Efficient way to convert 32 bit float to 32 bit int (SSE)

Hi, thank you all for the interesting discussion posts on denorms and fixed-point/floating-point processing. I have a problem that is very much related to the arguments posted by B.J., mentioning the lack of saturation arithmetics on x86/x64 processors. I need to convert a batch of 32 bit float

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Moin, Seems the problem only exists for positive numbers where some lower bits are zero instead of one. I would suggest to clip and convert first, then construct a mask to be or-ed to the result using the original float. Positive comparison value should have 24 bits (23 + implicit) to match flo

Re: Rendering note with dynamic low-pass filter by cross fading

Without time to listen to your example Yisheng I'm just voicing a theoretical/academic take; When it comes to LTI (linear time invariant) systems order doesn't matter. With time variant systems it does. As a very concrete practical example, I made some "rain" weather FX patches that employ short e

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Hello, I'm no SSE expert either but I would exploit IEEE 754r single precision floating point representation. Essentially you have that 0x4f00 represents 2147483648.f while 0x4eff represents 2147483520.f. OTOH, in 2's complement 32 bits, 0x7fff is 2147483647 and 0x8000 is -21

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Stefano’s solution is elegant because it exploits the fact that values outside the range are all set to 0x8000. But the implementation is a bit overcomplicated, this works as well with less instructions, same result: int main() { const __m128 sseFloatInput = _mm_set_ps(1000.f, -1000.f, 3

Re: Are there now 64-bit processors that deal with denorms routinely with no exception or interrupt?

I'm also very much emjoying this thread and realise that the focus is on microprocesor ALU behaviour. But to widen it a little, let's remember that compilers and languages are in many senses inseperable from silicon in modern computing. Returning to the spirit of the OP - that surely in 2023 we s

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Sorry for spamming, but I am obsessive about optimisations and cannot spare you the version with one less instruction: int main() { const __m128 sseFloatInput = _mm_set_ps(1000.f, -1000.f, 30.f, -30.f); const __m128 fcmp= _mm_set_ps(0x0FFp3f,0x0FFp3f,0x0F

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Yeah, Stefan's version is easier/better. It only needs an extra _mm_castps_si128() to compute m, which costs nothing. Best, Stefano D'Angelo Il 26/04/23 10:42, Stefan Stenzel ha scritto: Sorry for spamming, but I am obsessive about optimisations and cannot spare you the version with one less

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

That code snippet would be a good addition to the musicdsp source code archive: https://www.musicdsp.org/en/latest/Other/index.html Best, Steffan > On 26. Apr 2023, at 10:50, Stefano D'Angelo > wrote: > > Yeah, Stefan's version is easier/better. > > It only needs an extra _mm_castps_si12

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

OK, but first needs some bug fixing, here a corrected version with the proper constant for comparison: //src/dst must be aligned void float2intx4(__m128 *src, __m128i *dst) { const __m128 fcmp = _mm_set_ps1(0x00FFp7f); const __m128 sseFloatInput = *src; __m128i x = _mm_cvtp

Re: Efficient way to convert 32 bit float to 32 bit int (SSE)

Sorry, one more fix, I’ll keep silent now. //src/dst must be aligned void float2intx4(__m128 *src, __m128i *dst) { const __m128 fcmp = _mm_set_ps1(0x00FFp7f); const __m128 sseFloatInput = *src; __m128i x = _mm_cvtps_epi32(sseFloatInput); __m128i m = _mm_cmpgt_ps(sseFloat

AW: Efficient way to convert 32 bit float to 32 bit int (SSE)

Thank you very much for the quick answers, Stefan, Stefano and Steffan (what a coincidental match... :-)). That was very helpful. I had to add one more (no-op) instruction to get the types correctly (may be related to the Microsoft compiler, not sure): void float2intx4(__m128 *src, __m128i *dst

Re: Rendering note with dynamic low-pass filter by cross fading

Hey thanks for the feedback. I was worried it might not be the same effect.. By recalc the transform function I meant computing the a0, a1, b0 etc, which takes a large chunk of my rendering cycle.. In any case I will generate some wav files and see what the outcome is.. Sent from my iPhone > O

Re: Rendering note with dynamic low-pass filter by cross fading

Sorry I was unclear in my original email. This is separate from the amplitude envelope.. I’m talking about a modulating envelope that’s only connected to the resonant frequency of the filter.. it shifts it by about 1-semitone over the course of 1 second or so.. Sent from my iPhone > On Apr 26

Re: Rendering note with dynamic low-pass filter by cross fading

A less mathematical way to show why this doesn't quite work: If you have a resonant filter and you want to modulate it from frequency A to frequency B, you would expect that resonant peak to "sweep" from one frequency to the next. If you instead use two filters/streams and crossfade between them, y

Re: Rendering note with dynamic low-pass filter by cross fading

Yeah that’s a good way to look at it.Sent from my iPhoneOn Apr 26, 2023, at 11:16 AM, N G wrote:A less mathematical way to show why this doesn't quite work: If you have a resonant filter and you want to modulate it from frequency A to frequency B, you would expect that resonant peak to "sweep" fr

Re: Are there now 64-bit processors that deal with denorms routinely with no exception or interrupt?

The only place I’ve seen arbitrary precision arithmetic used in audio is in coding. If you’re doing arithmetic coding or FPC or similar, you need to do arithmetic operations on words of large, potentially variable size (hundreds of bits, say). This is still fixed point, and is not really in the