View Issue Details
ID | Project | Category | View Status | Date Submitted | Last Update |
---|---|---|---|---|---|
0009342 | ardour | bugs | public | 2023-05-21 20:58 | 2023-05-22 15:15 |
Reporter | flowz | Assigned To | x42 | ||
Priority | normal | Severity | minor | Reproducibility | always |
Status | resolved | Resolution | fixed | ||
Summary | 0009342: Noise shaping doesn't work | ||||
Description | The current noise shaping algorithm doesn't work as it doesn't use the error value within the error feedback function, because it always gets overridden by noise before any usage. I fixed the implementation by looking at Audacitys dither code which is ironically based on Ardours code. I attached 2 short 8 bit FLAC example exports for illustration, one before and one after the bugfix. Furthermore I attached a spectrogram for each of the files. I uploaded the fixed cpp file which I used to generate the "fixed" FLAC. Furthermore calling gdither_noise() twice gave me the very best results, much better than using the same logic from GDitherTri, but feel free to make you own experiments. | ||||
Steps To Reproduce | Export a file with noise shaping selected. Export 8 bit PCM for the best test results. | ||||
Tags | No tags attached. | ||||
|
gdither.cc (12,572 bytes)
/* * Copyright (C) 2002 Steve Harris <steve@plugin.org.uk> * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation; either version 2 of the License, or * (at your option) any later version. * * This program is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program; if not, write to the Free Software * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. * */ #include "gdither_types_internal.h" #include "gdither.h" /* this monstrosity is necessary to get access to lrintf() and random(). whoever is writing the glibc headers <cmath> and <cstdlib> should be hauled off to a programmer re-education camp. for the rest of their natural lives. or longer. <paul@linuxaudiosystems.com> */ #define _ISOC9X_SOURCE 1 #define _ISOC99_SOURCE 1 #ifdef __cplusplus #include <cmath> #else #include <math.h> #endif #undef __USE_SVID #define __USE_SVID 1 #ifdef __cplusplus #include <cstdlib> #else #include <stdlib.h> #endif #include <assert.h> #include <sys/types.h> /* Lipshitz's minimally audible FIR, only really works for 46kHz-ish signals */ static const float shaped_bs[] = { 2.033f, -2.165f, 1.959f, -1.590f, 0.6149f }; /* Some useful constants */ #define MAX_U8 255 #define MIN_U8 0 #define SCALE_U8 128.0f #define MAX_S16 32767 #define MIN_S16 -32768 #define SCALE_S16 32768.0f #define MAX_S24 8388607 #define MIN_S24 -8388608 #define SCALE_S24 8388608.0f inline static float gdither_noise () { static uint32_t rnd = 23232323; rnd = (rnd * 196314165) + 907633515; return rnd * 2.3283064365387e-10f; } GDither gdither_new(GDitherType type, uint32_t channels, GDitherSize bit_depth, int dither_depth) { GDither s; s = (GDither)calloc(1, sizeof(struct GDither_s)); s->type = type; s->channels = channels; s->bit_depth = (int)bit_depth; if (dither_depth <= 0 || dither_depth > (int)bit_depth) { dither_depth = (int)bit_depth; } s->dither_depth = dither_depth; s->scale = (float)(1LL << (dither_depth - 1)); if (bit_depth == GDitherFloat || bit_depth == GDitherDouble) { s->post_scale_fp = 1.0f / s->scale; s->post_scale = 0; } else { s->post_scale_fp = 0.0f; s->post_scale = 1 << ((int)bit_depth - dither_depth); } switch (bit_depth) { case GDither8bit: /* Unsigned 8 bit */ s->bias = 1.0f; s->clamp_u = 255; s->clamp_l = 0; break; case GDither16bit: /* Signed 16 bit */ s->bias = 0.0f; s->clamp_u = 32767; s->clamp_l = -32768; break; case GDither32bit: /* Signed 24 bit, in upper 24 bits of 32 bit word */ s->bias = 0.0f; s->clamp_u = 8388607; s->clamp_l = -8388608; break; case GDitherFloat: /* normalised float */ s->bias = 0.0f; s->clamp_u = lrintf(s->scale); s->clamp_l = lrintf(-s->scale); break; case GDitherDouble: /* normalised float */ s->bias = 0.0f; s->clamp_u = lrintf(s->scale); s->clamp_l = lrintf(-s->scale); break; case GDitherPerformanceTest: /* special performance test case */ s->scale = SCALE_S24; s->post_scale = 256; s->bias = 0.0f; s->clamp_u = 8388607; s->clamp_l = -8388608; break; default: /* Not a bit depth we can handle */ free(s); return NULL; break; } switch (type) { case GDitherNone: case GDitherRect: /* No state */ break; case GDitherTri: /* The last whitenoise sample */ s->tri_state = (float *) calloc(channels, sizeof(float)); break; case GDitherShaped: /* The error from the last few samples encoded */ s->shaped_state = (GDitherShapedState*) calloc(channels, sizeof(GDitherShapedState)); break; } return s; } void gdither_free(GDither s) { if (s) { free(s->tri_state); free(s->shaped_state); free(s); } } inline static void gdither_innner_loop(const GDitherType dt, const uint32_t stride, const float bias, const float scale, const uint32_t post_scale, const int bit_depth, const uint32_t channel, const uint32_t length, float *ts, GDitherShapedState *ss, float const *x, void *y, const int clamp_u, const int clamp_l) { uint32_t pos, i; uint8_t *o8 = (uint8_t*) y; int16_t *o16 = (int16_t*) y; int32_t *o32 = (int32_t*) y; float tmp, r, ideal; int64_t clamped; i = channel; for (pos = 0; pos < length; pos++, i += stride) { tmp = x[i] * scale + bias; switch (dt) { case GDitherNone: break; case GDitherRect: tmp -= gdither_noise (); break; case GDitherTri: r = gdither_noise () - 0.5f; tmp -= r - ts[channel]; ts[channel] = r; break; case GDitherShaped: assert (ss); /* Run FIR */ tmp += ss->buffer[ss->phase] * shaped_bs[0] + ss->buffer[(ss->phase - 1) & GDITHER_SH_BUF_MASK] * shaped_bs[1] + ss->buffer[(ss->phase - 2) & GDITHER_SH_BUF_MASK] * shaped_bs[2] + ss->buffer[(ss->phase - 3) & GDITHER_SH_BUF_MASK] * shaped_bs[3] + ss->buffer[(ss->phase - 4) & GDITHER_SH_BUF_MASK] * shaped_bs[4]; /* Capture signal for error calculation before adding white noise */ ideal = tmp; /* Add white noise */ tmp += (gdither_noise() + gdither_noise()) * 0.5f; /* Roll buffer and store last error */ ss->phase = (ss->phase + 1) & GDITHER_SH_BUF_MASK; ss->buffer[ss->phase] = ideal - (float)lrintf(tmp); break; } clamped = lrintf(tmp); if (clamped > clamp_u) { clamped = clamp_u; } else if (clamped < clamp_l) { clamped = clamp_l; } switch (bit_depth) { case GDither8bit: o8[i] = (uint8_t) (clamped * post_scale); break; case GDither16bit: o16[i] = (int16_t) (clamped * post_scale); break; case GDither32bit: o32[i] = (int32_t) (clamped * post_scale); break; } } } /* floating pint version of the inner loop function */ inline static void gdither_innner_loop_fp(const GDitherType dt, const uint32_t stride, const float bias, const