diff options
| author | Mans Rullgard <mans@mansr.com> | 2011-09-28 14:34:54 +0100 |
|---|---|---|
| committer | Mans Rullgard <mans@mansr.com> | 2011-09-28 15:56:09 +0100 |
| commit | baf6b738f23bbe7db3a8438372d01a958492e477 (patch) | |
| tree | d0a92c1496b69789afcd05ace8924c928afc5867 /libavcodec/arm/dsputil_neon.S | |
| parent | a92a1b93b4888821b2c0f01ab628d8ebaf3f6d61 (diff) | |
| download | ffmpeg-baf6b738f23bbe7db3a8438372d01a958492e477.tar.gz | |
ARM: NEON optimised vector_fmac_scalar()
Signed-off-by: Mans Rullgard <mans@mansr.com>
Diffstat (limited to 'libavcodec/arm/dsputil_neon.S')
| -rw-r--r-- | libavcodec/arm/dsputil_neon.S | 48 |
1 files changed, 48 insertions, 0 deletions
diff --git a/libavcodec/arm/dsputil_neon.S b/libavcodec/arm/dsputil_neon.S index aef8a129ee..13f0699eb2 100644 --- a/libavcodec/arm/dsputil_neon.S +++ b/libavcodec/arm/dsputil_neon.S @@ -587,6 +587,54 @@ NOVFP vdup.32 q8, r2 .unreq len endfunc +function ff_vector_fmac_scalar_neon, export=1 +VFP len .req r2 +VFP acc .req r3 +NOVFP len .req r3 +NOVFP acc .req r2 +VFP vdup.32 q15, d0[0] +NOVFP vdup.32 q15, r2 + bics r12, len, #15 + mov acc, r0 + beq 3f + vld1.32 {q0}, [r1,:128]! + vld1.32 {q8}, [acc,:128]! + vld1.32 {q1}, [r1,:128]! + vld1.32 {q9}, [acc,:128]! +1: vmla.f32 q8, q0, q15 + vld1.32 {q2}, [r1,:128]! + vld1.32 {q10}, [acc,:128]! + vmla.f32 q9, q1, q15 + vld1.32 {q3}, [r1,:128]! + vld1.32 {q11}, [acc,:128]! + vmla.f32 q10, q2, q15 + vst1.32 {q8}, [r0,:128]! + vmla.f32 q11, q3, q15 + vst1.32 {q9}, [r0,:128]! + subs r12, r12, #16 + beq 2f + vld1.32 {q0}, [r1,:128]! + vld1.32 {q8}, [acc,:128]! + vst1.32 {q10}, [r0,:128]! + vld1.32 {q1}, [r1,:128]! + vld1.32 {q9}, [acc,:128]! + vst1.32 {q11}, [r0,:128]! + b 1b +2: vst1.32 {q10}, [r0,:128]! + vst1.32 {q11}, [r0,:128]! + ands len, len, #15 + it eq + bxeq lr +3: vld1.32 {q0}, [r1,:128]! + vld1.32 {q8}, [acc,:128]! + vmla.f32 q8, q0, q15 + vst1.32 {q8}, [r0,:128]! + subs len, len, #4 + bgt 3b + bx lr + .unreq len +endfunc + function ff_butterflies_float_neon, export=1 1: vld1.32 {q0},[r0,:128] vld1.32 {q1},[r1,:128] |