Unified · Split
| @@ -40,8 +40,8 @@ function ff_nv24_to_yuv420p_chroma_neon, export=1 · 39 unmodified lines | |||
| 40 | mov w10, w6 // w10 = w | 40 | mov w10, w6 // w10 = w |
| 41 | 41 | ||
| 42 | 2: | 42 | 2: |
| 43 | ld2 { v0.16b, v1.16b }, [x4], #32 // v0 = U1, v1 = V1 | 43 | ld2 {v0.16b, v1.16b}, [x4], #32 // v0 = U1, v1 = V1 |
| 44 | ld2 { v2.16b, v3.16b }, [x9], #32 // v2 = U2, v3 = V2 | 44 | ld2 {v2.16b, v3.16b}, [x9], #32 // v2 = U2, v3 = V2 |
| 45 | 45 | ||
| 46 | uaddlp v0.8h, v0.16b // pairwise add U1 into v0 | 46 | uaddlp v0.8h, v0.16b // pairwise add U1 into v0 |
| 47 | uaddlp v1.8h, v1.16b // pairwise add V1 into v1 | 47 | uaddlp v1.8h, v1.16b // pairwise add V1 into v1 |
| @@ -51,8 +51,8 @@ function ff_nv24_to_yuv420p_chroma_neon, export=1 · 3 unmodified lines | |||
| 51 | shrn v0.8b, v0.8h, #2 // divide by 4 | 51 | shrn v0.8b, v0.8h, #2 // divide by 4 |
| 52 | shrn v1.8b, v1.8h, #2 // divide by 4 | 52 | shrn v1.8b, v1.8h, #2 // divide by 4 |
| 53 | 53 | ||
| 54 | st1 { v0.8b }, [x0], #8 // store U into dst1 | 54 | st1 {v0.8b}, [x0], #8 // store U into dst1 |
| 55 | st1 { v1.8b }, [x2], #8 // store V into dst2 | 55 | st1 {v1.8b}, [x2], #8 // store V into dst2 |
| 56 | 56 | ||
| 57 | subs w10, w10, #8 | 57 | subs w10, w10, #8 |
| 58 | b.gt 2b | 58 | b.gt 2b |