Profiling in Audio Weaver: Tips and Tricks

About This Application Note

This ‘Profiling in Audio Weaver’ application note contains instructions for computational and memory profiling of and Audio Weaver Designer (.awd) signal flow in Audio Weaver. The profiling works in real time on a running .awd layout (Figure 1) for both Native (PC) and on embedded targets.

This profiling happens while the audio processing is running in real-time. Additionally, there is a manual audio pump feature which does similar profiling but operates in non-real-time.

Real Time Profiling

In Audio Weaver, Design Mode refers to a non-running .awd, while Tuning Mode refers to a running .awd layout. You can enter Tuning Mode by clicking the run button in Audio Weaver Designer:

While an .awd is in Tuning Mode (running), the entire layout can be profiled in real time. Profiling a running .awd will yield the following information:

Overall memory usage and cycle consumption of the complete audio system signal flow
- Overall memory usage and cycle consumption of each Audio Weaver instance for multi-instance architectures
- For more information about Audio Weaver’s multi-instance architectures, please see this Application Note:
  - https://dspconcepts.atlassian.net/wiki/x/5wbosQ
Memory usage and cycle consumption of each individual module in the audio system signal flow
- Memory usage and cycle consumption of each individual module per Audio Weaver instance for multi-instance architectures in the audio system signal flow
Memory usage of all the wires (audio buffers) in the audio system signal flow
Memory usage of shared heap memory for multi-instance architectures

To profile a running .awd in real time, navigate to the ‘Tools > Profiling Running Layout’ menu in the Audio Weaver Designer toolbar while in Tuning Mode:

There are two options for profiling a running .awd layout in real time:

Profile Block by Block
Profile Peak

The ‘Block by Block’ profiling will provide information MHz and Memory consumption of each individual element of the .awd at the time of the profiling, while ‘Peak’ profiling will provide information about average and peak CPU cycle usage over a user-specified time.

Selecting ‘Profile Block by Block’ will automatically run profiling on the entire running .awd layout at the time of selection, while ‘Profile Peak’ is manually started once the desired ‘Sampling Period’ and ‘Test Length” fields are set:

Peak Profiling Sampling Period & Length of Test

Understanding Real Time Profiling Information

Block by Block Profiling

When selecting the ‘Profile Block by Block’ real time profiling option, Audio Weaver will run complete MHz and Memory profiling on the entire audio system signal flow (.awd) and display the profiling information in a pop-up window. The ‘Tick’ mentioned here is the time measured in terms of CPU clock cycles. It can be measured as Tick = (1/CPU Clock Frequency). MHz is the Million Instructions Per Second (MIPS) of the system.

In the upper right-hand corner of the profiling window provides the following information:

Total ticks per block process available (calculated and measured)
Average ticks per block
Instantaneous ticks per block
Peak ticks per block
Total memory usage of the system
Shared heap memory for multi-instance architectures

Profile Block by Block Window (intentional CPU overflow)

Profile Block by Block Terminology

The following table describes each field:

Field	Definition	Unit
Total ticks per block process available	Measurement from the end of the CPU first interrupt, to the end of the next CPU interrupt (i.e., how many clock cycles have elapsed on the processor). This is a rough indication of how many clock cycles are available for processing. (You will not be able to utilize 100% of cycles because the audio interrupt handle requires some processing.) This is an especially important number to check when you are bringing up new hardware. The value shown here should be close to: (Processor Speed) x (Block Size of Processing) / (Sample Rate) If this doesn’t match, then there could be a mismatch in your processor speed, the audio sample rate, or the underlying “fundamental block size” of your implementation. Lastly, there are two separate line items for ‘Total ticks per block process available’. Calculated: Manually calculated based on the processor clock speed and the block size and sample rate of the layout Measured: Processor clock ticks measured between blocks, based on the calling rate of the audio pump Both of these units should nearly match, however be aware that if there are CPU overflows (>100% processing load), the ‘Measured’ metric will be doubled. In this scenario, since the audio pump is not completed during the current block of processing, the measured metric would include the next block as well.	CPU clock cycles
Average ticks per block used	Average of CPU clock cycles per block of audio data (10x)	CPU clock cycles
Instantaneous ticks per block used	CPU clock cycles required to process the last block of audio. This is the instantaneous measurement without smoothing. This number will change every time you profile.	CPU clock cycles
Peak ticks per block used	Peak instantaneous CPU clock cycles consumed when processing a block of audio data. This is a “sticky measurement” and shows the peak value since system startup. If you reprofile, then it will reset this value.	CPU clock cycles
Fast Heap	Memory usage from the memory allocated in the Fast Heap	Words
Fast Heap B	Memory usage from the memory allocated in the Fast Heap B	Words
Slow Heap	Memory usage from the memory allocated in the Slow Heap	Words
Total Memory	Fast Heap + Fast Heap B + Slow Heap	Words
Shared Heap	Memory usage from the allocated Shared Heap	Words
Heaps At initialization time, memory to be used by the AWE Core instance for signal processing is allocated. The AWE Core refers to this memory as the heap. By default, AWE Core supports three heaps for which the BSP is responsible for allocating storage. Most commonly, heaps are allocated statically as large arrays. The heaps are: FASTA: storage accessible using the least time FASTB: a secondary bank of fast storage. Useful for memory that can be concurrently accessed with FASTA heaps SLOW: storage usually external and so more slowly accessed SHARED: Storage shared by and accessible to multiple Audio Weaver instances, used for multi-instance or IPC communication To calculate Memory in MB: ((Total Heap Memory) * 4)/1000000

