libavcodec: vp8 neon optimizations for aarch64
authorMagnus Röös <mla2.roos@gmail.com>
Thu, 31 Jan 2019 16:04:27 +0000 (17:04 +0100)
committerCarl Eugen Hoyos <ceffmpeg@gmail.com>
Thu, 31 Jan 2019 19:17:51 +0000 (20:17 +0100)
Partial port of the ARM Neon for aarch64.

Benchmarks from fate:

benchmarking with Linux Perf Monitoring API
nop: 58.6
checkasm: using random seed 1760970128
NEON:
 - vp8dsp.idct       [OK]
 - vp8dsp.mc         [OK]
 - vp8dsp.loopfilter [OK]
checkasm: all 21 tests passed
vp8_idct_add_c: 201.6
vp8_idct_add_neon: 83.1
vp8_idct_dc_add_c: 107.6
vp8_idct_dc_add_neon: 33.8
vp8_idct_dc_add4y_c: 426.4
vp8_idct_dc_add4y_neon: 59.4
vp8_loop_filter8uv_h_c: 688.1
vp8_loop_filter8uv_h_neon: 216.3
vp8_loop_filter8uv_inner_h_c: 649.3
vp8_loop_filter8uv_inner_h_neon: 195.3
vp8_loop_filter8uv_inner_v_c: 544.8
vp8_loop_filter8uv_inner_v_neon: 131.3
vp8_loop_filter8uv_v_c: 706.1
vp8_loop_filter8uv_v_neon: 141.1
vp8_loop_filter16y_h_c: 668.8
vp8_loop_filter16y_h_neon: 242.8
vp8_loop_filter16y_inner_h_c: 647.3
vp8_loop_filter16y_inner_h_neon: 224.6
vp8_loop_filter16y_inner_v_c: 647.8
vp8_loop_filter16y_inner_v_neon: 128.8
vp8_loop_filter16y_v_c: 721.8
vp8_loop_filter16y_v_neon: 154.3
vp8_loop_filter_simple_h_c: 387.8
vp8_loop_filter_simple_h_neon: 187.6
vp8_loop_filter_simple_v_c: 384.1
vp8_loop_filter_simple_v_neon: 78.6
vp8_put_epel8_h4v4_c: 3971.1
vp8_put_epel8_h4v4_neon: 855.1
vp8_put_epel8_h4v6_c: 5060.1
vp8_put_epel8_h4v6_neon: 989.6
vp8_put_epel8_h6v4_c: 4320.8
vp8_put_epel8_h6v4_neon: 1007.3
vp8_put_epel8_h6v6_c: 5449.3
vp8_put_epel8_h6v6_neon: 1158.1
vp8_put_epel16_h6_c: 6683.8
vp8_put_epel16_h6_neon: 831.8
vp8_put_epel16_h6v6_c: 11110.8
vp8_put_epel16_h6v6_neon: 2214.8
vp8_put_epel16_v6_c: 7024.8
vp8_put_epel16_v6_neon: 799.6
vp8_put_pixels8_c: 112.8
vp8_put_pixels8_neon: 78.1
vp8_put_pixels16_c: 131.3
vp8_put_pixels16_neon: 129.8

Signed-off-by: Magnus Röös <mla2.roos@gmail.com>
libavcodec/aarch64/Makefile
libavcodec/aarch64/vp8dsp.h [new file with mode: 0644]
libavcodec/aarch64/vp8dsp_init_aarch64.c [new file with mode: 0644]
libavcodec/aarch64/vp8dsp_neon.S [new file with mode: 0644]
libavcodec/vp8dsp.c
libavcodec/vp8dsp.h

index 72080c2..2f57bce 100644 (file)
@@ -48,6 +48,8 @@ NEON-OBJS-$(CONFIG_MPEGAUDIODSP)        += aarch64/mpegaudiodsp_neon.o
 NEON-OBJS-$(CONFIG_AAC_DECODER)         += aarch64/aacpsdsp_neon.o
 NEON-OBJS-$(CONFIG_DCA_DECODER)         += aarch64/synth_filter_neon.o
 NEON-OBJS-$(CONFIG_VORBIS_DECODER)      += aarch64/vorbisdsp_neon.o
+NEON-OBJS-$(CONFIG_VP8DSP)              += aarch64/vp8dsp_init_aarch64.o       \
+                                           aarch64/vp8dsp_neon.o
 NEON-OBJS-$(CONFIG_VP9_DECODER)         += aarch64/vp9itxfm_16bpp_neon.o       \
                                            aarch64/vp9itxfm_neon.o             \
                                            aarch64/vp9lpf_16bpp_neon.o         \
diff --git a/libavcodec/aarch64/vp8dsp.h b/libavcodec/aarch64/vp8dsp.h
new file mode 100644 (file)
index 0000000..960dfa8
--- /dev/null
@@ -0,0 +1,70 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_ARM_VP8DSP_H
+#define AVCODEC_ARM_VP8DSP_H
+
+#include "libavcodec/vp8dsp.h"
+
+#define VP8_LF_Y(hv, inner, opt)                                             \
+    void ff_vp8_##hv##_loop_filter16##inner##_##opt(uint8_t *dst,            \
+                                                    ptrdiff_t stride,        \
+                                                    int flim_E, int flim_I,  \
+                                                    int hev_thresh)
+
+#define VP8_LF_UV(hv, inner, opt)                                            \
+    void ff_vp8_##hv##_loop_filter8uv##inner##_##opt(uint8_t *dstU,          \
+                                                     uint8_t *dstV,          \
+                                                     ptrdiff_t stride,       \
+                                                     int flim_E, int flim_I, \
+                                                     int hev_thresh)
+
+#define VP8_LF_SIMPLE(hv, opt)                                          \
+    void ff_vp8_##hv##_loop_filter16_simple_##opt(uint8_t *dst,         \
+                                                  ptrdiff_t stride,     \
+                                                  int flim)
+
+#define VP8_LF_HV(inner, opt)                   \
+    VP8_LF_Y(h,  inner, opt);                   \
+    VP8_LF_Y(v,  inner, opt);                   \
+    VP8_LF_UV(h, inner, opt);                   \
+    VP8_LF_UV(v, inner, opt)
+
+#define VP8_LF(opt)                             \
+    VP8_LF_HV(,       opt);                     \
+    VP8_LF_HV(_inner, opt);                     \
+    VP8_LF_SIMPLE(h, opt);                      \
+    VP8_LF_SIMPLE(v, opt)
+
+#define VP8_MC(n, opt)                                                  \
+    void ff_put_vp8_##n##_##opt(uint8_t *dst, ptrdiff_t dststride,      \
+                                uint8_t *src, ptrdiff_t srcstride,      \
+                                int h, int x, int y)
+
+#define VP8_EPEL(w, opt)                        \
+    VP8_MC(pixels ## w, opt);                   \
+    VP8_MC(epel ## w ## _h4, opt);              \
+    VP8_MC(epel ## w ## _h6, opt);              \
+    VP8_MC(epel ## w ## _v4, opt);              \
+    VP8_MC(epel ## w ## _h4v4, opt);            \
+    VP8_MC(epel ## w ## _h6v4, opt);            \
+    VP8_MC(epel ## w ## _v6, opt);              \
+    VP8_MC(epel ## w ## _h4v6, opt);            \
+    VP8_MC(epel ## w ## _h6v6, opt)
+
+#endif /* AVCODEC_ARM_VP8DSP_H */
diff --git a/libavcodec/aarch64/vp8dsp_init_aarch64.c b/libavcodec/aarch64/vp8dsp_init_aarch64.c
new file mode 100644 (file)
index 0000000..92a8c96
--- /dev/null
@@ -0,0 +1,81 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdint.h>
+#include <string.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/aarch64/cpu.h"
+#include "libavcodec/vp8dsp.h"
+#include "vp8dsp.h"
+
+void ff_vp8_luma_dc_wht_neon(int16_t block[4][4][16], int16_t dc[16]);
+
+void ff_vp8_idct_add_neon(uint8_t *dst, int16_t block[16], ptrdiff_t stride);
+void ff_vp8_idct_dc_add_neon(uint8_t *dst, int16_t block[16], ptrdiff_t stride);
