Bricasti M7 Native Plugin

[Aiynzahev]

As for performance, you can optimize DSP much better than CPU.

That's a wrong way to say what you meant. The point of DSPs is having special registers which do specific calculations (for example, Fourier transform) in less CPU cycles than a general purpose Intel or AMD CPU. However, Intel has AAX vector extensions which enable much the same calculations to happen - and at a much greater speed any DSP runs on.

So yeah, I will say that Casey is full of it.

I programmed Blackfins for audio for several years at Analog Devices, including several reverb algorithms. My back of the envelope estimate is that the M7 is somewhat more powerful than a single core of any existing Intel processor. If the algorithm can't be divided into parallel threads (and many reverb algorithms can't), then it would be impossible to run this as a native plugin. Even if the algorithm could *barely* fit on a core, running that close to the max CPU of that core would result in stuttering in any real world situation.

The memory access speed of the M7 would also be difficult/impossible to reproduce on current Intel CPUs. Algorithms such as the Lexicon topologies tend to use around 1 to 2 seconds of delay memory, max. This will allow the memory to stay in the cache of modern CPUs. Casey has mentioned that the algorithm uses over 30 seconds of delay memory. This won't fit in the cache of modern CPUs, and would require a lot of swapping with the main memory, which is MUCH slower.

Blackfins also have hardware support for circular buffers, which is how delay lines are built. You can get faster delay lines on software reverbs w/o hardware circular buffer support by using bitwise tricks, but those hardware circular buffers allow for a HUGE number of delay taps to be calculated without breaking a sweat.

Over at Gearslutz, Casey has mentioned that he uses 32x16 bit multiplications where appropriate, as opposed to the 32x32 multiplications that are commonly used for audio. This suggests to me that the 32x16 bit multiplications are being used for delay taps, where a 16 bit coefficient would make no audible difference. It also suggests that Casey is working pretty close to the metal, as opposed to just running generic C code and letting the high power processors deal with the overhead.

Sean Costello

Wow thanks for those details. Puts things in perspective.

[eidenk]

Wrong counter-argument. GPUs are there because gaming industry "made them necessary". Computers were running just fine with non-3D accelerated graphics cards long before nVidia, ATi, Matrox, 3dfx, etc.

The point is you can't remotely get the same kind of 3D performance with a CPU than what you can get with a GPU because GPUs are tailored for 3D so why couldn't a particularly demanding DSP algorithm be run satisfactorily enough only with hardware specifically designed for it?

[valhallasound]

last time i heard Bricasti reverbs had no modulation whatsoever. "real spaces don't modulate, so why should a reverb" kind of thing. i could be wrong though.

This is incorrect. The M7 does have modulation, although I'm not clear on the exact type of modulation. The sound examples made from the hardware sound VERY different than the impulses.

Until you have precise knowledge on algo details of Bricasti vs other reverbs you have to take what Casey said at face value.
IIRC what he said was the memory througput was limiting factor. It seems that Bricasti is a sort of mega-delay with a huge number of individual taps which may go for several seconds. Thus it requires a lot of very fast memory.

well, one doesn't have to know precise details to suggest that it may be hogwash.
this was also being said about Lexicon algos, and yet they're now in plugin form (with imitations having been around for ages before that).

Lexicon algorithms were running on digital hardware that was state of the art in the 1979 to 1985 timeframe. They are about as computationally simple as you can get while still getting a good sound. The Lexicon algorithms use between 1 and 2 seconds of delay memory, which can now fit inside of the cache of an Intel CPU. In fact, you can get several instances on the cache before you run out of cache memory.

i suppose the reasons for keeping it in hardware is purely business-based (effort to port to a new platform vs. potential revenue from that platform, support overheads etc.).

No piracy, no dealing with the whims of Apple/Avid/Steinberg, no complaints about how many instances people can run in their DAW. The hardware will work 20 years from now, while plugins only have a few years that they can run before DAW/OS updates render them unusable. I'm in the middle of my plugin update cycle to deal with Avid's changes. Hardware sounds pretty great right now.

Sean Costello

[Burillo]

Well, it's your choice to think of Casey as liar or not. And I prefer latter.

i didn't say he was a liar, merely a salesman (at least, to some extent). i wouldn't expect otherwise, and i don't blame him for it.

