dnxhddec: implement slice multithreading

Around 3x speedup with 4 threads. Maybe more mb lines should be
batched per thread, but that's good enough for a first try.

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This commit is contained in:
Christophe Gisquet 2015-09-25 08:56:43 +02:00 committed by Michael Niedermayer
parent 265b106b92
commit 08a7510fca

View File

@ -32,10 +32,21 @@
#include "internal.h"
#include "thread.h"
typedef struct RowContext {
DECLARE_ALIGNED(16, int16_t, blocks)[12][64];
int luma_scale[64];
int chroma_scale[64];
GetBitContext gb;
int last_dc[3];
int last_qscale;
} RowContext;
typedef struct DNXHDContext {
AVCodecContext *avctx;
GetBitContext gb;
RowContext *rows;
BlockDSPContext bdsp;
const uint8_t* buf;
int buf_size;
int64_t cid; ///< compression id
unsigned int width, height;
enum AVPixelFormat pix_fmt;
@ -43,30 +54,29 @@ typedef struct DNXHDContext {
uint32_t mb_scan_index[68]; /* max for 1080p */
int cur_field; ///< current interlaced field
VLC ac_vlc, dc_vlc, run_vlc;
int last_dc[3];
IDCTDSPContext idsp;
DECLARE_ALIGNED(16, int16_t, blocks)[12][64];
ScanTable scantable;
const CIDEntry *cid_table;
int bit_depth; // 8, 10 or 0 if not initialized at all.
int is_444;
int mbaff;
int act;
void (*decode_dct_block)(struct DNXHDContext *ctx, int16_t *block,
void (*decode_dct_block)(const struct DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale);
int last_qscale;
int luma_scale[64];
int chroma_scale[64];
} DNXHDContext;
#define DNXHD_VLC_BITS 9
#define DNXHD_DC_VLC_BITS 7
static void dnxhd_decode_dct_block_8(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_8(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale);
static void dnxhd_decode_dct_block_10(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_10(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale);
static void dnxhd_decode_dct_block_10_444(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_10_444(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale);
static av_cold int dnxhd_decode_init(AVCodecContext *avctx)
@ -76,6 +86,11 @@ static av_cold int dnxhd_decode_init(AVCodecContext *avctx)
ctx->avctx = avctx;
ctx->cid = -1;
avctx->colorspace = AVCOL_SPC_BT709;
ctx->rows = av_mallocz_array(avctx->thread_count, sizeof(RowContext));
if (!ctx->rows)
return AVERROR(ENOMEM);
return 0;
}
@ -116,6 +131,20 @@ static int dnxhd_init_vlc(DNXHDContext *ctx, uint32_t cid)
return 0;
}
static av_cold int dnxhd_decode_init_thread_copy(AVCodecContext *avctx)
{
DNXHDContext *ctx = avctx->priv_data;
// make sure VLC tables will be loaded when cid is parsed
ctx->cid = -1;
ctx->rows = av_mallocz_array(avctx->thread_count, sizeof(RowContext));
if (!ctx->rows)
return AVERROR(ENOMEM);
return 0;
}
static int dnxhd_decode_header(DNXHDContext *ctx, AVFrame *frame,
const uint8_t *buf, int buf_size,
int first_field)
@ -240,7 +269,8 @@ static int dnxhd_decode_header(DNXHDContext *ctx, AVFrame *frame,
return 0;
}
static av_always_inline void dnxhd_decode_dct_block(DNXHDContext *ctx,
static av_always_inline void dnxhd_decode_dct_block(const DNXHDContext *ctx,
RowContext *row,
int16_t *block, int n,
int qscale,
int index_bits,
@ -254,61 +284,61 @@ static av_always_inline void dnxhd_decode_dct_block(DNXHDContext *ctx,
const uint8_t *ac_level = ctx->cid_table->ac_level;
const uint8_t *ac_flags = ctx->cid_table->ac_flags;
const int eob_index = ctx->cid_table->eob_index;
OPEN_READER(bs, &ctx->gb);
OPEN_READER(bs, &row->gb);