+void ff_vp8_idct_dc_add4y_neon(uint8_t *dst, int16_t block[4][16], ptrdiff_t stride);
+
+VP8_LF(neon);
+
+VP8_EPEL(16, neon);
+VP8_EPEL(8,  neon);
+
+
+av_cold void ff_vp78dsp_init_aarch64(VP8DSPContext *dsp)
+{
+    if (!have_neon(av_get_cpu_flags())) {
+        return;
+    }
+    dsp->put_vp8_epel_pixels_tab[0][0][0] = ff_put_vp8_pixels16_neon;
+    dsp->put_vp8_epel_pixels_tab[0][0][2] = ff_put_vp8_epel16_h6_neon;
+    dsp->put_vp8_epel_pixels_tab[0][2][0] = ff_put_vp8_epel16_v6_neon;
+    dsp->put_vp8_epel_pixels_tab[0][2][2] = ff_put_vp8_epel16_h6v6_neon;
+
+    dsp->put_vp8_epel_pixels_tab[1][0][0] = ff_put_vp8_pixels8_neon;
+    dsp->put_vp8_epel_pixels_tab[1][2][2] = ff_put_vp8_epel8_h6v6_neon;
+    dsp->put_vp8_epel_pixels_tab[1][2][1] = ff_put_vp8_epel8_h4v6_neon;
+    dsp->put_vp8_epel_pixels_tab[1][1][2] = ff_put_vp8_epel8_h6v4_neon;
+    dsp->put_vp8_epel_pixels_tab[1][1][1] = ff_put_vp8_epel8_h4v4_neon;
+}
+
+av_cold void ff_vp8dsp_init_aarch64(VP8DSPContext *dsp)
+{
+    if (!have_neon(av_get_cpu_flags())) {
+        return;
+    }
+
+    dsp->vp8_idct_add       = ff_vp8_idct_add_neon;
+    dsp->vp8_idct_dc_add    = ff_vp8_idct_dc_add_neon;
+    dsp->vp8_idct_dc_add4y  = ff_vp8_idct_dc_add4y_neon;
+
+    dsp->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16_neon;
+    dsp->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16_neon;
+    dsp->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_neon;
+    dsp->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_neon;
+
+    dsp->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16_inner_neon;
+    dsp->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16_inner_neon;
+    dsp->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_neon;
+    dsp->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_neon;
+
+    dsp->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter16_simple_neon;
+    dsp->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter16_simple_neon;
+}
diff --git a/libavcodec/aarch64/vp8dsp_neon.S b/libavcodec/aarch64/vp8dsp_neon.S
new file mode 100644 (file)
index 0000000..490b0ca
--- /dev/null
@@ -0,0 +1,1031 @@
+/*
+ * VP8 NEON optimisations
+ *
+ * Copyright (c) 2010 Rob Clark <rob@ti.com>
+ * Copyright (c) 2011 Mans Rullgard <mans@mansr.com>
+ * Copyright (c) 2018 Magnus Röös <mla2.roos@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/aarch64/asm.S"
+#include "neon.S"
+
+function ff_vp8_idct_add_neon, export=1
+        ld1             {v0.8b - v3.8b},  [x1]
+        mov             w4,  #20091
+        movk            w4,  #35468/2, lsl 16
+        dup             v4.2s, w4
+
+        smull           v26.4s, v1.4h,  v4.4h[0]
+        smull           v27.4s, v3.4h,  v4.4h[0]
+        sqdmulh         v20.4h, v1.4h,  v4.4h[1]
+        sqdmulh         v23.4h, v3.4h,  v4.4h[1]
+        sqshrn          v21.4h, v26.4s, #16
+        sqshrn          v22.4h, v27.4s, #16
+        add             v21.4h, v21.4h, v1.4h
+        add             v22.4h, v22.4h, v3.4h
+
+        add             v16.4h,  v0.4h,   v2.4h
+        sub             v17.4h,  v0.4h,   v2.4h
+
+        add             v18.4h,  v21.4h,  v23.4h
+        sub             v19.4h,  v20.4h,  v22.4h
+
+        add             v0.4h,   v16.4h,  v18.4h
+        add             v1.4h,   v17.4h,  v19.4h
+        sub             v3.4h,   v16.4h,  v18.4h
+        sub             v2.4h,   v17.4h,  v19.4h
+
+        transpose_4x4H  v0, v1, v2, v3, v24, v5, v6, v7
+
+        movi            v29.8h, #0
+        smull           v26.4s,     v1.4h,  v4.4h[0]
+        st1             {v29.8h},   [x1],   #16
+        smull           v27.4s,     v3.4h,  v4.4h[0]
+        st1             {v29.16b},  [x1]
+        sqdmulh         v21.4h,     v1.4h,  v4.4h[1]
+        sqdmulh         v23.4h,     v3.4h,  v4.4h[1]
+        sqshrn          v20.4h,     v26.4s, #16
+        sqshrn          v22.4h,     v27.4s, #16
+        add             v20.4h,     v20.4h, v1.4h
+        add             v22.4h,     v22.4h, v3.4h
+        add             v16.4h,     v0.4h,  v2.4h
+        sub             v17.4h,     v0.4h,  v2.4h
+
+        add             v18.4h,     v20.4h, v23.4h
+        ld1             {v24.d}[0], [x0],   x2
+        zip1            v16.2d,     v16.2d, v17.2d
+        sub             v19.4h,     v21.4h, v22.4h
+        ld1             {v25.d}[0], [x0],   x2
+        zip1            v18.2d,     v18.2d, v19.2d
+        add             v0.8h,      v16.8h, v18.8h
+        ld1             {v25.d}[1], [x0],   x2
+        sub             v1.8h,      v16.8h, v18.8h
+        ld1             {v24.d}[1], [x0],   x2
+        srshr           v0.8h,      v0.8h,  #3
+        trn1            v24.4s,     v24.4s, v25.4s
+        srshr           v1.8h,      v1.8h,  #3
+        sub             x0,  x0,  x2,  lsl #2
+
+        ext             v1.16b, v1.16b, v1.16b, #8
+        trn1            v3.2d,  v0.2d,  v1.2d
+        trn2            v0.2d,  v0.2d,  v1.2d
+        trn1            v1.8h,  v3.8h,  v0.8h
+        trn2            v3.8h,  v3.8h,  v0.8h
+        uzp1            v0.4s,  v1.4s,  v3.4s
+        uzp2            v1.4s,  v3.4s,  v1.4s
+
+        uaddw           v0.8h,  v0.8h, v24.8b
+        uaddw2          v1.8h,  v1.8h, v24.16b
+        sqxtun          v0.8b,  v0.8h
+        sqxtun2         v0.16b, v1.8h
+        st1             {v0.s}[0],  [x0], x2
+        st1             {v0.s}[1],  [x0], x2
+        st1             {v0.s}[3],  [x0], x2
+        st1             {v0.s}[2],  [x0], x2
+
+        ret
+endfunc
+
+function ff_vp8_idct_dc_add4y_neon, export=1
+        movi            v0.16b,  #0
+        mov             x3,  #32
+        ld1r            {v16.4h},    [x1]
+        st1             {v0.h}[0],   [x1], x3
+        ld1r            {v17.4h},    [x1]
+        st1             {v0.h}[0],   [x1], x3
+        zip1            v16.2d,      v16.2d, v17.2d
+        ld1r            {v18.4h},    [x1]
+        st1             {v0.h}[0],   [x1], x3
+        ld1r            {v19.4h},    [x1]
+        st1             {v0.h}[0],   [x1], x3
+        zip1            v18.2d,      v18.2d, v19.2d
+        srshr           v16.8h,      v16.8h,  #3            // dc >>= 3
+        ld1             {v0.16b},     [x0], x2
+        srshr           v18.8h,       v18.8h,  #3
