Laufzeit-Code vom Dev-Werkzeug getrennt: - game/: das eigentliche Spiel, startbar per `python -m game [--once]`. - formats.py (war tools/kellogg_formats.py) -- Asset-Parser, Kern-Bibliothek - audio.py (war tools/tfmx_audio.py) -- TFMX-Musik-Wrapper - intro.py (war tools/intro_sequence.py) -- Boot-/Intro-Sequenz - tfmx_player/(war tools/tfmx_player/) -- MIT-C-Renderer - __init__.py / __main__.py fuer `python -m game` - tools/: nur noch Dev-Werkzeuge (dat_extract, pcc_to_png, render_assets), importieren jetzt aus game.formats. - README.md: Ueberblick, Start, Struktur, Voraussetzungen, Rechtliches. - .gitignore: tfmx_player-Binary-Pfad auf game/ nachgezogen. Intro laeuft unveraendert (headless-Smoke-Test `python -m game --once` exit 0). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
633 lines
26 KiB
C
633 lines
26 KiB
C
// Copyright (c) 2026 Peter Fors
|
|
// SPDX-License-Identifier: MIT
|
|
//
|
|
// Amiga 500 Paula emulator for custom replayers.
|
|
//
|
|
// This is a hardware model, not a resampler. The channel mixer runs in the
|
|
// Paula clock domain (3546895 Hz PAL / 3579545 Hz NTSC). Each hardware
|
|
// channel has a period counter that, when it expires, latches the next 8-bit
|
|
// sample byte; between latches the channel holds that byte (the zero-order-
|
|
// hold staircase a real Paula produces). Volume is the real 6-bit PWM over a
|
|
// 64-clock window, not a multiply, so its quantization noise is reproduced.
|
|
// Channels 0+3 are summed to the left output, 1+2 to the right, hard-panned,
|
|
// in the Paula clock domain. The analog filter chain (always-on RC low-pass --
|
|
// ~4.4 kHz on A500, ~34 kHz on A1200 -- plus the switchable ~3.3 kHz LED
|
|
// Butterworth) runs at the Paula clock rate. Only the final stage decimates
|
|
// to the host rate, by box-filter integration of the Paula-clock samples that
|
|
// fall in each output window. Aliasing and quantization noise that a real
|
|
// Amiga produces are preserved.
|
|
//
|
|
// Paula has exactly four hardware channels (0..3). There is no software-
|
|
// mixer extension and no side bus: a real Amiga has no extra channel in its
|
|
// signal path. Every format with more than four voices built a mixed buffer
|
|
// on the CPU and DMA'd THAT through these four channels, so any such mixdown
|
|
// is the replayer's job and its output IS Paula channel sample data, <= 4
|
|
// channels, passing through this same hardware path.
|
|
//
|
|
// Output is ACCUMULATED into the caller's float buffer. The hardware output
|
|
// chain is modelled end to end, to the RCA jack, not just to the summer node:
|
|
//
|
|
// 1. Resistive averaging summer: the two channels on each side (0+3 left,
|
|
// 1+2 right) join through equal board resistors, so the per-side node
|
|
// is (ch_a + ch_b) / 2 (a ~6 dB attenuation).
|
|
// 2. The analog filter chain acts on that node: an always-on RC low-pass
|
|
// (~4.4 kHz on A500, ~34 kHz on A1200) plus the switchable ~3.3 kHz
|
|
// LED Butterworth.
|
|
// 3. Output buffer/amp: normalises int8 full scale to unity. The
|
|
// resistive divider's ~6 dB attenuation is preserved (not compensated)
|
|
// so the per-side level matches real hardware: a single full-scale
|
|
// channel lands at ~0.5, two correlated full-scale channels on the
|
|
// same side at ~1.0. No analog rail saturation is modelled -- at line
|
|
// out on a stock A500 the output op-amp runs with ~10 V of usable
|
|
// rail headroom against a ~1 V peak signal and never reaches its
|
|
// rails in practice. The output is then clamped to [-1, +1] purely
|
|
// as a digital safety guard for callers converting to fixed-point:
|
|
// small Butterworth step overshoot on transients (a few percent)
|
|
// cannot leak out as wrap/click noise after a (int16_t)(x * 32768)
|
|
// style cast. Absolute level is the host's concern.
|
|
//
|
|
// Host integration: the filter chain (always-on RC LP + switchable LED
|
|
// Butterworth + decimation anti-alias) runs IIR state at the Paula clock,
|
|
// and that state decays exponentially toward zero when channels go silent.
