More details about the hardware accelerator in stm32h730vb ?

[2217335612]

stm32h730VB，Mathematical acceleration • CORDIC for trigonometric functions acceleration • FMAC: Filter mathematical accelerator

Are those mathematical accelerations fixed-point or floating-point ? 32bit or 64bit ?​

[Danish1]

Cordic is fairly well described in the Reference Manual for stm32h730.

I haven't used it, but a quick glance tells me that the numbers are fixed-point, either q1.31 or q1.15

FMAC again looks to be fixed-point, q1.15 input, q4.22 output

(Edit: On first reading, I only answered cordic)

Hope this helps,

Danish

[2217335612]

Thanks, Danish. So, for the high-accuracy computation, should we use the floating-point unit instead of those hardware accelerators ?

[Tesla DeLorean]

Obviously going to depend on the type of math and computations you plan on doing, and the number space/range in question.

T​he MCU has a pretty well functioning double-precision FPU, which is better suited to non-DSP oriented tasks.

[2217335612]

Thanks, Tesla. But we are more confused now, since we intended to use the double-precision FPU for digital signal processing, and we believe that FPU is born for the digital signal processing......now you told us that FPU is better suited to non-DSP tasks.

[Tesla DeLorean]

Why should you be confused, you know the math you're doing.

Does a double-precision FPU not do real math better?

I'm saying that doing math on +/-1 range numbers (say q1.31 or q1.15) is next to useless for most applications of mathematics, which tend to expect a broader range of values, and intermediate precision.

32-bit floats might be great for DSP, other things they are not, you very quickly hit magnitude and precision limitations that come with a 24-bit mantissa. I might also argue that 64-bit doubles are overkill for "DSP" type applications, but are perfectly good for real math problems. None of these ARM FPU are in the same class as the INTEL/MOTOROLA ones that proceeded them in terms of engineering mathematics.