avfilter/vf_bwdif: Add neon for filter_edge

Adds clip and spatial macros for aarch64 neon
Exports C filter_edge needed for tail fixup of neon code
Adds neon for filter_edge

Signed-off-by: John Cox <jc@kynesim.co.uk>
Signed-off-by: Martin Storsjö <martin@martin.st>

2 files changed, 197 insertions, 0 deletions

diff --git a/libavfilter/aarch64/vf_bwdif_init_aarch64.c b/libavfilter/aarch64/vf_bwdif_init_aarch64.c
index 3ffaa07ab3..e75cf2f204 100644
--- a/libavfilter/aarch64/vf_bwdif_init_aarch64.c
+++ b/libavfilter/aarch64/vf_bwdif_init_aarch64.c
@@ -24,10 +24,29 @@
 #include "libavfilter/bwdif.h"
 #include "libavutil/aarch64/cpu.h"
 
+void ff_bwdif_filter_edge_neon(void *dst1, void *prev1, void *cur1, void *next1,
+                               int w, int prefs, int mrefs, int prefs2, int mrefs2,
+                               int parity, int clip_max, int spat);
+
 void ff_bwdif_filter_intra_neon(void *dst1, void *cur1, int w, int prefs, int mrefs,
                                 int prefs3, int mrefs3, int parity, int clip_max);
 
 
+static void filter_edge_helper(void *dst1, void *prev1, void *cur1, void *next1,
+                               int w, int prefs, int mrefs, int prefs2, int mrefs2,
+                               int parity, int clip_max, int spat)
+{
+    const int w0 = clip_max != 255 ? 0 : w & ~15;
+
+    ff_bwdif_filter_edge_neon(dst1, prev1, cur1, next1, w0, prefs, mrefs, prefs2, mrefs2,
+                              parity, clip_max, spat);
+
+    if (w0 < w)
+        ff_bwdif_filter_edge_c((char *)dst1 + w0, (char *)prev1 + w0, (char *)cur1 + w0, (char *)next1 + w0,
+                               w - w0, prefs, mrefs, prefs2, mrefs2,
+                               parity, clip_max, spat);
+}
+
 static void filter_intra_helper(void *dst1, void *cur1, int w, int prefs, int mrefs,
                                 int prefs3, int mrefs3, int parity, int clip_max)
 {
@@ -52,5 +71,6 @@ ff_bwdif_init_aarch64(BWDIFContext *s, int bit_depth)
         return;
 
     s->filter_intra = filter_intra_helper;
+    s->filter_edge  = filter_edge_helper;
 }
 
diff --git a/libavfilter/aarch64/vf_bwdif_neon.S b/libavfilter/aarch64/vf_bwdif_neon.S
index e288efbe6c..389302b813 100644
--- a/libavfilter/aarch64/vf_bwdif_neon.S
+++ b/libavfilter/aarch64/vf_bwdif_neon.S
@@ -66,6 +66,79 @@
         umlsl2          \a3\().4s, \s1\().8h, \k
 .endm
 