|
|
// Once any state slot crosses the float denormal threshold (~1.18e-38),
|
|
// every subsequent multiply touching it is denormal-slow on x86 (roughly
|
|
// two orders of magnitude); a single mix can blow past the host's audio
|
|
// buffer duration -- audible as underrun. Adding per-sample denormal-
|
|
// prevention bias inside the filter inner loops would cost a fadd per
|
|
// stage per Paula clock (millions/sec), so the agreed convention is:
|
|
// THE HOST AUDIO THREAD MUST RUN WITH MXCSR FTZ+DAZ ENABLED. Any thread
|
|
// that calls paula_mix_frames is in scope. The standard recipe is
|
|
// #include <pmmintrin.h>
|
|
// _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
|
|
// _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
|
|
// at the top of the audio thread proc. MXCSR is per-thread on x86 so this
|
|
// must be set inside the thread, not once at program start.
|
|
|
|
#pragma once
|
|
|
|
#include <stdint.h>
|
|
#include <string.h>
|
|
#include <math.h>
|
|
|
|
// L+R packed double, used through the filter chain to halve biquad cost: all
|
|
// three filter stages use identical coefficients per side, only state differs,
|
|
// so each biquad line becomes one packed instruction (one packed FMA on
|
|
// x86-64-v3). GCC/Clang vector extension; arithmetic operators are overloaded
|
|
// to the right SIMD ops per -march.
|
|
typedef double paula_v2df __attribute__((vector_size(16)));
|
|
|
|
// Opt-in mixer profiler. Compiled in only when PAULA_PROFILE is defined, so
|
|
// normal builds carry zero footprint. Accumulates process CPU time spent
|
|
// strictly inside paula_mix_frames (not replayer tick work) and the number of
|
|
// frames produced; paula_profile_report() turns that into a realtime factor.
|
|
#ifdef PAULA_PROFILE
|
|
#include <stdio.h>
|
|
#include <time.h>
|
|
static double paula_profile_cpu_ns = 0.0;
|
|
static uint64_t paula_profile_frames = 0;
|
|
#endif
|
|
|
|
// Paula has exactly four hardware channels. Formats with more voices must
|
|
// CPU-mix down to <= 4 themselves; there is no extra channel here.
|
|
#define PAULA_NUM_CHANNELS 4
|
|
#define PAULA_PAL_CLOCK 3546895
|
|
#define PAULA_NTSC_CLOCK 3579545
|
|
// Real Paula audio DMA floor. The Hardware Reference Manual's period-124
|
|
// figure is the rate at which all four channels can DMA without the bus
|
|
// falling behind during display fetch; a single channel goes lower. The true
|
|
// hardware floor is period 113 -- the ProTracker/Soundtracker note table
|
|
// bottoms at exactly 113 (B-3) because that is where Paula stops. Clamping to
|
|
// 124 detunes the whole top octave flat (period 113 -> ~160 cents). All four
|
|
// hardware channels clamp to this.
|
|
#define PAULA_DMA_MIN_PERIOD 113
|
|
|
|
// Period accumulator fixed-point: one Paula clock advances the accumulator by
|
|
// PAULA_PERIOD_ONE; a channel consumes one sample byte every period_q of
|
|
// these. period_q is integer-exact for the Paula register path and fractional
|
|
// for the Hz path, so both keep exact pitch.
|
|
#define PAULA_PERIOD_SHIFT 16
|
|
#define PAULA_PERIOD_ONE (1ull << PAULA_PERIOD_SHIFT)
|
|
|
|
struct paula_channel {
|
|
int8_t *sample;
|
|
uint32_t length; // bytes (becomes loop_start+loop_length after first wrap)
|
|
uint32_t loop_start; // bytes
|
|
uint32_t loop_length; // bytes, 0 => one-shot
|
|
uint32_t pos; // current byte index into sample
|
|
uint64_t period_q; // Paula clocks per sample byte, Q16
|
|
uint64_t period_acc; // period accumulator, Q16
|
|
int8_t *pending_sample; // deferred switch on next wrap (Paula AUDxLC trick)
|
|
uint32_t pending_pos;
|
|
uint32_t pending_length;
|
|
int8_t cur; // latched sample byte (zero-order-hold output)
|
|
uint16_t volume; // 0..64 Amiga scale
|
|
uint8_t pwm_cnt; // 0..63 volume-PWM phase
|
|
uint8_t active;
|
|
uint8_t muted;
|
|
uint8_t has_pending;
|
|
uint8_t backwards; // 1 -> step DOWN through sample (DBP E3, etc.)