As for performance, you can optimize DSP much better than CPU.

not if your CPU runs even unoptimized code an order of magnitude faster than whatever you come up with on your DSP chip. plus, surprise-surprise - you can also optimize a general purpose CPU as well. even if you hand-craft your DSP chips to do only the task you want them to do, it will still take time to do stuff, and clock speed will still matter the most. even if a particularly complex operation takes 4 cycles as opposed to 20 cycles on x86 chip, running it on a 600MHz custom DSP chip will still be slower than on a modern Core i7. and i'm pretty sure you won't be able to hand-craft, for example, hyper-threading (fake multicore) and out-of-order execution (instruction reordering) an x86 chip has. not saying that M7 has a 600MHz chips, just providing an example that "optimized code" doesn't always perform better than modern CPU.

anyway, the point is, all these talks about "custom DSP chips" being better than a modern CPU are, in most cases, hogwash. mind you, the Bricasti guys say it's not about the CPU but about memory bandwidth, which i find hard to believe as well, but which is at least plausible.

You have GPU's for a reason, it is a processor highly optimised for a task at hand.

i don't have a GPU or rather, it's integrated into my CPU. so there, a general-purpose CPU doing GPU's job.

You don't call NVIDIA and AMD liars for making a separate GPU card, don't you?

no, but that card is basically loads of cores running at less-than-half-of-CPU clock speeds. i.e. the processors themselves aren't that powerful - it's just there's A LOT of them.

[meloco_go]

i don't have a GPU or rather, it's integrated into my CPU. so there, a general-purpose CPU doing GPU's job.

But it is a different architecture core.

no, but that card is basically loads of cores running at less-than-half-of-CPU clock speeds. i.e. the processors themselves aren't that powerful - it's just there's A LOT of them.

Yep, a different ideology.
So why do you think Bricasti can't benefit from using a spealty designed topology?
The major curse of computers is OS. It runs all sorts of stuff which is not essential to audio and only hinder performance.

[izonin]

M7 v.1 doesn't have modulation, v.2 does.

From my experience CPU's can be slow for certain tasks, compared to dedicated hardware. For example, unbiased render engines run up to 20 times faster on top consumer GPU's than on the latest i7's.

[penguinfromdeep]

Great posts from Sean, that should be enough

[Aiynzahev]

on youtube go to zenpro audio and listen to the reverb demoes. They are all quite different imo and very distinct. TC electronic doesn't sound anything like Lexicon or Bricasti.

[Warp69]

M7 v.1 doesn't have modulation, v.2 does.

Both v1 and v2 have modulation.

[antithesist]

I love it when the warpster surfaces. Bad thread title, by the way. I almost crapped in my pants the first time I read it.

[valhallasound]

M7 v.1 doesn't have modulation, v.2 does.

Both v1 and v2 have modulation.

Different modulation types for each, correct? From what I have heard of v1, the modulation sounds Random Hall-ish, or at least not delay length/chorusing modulation.

Sean Costello

[kmonkey]

It doesnt read like he was saying it had identical results, only that he was having 'great success', and thus whether it replicates the M7 exactly or not is moot, for him.

Yeah that's fine regarding moot point but people should then stop think they're havin M7 because it is far from it. Posting and pointing M7 impulse to replace actual M7 isn't the right thing, it just isn't. You can tell then "try with m7 impulses but it's a moot point" so where is the point there?

I am responding in the sense of OP. He asked for M7 plugin. I am not trying to be smartass here. Simply there isn't M7 reverbs in plugin not even "like".

I've not used the M7 but I have heard demos. Hmm I would honestly recommend you try B2. It's got a very dense yet sophisticated and musical sound.

There really isn't much else I can think of right now. AE Phoenix might also be worth considering, though it's not got the dense sound of M7 it's a very sophisticated sound.

Nothing like the M7 though, wow.

B2 seems to be great plug, advanced workhorse too but it is not close to M7 in therms of reality and space simulation. Both 2C reverbs actually can't compete with M7. It's my IMHO obviously. Yes i have tried both. To my ear they are more close to TC VSS3 algorithm but they are just not there. And then...vss3 is far different from M7

[SuperFly76]

I love it when the warpster surfaces. Bad thread title, by the way. I almost crapped in my pants the first time I read it.

Me too. There should at least be a question mark there.

Good thread. I always appreciate when Sean drops by.

[mkdr]

The Bricasti M7 is running six dual-core Analog Devices Blackfin ADSP-BF561 processors that run at 600MHz and have a combined performance of 14400 MMACs (2400 per processor).

[Warp69]

Different modulation types for each, correct? From what I have heard of v1, the modulation sounds Random Hall-ish, or at least not delay length/chorusing modulation.

Correct - the predelay buffer have 8 voices chorusing modulation going into the reverberation system which have what you would define as Random Hall-ish modulation.