Lines Matching refs:v0
24 movi v0.4S, #0 // accumulator
27 …fmla v0.4S, v1.4S, v2.4S // accumulator += src[0..3] * f…
30 …faddp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
31 …faddp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
32 st1 {v0.S}[0], [x0], #4 // write accumulator
37 movi v0.4S, #0 // accumulator
42 …fmla v0.4S, v1.4S, v2.4S // accumulator += src[0..3] * f…
43 …fmla v0.4S, v3.4S, v4.4S // accumulator += src[4..7] * f…
46 …faddp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
47 …faddp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
48 st1 {v0.S}[0], [x0], #4 // write accumulator
53 movi v0.4S, #0 // accumulator
56 …smlal v0.4S, v1.4H, v2.4H // accumulator += src[0..3] * f…
59 …addp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
60 …addp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
61 st1 {v0.S}[0], [x0], #4 // write accumulator
66 movi v0.4S, #0 // accumulator
69 …smlal v0.4S, v1.4H, v2.4H // accumulator += src[0..3] * f…
70 …smlal2 v0.4S, v1.8H, v2.8H // accumulator += src[4..7] * f…
73 …addp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
74 …addp v0.4S, v0.4S, v0.4S // pair adding of the 4x32-bit …
75 st1 {v0.S}[0], [x0], #4 // write accumulator