+        ld1             {v1.16b},     [x0], x2
+        uaddw           v20.8h,       v16.8h,  v0.8b
+        ld1             {v2.16b},     [x0], x2
+        uaddw2          v0.8h,        v18.8h,   v0.16b
+        ld1             {v3.16b},     [x0], x2
+        uaddw           v21.8h, v16.8h,  v1.8b
+        uaddw2          v1.8h,  v18.8h,  v1.16b
+        uaddw           v22.8h, v16.8h,  v2.8b
+        uaddw2          v2.8h,  v18.8h,  v2.16b
+        uaddw           v23.8h, v16.8h,  v3.8b
+        uaddw2          v3.8h,  v18.8h,  v3.16b
+        sub             x0,  x0,  x2,  lsl #2
+        sqxtun          v20.8b,  v20.8h
+        sqxtun2         v20.16b, v0.8h
+        sqxtun          v21.8b,  v21.8h
+        sqxtun2         v21.16b, v1.8h
+        sqxtun          v22.8b,  v22.8h
+        st1             {v20.16b},    [x0], x2
+        sqxtun2         v22.16b, v2.8h
+        st1             {v21.16b},    [x0], x2
+        sqxtun          v23.8b,  v23.8h
+        st1             {v22.16b},    [x0], x2
+        sqxtun2         v23.16b, v3.8h
+        st1             {v23.16b},    [x0], x2
+
+        ret
+endfunc
+
+function ff_vp8_idct_dc_add_neon, export=1
+        mov             w3,       #0
+        ld1r            {v2.8h},  [x1]
+        strh            w3,       [x1]
+        srshr           v2.8h,  v2.8h,  #3
+        ld1             {v0.s}[0],  [x0], x2
+        ld1             {v0.s}[1],  [x0], x2
+        uaddw           v3.8h,  v2.8h,  v0.8b
+        ld1             {v1.s}[0],  [x0], x2
+        ld1             {v1.s}[1],  [x0], x2
+        uaddw           v4.8h,  v2.8h,  v1.8b
+        sqxtun          v0.8b,  v3.8h
+        sqxtun          v1.8b,  v4.8h
+        sub             x0,  x0,  x2, lsl #2
+        st1             {v0.s}[0],  [x0], x2
+        st1             {v0.s}[1],  [x0], x2
+        st1             {v1.s}[0],  [x0], x2
+        st1             {v1.s}[1],  [x0], x2
+        ret
+endfunc
+
+// Register layout:
+//   P3..Q3 -> v0..v7
+//   flim_E -> v22
+//   flim_I -> v23
+//   hev_thresh -> x5
+//
+.macro  vp8_loop_filter, inner=0, simple=0, hev_thresh
+    .if \simple
+        uabd            v17.16b, v3.16b,  v4.16b      // abs(P0-Q0)
+        uabd            v23.16b, v2.16b,  v5.16b      // abs(P1-Q1)
+        uqadd           v17.16b, v17.16b, v17.16b     // abs(P0-Q0) * 2
+        ushr            v18.16b, v23.16b, #1          // abs(P1-Q1) / 2
+        uqadd           v19.16b, v17.16b,  v18.16b    // (abs(P0-Q0)*2) + (abs(P1-Q1)/2)
+        movi            v21.16b, #0x80
+        cmhs            v16.16b, v22.16b, v19.16b    // (abs(P0-Q0)*2) + (abs(P1-Q1)/2) <= flim
+    .else
+        // calculate hev and normal_limit:
+        uabd            v20.16b, v2.16b,  v3.16b      // abs(P1-P0)
+        uabd            v21.16b, v5.16b,  v4.16b      // abs(Q1-Q0)
+        uabd            v18.16b, v0.16b,  v1.16b      // abs(P3-P2)
+        uabd            v19.16b, v1.16b,  v2.16b      // abs(P2-P1)
+        cmhs            v16.16b, v23.16b, v20.16b     // abs(P1-P0) <= flim_I
+        cmhs            v17.16b, v23.16b, v21.16b     // abs(Q1-Q0) <= flim_I
+        cmhs            v18.16b, v23.16b, v18.16b     // abs(P3-P2) <= flim_I
+        cmhs            v19.16b, v23.16b, v19.16b     // abs(P2-P1) <= flim_I
+        and             v16.16b, v17.16b, v16.16b
+        uabd            v17.16b, v7.16b,  v6.16b      // abs(Q3-Q2)
+        and             v16.16b, v16.16b, v19.16b
+        uabd            v19.16b, v6.16b,  v5.16b      // abs(Q2-Q1)
+        and             v16.16b, v16.16b, v18.16b
+        cmhs            v18.16b, v23.16b, v17.16b     // abs(Q3-Q2) <= flim_I
+        cmhs            v19.16b, v23.16b, v19.16b     // abs(Q2-Q1) <= flim_I
+        uabd            v17.16b, v3.16b,  v4.16b      // abs(P0-Q0)
+        uabd            v23.16b, v2.16b,  v5.16b      // abs(P1-Q1)
+        and             v16.16b, v16.16b, v18.16b
+        uqadd           v17.16b, v17.16b, v17.16b     // abs(P0-Q0) * 2
+        and             v16.16b, v16.16b, v19.16b
+        ushr            v18.16b, v23.16b, #1          // abs(P1-Q1) / 2
+        dup             v23.16b, \hev_thresh          // hev_thresh
+        uqadd           v19.16b, v17.16b, v18.16b     // (abs(P0-Q0)*2) + (abs(P1-Q1)/2)
+        cmhi            v20.16b, v20.16b, v23.16b     // abs(P1-P0) > hev_thresh
+        cmhs            v19.16b, v22.16b, v19.16b     // (abs(P0-Q0)*2) + (abs(P1-Q1)/2) <= flim_E
+        cmhi            v22.16b, v21.16b, v23.16b     // abs(Q1-Q0) > hev_thresh
+        and             v16.16b, v16.16b, v19.16b
+        movi            v21.16b, #0x80
+        orr             v17.16b, v20.16b, v22.16b
+    .endif
+
+        // at this point:
+        //   v16: normal_limit
+        //   v17: hev
+
+        // convert to signed value:
+        eor            v3.16b, v3.16b, v21.16b           // PS0 = P0 ^ 0x80
+        eor            v4.16b, v4.16b, v21.16b           // QS0 = Q0 ^ 0x80
+
+        movi           v20.8h, #3
+        ssubl          v18.8h, v4.8b,  v3.8b             // QS0 - PS0
+        ssubl2         v19.8h, v4.16b, v3.16b            //   (widened to 16bit)
+        eor            v2.16b, v2.16b, v21.16b           // PS1 = P1 ^ 0x80
+        eor            v5.16b, v5.16b, v21.16b           // QS1 = Q1 ^ 0x80
+        mul            v18.8h, v18.8h, v20.8h            // w = 3 * (QS0 - PS0)
+        mul            v19.8h, v19.8h, v20.8h
+
+        sqsub          v20.16b, v2.16b, v5.16b           // clamp(PS1-QS1)
+        movi           v22.16b, #4
+        movi           v23.16b, #3
+    .if \inner
+        and            v20.16b, v20.16b, v17.16b         // if(hev) w += clamp(PS1-QS1)
+    .endif
+        saddw          v18.8h,  v18.8h, v20.8b           // w += clamp(PS1-QS1)
+        saddw2         v19.8h,  v19.8h, v20.16b
+        sqxtn          v18.8b,  v18.8h                   // narrow result back into v18
+        sqxtn2         v18.16b, v19.8h
+    .if !\inner && !\simple
+        eor            v1.16b,  v1.16b,  v21.16b         // PS2 = P2 ^ 0x80
+        eor            v6.16b,  v6.16b,  v21.16b         // QS2 = Q2 ^ 0x80
+    .endif
+        and            v18.16b, v18.16b, v16.16b         // w &= normal_limit
+
+        // registers used at this point..