|
|
};
|
|
|
|
// Amiga model. Selects the always-on post-DAC RC low-pass corner: ~4.4 kHz on
|
|
// A500 (the classic muffled top end), ~34 kHz on A1200 (bright but not brick-
|
|
// walled -- the slight roll into the top octave that real hardware has,
|
|
// neither aliasing brightness nor A500 muffling). The LED filter exists on
|
|
// both. Default is the A500.
|
|
#define PAULA_MODEL_A500 0
|
|
#define PAULA_MODEL_A1200 1
|
|
|
|
struct paula {
|
|
struct paula_channel ch[PAULA_NUM_CHANNELS];
|
|
int32_t sample_rate; // host output rate
|
|
int32_t clock; // Paula clock (PAL/NTSC), internal mix rate
|
|
int32_t samples_per_tick;
|
|
int32_t tick_offset;
|
|
int32_t model;
|
|
|
|
// Box-filter decimation from the Paula clock domain to the host rate.
|
|
// Each output sample averages the decim_step (Q16) Paula clocks that
|
|
// fall in its window; decim_phase carries the fraction across calls so
|
|
// the clock count alternates with no pitch drift.
|
|
uint64_t decim_step;
|
|
uint64_t decim_phase;
|
|
|
|
// Always-on 1-pole RC low-pass, at the Paula clock rate. Corner depends
|
|
// on model: ~4.4 kHz for A500, ~34 kHz for A1200. State is L+R packed.
|
|
double fixed_lp_a;
|
|
paula_v2df fixed_lp;
|
|
|
|
// Switchable LED filter: 2-pole Butterworth low-pass (~3.3 kHz,
|
|
// Q=1/sqrt(2)), at the Paula clock rate, RBJ bilinear coefficients.
|
|
// Driven by the replayer via paula_set_lp_filter; biquad state (TDF-II,
|
|
// L+R packed) persists across toggles so flips don't click.
|
|
int32_t lp_filter_on;
|
|
double led_b0;
|
|
double led_b1;
|
|
double led_b2;
|
|
double led_a1;
|
|
double led_a2;
|
|
paula_v2df led_z1;
|
|
paula_v2df led_z2;
|
|
|
|
// Decimation anti-alias low-pass: 8th-order Butterworth (4 cascaded RBJ
|
|
// biquads) at 0.45*host_rate, run in the Paula clock domain just before the
|
|
// rate drop. This is a resampler reconstruction filter, NOT modelled
|
|
// hardware: it bandlimits to below the host Nyquist so the box-average
|
|
// decimation cannot fold ultrasonic ZOH images down into the audible band.
|
|
// Keyed to host_rate, so it runs for both models; on the A500 the analog
|
|
// chain has already removed everything near Nyquist, making it a no-op.
|
|
// State is L+R packed per stage.
|
|
double aa_b0[4];
|
|
double aa_a1[4];
|
|
double aa_a2[4];
|
|
paula_v2df aa_z1[4];
|
|
paula_v2df aa_z2[4];
|
|
};
|
|
|
|
// [=]===^=[ paula_recalc ]=======================================================================[=]
|
|
// Recompute every rate-dependent coefficient from p->clock and
|
|
// p->sample_rate. The analog filters run at the Paula clock, so their
|
|
// coefficients are bilinear-transformed for that rate, not the host rate.
|
|
static void paula_recalc(struct paula *p) {
|
|
double fs = (double)p->clock;
|
|
double dt = 1.0 / fs;
|
|
|
|
// Always-on RC low-pass. Corner is model-dependent: A500 ~4.4 kHz,
|
|
// A1200 ~34 kHz. a = dt / (RC + dt).
|
|
double lp_fc = (p->model == PAULA_MODEL_A1200) ? 34000.0 : 4400.0;
|
|
double lp_rc = 1.0 / (2.0 * 3.14159265358979323846 * lp_fc);
|
|
p->fixed_lp_a = dt / (lp_rc + dt);
|
|
|
|
// LED filter: 2-pole Butterworth low-pass, ~3.3 kHz, Q = 1/sqrt(2),
|
|
// RBJ cookbook low-pass mapped via the bilinear transform at fs.
|
|
double fc = 3300.0;
|
|
double q = 0.70710678118654752440;
|
|
double w0 = 2.0 * 3.14159265358979323846 * fc / fs;
|
|
double cw = cos(w0);
|
|
double sw = sin(w0);
|
|
double alpha = sw / (2.0 * q);
|
|
double a0 = 1.0 + alpha;
|
|
p->led_b0 = ((1.0 - cw) * 0.5) / a0;
|
|
p->led_b1 = (1.0 - cw) / a0;
|
|
p->led_b2 = ((1.0 - cw) * 0.5) / a0;
|
|
p->led_a1 = (-2.0 * cw) / a0;
|
|
p->led_a2 = (1.0 - alpha) / a0;
|
|
|
|
// Decimation anti-alias: 8th-order Butterworth low-pass at 0.45*host_rate,
|
|
// mapped via RBJ bilinear at the Paula clock. The four sections carry the
|
|
// standard 8th-order Butterworth section Q's; cascaded DC gain is unity.
