Forum Discussion

ADua0's avatar
ADua0
Icon for Occasional Contributor rankOccasional Contributor
6 years ago

Dsp utilization for arithmetic operations

I am trying doing accumulation and multiplication operations on 8 bit of data. "for(int i =0 ;i<4;i++) { sum[i] += inp1[i] * inp2[i]; } " now both my inp1 and inp2 are 8 bit of data , and...
HRZ's avatar
HRZ
Icon for Frequent Contributor rankFrequent Contributor
6 years ago

In my experience, the OpenCL compiler generates math IP cores based on the size of the result, not the operands. If your result is 20 bits, then you will not get one FMA (Fused Multiply and Add) per DSP. I think the result should be less than 18 bits (or be exactly 32-bit float) to get FMA, since there are two 18x19 multipliers in each DSP. Also, if you are using anything below Quartus/AOC v18.1, do not expect good mapping of math operations to DSPs.