QHL HVX is a set of libraries for the HVX extensions of the Qualcomm Hexagon DSPs with a v66 or above architectures instruction set. These libraries provide a set of functions performing various mathematical computations using HVX instructions. Floating-point libraries are only supported on v68 and above.

Since performance, alongside with accuracy, is crucial to each mathematical operation, most of the QHL HVX implementations are written in C with intrinsics.

The QHL HVX libraries are organized in the following groups:

Qualcomm Hexagon Math HVX (QHMATH HVX)
- Common mathematical computations on vectors of fixed-point or floating-point data
Qualcomm Hexagon Basic Linear Algebra Subprograms HVX (QHBLAS HVX)
- Basic linear algebra operations on vectors of fixed-point or floating-point data
Qualcomm Hexagon Digital Signal Processing HVX (QHDSP HVX)
- Digital signal processing operations on vectors of fixed-point or floating-point data

Dependencies

QHL HVX libraries are dependent on QHL libraries, that can be found under libs/qhl. In case of using QHL HVX libraries, user needs to be sure that QHL libraries are built and added to a linking line. QHL HVX libraries need dsp with HVX H/W feature and is primarily available in CDSP.

Naming convention

Naming convention used for QHL HVX library functions is as below:

<group_name>[_<output_format>]_<function description>_<input_format>
<output_format> is optional and will be used only if input and output formats differ in a function.
<function description> describes the function in brief

Variant	Allowed values
group_name	qhdsp_hvx, qhmath_hvx, qhcomplex_hvx, qhblas_hvx
output_format	af,ahf,ab,ah,aw,aub,auh,auw,ach,acw,acf,achf,aqf16,aqf32,f,hf,h,w,i
input_format	af,ahf,ab,ah,aw,aub,auh,auw,acub,ach,acw,acf,achf,aqf16,aqf32

The desciption for the short forms used in input and output formats is expanded in the table below: When a function takes multiple input arguments, the input format only refers to the format of the main argument.

Abbreviation	Description
a	array
c	complex
u	unsigned values
i	64-bit integer
f	single-precision float
hf	half-precision float
b	8-bit fixed-point or integer
h	16-bit fixed-point or integer
w	32-bit fixed-point or integer
qf16	16-bit QFLOAT
qf32	32-bit QFLOAT

Example: qhdsp_hvx_acw_c1dfft_acub: This function falls under the sub-group qhdsp_hvx. It takes an array of complex unsigned byte inputs(acub) and computes Complex 1-D FFT(c1dfft) outputs stored in an array of complex 32-bit integers(acw).

Build instructions

In order to build QHL HVX libraries, run the following command:

cd $HEXAGON_SDK_ROOT/libs/qhl_hvx
build_cmake hexagon DSP_ARCH=<v66/v68>

In order to remove all binaries, run the following command:

build_cmake hexagon_clean DSP_ARCH=<v66/v68>

Running an example

In order to see an example of how to invoke the libraries, please see $HEXAGON_SDK_ROOT/examples/qhl_hvx.

Running unit tests

The QHL HVX library comes with a set of unit tests that measures the accuracy and performance of each function. The following instructions demonstrate how to run the unit tests.

cd $HEXAGON_SDK_ROOT/libs/qhl_hvx/tests

Build and run accuracy tests on simulator

./run_accuracy_tests_simulator.<bash/cmd>  <v66/v68>

Build and run benchmarks on simulator

./run_benchmarks_simulator.<bash/cmd>  <v66/v68>

Build and run accuracy tests on device

./run_tests_dsp.<bash/cmd> <v66/v68>

Note: A v66 or v68 target must be connected to execute this command. HVX float tests require a v68 device.

Performance Numbers

The following table provides performance numbers for the QHL_HVX library functions. These numbers may vary depending on the exact system configuration.