Mercurial > libavcodec.hg
diff arm/dsputil_neon_s.S @ 9345:e0a7a6338526 libavcodec
ARM: NEON optimized put_signed_pixels_clamped
| author | conrad |
|---|---|
| date | Sat, 04 Apr 2009 21:02:48 +0000 |
| parents | 9ea1ea6db616 |
| children | 93c20dd3da43 |
line wrap: on
line diff
--- a/arm/dsputil_neon_s.S Sat Apr 04 20:18:58 2009 +0000 +++ b/arm/dsputil_neon_s.S Sat Apr 04 21:02:48 2009 +0000 @@ -273,6 +273,43 @@ pixfunc2 put_ pixels8_y2, _no_rnd, vhadd.u8 pixfunc2 put_ pixels8_xy2, _no_rnd, vshrn.u16, 1 +function ff_put_signed_pixels_clamped_neon, export=1 + vmov.u8 d31, #128 + vld1.64 {d16-d17}, [r0,:128]! + vqmovn.s16 d0, q8 + vld1.64 {d18-d19}, [r0,:128]! + vqmovn.s16 d1, q9 + vld1.64 {d16-d17}, [r0,:128]! + vqmovn.s16 d2, q8 + vld1.64 {d18-d19}, [r0,:128]! + vadd.u8 d0, d0, d31 + vld1.64 {d20-d21}, [r0,:128]! + vadd.u8 d1, d1, d31 + vld1.64 {d22-d23}, [r0,:128]! + vadd.u8 d2, d2, d31 + vst1.64 {d0}, [r1,:64], r2 + vqmovn.s16 d3, q9 + vst1.64 {d1}, [r1,:64], r2 + vqmovn.s16 d4, q10 + vst1.64 {d2}, [r1,:64], r2 + vqmovn.s16 d5, q11 + vld1.64 {d24-d25}, [r0,:128]! + vadd.u8 d3, d3, d31 + vld1.64 {d26-d27}, [r0,:128]! + vadd.u8 d4, d4, d31 + vadd.u8 d5, d5, d31 + vst1.64 {d3}, [r1,:64], r2 + vqmovn.s16 d6, q12 + vst1.64 {d4}, [r1,:64], r2 + vqmovn.s16 d7, q13 + vst1.64 {d5}, [r1,:64], r2 + vadd.u8 d6, d6, d31 + vadd.u8 d7, d7, d31 + vst1.64 {d6}, [r1,:64], r2 + vst1.64 {d7}, [r1,:64], r2 + bx lr + .endfunc + function ff_add_pixels_clamped_neon, export=1 mov r3, r1 vld1.64 {d16}, [r1,:64], r2