-/* void ff_rv34_inv_transform_neon(DCTELEM *block); */
-function ff_rv34_inv_transform_neon, export=1
- mov r2, r0
- rv34_inv_transform
- vrshrn.s32 d1, q2, #10 @ (z1 + z2) >> 10
- vrshrn.s32 d0, q1, #10 @ (z0 + z3) >> 10
- vrshrn.s32 d2, q3, #10 @ (z1 - z2) >> 10
- vrshrn.s32 d3, q15, #10 @ (z0 - z3) >> 10
- vst4.16 {d0[0], d1[0], d2[0], d3[0]}, [r2,:64], r1
- vst4.16 {d0[1], d1[1], d2[1], d3[1]}, [r2,:64], r1
- vst4.16 {d0[2], d1[2], d2[2], d3[2]}, [r2,:64], r1
- vst4.16 {d0[3], d1[3], d2[3], d3[3]}, [r2,:64], r1
+/* void rv34_idct_add_c(uint8_t *dst, int stride, DCTELEM *block) */
+function ff_rv34_idct_add_neon, export=1
+ mov r3, r0
+ rv34_inv_transform r2
+ vmov.i16 q12, #0
+ vrshrn.s32 d16, q1, #10 @ (z0 + z3) >> 10
+ vrshrn.s32 d17, q2, #10 @ (z1 + z2) >> 10
+ vrshrn.s32 d18, q3, #10 @ (z1 - z2) >> 10
+ vrshrn.s32 d19, q15, #10 @ (z0 - z3) >> 10
+ vld1.32 {d28[]}, [r0,:32], r1
+ vld1.32 {d29[]}, [r0,:32], r1
+ vtrn.32 q8, q9
+ vld1.32 {d28[1]}, [r0,:32], r1
+ vld1.32 {d29[1]}, [r0,:32], r1
+ vst1.16 {q12}, [r2,:128]! @ memset(block, 0, 16)
+ vst1.16 {q12}, [r2,:128] @ memset(block+16, 0, 16)
+ vtrn.16 d16, d17
+ vtrn.32 d28, d29
+ vtrn.16 d18, d19
+ vaddw.u8 q0, q8, d28
+ vaddw.u8 q1, q9, d29
+ vqmovun.s16 d28, q0
+ vqmovun.s16 d29, q1
+ vst1.32 {d28[0]}, [r3,:32], r1
+ vst1.32 {d28[1]}, [r3,:32], r1
+ vst1.32 {d29[0]}, [r3,:32], r1
+ vst1.32 {d29[1]}, [r3,:32], r1