[FFmpeg-devel] HQX codec path: faster decode with multi threaded decoding of slices; fixed clipping bug

[Michael Niedermayer]

On Wed, Apr 08, 2015 at 09:26:05AM +0200, Ferdinand Oeinck wrote:

[...]
> -    put_blocks(pic, 0, x,      y, flag, ctx->block[0], ctx->block[2], hqx_quant_luma);
> -    put_blocks(pic, 0, x + 8,  y, flag, ctx->block[1], ctx->block[3], hqx_quant_luma);
> -    put_blocks(pic, 2, x >> 1, y, flag, ctx->block[4], ctx->block[5], hqx_quant_chroma);
> -    put_blocks(pic, 1, x >> 1, y, flag, ctx->block[6], ctx->block[7], hqx_quant_chroma);
> +	put_blocks(pic, 0, x, y, flag, slice_data->block[0], slice_data->block[2], hqx_quant_luma);
> +	put_blocks(pic, 0, x + 8, y, flag, slice_data->block[1], slice_data->block[3], hqx_quant_luma);
> +	put_blocks(pic, 2, x >> 1, y, flag, slice_data->block[4], slice_data->block[5], hqx_quant_chroma);
> +	put_blocks(pic, 1, x >> 1, y, flag, slice_data->block[6], slice_data->block[7], hqx_quant_chroma);

tabs are forbidden in ffmpeg git


[...]
> +static int decode_slice_thread(AVCodecContext *avctx, void *arg, int slice, int threadnr)
> +{
> +	DECLARE_ALIGNED(16, int16_t, block)[16][64];
> +
> +	data_t * data = (data_t*) arg;
> +	uint32_t * slice_off = data->slice_off;
> +	unsigned data_size = data->data_size;
> +	if (slice_off[slice] < HQX_HEADER_SIZE ||
> +		slice_off[slice] >= slice_off[slice + 1] ||
> +		slice_off[slice + 1] > data_size) {
> +		av_log(avctx, AV_LOG_ERROR, "Invalid slice size.\n");
> +		return AVERROR_INVALIDDATA;
> +	}
> +	int ret = init_get_bits8(&data->gb[slice], data->src + slice_off[slice],
> +		slice_off[slice + 1] - slice_off[slice]);

please dont mix declarations and statements


> +	if (ret < 0)
> +		return ret;
> +	HQXContext *ctx = avctx->priv_data;
> +	ret = decode_slice(ctx, data->pic, &data->gb[slice], slice, data->decode_func);
> +	if (ret < 0) {
> +		av_log(avctx, AV_LOG_ERROR, "Error decoding slice %d.\n", slice);
> +	}
> +	return ret;
> +}
> +
>  static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>                              int *got_picture_ptr, AVPacket *avpkt)
>  {
> @@ -492,11 +527,10 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>      uint32_t info_tag, info_offset;
>      int data_start;
>      unsigned data_size;
> -    GetBitContext gb;
>      int i, ret;
>      int slice;
> -    uint32_t slice_off[17];
> -    mb_decode_func decode_func = 0;
> +	data_t arg_data;
> +    arg_data.decode_func = 0;

some of the variables become unused, dont leave unused variables


>  
>      if (avpkt->size < 8)
>          return AVERROR_INVALIDDATA;
> @@ -520,9 +554,11 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>      }
>  
>      data_start = src - avpkt->data;
> -    data_size  = avpkt->size - data_start;
> +	arg_data.src = src;
> +	arg_data.pic = data;
> +	arg_data.data_size = avpkt->size - data_start;
>  
> -    if (data_size < HQX_HEADER_SIZE) {
> +	if (arg_data.data_size < HQX_HEADER_SIZE) {
>          av_log(avctx, AV_LOG_ERROR, "Frame too small.\n");
>          return AVERROR_INVALIDDATA;
>      }
> @@ -537,7 +573,7 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>      ctx->width      = AV_RB16(src + 4);
>      ctx->height     = AV_RB16(src + 6);
>      for (i = 0; i < 17; i++)
> -        slice_off[i] = AV_RB24(src + 8 + i * 3);
> +        arg_data.slice_off[i] = AV_RB24(src + 8 + i * 3);
>  
>      if (ctx->dcb == 8) {
>          av_log(avctx, AV_LOG_ERROR, "Invalid DC precision %d.\n", ctx->dcb);
> @@ -559,22 +595,22 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>      switch (ctx->format) {
>      case HQX_422:
>          avctx->pix_fmt = AV_PIX_FMT_YUV422P16;
> -        decode_func = hqx_decode_422;
> +        arg_data.decode_func = hqx_decode_422;
>          break;
>      case HQX_444:
>          avctx->pix_fmt = AV_PIX_FMT_YUV444P16;
> -        decode_func = hqx_decode_444;
> +		arg_data.decode_func = hqx_decode_444;
>          break;
>      case HQX_422A:
>          avctx->pix_fmt = AV_PIX_FMT_YUVA422P16;
> -        decode_func = hqx_decode_422a;
> +		arg_data.decode_func = hqx_decode_422a;
>          break;
>      case HQX_444A:
>          avctx->pix_fmt = AV_PIX_FMT_YUVA444P16;
> -        decode_func = hqx_decode_444a;
> +		arg_data.decode_func = hqx_decode_444a;
>          break;
>      }
> -    if (!decode_func) {
> +	if (!arg_data.decode_func) {
>          av_log(avctx, AV_LOG_ERROR, "Invalid format: %d.\n", ctx->format);
>          return AVERROR_INVALIDDATA;
>      }
> @@ -583,6 +619,10 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>      if (ret < 0)
>          return ret;
>  
> +
> +	avctx->execute2(avctx, decode_slice_thread, &arg_data, NULL, 16);
> +

> +#if 0
>      for (slice = 0; slice < 16; slice++) {
>          if (slice_off[slice] < HQX_HEADER_SIZE ||
>              slice_off[slice] >= slice_off[slice + 1] ||
> @@ -599,7 +639,7 @@ static int hqx_decode_frame(AVCodecContext *avctx, void *data,
>              av_log(avctx, AV_LOG_ERROR, "Error decoding slice %d.\n", slice);
>          }
>      }
> -
> +#endif

please remove code that is not usefull anymore dont just comment it
out

[...]

--
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

In fact, the RIAA has been known to suggest that students drop out
of college or go to community college in order to be able to afford
settlements. -- The RIAA
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150408/81054903/attachment.asc>