Change #278862

Comments

swscale/loongarch: Enable range_convert optimization for LA
The base implementation has been updated to runtime coeff and offset.
Benchmark results:
  name                             ticks (vs ref)
  chrRangeFromJpeg8_1920_c:        237.9
  chrRangeFromJpeg8_1920_lsx:       39.3 ( 5.91x)
  chrRangeFromJpeg8_1920_lasx:      19.6 (11.95x)
  chrRangeFromJpeg16_1920_c:       240.3
  chrRangeToJpeg8_1920_c:          444.3
  chrRangeToJpeg8_1920_lsx:         63.7 ( 6.36x)
  chrRangeToJpeg8_1920_lasx:        28.9 (15.31x)
  chrRangeToJpeg16_1920_c:         502.9
  lumRangeFromJpeg8_1920_c:        140.3
  lumRangeFromJpeg8_1920_lsx:       29.0 ( 4.83x)
  lumRangeFromJpeg8_1920_lasx:      15.8 ( 8.39x)
  lumRangeFromJpeg16_1920_c:        71.7
  lumRangeToJpeg8_1920_c:          303.4
  lumRangeToJpeg8_1920_lsx:         29.2 (10.34x)
  lumRangeToJpeg8_1920_lasx:        14.8 (20.47x)
  lumRangeToJpeg16_1920_c:         294.1

Changed files

• libswscale/loongarch/swscale.S
• libswscale/loongarch/swscale_init_loongarch.c
• libswscale/loongarch/swscale_loongarch.h