virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-01-13 21:28:01 +02:00
Code Issues Releases Activity

CrystalHD: Improve interlaced h.264 support.

Browse Source 
As previously discussed, the CrystalHD hardware returns exceptionally
useless information about interlaced h.264 content - to the extent
that it's not possible to distinguish MBAFF and PAFF content until
it's too late.

This change introduces use of the h264_parser to help bridge the
gap; it can indicate if the input data is PAFF fields or not.

With this clarity, some of heuristics can be removed from the code,
making this less convoluted.

Finally, I found an MBAFF clip that acts like non h.264 content so
I had to make allowances for that.

Note that I still cannot distinguish between two forms of PAFF,
where the hardware either returns individual fields or a field-pair.
It's not clear that there's even a spec relevant difference between
the two forms, as opposed to hardware ideosyncracies.
This commit is contained in:
Philip Langdale 2011-03-14 21:47:35 -07:00
parent 907783f221
commit e44073ca5e
2 changed files with 135 additions and 83 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
configure vendored
View File

@ -1276,7 +1276,7 @@ h263_vaapi_hwaccel_select="vaapi h263_decoder"
h263i_decoder_select="h263_decoder" h263i_decoder_select="h263_decoder"
h263p_encoder_select="h263_encoder" h263p_encoder_select="h263_encoder"
h264_decoder_select="golomb h264dsp h264pred" h264_decoder_select="golomb h264dsp h264pred"
h264_crystalhd_decoder_select="crystalhd h264_mp4toannexb_bsf" h264_crystalhd_decoder_select="crystalhd h264_mp4toannexb_bsf h264_parser"
h264_dxva2_hwaccel_deps="dxva2api_h" h264_dxva2_hwaccel_deps="dxva2api_h"
h264_dxva2_hwaccel_select="dxva2 h264_decoder" h264_dxva2_hwaccel_select="dxva2 h264_decoder"
h264_vaapi_hwaccel_select="vaapi" h264_vaapi_hwaccel_select="vaapi"

