vp9: re-split the decoder/format/dsp interface header files.

The advantage here is that the internal software decoder interface is
not exposed to the DSP functions or the hardware accelerations.

libavcodec/aarch64/vp9dsp_init.h

@@ -21,7 +21,7 @@
 #ifndef AVCODEC_AARCH64_VP9DSP_INIT_H
 #define AVCODEC_AARCH64_VP9DSP_INIT_H
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 
 void ff_vp9dsp_init_10bpp_aarch64(VP9DSPContext *dsp);
 void ff_vp9dsp_init_12bpp_aarch64(VP9DSPContext *dsp);

libavcodec/aarch64/vp9dsp_init_aarch64.c

@@ -22,7 +22,7 @@
 
 #include "libavutil/attributes.h"
 #include "libavutil/aarch64/cpu.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "vp9dsp_init.h"
 
 #define declare_fpel(type, sz)                                          \

libavcodec/arm/vp9dsp_init.h

@@ -21,7 +21,7 @@
 #ifndef AVCODEC_ARM_VP9DSP_INIT_H
 #define AVCODEC_ARM_VP9DSP_INIT_H
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 
 void ff_vp9dsp_init_10bpp_arm(VP9DSPContext *dsp);
 void ff_vp9dsp_init_12bpp_arm(VP9DSPContext *dsp);

libavcodec/arm/vp9dsp_init_arm.c

@@ -22,7 +22,7 @@
 
 #include "libavutil/attributes.h"
 #include "libavutil/arm/cpu.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "vp9dsp_init.h"
 
 #define declare_fpel(type, sz)                                          \

libavcodec/mips/vp9_idct_msa.c

@@ -19,7 +19,7 @@
  */
 
 #include <string.h>
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavutil/mips/generic_macros_msa.h"
 #include "vp9dsp_mips.h"
 

libavcodec/mips/vp9_intra_msa.c

@@ -18,7 +18,7 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavutil/mips/generic_macros_msa.h"
 #include "vp9dsp_mips.h"
 

libavcodec/mips/vp9_lpf_msa.c

@@ -18,7 +18,7 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavutil/mips/generic_macros_msa.h"
 #include "vp9dsp_mips.h"
 

libavcodec/mips/vp9_mc_msa.c

@@ -18,7 +18,7 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavutil/mips/generic_macros_msa.h"
 #include "vp9dsp_mips.h"
 

libavcodec/mips/vp9dsp_init_mips.c

@@ -20,7 +20,7 @@
 
 #include "config.h"
 #include "libavutil/common.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "vp9dsp_mips.h"
 
 #if HAVE_MSA

libavcodec/vp9.c

@@ -30,7 +30,7 @@
 #include "vp56.h"
 #include "vp9.h"
 #include "vp9data.h"
-#include "vp9.h"
+#include "vp9dec.h"
 #include "libavutil/avassert.h"
 #include "libavutil/pixdesc.h"
 

libavcodec/vp9.h

@@ -27,9 +27,6 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#include "libavutil/buffer.h"
-#include "libavutil/internal.h"
-
 #include "avcodec.h"
 #include "thread.h"
 #include "vp56.h"
@@ -93,131 +90,6 @@ enum InterPredMode {
     NEWMV     = 13,
 };
 
