diff options
Diffstat (limited to 'libavcodec/arm/dsputil_neon_s.S')
-rw-r--r-- | libavcodec/arm/dsputil_neon_s.S | 45 |
1 files changed, 45 insertions, 0 deletions
diff --git a/libavcodec/arm/dsputil_neon_s.S b/libavcodec/arm/dsputil_neon_s.S index 3b39d2e75e..c305210126 100644 --- a/libavcodec/arm/dsputil_neon_s.S +++ b/libavcodec/arm/dsputil_neon_s.S @@ -273,6 +273,51 @@ function ff_put_h264_qpel8_mc00_neon, export=1 pixfunc2 put_ pixels8_y2, _no_rnd, vhadd.u8 pixfunc2 put_ pixels8_xy2, _no_rnd, vshrn.u16, 1 +function ff_add_pixels_clamped_neon, export=1 + mov r3, r1 + vld1.64 {d16}, [r1,:64], r2 + vld1.64 {d0-d1}, [r0,:128]! + vaddw.u8 q0, q0, d16 + vld1.64 {d17}, [r1,:64], r2 + vld1.64 {d2-d3}, [r0,:128]! + vqmovun.s16 d0, q0 + vld1.64 {d18}, [r1,:64], r2 + vaddw.u8 q1, q1, d17 + vld1.64 {d4-d5}, [r0,:128]! + vaddw.u8 q2, q2, d18 + vst1.64 {d0}, [r3,:64], r2 + vqmovun.s16 d2, q1 + vld1.64 {d19}, [r1,:64], r2 + vld1.64 {d6-d7}, [r0,:128]! + vaddw.u8 q3, q3, d19 + vqmovun.s16 d4, q2 + vst1.64 {d2}, [r3,:64], r2 + vld1.64 {d16}, [r1,:64], r2 + vqmovun.s16 d6, q3 + vld1.64 {d0-d1}, [r0,:128]! + vaddw.u8 q0, q0, d16 + vst1.64 {d4}, [r3,:64], r2 + vld1.64 {d17}, [r1,:64], r2 + vld1.64 {d2-d3}, [r0,:128]! + vaddw.u8 q1, q1, d17 + vst1.64 {d6}, [r3,:64], r2 + vqmovun.s16 d0, q0 + vld1.64 {d18}, [r1,:64], r2 + vld1.64 {d4-d5}, [r0,:128]! + vaddw.u8 q2, q2, d18 + vst1.64 {d0}, [r3,:64], r2 + vqmovun.s16 d2, q1 + vld1.64 {d19}, [r1,:64], r2 + vqmovun.s16 d4, q2 + vld1.64 {d6-d7}, [r0,:128]! + vaddw.u8 q3, q3, d19 + vst1.64 {d2}, [r3,:64], r2 + vqmovun.s16 d6, q3 + vst1.64 {d4}, [r3,:64], r2 + vst1.64 {d6}, [r3,:64], r2 + bx lr + .endfunc + function ff_float_to_int16_neon, export=1 subs r2, r2, #8 vld1.64 {d0-d1}, [r1,:128]! |