216
libavcodec/crystalhd.c
View File

@ -1,7 +1,7 @@
/* /*
* - CrystalHD decoder module - * - CrystalHD decoder module -
* *
* Copyright(C) 2010 Philip Langdale <ffmpeg.philipl@overt.org> * Copyright(C) 2010,2011 Philip Langdale <ffmpeg.philipl@overt.org>
* *
* This file is part of FFmpeg. * This file is part of FFmpeg.
* *
@ -84,6 +84,7 @@
#include <libcrystalhd/libcrystalhd_if.h> #include <libcrystalhd/libcrystalhd_if.h>
#include "avcodec.h" #include "avcodec.h"
#include "h264.h"
#include "libavutil/imgutils.h" #include "libavutil/imgutils.h"
#include "libavutil/intreadwrite.h" #include "libavutil/intreadwrite.h"
@ -102,16 +103,18 @@
****************************************************************************/ ****************************************************************************/
typedef enum { typedef enum {
RET_ERROR = -1, RET_ERROR = -1,
RET_OK = 0, RET_OK = 0,
RET_COPY_AGAIN = 1, RET_COPY_AGAIN = 1,
RET_SKIP_NEXT_COPY = 2, RET_SKIP_NEXT_COPY = 2,
RET_COPY_NEXT_FIELD = 3,
} CopyRet; } CopyRet;
typedef struct OpaqueList { typedef struct OpaqueList {
struct OpaqueList *next; struct OpaqueList *next;
uint64_t fake_timestamp; uint64_t fake_timestamp;
uint64_t reordered_opaque; uint64_t reordered_opaque;
uint8_t pic_type;
} OpaqueList; } OpaqueList;
typedef struct { typedef struct {
@ -119,6 +122,8 @@ typedef struct {
AVFrame pic; AVFrame pic;
HANDLE dev; HANDLE dev;
AVCodecParserContext *parser;
uint8_t is_70012; uint8_t is_70012;
uint8_t *sps_pps_buf; uint8_t *sps_pps_buf;
uint32_t sps_pps_size; uint32_t sps_pps_size;
@ -205,7 +210,8 @@ static inline void print_frame_info(CHDContext *priv, BC_DTS_PROC_OUT *output)
* OpaqueList functions * OpaqueList functions
****************************************************************************/ ****************************************************************************/
static uint64_t opaque_list_push(CHDContext *priv, uint64_t reordered_opaque) static uint64_t opaque_list_push(CHDContext *priv, uint64_t reordered_opaque,
uint8_t pic_type)
{ {
OpaqueList *newNode = av_mallocz(sizeof (OpaqueList)); OpaqueList *newNode = av_mallocz(sizeof (OpaqueList));
if (!newNode) { if (!newNode) {
@ -222,6 +228,7 @@ static uint64_t opaque_list_push(CHDContext *priv, uint64_t reordered_opaque)
} }
priv->tail = newNode; priv->tail = newNode;
newNode->reordered_opaque = reordered_opaque; newNode->reordered_opaque = reordered_opaque;
newNode->pic_type = pic_type;
return newNode->fake_timestamp; return newNode->fake_timestamp;
} }
@ -231,14 +238,15 @@ static uint64_t opaque_list_push(CHDContext *priv, uint64_t reordered_opaque)
* in presentation order. If frames are reordered, this means we must be * in presentation order. If frames are reordered, this means we must be
* able to remove elements that are not the first element. * able to remove elements that are not the first element.
*/ */
static uint64_t opaque_list_pop(CHDContext *priv, uint64_t fake_timestamp) static uint8_t opaque_list_pop(CHDContext *priv, uint64_t fake_timestamp,
uint64_t *reordered_opaque, uint8_t *pic_type)
{ {
OpaqueList *node = priv->head; OpaqueList *node = priv->head;
if (!priv->head) { if (!priv->head) {
av_log(priv->avctx, AV_LOG_ERROR, av_log(priv->avctx, AV_LOG_ERROR,
"CrystalHD: Attempted to query non-existent timestamps.\n"); "CrystalHD: Attempted to query non-existent timestamps.\n");
return AV_NOPTS_VALUE; return FALSE;
} }
/* /*
@ -246,14 +254,15 @@ static uint64_t opaque_list_pop(CHDContext *priv, uint64_t fake_timestamp)
* the head pointer rather than the previous element in the list. * the head pointer rather than the previous element in the list.
*/ */
if (priv->head->fake_timestamp == fake_timestamp) { if (priv->head->fake_timestamp == fake_timestamp) {
uint64_t reordered_opaque = node->reordered_opaque; *reordered_opaque = node->reordered_opaque;
*pic_type = node->pic_type;
priv->head = node->next; priv->head = node->next;
av_free(node); av_free(node);
if (!priv->head->next) if (!priv->head->next)
priv->tail = priv->head; priv->tail = priv->head;
return reordered_opaque; return TRUE;
} }
/* /*
@ -263,14 +272,15 @@ static uint64_t opaque_list_pop(CHDContext *priv, uint64_t fake_timestamp)