-enum MVJoint {
-    MV_JOINT_ZERO,
-    MV_JOINT_H,
-    MV_JOINT_V,
-    MV_JOINT_HV,
-};
-
-typedef struct ProbContext {
-    uint8_t y_mode[4][9];
-    uint8_t uv_mode[10][9];
-    uint8_t filter[4][2];
-    uint8_t mv_mode[7][3];
-    uint8_t intra[4];
-    uint8_t comp[5];
-    uint8_t single_ref[5][2];
-    uint8_t comp_ref[5];
-    uint8_t tx32p[2][3];
-    uint8_t tx16p[2][2];
-    uint8_t tx8p[2];
-    uint8_t skip[3];
-    uint8_t mv_joint[3];
-    struct {
-        uint8_t sign;
-        uint8_t classes[10];
-        uint8_t class0;
-        uint8_t bits[10];
-        uint8_t class0_fp[2][3];
-        uint8_t fp[3];
-        uint8_t class0_hp;
-        uint8_t hp;
-    } mv_comp[2];
-    uint8_t partition[4][4][3];
-} ProbContext;
-
-typedef void (*vp9_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
-                            const uint8_t *ref, ptrdiff_t ref_stride,
-                            int h, int mx, int my);
-typedef void (*vp9_scaled_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
-                                   const uint8_t *ref, ptrdiff_t ref_stride,
-                                   int h, int mx, int my, int dx, int dy);
-
-typedef struct VP9DSPContext {
-    /*
-     * dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32
-     * dimension 2: intra prediction modes
-     *
-     * dst/left/top is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
-     * stride is aligned by 16 pixels
-     * top[-1] is top/left; top[4,7] is top-right for 4x4
-     */
-    // FIXME(rbultje) maybe replace left/top pointers with HAVE_TOP/
-    // HAVE_LEFT/HAVE_TOPRIGHT flags instead, and then handle it in-place?
-    // also needs to fit in with what H.264/VP8/etc do
-    void (*intra_pred[N_TXFM_SIZES][N_INTRA_PRED_MODES])(uint8_t *dst,
-                                                         ptrdiff_t stride,
-                                                         const uint8_t *left,
-                                                         const uint8_t *top);
-
-    /*
-     * dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32, 4=lossless (3-4=dct only)
-     * dimension 2: 0=dct/dct, 1=dct/adst, 2=adst/dct, 3=adst/adst
-     *
-     * dst is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
-     * stride is aligned by 16 pixels
-     * block is 16-byte aligned
-     * eob indicates the position (+1) of the last non-zero coefficient,
-     * in scan-order. This can be used to write faster versions, e.g. a
-     * dc-only 4x4/8x8/16x16/32x32, or a 4x4-only (eob<10) 8x8/16x16/32x32,
-     * etc.
-     */
-    // FIXME also write idct_add_block() versions for whole (inter) pred
-    // blocks, so we can do 2 4x4s at once
-    void (*itxfm_add[N_TXFM_SIZES + 1][N_TXFM_TYPES])(uint8_t *dst,
-                                                      ptrdiff_t stride,
-                                                      int16_t *block, int eob);
-
-    /*
-     * dimension 1: width of filter (0=4, 1=8, 2=16)
-     * dimension 2: 0=col-edge filter (h), 1=row-edge filter (v)
-     *
-     * dst/stride are aligned by 8
-     */
-    void (*loop_filter_8[3][2])(uint8_t *dst, ptrdiff_t stride,
-                                int mb_lim, int lim, int hev_thr);
-
-    /*
-     * dimension 1: 0=col-edge filter (h), 1=row-edge filter (v)
-     *
-     * The width of filter is assumed to be 16; dst/stride are aligned by 16
-     */
-    void (*loop_filter_16[2])(uint8_t *dst, ptrdiff_t stride,
-                              int mb_lim, int lim, int hev_thr);
-
-    /*
-     * dimension 1/2: width of filter (0=4, 1=8) for each filter half
-     * dimension 3: 0=col-edge filter (h), 1=row-edge filter (v)
-     *
-     * dst/stride are aligned by operation size
-     * this basically calls loop_filter[d1][d3][0](), followed by
-     * loop_filter[d2][d3][0]() on the next 8 pixels
-     * mb_lim/lim/hev_thr contain two values in the lowest two bytes of the
-     * integer.
-     */
-    // FIXME perhaps a mix4 that operates on 32px (for AVX2)
-    void (*loop_filter_mix2[2][2][2])(uint8_t *dst, ptrdiff_t stride,
-                                      int mb_lim, int lim, int hev_thr);
-
-    /*
-     * dimension 1: hsize (0: 64, 1: 32, 2: 16, 3: 8, 4: 4)
-     * dimension 2: filter type (0: smooth, 1: regular, 2: sharp, 3: bilin)
-     * dimension 3: averaging type (0: put, 1: avg)
-     * dimension 4: x subpel interpolation (0: none, 1: 8tap/bilin)
-     * dimension 5: y subpel interpolation (0: none, 1: 8tap/bilin)
-     *
-     * dst/stride are aligned by hsize
-     */
-    vp9_mc_func mc[5][4][2][2][2];
-
-    /*
-     * for scalable MC, first 3 dimensions identical to above, the other two
-     * don't exist since it changes per stepsize.
-     */
-    vp9_scaled_mc_func smc[5][4][2];
-} VP9DSPContext;
-
 enum CompPredMode {
     PRED_SINGLEREF,
     PRED_COMPREF,
@@ -229,12 +101,6 @@ typedef struct VP9mvrefPair {
     int8_t ref[2];
 } VP9mvrefPair;
 
