Merge commit 'c33c1fa8af2b2e82418a06901b6ad17b3d61b73e'
[ffmpeg.git] / libavcodec / mpegvideo_xvmc.c
index 1749dfb..b469c4e 100644 (file)
  */
 
 #include <limits.h>
+#include <X11/extensions/XvMC.h>
 
 #include "avcodec.h"
-#include "dsputil.h"
+#include "mpegutils.h"
 #include "mpegvideo.h"
 
 #undef NDEBUG
 #include <assert.h>
 
 #include "xvmc.h"
-
-//set s->block
+#include "xvmc_internal.h"
+#include "version.h"
+
+/**
+ * Initialize the block field of the MpegEncContext pointer passed as
+ * parameter after making sure that the data is not corrupted.
+ * In order to implement something like direct rendering instead of decoding
+ * coefficients in s->blocks and then copying them, copy them directly
+ * into the data_blocks array provided by xvmc.
+ */
 void ff_xvmc_init_block(MpegEncContext *s)
 {
-    struct xvmc_render_state *render;
-    render = (struct xvmc_render_state*)s->current_picture.data[2];
-    assert(render);
-    if (!render || render->magic != AV_XVMC_RENDER_MAGIC) {
-        assert(0);
-        return; // make sure that this is a render packet
-    }
-    s->block = (DCTELEM *)(render->data_blocks + render->next_free_data_block_num * 64);
+    struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
+    assert(render && render->xvmc_id == AV_XVMC_ID);
+
+    s->block = (int16_t (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
 }
 
+static void exchange_uv(MpegEncContext *s)
+{
+    int16_t (*tmp)[64];
+
+    tmp           = s->pblocks[4];
+    s->pblocks[4] = s->pblocks[5];
+    s->pblocks[5] = tmp;
+}
+
+/**
+ * Fill individual block pointers, so there are no gaps in the data_block array
+ * in case not all blocks in the macroblock are coded.
+ */
 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
 {
-    int i, j;
+    int i, j = 0;
     const int mb_block_count = 4 + (1 << s->chroma_format);
 
-    j = 0;
     cbp <<= 12-mb_block_count;
     for (i = 0; i < mb_block_count; i++) {
         if (cbp & (1 << 11))
-            s->pblocks[i] = (short *)(&s->block[j++]);
+            s->pblocks[i] = &s->block[j++];
         else
             s->pblocks[i] = NULL;
-        cbp+=cbp;
+        cbp += cbp;
+    }
+    if (s->swap_uv) {
+        exchange_uv(s);
     }
 }
 
-// These functions should be called on every new field and/or frame.
-// They should be safe if they are called a few times for the same field!
-int ff_xvmc_field_start(MpegEncContext*s, AVCodecContext *avctx)
+/**
+ * Find and store the surfaces that are used as reference frames.
+ * This function should be called for every new field and/or frame.
+ * It should be safe to call the function a few times for the same field.
+ */
+static int ff_xvmc_field_start(AVCodecContext *avctx, const uint8_t *buf, uint32_t buf_size)
 {
-    struct xvmc_render_state *render, *last, *next;
+    struct MpegEncContext *s = avctx->priv_data;
+    struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
+    const int mb_block_count = 4 + (1 << s->chroma_format);
 
     assert(avctx);
-
-    render = (struct xvmc_render_state*)s->current_picture.data[2];
-    assert(render);
-    if (!render || render->magic != AV_XVMC_RENDER_MAGIC)
+    if (!render || render->xvmc_id != AV_XVMC_ID ||
+        !render->data_blocks || !render->mv_blocks ||
+        (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
+        (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
+        !render->p_surface) {
+        av_log(avctx, AV_LOG_ERROR,
+               "Render token doesn't look as expected.\n");
         return -1; // make sure that this is a render packet
+    }
+
+    if (render->filled_mv_blocks_num) {
+        av_log(avctx, AV_LOG_ERROR,
+               "Rendering surface contains %i unprocessed blocks.\n",
+               render->filled_mv_blocks_num);
+        return -1;
+    }
+    if (render->allocated_mv_blocks   < 1 ||
+        render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
+        render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
+        render->next_free_data_block_num >
+                        render->allocated_data_blocks -
+                        mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
+        av_log(avctx, AV_LOG_ERROR,
+               "Rendering surface doesn't provide enough block structures to work with.\n");
+        return -1;
+    }
 
     render->picture_structure = s->picture_structure;
     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
-
-    // make sure that all data is drawn by XVMC_end_frame
-    assert(render->filled_mv_blocks_num == 0);
-
-    render->p_future_surface = NULL;
-    render->p_past_surface   = NULL;
+    render->p_future_surface  = NULL;
+    render->p_past_surface    = NULL;
 