while (node->next) { while (node->next) {
OpaqueList *next = node->next; OpaqueList *next = node->next;
if (next->fake_timestamp == fake_timestamp) { if (next->fake_timestamp == fake_timestamp) {
uint64_t reordered_opaque = next->reordered_opaque; *reordered_opaque = node->reordered_opaque;
*pic_type = node->pic_type;
node->next = next->next; node->next = next->next;
av_free(next); av_free(next);
if (!node->next) if (!node->next)
priv->tail = node; priv->tail = node;
return reordered_opaque; return TRUE;
} else { } else {
node = next; node = next;
} }
@ -278,7 +288,7 @@ static uint64_t opaque_list_pop(CHDContext *priv, uint64_t fake_timestamp)
av_log(priv->avctx, AV_LOG_VERBOSE, av_log(priv->avctx, AV_LOG_VERBOSE,
"CrystalHD: Couldn't match fake_timestamp.\n"); "CrystalHD: Couldn't match fake_timestamp.\n");
return AV_NOPTS_VALUE; return FALSE;
} }
@ -315,6 +325,8 @@ static av_cold int uninit(AVCodecContext *avctx)
DtsCloseDecoder(device); DtsCloseDecoder(device);
DtsDeviceClose(device); DtsDeviceClose(device);
av_parser_close(priv->parser);
av_free(priv->sps_pps_buf); av_free(priv->sps_pps_buf);
if (priv->pic.data[0]) if (priv->pic.data[0])
@ -477,6 +489,13 @@ static av_cold int init(AVCodecContext *avctx)
goto fail; goto fail;
} }
if (avctx->codec->id == CODEC_ID_H264) {
priv->parser = av_parser_init(avctx->codec->id);
if (!priv->parser)
av_log(avctx, AV_LOG_WARNING,
"Cannot open the h.264 parser! Interlaced h.264 content "
"will not be detected reliably.\n");
}
av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: Init complete.\n"); av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: Init complete.\n");
return 0; return 0;
@ -492,30 +511,24 @@ static av_cold int init(AVCodecContext *avctx)
* us to distinguish between specific cases that require different handling. * us to distinguish between specific cases that require different handling.
* So, for now, we have to hard-code the behaviour we want. * So, for now, we have to hard-code the behaviour we want.
* *
* The default behaviour is to assume MBAFF with input and output fieldpairs. * Specifically, there are PAFF samples where input is always separate fields
* * but the hardware returns separate fields on one occasion and a field-pair
* Define ASSUME_PAFF_OVER_MBAFF to treat input as PAFF with separate input * on another. The code assumes the first case and define
* and output fields. * ASSUME_TWO_INPUTS_ONE_OUTPUT to assume the second case.
*
* Define ASSUME_TWO_INPUTS_ONE_OUTPUT to treat input as separate fields but
* output as a single fieldpair.
*
* Define both to mess up your playback.
*/ */
#define ASSUME_PAFF_OVER_MBAFF 0
#define ASSUME_TWO_INPUTS_ONE_OUTPUT 0 #define ASSUME_TWO_INPUTS_ONE_OUTPUT 0
static inline CopyRet copy_frame(AVCodecContext *avctx, static inline CopyRet copy_frame(AVCodecContext *avctx,
BC_DTS_PROC_OUT *output, BC_DTS_PROC_OUT *output,
void *data, int *data_size, void *data, int *data_size)
uint8_t second_field)
{ {
BC_STATUS ret; BC_STATUS ret;
BC_DTS_STATUS decoder_status; BC_DTS_STATUS decoder_status;
uint8_t is_paff; uint8_t ignore_interlaced;
uint8_t next_frame_same;
uint8_t interlaced; uint8_t interlaced;
CHDContext *priv = avctx->priv_data; CHDContext *priv = avctx->priv_data;
int64_t pkt_pts = AV_NOPTS_VALUE;
uint8_t pic_type = 0;
uint8_t bottom_field = (output->PicInfo.flags & VDEC_FLAG_BOTTOMFIELD) == uint8_t bottom_field = (output->PicInfo.flags & VDEC_FLAG_BOTTOMFIELD) ==
VDEC_FLAG_BOTTOMFIELD; VDEC_FLAG_BOTTOMFIELD;
@ -529,6 +542,27 @@ static inline CopyRet copy_frame(AVCodecContext *avctx,
uint8_t *dst; uint8_t *dst;
int dStride; int dStride;
if (output->PicInfo.timeStamp != 0) {
uint8_t pop_ret;
pop_ret = opaque_list_pop(priv, output->PicInfo.timeStamp,
&pkt_pts, &pic_type);
if (!pop_ret) {
/*
* We will encounter a situation where a timestamp cannot be
* popped if a second field is being returned. In this case,
* each field has the same timestamp and the first one will
* cause it to be popped. To keep subsequent calculations
* simple, pic_type should be set a FIELD value - doesn't
* matter which, but I chose BOTTOM.
*/