+        //   v0 -> P3  (don't corrupt)
+        //   v1-v6 -> PS2-QS2
+        //   v7 -> Q3  (don't corrupt)
+        //   v17 -> hev
+        //   v18 -> w
+        //   v21 -> #0x80
+        //   v22 -> #4
+        //   v23 -> #3
+        //   v16, v19, v29 -> unused
+        //
+        // filter_common:   is4tap==1
+        //   c1 = clamp(w + 4) >> 3;
+        //   c2 = clamp(w + 3) >> 3;
+        //   Q0 = s2u(QS0 - c1);
+        //   P0 = s2u(PS0 + c2);
+
+    .if \simple
+        sqadd          v19.16b, v18.16b, v22.16b           // c1 = clamp((w&hev)+4)
+        sqadd          v20.16b, v18.16b, v23.16b           // c2 = clamp((w&hev)+3)
+        sshr           v19.16b, v19.16b, #3                // c1 >>= 3
+        sshr           v20.16b, v20.16b, #3                // c2 >>= 3
+        sqsub          v4.16b,  v4.16b,  v19.16b           // QS0 = clamp(QS0-c1)
+        sqadd          v3.16b,  v3.16b,  v20.16b           // PS0 = clamp(PS0+c2)
+        eor            v4.16b,  v4.16b,  v21.16b           // Q0 = QS0 ^ 0x80
+        eor            v3.16b,  v3.16b,  v21.16b           // P0 = PS0 ^ 0x80
+        eor            v5.16b,  v5.16b,  v21.16b           // Q1 = QS1 ^ 0x80
+        eor            v2.16b,  v2.16b,  v21.16b           // P1 = PS1 ^ 0x80
+    .elseif \inner
+        // the !is4tap case of filter_common, only used for inner blocks
+        //   c3 = ((c1&~hev) + 1) >> 1;
+        //   Q1 = s2u(QS1 - c3);
+        //   P1 = s2u(PS1 + c3);
+        sqadd          v19.16b, v18.16b, v22.16b           // c1 = clamp((w&hev)+4)
+        sqadd          v20.16b, v18.16b, v23.16b           // c2 = clamp((w&hev)+3)
+        sshr           v19.16b, v19.16b, #3                // c1 >>= 3
+        sshr           v20.16b, v20.16b, #3                // c2 >>= 3
+        sqsub          v4.16b,  v4.16b,  v19.16b           // QS0 = clamp(QS0-c1)
+        sqadd          v3.16b,  v3.16b,  v20.16b           // PS0 = clamp(PS0+c2)
+        bic            v19.16b, v19.16b, v17.16b           // c1 & ~hev
+        eor            v4.16b,  v4.16b,  v21.16b           // Q0 = QS0 ^ 0x80
+        srshr          v19.16b, v19.16b, #1                // c3 >>= 1
+        eor            v3.16b,  v3.16b,  v21.16b           // P0 = PS0 ^ 0x80
+        sqsub          v5.16b,  v5.16b,  v19.16b           // QS1 = clamp(QS1-c3)
+        sqadd          v2.16b,  v2.16b,  v19.16b           // PS1 = clamp(PS1+c3)
+        eor            v5.16b,  v5.16b,  v21.16b           // Q1 = QS1 ^ 0x80
+        eor            v2.16b,  v2.16b,  v21.16b           // P1 = PS1 ^ 0x80
+    .else
+        and            v20.16b, v18.16b, v17.16b           // w & hev
+        sqadd          v19.16b, v20.16b, v22.16b           // c1 = clamp((w&hev)+4)
+        sqadd          v20.16b, v20.16b, v23.16b           // c2 = clamp((w&hev)+3)
+        sshr           v19.16b, v19.16b, #3                // c1 >>= 3
+        sshr           v20.16b, v20.16b, #3                // c2 >>= 3
+        bic            v18.16b, v18.16b, v17.16b           // w &= ~hev
+        sqsub          v4.16b,  v4.16b,  v19.16b           // QS0 = clamp(QS0-c1)
+        sqadd          v3.16b,  v3.16b,  v20.16b           // PS0 = clamp(PS0+c2)
+
+        // filter_mbedge:
+        //   a = clamp((27*w + 63) >> 7);
+        //   Q0 = s2u(QS0 - a);
+        //   P0 = s2u(PS0 + a);
+        //   a = clamp((18*w + 63) >> 7);
+        //   Q1 = s2u(QS1 - a);
+        //   P1 = s2u(PS1 + a);
+        //   a = clamp((9*w + 63) >> 7);
+        //   Q2 = s2u(QS2 - a);
+        //   P2 = s2u(PS2 + a);
+        movi           v17.8h,  #63
+        sshll          v22.8h,  v18.8b, #3
+        sshll2         v23.8h,  v18.16b, #3
+        saddw          v22.8h,  v22.8h, v18.8b
+        saddw2         v23.8h,  v23.8h, v18.16b
+        add            v16.8h,  v17.8h, v22.8h
+        add            v17.8h,  v17.8h, v23.8h           //  9*w + 63
+        add            v19.8h,  v16.8h, v22.8h
+        add            v20.8h,  v17.8h, v23.8h           // 18*w + 63
+        add            v22.8h,  v19.8h, v22.8h
+        add            v23.8h,  v20.8h, v23.8h           // 27*w + 63
+        sqshrn         v16.8b,  v16.8h,  #7
+        sqshrn2        v16.16b, v17.8h, #7              // clamp(( 9*w + 63)>>7)
+        sqshrn         v19.8b,  v19.8h, #7
+        sqshrn2        v19.16b, v20.8h, #7              // clamp((18*w + 63)>>7)
+        sqshrn         v22.8b,  v22.8h, #7
+        sqshrn2        v22.16b, v23.8h, #7              // clamp((27*w + 63)>>7)
+        sqadd          v1.16b,  v1.16b,  v16.16b        // PS2 = clamp(PS2+a)
+        sqsub          v6.16b,  v6.16b,  v16.16b        // QS2 = clamp(QS2-a)
+        sqadd          v2.16b,  v2.16b,  v19.16b        // PS1 = clamp(PS1+a)
+        sqsub          v5.16b,  v5.16b,  v19.16b        // QS1 = clamp(QS1-a)
+        sqadd          v3.16b,  v3.16b,  v22.16b        // PS0 = clamp(PS0+a)
+        sqsub          v4.16b,  v4.16b,  v22.16b        // QS0 = clamp(QS0-a)
+        eor            v3.16b,  v3.16b,  v21.16b        // P0 = PS0 ^ 0x80
+        eor            v4.16b,  v4.16b,  v21.16b        // Q0 = QS0 ^ 0x80
+        eor            v2.16b,  v2.16b,  v21.16b        // P1 = PS1 ^ 0x80
+        eor            v5.16b,  v5.16b,  v21.16b        // Q1 = QS1 ^ 0x80
+        eor            v1.16b,  v1.16b,  v21.16b        // P2 = PS2 ^ 0x80
+        eor            v6.16b,  v6.16b,  v21.16b        // Q2 = QS2 ^ 0x80
+    .endif
+.endm
+
+.macro  vp8_v_loop_filter16 name, inner=0, simple=0
+function ff_vp8_v_loop_filter16\name\()_neon, export=1
+        sub             x0,  x0,  x1,  lsl #1+!\simple
+
+        // Load pixels:
+    .if !\simple
+        ld1             {v0.16b},     [x0], x1 // P3
+        ld1             {v1.16b},     [x0], x1 // P2
+    .endif
+        ld1             {v2.16b},     [x0], x1 // P1
+        ld1             {v3.16b},     [x0], x1 // P0
+        ld1             {v4.16b},     [x0], x1 // Q0
+        ld1             {v5.16b},     [x0], x1 // Q1
+    .if !\simple
+        ld1             {v6.16b},     [x0], x1 // Q2
+        ld1             {v7.16b},     [x0]     // Q3
+        dup             v23.16b, w3                 // flim_I
+    .endif
+        dup             v22.16b, w2                 // flim_E
+