|
|
double aa_q[4] = {0.50979558, 0.60134489, 0.89997622, 2.56291545};
|
|
double aa_w0 = 2.0 * 3.14159265358979323846 * (0.45 * (double)p->sample_rate) / fs;
|
|
double aa_cw = cos(aa_w0);
|
|
double aa_sw = sin(aa_w0);
|
|
for(uint32_t st = 0; st < 4; ++st) {
|
|
double al = aa_sw / (2.0 * aa_q[st]);
|
|
double a0 = 1.0 + al;
|
|
p->aa_b0[st] = ((1.0 - aa_cw) * 0.5) / a0;
|
|
p->aa_a1[st] = (-2.0 * aa_cw) / a0;
|
|
p->aa_a2[st] = (1.0 - al) / a0;
|
|
}
|
|
|
|
// Box-filter decimation step: Paula clocks per host output sample, Q16.
|
|
p->decim_step = ((uint64_t)p->clock << PAULA_PERIOD_SHIFT) / (uint64_t)p->sample_rate;
|
|
}
|
|
|
|
// [=]===^=[ paula_init ]=========================================================================[=]
|
|
static void paula_init(struct paula *p, int32_t sample_rate, int32_t tick_rate_hz) {
|
|
memset(p, 0, sizeof(*p));
|
|
p->sample_rate = sample_rate;
|
|
p->clock = PAULA_PAL_CLOCK;
|
|
p->samples_per_tick = sample_rate / tick_rate_hz;
|
|
p->model = PAULA_MODEL_A500;
|
|
// Hard-panned: channels 0+3 -> left, 1+2 -> right (fixed Paula wiring).
|
|
paula_recalc(p);
|
|
}
|
|
|
|
// [=]===^=[ paula_set_clock ]====================================================================[=]
|
|
// Select the Paula clock (PAULA_PAL_CLOCK / PAULA_NTSC_CLOCK). Recomputes the
|
|
// rate-dependent coefficients. Default after paula_init is PAL.
|
|
static void paula_set_clock(struct paula *p, int32_t clock_hz) {
|
|
p->clock = clock_hz > 0 ? clock_hz : PAULA_PAL_CLOCK;
|
|
paula_recalc(p);
|
|
}
|
|
|
|
// [=]===^=[ paula_set_model ]====================================================================[=]
|
|
// Select the emulated machine. The always-on post-DAC RC low-pass corner
|
|
// changes with model (A500 ~4.4 kHz, A1200 ~34 kHz); the LED filter exists on
|
|
// both. Default is the A500.
|
|
static void paula_set_model(struct paula *p, int32_t model) {
|
|
p->model = (model == PAULA_MODEL_A1200) ? PAULA_MODEL_A1200 : PAULA_MODEL_A500;
|
|
paula_recalc(p);
|
|
}
|
|
|
|
// [=]===^=[ paula_set_lp_filter ]================================================================[=]
|
|
// Enable or disable the switchable Amiga LED filter (the power-LED-gated
|
|
// 2-pole low-pass). Replayers call this to mirror the module's own filter
|
|
// state. The always-on RC low-pass is not affected (A500 ~4.4 kHz, A1200
|
|
// ~34 kHz).
|
|
static void paula_set_lp_filter(struct paula *p, int32_t on) {
|
|
p->lp_filter_on = on ? 1 : 0;
|
|
}
|
|
|
|
// [=]===^=[ paula_set_period ]===================================================================[=]
|
|
// Amiga AUDxPER (DMA) period. All four hardware channels clamp to the real
|
|
// Paula DMA minimum period.
|
|
static void paula_set_period(struct paula *p, int32_t idx, uint16_t period) {
|
|
if(period != 0 && period < PAULA_DMA_MIN_PERIOD) {
|
|
period = PAULA_DMA_MIN_PERIOD;
|
|
}
|
|
if(period == 0) {
|
|
p->ch[idx].period_q = 0;
|
|
return;
|
|
}
|
|
p->ch[idx].period_q = (uint64_t)period << PAULA_PERIOD_SHIFT;
|
|
}
|
|
|
|
// [=]===^=[ paula_set_freq_hz ]==================================================================[=]
|
|
// Set a channel's DMA rate directly in Hz, for replayers that DMA a
|
|
// CPU-built mixdown buffer through a Paula channel (DigiBoosterPro, FaceThe-
|
|
// Music). The period is fractional in the Paula clock domain so pitch stays
|
|
// exact. No DMA period floor: mixdown rates are well above it anyway.
