Mcu32
Senior Member

Are the same build flags used?

I told you about this new information a long time ago :) But anyway, I'm glad that a newer toolchain is finally being used. It would be nice to see this for decades old components like the CMSIS or libraries, too.

@ JarkkoL This, for a micro controller quite impressive piece of silicon, is however miles away from a modern desktop CPU. It is also helpful to simply compare the orders of magnitude lower number of transistors or...

Even with branch prediction, a register must be incremented or decremented and a comparison/jump must be performed. This is not without cost. (the jump takes 1 cycle if the prediction is correct. +1 cyle minimum for...

There are no DSP instructions for float. It's integer only. So, for a float multiplication just use "*" (which is pretty fast) :-)

No, it is tied to the DMA buffer (half-full/full) And I'd say yes, it's likely that the code you posted is not the problem.

Well, this is a forum for teensy, and I would guess that it would run on a T4. (btw, the architecture is also known ;-), and is has no "no-overhead loops" like esp32) Frankzappa, don't get confused! Normally the...

you're welcome :)

https://gcc.gnu.org/onlinedocs/gcc/Loop-Specific-Pragmas.html the last

p.s. you know gcc has a #pragma for unrolling? the compiler can do it for you.

It will be not _that_ much faster, because the count of loops is important, too. an unrolled loop has much less jumps and compares. jumps are a bit expensive, too (not as much as a division) But both together is ok :)

1 cycle for a div is not even for 16/32-bit-integer true :) udiv is 3 cycle minimum, in most cases more. a div is expensive.

No, a floating point multiply needs up to 3 cycles, division up to 18. https://www.quinapalus.com/cm7cycles.html

... amazing. something like this has been in demand here for more than a decade :)

...and, best, sequentially. No random access.

Divisions are slow. I have not tried it, but maybe it helps to multiply with 1.0f/max instead: std::max(max, 10.0f); max = 1.0f/max; for (int i = 0; i < m_signalLength; i++) { m_targetSignal *=...

might be the boards are not soldered together.