AD535 | BIOS | Booting | Bootloader | C621 | C6211 | C6415 | C671 | C6711 | C6711DSK | C6713 | CCS | Chassaing | COFF | DAT | DM64 | DM642 | DMA | DSK671 | DSK6711 | EDM | EDMA | EMIF | Emulator | EVM | EVM620 | FFT | FIR | GPIO | Halting | HPI | HWI | IDK | JTAG | LDB | LDH | LDW | Linker | LMS | LOG_printf | Matlab | McBSP | MEM_alloc | MIPS | PCI | PCM3003 | Pipeline | Profiling | QDM | Reset | ROM | RTDX | Sampling | SDRAM | Stack | TEB | THS1206 | TMS320C621 | TMS320C6416 | TMS320C6711 | TMS320C6713 | UART | Vector Table | XBUS | XDS560

Hi, we need a efficient routine to calculate the sin and cos of any angle(a 32 bits float value) in a TMS320C6701. The problem is that we call the funtion sin and cos into a C loop. Then, the consecuence is that the compiler doesn´t generate a software pipelined code. We try with a Taylor serie and splitted quadratic interpolation, but the code is too large to be pipelined. We need about 6 digits or more of accuracy. Any advices will be greatly appreciated. Thanks...

______________________________
New Code Sharing Section now Live on DSPRelated.com. Learn about the Reward Program for Contributors here.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

Hi, You can try table lookup method if the memory is not a constraint. regards Kiran >From: ruben <> >Reply-To: >To: >Subject: [c6x] SINE AND COSINE OF ANY ANGLE >Date: Thu, 25 May 2000 19:18:54 +0200 > >Hi, we need a efficient routine to calculate the sin and cos of any >angle(a 32 bits float value) in a TMS320C6701. The problem is that we >call the funtion sin and cos into a C loop. Then, the consecuence is >that the compiler doesn´t generate a software pipelined code. We try >with a Taylor serie and splitted quadratic interpolation, but the code >is too large to be pipelined. We need about 6 digits or more of >accuracy. > >Any advices will be greatly appreciated. >Thanks... ________________________________________________________________________ Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com

(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

On 25 May 2000, at 19:18, ruben wrote: > Hi, we need a efficient routine to calculate the sin and cos of any > angle(a 32 bits float value) in a TMS320C6701. The problem is that we > call the funtion sin and cos into a C loop. Then, the consecuence is > that the compiler doesn´t generate a software pipelined code. We try > with a Taylor serie and splitted quadratic interpolation, but the code > is too large to be pipelined. We need about 6 digits or more of > accuracy. Decimal or binary digits? > Any advices will be greatly appreciated. As I understand it, longer code should not be a problem as long as the pipeline remains full, and does not stall. It doesnt matter that the code doesn't all fit in the pipeline at once. YOu should be able to extract the source for sin and cos from the library that you are using. You'll probably find that its a series expansion. ================================================= Eliot Blennerhassett *:-{)> AudioScience, Inc. (New Zealand Office) 6 Centaurus Rd Christchurch 8002 Mobile: +64 21 1183531 New Zealand Ph/fax: +64 3 3327818 e...@audioscience.com <http://www.audioscience.com> =================================================

______________________________
Start your Android Ice Cream Sandwich development on TI's AM35x Sitara ARM Cortex-A8 processor today.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