+        vp8_loop_filter inner=\inner, simple=\simple, hev_thresh=w4
+
+        // back up to P2:  dst -= stride * 6
+        sub             x0,  x0,  x1,  lsl #2
+    .if !\simple
+        sub             x0,  x0,  x1,  lsl #1
+
+        // Store pixels:
+        st1             {v1.16b},     [x0], x1 // P2
+    .endif
+        st1             {v2.16b},     [x0], x1 // P1
+        st1             {v3.16b},     [x0], x1 // P0
+        st1             {v4.16b},     [x0], x1 // Q0
+        st1             {v5.16b},     [x0], x1 // Q1
+    .if !\simple
+        st1             {v6.16b},     [x0]     // Q2
+    .endif
+
+        ret
+endfunc
+.endm
+
+vp8_v_loop_filter16
+vp8_v_loop_filter16 _inner,  inner=1
+vp8_v_loop_filter16 _simple, simple=1
+
+.macro  vp8_v_loop_filter8uv name, inner=0
+function ff_vp8_v_loop_filter8uv\name\()_neon, export=1
+        sub             x0,  x0,  x2,  lsl #2
+        sub             x1,  x1,  x2,  lsl #2
+        // Load pixels:
+        ld1          {v0.d}[0],     [x0], x2  // P3
+        ld1          {v0.d}[1],     [x1], x2  // P3
+        ld1          {v1.d}[0],     [x0], x2  // P2
+        ld1          {v1.d}[1],     [x1], x2  // P2
+        ld1          {v2.d}[0],     [x0], x2  // P1
+        ld1          {v2.d}[1],     [x1], x2  // P1
+        ld1          {v3.d}[0],     [x0], x2  // P0
+        ld1          {v3.d}[1],     [x1], x2  // P0
+        ld1          {v4.d}[0],     [x0], x2  // Q0
+        ld1          {v4.d}[1],     [x1], x2  // Q0
+        ld1          {v5.d}[0],     [x0], x2  // Q1
+        ld1          {v5.d}[1],     [x1], x2  // Q1
+        ld1          {v6.d}[0],     [x0], x2  // Q2
+        ld1          {v6.d}[1],     [x1], x2  // Q2
+        ld1          {v7.d}[0],     [x0]      // Q3
+        ld1          {v7.d}[1],     [x1]      // Q3
+
+        dup          v22.16b, w3                 // flim_E
+        dup          v23.16b, w4                 // flim_I
+
+        vp8_loop_filter inner=\inner, hev_thresh=w5
+
+        // back up to P2:  u,v -= stride * 6
+        sub          x0,  x0,  x2,  lsl #2
+        sub          x1,  x1,  x2,  lsl #2
+        sub          x0,  x0,  x2,  lsl #1
+        sub          x1,  x1,  x2,  lsl #1
+
+        // Store pixels:
+
+        st1          {v1.d}[0],     [x0], x2  // P2
+        st1          {v1.d}[1],     [x1], x2  // P2
+        st1          {v2.d}[0],     [x0], x2  // P1
+        st1          {v2.d}[1],     [x1], x2  // P1
+        st1          {v3.d}[0],     [x0], x2  // P0
+        st1          {v3.d}[1],     [x1], x2  // P0
+        st1          {v4.d}[0],     [x0], x2  // Q0
+        st1          {v4.d}[1],     [x1], x2  // Q0
+        st1          {v5.d}[0],     [x0], x2  // Q1
+        st1          {v5.d}[1],     [x1], x2  // Q1
+        st1          {v6.d}[0],     [x0]      // Q2
+        st1          {v6.d}[1],     [x1]      // Q2
+
+        ret
+endfunc
+.endm
+
+vp8_v_loop_filter8uv
+vp8_v_loop_filter8uv _inner, inner=1
+
+.macro  vp8_h_loop_filter16 name, inner=0, simple=0
+function ff_vp8_h_loop_filter16\name\()_neon, export=1
+
+        sub             x0,  x0,  #4
+        // Load pixels:
+        ld1             {v0.d}[0], [x0], x1
+        ld1             {v1.d}[0], [x0], x1
+        ld1             {v2.d}[0], [x0], x1
+        ld1             {v3.d}[0], [x0], x1
+        ld1             {v4.d}[0], [x0], x1
+        ld1             {v5.d}[0], [x0], x1
+        ld1             {v6.d}[0], [x0], x1
+        ld1             {v7.d}[0], [x0], x1
+        ld1             {v0.d}[1], [x0], x1
+        ld1             {v1.d}[1], [x0], x1
+        ld1             {v2.d}[1], [x0], x1
+        ld1             {v3.d}[1], [x0], x1
+        ld1             {v4.d}[1], [x0], x1
+        ld1             {v5.d}[1], [x0], x1
+        ld1             {v6.d}[1], [x0], x1
+        ld1             {v7.d}[1], [x0], x1
+
+        transpose_8x16b   v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7, v30, v31
+
+        dup             v22.16b, w2                 // flim_E
+    .if !\simple
+        dup             v23.16b, w3                 // flim_I
+    .endif
+
+        vp8_loop_filter inner=\inner, simple=\simple, hev_thresh=w4
+
+        sub             x0,  x0,  x1, lsl #4    // backup 16 rows
+
+        transpose_8x16b   v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7, v30, v31
+
+        // Store pixels:
+        st1             {v0.d}[0], [x0], x1
+        st1             {v1.d}[0], [x0], x1
+        st1             {v2.d}[0], [x0], x1
+        st1             {v3.d}[0], [x0], x1
+        st1             {v4.d}[0], [x0], x1
+        st1             {v5.d}[0], [x0], x1
+        st1             {v6.d}[0], [x0], x1
+        st1             {v7.d}[0], [x0], x1
+        st1             {v0.d}[1], [x0], x1
+        st1             {v1.d}[1], [x0], x1
+        st1             {v2.d}[1], [x0], x1
+        st1             {v3.d}[1], [x0], x1
+        st1             {v4.d}[1], [x0], x1
+        st1             {v5.d}[1], [x0], x1
+        st1             {v6.d}[1], [x0], x1
+        st1             {v7.d}[1], [x0]
+
+        ret
+endfunc
+.endm
+
+vp8_h_loop_filter16
+vp8_h_loop_filter16 _inner,  inner=1
+vp8_h_loop_filter16 _simple, simple=1
+
+.macro  vp8_h_loop_filter8uv name, inner=0
+function ff_vp8_h_loop_filter8uv\name\()_neon, export=1
+        sub             x0,  x0,  #4
+        sub             x1,  x1,  #4
+
+        // Load pixels:
+        ld1          {v0.d}[0],     [x0], x2 // load u
+        ld1          {v0.d}[1],     [x1], x2 // load v
+        ld1          {v1.d}[0],     [x0], x2
+        ld1          {v1.d}[1],     [x1], x2
+        ld1          {v2.d}[0],     [x0], x2
+        ld1          {v2.d}[1],     [x1], x2
+        ld1          {v3.d}[0],     [x0], x2
+        ld1          {v3.d}[1],     [x1], x2
+        ld1          {v4.d}[0],     [x0], x2
+        ld1          {v4.d}[1],     [x1], x2
+        ld1          {v5.d}[0],     [x0], x2
+        ld1          {v5.d}[1],     [x1], x2
+        ld1          {v6.d}[0],     [x0], x2
+        ld1          {v6.d}[1],     [x1], x2
+        ld1          {v7.d}[0],     [x0], x2
+        ld1          {v7.d}[1],     [x1], x2
+
+        transpose_8x16b   v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7, v30, v31
+