float scale, const float post_scale, const int bit_depth, const uint32_t channel, const uint32_t length, float *ts, GDitherShapedState *ss, float const *x, void *y, const int clamp_u, const int clamp_l) { uint32_t pos, i; float *oflt = (float*) y; double *odbl = (double*) y; float tmp, r, ideal; double clamped; i = channel; for (pos = 0; pos < length; pos++, i += stride) { tmp = x[i] * scale + bias; switch (dt) { case GDitherNone: break; case GDitherRect: tmp -= gdither_noise (); break; case GDitherTri: r = gdither_noise () - 0.5f; tmp -= r - ts[channel]; ts[channel] = r; break; case GDitherShaped: assert (ss); /* Run FIR */ tmp += ss->buffer[ss->phase] * shaped_bs[0] + ss->buffer[(ss->phase - 1) & GDITHER_SH_BUF_MASK] * shaped_bs[1] + ss->buffer[(ss->phase - 2) & GDITHER_SH_BUF_MASK] * shaped_bs[2] + ss->buffer[(ss->phase - 3) & GDITHER_SH_BUF_MASK] * shaped_bs[3] + ss->buffer[(ss->phase - 4) & GDITHER_SH_BUF_MASK] * shaped_bs[4]; /* Capture signal for error calculation before adding white noise */ ideal = tmp; /* Add white noise */ tmp += (gdither_noise() + gdither_noise()) * 0.5f; /* Roll buffer and store last error */ ss->phase = (ss->phase + 1) & GDITHER_SH_BUF_MASK; ss->buffer[ss->phase] = ideal - (float)lrintf(tmp); break; } clamped = (double)lrintf(tmp); if (clamped > clamp_u) { clamped = clamp_u; } else if (clamped < clamp_l) { clamped = clamp_l; } switch (bit_depth) { case GDitherFloat: oflt[i] = (float) (clamped * post_scale); break; case GDitherDouble: odbl[i] = (double) (clamped * post_scale); break; } } } #define GDITHER_CONV_BLOCK 512 void gdither_run(GDither s, uint32_t channel, uint32_t length, double const *x, void *y) { float conv[GDITHER_CONV_BLOCK]; uint32_t i, pos; char *ycast = (char *)y; int step; switch (s->bit_depth) { case GDither8bit: step = 1; break; case GDither16bit: step = 2; break; case GDither32bit: case GDitherFloat: step = 4; break; case GDitherDouble: step = 8; break; default: step = 0; break; } pos = 0; while (pos < length) { for (i=0; (i + pos) < length && i < GDITHER_CONV_BLOCK; i++) { conv[i] = x[pos + i]; } gdither_runf(s, channel, i, conv, ycast + s->channels * step); pos += i; } } void gdither_runf(GDither s, uint32_t channel, uint32_t length, float const *x, void *y) { uint32_t pos, i; float tmp; int64_t clamped; GDitherShapedState *ss = NULL; if (!s || channel >= s->channels) { return; } if (s->shaped_state) { ss = s->shaped_state + channel; } if (s->type == GDitherNone && s->bit_depth == 23) { int32_t *o32 = (int32_t*) y; for (pos = 0; pos < length; pos++) { i = channel + (pos * s->channels); tmp = x[i] * 8388608.0f; clamped = lrintf(tmp); if (clamped > 8388607) { clamped = 8388607; } else if (clamped < -8388608) { clamped = -8388608; } o32[i] = (int32_t) (clamped * 256); } return; } /* some common case handling code - looks a bit weird, but it allows * the compiler to optimise out the branches in the inner loop */ if (s->bit_depth == 8 && s->dither_depth == 8) { switch (s->type) { case GDitherNone: gdither_innner_loop(GDitherNone, s->channels, 128.0f, SCALE_U8, 1, 8, channel, length, NULL, NULL, x, y, MAX_U8, MIN_U8); break; case GDitherRect: gdither_innner_loop(GDitherRect, s->channels, 128.0f, SCALE_U8, 1, 8, channel, length, NULL, NULL, x, y, MAX_U8, MIN_U8); break; case GDitherTri: gdither_innner_loop(GDitherTri, s->channels, 