@@ -34,16 +34,15 @@ function ff_yuv2planeX_8_neon, export=1
mov x9, x2 // srcp = src
mov x10, x0 // filterp = filter
3: ldp x11, x12, [x9], #16 // get 2 pointers: src[j] and src[j+1]
+ ldr s7, [x10], #4 // read 2x16-bit coeff X and Y at filter[j] and filter[j+1]
add x11, x11, x7, lsl #1 // &src[j ][i]
add x12, x12, x7, lsl #1 // &src[j+1][i]
ld1 {v5.8H}, [x11] // read 8x16-bit @ src[j ][i + {0..7}]: A,B,C,D,E,F,G,H
ld1 {v6.8H}, [x12] // read 8x16-bit @ src[j+1][i + {0..7}]: I,J,K,L,M,N,O,P
- ld1r {v7.8H}, [x10], #2 // read 1x16-bit coeff X at filter[j ] and duplicate across lanes
- ld1r {v16.8H}, [x10], #2 // read 1x16-bit coeff Y at filter[j+1] and duplicate across lanes
- smlal v3.4S, v5.4H, v7.4H // val0 += {A,B,C,D} * X
- smlal2 v4.4S, v5.8H, v7.8H // val1 += {E,F,G,H} * X
- smlal v3.4S, v6.4H, v16.4H // val0 += {I,J,K,L} * Y
- smlal2 v4.4S, v6.8H, v16.8H // val1 += {M,N,O,P} * Y
+ smlal v3.4S, v5.4H, v7.H[0] // val0 += {A,B,C,D} * X
+ smlal2 v4.4S, v5.8H, v7.H[0] // val1 += {E,F,G,H} * X
+ smlal v3.4S, v6.4H, v7.H[1] // val0 += {I,J,K,L} * Y
+ smlal2 v4.4S, v6.8H, v7.H[1] // val1 += {M,N,O,P} * Y
subs w8, w8, #2 // tmpfilterSize -= 2
b.gt 3b // loop until filterSize consumed
Use scalar times vector multiply accumlate instructions instead of vector times vector to remove the need for replicating load instructions which are slightly slower. On AWS c7g (Graviton 3, Neoverse V1) instances: yuv2yuvX_8_0_512_accurate_neon: 1144.8 987.4 yuv2yuvX_16_0_512_accurate_neon: 2080.5 1869.4 Signed-off-by: Jonathan Swinney <jswinney@amazon.com> --- libswscale/aarch64/output.S | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)