rv40: NEON optimised weak loop filter

Signed-off-by: Mans Rullgard <mans@mansr.com>

libavcodec/arm/rv40dsp_init_neon.c

@@ -61,6 +61,13 @@ int ff_rv40_v_loop_filter_strength_neon(uint8_t *src, int stride,
                                         int beta, int beta2, int edge,
                                         int *p1, int *q1);
 
+void ff_rv40_h_weak_loop_filter_neon(uint8_t *src, int stride, int filter_p1,
+                                     int filter_q1, int alpha, int beta,
+                                     int lim_p0q0, int lim_q1, int lim_p1);
+void ff_rv40_v_weak_loop_filter_neon(uint8_t *src, int stride, int filter_p1,
+                                     int filter_q1, int alpha, int beta,
+                                     int lim_p0q0, int lim_q1, int lim_p1);
+
 void ff_rv40dsp_init_neon(RV34DSPContext *c, DSPContext* dsp)
 {
     c->put_pixels_tab[0][ 1] = ff_put_rv40_qpel16_mc10_neon;
@@ -126,4 +133,6 @@ void ff_rv40dsp_init_neon(RV34DSPContext *c, DSPContext* dsp)
 
     c->rv40_loop_filter_strength[0] = ff_rv40_h_loop_filter_strength_neon;
     c->rv40_loop_filter_strength[1] = ff_rv40_v_loop_filter_strength_neon;
+    c->rv40_weak_loop_filter[0]     = ff_rv40_h_weak_loop_filter_neon;
+    c->rv40_weak_loop_filter[1]     = ff_rv40_v_weak_loop_filter_neon;
 }

libavcodec/arm/rv40dsp_neon.S

@@ -808,3 +808,113 @@ function ff_rv40_v_loop_filter_strength_neon, export=1
         vmov.u16        r0,  d0[0]
         bx              lr
 endfunc
+
+.macro  rv40_weak_loop_filter
+        vdup.16         d30, r2                 @ filter_p1
+        vdup.16         d31, r3                 @ filter_q1
+        ldrd            r2,  r3,  [sp]
+        vdup.16         d28, r2                 @ alpha
+        vdup.16         d29, r3                 @ beta
+        ldr             r12, [sp, #8]
+        vdup.16         d25, r12                @ lim_p0q0
+        ldrd            r2,  r3,  [sp, #12]
+        vsubl.u8        q9,  d5,  d4            @ x, t
+        vabdl.u8        q8,  d5,  d4            @ x, abs(t)
+        vneg.s16        q15, q15
+        vceq.i16        d16, d19, #0            @ !t
+        vshl.s16        d19, d19, #2            @ t << 2
+        vmul.u16        d18, d17, d28           @ alpha * abs(t)
+        vand            d24, d30, d31           @ filter_p1 & filter_q1
+        vsubl.u8        q1,  d0,  d4            @ p1p2, p1p0
+        vsubl.u8        q3,  d1,  d5            @ q1q2, q1q0
+        vmov.i16        d22, #3
+        vshr.u16        d18, d18, #7
+        vadd.i16        d22, d22, d24           @ 3 - (filter_p1 & filter_q1)
+        vsubl.u8        q10, d0,  d1            @ src[-2] - src[1]
+        vcle.u16        d18, d18, d22
+        vand            d20, d20, d24
+        vneg.s16        d23, d25                @ -lim_p0q0
+        vadd.s16        d19, d19, d20
+        vbic            d16, d18, d16           @ t && u <= 3 - (fp1 & fq1)
+        vtrn.32         d4,  d5                 @ -3,  2, -1,  0
+        vrshr.s16       d19, d19, #3
+        vmov            d28, d29                @ beta
+        vswp            d3,  d6                 @ q1q2, p1p0
+        vmin.s16        d19, d19, d25
+        vand            d30, d30, d16
+        vand            d31, d31, d16
+        vadd.s16        q10, q1,  q3            @ p1p2 + p1p0, q1q2 + q1q0
+        vmax.s16        d19, d19, d23           @ diff
+        vabs.s16        q1,  q1                 @ abs(p1p2), abs(q1q2)
+        vand            d18, d19, d16           @ diff
+        vcle.u16        q1,  q1,  q14
+        vneg.s16        d19, d18                @ -diff
+        vdup.16         d26, r3                 @ lim_p1
+        vaddw.u8        q2,  q9,  d5            @ src[-1]+diff, src[0]-diff
+        vhsub.s16       q11, q10, q9
+        vand            q1,  q1,  q15
+        vqmovun.s16     d4,  q2                 @ -1,  0
+        vand            q9,  q11, q1
+        vdup.16         d27, r2                 @ lim_q1
+        vneg.s16        q9,  q9
+        vneg.s16        q14, q13
+        vmin.s16        q9,  q9,  q13
+        vtrn.32         d0,  d1                 @ -2,  1,  -2,  1
+        vmax.s16        q9,  q9,  q14
+        vaddw.u8        q3,  q9,  d0
+        vqmovun.s16     d5,  q3                 @ -2,  1
+.endm
+
+function ff_rv40_h_weak_loop_filter_neon, export=1
+        sub             r0,  r0,  r1,  lsl #1
+        sub             r0,  r0,  r1
+
+        vld1.32         {d4[]},   [r0,:32], r1
+        vld1.32         {d0[]},   [r0,:32], r1
+        vld1.32         {d4[1]},  [r0,:32], r1
+        vld1.32         {d5[]},   [r0,:32], r1
+        vld1.32         {d1[]},   [r0,:32], r1
+        vld1.32         {d5[0]},  [r0,:32]
+
+        sub             r0,  r0,  r1,  lsl #2
+
+        rv40_weak_loop_filter
+
+        vst1.32         {d5[0]},  [r0,:32], r1
+        vst1.32         {d4[0]},  [r0,:32], r1
+        vst1.32         {d4[1]},  [r0,:32], r1
+        vst1.32         {d5[1]},  [r0,:32], r1
+
+        bx              lr
+endfunc
+
+function ff_rv40_v_weak_loop_filter_neon, export=1
+        sub             r12, r0,  #3
+        sub             r0,  r0,  #2
+
+        vld1.8          {d4},     [r12], r1
+        vld1.8          {d5},     [r12], r1
+        vld1.8          {d2},     [r12], r1
+        vld1.8          {d3},     [r12], r1
+
+        vtrn.16         q2,  q1
+        vtrn.8          d4,  d5
+        vtrn.8          d2,  d3
+
+        vrev64.32       d5,  d5
+        vtrn.32         q2,  q1
+        vdup.32         d0,  d3[0]
+        vdup.32         d1,  d2[0]
+
+        rv40_weak_loop_filter
+
+        vtrn.32         q2,  q3
+        vswp            d4,  d5
+
+        vst4.8          {d4[0],d5[0],d6[0],d7[0]}, [r0], r1
+        vst4.8          {d4[1],d5[1],d6[1],d7[1]}, [r0], r1
+        vst4.8          {d4[2],d5[2],d6[2],d7[2]}, [r0], r1
+        vst4.8          {d4[3],d5[3],d6[3],d7[3]}, [r0], r1
+
+        bx              lr
+endfunc