The rest of the profiling window provides profiling information for each individual module and wire (audio buffer) in the running .awd layout:

Profile Block by Block Individual Module & Wire Profiling (All Audio Weaver Instances)

The ‘Top_0’ Module Name line item contains profiling information for the entire .awd layout’s processing. In a multi-instance architecture, aggregate profiling information for each discrete Audio Weaver instance is labeled as ‘Top_<AWE Instance #>’. If the .awd utilizes multiple audio processing threads, aggregate profiling information for each discrete thread is labeled as ‘Top_<AWE Instance #>_<Thread ID>’:

Top Profiling for AWE Instances and Threads

For more information on multi-threading in Audio Weaver, see the Threading and Clock Dividers Application Note.

In a multi-instance architecture, by default the profiling pop-up window displays profiling information for all of the Audio Weaver instances used:

Block by Block Processing: All AWE Instances

To display profiling information for only one specific Audio Weaver instance, you can do so by selecting the desired Audio Weaver instance from the Instance drop down menu in the upper left-hand corner:

Block by Block Profiling: AWE Instance Selection

Block by Block Profiling – Discrete AWE Instance

Peak Profiling

When selecting the Profile Peak real time profiling option, the Peak Profile Window will pop up:

As discussed above, the ‘Sampling Period’ and ‘Test Length’ fields must be entered in order to run the peak profiling. The time unit for both fields are seconds, and the default values are 0.5s sampling period and 10s test length. When ready to start the peak profiling, simply click the ‘Start’ button.

The Peak Profile window will update in real time as the profiling is executed. You can click the Stop button at any time during the profiling, otherwise the peak profiling will complete at the end of the specified length of test:

The ‘Peak vs Average Cycles’ graph at the top of the Peak Profiling pop-up window displays peak and average CPU usage percentages of the processor Audio Weaver is running on. If the running .awd has a multi-instance architecture or contains multi-threading, multiple profiling measurements for each Audio Weaver instance and thread will also display. The x-axis is time in seconds and the y-axis is CPU percentage:

The bottom portion of the Peak Profiling pop-up window displays a Legend for the Peak vs Average Cycles line graph and an Instance List to select which Audio Weaver instance profiling measurements to display:

Manual Profiling

As mentioned earlier in this application note, Audio Weaver also features Manual Profiling, which collects the same exact profiling information as the real-time block by block profiling but in a different manner. Rather than profiling the runtime input audio stream, during manual profiling, real-time processing is halted on the target and all layouts present in the AWD/AWJ signal flow are processed one at a time through tuning commands (pump_layout), for the user specified amount of audio frames. The yielded profiling results may be preferred in some cases, as this decouples discrete layout processing (and thus layout thread priority) from real-time audio device interrupts. This allows for accurate profiling even in cases when the target is not configured for real-time audio, or if a layout is unable to finish executing in the allotted clock cycles (calculated total ticks per block process available).

To Manual Profile an .awd layout, navigate to ‘Tools > Profile Running Layout > Manual Profile Layout’ in the Audio Weaver Designer toolbar while in Tuning Mode:

At the bottom of the Non-RT MHz and Memory Profiling pop-up window, set the Number of audio frames (blocks) to process and an audio file to be used as input audio data:

If no audio file is used as input audio data, the Manual Profiler will use audio data from the specified source selected in the Layout Properties.

