[FFmpeg-devel] [PATCH 3/3] libavcodec/huffyuv: >8 bit support
Michael Niedermayer
michaelni at gmx.at
Mon Jan 13 01:22:15 CET 2014
This adds only yuv420p10, others are trivial to add and will be added
in a subsequent commit.
Currently the implementation is not optimized, optimizations will be
added later
Signed-off-by: Michael Niedermayer <michaelni at gmx.at>
---
libavcodec/huffyuv.c | 8 ++--
libavcodec/huffyuv.h | 12 +++--
libavcodec/huffyuvdec.c | 116 ++++++++++++++++++++++++++++++++++++--------
libavcodec/huffyuvenc.c | 122 ++++++++++++++++++++++++++++++++++++-----------
4 files changed, 201 insertions(+), 57 deletions(-)
diff --git a/libavcodec/huffyuv.c b/libavcodec/huffyuv.c
index b4ef76f..c183bdf 100644
--- a/libavcodec/huffyuv.c
+++ b/libavcodec/huffyuv.c
@@ -35,13 +35,13 @@
#include "avcodec.h"
#include "huffyuv.h"
-int ff_huffyuv_generate_bits_table(uint32_t *dst, const uint8_t *len_table)
+int ff_huffyuv_generate_bits_table(uint32_t *dst, const uint8_t *len_table, int n)
{
int len, index;
uint32_t bits = 0;
for (len = 32; len > 0; len--) {
- for (index = 0; index < 256; index++) {
+ for (index = 0; index < n; index++) {
if (len_table[index] == len)
dst[index] = bits++;
}
@@ -60,9 +60,10 @@ av_cold int ff_huffyuv_alloc_temp(HYuvContext *s)
if (s->bitstream_bpp<24) {
for (i=0; i<3; i++) {
- s->temp[i]= av_malloc(s->width + 16);
+ s->temp[i]= av_malloc(2*s->width + 16);
if (!s->temp[i])
return AVERROR(ENOMEM);
+ s->temp16[i] = (uint16_t*)s->temp[i];
}
} else {
s->temp[0]= av_mallocz(4*s->width + 16);
@@ -93,5 +94,6 @@ av_cold void ff_huffyuv_common_end(HYuvContext *s)
for(i = 0; i < 3; i++) {
av_freep(&s->temp[i]);
+ s->temp16[i] = NULL;
}
}
diff --git a/libavcodec/huffyuv.h b/libavcodec/huffyuv.h
index 0bd9f91..f4e0fb9 100644
--- a/libavcodec/huffyuv.h
+++ b/libavcodec/huffyuv.h
@@ -38,6 +38,9 @@
#define VLC_BITS 11
+#define MAX_BITS 14
+#define MAX_N (1<<MAX_BITS)
+
#if HAVE_BIGENDIAN
#define B 3
#define G 2
@@ -80,9 +83,10 @@ typedef struct HYuvContext {
int picture_number;
int last_slice_end;
uint8_t *temp[3];
- uint64_t stats[4][256];
- uint8_t len[4][256];
- uint32_t bits[4][256];
+ uint16_t *temp16[3]; ///< identical to temp but 16bit type
+ uint64_t stats[4][MAX_N];
+ uint8_t len[4][MAX_N];
+ uint32_t bits[4][MAX_N];
uint32_t pix_bgr_map[1<<VLC_BITS];
VLC vlc[8]; //Y,U,V,A,YY,YU,YV,AA
uint8_t *bitstream_buffer;
@@ -93,6 +97,6 @@ typedef struct HYuvContext {
void ff_huffyuv_common_init(AVCodecContext *s);
void ff_huffyuv_common_end(HYuvContext *s);
int ff_huffyuv_alloc_temp(HYuvContext *s);
-int ff_huffyuv_generate_bits_table(uint32_t *dst, const uint8_t *len_table);
+int ff_huffyuv_generate_bits_table(uint32_t *dst, const uint8_t *len_table, int n);
#endif /* AVCODEC_HUFFYUV_H */
diff --git a/libavcodec/huffyuvdec.c b/libavcodec/huffyuvdec.c
