Conversation

Jarkko Sakkinen

Edited 23 days ago
Doing my first peaces of SIMD to the audio pipeline :-) Both SSE and NEON.

Last time I did anything resembling this was when using FPU to divide 1/w while ALU is linearly interpolating the texture (on Pentium which has also nice and fast ALU pipeline, a leap over 486). This used to be common approach for software rendered perspective corrected texture mapping.
1
0
0
To get more texel throughput from Pentium, the next trick is to preprocess texture bitmap in 8x8 pixel tiles. That will significantly reduce cache misses :-)
1
0
0
For x86 I actually (after investigating) solely base on AVX2. There's no legacy history in this software so I don't care about legacy SIMDs (or want to implement them at least myself).
0
0
0