if (!ctx->is_444) {
if (n & 2) {
component = 1 + (n & 1);
scale = ctx->chroma_scale;
scale = row->chroma_scale;
weight_matrix = ctx->cid_table->chroma_weight;
} else {
component = 0;
scale = ctx->luma_scale;
scale = row->luma_scale;
weight_matrix = ctx->cid_table->luma_weight;
}
} else {
component = (n >> 1) % 3;
if (component) {
scale = ctx->chroma_scale;
scale = row->chroma_scale;
weight_matrix = ctx->cid_table->chroma_weight;
} else {
scale = ctx->luma_scale;
scale = row->luma_scale;
weight_matrix = ctx->cid_table->luma_weight;
}
}
UPDATE_CACHE(bs, &ctx->gb);
GET_VLC(len, bs, &ctx->gb, ctx->dc_vlc.table, DNXHD_DC_VLC_BITS, 1);
UPDATE_CACHE(bs, &row->gb);
GET_VLC(len, bs, &row->gb, ctx->dc_vlc.table, DNXHD_DC_VLC_BITS, 1);
if (len) {
level = GET_CACHE(bs, &ctx->gb);
LAST_SKIP_BITS(bs, &ctx->gb, len);
level = GET_CACHE(bs, &row->gb);
LAST_SKIP_BITS(bs, &row->gb, len);
sign = ~level >> 31;
level = (NEG_USR32(sign ^ level, len) ^ sign) - sign;
ctx->last_dc[component] += level;
row->last_dc[component] += level;
}
block[0] = ctx->last_dc[component];
block[0] = row->last_dc[component];
i = 0;
UPDATE_CACHE(bs, &ctx->gb);
GET_VLC(index1, bs, &ctx->gb, ctx->ac_vlc.table,
UPDATE_CACHE(bs, &row->gb);
GET_VLC(index1, bs, &row->gb, ctx->ac_vlc.table,
DNXHD_VLC_BITS, 2);
while (index1 != eob_index) {
level = ac_level[index1];
flags = ac_flags[index1];
sign = SHOW_SBITS(bs, &ctx->gb, 1);
SKIP_BITS(bs, &ctx->gb, 1);
sign = SHOW_SBITS(bs, &row->gb, 1);
SKIP_BITS(bs, &row->gb, 1);
if (flags & 1) {
level += SHOW_UBITS(bs, &ctx->gb, index_bits) << 7;
SKIP_BITS(bs, &ctx->gb, index_bits);
level += SHOW_UBITS(bs, &row->gb, index_bits) << 7;
SKIP_BITS(bs, &row->gb, index_bits);
}
if (flags & 2) {
UPDATE_CACHE(bs, &ctx->gb);
GET_VLC(index2, bs, &ctx->gb, ctx->run_vlc.table,
UPDATE_CACHE(bs, &row->gb);
GET_VLC(index2, bs, &row->gb, ctx->run_vlc.table,
DNXHD_VLC_BITS, 2);
i += ctx->cid_table->run[index2];
}
@ -326,34 +356,37 @@ static av_always_inline void dnxhd_decode_dct_block(DNXHDContext *ctx,
block[j] = (level ^ sign) - sign;
UPDATE_CACHE(bs, &ctx->gb);
GET_VLC(index1, bs, &ctx->gb, ctx->ac_vlc.table,
UPDATE_CACHE(bs, &row->gb);
GET_VLC(index1, bs, &row->gb, ctx->ac_vlc.table,
DNXHD_VLC_BITS, 2);
}
CLOSE_READER(bs, &ctx->gb);
CLOSE_READER(bs, &row->gb);
}
static void dnxhd_decode_dct_block_8(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_8(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale)
{
dnxhd_decode_dct_block(ctx, block, n, qscale, 4, 32, 6);
dnxhd_decode_dct_block(ctx, row, block, n, qscale, 4, 32, 6);
}
static void dnxhd_decode_dct_block_10(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_10(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale)
{
dnxhd_decode_dct_block(ctx, block, n, qscale, 6, 8, 4);
dnxhd_decode_dct_block(ctx, row, block, n, qscale, 6, 8, 4);
}
static void dnxhd_decode_dct_block_10_444(DNXHDContext *ctx, int16_t *block,
static void dnxhd_decode_dct_block_10_444(const DNXHDContext *ctx,
RowContext *row, int16_t *block,
int n, int qscale)
{
dnxhd_decode_dct_block(ctx, block, n, qscale, 6, 32, 6);
dnxhd_decode_dct_block(ctx, row, block, n, qscale, 6, 32, 6);
}
static int dnxhd_decode_macroblock(DNXHDContext *ctx, AVFrame *frame,
int x, int y)
static int dnxhd_decode_macroblock(const DNXHDContext *ctx, RowContext *row,
AVFrame *frame, int x, int y)
{
int shift1 = ctx->bit_depth == 10;
int dct_linesize_luma = frame->linesize[0];
@ -364,11 +397,11 @@ static int dnxhd_decode_macroblock(DNXHDContext *ctx, AVFrame *frame,
int interlaced_mb = 0;
if (ctx->mbaff) {
interlaced_mb = get_bits1(&ctx->gb);
qscale = get_bits(&ctx->gb, 10);
interlaced_mb = get_bits1(&row->gb);
qscale = get_bits(&row->gb, 10);
} else
qscale = get_bits(&ctx->gb, 11);
act = get_bits1(&ctx->gb);
qscale = get_bits(&row->gb, 11);
act = get_bits1(&row->gb);
if (act) {
static int warned = 0;
if (!warned) {
@ -378,22 +411,22 @@ static int dnxhd_decode_macroblock(DNXHDContext *ctx, AVFrame *frame,
}
}
if (qscale != ctx->last_qscale) {
if (qscale != row->last_qscale) {
for (i = 0; i < 64; i++) {
ctx->luma_scale[i] = qscale * ctx->cid_table->luma_weight[i];
ctx->chroma_scale[i] = qscale * ctx->cid_table->chroma_weight[i];
row->luma_scale[i] = qscale * ctx->cid_table->luma_weight[i];
row->chroma_scale[i] = qscale * ctx->cid_table->chroma_weight[i];
}
ctx->last_qscale = qscale;
row->last_qscale = qscale;
}
for (i = 0; i < 8; i++) {
ctx->bdsp.clear_block(ctx->blocks[i]);
ctx->decode_dct_block(ctx, ctx->blocks[i], i, qscale);
ctx->bdsp.clear_block(row->blocks[i]);
ctx->decode_dct_block(ctx, row, row->blocks[i], i, qscale);
}
if (ctx->is_444) {
for (; i < 12; i++) {
ctx->bdsp.clear_block(ctx->blocks[i]);
ctx->decode_dct_block(ctx, ctx->blocks[i], i, qscale);
ctx->bdsp.clear_block(row->blocks[i]);
ctx->decode_dct_block(ctx, row, row->blocks[i], i, qscale);
}
}
@ -419,55 +452,58 @@ static int dnxhd_decode_macroblock(DNXHDContext *ctx, AVFrame *frame,
dct_y_offset = interlaced_mb ? frame->linesize[0] : (dct_linesize_luma << 3);
dct_x_offset = 8 << shift1;
if (!ctx->is_444) {
ctx->idsp.idct_put(dest_y, dct_linesize_luma, ctx->blocks[0]);
ctx->idsp.idct_put(dest_y + dct_x_offset, dct_linesize_luma, ctx->blocks[1]);
ctx->idsp.idct_put(dest_y + dct_y_offset, dct_linesize_luma, ctx->blocks[4]);
ctx->idsp.idct_put(dest_y + dct_y_offset + dct_x_offset, dct_linesize_luma, ctx->blocks[5]);
ctx->idsp.idct_put(dest_y, dct_linesize_luma, row->blocks[0]);
ctx->idsp.idct_put(dest_y + dct_x_offset, dct_linesize_luma, row->blocks[1]);
ctx->idsp.idct_put(dest_y + dct_y_offset, dct_linesize_luma, row->blocks[4]);
ctx->idsp.idct_put(dest_y + dct_y_offset + dct_x_offset, dct_linesize_luma, row->blocks[5]);
if (!(ctx->avctx->flags & AV_CODEC_FLAG_GRAY)) {
dct_y_offset = interlaced_mb ? frame->linesize[1] : (dct_linesize_chroma << 3);
ctx->idsp.idct_put(dest_u, dct_linesize_chroma, ctx->blocks[2]);
ctx->idsp.idct_put(dest_v, dct_linesize_chroma, ctx->blocks[3]);
ctx->idsp.idct_put(dest_u + dct_y_offset, dct_linesize_chroma, ctx->blocks[6]);
ctx->idsp.idct_put(dest_v + dct_y_offset, dct_linesize_chroma, ctx->blocks[7]);
ctx->idsp.idct_put(dest_u, dct_linesize_chroma, row->blocks[2]);
ctx->idsp.idct_put(dest_v, dct_linesize_chroma, row->blocks[3]);
ctx->idsp.idct_put(dest_u + dct_y_offset, dct_linesize_chroma, row->blocks[6]);
ctx->idsp.idct_put(dest_v + dct_y_offset, dct_linesize_chroma, row->blocks[7]);
}
} else {
ctx->idsp.idct_put(dest_y, dct_linesize_luma, ctx->blocks[0]);
ctx->idsp.idct_put(dest_y + dct_x_offset, dct_linesize_luma, ctx->blocks[1]);
ctx->idsp.idct_put(dest_y + dct_y_offset, dct_linesize_luma, ctx->blocks[6]);
ctx->idsp.idct_put(dest_y + dct_y_offset + dct_x_offset, dct_linesize_luma, ctx->blocks[7]);
ctx->idsp.idct_put(dest_y, dct_linesize_luma, row->blocks[0]);
ctx->idsp.idct_put(dest_y + dct_x_offset, dct_linesize_luma, row->blocks[1]);
ctx->idsp.idct_put(dest_y + dct_y_offset, dct_linesize_luma, row->blocks[6]);
