TDA3: Vision SDK - Package content understanding

Pavan Divakar72 said:

1. What is the difference between vlib and vxlib present within <Vision_SDK>/ti_components? Are they the same modules with different interfaces? There seems to be an overlap in the type of packages in both.

VLIB is TI's legacy library of optimized vision kernels originally optimized for C64+ DSP.  It started circa 2007.  Along the way, we added support for C674 and C66 DSP libraries, and added several new functions for C66 only version.  New features to VLIB stopped in 2015, and it has been in maintenance mode since then.  It is released in binary only.

VXLIB is a new library released in the last year, and it started as a library of optimized vision kernels for C66 DSP that implements the Khronos OpenVX 1.1 standard kernel functions (though we plan to add TI extension kernels to be integrated in OpenVX).  If you look at the User guide, you can see how each kernel maps to which part of the Khronos OpenVX 1.1 specified kernels.  It follows a similar directory structure, test framework, and documentation style as VLIB for consistency, but it offers some additional enhancements:

1. Available as full source so user can modify and recompile as needed.

2. All API's follow consistent naming, parameter naming, and data structure conventions.

3. All images use a common data structure to define width, height, data type, and stride/pitch.  In VLIB some kernels supported stride/pitch, and others didn't.  The stride/pitch parameter allows for processing a region within a larger buffer.

4. Most kernels internally have a separate core function which implements a single loop, and can be called for the whole image if the strides are equal to the widths, or once for each line, if the strides are larger the  the width.  This enables the most optimal execution while enabling flexibility for the user.

5. Maintains optimization while offering more flexibility:

- 5a. doesn't constrain the user as to having a width as a multiple of some SIMD number (as much of VLIB did).

- 5b. doesn't constrain the user to requiring buffer pointers to be a multiple of some value (as much of VLIB did).

6. All kernels offer a checkParams function which can be independently called, or called from within the main function call for parameter checking (build option).

7. Most kernels are written in such a way that they can be called multiple times in a tiled fashion across the image, assuming a DMA breaks up the image into blocks, and processes the image one block at a time. To this end, many also come with BAM plugins, which enable them to be called by TI's block based DMA framework (BAM).  This library is in the algframework of the vision SDK.

Note: Both VXLIB and VLIB do not internally perform DMA transactions.  They operate directly on memory pointers given to it on the API.  If DMA's are to be used, it is done at a layer above VXLIB/VLIB by the user or other frameworks such as BAM.

Pavan Divakar72 said:

2. What is the difference between TIOVX library and vlib/vxlib? Is it just the interface for the same set of kernels/modules?

TIOVX is TI's implementation of Khronos OpenVX 1.1 framework.  Using this framework, applications can define a graph of kernel processing function from the host CPU, and OpenVX will manage the dispatch and execution of the graph across the system for each frame.

VXLIB is the underlying C66 DSP kernel library which implements the Khronos OpenVX kernels.

Whereas TIOVX is the framework which has components compiled on all processing cores in the system, VXLIB is compiled only on the C66 DSPs.  TIOVX has a layer (in the kernels folder) which define the interface between the framework and the VXLIB calls.  In this layer, the framework can call one or many VXLIB functions for each OpenVX kernel.  Additionally, it can call the kernels directly using cache, or indirectly using BAM, which manages the DMA acceleration of the data access into L2 RAM before calling the VXLIB functions on each block.  By default in TIOVX, the BAM acceleration is enabled for kernels which it is supported.

Pavan Divakar72 said:

3. How to decide which library to use in a particular scenario if the same kernel / function is available in all libraries (TIOVX, VLIB, VXLIB)? Any guidelines for this?

VXLIB and VLIB can be used on the C66 DSP independent of TIOVX.  If you already have your calling code running on the DSP and want to use acceleration from VXLIB or VLIB, you can use these libraries with direct calls.

If you have your application running on an ARM processor and want to call into one of these libraries (running on C66 DSP), it is probably easier for you to take advantage of the integration TI has already done and use the TIOVX interface so you don't have to worry about inter-processor communication, cache maintenance, dma acceleration, and synchronization.