+        dup             v22.16b, w3                 // flim_E
+        dup             v23.16b, w4                 // flim_I
+
+        vp8_loop_filter inner=\inner, hev_thresh=w5
+
+        sub             x0,  x0,  x2, lsl #3    // backup u 8 rows
+        sub             x1,  x1,  x2, lsl #3    // backup v 8 rows
+
+        transpose_8x16b   v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7, v30, v31
+
+        // Store pixels:
+        st1          {v0.d}[0],     [x0], x2 // load u
+        st1          {v0.d}[1],     [x1], x2 // load v
+        st1          {v1.d}[0],     [x0], x2
+        st1          {v1.d}[1],     [x1], x2
+        st1          {v2.d}[0],     [x0], x2
+        st1          {v2.d}[1],     [x1], x2
+        st1          {v3.d}[0],     [x0], x2
+        st1          {v3.d}[1],     [x1], x2
+        st1          {v4.d}[0],     [x0], x2
+        st1          {v4.d}[1],     [x1], x2
+        st1          {v5.d}[0],     [x0], x2
+        st1          {v5.d}[1],     [x1], x2
+        st1          {v6.d}[0],     [x0], x2
+        st1          {v6.d}[1],     [x1], x2
+        st1          {v7.d}[0],     [x0]
+        st1          {v7.d}[1],     [x1]
+
+        ret
+
+endfunc
+.endm
+
+vp8_h_loop_filter8uv
+vp8_h_loop_filter8uv _inner, inner=1
+
+
+function ff_put_vp8_pixels16_neon, export=1
+1:
+        subs            w4, w4, #4
+        ld1             {v0.16b},     [x2], x3
+        ld1             {v1.16b},     [x2], x3
+        ld1             {v2.16b},     [x2], x3
+        ld1             {v3.16b},     [x2], x3
+        st1             {v0.16b},     [x0], x1
+        st1             {v1.16b},     [x0], x1
+        st1             {v2.16b},     [x0], x1
+        st1             {v3.16b},     [x0], x1
+        bgt             1b
+        ret
+endfunc
+
+function ff_put_vp8_pixels8_neon, export=1
+1:
+        subs            w4, w4, #4
+        ld1             {v0.8b},   [x2], x3
+        ld1             {v0.d}[1], [x2], x3
+        ld1             {v1.8b},   [x2], x3
+        ld1             {v1.d}[1], [x2], x3
+        st1             {v0.8b},   [x0], x1
+        st1             {v0.d}[1], [x0], x1
+        st1             {v1.8b},   [x0], x1
+        st1             {v1.d}[1], [x0], x1
+        bgt             1b
+        ret
+endfunc
+
+/* 4/6-tap 8th-pel MC */
+
+.macro  vp8_epel8_h6    d,   s0,   s1
+        ext             v22.8b, \s0\().8b,  \s1\().8b,  #1
+        uxtl            v18.8h, \s0\().8b
+        ext             v23.8b, \s0\().8b,  \s1\().8b,  #2
+        uxtl            v19.8h, v22.8b
+        ext             v24.8b, \s0\().8b,  \s1\().8b,  #3
+        uxtl            v21.8h, v23.8b
+        ext             v25.8b, \s0\().8b,  \s1\().8b,  #4
+        uxtl            v22.8h, v24.8b
+        ext             v26.8b, \s0\().8b,  \s1\().8b,  #5
+        uxtl            v25.8h, v25.8b
+        mul             v21.8h, v21.8h, v0.8h[2]
+        uxtl            v26.8h, v26.8b
+        mul             v22.8h, v22.8h, v0.8h[3]
+        mls             v21.8h, v19.8h, v0.8h[1]
+        mls             v22.8h, v25.8h, v0.8h[4]
+        mla             v21.8h, v18.8h, v0.8h[0]
+        mla             v22.8h, v26.8h, v0.8h[5]
+        sqadd           v22.8h, v21.8h, v22.8h
+        sqrshrun        \d\().8b, v22.8h, #7
+.endm
+
+.macro  vp8_epel16_h6   d0,  v0,  v1
+        ext             v22.16b, \v0\().16b, \v1\().16b, #3
+        ext             v23.16b, \v0\().16b, \v1\().16b, #4
+        uxtl            v19.8h,  v22.8b
+        uxtl2           v22.8h,  v22.16b
+        ext             v3.16b,  \v0\().16b, \v1\().16b, #2
+        uxtl            v20.8h,  v23.8b
+        uxtl2           v23.8h,  v23.16b
+        ext             v16.16b, \v0\().16b, \v1\().16b, #1
+        uxtl            v18.8h,  v3.8b
+        uxtl2           v3.8h,   v3.16b
+        ext             v2.16b,  \v0\().16b, \v1\().16b, #5
+        uxtl            v21.8h,  v2.8b
+        uxtl2           v2.8h,   v2.16b
+        uxtl            v17.8h,  v16.8b
+        uxtl2           v16.8h,  v16.16b
+        mul             v19.8h,  v19.8h, v0.8h[3]
+        mul             v18.8h,  v18.8h, v0.8h[2]
+        mul             v3.8h,   v3.8h,  v0.8h[2]
+        mul             v22.8h,  v22.8h, v0.8h[3]
+        mls             v19.8h,  v20.8h, v0.8h[4]
+        uxtl            v20.8h,  \v0\().8b
+        uxtl2           v1.8h,   \v0\().16b
+        mls             v18.8h,  v17.8h, v0.8h[1]
+        mls             v3.8h,   v16.8h, v0.8h[1]
+        mls             v22.8h,  v23.8h, v0.8h[4]
+        mla             v18.8h,  v20.8h, v0.8h[0]
+        mla             v19.8h,  v21.8h, v0.8h[5]
+        mla             v3.8h,   v1.8h,  v0.8h[0]
+        mla             v22.8h,  v2.8h,  v0.8h[5]
+        sqadd           v19.8h,  v18.8h, v19.8h
+        sqadd           v22.8h,  v3.8h,  v22.8h
+        sqrshrun        \d0\().8b,  v19.8h, #7
+        sqrshrun2       \d0\().16b, v22.8h, #7
+.endm
+
+.macro  vp8_epel8_v6    d0,  s0,  s1,  s2, s3, s4, s5
+        uxtl            \s2\().8h, \s2\().8b
+        uxtl            \s3\().8h, \s3\().8b
+        uxtl            \s1\().8h, \s1\().8b
+        uxtl            \s4\().8h, \s4\().8b
+        uxtl            \s0\().8h, \s0\().8b
+        uxtl            \s5\().8h, \s5\().8b
+        mul             \s2\().8h, \s2\().8h, v0.8h[2]
+        mul             \s3\().8h, \s3\().8h, v0.8h[3]
+        mls             \s2\().8h, \s1\().8h, v0.8h[1]
+        mls             \s3\().8h, \s4\().8h, v0.8h[4]
+        mla             \s2\().8h, \s0\().8h, v0.8h[0]
+        mla             \s3\().8h, \s5\().8h, v0.8h[5]
+        sqadd           \s3\().8h, \s2\().8h, \s3\().8h
+        sqrshrun        \d0\().8b, \s3\().8h, #7
+.endm
+
+.macro  vp8_epel8_v6_y2 d0, d1, s0, s1, s2, s3, s4, s5, s6
+        uxtl            \s0\().8h, \s0\().8b
+        uxtl            \s3\().8h, \s3\().8b
+        uxtl            \s6\().8h, \s6\().8b
+        uxtl            \s1\().8h, \s1\().8b
+        uxtl            \s4\().8h, \s4\().8b
+        uxtl            \s2\().8h, \s2\().8b
+        uxtl            \s5\().8h, \s5\().8b
+        mul             \s0\().8h, \s0\().8h, v0.8h[0]
+        mul             v31.8h   , \s3\().8h, v0.8h[3]
+        mul             \s3\().8h, \s3\().8h, v0.8h[2]
