUpscale2_row_ixX_ixX_oxX
MMALIB_CNN_pixelShuffleUpscale4_row_ixX_ixX_oxX
MMALIB_CNN_pixelShuffleUpscale8_row_ixX_ixX_oxX
MMALIB_CNN_fullyConnected_ixX_ixX_oxX
MMALIB_CNN_generateFillSeamPredicateRegisters
MMALIB_CNN_tensor_convert_ixX_oxX
MMALIB_CNN_generateFillBiasPredicateRegisters

LINALG

8, 16 and 32 bit signed integer support
MMALIB_LINALG_matrixMatrixMultiply_ixX_ixX_oxX
MMALIB_LINALG_pointwiseMatrixMatrixMultiply_ixX_ixX_oxX
MMALIB_LINALG_matrixMatrixMultiplyAccumulate_ixX_ixX_ixX_oxX
MMALIB_LINALG_matrixTranspose_ixX_oxX

FFT

Support for 16-bit and 32-bit signed complex input/output
Support for real and imaginary parts of data to be in either interleaved or non-interleaved format. Optimal performance assumes interleaved data format
Support for FFT sizes from 2 through 8192. FFT sizes are assumed to be powers of 2
Support for batch processing of multiple FFTs in one kernel call. For optimal performance, input data and output data for the entire batch, and twiddle factor data are assumed to fit in L2 memory
[New] support for C7x only kernels; host-emulation mode support is TBD
MMALIB_FFT_dftSmall_ixX_cxX_oxX
MMALIB_FFT_dftLarge_ixX_cxX_oxX
MMALIB_FFT_highRadixDecompositions_ixX_cxX_oxX
MMALIB_FFT_fft_ixX_cxX_oxX
MMALIB_fft1dBatched_i16sc_c16sc_o16sc [C7x only]
MMALIB_fft1dBatched_i32fc_c32fc_o32fc [C7x only]
MMALIB_fft1d_i16sc_c16sc_o16sc [C7x only]
MMALIB_fft1d_i32fc_c32fc_o32fc [C7x only]
Utility functions (with suffixes _getSizes) to compute sizes required for data and twiddle factor buffers
Utility functions (with suffixes _twGen) to compute relevant DFT matrices and twiddle factors

DSP

MMALIB_DSP_fir_ixX_ixX_oxX
MMALIB_DSP_firSmall_ixX_ixX_oxX

Details

CNN style 2D convolution

K = 1: optimized 8 bit, 16 bit
K = 2: optimized 8 bit, 16 bit
K = 3: optimized 8 bit, 16 bit
K > 3: optimized
Dilation : optimized
1x1 : supported as regular convolution

CNN style 2D convolution 1x1 stride by 2

K = 1: optimized 8 bit, 16 bit
K = 2: optimized 8 bit, 16 bit
K = 3: optimized 8 bit, 16 bit
K > 3: optimized

K = 1: optimized 8 bit tested, 16 bit
K = 2: optimized 8 bit tested, 16 bit
K = 3: optimized 8 bit tested, 16 bit
K > 3: optimized

CNN style 2D convolution 3x3 stride by 3

K = 1: optimized 8 bit tested, 16 bit
K = 2: optimized 8 bit tested, 16 bit
K = 3: optimized 8 bit tested, 16 bit
K > 3: optimized

CNN style 2D convolution 5x5 stride by 2

K = 1: optimized 8 bit tested, 16 bit
K = 2: optimized 8 bit tested, 16 bit
K = 3: optimized 8 bit tested, 16 bit
K > 3: optimized

CNN style 2D convolution 7x7 stride by 2

K = 1: currently natural C
K = 2: currently natural C
K = 3: optimized 8 bit tested, 16 bit
K > 3: optimized

CNN style 2D convolution 11x11 stride by 4

K = 1: currently natural C
K = 2: currently natural C
K = 3: currently natural C
K > 3: optimized

All:

Condition for Mflag =1. The seam insertion predicate restricted to support (image width + pad)*image height less than 9Kbytes fits in L1 and otherwise use L2 buffer
Generic strided convolution is supported as natural C.
Strided convolution with dilation is supported as natural C.
K and KBlocks is total filter coefficient per output channle/MMA block size (8 bit 64 and 16 bit 32), MBlocks is Number of output channels/MMA Block size, NBlocks is Feature map (width*height/MMA Size)
Mflag=1 is the condition with number of output channels do not fit in L2.
Please refer user guide for more details

CNN style 2D convolution 3x3 5x5 7x7 Ni = No = 1, small

stride-by-1
- optimized-C and natural-C
stride-by-1 with dilation
- optimized-C and natural-C
3x3 stride-by-2
- optimized-C and natural-C
- dilation with stride is not supported
- optimized-C requires the number of rows in a feature map to be even
All:
- Bias can be embedded into kernel coefficients or a seperator when calling the reorder weights function
- Split processing of a single group across both sides of the MMA
- Multiple groups processed in single kernel call
- Supports multiple rows of bias provided the rows do not push the computation beyond the MMA block size

CNN style 2D deconvolution

Supports 8- and 16-bit datatypes
Supports signed datatype with kernel matrix
Support only for optimized code for 2x2, 4x4, 8x8 stride 2

Fully Connected

Supports 8- and 16-bit datatypes
Supports signed and unsigned combinations of datatypes with kernel matrix always assumed to be signed
Support of multiple Bias rows using 2D DECDIM feature

DFT building blocks

16-bit, interleaved/non-interleaved, any FFT_SIZE, any BATCH_SIZE: works
32-bit, interleaved/non-interleaved, any FFT_SIZE, any BATCH_SIZE: works

FFT building blocks

16-bit, interleaved/non-interleaved, 64 < FFT_SIZE < 8192, BATCH_SIZE = 1,2,4,8,16,32,64: works
32-bit, interleaved/non-interleaved, 32 < FFT_SIZE < 8192, BATCH_SIZE = 1,2,4,8,16,32,64: works

FFT coordination

16-bit, interleaved/non-interleaved, any FFT_SIZE, any BATCH_SIZE: works
32-bit, interleaved/non-interleaved, any FFT_SIZE, any BATCH_SIZE: works

Upgrade and Compatibility Information

File

Change description

User application change required

User application recompile required

Device Support

SoC	Host (OS)	Target (OS)	Test Platform
J7ES C7x, MMA	No OS	No OS

Validation Information

This release was built and validated using the following tools:

Build Tools (NOT included in MMALIB):

C7x CGT C7000-CGT-2.0.1 STS

Fixed Issues

[MMALIB-524] MMALIB shall support Tensor convert needed by TVM
[MMALIB-526] MMALIB shall support 1D FFT single/ batch on C7x

Known Issues

None

Technical Support

For technical support, please post your questions on TI E2E Forum for Automotive ADAS SoCs.

For additional assistance, contact local TI Field Application Engineer

Package Versioning

Each package version is composed of 4 period-delimited numbers - represented here by the letters M, m, p and b [M.m.p.b]. The table below provides a descriptive reference regarding package version numbering.

Digit	Meaning	Description
1 (M=Major)	Major revision	Incremented when the new version is substantially different from the previous For example, a new module added or an existing module's algorithm significantly altered.
2 (m=minor)	Minor revision	Incremented when the new version has changed but not in a major way. For example, some minor changes in the API or feature set.
3 (p=patch)	Patch number	Incremented for all other source code changes. This include any packaging support code.
4 (b=build)	Build number	Incremented for each release delivery to CM. Reset for any change to M, m or p