-typedef struct VP9Filter {
-    uint8_t level[8 * 8];
-    uint8_t /* bit=col */ mask[2 /* 0=y, 1=uv */][2 /* 0=col, 1=row */]
-                              [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];
-} VP9Filter;
-
 typedef struct VP9Frame {
     ThreadFrame tf;
     AVBufferRef *extradata;
@@ -270,16 +136,6 @@ enum BlockSize {
     N_BS_SIZES,
 };
 
-typedef struct VP9Block {
-    uint8_t seg_id, intra, comp, ref[2], mode[4], uvmode, skip;
-    enum FilterMode filter;
-    VP56mv mv[4 /* b_idx */][2 /* ref */];
-    enum BlockSize bs;
-    enum TxfmMode tx, uvtx;
-    enum BlockLevel bl;
-    enum BlockPartition bp;
-} VP9Block;
-
 typedef struct VP9BitstreamHeader {
     // bitstream header
     uint8_t profile;
@@ -355,137 +211,4 @@ typedef struct VP9SharedContext {
     VP9Frame frames[3];
 } VP9SharedContext;
 
-typedef struct VP9Context {
-    VP9SharedContext s;
-
-    VP9DSPContext dsp;
-    VideoDSPContext vdsp;
-    GetBitContext gb;
-    VP56RangeCoder c;
-    VP56RangeCoder *c_b;
-    unsigned c_b_size;
-    VP9Block *b_base, *b;
-    int pass;
-    int row, row7, col, col7;
-    uint8_t *dst[3];
-    ptrdiff_t y_stride, uv_stride;
-
-    uint8_t ss_h, ss_v;
-    uint8_t last_bpp, bpp_index, bytesperpixel;
-    uint8_t last_keyframe;
-    // sb_cols/rows, rows/cols and last_fmt are used for allocating all internal
-    // arrays, and are thus per-thread. w/h and gf_fmt are synced between threads
-    // and are therefore per-stream. pix_fmt represents the value in the header
-    // of the currently processed frame.
-    int w, h;
-    enum AVPixelFormat pix_fmt, last_fmt, gf_fmt;
-    unsigned sb_cols, sb_rows, rows, cols;
-    ThreadFrame next_refs[8];
-
-    struct {
-        uint8_t lim_lut[64];
-        uint8_t mblim_lut[64];
-    } filter_lut;
-    unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;
-    struct {
-        ProbContext p;
-        uint8_t coef[4][2][2][6][6][3];
-    } prob_ctx[4];
-    struct {
-        ProbContext p;
-        uint8_t coef[4][2][2][6][6][11];
-    } prob;
-    struct {
-        unsigned y_mode[4][10];
-        unsigned uv_mode[10][10];
-        unsigned filter[4][3];
-        unsigned mv_mode[7][4];
-        unsigned intra[4][2];
-        unsigned comp[5][2];
-        unsigned single_ref[5][2][2];
-        unsigned comp_ref[5][2];
-        unsigned tx32p[2][4];
-        unsigned tx16p[2][3];
-        unsigned tx8p[2][2];
-        unsigned skip[3][2];
-        unsigned mv_joint[4];
-        struct {
-            unsigned sign[2];
-            unsigned classes[11];
-            unsigned class0[2];
-            unsigned bits[10][2];
-            unsigned class0_fp[2][4];
-            unsigned fp[4];
-            unsigned class0_hp[2];
-            unsigned hp[2];
-        } mv_comp[2];
-        unsigned partition[4][4][4];
-        unsigned coef[4][2][2][6][6][3];
-        unsigned eob[4][2][2][6][6][2];
-    } counts;
-
-    // contextual (left/above) cache
-    DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];
-    DECLARE_ALIGNED(16, uint8_t, left_mode_ctx)[16];
-    DECLARE_ALIGNED(16, VP56mv, left_mv_ctx)[16][2];
-    DECLARE_ALIGNED(16, uint8_t, left_uv_nnz_ctx)[2][16];
-    DECLARE_ALIGNED(8, uint8_t, left_partition_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_skip_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_txfm_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_segpred_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_intra_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_comp_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_ref_ctx)[8];
-    DECLARE_ALIGNED(8, uint8_t, left_filter_ctx)[8];
-    uint8_t *above_partition_ctx;
-    uint8_t *above_mode_ctx;
-    // FIXME maybe merge some of the below in a flags field?
-    uint8_t *above_y_nnz_ctx;
-    uint8_t *above_uv_nnz_ctx[2];
-    uint8_t *above_skip_ctx; // 1bit
-    uint8_t *above_txfm_ctx; // 2bit
-    uint8_t *above_segpred_ctx; // 1bit
-    uint8_t *above_intra_ctx; // 1bit
-    uint8_t *above_comp_ctx; // 1bit
-    uint8_t *above_ref_ctx; // 2bit
-    uint8_t *above_filter_ctx;
-    VP56mv (*above_mv_ctx)[2];
-
-    // whole-frame cache
-    uint8_t *intra_pred_data[3];
-    VP9Filter *lflvl;
-    DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];
-
-    // block reconstruction intermediates
-    int block_alloc_using_2pass;
-    int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
-    uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];
-    struct { int x, y; } min_mv, max_mv;
-    DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];
-    DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];
-    uint16_t mvscale[3][2];
-    uint8_t mvstep[3][2];
-} VP9Context;
-
-extern const int16_t ff_vp9_subpel_filters[3][16][8];
-
-void ff_vp9dsp_init(VP9DSPContext *dsp, int bpp, int bitexact);
-
-void ff_vp9dsp_init_8(VP9DSPContext *dsp);
-void ff_vp9dsp_init_10(VP9DSPContext *dsp);
-void ff_vp9dsp_init_12(VP9DSPContext *dsp);
-
-void ff_vp9dsp_init_aarch64(VP9DSPContext *dsp, int bpp);
-void ff_vp9dsp_init_arm(VP9DSPContext *dsp, int bpp);
-void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact);
-void ff_vp9dsp_init_mips(VP9DSPContext *dsp, int bpp);
-
-void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb);
-
-void ff_vp9_adapt_probs(VP9Context *s);
-
-void ff_vp9_decode_block(AVCodecContext *ctx, int row, int col,
-                         VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
-                         enum BlockLevel bl, enum BlockPartition bp);
-
 #endif /* AVCODEC_VP9_H */