pic_type = PICT_BOTTOM_FIELD;
}
av_log(avctx, AV_LOG_VERBOSE, "output \"pts\": %"PRIu64"\n",
output->PicInfo.timeStamp);
av_log(avctx, AV_LOG_VERBOSE, "output picture type %d\n",
pic_type);
}
ret = DtsGetDriverStatus(priv->dev, &decoder_status); ret = DtsGetDriverStatus(priv->dev, &decoder_status);
if (ret != BC_STS_SUCCESS) { if (ret != BC_STS_SUCCESS) {
av_log(avctx, AV_LOG_ERROR, av_log(avctx, AV_LOG_ERROR,
@ -536,17 +570,19 @@ static inline CopyRet copy_frame(AVCodecContext *avctx,
return RET_ERROR; return RET_ERROR;
} }
is_paff = ASSUME_PAFF_OVER_MBAFF || /*
!(output->PicInfo.flags & VDEC_FLAG_UNKNOWN_SRC); * Testing has, so far, shown that we can't trust the interlaced flag for
next_frame_same = output->PicInfo.picture_number == * H.264 content when VDEC_FLAG_UNKNOWN_SRC is set.
(decoder_status.picNumFlags & ~0x40000000); */
interlaced = ((output->PicInfo.flags & ignore_interlaced = avctx->codec->id == CODEC_ID_H264 &&
VDEC_FLAG_INTERLACED_SRC) && is_paff) || (output->PicInfo.flags & VDEC_FLAG_UNKNOWN_SRC) &&
next_frame_same || bottom_field || second_field; (pic_type == 0 || pic_type == PICT_FRAME ||
ASSUME_TWO_INPUTS_ONE_OUTPUT);
interlaced = (output->PicInfo.flags & VDEC_FLAG_INTERLACED_SRC) &&
!ignore_interlaced;
av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: next_frame_same: %u | %u | %u\n", av_log(avctx, AV_LOG_VERBOSE, "Interlaced state: %d | ignore_interlaced %d\n",
next_frame_same, output->PicInfo.picture_number, interlaced, ignore_interlaced);
decoder_status.picNumFlags & ~0x40000000);
if (priv->pic.data[0] && !priv->need_second_field) if (priv->pic.data[0] && !priv->need_second_field)
avctx->release_buffer(avctx, &priv->pic); avctx->release_buffer(avctx, &priv->pic);
@ -597,8 +633,7 @@ static inline CopyRet copy_frame(AVCodecContext *avctx,
for (sY = 0; sY < height; dY++, sY++) { for (sY = 0; sY < height; dY++, sY++) {
memcpy(&(dst[dY * dStride]), &(src[sY * sStride]), bwidth); memcpy(&(dst[dY * dStride]), &(src[sY * sStride]), bwidth);
if (interlaced) dY++;
dY++;
} }
} else { } else {
av_image_copy_plane(dst, dStride, src, sStride, bwidth, height); av_image_copy_plane(dst, dStride, src, sStride, bwidth, height);
@ -608,11 +643,7 @@ static inline CopyRet copy_frame(AVCodecContext *avctx,
if (interlaced) if (interlaced)
priv->pic.top_field_first = !bottom_first; priv->pic.top_field_first = !bottom_first;
if (output->PicInfo.timeStamp != 0) { priv->pic.pkt_pts = pkt_pts;
priv->pic.pkt_pts = opaque_list_pop(priv, output->PicInfo.timeStamp);
av_log(avctx, AV_LOG_VERBOSE, "output \"pts\": %"PRIu64"\n",
priv->pic.pkt_pts);
}
if (!priv->need_second_field) { if (!priv->need_second_field) {
*data_size = sizeof(AVFrame); *data_size = sizeof(AVFrame);
@ -625,13 +656,18 @@ static inline CopyRet copy_frame(AVCodecContext *avctx,
return RET_SKIP_NEXT_COPY; return RET_SKIP_NEXT_COPY;
} }
return RET_OK; /*
* Testing has shown that in all cases where we don't want to return the
* full frame immediately, VDEC_FLAG_UNKNOWN_SRC is set.
*/
return priv->need_second_field &&
!(output->PicInfo.flags & VDEC_FLAG_UNKNOWN_SRC) ?
RET_COPY_NEXT_FIELD : RET_OK;
} }
static inline CopyRet receive_frame(AVCodecContext *avctx, static inline CopyRet receive_frame(AVCodecContext *avctx,
void *data, int *data_size, void *data, int *data_size)
uint8_t second_field)
{ {
BC_STATUS ret; BC_STATUS ret;
BC_DTS_PROC_OUT output = { BC_DTS_PROC_OUT output = {
@ -688,7 +724,7 @@ static inline CopyRet receive_frame(AVCodecContext *avctx,
priv->last_picture = output.PicInfo.picture_number - 1; priv->last_picture = output.PicInfo.picture_number - 1;
} }
copy_ret = copy_frame(avctx, &output, data, data_size, second_field); copy_ret = copy_frame(avctx, &output, data, data_size);
if (*data_size > 0) { if (*data_size > 0) {
avctx->has_b_frames--; avctx->has_b_frames--;
priv->last_picture++; priv->last_picture++;
@ -724,11 +760,29 @@ static int decode(AVCodecContext *avctx, void *data, int *data_size, AVPacket *a
CHDContext *priv = avctx->priv_data; CHDContext *priv = avctx->priv_data;
HANDLE dev = priv->dev; HANDLE dev = priv->dev;
int len = avpkt->size; int len = avpkt->size;
uint8_t pic_type = 0;