128.0f, SCALE_U8, 1, 8, channel, length, s->tri_state, NULL, x, y, MAX_U8, MIN_U8); break; case GDitherShaped: gdither_innner_loop(GDitherShaped, s->channels, 128.0f, SCALE_U8, 1, 8, channel, length, NULL, ss, x, y, MAX_U8, MIN_U8); break; } } else if (s->bit_depth == 16 && s->dither_depth == 16) { switch (s->type) { case GDitherNone: gdither_innner_loop(GDitherNone, s->channels, 0.0f, SCALE_S16, 1, 16, channel, length, NULL, NULL, x, y, MAX_S16, MIN_S16); break; case GDitherRect: gdither_innner_loop(GDitherRect, s->channels, 0.0f, SCALE_S16, 1, 16, channel, length, NULL, NULL, x, y, MAX_S16, MIN_S16); break; case GDitherTri: gdither_innner_loop(GDitherTri, s->channels, 0.0f, SCALE_S16, 1, 16, channel, length, s->tri_state, NULL, x, y, MAX_S16, MIN_S16); break; case GDitherShaped: gdither_innner_loop(GDitherShaped, s->channels, 0.0f, SCALE_S16, 1, 16, channel, length, NULL, ss, x, y, MAX_S16, MIN_S16); break; } } else if (s->bit_depth == 32 && s->dither_depth == 24) { switch (s->type) { case GDitherNone: gdither_innner_loop(GDitherNone, s->channels, 0.0f, SCALE_S24, 256, 32, channel, length, NULL, NULL, x, y, MAX_S24, MIN_S24); break; case GDitherRect: gdither_innner_loop(GDitherRect, s->channels, 0.0f, SCALE_S24, 256, 32, channel, length, NULL, NULL, x, y, MAX_S24, MIN_S24); break; case GDitherTri: gdither_innner_loop(GDitherTri, s->channels, 0.0f, SCALE_S24, 256, 32, channel, length, s->tri_state, NULL, x, y, MAX_S24, MIN_S24); break; case GDitherShaped: gdither_innner_loop(GDitherShaped, s->channels, 0.0f, SCALE_S24, 256, 32, channel, length, NULL, ss, x, y, MAX_S24, MIN_S24); break; } } else if (s->bit_depth == GDitherFloat || s->bit_depth == GDitherDouble) { gdither_innner_loop_fp(s->type, s->channels, s->bias, s->scale, s->post_scale_fp, s->bit_depth, channel, length, s->tri_state, ss, x, y, s->clamp_u, s->clamp_l); } else { /* no special case handling, just process it from the struct */ gdither_innner_loop(s->type, s->channels, s->bias, s->scale, s->post_scale, s->bit_depth, channel, length, s->tri_state, ss, x, y, s->clamp_u, s->clamp_l); } } |
|
While this ostensibly fixes the issue, it does not seem to be right. With this change "ideal" is really the error. The key difference in your patch seems to be to use triangular shaped noise (gdither_noise() + gdither_noise()) * 0.5f; It has been a while since I've looked into shaped dithering, I'll have to refresh my memory. |
|
I have applied your patch with minor changes. Thank you for tracking this down! |
Date Modified | Username | Field | Change |
---|---|---|---|
2023-05-21 20:58 | flowz | New Issue | |
2023-05-21 20:58 | flowz | File Added: diffview.png | |
2023-05-21 20:58 | flowz | File Added: noise_shape_bugged.flac | |
2023-05-21 20:58 | flowz | File Added: noise_shape_fixed.flac | |
2023-05-21 20:58 | flowz | File Added: noise_shape_bugged.flac.png | |
2023-05-21 20:58 | flowz | File Added: noise_shape_fixed.flac.png | |
2023-05-21 20:58 | flowz | File Added: gdither.cc | |
2023-05-22 01:23 | x42 | Note Added: 0027663 | |
2023-05-22 01:24 | x42 | Assigned To | => x42 |
2023-05-22 01:24 | x42 | Status | new => confirmed |
2023-05-22 15:15 | x42 | Status | confirmed => resolved |
2023-05-22 15:15 | x42 | Resolution | open => fixed |
2023-05-22 15:15 | x42 | Note Added: 0027668 |