index 15fce93..e45bcd1 100644
--- a/libavcodec/huffyuvdec.c
+++ b/libavcodec/huffyuvdec.c
@@ -88,16 +88,16 @@ static const unsigned char classic_add_chroma[256] = {
6, 12, 8, 10, 7, 9, 6, 4, 6, 2, 2, 3, 3, 3, 3, 2,
};
-static int read_len_table(uint8_t *dst, GetBitContext *gb)
+static int read_len_table(uint8_t *dst, GetBitContext *gb, int n)
{
int i, val, repeat;
- for (i = 0; i < 256;) {
+ for (i = 0; i < n;) {
repeat = get_bits(gb, 3);
val = get_bits(gb, 5);
if (repeat == 0)
repeat = get_bits(gb, 8);
- if (i + repeat > 256 || get_bits_left(gb) < 0) {
+ if (i + repeat > n || get_bits_left(gb) < 0) {
av_log(NULL, AV_LOG_ERROR, "Error reading huffman table\n");
return -1;
}
@@ -118,19 +118,19 @@ static int generate_joint_tables(HYuvContext *s)
int p, i, y, u;
for (p = 0; p < 4; p++) {
int p0 = s->version > 2 ? p : 0;
- for (i = y = 0; y < 256; y++) {
+ for (i = y = 0; y < s->n; y++) {
int len0 = s->len[p0][y];
int limit = VLC_BITS - len0;
if(limit <= 0 || !len0)
continue;
- for (u = 0; u < 256; u++) {
+ for (u = 0; u < s->n; u++) {
int len1 = s->len[p][u];
if (len1 > limit || !len1)
continue;
av_assert0(i < (1 << VLC_BITS));
len[i] = len0 + len1;
bits[i] = (s->bits[p0][y] << len1) + s->bits[p][u];
- symbols[i] = (y << 8) + u;
+ symbols[i] = (y << 8) + u; //FIXME
if(symbols[i] != 0xffff) // reserved to mean "invalid"
i++;
}
@@ -199,13 +199,13 @@ static int read_huffman_tables(HYuvContext *s, const uint8_t *src, int length)
count = 1 + s->alpha + 2*s->chroma;
for (i = 0; i < count; i++) {
- if (read_len_table(s->len[i], &gb) < 0)
+ if (read_len_table(s->len[i], &gb, s->n) < 0)
return -1;
- if (ff_huffyuv_generate_bits_table(s->bits[i], s->len[i]) < 0) {
+ if (ff_huffyuv_generate_bits_table(s->bits[i], s->len[i], s->n) < 0) {
return -1;
}
ff_free_vlc(&s->vlc[i]);
- if ((ret = init_vlc(&s->vlc[i], VLC_BITS, 256, s->len[i], 1, 1,
+ if ((ret = init_vlc(&s->vlc[i], VLC_BITS, s->n, s->len[i], 1, 1,
s->bits[i], 4, 4, 0)) < 0)
return ret;
}
@@ -224,12 +224,12 @@ static int read_old_huffman_tables(HYuvContext *s)
init_get_bits(&gb, classic_shift_luma,
classic_shift_luma_table_size * 8);
- if (read_len_table(s->len[0], &gb) < 0)
+ if (read_len_table(s->len[0], &gb, 256) < 0)
return -1;
init_get_bits(&gb, classic_shift_chroma,
classic_shift_chroma_table_size * 8);
- if (read_len_table(s->len[1], &gb) < 0)
+ if (read_len_table(s->len[1], &gb, 256) < 0)
return -1;
for(i=0; i<256; i++) s->bits[0][i] = classic_add_luma [i];
@@ -397,6 +397,9 @@ static av_cold int decode_init(AVCodecContext *avctx)
case 0x675:
avctx->pix_fmt = AV_PIX_FMT_YUV420P;
break;
+ case 0x695:
+ avctx->pix_fmt = AV_PIX_FMT_YUV420P10;
+ break;
case 0x67A:
avctx->pix_fmt = AV_PIX_FMT_YUV410P;
break;
@@ -409,6 +412,11 @@ static av_cold int decode_init(AVCodecContext *avctx)
