|
Hello, I'm porting a voice codec to Blackfin DSP (BF535). Until recently I didn't have the Blackfin evaluation board (ADDS- BF535-EZLITE). So I used Visual DSP++ 3.1 for Blackfin to simulate the DSP and evaluate the codec performance. The cycle count that I obtained from simulation was satisfactory. Lately I received the BF535 EZ-KIT Lite board. Using the on-board USB JTAG connection, I launched the same DXE that I used in simulation. It was quite a surprise that now the algorithm takes as many as several times more cycles to perform. Besides that, according to wall clock, encoding a 20 ms frame took about 6 minutes! The same behavior is reproducible with very simple code (say, 50 NOPs that should take 50 cycles), so the problem doesn't depend on the code I run. It looks like that the on-board USB emulator adds an overhead of several "communication" instructions per each "real" instruction, and the clock (according to CYCLES register) doesn't stop when the communication instructions are executed. Could you please explain me what I'm doing wrong? Why the cycle count becomes several times larger with the EZ-KIT? Do I need an emulator to run my software in *actual* real-time? If yes, what model would you recommend? HPPCI-ICE or APEX-ICE? I have a Corelis JTAG PCI board. Can I use it with the EZ-KIT or only ADI emulators will do? Thanks in advance, Igor |
|
|
Blackfin EZ-KIT Lite profiling
Started by ●July 31, 2003
Reply by ●July 31, 20032003-07-31
|
On Thu, 31 Jul 2003, igor_mikh wrote: > I'm porting a voice codec to Blackfin DSP (BF535). > Until recently I didn't have the Blackfin evaluation board (ADDS- > BF535-EZLITE). > So I used Visual DSP++ 3.1 for Blackfin to simulate the DSP and > evaluate > the codec performance. The cycle count that I obtained from simulation > was satisfactory. > > Lately I received the BF535 EZ-KIT Lite board. Using the on-board USB > JTAG > connection, I launched the same DXE that I used in simulation. > It was quite a surprise that now the algorithm takes as many as > several times more cycles to perform. Besides that, according to wall > clock, > encoding a 20 ms frame took about 6 minutes! > > The same behavior is reproducible with very simple code (say, > 50 NOPs that should take 50 cycles), so the problem doesn't depend > on the code I run. It looks like that the on-board USB emulator adds > an overhead of several "communication" instructions per each "real" > instruction, and the clock (according to CYCLES register) doesn't > stop when the communication instructions are executed. > > Could you please explain me what I'm doing wrong? Why the cycle count > becomes several times larger with the EZ-KIT? It sounds like it is single stepping rather than running to a break point. I'm not familiar with the blackfin board or driver software, but you should be able to set it up so that it runs in real time to some break point, and it might even count clock cycles for you too. You'll have to look into the manual and find out how to set up "run to break point". > Do I need an emulator to run my software in *actual* real-time? > If yes, what model would you recommend? HPPCI-ICE or APEX-ICE? Maybe. I thought the usb-jtag was an emulator. If not, that's good for me to know :-) Basicly you need to connect to the jtag port, and the usb processor does just that. It may not allow run to break point though. > I have a Corelis JTAG PCI board. Can I use it with the EZ-KIT > or only ADI emulators will do? At this point ADI hasn't released any info on the blackfin jtag command set, so only adi emulators will work. Patience, persistence, truth, Dr. mike |
Reply by ●August 1, 20032003-08-01
|
On Thu, 31 Jul 2003 Mike Rosing wrote : >On Thu, 31 Jul 2003, igor_mikh wrote: > > > I'm porting a voice codec to Blackfin DSP (BF535). > > Until recently I didn't have the Blackfin evaluation board >(ADDS- > > BF535-EZLITE). > > So I used Visual DSP++ 3.1 for Blackfin to simulate the DSP >and > > evaluate > > the codec performance. The cycle count that I obtained from >simulation > > was satisfactory. > > > > Lately I received the BF535 EZ-KIT Lite board. Using the >on-board USB > > JTAG > > connection, I launched the same DXE that I used in >simulation. > > It was quite a surprise that now the algorithm takes as many >as > > several times more cycles to perform. Besides that, according >to wall > > clock, > > encoding a 20 ms frame took about 6 minutes! > > > > The same behavior is reproducible with very simple code >(say, > > 50 NOPs that should take 50 cycles), so the problem doesn't >depend > > on the code I run. It looks like that the on-board USB >emulator adds > > an overhead of several "communication" instructions per each >"real" > > instruction, and the clock (according to CYCLES register) >doesn't > > stop when the communication instructions are executed. > > > > Could you please explain me what I'm doing wrong? Why the >cycle count > > becomes several times larger with the EZ-KIT? > >It sounds like it is single stepping rather than running to a >break point. >I'm not familiar with the blackfin board or driver software, but >you >should be able to set it up so that it runs in real time to some >break >point, and it might even count clock cycles for you too. You'll >have to >look into the manual and find out how to set up "run to break >point". > > > Do I need an emulator to run my software in *actual* >real-time? > > If yes, what model would you recommend? HPPCI-ICE or >APEX-ICE? > >Maybe. I thought the usb-jtag was an emulator. If not, that's >good for me to know :-) Basicly you need to connect to the jtag >port, >and the usb processor does just that. It may not allow run to >break >point though. > > > I have a Corelis JTAG PCI board. Can I use it with the >EZ-KIT > > or only ADI emulators will do? > >At this point ADI hasn't released any info on the blackfin jtag >command >set, so only adi emulators will work. > >Patience, persistence, truth, >Dr. mike > Hi Igor and Dr.Mike, I am not worked on BF533 but instead worked on ADSP21535 which is similar to BF533 that has less of onchip memory and speed. As Igor says it gives more cycles then it actually gives. U can check one by one as follows.. Check if U are still on the Simulator mode or EZ-KIT mode. If You are in the Simulator mode then there is no question regarding this,its obvious. If you are in EZ-KIT mode then there is some problem sometimes. Try restarting ur system,maybe due to the interaction between the PC and the KIT may be slow. But when You see the actual cycles it should be less.Its better always to see the Cycles Register and go ahead with debugging. If still persists then see if everything is in the internal memory,the code,the data and all the required paramaters which are MIPS crytical. If the code is outside blackfin takes 7 cycles to get them. similar with the data also.This is done using the linker file available for You. Check the manual as to where exactly the data banks has to be assigned and go accordingly. Sometimes it may be required that the data banks should be partitioned into many banks to use the full effeciency of the processor.I mean to use the parallel instructions provided. If this also does not solve ur problem and still U have lots of cycles there there is some problem in ur code. May be the code goes into a loop where it ends after 6 mins. This happens sometimes normally when You dont check for zero condition at the initial start of the loop and try to decrement the variable and feed it to the loop Counter. so what the loop counter gets is FFFF instead of zero.Many do this mistake in trying to optimise the code,even I did :-). As far as problems with break points are concerned yess it may occur due to overlays. but in this case the code is not executed till the next break point,instead goes in the Error handler routine and get stuck there. If ur code runs fine till the next break point then this is not the problem of breatpoint. Anyway it i good to remove all the breakpoint and run or keep less breakpoints. I hope this will solve ur problem. Rgrds, Raj ___________________________________________________ Download the hottest & happening ringtones here! OR SMS: Top tone to 7333 Click here now: http://sms.rediff.com/cgi-bin/ringtone/ringhome.pl |