Hi! Two things may help: 1) There's a little-known constant you can define - I think it's TI_ENHANCED_MATH_H - which allows the compiler to use float versions of sin and cos. Normally sin & cos are doubles - read the manual on this, it will make quite a difference. 2) Get the archiver out, and extract the source code of sin from the rts library. The archiver is ar6x.exe, the source is rts.src... sin is likely to be in either sin.c or sin.asm. If it's in sin.c, include that function in your own code. Now, the compiler (if you set the optimiser high enough) can in-line the function, so there is no function call. If it's in sin.asm, there may be a C-code version of the function included in the .asm file. Again, you could include taht in your source. With both of these, benchmark the code before and after - it may not hel;p you very much. Remember sin & cos are still big functions, so unless the calling functions are very complex, you'll be spending 99% of your time in sin, and halving the time to do that last 1% might not help... ATB Ralph ******************************************************** Ralph Weir mailto: Hunt Engineering http://www.hunteng.co.uk Direct: +44 (0)1283 819991 Mobile: +44 (0)7703 472329 Fax: +44 (0)1283 819992 Somerset Office: +44 (0)1278 760188 +44 (0)1278 760199 ******************************************************** > -----Original Message----- > From: ruben [mailto:] > Sent: 25 May 2000 18:19 > To: > Subject: [c6x] SINE AND COSINE OF ANY ANGLE > Hi, we need a efficient routine to calculate the sin and cos of any > angle(a 32 bits float value) in a TMS320C6701. The problem is that we > call the funtion sin and cos into a C loop. Then, the consecuence is > that the compiler doesnt generate a software pipelined code. We try > with a Taylor serie and splitted quadratic interpolation, but the code > is too large to be pipelined. We need about 6 digits or more of > accuracy. > > Any advices will be greatly appreciated. > Thanks... > > ------------------------------------------------------------------------ > Tradiant is creating the premiere marketplace for shippers and ocean > carriers to conduct the business of moving containerized freight online. > Our comprehensive suite of online services saves time and money. > Visit Tradiant today. www.tradiant.com > http://click.egroups.com/1/5136/7/_/14737/_/960483692/ > ------------------------------------------------------------------------ > > To Join: Send an email to > > To Post: Send an email to > > To Leave: Send an email to > > Archives: http://www.egroups.com/group/c6x > > Other Groups: http://www.dsprelated.com

______________________________
New Code Sharing Section now Live on DSPRelated.com. Learn about the Reward Program for Contributors here.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

wrote: > > On 25 May 2000, at 19:18, ruben wrote: > > > Hi, we need a efficient routine to calculate the sin and cos of any > > angle(a 32 bits float value) in a TMS320C6701. The problem is that we > > call the funtion sin and cos into a C loop. Then, the consecuence is > > that the compiler doesn´t generate a software pipelined code. We try > > with a Taylor serie and splitted quadratic interpolation, but the code > > > is too large to be pipelined. We need about 6 digits or more of > > accuracy. > > Decimal or binary digits? > > > Any advices will be greatly appreciated. > > As I understand it, longer code should not be a problem as long as the > > pipeline remains full, and does not stall. It doesnt matter that the > code > doesn't all fit in the pipeline at once. I think that he is talking about so called pipeline optimization, that is the compiler is not able to prepare second, third or higher loop iteration before the first is finished. U seem to be talking about runtime pipeline. The problem is that no loop that contains a call or another loop can't be optimized. I have the same problem on 6202. Because for this accuracy we need 32-bit multiplies, the sin(x) takes 6 16-bit multipiles and 8 additions/subtractions per iteration I later actually maanged to schedule 3 iterations in prallel but it still cannot find better than 6 cycles per iteration. With 8 iterations and overhead on 200MHz engine it takes dear 300 ns to get down to 7 digit accuracy. > YOu should be able to extract the source for sin and cos from the > library > that you are using. You'll probably find that its a series expansion. But that's just what he is doing and the resulting code is so long, that it cannot be optimized that way eirher. -- Petr Vicherek Ericsson Mobile Networks Phone: +31 53-450-5287

______________________________
Start your Android Ice Cream Sandwich development on TI's AM35x Sitara ARM Cortex-A8 processor today.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

Hello, Ruben and Kiran Kiran Gunnam wrote: > > Hi, > You can try table lookup method if the memory is not a constraint. > regards > Kiran That's not the approach for 6-digit precision. Besides memory for huge tables, the lookup and linear (or rather quadratic) interpolation would take about as long as the taylor series. Try to make it iterate faster. This can be done by splitting the interval of <0; Pi/2> to two or more smaller pieces and use two different taylors develop around different origin. For instance sin(x) in <0; Pi/4> and cos(Pi/2 - x) for x in <Pi/4; Pi/2>. This might reduce the number of necessary coefficients by half. Then hardcode a shorter expansion. > >From: ruben <> > >Date: Thu, 25 May 2000 19:18:54 +0200 > > > >Hi, we need a efficient routine to calculate the sin and cos of any > >angle(a 32 bits float value) in a TMS320C6701. The problem is that we > >call the funtion sin and cos into a C loop. Then, the consecuence is > >that the compiler doesn´t generate a software pipelined code. We try > >with a Taylor serie and splitted quadratic interpolation, but the code > >is too large to be pipelined. We need about 6 digits or more of > >accuracy. -- Petr Vicherek Ericsson Mobile Networks Phone: +31 53-450-5287

