Long compile times are a necessary evil of FPGA code. Even with the vast improvements of Vivado, compile time still ranks as the biggest killer of large project efficiency. As compile times approach 3-4 hours, their successful completion becomes paramount. All too often I find that the Xilinx compiler running on the compile worker has completed successfully however some small communication glitch either between my development machine and the farmer or the farmer and the worker has caused the compile to be lost. It is quite frustrating to know you have a completed bitfile from Xilinx but the NI tools will not perform the final processing steps required to create the lvbitx file. The only solution is to restart the compile costing another 3-4 hours of productivity.
Typical workflow in our company for these large projects is to spend mornings testing and stressing the compile(s) from overnight. Then make any bug fixes and incremental feature improvements and try to start a compile by mid-morning. By mid-afternoon when the compile is complete do the process again so that you can process another build for overnight. If one of the compiles fails because of timing or resource problems, there's nothing that can be done. But if it fails because of glitches in NI's compile wrapper code, that becomes a waste of a half of a day of productivity.
I propose that the current methods for compiling bitfiles be modified. The goal is to improve user productivity. Some of my suggestions include:
Proper unit testing is a key component of large LVFPGA project success. As the number of modules grows, so does the number of units tests increase. We are working on automatic continuous integration methods that will continue to monitor the accuracy of all modules and assemblies throughout the project life cycle. IP Integration node and NI's wrapper around Xilinx IP makes this task more difficult. When our continuous integration machine refreshes the source code from the repository it has to regenerate these nodes in order to compile. The "Regenerate IP Integration Node Support Files..." tool does not always properly update all IP. It seems like the NI wrappers around Xilinx IP are most problematic. The only way a computer can properly update its version of the node is to open the VI and click through the Xilinx IP generation pages to get it to update.
As a rule we do not commit build products into our source code repository. One option would be to submit all the support files generated when first creating the IP integration node so that others users might not have to update. This can become messy and quickly unwielding. The problem arrises when changes to the IP integration node cause more or fewer support files to be generated. This then requires deleting some files from the repository and adding others which is tedious and leads to errors. One compromise would be if the IP integration created one single compact file that contains all the information needed (expect the input vhd, etc. files) that could be committed to the repository. Even better would be to roll this information into the VI itself. Neither of these might be ideal if the support file contents take up a lot of disk space.
Regardless there needs to be a better solution to allow for automated continuous integration testing.
When there are many controls on the front panel of the FPGA, selecting the control from a Read/Write Control node in the host can become a pain. It is one very large list of controls on the front panel of the FPGA. This list has no scrollbar, no browse, or search feature, and no obvious way of grouping controls.
Here is one example of a front panel, and a video showing how long it takes to scroll through the list of contorls.
And here is the video of me scrolling through the controls: http://screencast.com/t/PLzptTwq58aw
There is plenty of room for improvement. Here are just a few ways I think NI could make this better.
Browse and Search
When using a Property Node, or Invoke Node, the very top option is to "Browse..." From here a list of all properties, or methods can be seen in a resizeable window. Here you can also search, and sort alphabetically. The Read/Write Control node could have similar functionality making selection of controls easier.
Front Panel Selection From FPGA
There could be an option for creating a node by selecting the controls on the front panel of the FPGA. A solution that may work today, is to select the controls, then invoke a custom QuickDrop command that creates the node and puts it in the clipboard so it can be pasted in the host VI. If this were to become an option, I'd hope there is a way to combine two nodes into one, by concatenating the controls of one onto the other.
Front Panel Selection From Host
Lets say you already have the Read/Write Control node on the host. There could be an option by right clicking that would open a new window, showing a static image of the front panel of the FPGA, which the user could then click on. This would be great because the developer probably already knows the control they want based on the front panel location. I don't know how possible this is because you could load a bit file which won't have any front panel information.
Easier Grouping of Controls
Right now there is a way to group controls of an FPGA. This feature is never talked about, and doesn't work on dynamic bit files. Here is a discussion where I describe the steps to make controls grouped on the host. Still this isn't supported on all FPGA setups, and you have to conform to a specific naming convention. Why can't controls that are grouped on the front panel, just be grouped in the host?
This idea exchange is really for any kind of improvement to the FPGA control selection.
Hi How about facility of import and export of I/O Label in FPGA-Real time project as shown image instead of manually renaming each I/O
Xilinx log window should use a fixed-width font.
Which of these two string indicators with identical content is easier to read?
Basically I want a VI like open FPGA VI ref which takes a RIO interface and returns a reference, except that it doesn't deploy a reference if one doesn't exist. It would instead pop out a boolean or error if you try to get a reference and there is no bitfile already deployed.
Two use cases I have in mind:
-Imagine if you need a cRIO to start working ASAP so you deploy your bitfile to flash and tell it to run on power-up. You still have to package the exact same bitfile with your RTEXE, even though its already deployed. This increases the size of your RTexe significantly. Lets say you version your RTexe and don't version the FPGA deployed to flash. Depending on what the signature check is, obtaining a reference to your bitfile may cause the "new" bitfile to be redeployed, eliminating the advantage of loading your bitfile onto flash in the first place.
-Imagine if you have a framework like veristand where you need to use a single bitfile in multiple locations which were written by different developers and possibly released at different times. The tools on NI labs (https://decibel.ni.com/content/docs/DOC-35574) help a lot and let you, once you have a reference, confirm the reference has all the interfaces you need to run your code. However, if you need to share references between code modules you must still be sure to obtain it in just one place and then share the reference yourself using a global or FGV.
Having the RIO driver solve this would be very helpful.
How amazing yould it be to have the ability to visualise resource usage on a FPGA target using a similar view to that shown above (courtesy of Windirstat)
I only recently shaved a significant portion off my FPGA usage by finding out that I had a massively oversized FIFO in my code for almost a year without noticing. I feel that this kind of visualisation (with mouse over showing what is actually occupying the space) with differentiation between Registers, LUTs, BRAM, DSPs and so on would greatly aid those of us trying to squeeze as much as possible out of our FPGA designs.
I think providing this information based on the "estimated resource utilisation" i.e. before Xilinx optimises stuff away would be OK. I'm not sure if the final resource utilisation can be mapped as accurately in this way.
It would also be nice to see CLIP utilisation and NI-internal utilisation at a glance as this is apparently hugely different between targets.
We need a way to simply reinterpret the bits in our FPGAs. I currently have a situation where I need to change my SGL values into U32 for the sake of sending data up to the host. Currently, the only way is to make an IP node. That is just silly. We should be able to use the Type Cast simply for the purpose of reinterpreting the bits.
It is time-consuming that we have to compile all LabVIEW FPGA code even if there is tiny little change on FPGA code.
I understand there is sampling probe, Desktop execution node and simulation tools to reduce such time.
Our customer in Japan, would like to use incremental compile function also on LabVIEW.(Please see below)
I agree his opinion.
What do you think?
Application Engineer at National Instruments Japan.
User Lorn has found a brilliant tip for *DRASTICALLY* speeding up FPGA compile times under Windows for PCs with the turbo boost feature. What's more, it's extremely simple to implement.
Please let's see this in future versions of LabVIEW as standard.
A smaller (and cheaper) sbRIO based on the Xilinx Zynq chip. Target size is SO-DIMM form factor (68 x 30 mm (half the area of a credit card), 200 pins). Such a board would be OEM friendly and can be plugged into a product (rather than the current sbRIO offerings that requires the product to be developed around the sbRIO rather than the sbRIO fitting into your product). Also, a Base Board that is (only) used during development. Below is what the proposed sbRIO and Base Board would roughly look like (courtesy of Enclustra FPGA Solutions)
Vision is available under LabVIEW 64 bit and this makes sense since vision can generate very large amounts of data. I think now is the time to bring FPGA over to LabVIEW 64 bit as well. With FPGA systems you can also generate very large data sets. Also with cards like the PCIE 1473R, you have a VISION requiring card that generates lots of data, but it requires FPGA, so you can only use it in LabVIEW 32 bit. This is not a good thing. It has been 5 years since LabVIEW 64 bit has been released it is time to finish moving the addons over to 64 bit.
