ESP32-S3 throws IllegalInstruction-error for dsps_dotprod_f32 due to (ASM) dsps_dotprod_f32_ae32?

[GiorgosXou]

Describe the bug
Based on this macro-condition dsps_dotprod_f32 defaults to dsps_dotprod_f32_ae32 for ESP32-S3. However, (despite the fact that it compiles) it throws IllegalInstruction-error in log, therefore it doesn't work. Unfortunately, I don't have an ESP32-S3 to verify this 100%, but as far as I understand this should be a wokwi issue.

To Reproduce
Just test this wokwi example here:

#include "esp_dsp.h"

void setup() {
  // put your setup code here, to run once:
  Serial.begin(115200);
  Serial.println("Hello, ESP32-S3!");
  
  float x[]  = {1,2,3,4};
  float y[]  = {5,4,3,2};
  float dest = 1;

  dsps_dotprod_f32(x, y, &dest, 4); // FAILS (ASM)
  // dsps_dotprod_f32_ae32(x, y, &dest, 4); // FAILS (ASM)
  // dsps_dotprod_f32_ansi(x, y, &dest, 4); // WORKS (C-CODE)
  
  Serial.println(dest);
}

void loop() {}

Expected behavior
Based on this example, dsps_dotprod_f32 should work just like the ...ansi-version works without any error.

Environment (please complete the following information):

• OS: Arch Linux
• Browser: Ungoogled-chromium
• Version: -

Additional context
As said, unfortunately: I don't have an ESP32-S3 to verify this 100% (an order I had never came), but as far as I understand this should be a wokwi issue.

[urish]

Thanks for reporting!

I looked into this, and it seems that the DSP code uses one of the ESP32-S3 PIE (Processor Instruction Extensions), ee.ldf.128.ip, which is not currently implemented in Wokwi.

Out of curiosity, what project are you working on?

[GiorgosXou]

Out of curiosity, what project are you working on?

Thanks for asking. I'm working on my NeuralNetworks-library and while I was messing around with DSPs and SIMD-instructions for adding extensive support for other CPU-architectures^{^1} + while I was developing RNN support^{^2}, I went back to wokwi to test if everything works fine (since my order never came). And that's how I found this issue, in a nutshell (:

^{[1]: (eg. see issue for arm_dot_prod_f32)
[2]: (for a private-company I'm recently collaborating with [not officially yet])}

[urish]

Thanks for explaining the use case!

Implementing the SIMD + DSP instruction simulation would take a considerable amount of effort. Is this something that would be helpful for your work?

[GiorgosXou]

it'd be cool, but there's no such rush at all, you don't need to. I re-ordered an ESP, so in a few days: (to be more specific about the use case) I'll be able to test a custom function I was planning to make based on those simd instructions on hardware. Stay safe ✌