ARM: NEON optimised H.264 loop filter
authorMåns Rullgård <mans@mansr.com>
Mon, 15 Dec 2008 22:12:44 +0000 (22:12 +0000)
committerMåns Rullgård <mans@mansr.com>
Mon, 15 Dec 2008 22:12:44 +0000 (22:12 +0000)
Originally committed as revision 16148 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/armv4l/dsputil_neon.c
libavcodec/armv4l/h264dsp_neon.S

index a5e241a..adea7a3 100644 (file)
@@ -52,6 +52,15 @@ void ff_put_h264_chroma_mc4_neon(uint8_t *, uint8_t *, int, int, int, int);
 void ff_avg_h264_chroma_mc8_neon(uint8_t *, uint8_t *, int, int, int, int);
 void ff_avg_h264_chroma_mc4_neon(uint8_t *, uint8_t *, int, int, int, int);
 
 void ff_avg_h264_chroma_mc8_neon(uint8_t *, uint8_t *, int, int, int, int);
 void ff_avg_h264_chroma_mc4_neon(uint8_t *, uint8_t *, int, int, int, int);
 
+void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, int stride, int alpha,
+                                     int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, int stride, int alpha,
+                                     int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
+                                       int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
+                                       int beta, int8_t *tc0);
+
 void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
 {
     c->put_pixels_tab[0][0] = ff_put_pixels16_neon;
 void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
 {
     c->put_pixels_tab[0][0] = ff_put_pixels16_neon;
@@ -84,4 +93,9 @@ void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
     c->put_h264_qpel_pixels_tab[1][0] = ff_put_h264_qpel8_mc00_neon;
 
     c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
     c->put_h264_qpel_pixels_tab[1][0] = ff_put_h264_qpel8_mc00_neon;
 
