Very bad FPU performance

[portasynthinca3]

Hardware:

Board: ESP-WROVER-32 board
Core Installation/update date: 03/mar/2019 (version 1.0.1)
IDE name: Arduino IDE
Flash Frequency: 80 MHz
PSRAM enabled: yes
Upload Speed: 921600
Computer OS: Windows 10

Description:

I've noticed a very bad performance of the FPU unit. The datasheet says it's only able to work with 32-bit floating point numbers. So, the "float" data type. I prepared a sketch below to see which will work faster: the FPU (using the "float" data type) or the CPU (using the "double" data type). I expect that the "float" data type operations will complete faster than the ones using "double". In reality, the complete opposite of this can be seen ("double" data type operations execute almost 2 times faster!). After searching for this problem on the Internet, I saw that this may be caused by the FPU startup procedure. So, I repeated the same code twice, but the difference has got even bigger!

Sketch

void setup() {
  Serial.begin(115200);
  
  float f = 2;
  uint32_t st_f = micros();
  for(uint64_t i = 0; i < 10000000; i++)
    f *= f * (double)f;
  Serial.println(f);
  Serial.println(String(micros() - st_f));
  
  double d = 2;
  uint32_t st_d = micros();
  for(uint64_t i = 0; i < 10000000; i++)
    d *= d * (double)d;
  Serial.println(d);
  Serial.println(String(micros() - st_d));

  Serial.println("---RETRY---");

  f = 2;
  st_f = micros();
  for(uint64_t i = 0; i < 10000000; i++)
    f *= f * (double)f;
  Serial.println(f);
  Serial.println(String(micros() - st_f));
  
  d = 2;
  st_d = micros();
  for(uint64_t i = 0; i < 10000000; i++)
    d *= d * (double)d;
  Serial.println(d);
  Serial.println(String(micros() - st_d));
}

void loop() {
}

Debug Messages:

inf
6925317
inf
4322672
---RETRY---
inf
7134875
inf
3987262

[Erich11]

I'm not sure what you are trying to do.
In each loop, you increase your value by the power of three, starting with 2.

So by using floats, after 5 ops your resulting value would be 1.41348E+73 which is larger than max float value.
Same for double after 7 ops.

Testing with Arduino, ESP32 PICO V4 my results with simple multiplication are:
Float: 77.7 ns per operation
Double: 741 ns per operation

#define loops   10000

void setup() {
  Serial.begin(115200);

  volatile float f, f1;
  volatile double d, d1;
  uint32_t st_f;

  f = 7.31;
  st_f = micros();
  for (uint64_t i=0; i<loops; i++)
    f1 = i*f;
  Serial.println("SPFPO MUL");
  Serial.println(String(f1));
  Serial.println(String(micros() - st_f));


  d = 7.31;
  st_f = micros();
  for (uint64_t i=0; i<loops; i++)
    d1 = i*d;
  Serial.println("DPFPO MUL");
  Serial.println(String(d1));
  Serial.println(String(micros() - st_f));
  
}

SPFPO MUL
73092.69
777
DPFPO MUL
73092.69
7411

Some more tests confirm the results from https://www.esp32.com/viewtopic.php?t=800

Operation	uint32_t	float	double
ADD	46 ns	54 ns	508 ns
MUL	50 ns	54 ns	728 ns
DIV	410 ns	239 ns	2747 ns

[portasynthinca3]

Thank you very much for answering!
With this code I tried to measure the performace of the FPU, and now I understand that's not a correct way for doing this. Again, thank you!