+//      int b = m2s1 - m1;
+//      int f = p2s1 - p1;
+//      int dc = c0s1 - m1;
+//      int de = c0s1 - p1;
+//      int sp_max = FFMIN(p1 - c0s1, m1 - c0s1);
+//      sp_max = FFMIN(sp_max, FFMAX(-b,-f));
+//      int sp_min = FFMIN(c0s1 - p1, c0s1 - m1);
+//      sp_min = FFMIN(sp_min, FFMAX(b,f));
+//      diff = diff == 0 ? 0 : FFMAX3(diff, sp_min, sp_max);
+.macro SPAT_CHECK diff, m2s1, m1, c0s1, p1, p2s1, t0, t1, t2, t3
+        uqsub           \t0\().16b, \p1\().16b, \c0s1\().16b
+        uqsub           \t2\().16b, \m1\().16b, \c0s1\().16b
+        umin            \t2\().16b, \t0\().16b, \t2\().16b
+
+        uqsub           \t1\().16b, \m1\().16b, \m2s1\().16b
+        uqsub           \t3\().16b, \p1\().16b, \p2s1\().16b
+        umax            \t3\().16b, \t3\().16b, \t1\().16b
+        umin            \t3\().16b, \t3\().16b, \t2\().16b
+
+        uqsub           \t0\().16b, \c0s1\().16b, \p1\().16b
+        uqsub           \t2\().16b, \c0s1\().16b, \m1\().16b
+        umin            \t2\().16b, \t0\().16b, \t2\().16b
+
+        uqsub           \t1\().16b, \m2s1\().16b, \m1\().16b
+        uqsub           \t0\().16b, \p2s1\().16b, \p1\().16b
+        umax            \t0\().16b, \t0\().16b, \t1\().16b
+        umin            \t2\().16b, \t2\().16b, \t0\().16b
+
+        cmeq            \t1\().16b, \diff\().16b, #0
+        umax            \diff\().16b, \diff\().16b, \t3\().16b
+        umax            \diff\().16b, \diff\().16b, \t2\().16b
+        bic             \diff\().16b, \diff\().16b, \t1\().16b
+.endm
+
+//      i0 = s0;
+//      if (i0 > d0 + diff0)
+//          i0 = d0 + diff0;
+//      else if (i0 < d0 - diff0)
+//          i0 = d0 - diff0;
+//
+// i0 = s0 is safe
+.macro DIFF_CLIP i0, s0, d0, diff, t0, t1
+        uqadd           \t0\().16b, \d0\().16b, \diff\().16b
+        uqsub           \t1\().16b, \d0\().16b, \diff\().16b
+        umin            \i0\().16b, \s0\().16b, \t0\().16b
+        umax            \i0\().16b, \i0\().16b, \t1\().16b
+.endm
+
+//      i0 = FFABS(m1 - p1) > td0 ? i1 : i2;
+//      DIFF_CLIP
+//
+// i0 = i1 is safe
+.macro INTERPOL i0, i1, i2, m1, d0, p1, td0, diff, t0, t1, t2
+        uabd            \t0\().16b, \m1\().16b, \p1\().16b
+        cmhi            \t0\().16b, \t0\().16b, \td0\().16b
+        bsl             \t0\().16b, \i1\().16b, \i2\().16b
+        DIFF_CLIP       \i0, \t0, \d0, \diff, \t1, \t2
+.endm
+
+.macro PUSH_VREGS
+        stp             d8,  d9,  [sp, #-64]!
+        stp             d10, d11, [sp, #16]
+        stp             d12, d13, [sp, #32]
+        stp             d14, d15, [sp, #48]
+.endm
+
+.macro POP_VREGS
+        ldp             d14, d15, [sp, #48]
+        ldp             d12, d13, [sp, #32]
+        ldp             d10, d11, [sp, #16]
+        ldp             d8,  d9,  [sp], #64
+.endm
+
 .macro LDR_COEFFS d, t0
         movrel          \t0, coeffs, 0
         ld1             {\d\().8h}, [\t0]
@@ -83,6 +156,110 @@ endconst
 