Exporting Profiling Data to CSV

All the Audio Weaver profiling utilities allow you to export the profiling data to a comma separated value (.csv) file. All three of the profiling windows (Profile Block by Block, Profile Peak, and Manual Profile Layout) have an ‘Export to File’ Button:

Once the ‘Export to File’ button is clicked, users are prompted to choose a file name and path for the profiling data export file:

The default profiling .csv file name follows this naming convention:

For multi-instance architectures, every Audio Weaver profiling utility enables users to export either profiling data for all instances or individual profiling data for a selected Audio Weaver instance:

Block by Block and Manual Profiling – Instance Export Option

Peak Profiling – Instance Export Options

For block by block and manual profiling of multi-instance architectures, if profiling data for all instances are selected for export, Audio Weaver will generate one aggregate profiling CSV file and individual CSV files for each Audio Weaver instance in the system:

CSV Export for All Audio Weaver Instances

Once exported and saved, users can open the .csv file in Microsoft Excel or equivalent.

The exported block by block and manual profiling .csv files are structured identically to their pop-up windows in Audio Weaver:

Block by Block and Manual Profiling - Exported CSV

The Profile Peak .csv file is structured as followed, where the Peak and Average Cycle CPU% profile values are indexed by the specified time period in seconds:

Profiling Diagrams

Below are some diagrams that further illustrate the real time profiling function in Audio Weaver.

Profiling Details: Single Threaded Profiling

Profiling Details: Total ticks Per Block Process Available

CycleBurner, BiquadLoading, and FIRLoading Modules

The CycleBurner, BiquadLoading, FIRLoading and MemoryLoading Modules intentionally consume processing cycles on the target processor. They are useful for load/stress testing.

CycleBurner Module

The CycleBurner intentionally consumes a user-specified number of clock cycles. You can increase the number of clock cycles used by CycleBurner by adjusting the ‘numCyclesPerBlock’ slider of CycleBurner’s inspector. Using this technique we can increase the number of cycles until we start to hear distortion in the program material that is being processed by the signal flow.

If a higher priority tasks pre-empt the CycleBurner, the execution time of the CycleBurner will remain constant. This assumes that the module is pre-empted and the pre-emption completes before the Cycle Burner completes.

BiquadLoading Module

The BiquadLoading Module simulates the loading of biquad filters. This module is used for CPU load testing purposes and implements a large number of cascade Biquad filters. At instantiation time, you specify the maximum number of filter stages that you would like to simulate. Then at run-time you can vary the number of filters running. This makes the CPU work harder. Internally, the module uses the BiquadCascade module to implement the filtering.

If the BiquadLoading Module is pre-empted, then the execution time of the modules will be increased.

FIRLoading Module

The FIRLoading Module simulates FIR filter loading. This module is used for CPU load testing purposes and implements a large number of FIR filter taps. At instantiation time, you specify the maximum number of filter taps that you would like to simulate. Then at run-time you can vary the number of filter taps. This makes the CPU work harder since more filter taps are running. Internally, the module uses the FIR module to implement the filtering.

If the FIRLoading Module is pre-empted, then the execution time of the modules will be increased.

MemoryLoading Module

This module is used to check the memory bandwidth of the target. At instantiation time you specify the size of the memory buffer (memSize) and in which heap it should be allocated (memHeap). Then at run time, the module writes a block counter value into every value of the array. It repeats this blockWriteCount times per block process. That is, every time the processing function is called, the module performs a total blockWriteCount*memSize memory write operations. All write operations write the current value of the block counter.

Additional Notes on Audio Weaver Profiling

Since system and other interrupts can and sometimes will preempt audio processing, every profiling measurement has a first order low pass filter to mitigate preemption spikes. The low pass filters are designed so that the time for the profiling values to stabilize is less than 5 seconds. The output of these low pass filters represents the average cycle profiling numbers.
In order for profiling to be accurate for embedded targets, the AWE Core relies on a suitable operation to retrieve the system cycle count. This operation can be implemented internally in the AWE Core library configuration only if the method supported by the target is known at compile time. If the cycle count retrieve function is not implemented in the AWE Core library configuration, then the user must supply functions that retrieve the system cycle count as part of the application. The core speed and the units of the retrieved cycle counts must be correctly defined by the user for the profiling to be accurate.

Audio Weaver Server Profiling

While an .awd is running in Audio Weaver, the Audio Weaver Server provides real time CPU and Memory profiling of the entire layout:

Server Profiling on Multi-CPU Targets

AWE Core may be running in multiple threads on several CPUs in some instances. The application is not aware of which CPU the thread is running on by default. This means that without some extra work, the default display of CPU % cannot be trusted for multi CPU targets as the assumption will be that all threads run on a single CPU.

If CPU affinities are enabled on the target, server will display text percentages for each CPU that a layout is running on. The CPU percentage indicator bar will show the load for the CPU which has the highest load. Any CPUs without an associated layout will not be shown.

Setting and getting CPU affinities is done through the AWE Core functions awe_fwSetLayoutCoreAffinity and awe_fwGetLayoutCoreAffinity. These functions are described in the AWE Core API documentation. Customers utilizing AWE Core OS have this feature built in as described here.