case 0x775:
avctx->pix_fmt = AV_PIX_FMT_YUVA420P;
break;
+ case 0x795:
+ avctx->pix_fmt = AV_PIX_FMT_YUVA420P10;
+ break;
+ default:
+ return AVERROR_INVALIDDATA;
}
}
@@ -499,19 +507,35 @@ static void decode_422_bitstream(HYuvContext *s, int count)
dst1 = get_vlc2(&s->gb, s->vlc[plane].table, VLC_BITS, 3);\
}\
}
+#define READ_2PIX_PLANE16(dst0, dst1, plane){\
+ dst0 = get_vlc2(&s->gb, s->vlc[plane].table, VLC_BITS, 3);\
+ dst1 = get_vlc2(&s->gb, s->vlc[plane].table, VLC_BITS, 3);\
+}
static void decode_plane_bitstream(HYuvContext *s, int count, int plane)
{
int i;
count/=2;
- if (count >= (get_bits_left(&s->gb)) / (31 * 2)) {
- for (i = 0; i < count && get_bits_left(&s->gb) > 0; i++) {
- READ_2PIX_PLANE(s->temp[0][2 * i], s->temp[0][2 * i + 1], plane);
+ if (s->bps <= 8) {
+ if (count >= (get_bits_left(&s->gb)) / (31 * 2)) {
+ for (i = 0; i < count && get_bits_left(&s->gb) > 0; i++) {
+ READ_2PIX_PLANE(s->temp[0][2 * i], s->temp[0][2 * i + 1], plane);
+ }
+ } else {
+ for(i=0; i<count; i++){
+ READ_2PIX_PLANE(s->temp[0][2 * i], s->temp[0][2 * i + 1], plane);
+ }
}
} else {
- for(i=0; i<count; i++){
- READ_2PIX_PLANE(s->temp[0][2 * i], s->temp[0][2 * i + 1], plane);
+ if (count >= (get_bits_left(&s->gb)) / (31 * 2)) {
+ for (i = 0; i < count && get_bits_left(&s->gb) > 0; i++) {
+ READ_2PIX_PLANE16(s->temp16[0][2 * i], s->temp16[0][2 * i + 1], plane);
+ }
+ } else {
+ for(i=0; i<count; i++){
+ READ_2PIX_PLANE16(s->temp16[0][2 * i], s->temp16[0][2 * i + 1], plane);
+ }
}
}
}
@@ -601,6 +625,56 @@ static void draw_slice(HYuvContext *s, AVFrame *frame, int y)
s->last_slice_end = y + h;
}
+static int left_prediction(HYuvContext *s, uint8_t *dst, const uint8_t *src, int w, int acc)
+{
+ if (s->bps <= 8) {
+ return s->dsp.add_hfyu_left_prediction(dst, src, w, acc);
+ } else {
+ //FIXME optimize
+ unsigned mask = s->n-1;
+ int i;
+ const uint16_t *src16 = (const uint16_t *)src;
+ uint16_t *dst16 = ( uint16_t *)dst;
+
+ for(i=0; i<w-1; i++){
+ acc+= src16[i];
+ dst16[i]= acc & mask;
+ i++;
+ acc+= src16[i];
+ dst16[i]= acc & mask;
+ }
+
+ for(; i<w; i++){
+ acc+= src16[i];
+ dst16[i]= acc & mask;
+ }
+
+ return acc;
+ }
+}
+
+static void add_bytes(HYuvContext *s, uint8_t *dst, uint8_t *src, int w)
+{
+ if (s->bps <= 8) {
+ s->dsp.add_bytes(dst, src, w);
+ } else {
+ //FIXME optimize
+ const uint16_t *src16 = (const uint16_t *)src;
+ uint16_t *dst16 = ( uint16_t *)dst;
+ long i;
+ unsigned long msb = 0x1000100010001ULL << (s->bps-1);
+ unsigned long lsb = msb - 0x1000100010001ULL;
+ unsigned long mask = lsb + msb;
+ for (i = 0; i <= w - (int)sizeof(long)/2; i += sizeof(long)/2) {
+ long a = *(long*)(src16+i);
+ long b = *(long*)(dst16+i);
+ *(long*)(dst16+i) = ((a&lsb) + (b&lsb)) ^ ((a^b)&msb);
+ }
+ for(; i<w; i++)
