LoongArch: LSX and LASX SIMD implementation. (!83) · Merge requests · Pixman / pixman

Add LoongArch SIMD support. Add LSX and LASX optimizations.

Benchmark results, before is upstream/master 47d3fbe3,

LSX build: ./autogen.sh --disable-lasx && make -j4 LASX build: ./autogen.sh && make -j4

For example, the highest improvement is add_n_888.

./tests/lowlevel-blt-bench add_n_888

before: add_n_8 = L1: 186.07 L2: 200.18 M:198.43 ( 1.41%) HT:161.37 VT:156.22 R:156.65 RT:103.67 ( 654Kops/s)

LSX: add_n_8 = L1:13782.81 L2:21067.23 M:14209.75 ( 98.95%) HT:1712.74 VT:3345.05 R:1661.89 RT:469.35 (2054Kops/s)

LASX: add_n_8 = L1:13034.63 L2:19725.46 M:16530.90 (117.71%) HT:1104.39 VT:2264.26 R:1077.33 RT:442.79 (2020Kops/s)

./test/lowlevel-blt-bench all, 10 iterations:

2.5 GHz LoongArch 3A5000, Linux, 64-bit, MEAN:

LoongArch: LSX and LASX SIMD implementation.