I move from ADSP21992 to ADSP21489. My code in ADSP21992 was written in asm. I used fixed point and my float point.
Then I need more perfomance processor and developing my code in C/С++. I draw attention ADSP21xxx because it have 40bit float point and it's AD.
Unfortunately the compiler don't full support 40bit float point especially because the compiler use the function "Rn = pass Rx" unpredictably.
My algorithms are sensitive to the mantissa width.
Now I deceided to use fixed point. But I want to use asm's construction because the optimization don't work.
Existing built-in function are scanty.
unsigned fract * fract
unsigned fract * unsigned fract
fract + unsigned fract * fract
unsigned fract + unsigned fract * unsigned fract
convert unsigned fract to float
It would super if VDSP++ have almost instructions of the processor
My colleague which work with Code Composer of TI's DSP has almost instructions of the processor as built-in functions (intrisinc).
Also these built-in functions in Code Composer let to use SIMD without the compiler.
And he may reach the maximum perfomance, but I cant' do these because the compiler isn't versatile.