libavcodec/vp9block.c

@@ -29,6 +29,7 @@
 #include "vp56.h"
 #include "vp9.h"
 #include "vp9data.h"
+#include "vp9dec.h"
 
 static const uint8_t bwh_tab[2][N_BS_SIZES][2] = {
     {

libavcodec/vp9data.h

@@ -24,7 +24,7 @@
 
 #include <stdint.h>
 
-#include "vp9.h"
+#include "vp9dec.h"
 
 extern const int8_t ff_vp9_partition_tree[3][2];
 extern const uint8_t ff_vp9_default_kf_partition_probs[4][4][3];

libavcodec/vp9dec.h

@@ -0,0 +1,206 @@
+/*
+ * VP9 compatible video decoder
+ *
+ * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
+ * Copyright (C) 2013 Clément Bœsch <u pkh me>
+ *
+ * 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_VP9DEC_H
+#define AVCODEC_VP9DEC_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include "libavutil/buffer.h"
+#include "libavutil/internal.h"
+
+#include "vp9.h"
+#include "vp9dsp.h"
+
+enum MVJoint {
+    MV_JOINT_ZERO,
+    MV_JOINT_H,
+    MV_JOINT_V,
+    MV_JOINT_HV,
+};
+
+typedef struct ProbContext {
+    uint8_t y_mode[4][9];
+    uint8_t uv_mode[10][9];
+    uint8_t filter[4][2];
+    uint8_t mv_mode[7][3];
+    uint8_t intra[4];
+    uint8_t comp[5];
+    uint8_t single_ref[5][2];
+    uint8_t comp_ref[5];
+    uint8_t tx32p[2][3];
+    uint8_t tx16p[2][2];
+    uint8_t tx8p[2];
+    uint8_t skip[3];
+    uint8_t mv_joint[3];
+    struct {
+        uint8_t sign;
+        uint8_t classes[10];
+        uint8_t class0;
+        uint8_t bits[10];
+        uint8_t class0_fp[2][3];
+        uint8_t fp[3];
+        uint8_t class0_hp;
+        uint8_t hp;
+    } mv_comp[2];
+    uint8_t partition[4][4][3];
+} ProbContext;
+
+typedef struct VP9Filter {
+    uint8_t level[8 * 8];
+    uint8_t /* bit=col */ mask[2 /* 0=y, 1=uv */][2 /* 0=col, 1=row */]
+                              [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];
+} VP9Filter;
+
+typedef struct VP9Block {
+    uint8_t seg_id, intra, comp, ref[2], mode[4], uvmode, skip;
+    enum FilterMode filter;
+    VP56mv mv[4 /* b_idx */][2 /* ref */];
+    enum BlockSize bs;
+    enum TxfmMode tx, uvtx;
+    enum BlockLevel bl;
+    enum BlockPartition bp;
+} VP9Block;
+
+typedef struct VP9Context {
+    VP9SharedContext s;
+
+    VP9DSPContext dsp;
+    VideoDSPContext vdsp;
+    GetBitContext gb;
+    VP56RangeCoder c;
+    VP56RangeCoder *c_b;
+    unsigned c_b_size;
+    VP9Block *b_base, *b;
+    int pass;
+    int row, row7, col, col7;
+    uint8_t *dst[3];
+    ptrdiff_t y_stride, uv_stride;
+
+    uint8_t ss_h, ss_v;
+    uint8_t last_bpp, bpp_index, bytesperpixel;
+    uint8_t last_keyframe;
+    // sb_cols/rows, rows/cols and last_fmt are used for allocating all internal
+    // arrays, and are thus per-thread. w/h and gf_fmt are synced between threads
+    // and are therefore per-stream. pix_fmt represents the value in the header
+    // of the currently processed frame.
+    int w, h;
+    enum AVPixelFormat pix_fmt, last_fmt, gf_fmt;
+    unsigned sb_cols, sb_rows, rows, cols;
+    ThreadFrame next_refs[8];
+
+    struct {
+        uint8_t lim_lut[64];
+        uint8_t mblim_lut[64];
+    } filter_lut;
+    unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;
+    struct {
+        ProbContext p;