+ dst16[i] = (dst16[i] + src16[i]) & mask;
+ }
+}
+
static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
AVPacket *avpkt)
{
@@ -663,16 +737,16 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
case LEFT:
case PLANE:
decode_plane_bitstream(s, w, plane);
- left = s->dsp.add_hfyu_left_prediction(p->data[plane], s->temp[0], w, 0);
+ left = left_prediction(s, p->data[plane], s->temp[0], w, 0);
for (y = 1; y < h; y++) {
uint8_t *dst = p->data[plane] + p->linesize[plane]*y;
decode_plane_bitstream(s, w, plane);
- left = s->dsp.add_hfyu_left_prediction(dst, s->temp[0], w, left);
+ left = left_prediction(s, dst, s->temp[0], w, left);
if (s->predictor == PLANE) {
if (y > s->interlaced) {
- s->dsp.add_bytes(dst, dst - fake_stride, w);
+ add_bytes(s, dst, dst - fake_stride, w);
}
}
}
@@ -680,14 +754,14 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
break;
case MEDIAN:
decode_plane_bitstream(s, w, plane);
- left= s->dsp.add_hfyu_left_prediction(p->data[plane], s->temp[0], w, 0);
+ left= left_prediction(s, p->data[plane], s->temp[0], w, 0);
y = 1;
/* second line is left predicted for interlaced case */
if (s->interlaced) {
decode_plane_bitstream(s, w, plane);
- left = s->dsp.add_hfyu_left_prediction(p->data[plane] + p->linesize[plane], s->temp[0], w, left);
+ left = left_prediction(s, p->data[plane] + p->linesize[plane], s->temp[0], w, left);
y++;
}
diff --git a/libavcodec/huffyuvenc.c b/libavcodec/huffyuvenc.c
index bd633a2..3030f10 100644
--- a/libavcodec/huffyuvenc.c
+++ b/libavcodec/huffyuvenc.c
@@ -33,25 +33,58 @@
#include "put_bits.h"
#include "libavutil/pixdesc.h"
+static inline void diff_bytes(HYuvContext *s, uint8_t *dst,
+ const uint8_t *src0, const uint8_t *src1, int w)
+{
+ int i;
+ if (s->bps <= 8) {
+ s->dsp.diff_bytes(dst, src0, src1, w);
+ } else {
+ const uint16_t *src016 = (const uint16_t *)src0;
+ const uint16_t *src116 = (const uint16_t *)src1;
+ uint16_t *dst16 = ( uint16_t *)dst;
+
+ for (i = 0; i < w; i++) {
+ dst16[i] = src016[i] - src116[i];
+ }
+
+ //FIXME optimize
+ }
+}
+
static inline int sub_left_prediction(HYuvContext *s, uint8_t *dst,
const uint8_t *src, int w, int left)
{
int i;
- if (w < 32) {
- for (i = 0; i < w; i++) {
- const int temp = src[i];
- dst[i] = temp - left;
- left = temp;
+ if (s->bps <= 8) {
+ if (w < 32) {
+ for (i = 0; i < w; i++) {
+ const int temp = src[i];
+ dst[i] = temp - left;
+ left = temp;
+ }
+ return left;
+ } else {
+ for (i = 0; i < 16; i++) {
+ const int temp = src[i];
+ dst[i] = temp - left;
+ left = temp;
+ }
+ s->dsp.diff_bytes(dst + 16, src + 16, src + 15, w - 16);
+ return src[w-1];
}
- return left;
} else {
- for (i = 0; i < 16; i++) {
- const int temp = src[i];
- dst[i] = temp - left;
+ const uint16_t *src16 = (const uint16_t *)src;