av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: decode_frame\n"); av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: decode_frame\n");
if (len) { if (len) {
int32_t tx_free = (int32_t)DtsTxFreeSize(dev); int32_t tx_free = (int32_t)DtsTxFreeSize(dev);
if (priv->parser) {
uint8_t *pout = NULL;
int psize = len;
H264Context *h = priv->parser->priv_data;
while (psize) {
ret = av_parser_parse2(priv->parser, avctx, &pout, &psize,
avpkt->data, len, avctx->pkt->pts,
avctx->pkt->dts, len - psize);
}
av_log(avctx, AV_LOG_VERBOSE,
"CrystalHD: parser picture type %d\n",
h->s.picture_structure);
pic_type = h->s.picture_structure;
}
if (len < tx_free - 1024) { if (len < tx_free - 1024) {
/* /*
* Despite being notionally opaque, either libcrystalhd or * Despite being notionally opaque, either libcrystalhd or
@ -739,7 +793,7 @@ static int decode(AVCodecContext *avctx, void *data, int *data_size, AVPacket *a
* avoiding mangling so we need to build a mapping to values * avoiding mangling so we need to build a mapping to values
* we know will not be mangled. * we know will not be mangled.
*/ */
uint64_t pts = opaque_list_push(priv, avctx->pkt->pts); uint64_t pts = opaque_list_push(priv, avctx->pkt->pts, pic_type);
if (!pts) { if (!pts) {
return AVERROR(ENOMEM); return AVERROR(ENOMEM);
} }
@ -806,41 +860,39 @@ static int decode(AVCodecContext *avctx, void *data, int *data_size, AVPacket *a
} }
do { do {
rec_ret = receive_frame(avctx, data, data_size, 0); rec_ret = receive_frame(avctx, data, data_size);
if (rec_ret == 0 && *data_size == 0) { if (rec_ret == RET_OK && *data_size == 0) {
if (avctx->codec->id == CODEC_ID_H264) { /*
/* * This case is for when the encoded fields are stored
* This case is for when the encoded fields are stored * separately and we get a separate avpkt for each one. To keep
* separately and we get a separate avpkt for each one. To keep * the pipeline stable, we should return nothing and wait for
* the pipeline stable, we should return nothing and wait for * the next time round to grab the second field.
* the next time round to grab the second field. * H.264 PAFF is an example of this.
* H.264 PAFF is an example of this. */
*/ av_log(avctx, AV_LOG_VERBOSE, "Returning after first field.\n");
av_log(avctx, AV_LOG_VERBOSE, "Returning after first field.\n"); avctx->has_b_frames--;
avctx->has_b_frames--; } else if (rec_ret == RET_COPY_NEXT_FIELD) {
} else { /*
/* * This case is for when the encoded fields are stored in a
* This case is for when the encoded fields are stored in a * single avpkt but the hardware returns then separately. Unless
* single avpkt but the hardware returns then separately. Unless * we grab the second field before returning, we'll slip another
* we grab the second field before returning, we'll slip another * frame in the pipeline and if that happens a lot, we're sunk.
* frame in the pipeline and if that happens a lot, we're sunk. * So we have to get that second field now.
* So we have to get that second field now. * Interlaced mpeg2 and vc1 are examples of this.
* Interlaced mpeg2 and vc1 are examples of this. */
*/ av_log(avctx, AV_LOG_VERBOSE, "Trying to get second field.\n");
av_log(avctx, AV_LOG_VERBOSE, "Trying to get second field.\n"); while (1) {
while (1) { usleep(priv->decode_wait);
usleep(priv->decode_wait); ret = DtsGetDriverStatus(dev, &decoder_status);
ret = DtsGetDriverStatus(dev, &decoder_status); if (ret == BC_STS_SUCCESS &&
if (ret == BC_STS_SUCCESS && decoder_status.ReadyListCount > 0) {
decoder_status.ReadyListCount > 0) { rec_ret = receive_frame(avctx, data, data_size);
rec_ret = receive_frame(avctx, data, data_size, 1); if ((rec_ret == RET_OK && *data_size > 0) ||
if ((rec_ret == 0 && *data_size > 0) || rec_ret == RET_ERROR)
rec_ret == RET_ERROR) break;
break;
}
} }
av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: Got second field.\n");
} }
av_log(avctx, AV_LOG_VERBOSE, "CrystalHD: Got second field.\n");
} else if (rec_ret == RET_SKIP_NEXT_COPY) { } else if (rec_ret == RET_SKIP_NEXT_COPY) {
/* /*
* Two input packets got turned into a field pair. Gawd. * Two input packets got turned into a field pair. Gawd.