     c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
+
+    c->h264_v_loop_filter_luma = ff_h264_v_loop_filter_luma_neon;
+    c->h264_h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
+    c->h264_v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
+    c->h264_h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
 }
 }
index ca32770..63f93e1 100644 (file)
@@ -296,3 +296,294 @@ function ff_put_h264_chroma_mc4_neon, export=1
 function ff_avg_h264_chroma_mc4_neon, export=1
         h264_chroma_mc4 avg=1
         .endfunc
 function ff_avg_h264_chroma_mc4_neon, export=1
         h264_chroma_mc4 avg=1
         .endfunc
+
+        /* H.264 loop filter */
+
+        .macro h264_loop_filter_start
+        ldr             ip,  [sp]
+        tst             r2,  r2
+        ldr             ip,  [ip]
+        tstne           r3,  r3
+        vmov.32         d24[0], ip
+        and             ip,  ip,  ip, lsl #16
+        bxeq            lr
+        ands            ip,  ip,  ip, lsl #8
+        bxlt            lr
+        .endm
+
+        .macro align_push_regs
+        and             ip,  sp,  #15
+        add             ip,  ip,  #32
+        sub             sp,  sp,  ip
+        vst1.64         {d12-d15}, [sp,:128]
+        sub             sp,  sp,  #32
+        vst1.64         {d8-d11},  [sp,:128]
+        .endm
+
+        .macro align_pop_regs
+        vld1.64         {d8-d11},  [sp,:128]!
+        vld1.64         {d12-d15}, [sp,:128], ip
+        .endm
+
+        .macro h264_loop_filter_luma
+        vdup.8          q11, r2         @ alpha
+        vmovl.u8        q12, d24
+        vabd.u8         q6,  q8,  q0    @ abs(p0 - q0)
+        vmovl.u16       q12, d24
+        vabd.u8         q14, q9,  q8    @ abs(p1 - p0)
+        vsli.16         q12, q12, #8
+        vabd.u8         q15, q1,  q0    @ abs(q1 - q0)
+        vsli.32         q12, q12, #16
+        vclt.u8         q6,  q6,  q11   @ < alpha
+        vdup.8          q11, r3         @ beta
+        vclt.s8         q7,  q12, #0
+        vclt.u8         q14, q14, q11   @ < beta
+        vclt.u8         q15, q15, q11   @ < beta
+        vbic            q6,  q6,  q7
+        vabd.u8         q4,  q10, q8    @ abs(p2 - p0)
+        vand            q6,  q6,  q14
+        vabd.u8         q5,  q2,  q0    @ abs(q2 - q0)
+        vclt.u8         q4,  q4,  q11   @ < beta
+        vand            q6,  q6,  q15
+        vclt.u8         q5,  q5,  q11   @ < beta
+        vand            q4,  q4,  q6
+        vand            q5,  q5,  q6
+        vand            q12, q12, q6
+        vrhadd.u8       q14, q8,  q0
+        vsub.i8         q6,  q12, q4
+        vqadd.u8        q7,  q9,  q12
+        vhadd.u8        q10, q10, q14
+        vsub.i8         q6,  q6,  q5
+        vhadd.u8        q14, q2,  q14
+        vmin.u8         q7,  q7,  q10
+        vqsub.u8        q11, q9,  q12
+        vqadd.u8        q2,  q1,  q12
+        vmax.u8         q7,  q7,  q11
+        vqsub.u8        q11, q1,  q12
+        vmin.u8         q14, q2,  q14
+        vmovl.u8        q2,  d0
+        vmax.u8         q14, q14, q11
+        vmovl.u8        q10, d1
+        vsubw.u8        q2,  q2,  d16
+        vsubw.u8        q10, q10, d17
+        vshl.i16        q2,  q2,  #2
+        vshl.i16        q10, q10, #2
+        vaddw.u8        q2,  q2,  d18
+        vaddw.u8        q10, q10, d19
+        vsubw.u8        q2,  q2,  d2
+        vsubw.u8        q10, q10, d3
+        vrshrn.i16      d4,  q2,  #3
+        vrshrn.i16      d5,  q10, #3
+        vbsl            q4,  q7,  q9
+        vbsl            q5,  q14, q1
+        vneg.s8         q7,  q6
+        vmovl.u8        q14, d16
+        vmin.s8         q2,  q2,  q6
+        vmovl.u8        q6,  d17
+        vmax.s8         q2,  q2,  q7
+        vmovl.u8        q11, d0
+        vmovl.u8        q12, d1
+        vaddw.s8        q14, q14, d4
+        vaddw.s8        q6,  q6,  d5
+        vsubw.s8        q11, q11, d4
+        vsubw.s8        q12, q12, d5
+        vqmovun.s16     d16, q14
+        vqmovun.s16     d17, q6
+        vqmovun.s16     d0,  q11
+        vqmovun.s16     d1,  q12
+        .endm
+
+function ff_h264_v_loop_filter_luma_neon, export=1
+        h264_loop_filter_start
+
+        vld1.64         {d0, d1},  [r0,:128], r1
+        vld1.64         {d2, d3},  [r0,:128], r1
+        vld1.64         {d4, d5},  [r0,:128], r1
+        sub             r0,  r0,  r1, lsl #2
+        sub             r0,  r0,  r1, lsl #1
+        vld1.64         {d20,d21}, [r0,:128], r1
+        vld1.64         {d18,d19}, [r0,:128], r1
+        vld1.64         {d16,d17}, [r0,:128], r1
+
+        align_push_regs
+
+        h264_loop_filter_luma
+
+        sub             r0,  r0,  r1, lsl #1
+        vst1.64         {d8, d9},  [r0,:128], r1
+        vst1.64         {d16,d17}, [r0,:128], r1
+        vst1.64         {d0, d1},  [r0,:128], r1
+        vst1.64         {d10,d11}, [r0,:128]
+
+        align_pop_regs
+        bx              lr
+        .endfunc
+
+function ff_h264_h_loop_filter_luma_neon, export=1
+        h264_loop_filter_start
+
+        sub             r0,  r0,  #4
+        vld1.64         {d6},  [r0], r1
+        vld1.64         {d20}, [r0], r1
+        vld1.64         {d18}, [r0], r1
+        vld1.64         {d16}, [r0], r1
+        vld1.64         {d0},  [r0], r1
+        vld1.64         {d2},  [r0], r1