|
|
static void paula_set_freq_hz(struct paula *p, int32_t idx, uint32_t freq_hz) {
|
|
if(freq_hz == 0) {
|
|
p->ch[idx].period_q = 0;
|
|
return;
|
|
}
|
|
p->ch[idx].period_q = ((uint64_t)p->clock << PAULA_PERIOD_SHIFT) / (uint64_t)freq_hz;
|
|
}
|
|
|
|
// [=]===^=[ paula_set_volume ]===================================================================[=]
|
|
static void paula_set_volume(struct paula *p, int32_t idx, uint16_t volume) {
|
|
if(volume > 64) {
|
|
volume = 64;
|
|
}
|
|
p->ch[idx].volume = volume;
|
|
}
|
|
|
|
// Volume is passed in 0..256 range in NostalgicPlayer convention; divide to 0..64.
|
|
// [=]===^=[ paula_set_volume_256 ]===============================================================[=]
|
|
static void paula_set_volume_256(struct paula *p, int32_t idx, uint16_t volume) {
|
|
if(volume > 256) {
|
|
volume = 256;
|
|
}
|
|
p->ch[idx].volume = volume >> 2;
|
|
}
|
|
|
|
// [=]===^=[ paula_play_sample ]==================================================================[=]
|
|
static void paula_play_sample(struct paula *p, int32_t idx, int8_t *sample, uint32_t length) {
|
|
struct paula_channel *c = &p->ch[idx];
|
|
c->sample = sample;
|
|
c->length = length;
|
|
c->pos = (c->backwards && length > 0) ? (length - 1) : 0;
|
|
c->loop_start = 0;
|
|
c->loop_length = 0;
|
|
c->has_pending = 0;
|
|
c->pending_sample = 0;
|
|
c->period_acc = 0;
|
|
c->active = (sample != 0) && (length > 0);
|
|
c->cur = c->active ? sample[c->pos] : 0;
|
|
}
|
|
|
|
// [=]===^=[ paula_set_backwards ]================================================================[=]
|
|
// Set or clear the backwards-playback flag for a channel. Takes effect on the
|
|
// next paula_play_sample (which seeds pos at the high end) and reverses the
|
|
// per-byte advance direction.
|
|
static void paula_set_backwards(struct paula *p, int32_t idx, int32_t on) {
|
|
p->ch[idx].backwards = on ? 1 : 0;
|
|
}
|
|
|
|
// [=]===^=[ paula_set_pos ]======================================================================[=]
|
|
// Move the channel's read position to `byte_offset` within the current sample
|
|
// and re-latch the held byte. Used by effects like ProTracker 9xx (sample
|
|
// offset). Clamps to [0, length-1].
|
|
static void paula_set_pos(struct paula *p, int32_t idx, uint32_t byte_offset) {
|
|
struct paula_channel *c = &p->ch[idx];
|
|
if(c->sample == 0 || c->length == 0) {
|
|
c->pos = 0;
|
|
c->cur = 0;
|
|
return;
|
|
}
|
|
if(byte_offset >= c->length) {
|
|
byte_offset = c->length - 1;
|
|
}
|
|
c->pos = byte_offset;
|
|
c->cur = c->sample[byte_offset];
|
|
}
|
|
|
|
// [=]===^=[ paula_queue_sample ]=================================================================[=]
|
|
// If the channel is active, the new sample takes effect when the current one
|
|
// reaches length (Amiga "write AUDxLC/AUDxLEN mid-DMA"). If inactive, it
|
|
// starts immediately. Plays from sample[start_offset] for `length` bytes,
|
|
// then wraps using the channel's current loop_start / loop_length.