+        uint8_t coef[4][2][2][6][6][3];
+    } prob_ctx[4];
+    struct {
+        ProbContext p;
+        uint8_t coef[4][2][2][6][6][11];
+    } prob;
+    struct {
+        unsigned y_mode[4][10];
+        unsigned uv_mode[10][10];
+        unsigned filter[4][3];
+        unsigned mv_mode[7][4];
+        unsigned intra[4][2];
+        unsigned comp[5][2];
+        unsigned single_ref[5][2][2];
+        unsigned comp_ref[5][2];
+        unsigned tx32p[2][4];
+        unsigned tx16p[2][3];
+        unsigned tx8p[2][2];
+        unsigned skip[3][2];
+        unsigned mv_joint[4];
+        struct {
+            unsigned sign[2];
+            unsigned classes[11];
+            unsigned class0[2];
+            unsigned bits[10][2];
+            unsigned class0_fp[2][4];
+            unsigned fp[4];
+            unsigned class0_hp[2];
+            unsigned hp[2];
+        } mv_comp[2];
+        unsigned partition[4][4][4];
+        unsigned coef[4][2][2][6][6][3];
+        unsigned eob[4][2][2][6][6][2];
+    } counts;
+
+    // contextual (left/above) cache
+    DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];
+    DECLARE_ALIGNED(16, uint8_t, left_mode_ctx)[16];
+    DECLARE_ALIGNED(16, VP56mv, left_mv_ctx)[16][2];
+    DECLARE_ALIGNED(16, uint8_t, left_uv_nnz_ctx)[2][16];
+    DECLARE_ALIGNED(8, uint8_t, left_partition_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_skip_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_txfm_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_segpred_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_intra_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_comp_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_ref_ctx)[8];
+    DECLARE_ALIGNED(8, uint8_t, left_filter_ctx)[8];
+    uint8_t *above_partition_ctx;
+    uint8_t *above_mode_ctx;
+    // FIXME maybe merge some of the below in a flags field?
+    uint8_t *above_y_nnz_ctx;
+    uint8_t *above_uv_nnz_ctx[2];
+    uint8_t *above_skip_ctx; // 1bit
+    uint8_t *above_txfm_ctx; // 2bit
+    uint8_t *above_segpred_ctx; // 1bit
+    uint8_t *above_intra_ctx; // 1bit
+    uint8_t *above_comp_ctx; // 1bit
+    uint8_t *above_ref_ctx; // 2bit
+    uint8_t *above_filter_ctx;
+    VP56mv (*above_mv_ctx)[2];
+
+    // whole-frame cache
+    uint8_t *intra_pred_data[3];
+    VP9Filter *lflvl;
+    DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];
+
+    // block reconstruction intermediates
+    int block_alloc_using_2pass;
+    int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
+    uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];
+    struct { int x, y; } min_mv, max_mv;
+    DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];
+    DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];
+    uint16_t mvscale[3][2];
+    uint8_t mvstep[3][2];
+} VP9Context;
+
+void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb);
+
+void ff_vp9_adapt_probs(VP9Context *s);
+
+void ff_vp9_decode_block(AVCodecContext *ctx, int row, int col,
+                         VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
+                         enum BlockLevel bl, enum BlockPartition bp);
+
+#endif /* AVCODEC_VP9DEC_H */