+        mul             \s6\().8h, \s6\().8h, v0.8h[5]
+
+        mls             \s0\().8h, \s1\().8h, v0.8h[1]
+        mls             v31.8h   , \s4\().8h, v0.8h[4]
+        mls             \s3\().8h, \s2\().8h, v0.8h[1]
+        mls             \s6\().8h, \s5\().8h, v0.8h[4]
+
+        mla             \s0\().8h, \s2\().8h, v0.8h[2]
+        mla             v31.8h   , \s5\().8h, v0.8h[5]
+        mla             \s3\().8h, \s1\().8h, v0.8h[0]
+        mla             \s6\().8h, \s4\().8h, v0.8h[3]
+        sqadd           v31.8h   , \s0\().8h, v31.8h
+        sqadd           \s6\().8h, \s3\().8h, \s6\().8h
+        sqrshrun        \d0\().8b, v31.8h,    #7
+        sqrshrun        \d1\().8b, \s6\().8h, #7
+.endm
+
+.macro  vp8_epel8_h4    d,   v0,   v1
+        ext             v22.8b, \v0\().8b,  \v1\().8b,  #1
+        uxtl            v19.8h, \v0\().8b
+        ext             v23.8b, \v0\().8b,  \v1\().8b,  #2
+        uxtl            v20.8h, v22.8b
+        ext             v25.8b, \v0\().8b,  \v1\().8b,  #3
+        uxtl            v22.8h, v23.8b
+        uxtl            v25.8h, v25.8b
+        mul             v20.8h, v20.8h, v0.8h[2]
+        mul             v22.8h, v22.8h, v0.8h[3]
+        mls             v20.8h, v19.8h, v0.8h[1]
+        mls             v22.8h, v25.8h, v0.8h[4]
+        sqadd           v22.8h, v20.8h, v22.8h
+        sqrshrun        \d\().8b, v22.8h, #7
+.endm
+
+.macro  vp8_epel8_v4_y2 d0, s0, s1, s2, s3, s4
+        uxtl            \s0\().8h,  \s0\().8b
+        uxtl            \s1\().8h,  \s1\().8b
+        uxtl            \s2\().8h,  \s2\().8b
+        uxtl            \s3\().8h,  \s3\().8b
+        uxtl            \s4\().8h,  \s4\().8b
+        mul             v21.8h,     \s1\().8h, v0.8h[2]
+        mul             v23.8h,     \s2\().8h, v0.8h[3]
+        mul             \s2\().8h,  \s2\().8h, v0.8h[2]
+        mul             v22.8h,     \s3\().8h, v0.8h[3]
+        mls             v21.8h,     \s0\().8h, v0.8h[1]
+        mls             v23.8h,     \s3\().8h, v0.8h[4]
+        mls             \s2\().8h,  \s1\().8h, v0.8h[1]
+        mls             v22.8h,     \s4\().8h, v0.8h[4]
+        sqadd           v21.8h,     v21.8h,    v23.8h
+        sqadd           \s2\().8h,  \s2\().8h, v22.8h
+        sqrshrun        \d0\().8b,  v21.8h,    #7
+        sqrshrun2       \d0\().16b, \s2\().8h, #7
+.endm
+
+
+// note: worst case sum of all 6-tap filter values * 255 is 0x7f80 so 16 bit
+// arithmatic can be used to apply filters
+const   subpel_filters, align=4
+        .short     0,   6, 123,  12,   1,   0,   0,   0
+        .short     2,  11, 108,  36,   8,   1,   0,   0
+        .short     0,   9,  93,  50,   6,   0,   0,   0
+        .short     3,  16,  77,  77,  16,   3,   0,   0
+        .short     0,   6,  50,  93,   9,   0,   0,   0
+        .short     1,   8,  36, 108,  11,   2,   0,   0
+        .short     0,   1,  12, 123,   6,   0,   0,   0
+endconst
+
+function ff_put_vp8_epel16_v6_neon, export=1
+        sub             x2,  x2,  x3,  lsl #1
+
+        sxtw            x4,  w4
+        sxtw            x6,  w6
+        movrel          x17,  subpel_filters-16
+        add             x6,  x17,  x6, lsl #4  // y
+        ld1             {v0.8h},     [x6]
+1:
+        ld1             {v1.1d - v2.1d},    [x2], x3
+        ld1             {v3.1d - v4.1d},    [x2], x3
+        ld1             {v16.1d - v17.1d},  [x2], x3
+        ld1             {v18.1d - v19.1d},  [x2], x3
+        ld1             {v20.1d - v21.1d},  [x2], x3
+        ld1             {v22.1d - v23.1d},  [x2], x3
+        ld1             {v24.1d - v25.1d},  [x2]
+        sub             x2,  x2,  x3, lsl #2
+
+        vp8_epel8_v6_y2 v1, v3, v1, v3, v16, v18, v20, v22, v24
+        vp8_epel8_v6_y2 v2, v4, v2, v4, v17, v19, v21, v23, v25
+
+        st1             {v1.1d - v2.1d}, [x0], x1
+        st1             {v3.1d - v4.1d}, [x0], x1
+        subs            x4, x4, #2
+        bne             1b
+
+        ret
+endfunc
+
+function ff_put_vp8_epel16_h6_neon, export=1
+        sub             x2,  x2,  #2
+        sxtw            x5,  w5 // x
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        add             x5,  x17,  x5, lsl #4 // x
+        ld1             {v0.8h},  [x5]
+1:
+        ld1             {v1.16b, v2.16b}, [x2], x3
+        vp8_epel16_h6   v1, v1, v2
+        st1             {v1.16b}, [x0], x1
+
+        subs            w4, w4, #1
+        bne             1b
+        ret
+endfunc
+
+
+function ff_put_vp8_epel16_h6v6_neon, export=1
+        sub             x2,  x2,  x3,  lsl #1
+        sub             x2,  x2,  #2
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        sxtw            x5,  w5 // x
+        add             x16,  x17,  x5, lsl #4 // x
+        sub             sp,  sp,  #336+16
+        ld1             {v0.8h},  [x16]
+        add             x7,  sp,  #15
+        sxtw            x4,  w4
+        add             x16, x4, #5   // h
+        bic             x7,  x7,  #15
+1:
+        ld1             {v1.16b, v2.16b}, [x2], x3
+        vp8_epel16_h6   v1, v1, v2
+        st1             {v1.16b}, [x7], #16
+        subs            x16, x16, #1
+        bne             1b
+
+
+        // second pass (vertical):
+        sxtw            x6,  w6
+        add             x6,  x17,  x6, lsl #4  // y
+        add             x7,  sp,  #15
+        ld1             {v0.8h},     [x6]
+        bic             x7,  x7,  #15
+2:
+        ld1             {v1.8b - v4.8b},    [x7], #32
+        ld1             {v16.8b - v19.8b},  [x7], #32
+        ld1             {v20.8b - v23.8b},  [x7]
+        sub             x7,  x7,  #48
+
+        vp8_epel8_v6    v5, v1, v3, v16, v18, v20, v22
+        vp8_epel8_v6    v2, v2, v4, v17, v19, v21, v23
+        trn1            v2.2d, v5.2d, v2.2d
+
+        st1             {v2.16b}, [x0], x1
+        subs            x4, x4, #1
+        bne             2b
+
+        add             sp,  sp,  #336+16
+        ret
+endfunc
+
+function ff_put_vp8_epel8_h6v6_neon, export=1
+        sub             x2,  x2,  x3,  lsl #1
+        sub             x2,  x2,  #2
+        sxtw            x4,  w4
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        sxtw            x5,  w5
+        add             x5,  x17,  x5, lsl #4 // x