+ uint16_t *dst16 = ( uint16_t *)dst;
+
+ for (i = 0; i < w; i++) {
+ const int temp = src16[i];
+ dst16[i] = temp - left;
left = temp;
}
- s->dsp.diff_bytes(dst + 16, src + 16, src + 15, w - 16);
- return src[w-1];
+ return left;
+
+ //FIXME optimize
}
}
@@ -122,15 +155,16 @@ static int store_table(HYuvContext *s, const uint8_t *len, uint8_t *buf)
{
int i;
int index = 0;
+ int n = s->n;
- for (i = 0; i < 256;) {
+ for (i = 0; i < n;) {
int val = len[i];
int repeat = 0;
- for (; i < 256 && len[i] == val && repeat < 255; i++)
+ for (; i < n && len[i] == val && repeat < 255; i++)
repeat++;
- av_assert0(val < 32 && val >0 && repeat<256 && repeat>0);
+ av_assert0(val < 32 && val >0 && repeat < 256 && repeat>0);
if (repeat > 7) {
buf[index++] = val;
buf[index++] = repeat;
@@ -152,10 +186,10 @@ static int store_huffman_tables(HYuvContext *s, uint8_t *buf)
count = 1 + s->alpha + 2*s->chroma;
for (i = 0; i < count; i++) {
- if ((ret = ff_huff_gen_len_table(s->len[i], s->stats[i], 256)) < 0)
+ if ((ret = ff_huff_gen_len_table(s->len[i], s->stats[i], s->n)) < 0)
return ret;
- if (ff_huffyuv_generate_bits_table(s->bits[i], s->len[i]) < 0) {
+ if (ff_huffyuv_generate_bits_table(s->bits[i], s->len[i], s->n) < 0) {
return -1;
}
@@ -173,8 +207,8 @@ static av_cold int encode_init(AVCodecContext *avctx)
ff_huffyuv_common_init(avctx);
- avctx->extradata = av_mallocz(1024*30); // 256*3+4 == 772
- avctx->stats_out = av_mallocz(1024*30); // 21*256*3(%llu ) + 3(\n) + 1(0) = 16132
+ avctx->extradata = av_mallocz(3*MAX_N + 4);
+ avctx->stats_out = av_mallocz(21*MAX_N*3 + 4); // 21*256*3(%llu ) + 3(\n) + 1(0) = 16132
if (!avctx->extradata || !avctx->stats_out) {
av_freep(&avctx->stats_out);
return AVERROR(ENOMEM);
@@ -216,6 +250,7 @@ static av_cold int encode_init(AVCodecContext *avctx)
case AV_PIX_FMT_YUVA422P:
case AV_PIX_FMT_GBRAP:
case AV_PIX_FMT_GRAY8A:
+ case AV_PIX_FMT_YUV420P10:
s->version = 3;
break;
case AV_PIX_FMT_RGB32:
@@ -301,14 +336,14 @@ static av_cold int encode_init(AVCodecContext *avctx)
char *p = avctx->stats_in;
for (i = 0; i < 4; i++)
- for (j = 0; j < 256; j++)
+ for (j = 0; j < s->n; j++)
s->stats[i][j] = 1;
for (;;) {
for (i = 0; i < 4; i++) {
char *next;
- for (j = 0; j < 256; j++) {
+ for (j = 0; j < s->n; j++) {
s->stats[i][j] += strtol(p, &next, 0);
if (next == p) return -1;
p = next;
@@ -318,8 +353,8 @@ static av_cold int encode_init(AVCodecContext *avctx)
}
} else {
for (i = 0; i < 4; i++)
- for (j = 0; j < 256; j++) {
- int d = FFMIN(j, 256 - j);
+ for (j = 0; j < s->n; j++) {
+ int d = FFMIN(j, s->n - j);
s->stats[i][j] = 100000000 / (d + 1);
}
@@ -333,14 +368,14 @@ static av_cold int encode_init(AVCodecContext *avctx)
if (s->context) {
for (i = 0; i < 4; i++) {
int pels = s->width * s->height / (i ? 40 : 10);
- for (j = 0; j < 256; j++) {
