[Flexradio] Change to floats in Preview 9
Eric Bob, Ignoring the EQ conundrum, the change to floats in Preview 9 has dropped my CPU utilization on RX to a range from 0.8% to 2.3% where 1.6% seems to be the average with the logger running too. Adding MixW to the mix (a pun!) drove it up to a whopping 6.0% to 8.3% where 7.0% seems to be the average. That is about a 60% decrease from Preview 7 and Preview 8. VERY NICE! Great job! -Tim --- Tim Ellison mailto:[EMAIL PROTECTED] Integrated Technical Services http://www.itsco.com/ Apex, NC USA 919.674.0044 Ext. 25 / 919.674.0045 (FAX) 919.215.6375 - cell PGP public key available at all public KeyServers
Re: [Flexradio] Change to floats in Preview 9, speed up, and EQ
Gerald, Frank, Eric, and I have come to an agreement on what the new EQ will look like. It will not be like a ISO centered RANE lookalike( but not function-alike !) but will provide the necessary shaping so that you do not get this very flat response that sounds so different on TX from that which people are accustomed to (they are accustomed to at least a bit of preemphasis and some other shaping). The new EQ will be 10 bands or less and not work above 6 KHz. We will concentrate on those areas where SSB, AM, and FM needs the shaping. It will be implemented using 512 sample buffers to limit latency to 11 ms. This was NOT that different from the delay through the low frequency filters in the IIR version. Expect this out in preview 10. Your results are consistent with mine. We are taking cache hits 1/2 as often on average and the total memory bandwidth demands are down under 50% from before. Slow off chip (not cache) memory was a big limiting factor before. The use of floats in the optimized FFTW routines more than make up for the slightly loss of speed when the floating point unit is used to do floats/doubles. Many functions automatically promote to doubles so this can be a net loss. In this case, the overwhelming increase in speed in FFTW3 more than makes up for the occasional sin/cos promotion to double and then conversion back to float. Also, we just left the oscillators running as doubles so the phase wrap glitch occurs once a week! On my wife's sempron, with almost no cache, the lowered memory bandwidth demand dropped it from 65% to 25%. Thanks and again, our apologies for not testing the EQ after the change. Bob N4HY Tim Ellison wrote: Eric Bob, Ignoring the EQ conundrum, the change to floats in Preview 9 has dropped my CPU utilization on RX to a range from 0.8% to 2.3% where 1.6% seems to be the average with the logger running too. Adding MixW to the mix (a pun!) drove it up to a whopping 6.0% to 8.3% where 7.0% seems to be the average. That is about a 60% decrease from Preview 7 and Preview 8. VERY NICE! Great job! -Tim --- Tim Ellison mailto:[EMAIL PROTECTED] Integrated Technical Services http://www.itsco.com/ Apex, NC USA 919.674.0044 Ext. 25 / 919.674.0045 (FAX) 919.215.6375 - cell -- AMSAT VP Engineering. Member: ARRL, AMSAT-DL, TAPR, Packrats, NJQRP/AMQRP, QRP ARCI, QCWA, FRC. ARRL SDR Wrk Grp Chairman Laziness is the number one inspiration for ingenuity. Guilty as charged!
Re: [Flexradio] Change to floats in Preview 9, speed up, and EQ
Hi all, audiophiles included (like myself), Just wanted to remind that any HF equipment doesn't deserve to be called SSB transceiver if it has no equalization of the transmitted signal. It is extremely important in SSB DX work when you want to get through the QRM and other noise. As Bob let us understand, SSB, AM and FM DO NEED shaping. We do not need flat frequency response for the best intelligibility with the legally or technically limited powers and bandwidths. In addition to the frequency shaping we need amplitude compression or even clipping. Read this article from 1970's, it is still true: http://kotisivu.dnainternet.net/ahti/sdr-1000/filtclip.pdf . Now with the DSP tools we can make everything in a more elegant and efficient way. Merry Christmas and Happy New Year with the wonderful presents from FlexRadio, 73, Ahti OH2RZ - Original Message - From: Robert McGwier [EMAIL PROTECTED] To: Tim Ellison [EMAIL PROTECTED] Cc: FlexRadio@flex-radio.biz Sent: Friday, December 23, 2005 8:29 PM Subject: Re: [Flexradio] Change to floats in Preview 9, speed up, and EQ Gerald, Frank, Eric, and I have come to an agreement on what the new EQ will look like. It will not be like a ISO centered RANE lookalike( but not function-alike !) but will provide the necessary shaping so that you do not get this very flat response that sounds so different on TX from that which people are accustomed to (they are accustomed to at least a bit of preemphasis and some other shaping). The new EQ will be 10 bands or less and not work above 6 KHz. We will concentrate on those areas where SSB, AM, and FM needs the shaping. It will be implemented using 512 sample buffers to limit latency to 11 ms. This was NOT that different from the delay through the low frequency filters in the IIR version. Expect this out in preview 10. Your results are consistent with mine. We are taking cache hits 1/2 as often on average and the total memory bandwidth demands are down under 50% from before. Slow off chip (not cache) memory was a big limiting factor before. The use of floats in the optimized FFTW routines more than make up for the slightly loss of speed when the floating point unit is used to do floats/doubles. Many functions automatically promote to doubles so this can be a net loss. In this case, the overwhelming increase in speed in FFTW3 more than makes up for the occasional sin/cos promotion to double and then conversion back to float. Also, we just left the oscillators running as doubles so the phase wrap glitch occurs once a week! On my wife's sempron, with almost no cache, the lowered memory bandwidth demand dropped it from 65% to 25%. Thanks and again, our apologies for not testing the EQ after the change. Bob N4HY