libavcodec/vp9dsp.c

@@ -23,7 +23,7 @@
 
 #include "libavutil/avassert.h"
 #include "libavutil/common.h"
-#include "vp9.h"
+#include "vp9dsp.h"
 
 const DECLARE_ALIGNED(16, int16_t, ff_vp9_subpel_filters)[3][16][8] = {
     [FILTER_8TAP_REGULAR] = {

libavcodec/vp9dsp.h

@@ -0,0 +1,136 @@
+/*
+ * VP9 compatible video decoder
+ *
+ * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
+ * Copyright (C) 2013 Clément Bœsch <u pkh me>
+ *
+ * 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_VP9DSP_H
+#define AVCODEC_VP9DSP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include "libavcodec/vp9.h"
+
+typedef void (*vp9_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
+                            const uint8_t *ref, ptrdiff_t ref_stride,
+                            int h, int mx, int my);
+typedef void (*vp9_scaled_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
+                                   const uint8_t *ref, ptrdiff_t ref_stride,
+                                   int h, int mx, int my, int dx, int dy);
+
+typedef struct VP9DSPContext {
+    /*
+     * dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32
+     * dimension 2: intra prediction modes
+     *
+     * dst/left/top is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
+     * stride is aligned by 16 pixels
+     * top[-1] is top/left; top[4,7] is top-right for 4x4
+     */
+    // FIXME(rbultje) maybe replace left/top pointers with HAVE_TOP/
+    // HAVE_LEFT/HAVE_TOPRIGHT flags instead, and then handle it in-place?
+    // also needs to fit in with what H.264/VP8/etc do
+    void (*intra_pred[N_TXFM_SIZES][N_INTRA_PRED_MODES])(uint8_t *dst,
+                                                         ptrdiff_t stride,
+                                                         const uint8_t *left,
+                                                         const uint8_t *top);
+
+    /*
+     * dimension 1: 0=4x4, 1=8x8, 2=16x16, 3=32x32, 4=lossless (3-4=dct only)
+     * dimension 2: 0=dct/dct, 1=dct/adst, 2=adst/dct, 3=adst/adst
+     *
+     * dst is aligned by transform-size (i.e. 4, 8, 16 or 32 pixels)
+     * stride is aligned by 16 pixels
+     * block is 16-byte aligned
+     * eob indicates the position (+1) of the last non-zero coefficient,
+     * in scan-order. This can be used to write faster versions, e.g. a
+     * dc-only 4x4/8x8/16x16/32x32, or a 4x4-only (eob<10) 8x8/16x16/32x32,
+     * etc.
+     */
+    // FIXME also write idct_add_block() versions for whole (inter) pred
+    // blocks, so we can do 2 4x4s at once
+    void (*itxfm_add[N_TXFM_SIZES + 1][N_TXFM_TYPES])(uint8_t *dst,
+                                                      ptrdiff_t stride,
+                                                      int16_t *block, int eob);
+
+    /*
+     * dimension 1: width of filter (0=4, 1=8, 2=16)
+     * dimension 2: 0=col-edge filter (h), 1=row-edge filter (v)
+     *
+     * dst/stride are aligned by 8
+     */
+    void (*loop_filter_8[3][2])(uint8_t *dst, ptrdiff_t stride,
+                                int mb_lim, int lim, int hev_thr);
+
+    /*
+     * dimension 1: 0=col-edge filter (h), 1=row-edge filter (v)
+     *
+     * The width of filter is assumed to be 16; dst/stride are aligned by 16
+     */
+    void (*loop_filter_16[2])(uint8_t *dst, ptrdiff_t stride,
+                              int mb_lim, int lim, int hev_thr);
+
+    /*
+     * dimension 1/2: width of filter (0=4, 1=8) for each filter half
+     * dimension 3: 0=col-edge filter (h), 1=row-edge filter (v)
+     *
+     * dst/stride are aligned by operation size
+     * this basically calls loop_filter[d1][d3][0](), followed by
+     * loop_filter[d2][d3][0]() on the next 8 pixels
+     * mb_lim/lim/hev_thr contain two values in the lowest two bytes of the
+     * integer.
+     */
+    // FIXME perhaps a mix4 that operates on 32px (for AVX2)
+    void (*loop_filter_mix2[2][2][2])(uint8_t *dst, ptrdiff_t stride,
+                                      int mb_lim, int lim, int hev_thr);
+
+    /*
+     * dimension 1: hsize (0: 64, 1: 32, 2: 16, 3: 8, 4: 4)
+     * dimension 2: filter type (0: smooth, 1: regular, 2: sharp, 3: bilin)
+     * dimension 3: averaging type (0: put, 1: avg)
+     * dimension 4: x subpel interpolation (0: none, 1: 8tap/bilin)
+     * dimension 5: y subpel interpolation (0: none, 1: 8tap/bilin)
+     *
+     * dst/stride are aligned by hsize
+     */
+    vp9_mc_func mc[5][4][2][2][2];
+
+    /*
+     * for scalable MC, first 3 dimensions identical to above, the other two
+     * don't exist since it changes per stepsize.
+     */
+    vp9_scaled_mc_func smc[5][4][2];
+} VP9DSPContext;
+
+extern const int16_t ff_vp9_subpel_filters[3][16][8];
+
+void ff_vp9dsp_init(VP9DSPContext *dsp, int bpp, int bitexact);
+
+void ff_vp9dsp_init_8(VP9DSPContext *dsp);
+void ff_vp9dsp_init_10(VP9DSPContext *dsp);
+void ff_vp9dsp_init_12(VP9DSPContext *dsp);
+
+void ff_vp9dsp_init_aarch64(VP9DSPContext *dsp, int bpp);
+void ff_vp9dsp_init_arm(VP9DSPContext *dsp, int bpp);
+void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact);
+void ff_vp9dsp_init_mips(VP9DSPContext *dsp, int bpp);
+
+#endif /* AVCODEC_VP9DSP_H */

