STM32U5 SIMD usage (or any other trick to speedup math)

[Linas L]

Hello,
In my application i need to run double for loop for data inside SRAM. and I would love to speed up my program as much as possible.

uint32_t xs = 0, yx = 0, sum =0, i = 0;

for(uint32_t x = 0 ; x<600 ; x++)
{
    for(uint32_t y = 0 ; y<600 ; y++)
    {
        xs+=DATA[i]*x;
        ys+=DATA[i]*y;
        sum+=DATA[i];
        i++;
    }
}

Any insight into how I can make this faster ? Optimization is already at highest setting, and I am getting 33Hz loop speed with data acquisition.
Last resort is going into inline ASM somehow, just don't understand how do I know if compiler is using some particular CPU registers, that could hold registers longer.

[TDK]

You can also get a sense for how optimized a loop is by looking at the generated assembly code. Use that to guide how the code is written.

You can use a pointer to DATA instead of an index. Might save a little.

const uint32_t* ptr = &DATA[0];

...

xs += *ptr * x;

...

++ptr;

[GLASS]

And try also :

• Manual Loop unrolling (see for example or use if possible CMSIS DSP)
• Locate DATA in a SRAM and temporary var in DTCMRAM
• Give a look to FMAC/DMA of STM32U5 if DATA is not 32bits.