     switch(s->pict_type) {
-        case  FF_I_TYPE:
+        case  AV_PICTURE_TYPE_I:
             return 0; // no prediction from other frames
-        case  FF_B_TYPE:
-            next = (struct xvmc_render_state*)s->next_picture.data[2];
-            assert(next);
+        case  AV_PICTURE_TYPE_B:
+            next = (struct xvmc_pix_fmt*)s->next_picture.f->data[2];
             if (!next)
                 return -1;
-            if (next->magic != AV_XVMC_RENDER_MAGIC)
+            if (next->xvmc_id != AV_XVMC_ID)
                 return -1;
             render->p_future_surface = next->p_surface;
             // no return here, going to set forward prediction
-        case  FF_P_TYPE:
-            last = (struct xvmc_render_state*)s->last_picture.data[2];
-            if (!last) // && !s->first_field)
+        case  AV_PICTURE_TYPE_P:
+            last = (struct xvmc_pix_fmt*)s->last_picture.f->data[2];
+            if (!last)
                 last = render; // predict second field from the first
-            if (last->magic != AV_XVMC_RENDER_MAGIC)
+            if (last->xvmc_id != AV_XVMC_ID)
                 return -1;
             render->p_past_surface = last->p_surface;
             return 0;
@@ -106,20 +147,31 @@ int ff_xvmc_field_start(MpegEncContext*s, AVCodecContext *avctx)
 return -1;
 }
 
-void ff_xvmc_field_end(MpegEncContext *s)
+/**
+ * Complete frame/field rendering by passing any remaining blocks.
+ * Normally ff_draw_horiz_band() is called for each slice, however,
+ * some leftover blocks, for example from error_resilience(), may remain.
+ * It should be safe to call the function a few times for the same field.
+ */
+static int ff_xvmc_field_end(AVCodecContext *avctx)
 {
-    struct xvmc_render_state *render;
-    render = (struct xvmc_render_state*)s->current_picture.data[2];
+    struct MpegEncContext *s = avctx->priv_data;
+    struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
     assert(render);
 
     if (render->filled_mv_blocks_num > 0)
-        ff_draw_horiz_band(s,0,0);
+        ff_mpeg_draw_horiz_band(s, 0, 0);
+    return 0;
 }
 
-void ff_xvmc_decode_mb(MpegEncContext *s)
+/**
+ * Synthesize the data needed by XvMC to render one macroblock of data.
+ * Fill all relevant fields, if necessary do IDCT.
+ */
+static void ff_xvmc_decode_mb(struct MpegEncContext *s)
 {
     XvMCMacroBlock *mv_block;
-    struct xvmc_render_state *render;
+    struct xvmc_pix_fmt *render;
     int i, cbp, blocks_per_mb;
 
     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
@@ -130,7 +182,7 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
         return;
     }
 
-    // from MPV_decode_mb(), update DC predictors for P macroblocks
+    // from ff_mpv_decode_mb(), update DC predictors for P macroblocks
     if (!s->mb_intra) {
         s->last_dc[0] =
         s->last_dc[1] =
@@ -146,14 +198,14 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
     s->current_picture.qscale_table[mb_xy] = s->qscale;
 
     // start of XVMC-specific code
-    render = (struct xvmc_render_state*)s->current_picture.data[2];
+    render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
     assert(render);
-    assert(render->magic == AV_XVMC_RENDER_MAGIC);
+    assert(render->xvmc_id == AV_XVMC_ID);
     assert(render->mv_blocks);
 
     // take the next free macroblock
     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
-                                  render->filled_mv_blocks_num ];
+                                  render->filled_mv_blocks_num];
 
     mv_block->x        = s->mb_x;
     mv_block->y        = s->mb_y;
@@ -165,7 +217,7 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
 
         if (s->mv_dir & MV_DIR_FORWARD) {
             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
-            //pmv[n][dir][xy]=mv[dir][n][xy]
+            // PMV[n][dir][xy] = mv[dir][n][xy]
             mv_block->PMV[0][0][0] = s->mv[0][0][0];
             mv_block->PMV[0][0][1] = s->mv[0][0][1];
             mv_block->PMV[1][0][0] = s->mv[0][1][0];
@@ -179,7 +231,7 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
             mv_block->PMV[1][1][1] = s->mv[1][1][1];
         }
 
-        switch(s->mv_type){
+        switch(s->mv_type) {
             case  MV_TYPE_16X16:
                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
                 break;
@@ -199,21 +251,21 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
                 if (s->picture_structure == PICT_FRAME) {
 
-                    mv_block->PMV[0][0][0] = s->mv[0][0][0]; // top from top
-                    mv_block->PMV[0][0][1] = s->mv[0][0][1]<<1;
+                    mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
+                    mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
 
-                    mv_block->PMV[0][1][0] = s->mv[0][0][0]; // bottom from bottom
-                    mv_block->PMV[0][1][1] = s->mv[0][0][1]<<1;
+                    mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
+                    mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
 