|
|
static void paula_queue_sample(struct paula *p, int32_t idx, int8_t *sample, uint32_t start_offset, uint32_t length) {
|
|
struct paula_channel *c = &p->ch[idx];
|
|
if(!c->active && sample != 0 && length > 0) {
|
|
c->sample = sample;
|
|
c->pos = start_offset;
|
|
c->length = start_offset + length;
|
|
c->has_pending = 0;
|
|
c->pending_sample = 0;
|
|
c->period_acc = 0;
|
|
c->active = 1;
|
|
c->cur = sample[start_offset];
|
|
return;
|
|
}
|
|
c->pending_sample = sample;
|
|
c->pending_pos = start_offset;
|
|
c->pending_length = start_offset + length;
|
|
c->has_pending = (sample != 0) && (length > 0);
|
|
}
|
|
|
|
// [=]===^=[ paula_set_loop ]=====================================================================[=]
|
|
static void paula_set_loop(struct paula *p, int32_t idx, uint32_t start, uint32_t length) {
|
|
struct paula_channel *c = &p->ch[idx];
|
|
c->loop_start = start;
|
|
c->loop_length = length;
|
|
}
|
|
|
|
// [=]===^=[ paula_mute ]=========================================================================[=]
|
|
static void paula_mute(struct paula *p, int32_t idx) {
|
|
p->ch[idx].active = 0;
|
|
}
|
|
|
|
// [=]===^=[ paula_ch_advance ]===================================================================[=]
|
|
// Consume one sample byte for a channel: step the read position one byte
|
|
// (forward or backward), apply the pending-sample swap / loop wrap / one-shot
|
|
// stop exactly as Paula DMA does, and re-latch the held byte.
|
|
__attribute__((always_inline))
|
|
static inline void paula_ch_advance(struct paula_channel *c) {
|
|
if(!c->backwards) {
|
|
uint32_t np = c->pos + 1;
|
|
if(np >= c->length) {
|
|
if(c->has_pending) {
|
|
c->sample = c->pending_sample;
|
|
np = c->pending_pos;
|
|
c->length = c->pending_length;
|
|
c->has_pending = 0;
|
|
c->pending_sample = 0;
|
|
} else if(c->loop_length > 0) {
|
|
uint32_t over = np - c->length;
|
|
np = c->loop_start + (over % c->loop_length);
|
|
c->length = c->loop_start + c->loop_length;
|
|
} else {
|
|
c->active = 0;
|
|
return;
|
|
}
|
|
}
|
|
c->pos = np;
|
|
} else {
|
|
if(c->pos == 0 || (c->loop_length > 0 && c->pos <= c->loop_start)) {
|
|
if(c->has_pending) {
|
|
c->sample = c->pending_sample;
|
|
c->length = c->pending_length;
|
|
c->pos = c->pending_length - 1;
|
|
c->has_pending = 0;
|
|
c->pending_sample = 0;
|
|
} else if(c->loop_length > 0) {
|
|
c->pos = c->loop_start + c->loop_length - 1;
|
|
} else {
|
|
c->active = 0;
|
|
return;
|
|
}
|
|
} else {
|
|
c->pos = c->pos - 1;
|
|
}
|
|
}
|
|
c->cur = c->sample[c->pos];
|
|
}
|
|
|
|
// [=]===^=[ paula_ch_sample ]====================================================================[=]
|
|
// Consume one Paula clock for a single channel: bump the period accumulator,
|
|
// advance the read position by as many bytes as the accumulator demands (may
|
|
// deactivate a one-shot channel), then return the PWM-gated sample value the
|
|
// channel contributes this clock. Returns 0.0 for a channel that is or just
|
|
// went inactive, so the caller's accumulator can stay branch-free.
|
|
// always_inline: called per Paula clock per active channel (~14M/s of audio),
|
|
// and the inliner's -O2 size budget refuses on its own.
|
|
__attribute__((always_inline))
|
|
static inline double paula_ch_sample(struct paula_channel *c) {
|
|
if(!c->active) {
|
|
return 0.0;
|
|
}
|
|
c->period_acc += PAULA_PERIOD_ONE;
|
|
while(c->period_acc >= c->period_q) {
|
|
c->period_acc -= c->period_q;
|
|
paula_ch_advance(c);
|
|
if(!c->active) {
|
|
return 0.0;
|
|
}
|
|
}
|
|
c->pwm_cnt = (uint8_t)((c->pwm_cnt + 1) & 63);
|
|
int32_t v = (c->pwm_cnt < c->volume) ? (int32_t)c->cur : 0;
|
|
return (double)v;
|
|
}
|
|
|
|
// [=]===^=[ paula_mix_frames ]===================================================================[=]
|
|
// Accumulates `frames` float stereo frames into `output`. Caller must
|
|
// pre-clear. The inner loop runs at the Paula clock; each output frame is the
|
|
// box-filter average of the Paula-clock samples in its window.
|
|
//
|
|
// L+R run packed as paula_v2df through the analog/AA chain: all three filter
|
|
// stages share coefficients across sides, only state differs, so each biquad
|
|
// line is one packed instruction (one packed FMA on x86-64-v3). The output
|
|
// safety clamp at the box-average store is also packed (one minpd, one
|
|
// maxpd).
