diff options
Diffstat (limited to 'libavutil/arm/float_dsp_neon.S')
-rw-r--r-- | libavutil/arm/float_dsp_neon.S | 24 |
1 files changed, 24 insertions, 0 deletions
diff --git a/libavutil/arm/float_dsp_neon.S b/libavutil/arm/float_dsp_neon.S index 100eb02..d00e59d 100644 --- a/libavutil/arm/float_dsp_neon.S +++ b/libavutil/arm/float_dsp_neon.S @@ -220,3 +220,27 @@ function ff_vector_fmul_add_neon, export=1 2: vst1.32 {q12-q13},[r0,:128]! bx lr endfunc + +function ff_vector_fmul_reverse_neon, export=1 + add r2, r2, r3, lsl #2 + sub r2, r2, #32 + mov r12, #-32 + vld1.32 {q0-q1}, [r1,:128]! + vld1.32 {q2-q3}, [r2,:128], r12 +1: pld [r1, #32] + vrev64.32 q3, q3 + vmul.f32 d16, d0, d7 + vmul.f32 d17, d1, d6 + pld [r2, #-32] + vrev64.32 q2, q2 + vmul.f32 d18, d2, d5 + vmul.f32 d19, d3, d4 + subs r3, r3, #8 + beq 2f + vld1.32 {q0-q1}, [r1,:128]! + vld1.32 {q2-q3}, [r2,:128], r12 + vst1.32 {q8-q9}, [r0,:128]! + b 1b +2: vst1.32 {q8-q9}, [r0,:128]! + bx lr +endfunc |