+        vld1.64         {d4},  [r0], r1
+        vld1.64         {d26}, [r0], r1
+        vld1.64         {d7},  [r0], r1
+        vld1.64         {d21}, [r0], r1
+        vld1.64         {d19}, [r0], r1
+        vld1.64         {d17}, [r0], r1
+        vld1.64         {d1},  [r0], r1
+        vld1.64         {d3},  [r0], r1
+        vld1.64         {d5},  [r0], r1
+        vld1.64         {d27}, [r0], r1
+
+        vtrn.32         q3,  q0
+        vtrn.32         q10, q1
+        vtrn.32         q9,  q2
+        vtrn.32         q8,  q13
+        vtrn.16         q3,  q9
+        vtrn.16         q10, q8
+        vtrn.16         q0,  q2
+        vtrn.16         q1,  q13
+        vtrn.8          q3,  q10
+        vtrn.8          q9,  q8
+        vtrn.8          q0,  q1
+        vtrn.8          q2,  q13
+
+        align_push_regs
+        sub             sp,  sp,  #16
+        vst1.64         {d4, d5},  [sp,:128]
+        sub             sp,  sp,  #16
+        vst1.64         {d20,d21}, [sp,:128]
+
+        h264_loop_filter_luma
+
+        vld1.64         {d20,d21}, [sp,:128]!
+        vld1.64         {d4, d5},  [sp,:128]!
+
+        vtrn.32         q3,  q0
+        vtrn.32         q10, q5
+        vtrn.32         q4,  q2
+        vtrn.32         q8,  q13
+        vtrn.16         q3,  q4
+        vtrn.16         q10, q8
+        vtrn.16         q0,  q2
+        vtrn.16         q5,  q13
+        vtrn.8          q3,  q10
+        vtrn.8          q4,  q8
+        vtrn.8          q0,  q5
+        vtrn.8          q2,  q13
+
+        sub             r0,  r0,  r1, lsl #4
+        vst1.64         {d6},  [r0], r1
+        vst1.64         {d20}, [r0], r1
+        vst1.64         {d8},  [r0], r1
+        vst1.64         {d16}, [r0], r1
+        vst1.64         {d0},  [r0], r1
+        vst1.64         {d10}, [r0], r1
+        vst1.64         {d4},  [r0], r1
+        vst1.64         {d26}, [r0], r1
+        vst1.64         {d7},  [r0], r1
+        vst1.64         {d21}, [r0], r1
+        vst1.64         {d9},  [r0], r1
+        vst1.64         {d17}, [r0], r1
+        vst1.64         {d1},  [r0], r1
+        vst1.64         {d11}, [r0], r1
+        vst1.64         {d5},  [r0], r1
+        vst1.64         {d27}, [r0], r1
+
+        align_pop_regs
+        bx              lr
+        .endfunc
+
+        .macro h264_loop_filter_chroma
+        vdup.8          d22, r2         @ alpha
+        vmovl.u8        q12, d24
+        vabd.u8         d26, d16, d0    @ abs(p0 - q0)
+        vmovl.u8        q2,  d0
+        vabd.u8         d28, d18, d16   @ abs(p1 - p0)
+        vsubw.u8        q2,  q2,  d16
+        vsli.16         d24, d24, #8
+        vshl.i16        q2,  q2,  #2
+        vabd.u8         d30, d2,  d0    @ abs(q1 - q0)
+        vaddw.u8        q2,  q2,  d18
+        vclt.u8         d26, d26, d22   @ < alpha
+        vsubw.u8        q2,  q2,  d2
+        vdup.8          d22, r3         @ beta
+        vclt.s8         d25, d24, #0
+        vrshrn.i16      d4,  q2,  #3
+        vclt.u8         d28, d28, d22   @ < beta
+        vbic            d26, d26, d25
+        vclt.u8         d30, d30, d22   @ < beta
+        vand            d26, d26, d28
+        vneg.s8         d25, d24
+        vand            d26, d26, d30
+        vmin.s8         d4,  d4,  d24
+        vmovl.u8        q14, d16
+        vand            d4,  d4,  d26
+        vmax.s8         d4,  d4,  d25
+        vmovl.u8        q11, d0
+        vaddw.s8        q14, q14, d4
+        vsubw.s8        q11, q11, d4
+        vqmovun.s16     d16, q14
+        vqmovun.s16     d0,  q11
+        .endm
+
+function ff_h264_v_loop_filter_chroma_neon, export=1
+        h264_loop_filter_start
+
+        sub             r0,  r0,  r1, lsl #1
+        vld1.64         {d18}, [r0,:64], r1
+        vld1.64         {d16}, [r0,:64], r1
+        vld1.64         {d0},  [r0,:64], r1
+        vld1.64         {d2},  [r0,:64]
+
+        h264_loop_filter_chroma
+
+        sub             r0,  r0,  r1, lsl #1
+        vst1.64         {d16}, [r0,:64], r1
+        vst1.64         {d0},  [r0,:64], r1
+
+        bx              lr
+        .endfunc
+
+function ff_h264_h_loop_filter_chroma_neon, export=1
+        h264_loop_filter_start
+
+        sub             r0,  r0,  #2
+        vld1.32         {d18[0]}, [r0], r1
+        vld1.32         {d16[0]}, [r0], r1
+        vld1.32         {d0[0]},  [r0], r1
+        vld1.32         {d2[0]},  [r0], r1
+        vld1.32         {d18[1]}, [r0], r1
+        vld1.32         {d16[1]}, [r0], r1
+        vld1.32         {d0[1]},  [r0], r1
+        vld1.32         {d2[1]},  [r0], r1
+
+        vtrn.16         d18, d0
+        vtrn.16         d16, d2
+        vtrn.8          d18, d16
+        vtrn.8          d0,  d2
+
+        h264_loop_filter_chroma
+
+        vtrn.16         d18, d0
+        vtrn.16         d16, d2
+        vtrn.8          d18, d16
+        vtrn.8          d0,  d2
+
+        sub             r0,  r0,  r1, lsl #3
+        vst1.32         {d18[0]}, [r0], r1
+        vst1.32         {d16[0]}, [r0], r1
+        vst1.32         {d0[0]},  [r0], r1
+        vst1.32         {d2[0]},  [r0], r1
+        vst1.32         {d18[1]}, [r0], r1
+        vst1.32         {d16[1]}, [r0], r1
+        vst1.32         {d0[1]},  [r0], r1
+        vst1.32         {d2[1]},  [r0], r1
+
+        bx              lr
+        .endfunc