[alsa-devel] [PATCH 1/1] alsabat: truncate sample frames for faster FFT analysis

[Takashi Iwai]

On Mon, 29 Feb 2016 03:33:45 +0100,
han.lu at intel.com wrote:
> 
> From: "Lu, Han" <han.lu at intel.com>
> 
> Truncate the sample frames to powers of 2, since the FFTW algorithm
> runs especially fast in this case, and other sizes may be computed
> by means of a slow, general-purpose algorithm.
> In my test environment applying the patch, a sound clip of 33072
> frames is cut off to 32768 frames before analysis, and the time
> cost is reduced from 6.128s to 0.224s.
> 
> Signed-off-by: Lu, Han <han.lu at intel.com>

Applied, thanks.


Takashi

> 
> diff --git a/bat/analyze.c b/bat/analyze.c
> index 60e2d1c..5cfdac3 100644
> --- a/bat/analyze.c
> +++ b/bat/analyze.c
> @@ -256,6 +256,20 @@ static int reorder_data(struct bat *bat)
>  	return 0;
>  }
>  
> +/* truncate sample frames for faster FFT analysis process */
> +static int truncate_frames(struct bat *bat)
> +{
> +	int shift = SHIFT_MAX;
> +
> +	for (; shift > SHIFT_MIN; shift--)
> +		if (bat->frames & (1 << shift)) {
> +			bat->frames = 1 << shift;
> +			return 0;
> +		}
> +
> +	return -EINVAL;
> +}
> +
>  int analyze_capture(struct bat *bat)
>  {
>  	int err = 0;
> @@ -263,6 +277,13 @@ int analyze_capture(struct bat *bat)
>  	int c;
>  	struct analyze a;
>  
> +	err = truncate_frames(bat);
> +	if (err < 0) {
> +		fprintf(bat->err, _("Invalid frame number for analysis: %d\n"),
> +				bat->frames);
> +		return err;
> +	}
> +
>  	fprintf(bat->log, _("\nBAT analysis: signal has %d frames at %d Hz,"),
>  			bat->frames, bat->rate);
>  	fprintf(bat->log, _(" %d channels, %d bytes per sample.\n"),
> diff --git a/bat/common.h b/bat/common.h
> index b8f6670..ff03fc1 100644
> --- a/bat/common.h
> +++ b/bat/common.h
> @@ -68,6 +68,12 @@
>  #define FOUND_DC			(1<<1)
>  #define FOUND_WRONG_PEAK		(1<<0)
>  
> +/* Truncate sample frames to (1 << N), for faster FFT analysis process. The
> + * valid range of N is (SHIFT_MIN, SHIFT_MAX). When N increases, the analysis
> + * will be more time-consuming, and the result will be more accurate. */
> +#define SHIFT_MAX			(sizeof(int) * 8 - 2)
> +#define SHIFT_MIN			8
> +
>  struct wav_header {
>  	unsigned int magic; /* 'RIFF' */
>  	unsigned int length; /* file len */
> -- 
> 2.5.0
>