I used this harness to test my work for quake3, when converting altivec optimisations to SSE and neon.
Binary dumps can be found in the sister repo: https://github.com/whisperglen/perftesthelpers_testdata. They need to be uncompressed and placed in the same folder as the executable (I use testdata folder).
I implemented these to learn about SIMD on intel and arm, there is no real benefit from running them in a quake3 engine.
Even on the low spec x86 cpu I used (AMD bobcat) I did not notice a significant improvement in fps.
Commited so far:
-rsqrt conversion
-sndmix conversion (S_WriteLinearBlastStereo16 and S_PaintChannelFrom16)
-dotproduct experiments (SSE4.1 is good enough, SSE3 improves when run in a row on the same dataset)
-diffusecolor
-lerpmeshvertexes
-projectdlighttexture
These results are from a desktop with an i7-7700. When running on a laptop i7-6500u the SSE code performs a little better.
Results (lower than 1.0 is an improvement):
| function | x86 | x64 | x64/x86 |
|---|---|---|---|
| rsqrt_math | 1 | 1 | 0.67 |
| rsqrt_q3 | 0.64 | 0.98 | 0.65 |
| rsqrt_sse_precise | 0.54 | 0.80 | 0.53 |
| rsqrt_sse | 0.40 | 0.42 | 0.28 |
| sndmix_scalar | 1 | 1 | 0.67 |
| sndmix_mmx | 0.22 | n/a | n/a |
| sndmix_sseasm | 0.13 | n/a | n/a |
| sndmix_sse | 0.14 | 0.16 | 0.10 |
| dotprod | 1 | 1 | 1.1 |
| dotprod_sse | 1.1 | 0.96 | 1.05 |
| dotprod_sse_dp | 0.64 | 0.86 | 1.03 |
| diffusecolor | 1 | 1 | 0.95 |
| diffusecolor_sse | 0.4 | 0.38 | 0.36 |
| lerpmesh | 1 | 1 | 0.82 |
| lerpmesh_sse | 0.82 | 0.79 | 0.62 |
| projectdlight | 1 | 1 | 0.94 |
| projectdlight_sse | 1.02 | 1.02 | 0.96 |
| sndpaint | 1 | 1 | 0.91 |
| sndpaint_sse | 0.31 | 0.32 | 0.29 |
| function | arm64 |
|---|---|
| rsqrt_math | 1 |
| rsqrt_q3 | 1.53 |
| rsqrt_neon | 0.68 |
| sndmix_scalar | 1 |
| sndmix_neon | 0.15 |
| dotprod | 1 |
| dotprod_neon | 1.03 |
| sndpaint | 1 |
| sndpaint_neon | 0.56 |