Conversation

Notices

1. Embed this notice
   CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 04:47:00 JST CrunkLord420
   libfixmath can do decimal multiplication 5.5x faster than the built in FPU.
   
   It takes 4ms to do 1024 floating point divisions on a ESP32 S3. Wew lad. libfixmath can do it in 900us.
   • Embed this notice
     Pawlicker (purpcat@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 04:52:34 JST Pawlicker

     in reply to
     @crunklord420 reminds me of ia64 and how the software x86 emu beat the hardware x86 instruction decoder block
     Then they removed it on later chips and it broke certain things 💀
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 04:52:34 JST CrunkLord420

     in reply to

     • Pawlicker
     @PurpCat I mean the Espressif documents specifically warn how slow floating point division is. It's not surprising that fixed point math is faster. But the ESP32 does advertise itself as having a FPU, it's just not very fast.
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 05:24:56 JST CrunkLord420

     in reply to

     • THOT POLICE
     • Pawlicker
     @s8n @PurpCat then why is it?
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 10-Sep-2025 05:24:58 JST THOT POLICE

     in reply to

     • Pawlicker
     @crunklord420 @PurpCat libfixmath should not be faster than the FPU
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 05:35:25 JST CrunkLord420

     in reply to

     • THOT POLICE
     • Pawlicker
     @s8n @PurpCat
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 10-Sep-2025 05:35:26 JST THOT POLICE

     in reply to

     • Pawlicker
     @crunklord420 @PurpCat I'm looking into it, my theory is there is no actual FPU and the cpu on the SoC just has support for it
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 05:37:22 JST CrunkLord420

     in reply to

     • THOT POLICE
     • Pawlicker
     @s8n @PurpCat I don't know why you say that since the datasheet says it has it.
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 10-Sep-2025 05:37:23 JST THOT POLICE

     in reply to

     • Pawlicker
     @crunklord420 @PurpCat idk what's in that screenshot, but "you don't have an FPU" is my final answer. If it's stupid then that's what you get
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Wednesday, 10-Sep-2025 05:45:01 JST CrunkLord420

     in reply to

     • THOT POLICE
     • Pawlicker
     @s8n @PurpCat https://docs.espressif.com/projects/esp-idf/en/v5.5.1/esp32s3/api-reference/system/freertos_idf.html#misc
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 10-Sep-2025 05:45:02 JST THOT POLICE

     in reply to

     • Pawlicker
     @crunklord420 @PurpCat in the datasheet for the part it does in fact say that, but it also says this and an FPU is not anywhere in the block diagram for the part that's actually in the esp32
   • Embed this notice
     CrunkLord420 (crunklord420@clubcyberia.co)'s status on Thursday, 11-Sep-2025 01:59:55 JST CrunkLord420

     in reply to

     • Captain Arepa
     @captain_arepa well fixed precision and actual floating point aren't exactly the same thing. You can notice significant precision issues if you don't control your range.
   • Embed this notice
     Captain Arepa (captain_arepa@pl.cachapa.cc)'s status on Thursday, 11-Sep-2025 01:59:56 JST Captain Arepa

     in reply to
     @crunklord420
     
     > faster than the built in FPU.
     > It takes 4ms to do 1024 floating point divisions on a ESP32 S3.
     > libfixmath can do it in 900us.
     
     So, around a 75% time reduction... How the fuck? :akko_thonk: