comparison dnxhdenc.c @ 10387:19a4f1ecd8fe libavcodec

Move dnxhdenc to execute2 multithreading API. This allows for some simplifications like removing some outer loops and gives much better performance for thread_count > number of idle CPUs.

10386:98501365c3aa

        return -1;
    if (dnxhd_init_rc(ctx) < 0)
        return -1;

    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t), fail);
    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t) , fail);

    ctx->frame.key_frame = 1;
    ctx->frame.pict_type = FF_I_TYPE;
    ctx->m.avctx->coded_frame = &ctx->frame;

    if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
        av_log(avctx, AV_LOG_ERROR, "too many threads\n");
        return -1;
    }

    ctx->thread[0] = ctx;
    for (i = 1; i < avctx->thread_count; i++) {
        ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
        memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
    }

    for (i = 0; i < avctx->thread_count; i++) {
        ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;
        ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
    }

    return 0;
 fail: //for FF_ALLOCZ_OR_GOTO
    return -1;

        ctx->m.q_intra_matrix   = ctx->qmatrix_l;
        return 0;
    }
}

static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
{
    DNXHDEncContext *ctx = *(void**)arg;
    int mb_y, mb_x;
    int qscale = ctx->thread[0]->qscale;

    for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
        ctx->m.last_dc[0] =
        ctx->m.last_dc[1] =
        ctx->m.last_dc[2] = 1024;

        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {

                }
            }
            ctx->mb_rc[qscale][mb].ssd = ssd;
            ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
        }
    }
    return 0;
}

static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
{
    DNXHDEncContext *ctx = *(void**)arg;
    int mb_y, mb_x;

    for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
        ctx->m.last_dc[0] =
        ctx->m.last_dc[1] =
        ctx->m.last_dc[2] = 1024;
        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
            unsigned mb = mb_y * ctx->m.mb_width + mb_x;

                //STOP_TIMER("encode_block");
            }
        }
        if (put_bits_count(&ctx->m.pb)&31)
            put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
    }
    flush_put_bits(&ctx->m.pb);
    return 0;
}

static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
{
    int mb_y, mb_x;
    int i, offset = 0;
    for (i = 0; i < ctx->m.avctx->thread_count; i++) {
        int thread_size = 0;
        for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
            ctx->slice_size[mb_y] = 0;
            for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
                unsigned mb = mb_y * ctx->m.mb_width + mb_x;
                ctx->slice_size[mb_y] += ctx->mb_bits[mb];
            }
            ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
            ctx->slice_size[mb_y] >>= 3;
            thread_size += ctx->slice_size[mb_y];
        }
        init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
        offset += thread_size;
    }
}

static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
{
    DNXHDEncContext *ctx = *(void**)arg;
    int mb_y, mb_x;
    for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
            unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
            uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
            int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
            int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
            ctx->mb_cmp[mb].value = varc;
            ctx->mb_cmp[mb].mb = mb;
        }
    }
    return 0;
}

static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
{

    int last_lower = INT_MAX, last_higher = 0;
    int x, y, q;

    for (q = 1; q < avctx->qmax; q++) {
        ctx->qscale = q;
        avctx->execute(avctx, dnxhd_calc_bits_thread, &ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
    }
    up_step = down_step = 2<<LAMBDA_FRAC_BITS;
    lambda = ctx->lambda;

    for (;;) {

    qscale = ctx->qscale;
    for (;;) {
        bits = 0;
        ctx->qscale = qscale;
        // XXX avoid recalculating bits
        ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, &ctx->thread[0], NULL, ctx->m.avctx->thread_count, sizeof(void*));
        for (y = 0; y < ctx->m.mb_height; y++) {
            for (x = 0; x < ctx->m.mb_width; x++)
                bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
            bits = (bits+31)&~31; // padding
            if (bits > ctx->frame_bits)

        }
        max_bits += 31; //worst padding
    }
    if (!ret) {
        if (RC_VARIANCE)
            avctx->execute(avctx, dnxhd_mb_var_thread, &ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
        radix_sort(ctx->mb_cmp, ctx->m.mb_num);
        for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
            int mb = ctx->mb_cmp[x].mb;
            max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
            ctx->mb_qscale[mb] = ctx->qscale+1;

    if (ret < 0) {
        av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
        return -1;
    }

    dnxhd_setup_threads_slices(ctx, buf);

    offset = 0;
    for (i = 0; i < ctx->m.mb_height; i++) {
        AV_WB32(ctx->msip + i * 4, offset);
        offset += ctx->slice_size[i];
        assert(!(ctx->slice_size[i] & 3));
    }

    avctx->execute(avctx, dnxhd_encode_thread, &ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));

    assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
    memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);

    AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF

    av_freep(&ctx->mb_bits);
    av_freep(&ctx->mb_qscale);
    av_freep(&ctx->mb_rc);
    av_freep(&ctx->mb_cmp);
    av_freep(&ctx->slice_size);

    av_freep(&ctx->qmatrix_c);
    av_freep(&ctx->qmatrix_l);
    av_freep(&ctx->qmatrix_c16);
    av_freep(&ctx->qmatrix_l16);

10387:19a4f1ecd8fe

        return -1;
    if (dnxhd_init_rc(ctx) < 0)
        return -1;

    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t), fail);
    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t) , fail);

    ctx->frame.key_frame = 1;
    ctx->frame.pict_type = FF_I_TYPE;
    ctx->m.avctx->coded_frame = &ctx->frame;

    if (avctx->thread_count > MAX_THREADS) {
        av_log(avctx, AV_LOG_ERROR, "too many threads\n");
        return -1;
    }

    ctx->thread[0] = ctx;
    for (i = 1; i < avctx->thread_count; i++) {
        ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
        memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
    }

    return 0;
 fail: //for FF_ALLOCZ_OR_GOTO
    return -1;

        ctx->m.q_intra_matrix   = ctx->qmatrix_l;
        return 0;
    }
}

static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
{
    DNXHDEncContext *ctx = avctx->priv_data;
    int mb_y = jobnr, mb_x;
    int qscale = ctx->qscale;
    ctx = ctx->thread[threadnr];

        ctx->m.last_dc[0] =
        ctx->m.last_dc[1] =
        ctx->m.last_dc[2] = 1024;

        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {

                }
            }
            ctx->mb_rc[qscale][mb].ssd = ssd;
            ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
        }
    return 0;
}

static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
{
    DNXHDEncContext *ctx = avctx->priv_data;
    int mb_y = jobnr, mb_x;
    ctx = ctx->thread[threadnr];
    init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);

        ctx->m.last_dc[0] =
        ctx->m.last_dc[1] =
        ctx->m.last_dc[2] = 1024;
        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
            unsigned mb = mb_y * ctx->m.mb_width + mb_x;

                //STOP_TIMER("encode_block");
            }
        }
        if (put_bits_count(&ctx->m.pb)&31)
            put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
    flush_put_bits(&ctx->m.pb);
    return 0;
}

static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
{
    int mb_y, mb_x;
    int offset = 0;
    for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
        int thread_size;
        ctx->slice_offs[mb_y] = offset;
            ctx->slice_size[mb_y] = 0;
            for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
                unsigned mb = mb_y * ctx->m.mb_width + mb_x;
                ctx->slice_size[mb_y] += ctx->mb_bits[mb];
            }
            ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
            ctx->slice_size[mb_y] >>= 3;
            thread_size = ctx->slice_size[mb_y];
        offset += thread_size;
    }
}

static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
{
    DNXHDEncContext *ctx = avctx->priv_data;
    int mb_y = jobnr, mb_x;
    ctx = ctx->thread[threadnr];
        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
            unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
            uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
            int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
            int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
            ctx->mb_cmp[mb].value = varc;
            ctx->mb_cmp[mb].mb = mb;
        }
    return 0;
}

static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
{

    int last_lower = INT_MAX, last_higher = 0;
    int x, y, q;

    for (q = 1; q < avctx->qmax; q++) {
        ctx->qscale = q;
        avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
    }
    up_step = down_step = 2<<LAMBDA_FRAC_BITS;
    lambda = ctx->lambda;

    for (;;) {

    qscale = ctx->qscale;
    for (;;) {
        bits = 0;
        ctx->qscale = qscale;
        // XXX avoid recalculating bits
        ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
        for (y = 0; y < ctx->m.mb_height; y++) {
            for (x = 0; x < ctx->m.mb_width; x++)
                bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
            bits = (bits+31)&~31; // padding
            if (bits > ctx->frame_bits)

        }
        max_bits += 31; //worst padding
    }
    if (!ret) {
        if (RC_VARIANCE)
            avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
        radix_sort(ctx->mb_cmp, ctx->m.mb_num);
        for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
            int mb = ctx->mb_cmp[x].mb;
            max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
            ctx->mb_qscale[mb] = ctx->qscale+1;

    if (ret < 0) {
        av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
        return -1;
    }

    dnxhd_setup_threads_slices(ctx);

    offset = 0;
    for (i = 0; i < ctx->m.mb_height; i++) {
        AV_WB32(ctx->msip + i * 4, offset);
        offset += ctx->slice_size[i];
        assert(!(ctx->slice_size[i] & 3));
    }

    avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);

    assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
    memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);

    AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF

    av_freep(&ctx->mb_bits);
    av_freep(&ctx->mb_qscale);
    av_freep(&ctx->mb_rc);
    av_freep(&ctx->mb_cmp);
    av_freep(&ctx->slice_size);
    av_freep(&ctx->slice_offs);

    av_freep(&ctx->qmatrix_c);
    av_freep(&ctx->qmatrix_l);
    av_freep(&ctx->qmatrix_c16);
    av_freep(&ctx->qmatrix_l16);