-                    mv_block->PMV[1][0][0] = s->mv[0][2][0]; // dmv00, top from bottom
-                    mv_block->PMV[1][0][1] = s->mv[0][2][1]<<1; // dmv01
+                    mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
+                    mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
 
-                    mv_block->PMV[1][1][0] = s->mv[0][3][0]; // dmv10, bottom from top
-                    mv_block->PMV[1][1][1] = s->mv[0][3][1]<<1; // dmv11
+                    mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
+                    mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
 
                 } else {
-                    mv_block->PMV[0][1][0] = s->mv[0][2][0]; // dmv00
-                    mv_block->PMV[0][1][1] = s->mv[0][2][1]; // dmv01
+                    mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
+                    mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
                 }
                 break;
             default:
@@ -246,16 +298,16 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
             cbp++;
     }
 
-    if (s->flags & CODEC_FLAG_GRAY) {
-        if (s->mb_intra) { // intra frames are always full chroma blocks
+    if (s->avctx->flags & AV_CODEC_FLAG_GRAY) {
+        if (s->mb_intra) {                                   // intra frames are always full chroma blocks
             for (i = 4; i < blocks_per_mb; i++) {
-                memset(s->pblocks[i], 0, sizeof(short)*8*8); // so we need to clear them
+                memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
                 if (!render->unsigned_intra)
-                    s->pblocks[i][0] = 1 << 10;
+                    *s->pblocks[i][0] = 1 << 10;
             }
         } else {
             cbp &= 0xf << (blocks_per_mb - 4);
-            blocks_per_mb = 4; // luminance blocks only
+            blocks_per_mb = 4;                               // luminance blocks only
         }
     }
     mv_block->coded_block_pattern = cbp;
@@ -265,29 +317,60 @@ void ff_xvmc_decode_mb(MpegEncContext *s)
     for (i = 0; i < blocks_per_mb; i++) {
         if (s->block_last_index[i] >= 0) {
             // I do not have unsigned_intra MOCO to test, hope it is OK.
-            if (s->mb_intra && (render->idct || (!render->idct && !render->unsigned_intra)))
-                s->pblocks[i][0] -= 1 << 10;
+            if (s->mb_intra && (render->idct || !render->unsigned_intra))
+                *s->pblocks[i][0] -= 1 << 10;
             if (!render->idct) {
-                s->dsp.idct(s->pblocks[i]);
+                s->idsp.idct(*s->pblocks[i]);
                 /* It is unclear if MC hardware requires pixel diff values to be
                  * in the range [-255;255]. TODO: Clipping if such hardware is
                  * ever found. As of now it would only be an unnecessary
                  * slowdown. */
             }
             // copy blocks only if the codec doesn't support pblocks reordering
-            if (s->avctx->xvmc_acceleration == 1) {
+            if (!s->pack_pblocks) {
                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
-                       s->pblocks[i],sizeof(short)*8*8);
+                       s->pblocks[i], sizeof(*s->pblocks[i]));
             }
             render->next_free_data_block_num++;
         }
     }
     render->filled_mv_blocks_num++;
 
-    assert(render->filled_mv_blocks_num     <= render->total_number_of_mv_blocks);
-    assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
+    assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
+    assert(render->next_free_data_block_num <= render->allocated_data_blocks);
+    /* The above conditions should not be able to fail as long as this function
+     * is used and the following 'if ()' automatically calls a callback to free
+     * blocks. */
 
 
-    if (render->filled_mv_blocks_num >= render->total_number_of_mv_blocks)
-        ff_draw_horiz_band(s, 0, 0);
+    if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
+        ff_mpeg_draw_horiz_band(s, 0, 0);
 }
+
+#if CONFIG_MPEG1_XVMC_HWACCEL
+AVHWAccel ff_mpeg1_xvmc_hwaccel = {
+    .name           = "mpeg1_xvmc",
+    .type           = AVMEDIA_TYPE_VIDEO,
+    .id             = AV_CODEC_ID_MPEG1VIDEO,
+    .pix_fmt        = AV_PIX_FMT_XVMC,
+    .start_frame    = ff_xvmc_field_start,
+    .end_frame      = ff_xvmc_field_end,
+    .decode_slice   = NULL,
+    .decode_mb      = ff_xvmc_decode_mb,
+    .priv_data_size = 0,
+};
+#endif
+
+#if CONFIG_MPEG2_XVMC_HWACCEL
+AVHWAccel ff_mpeg2_xvmc_hwaccel = {
+    .name           = "mpeg2_xvmc",
+    .type           = AVMEDIA_TYPE_VIDEO,
+    .id             = AV_CODEC_ID_MPEG2VIDEO,
+    .pix_fmt        = AV_PIX_FMT_XVMC,
+    .start_frame    = ff_xvmc_field_start,
+    .end_frame      = ff_xvmc_field_end,
+    .decode_slice   = NULL,
+    .decode_mb      = ff_xvmc_decode_mb,
+    .priv_data_size = 0,
+};
+#endif