On 1/12/2018 8:12 PM, Даниил Чередник wrote:
> Hysterically dcaenc uses own implementation of time->frequency
> transformation used by psychoacoustic. But actually function named fft in
> original dcaenc code is not fft. Power spectrum looks similar to mdct, and
> Alexander E. Patrakov told me it is MDCT. But for me it is still a bit
> strange, because of output size, and absent phase shift sensitivity. I was
> thinking about MCLT. But again, result of transformation original function
> was different. So I decided to use ffmpeg mdct transformation here.
>
>
> Results:
>
> I could not hear the difference between original and modified version.
>
> I got approximately 10% performance boost.
> From 39e7f15886f1c083f3a3d37d52778882c8949a93 Mon Sep 17 00:00:00 2001
> From: Daniil Cherednik <dan.chered...@gmail.com>
> Date: Sun, 7 Jan 2018 22:39:22 +0000
> Subject: [PATCH] avcodec/dcaenc: Use ffmpeg mdct instead of own implementation
>
> Signed-off-by: Daniil Cherednik <dan.chered...@gmail.com>
> ---
> libavcodec/dcaenc.c | 107
> ++++++++++++++------------------------------------
> tests/fate/acodec.mak | 4 +-
> 2 files changed, 32 insertions(+), 79 deletions(-)
>
> diff --git a/libavcodec/dcaenc.c b/libavcodec/dcaenc.c
> index dd601ffae0..b924c58185 100644
> --- a/libavcodec/dcaenc.c
> +++ b/libavcodec/dcaenc.c
> @@ -21,6 +21,9 @@
> * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA
> */
>
> +#define FFT_FLOAT 0
> +#define FFT_FIXED_32 1
> +
> #include "libavutil/avassert.h"
> #include "libavutil/channel_layout.h"
> #include "libavutil/common.h"
> @@ -33,6 +36,7 @@
> #include "dca_core.h"
> #include "dcadata.h"
> #include "dcaenc.h"
> +#include "fft.h"
> #include "internal.h"
> #include "mathops.h"
> #include "put_bits.h"
> @@ -56,6 +60,7 @@ typedef struct DCAEncContext {
> AVClass *class;
> PutBitContext pb;
> DCAADPCMEncContext adpcm_ctx;
> + FFTContext mdct;
> CompressionOptions options;
> int frame_size;
> int frame_bits;
> @@ -154,6 +159,7 @@ static int encode_init(AVCodecContext *avctx)
> DCAEncContext *c = avctx->priv_data;
> uint64_t layout = avctx->channel_layout;
> int i, j, min_frame_bits;
> + int rv;
We normally use ret for variables meant to hold a return value.
>
> if (subband_bufer_alloc(c))
> return AVERROR(ENOMEM);
> @@ -231,6 +237,9 @@ static int encode_init(AVCodecContext *avctx)
>
> avctx->frame_size = 32 * SUBBAND_SAMPLES;
>
> + if ((rv = ff_mdct_init(&c->mdct, 9, 0, 1.0)) < 0)
> + return rv;
> +
> if (!cos_table[0]) {
> int j, k;
>
> @@ -297,6 +306,7 @@ static av_cold int encode_close(AVCodecContext *avctx)
> {
> if (avctx->priv_data) {
> DCAEncContext *c = avctx->priv_data;
> + ff_mdct_end(&c->mdct);
> subband_bufer_free(c);
> ff_dcaadpcm_free(&c->adpcm_ctx);
> }
> @@ -398,78 +408,6 @@ static void lfe_downsample(DCAEncContext *c, const
> int32_t *input)
> }
> }
>
> -typedef struct {
> - int32_t re;
> - int32_t im;
> -} cplx32;
> -
> -static void fft(const int32_t in[2 * 256], cplx32 out[256])
> -{
> - cplx32 buf[256], rin[256], rout[256];
> - int i, j, k, l;
> -
> - /* do two transforms in parallel */
> - for (i = 0; i < 256; i++) {
> - /* Apply the Hann window */
> - rin[i].re = mul32(in[2 * i], 0x3fffffff - (cos_t(8 * i + 2) >> 1));
> - rin[i].im = mul32(in[2 * i + 1], 0x3fffffff - (cos_t(8 * i + 6) >>
> 1));
> - }
> - /* pre-rotation */
> - for (i = 0; i < 256; i++) {
> - buf[i].re = mul32(cos_t(4 * i + 2), rin[i].re)
> - - mul32(sin_t(4 * i + 2), rin[i].im);
> - buf[i].im = mul32(cos_t(4 * i + 2), rin[i].im)
> - + mul32(sin_t(4 * i + 2), rin[i].re);
> - }
> -
> - for (j = 256, l = 1; j != 1; j >>= 1, l <<= 1) {
> - for (k = 0; k < 256; k += j) {
> - for (i = k; i < k + j / 2; i++) {
> - cplx32 sum, diff;
> - int t = 8 * l * i;
> -
> - sum.re = buf[i].re + buf[i + j / 2].re;
> - sum.im = buf[i].im + buf[i + j / 2].im;
> -
> - diff.re = buf[i].re - buf[i + j / 2].re;