|
|
static void paula_mix_frames(struct paula *p, float *output, int32_t frames) {
|
|
#ifdef PAULA_PROFILE
|
|
struct timespec prof_t0;
|
|
clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &prof_t0);
|
|
#endif
|
|
int32_t led = p->lp_filter_on;
|
|
paula_v2df fa = {p->fixed_lp_a, p->fixed_lp_a};
|
|
paula_v2df fl = p->fixed_lp;
|
|
paula_v2df lb0 = {p->led_b0, p->led_b0};
|
|
paula_v2df lb1 = {p->led_b1, p->led_b1};
|
|
paula_v2df lb2 = {p->led_b2, p->led_b2};
|
|
paula_v2df la1 = {p->led_a1, p->led_a1};
|
|
paula_v2df la2 = {p->led_a2, p->led_a2};
|
|
paula_v2df lz1 = p->led_z1;
|
|
paula_v2df lz2 = p->led_z2;
|
|
paula_v2df ab0[4];
|
|
paula_v2df aa1[4];
|
|
paula_v2df aa2[4];
|
|
paula_v2df az1[4];
|
|
paula_v2df az2[4];
|
|
for(uint32_t st = 0; st < 4; ++st) {
|
|
ab0[st] = (paula_v2df){p->aa_b0[st], p->aa_b0[st]};
|
|
aa1[st] = (paula_v2df){p->aa_a1[st], p->aa_a1[st]};
|
|
aa2[st] = (paula_v2df){p->aa_a2[st], p->aa_a2[st]};
|
|
az1[st] = p->aa_z1[st];
|
|
az2[st] = p->aa_z2[st];
|
|
}
|
|
paula_v2df two = {2.0, 2.0};
|
|
paula_v2df half = {0.5, 0.5};
|
|
// amp_gain normalises int8 full scale (128) to 1.0 and deliberately does
|
|
// NOT make up the resistive divider's 6 dB attenuation -- the per-side
|
|
// level then matches real hardware (single full-scale channel at ~0.5,
|
|
// two correlated full-scale channels on a side at ~1.0). Output is the
|
|
// box-filter average over the window (/n), clamped to [-1, +1] at the
|
|
// float store as a digital safety guard for fixed-point conversion.
|
|
paula_v2df amp = {1.0 / 128.0, 1.0 / 128.0};
|
|
uint64_t phase = p->decim_phase;
|
|
uint64_t dstep = p->decim_step;
|
|
|
|
// Active-channel working set, split by side. The selection predicate
|
|
// (active / unmuted / has sample / nonzero period) is stable within a
|
|
// mix call: only `active` can drop when a one-shot sample ends mid-call,
|
|
// which paula_ch_sample handles per channel. Splitting by side kills the
|
|
// per-Paula-clock "ci == 0 || ci == 3" branch -- each per-side scalar
|
|
// accumulator now stays in a register through its sweep. Output is bit-
|
|
// identical because pl and pr are separate accumulators with fixed
|
|
// channel assignments (0+3 -> pl, 1+2 -> pr): the per-side sum only
|
|
// depends on which channels are active, not on iteration order.
|
|
struct paula_channel *hw_l[PAULA_NUM_CHANNELS];
|
|
struct paula_channel *hw_r[PAULA_NUM_CHANNELS];
|
|
uint32_t nl = 0;
|
|
uint32_t nr = 0;
|
|
for(int32_t ci = 0; ci < PAULA_NUM_CHANNELS; ++ci) {
|
|
struct paula_channel *c = &p->ch[ci];
|
|
if(!c->active || c->muted || c->sample == 0 || c->period_q == 0) {
|
|
continue;
|
|
}
|
|
if(ci == 0 || ci == 3) {
|
|
hw_l[nl++] = c;
|
|
} else {
|
|
hw_r[nr++] = c;
|
|
}
|
|
}
|
|
|
|
for(int32_t i = 0; i < frames; ++i) {
|
|
phase += dstep;
|
|
uint32_t n = (uint32_t)(phase >> PAULA_PERIOD_SHIFT);
|
|
phase &= (PAULA_PERIOD_ONE - 1);
|
|
if(n == 0) {
|
|
n = 1;
|
|
}
|
|
paula_v2df s = {0.0, 0.0};
|
|
for(uint32_t k = 0; k < n; ++k) {
|
|
double pl = 0.0;
|
|
double pr = 0.0;
|
|
for(uint32_t j = 0; j < nl; ++j) {
|
|
pl += paula_ch_sample(hw_l[j]);
|
|
}
|
|
for(uint32_t j = 0; j < nr; ++j) {
|
|
pr += paula_ch_sample(hw_r[j]);
|
|
}
|
|
// Passive resistive averaging summer: the per-side filter node
|
|
// is (ch_a + ch_b) / 2, so it cannot exceed a single channel's
|
|
// full scale and the hardware path never clips.