ctx->idsp.idct_put(dest_y + dct_y_offset + dct_x_offset, dct_linesize_luma, row->blocks[7]);
if (!(ctx->avctx->flags & AV_CODEC_FLAG_GRAY)) {
dct_y_offset = interlaced_mb ? frame->linesize[1] : (dct_linesize_chroma << 3);
ctx->idsp.idct_put(dest_u, dct_linesize_chroma, ctx->blocks[2]);
ctx->idsp.idct_put(dest_u + dct_x_offset, dct_linesize_chroma, ctx->blocks[3]);
ctx->idsp.idct_put(dest_u + dct_y_offset, dct_linesize_chroma, ctx->blocks[8]);
ctx->idsp.idct_put(dest_u + dct_y_offset + dct_x_offset, dct_linesize_chroma, ctx->blocks[9]);
ctx->idsp.idct_put(dest_v, dct_linesize_chroma, ctx->blocks[4]);
ctx->idsp.idct_put(dest_v + dct_x_offset, dct_linesize_chroma, ctx->blocks[5]);
ctx->idsp.idct_put(dest_v + dct_y_offset, dct_linesize_chroma, ctx->blocks[10]);
ctx->idsp.idct_put(dest_v + dct_y_offset + dct_x_offset, dct_linesize_chroma, ctx->blocks[11]);
ctx->idsp.idct_put(dest_u, dct_linesize_chroma, row->blocks[2]);
ctx->idsp.idct_put(dest_u + dct_x_offset, dct_linesize_chroma, row->blocks[3]);
ctx->idsp.idct_put(dest_u + dct_y_offset, dct_linesize_chroma, row->blocks[8]);
ctx->idsp.idct_put(dest_u + dct_y_offset + dct_x_offset, dct_linesize_chroma, row->blocks[9]);
ctx->idsp.idct_put(dest_v, dct_linesize_chroma, row->blocks[4]);
ctx->idsp.idct_put(dest_v + dct_x_offset, dct_linesize_chroma, row->blocks[5]);
ctx->idsp.idct_put(dest_v + dct_y_offset, dct_linesize_chroma, row->blocks[10]);
ctx->idsp.idct_put(dest_v + dct_y_offset + dct_x_offset, dct_linesize_chroma, row->blocks[11]);
}
}
return 0;
}
static int dnxhd_decode_macroblocks(DNXHDContext *ctx, AVFrame *frame,
const uint8_t *buf, int buf_size)
static int dnxhd_decode_row(AVCodecContext *avctx, void *data,
int rownb, int threadnb)
{
int x, y;
for (y = 0; y < ctx->mb_height; y++) {
ctx->last_dc[0] =
ctx->last_dc[1] =
ctx->last_dc[2] = 1 << (ctx->bit_depth + 2); // for levels +2^(bitdepth-1)
init_get_bits(&ctx->gb, buf + ctx->mb_scan_index[y], (buf_size - ctx->mb_scan_index[y]) << 3);
for (x = 0; x < ctx->mb_width; x++) {
//START_TIMER;
dnxhd_decode_macroblock(ctx, frame, x, y);
//STOP_TIMER("decode macroblock");
}
const DNXHDContext *ctx = avctx->priv_data;
uint32_t offset = ctx->mb_scan_index[rownb];
RowContext *row = ctx->rows + threadnb;
int x;
row->last_dc[0] =
row->last_dc[1] =
row->last_dc[2] = 1 << (ctx->bit_depth + 2); // for levels +2^(bitdepth-1)
init_get_bits(&row->gb, ctx->buf + offset, (ctx->buf_size - offset) << 3);
for (x = 0; x < ctx->mb_width; x++) {
//START_TIMER;
dnxhd_decode_macroblock(ctx, row, data, x, rownb);
//STOP_TIMER("decode macroblock");
}
return 0;
}
@ -512,7 +548,9 @@ decode_coding_unit:
picture->key_frame = 1;
}
dnxhd_decode_macroblocks(ctx, picture, buf + 0x280, buf_size - 0x280);
ctx->buf_size = buf_size - 0x280;
ctx->buf = buf + 0x280;
avctx->execute2(avctx, dnxhd_decode_row, picture, NULL, ctx->mb_height);
if (first_field && picture->interlaced_frame) {
buf += ctx->cid_table->coding_unit_size;
@ -532,6 +570,9 @@ static av_cold int dnxhd_decode_close(AVCodecContext *avctx)
ff_free_vlc(&ctx->ac_vlc);
ff_free_vlc(&ctx->dc_vlc);
ff_free_vlc(&ctx->run_vlc);
av_freep(&ctx->rows);
return 0;
}
@ -544,5 +585,7 @@ AVCodec ff_dnxhd_decoder = {
.init = dnxhd_decode_init,
.close = dnxhd_decode_close,
.decode = dnxhd_decode_frame,
.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS |
AV_CODEC_CAP_SLICE_THREADS,
.init_thread_copy = ONLY_IF_THREADS_ENABLED(dnxhd_decode_init_thread_copy),
};