- int d = FFMIN(j, 256 - j);
+ for (j = 0; j < s->n; j++) {
+ int d = FFMIN(j, s->n - j);
s->stats[i][j] = pels/(d + 1);
}
}
} else {
for (i = 0; i < 4; i++)
- for (j = 0; j < 256; j++)
+ for (j = 0; j < s->n; j++)
s->stats[i][j]= 0;
}
@@ -412,7 +447,7 @@ static int encode_plane_bitstream(HYuvContext *s, int count, int plane)
{
int i;
- if (s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb) >> 3) < 4 * count) {
+ if (s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb) >> 3) < count * s->bps / 2) {
av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
return -1;
}
@@ -420,6 +455,9 @@ static int encode_plane_bitstream(HYuvContext *s, int count, int plane)
#define LOAD2\
int y0 = s->temp[0][2 * i];\
int y1 = s->temp[0][2 * i + 1];
+#define LOAD2_16\
+ int y0 = s->temp16[0][2 * i] & mask;\
+ int y1 = s->temp16[0][2 * i + 1] & mask;
#define STAT2\
s->stats[plane][y0]++;\
s->stats[plane][y1]++;
@@ -429,6 +467,7 @@ static int encode_plane_bitstream(HYuvContext *s, int count, int plane)
count /= 2;
+ if (s->bps <= 8) {
if (s->flags & CODEC_FLAG_PASS1) {
for (i = 0; i < count; i++) {
LOAD2;
@@ -450,6 +489,30 @@ static int encode_plane_bitstream(HYuvContext *s, int count, int plane)
WRITE2;
}
}
+ } else {
+ int mask = s->n - 1;
+ if (s->flags & CODEC_FLAG_PASS1) {
+ for (i = 0; i < count; i++) {
+ LOAD2_16;
+ STAT2;
+ }
+ }
+ if (s->avctx->flags2 & CODEC_FLAG2_NO_OUTPUT)
+ return 0;
+
+ if (s->context) {
+ for (i = 0; i < count; i++) {
+ LOAD2_16;
+ STAT2;
+ WRITE2;
+ }
+ } else {
+ for (i = 0; i < count; i++) {
+ LOAD2_16;
+ WRITE2;
+ }
+ }
+ }
#undef LOAD2
#undef STAT2
#undef WRITE2
@@ -574,7 +637,7 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
return size;
for (i = 0; i < 4; i++)
- for (j = 0; j < 256; j++)
+ for (j = 0; j < s->n; j++)
s->stats[i][j] >>= 1;
}
@@ -784,7 +847,7 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
uint8_t *dst = p->data[plane] + p->linesize[plane] * y;
if (s->predictor == PLANE && s->interlaced < y) {
- s->dsp.diff_bytes(s->temp[1], dst, dst - fake_stride, w);
+ diff_bytes(s, s->temp[1], dst, dst - fake_stride, w);
left = sub_left_prediction(s, s->temp[0], s->temp[1], w , left);
} else {
@@ -810,7 +873,7 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
char *p = avctx->stats_out;
char *end = p + 1024*30;
for (i = 0; i < 4; i++) {
- for (j = 0; j < 256; j++) {
+ for (j = 0; j < s->n; j++) {
snprintf(p, end-p, "%"PRIu64" ", s->stats[i][j]);
p += strlen(p);
s->stats[i][j]= 0;
@@ -883,6 +946,7 @@ AVCodec ff_ffvhuff_encoder = {
AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
AV_PIX_FMT_GBRAP,
AV_PIX_FMT_GRAY8A,
+ AV_PIX_FMT_YUV420P10,
AV_PIX_FMT_RGB24,
AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE
},
--
1.7.9.5
More information about the ffmpeg-devel
mailing list