 // ============================================================================
 //
+// void ff_bwdif_filter_edge_neon(
+//      void *dst1,     // x0
+//      void *prev1,    // x1
+//      void *cur1,     // x2
+//      void *next1,    // x3
+//      int w,          // w4
+//      int prefs,      // w5
+//      int mrefs,      // w6
+//      int prefs2,     // w7
+//      int mrefs2,     // [sp, #0]
+//      int parity,     // [sp, #SP_INT]
+//      int clip_max,   // [sp, #SP_INT*2]  unused
+//      int spat);      // [sp, #SP_INT*3]
+
+function ff_bwdif_filter_edge_neon, export=1
+        // Sanity check w
+        cmp             w4, #0
+        ble             99f
+
+// #define prev2 cur
+//     const uint8_t * restrict next2 = parity ? prev : next;
+
+        ldr             w8,  [sp, #0]                   // mrefs2
+
+        ldr             w17, [sp, #SP_INT]              // parity
+        ldr             w16, [sp, #SP_INT*3]            // spat
+        cmp             w17, #0
+        csel            x17, x1, x3, ne
+
+//     for (x = 0; x < w; x++) {
+
+10:
+//        int m1 = cur[mrefs];
+//        int d = (prev2[0] + next2[0]) >> 1;
+//        int p1 = cur[prefs];
+//        int temporal_diff0 = FFABS(prev2[0] - next2[0]);
+//        int temporal_diff1 =(FFABS(prev[mrefs] - m1) + FFABS(prev[prefs] - p1)) >> 1;
+//        int temporal_diff2 =(FFABS(next[mrefs] - m1) + FFABS(next[prefs] - p1)) >> 1;
+//        int diff = FFMAX3(temporal_diff0 >> 1, temporal_diff1, temporal_diff2);
+        ldr             q31, [x2]
+        ldr             q21, [x17]
+        uhadd           v16.16b, v31.16b, v21.16b       // d0 = v16
+        uabd            v17.16b, v31.16b, v21.16b       // td0 = v17
+        ldr             q24, [x2, w6, sxtw]             // m1 = v24
+        ldr             q22, [x2, w5, sxtw]             // p1 = v22
+
+        ldr             q0,  [x1, w6, sxtw]             // prev[mrefs]
+        ldr             q2,  [x1, w5, sxtw]             // prev[prefs]
+        ldr             q1,  [x3, w6, sxtw]             // next[mrefs]
+        ldr             q3,  [x3, w5, sxtw]             // next[prefs]
+
+        ushr            v29.16b, v17.16b, #1
+
+        uabd            v31.16b, v0.16b,  v24.16b
+        uabd            v30.16b, v2.16b,  v22.16b
+        uhadd           v0.16b,  v31.16b, v30.16b       // td1 = q0
+
+        uabd            v31.16b, v1.16b,  v24.16b
+        uabd            v30.16b, v3.16b,  v22.16b
+        uhadd           v1.16b,  v31.16b, v30.16b       // td2 = q1
+
+        umax            v0.16b,  v0.16b,  v29.16b
+        umax            v0.16b,  v0.16b,  v1.16b        // diff = v0
+
+//        if (spat) {
+//            SPAT_CHECK()
+//        }
+//        i0 = (m1 + p1) >> 1;
+        cbz             w16, 1f
+
+        ldr             q31, [x2,  w8, sxtw]
+        ldr             q18, [x17, w8, sxtw]
+        ldr             q30, [x2,  w7, sxtw]
+        ldr             q19, [x17, w7, sxtw]
+        uhadd           v18.16b, v18.16b, v31.16b
+        uhadd           v19.16b, v19.16b, v30.16b
+
+        SPAT_CHECK      v0, v18, v24, v16, v22, v19, v31, v30, v29, v28
+
+1:
+        uhadd           v2.16b,  v22.16b, v24.16b
+
+        // i0 = v2, s0 = v2, d0 = v16, diff = v0, t0 = v31, t1 = v30
+        DIFF_CLIP       v2, v2, v16, v0, v31, v30
+
+//        dst[0] = av_clip(interpol, 0, clip_max);
+        str             q2, [x0], #16
+
+//        dst++;
+//        cur++;
+//    }
+        subs            w4,  w4,  #16
+        add             x1,  x1,  #16
+        add             x2,  x2,  #16
+        add             x3,  x3,  #16
+        add             x17, x17, #16
+        bgt             10b
+
+99:
+        ret
+endfunc
+
+// ============================================================================
+//
 // void ff_bwdif_filter_intra_neon(
 //      void *dst1,     // x0
 //      void *cur1,     // x1