swscale/aarch64: use multiply accumulate and shift-right narrow
authorSebastian Pop <spop@amazon.com>
Mon, 9 Dec 2019 14:25:01 +0000 (14:25 +0000)
committerMichael Niedermayer <michael@niedermayer.cc>
Sat, 4 Jan 2020 19:59:31 +0000 (20:59 +0100)
This patch rewrites the innermost loop of ff_yuv2planeX_8_neon to avoid zips and
horizontal adds by using fused multiply adds. The patch also uses ld1r to load
one element and replicate it across all lanes of the vector. The patch also
improves the clipping code by removing the shift right instructions and
performing the shift with the shift-right narrow instructions.

I see 8% difference on an m6g instance with neoverse-n1 CPUs:
$ ffmpeg -nostats -f lavfi -i testsrc2=4k:d=2 -vf bench=start,scale=1024x1024,bench=stop -f null -
before: t:0.014015 avg:0.014096 max:0.015018 min:0.013971
after:  t:0.012985 avg:0.013013 max:0.013996 min:0.012818

Tested with `make check` on aarch64-linux.

Signed-off-by: Sebastian Pop <spop@amazon.com>
Reviewed-by: Clément Bœsch <u@pkh.me>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
libswscale/aarch64/output.S

index 90d3b57..25bf28b 100644 (file)
@@ -38,29 +38,21 @@ function ff_yuv2planeX_8_neon, export=1
         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
-        ldr                 w11, [x10], #4                  // read 2x16-bit coeffs (X, Y) at (filter[j], filter[j+1])
-        zip1                v16.8H, v5.8H, v6.8H            // A,I,B,J,C,K,D,L
-        zip2                v17.8H, v5.8H, v6.8H            // E,M,F,N,F,O,H,P
-        dup                 v7.4S, w11                      // X,Y,X,Y,X,Y,X,Y
-        smull               v18.4S, v16.4H, v7.4H           // A.X I.Y B.X J.Y
-        smull               v20.4S, v17.4H, v7.4H           // E.X M.Y F.X N.Y
-        smull2              v19.4S, v16.8H, v7.8H           // C.X K.Y D.X L.Y
-        smull2              v21.4S, v17.8H, v7.8H           // G.X O.Y H.X P.Y
-        addp                v16.4S, v18.4S, v19.4S          // A.X+I.Y B.X+J.Y C.X+K.Y D.X+L.Y
-        addp                v17.4S, v20.4S, v21.4S          // E.X+M.Y F.X+N.Y F.X+O.Y H.X+P.Y
-        add                 v3.4S, v3.4S, v16.4S            // update val accumulator for part 1
-        add                 v4.4S, v4.4S, v17.4S            // update val accumulator for part 2
+        ld1r                {v7.8H}, [x10], #2              // read 1x16-bit coeff X at filter[j  ] and duplicate across lanes
+        ld1r                {v8.8H}, [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, v8.4H             // val0 += {I,J,K,L} * Y
+        smlal2              v4.4S, v6.8H, v8.8H             // val1 += {M,N,O,P} * Y
         subs                w8, w8, #2                      // tmpfilterSize -= 2
         b.gt                3b                              // loop until filterSize consumed
-        sshr                v3.4S, v3.4S, #19               // val>>19 (part 1)
-        sshr                v4.4S, v4.4S, #19               // val>>19 (part 2)
-        sqxtun              v3.4H, v3.4S                    // clip16(val>>19) (part 1)
-        sqxtun              v4.4H, v4.4S                    // clip16(val>>19) (part 2)
-        mov                 v3.D[1], v4.D[0]                // merge part 1 and part 2
-        uqxtn               v3.8B, v3.8H                    // clip8(val>>19)
-        st1                 {v3.1D}, [x3], #8               // write to destination
-        add                 x7, x7, #8                      // i += 8
+
+        sqshrun             v3.4h, v3.4s, #16               // clip16(val0>>16)
+        sqshrun2            v3.8h, v4.4s, #16               // clip16(val1>>16)
+        uqshrn              v3.8b, v3.8h, #3                // clip8(val>>19)
+        st1                 {v3.8b}, [x3], #8               // write to destination
         subs                w4, w4, #8                      // dstW -= 8
+        add                 x7, x7, #8                      // i += 8
         b.gt                2b                              // loop until width consumed
         ret
 endfunc