I've searched but can't see anything similar - please add a method for setting the timeout for FPGA nodes. This includes the 'Open FPGA reference' and FPGA IO nodes.
If you disconnect a cRIO FPGA (e.g. NI 9148) from the network, it takes 20-30s for the IO node or Open FPGA reference to execute. This is really bad for the user experience as if they try to exit their application in this time it may take half a minute for the application to exit. It also means you may have to wait that length of time to realise that your FPGA has disconnected under most use cases (you can obviously have an external watchdog loop to check that the node is executing in a timely manner)
Please allow me to configure the timeouts for these nodes similar to the TCP/UDP or VISA nodes. They are very similar in how they operate to the FPGA nodes (i.e. a hardware device driver which is susceptible to disconnects!) so I don't understand why these have been omitted.
I wouldn't mind having to set the timeout as part of opening the FPGA reference and then internally have it use the same timeout for other IO nodes as follows:
When accessing a FPGA control / indicator from the HOST, a ~property node is used, and the developer has to pick the right control / indicator from a list
When there's lots of items on the list, it can be a pain to scan the list for the right one.
At the moment they are listed in order they were created on the FPGA.
Could they instead be listed in alphabetical order? Or give the developer the choice?
Currently when you build a VI the bit file path is stored as relative (you can see it in the project XML). This means if you change the project location either:
You have to recompile the FPGA to use VI mode or run interactively. It seems the bitfile could be stored as a relative path like all VIs in the projects.
The LabVIEW FPGA module has supported static dispatch of LabVIEW Class types since 2009. This essentially means all class wires must be analyzable and statically determinable at compile-time to a single type of class. However, this class can be a derived class of the original wire type which means, for instance, invoking a dynamic dispatch method can be supported since the compiler knows exactly which function will always be called.
This is not sufficient for many applications. Implementations that require message passing or other more event oriented programming models tend to use enums and flattened bit vectors to pass different pieces of data around on the same wire. All of this packing and unpacking can automatically be handled by the compiler if we can use run-time dynamic dispatch to describe the application.
We call for the LabVIEW FPGA module to add support for true run-time dynamic dispatch to take care of this tedious, annoying, and down-right boring job of figuring out how to pack and unpack bits everywhere. Whose with me?
Many data streams contain information for multiple channels or multiple samples. Today one must pack this data into larger integer types or interleave the data manually into multiple writes to the DMA FIFO API. It would be much simpler if the DMA natively support cluster and array data types. The local FIFO, Memory, and Register APIs already support this; extend it to DMA.
The FPGA compilation results should be copied to a file in the folder with the bitfile. This is needed to track the history of compilation results, especially useful when using source code control. Right now they get overwritten with each recompile.
Adding a Post-build action VI to the FPGA build spec, would also enable something like this.
For debugging, using FPGA VIs in interactive mode can be very valuable. I have, to this day, not been able to find out how LV determines if a bitfile and a VI match.
Therefore whenever I click on the run button for a VI, I'm never quite sure if the bitfile will match or not and often have to wait 1-5 minutes before I can resume working with LabVIEW. This is a very high price to pay for something which I end up cancelling. I would like very much if the IDE would TELL ME that the bitfile and VI don't match before starting a new compilation and thus wasting my time.
This is opposed to a CTRL_Click of the run arrow which explicitly tells the IDE to compile.
I just manually transferred a fairly large LabVIEW FPGA project from one target to another (7965R to 7966R). It would be nice to be able to click on the RIO target in the project and have an option to "Migrate to New FPGA Target" in the context menu. The menu would open a new dialog where you could select the new RIO target and then it is automatically added to the project and populated the VIs, FIFOs, derived clocks, memory blocks, etc. from the original target. The user can choose whether or not to delete the original RIO target.
This would also make it very easy for users to transfer sample code from the LabVIEW Example Finder to the correct FPGA target (insead of having the folder labeled "Move These Files").