libavcodec/vp9dsp_template.c

@@ -23,7 +23,7 @@
 
 #include "libavutil/common.h"
 #include "bit_depth_template.c"
-#include "vp9.h"
+#include "vp9dsp.h"
 
 #if BIT_DEPTH != 12
 

libavcodec/vp9mvs.c

@@ -25,6 +25,7 @@
 #include "vp56.h"
 #include "vp9.h"
 #include "vp9data.h"
+#include "vp9dec.h"
 
 static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,
                                       VP9Context *s)

libavcodec/vp9prob.c

@@ -24,6 +24,7 @@
 #include "vp56.h"
 #include "vp9.h"
 #include "vp9data.h"
+#include "vp9dec.h"
 
 static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,
                                         int max_count, int update_factor)

libavcodec/x86/vp9dsp_init.c

@@ -24,7 +24,7 @@
 #include "libavutil/cpu.h"
 #include "libavutil/mem.h"
 #include "libavutil/x86/cpu.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavcodec/x86/vp9dsp_init.h"
 
 #if HAVE_YASM

libavcodec/x86/vp9dsp_init.h

@@ -23,7 +23,7 @@
 #ifndef AVCODEC_X86_VP9DSP_INIT_H
 #define AVCODEC_X86_VP9DSP_INIT_H
 
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 
 // hack to force-expand BPC
 #define cat(a, bpp, b) a##bpp##b

libavcodec/x86/vp9dsp_init_16bpp.c

@@ -24,7 +24,7 @@
 #include "libavutil/cpu.h"
 #include "libavutil/mem.h"
 #include "libavutil/x86/cpu.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavcodec/x86/vp9dsp_init.h"
 
 #if HAVE_YASM

libavcodec/x86/vp9dsp_init_16bpp_template.c

@@ -24,7 +24,7 @@
 #include "libavutil/cpu.h"
 #include "libavutil/mem.h"
 #include "libavutil/x86/cpu.h"
-#include "libavcodec/vp9.h"
+#include "libavcodec/vp9dsp.h"
 #include "libavcodec/x86/vp9dsp_init.h"
 
 #if HAVE_YASM