> - diff.im = buf[i].im - buf[i + j / 2].im;
> -
> - buf[i].re = half32(sum.re);
> - buf[i].im = half32(sum.im);
> -
> - buf[i + j / 2].re = mul32(diff.re, cos_t(t))
> - - mul32(diff.im, sin_t(t));
> - buf[i + j / 2].im = mul32(diff.im, cos_t(t))
> - + mul32(diff.re, sin_t(t));
> - }
> - }
> - }
> - /* post-rotation */
> - for (i = 0; i < 256; i++) {
> - int b = ff_reverse[i];
> - rout[i].re = mul32(buf[b].re, cos_t(4 * i))
> - - mul32(buf[b].im, sin_t(4 * i));
> - rout[i].im = mul32(buf[b].im, cos_t(4 * i))
> - + mul32(buf[b].re, sin_t(4 * i));
> - }
> - for (i = 0; i < 256; i++) {
> - /* separate the results of the two transforms */
> - cplx32 o1, o2;
> -
> - o1.re = rout[i].re - rout[255 - i].re;
> - o1.im = rout[i].im + rout[255 - i].im;
> -
> - o2.re = rout[i].im - rout[255 - i].im;
> - o2.im = -rout[i].re - rout[255 - i].re;
> -
> - /* combine them into one long transform */
> - out[i].re = mul32( o1.re + o2.re, cos_t(2 * i + 1))
> - + mul32( o1.im - o2.im, sin_t(2 * i + 1));
> - out[i].im = mul32( o1.im + o2.im, cos_t(2 * i + 1))
> - + mul32(-o1.re + o2.re, sin_t(2 * i + 1));
> - }
> -}
> -
> static int32_t get_cb(int32_t in)
> {
> int i, res;
> @@ -494,21 +432,36 @@ static int32_t add_cb(int32_t a, int32_t b)
> return a + cb_to_add[a - b];
> }
>
> -static void adjust_jnd(int samplerate_index,
> +static void calc_power(DCAEncContext *c,
> + const int32_t in[2 * 256], int32_t power[256])
> +{
> + int i;
> + DECLARE_ALIGNED(32, int32_t, data)[512];
> + DECLARE_ALIGNED(32, int32_t, coeff)[256];
LOCAL_ALIGNED_32(int32_t, data, [512]);
LOCAL_ALIGNED_32(int32_t, coeff, [256]);
> + for (i = 0; i < 512; i++) {
> + data[i] = norm__(mul32(in[i], 0x3fffffff - (cos_t(4 * i + 2) >> 1)),
> 4);
> + }
> + c->mdct.mdct_calc(&c->mdct, coeff, data);
> + for (i = 0; i < 256; i++) {
> + const int32_t cb = get_cb(coeff[i]);
> + power[i] = add_cb(cb, cb);
> + }
> +}
> +
> +static void adjust_jnd(DCAEncContext *c,
> const int32_t in[512], int32_t out_cb[256])
> {
> int32_t power[256];
> - cplx32 out[256];
> int32_t out_cb_unnorm[256];
> int32_t denom;
> const int32_t ca_cb = -1114;
> const int32_t cs_cb = 928;
> + const int samplerate_index = c->samplerate_index;
> int i, j;
>
> - fft(in, out);
> + calc_power(c, in, power);
>
> for (j = 0; j < 256; j++) {
> - power[j] = add_cb(get_cb(out[j].re), get_cb(out[j].im));
> out_cb_unnorm[j] = -2047; /* and can only grow */
> }
>
> @@ -586,7 +539,7 @@ static void calc_masking(DCAEncContext *c, const int32_t
> *input)
> data[i] = c->history[ch][k];
> for (k -= 512; i < 512; i++, k++)
> data[i] = input[k * c->channels + chi];
> - adjust_jnd(c->samplerate_index, data, c->masking_curve_cb[ssf]);
> + adjust_jnd(c, data, c->masking_curve_cb[ssf]);
> }
> for (i = 0; i < 256; i++) {
> int32_t m = 2048;
> diff --git a/tests/fate/acodec.mak b/tests/fate/acodec.mak
> index 5c3fea90c5..80d26de0f9 100644
> --- a/tests/fate/acodec.mak
> +++ b/tests/fate/acodec.mak
> @@ -104,14 +104,14 @@ fate-acodec-dca: tests/data/asynth-44100-2.wav
> fate-acodec-dca: SRC = tests/data/asynth-44100-2.wav
> fate-acodec-dca: CMD = md5 -i $(TARGET_PATH)/$(SRC) -c:a dca -strict -2 -f
> dts -flags +bitexact
> fate-acodec-dca: CMP = oneline
> -fate-acodec-dca: REF = 7cd79a3717943a06b217f1130223a86f
> +fate-acodec-dca: REF = 2aa580ac67820fce4f581b96ebb34acc
>
> FATE_ACODEC-$(call ENCDEC, DCA, WAV) += fate-acodec-dca2
> fate-acodec-dca2: CMD = enc_dec_pcm dts wav s16le $(SRC) -c:a dca -strict -2
> -flags +bitexact
> fate-acodec-dca2: REF = $(SRC)
> fate-acodec-dca2: CMP = stddev
> fate-acodec-dca2: CMP_SHIFT = -2048
> -fate-acodec-dca2: CMP_TARGET = 527
> +fate-acodec-dca2: CMP_TARGET = 535
> fate-acodec-dca2: SIZE_TOLERANCE = 1632
>
> FATE_ACODEC-$(call ENCDEC, FLAC, FLAC) += fate-acodec-flac
> fate-acodec-flac-exact-rice
> --
> 2.13.5
>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