______________________________
New Code Sharing Section now Live on DSPRelated.com. Learn about the Reward Program for Contributors here.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

> >From: ruben <> > >Date: Thu, 25 May 2000 19:18:54 +0200 > > > >Hi, we need a efficient routine to calculate the sin and cos of any > >angle(a 32 bits float value) in a TMS320C6701. The problem is that we > >call the funtion sin and cos into a C loop. Then, the consecuence is > >that the compiler doesn´t generate a software pipelined code. We try > >with a Taylor serie and splitted quadratic interpolation, but the code > >is too large to be pipelined. We need about 6 digits or more of > >accuracy. > > > >Any advices will be greatly appreciated. Yet another idea, if this will be possible. If you really need sin/cos od ANY (random) angle, then try break the code too large to smaller pieces. For instance pre-calculate the sin/cos in one loop using taylor series and then use it in the next loop. If these angles are in some relation, then you should be able to find some iterative formula how to calculate one sin/cos from the previously calculated value. For instance formula: (1) x = k0*x + k1*y; y = k0*y - k1*x; generates orthogonal sine and cosine waves if k0=cos(inc) and k1=sin(inc). This has low THD, but due to inaccuracies amplitude slowly drifts and needs to be adjusted. If you use two formulas like (2) z = K*y - x; x = y; y = x; Where K = 2*cos(inc) then the amplitude stay steady but you have to care to have very good accuracy for K if the increment 'inc' is small to achieve really low THD. For larger 'inc' (1) is the ideal formula, unless you do not care much about frequency deviation since quantized cos(inc) will be always a little off the accurate irrational value. -- Petr Vicherek Ericsson Mobile Networks Phone: +31 53-450-5287

______________________________
New Code Sharing Section now Live on DSPRelated.com. Learn about the Reward Program for Contributors here.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )

Thanks Petr and all who respond to me. Well, actually when the original message was posted I had a solution. The late was for reasons not concerning me. I got to split the loops in smallers ones and with a table for cuadratic interpolation (thanks to Jerry Avins for the idea) I have an acceptable speed (but not as fast as i wish) and accuracy ( ERROR<0.0001). Because there's not a linear relation into angles consecutives (Its a cuadratic relation, with not trivial coeficients), It cant be able to use a recursive formula. And using the code from TI like Ralph Weir suggested can´t be done because this code its not suitable to be inlined and optimized due to serious calls to others functions and its big size. Someone said about the newest version of Code Composer and Compiler Tools...well, It looks like my city is too far away from TI, since always the new issues take three or four months to get it. Ruben. Petr Vicherek wrote: > Yet another idea, if this will be possible. > > If you really need sin/cos od ANY (random) angle, then try > break the code too large to smaller pieces. For instance pre-calculate > the sin/cos in one loop using taylor series and then use it in the > next loop. > > If these angles are in some relation, then you should be able to > find some iterative formula how to calculate one sin/cos from the > previously calculated value. > > For instance formula: > > (1) x = k0*x + k1*y; > y = k0*y - k1*x; > > generates orthogonal sine and cosine waves if k0=cos(inc) and k1=sin(inc). > This has low THD, but due to inaccuracies amplitude slowly drifts > and needs to be adjusted. > > If you use two formulas like > > (2) z = K*y - x; x = y; y = x; > > Where K = 2*cos(inc) > > then the amplitude stay steady but you have to care to have very > good accuracy for K if the increment 'inc' is small to achieve really > low THD. For larger 'inc' (1) is the ideal formula, unless you do not > care much about frequency deviation since quantized cos(inc) will be always > a little off the accurate irrational value. > > -- > Petr Vicherek Ericsson Mobile Networks Phone: +31 53-450-5287 > > ------------------------------------------------------------------------ > Click here for savings: beMANY! > http://click.egroups.com/1/4557/7/_/14737/_/960550370/ > ------------------------------------------------------------------------ > > To Join: Send an email to > > To Post: Send an email to > > To Leave: Send an email to > > Archives: http://www.egroups.com/group/c6x > > Other Groups: http://www.dsprelated.com

______________________________
New Code Sharing Section now Live on DSPRelated.com. Learn about the Reward Program for Contributors here.



(You need to be a member of c6x -- send a blank email to c6x-subscribe@yahoogroups.com )