+        sub             sp,  sp,  #168+16
+        ld1             {v0.8h},  [x5]
+        add             x7,  sp,  #15
+        add             x16, x4,  #5   // h
+        bic             x7,  x7,  #15
+1:
+        ld1             {v1.8b, v2.8b}, [x2], x3
+
+        vp8_epel8_h6    v1, v1, v2
+
+        st1             {v1.8b}, [x7], #8
+        subs            x16, x16, #1
+        bne             1b
+
+        // second pass (vertical):
+        sxtw            x6,  w6
+        add             x6,  x17,  x6, lsl #4  // y
+        add             x7,  sp,   #15
+        ld1             {v0.8h},   [x6]
+        bic             x7,  x7,   #15
+2:
+        ld1             {v1.8b - v4.8b}, [x7], #32
+        ld1             {v5.8b - v7.8b}, [x7]
+
+        sub             x7,  x7,  #16
+
+        vp8_epel8_v6_y2 v1, v2, v1, v2, v3, v4, v5, v6, v7
+
+        st1             {v1.8b}, [x0], x1
+        st1             {v2.8b}, [x0], x1
+        subs            x4, x4, #2
+        bne             2b
+
+        add             sp,  sp,  #168+16
+        ret
+endfunc
+
+function ff_put_vp8_epel8_h4v6_neon, export=1
+        sub             x2,  x2,  x3,  lsl #1
+        sub             x2,  x2,  #1
+        sxtw            x4,  w4
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        sxtw            x5,  w5
+        add             x5,  x17,  x5, lsl #4 // x
+        sub             sp,  sp,  #168+16
+        ld1             {v0.8h},  [x5]
+        add             x7,  sp,  #15
+        add             x16, x4, #5   // h
+        bic             x7,  x7,  #15
+1:
+        ld1             {v1.8b, v2.8b}, [x2], x3
+
+        vp8_epel8_h4    v1, v1, v2
+
+        st1             {v1.8b}, [x7], #8
+        subs            x16, x16, #1
+        bne             1b
+
+        // second pass (vertical):
+        sxtw            x6,  w6
+        add             x6,  x17,  x6, lsl #4  // y
+        add             x7,  sp,   #15
+        ld1             {v0.8h},   [x6]
+        bic             x7,  x7,   #15
+2:
+        ld1             {v1.8b - v4.8b}, [x7], #32
+        ld1             {v5.8b - v7.8b}, [x7]
+
+        sub             x7,  x7,  #16
+
+        vp8_epel8_v6_y2 v1, v2, v1, v2, v3, v4, v5, v6, v7
+
+        st1             {v1.8b}, [x0], x1
+        st1             {v2.8b}, [x0], x1
+        subs            x4, x4, #2
+        bne             2b
+
+        add             sp,  sp,  #168+16
+        ret
+endfunc
+
+function ff_put_vp8_epel8_h4v4_neon, export=1
+        sub             x2,  x2,  x3
+        sub             x2,  x2,  #1
+        sxtw            x4,  w4
+
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        sxtw            x5,  w5
+        add             x5,  x17,  x5, lsl #4 // x
+        sub             sp,  sp,  #168+16
+        ld1             {v0.8h},  [x5]
+        add             x7,  sp,  #15
+        add             x16, x4, #3   // h
+        bic             x7,  x7,  #15
+1:
+        ld1             {v1.8b, v2.8b}, [x2], x3
+
+        vp8_epel8_h4    v1, v1, v2
+
+        st1             {v1.8b}, [x7], #8
+        subs            x16, x16, #1
+        bne             1b
+
+        // second pass (vertical):
+        sxtw            x6,  w6
+        add             x6,  x17,  x6, lsl #4  // y
+        add             x7,  sp,   #15
+        ld1             {v0.8h},   [x6]
+        bic             x7,  x7,   #15
+2:
+        ld1             {v1.8b - v2.8b}, [x7], #16
+        ld1             {v3.8b - v5.8b}, [x7]
+
+        vp8_epel8_v4_y2 v1, v1, v2, v3, v4, v5
+
+        st1             {v1.d}[0], [x0], x1
+        st1             {v1.d}[1], [x0], x1
+        subs            x4, x4, #2
+        bne             2b
+
+        add             sp,  sp,  #168+16
+        ret
+endfunc
+
+function ff_put_vp8_epel8_h6v4_neon, export=1
+        sub             x2,  x2,  x3
+        sub             x2,  x2,  #2
+        sxtw            x4,  w4
+
+
+        // first pass (horizontal):
+        movrel          x17,  subpel_filters-16
+        sxtw            x5,  w5
+        add             x5,  x17,  x5, lsl #4 // x
+        sub             sp,  sp,  #168+16
+        ld1             {v0.8h},  [x5]
+        add             x7,  sp,  #15
+        add             x16, x4, #3   // h
+        bic             x7,  x7,  #15
+1:
+        ld1             {v1.8b, v2.8b}, [x2], x3
+
+        vp8_epel8_h6    v1, v1, v2
+
+        st1             {v1.8b}, [x7], #8
+        subs            x16, x16, #1
+        bne             1b
+
+        // second pass (vertical):
+        sxtw            x6,  w6
+        add             x6,  x17,  x6, lsl #4  // y
+        add             x7,  sp,   #15
+        ld1             {v0.8h},   [x6]
+        bic             x7,  x7,   #15
+2:
+        ld1             {v1.8b - v2.8b}, [x7], #16
+        ld1             {v3.8b - v5.8b}, [x7]
+
+        vp8_epel8_v4_y2 v1, v1, v2, v3, v4, v5
+
+        st1             {v1.d}[0], [x0], x1
+        st1             {v1.d}[1], [x0], x1
+        subs            x4, x4, #2
+        bne             2b
+
+        add             sp,  sp,  #168+16
+        ret
+endfunc
index fed5c67..30452bf 100644 (file)
@@ -679,6 +679,8 @@ av_cold void ff_vp78dsp_init(VP8DSPContext *dsp)
         ff_vp78dsp_init_ppc(dsp);
     if (ARCH_X86)
         ff_vp78dsp_init_x86(dsp);
+    if (ARCH_AARCH64)
+        ff_vp78dsp_init_aarch64(dsp);
 }
 
 #if CONFIG_VP7_DECODER
@@ -739,5 +741,7 @@ av_cold void ff_vp8dsp_init(VP8DSPContext *dsp)
         ff_vp8dsp_init_x86(dsp);
     if (ARCH_MIPS)
         ff_vp8dsp_init_mips(dsp);
+    if (ARCH_AARCH64)
+        ff_vp8dsp_init_aarch64(dsp);
 }
 #endif /* CONFIG_VP8_DECODER */
index eaae4ae..cfe1524 100644 (file)
@@ -91,11 +91,13 @@ void ff_put_vp8_pixels4_c(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 void ff_vp7dsp_init(VP8DSPContext *c);
 
 void ff_vp78dsp_init(VP8DSPContext *c);
+void ff_vp78dsp_init_aarch64(VP8DSPContext *c);
 void ff_vp78dsp_init_arm(VP8DSPContext *c);
 void ff_vp78dsp_init_ppc(VP8DSPContext *c);
 void ff_vp78dsp_init_x86(VP8DSPContext *c);
 
 void ff_vp8dsp_init(VP8DSPContext *c);
+void ff_vp8dsp_init_aarch64(VP8DSPContext *c);
 void ff_vp8dsp_init_arm(VP8DSPContext *c);
 void ff_vp8dsp_init_x86(VP8DSPContext *c);
 void ff_vp8dsp_init_mips(VP8DSPContext *c);