|
|
paula_v2df x = (paula_v2df){pl, pr} * half;
|
|
// Always-on RC pole (model-dependent corner baked into fa).
|
|
fl = fl + (x - fl) * fa;
|
|
x = fl;
|
|
if(led) {
|
|
paula_v2df y = lb0 * x + lz1;
|
|
lz1 = lb1 * x - la1 * y + lz2;
|
|
lz2 = lb2 * x - la2 * y;
|
|
x = y;
|
|
}
|
|
// Downstream output buffer/amp: int8 -> unity normalisation.
|
|
// No rail saturation is modelled (see header for rationale).
|
|
x = x * amp;
|
|
// Anti-alias before the rate drop: 4 cascaded Butterworth biquads
|
|
// (TDF-II), L+R packed. Bandlimits below host Nyquist so the box-
|
|
// average decimation below cannot fold ultrasonic images down.
|
|
// RBJ low-pass identities baked in here: b1 = 2*b0 and b2 = b0,
|
|
// so only ab0[] is stored. Do not reuse this loop for a non-LP
|
|
// section -- it will silently produce wrong output.
|
|
for(uint32_t st = 0; st < 4; ++st) {
|
|
paula_v2df y = ab0[st] * x + az1[st];
|
|
az1[st] = two * ab0[st] * x - aa1[st] * y + az2[st];
|
|
az2[st] = ab0[st] * x - aa2[st] * y;
|
|
x = y;
|
|
}
|
|
s = s + x;
|
|
}
|
|
// Box-filter average over the window, then clamp to [-1, +1] as a
|
|
// digital safety guard so callers casting to fixed-point cannot get
|
|
// wrap/click from small Butterworth transient overshoot. Branchless
|
|
// packed -- one minpd, one maxpd via the GCC vector built-ins.
|
|
double inv = 1.0 / (double)n;
|
|
paula_v2df out = s * (paula_v2df){inv, inv};
|
|
paula_v2df hi = {1.0, 1.0};
|
|
paula_v2df lo = {-1.0, -1.0};
|
|
out = __builtin_ia32_minpd(out, hi);
|
|
out = __builtin_ia32_maxpd(out, lo);
|
|
output[2 * i] += (float)out[0];
|
|
output[2 * i + 1] += (float)out[1];
|
|
}
|
|
|
|
p->fixed_lp = fl;
|
|
p->led_z1 = lz1;
|
|
p->led_z2 = lz2;
|
|
for(uint32_t st = 0; st < 4; ++st) {
|
|
p->aa_z1[st] = az1[st];
|
|
p->aa_z2[st] = az2[st];
|
|
}
|
|
p->decim_phase = phase;
|
|
|
|
#ifdef PAULA_PROFILE
|
|
struct timespec prof_t1;
|
|
clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &prof_t1);
|
|
paula_profile_cpu_ns += (double)(prof_t1.tv_sec - prof_t0.tv_sec) * 1.0e9 + (double)(prof_t1.tv_nsec - prof_t0.tv_nsec);
|
|
paula_profile_frames += (uint64_t)frames;
|
|
#endif
|
|
}
|
|
|
|
#ifdef PAULA_PROFILE
|
|
// [=]===^=[ paula_profile_report ]===============================================================[=]
|
|
// Print the accumulated mixer cost as a realtime factor. Call once at exit.
|
|
static void paula_profile_report(int32_t sample_rate) {
|
|
if(paula_profile_frames == 0) {
|
|
fprintf(stderr, "paula_mix_frames: never called (this player has its own mixer, not paula.h)\n");
|
|
return;
|
|
}
|
|
double cpu_s = paula_profile_cpu_ns * 1.0e-9;
|
|
double audio_s = (sample_rate > 0) ? (double)paula_profile_frames / (double)sample_rate : 0.0;
|
|
double rt = (cpu_s > 0.0) ? audio_s / cpu_s : 0.0;
|
|
double core_pct = (audio_s > 0.0) ? 100.0 * cpu_s / audio_s : 0.0;
|
|
fprintf(stderr, "paula_mix_frames: %.3fs CPU for %.1fs audio -> %.1fx realtime (%.2f%% of one core)\n",
|
|
cpu_s, audio_s, rt, core_pct);
|
|
}
|
|
#endif
|