Audio: Paula, the A500 filters and the stereo mix¶

This note explains the AExp audio path: Paula, the Amiga-specific output filters and the stereo crossfeed (both faithful ports from MiSTer's Minimig.sv), the OSM master volume, and how all of that differs from the generic MiSTer "audio improvements" filter that the M2M framework also carries (and which AExp keeps disabled).

It is written for a developer who knows neither MiSTer nor the Amiga audio hardware in detail. The short version is:

The audio chain is: raw Paula -> A500 fixed low-pass (switchable) -> LED filter (follows the simulated power LED) -> stereo crossfeed -> OSM master volume -> both HDMI and analog output.
The filters and the crossfeed live in CORE/vhdl/audio_filters.vhd and are bit-faithful to MiSTer Minimig.sv; the OSM defaults (A500 Filter on, LED Filter on, Full Stereo, 100% volume) reproduce a real A500.
With both filters off and Full Stereo, the path is bit-transparent raw Paula.
The generic M2M audio_out filter is a different thing: MiSTer output conditioning, not an Amiga model. AExp has the right coefficients for it in CORE/vhdl/globals.vhd but keeps qnice_audio_filter_o tied to '0'.
The end-user description of all of this is audio.md.

1. What is Paula?¶

The Amiga 500 audio chip is called Paula. In a real A500, Paula plays four 8-bit DMA audio channels: two are mixed to the left output and two to the right output. The Amiga software writes sample data into RAM, Paula fetches it at the programmed period, and the analog output circuitry turns the result into left/right audio.

Minimig models Paula and exposes a left and right digital sample stream:

ldata[14:0]
rdata[14:0]

Those are 15-bit signed Paula samples (channels 0+3 summed to the left, 1+2 to the right — the sum of two sign-extended 14-bit voice products, so it cannot overflow). MiSTer widens them to 16-bit signed PCM by shifting left once ({ldata, 1'b0}); AExp does the same widening at the input of audio_filters.vhd.

2. What MiSTer does¶

MiSTer has two relevant audio layers:

The core layer. For Minimig, this contains Paula and the Amiga-model filters.
The MiSTer system/output layer. This contains a generic final audio_out block used to prepare audio for the MiSTer output hardware.

This distinction matters. The MiSTer final audio_out filter and the Amiga-specific A500 filters are not the same thing.

2.1 The Minimig/Amiga-specific filters¶

In MiSTer Minimig, Paula does not go straight to the final output stage. Minimig.sv first builds paula_smp_l and paula_smp_r, then optionally passes them through Amiga-model filters.

The local source is:

CORE/Minimig_MiSTerMEGA65/Minimig.sv

The upstream source is:

https://github.com/MiSTer-devel/Minimig-AGA_MiSTer/blob/MiSTer/Minimig.sv

The relevant stages are:

lpf4400: a first-order low-pass filter around 4.4 kHz. This models the fixed A500 output RC filter (the A1200 dropped it).
lpf3275: a filter described in the source as "LPF 3000Hz 1st + 3400Hz 1st". This models the switchable Amiga LED filter.

MiSTer controls those filters with these signals:

wire flt_en    = ~status[48] ? pwr_led : status[47];
wire aud_1200  = status[49];
wire paula_pwm = status[50];

In plain language:

paula_pwm selects an alternate PWM-volume Paula path (not ported).
aud_1200 bypasses the A500 fixed low-pass stage.
flt_en controls the LED filter; the default "Auto(LED)" mode follows the Amiga power LED state live.

pwr_led is CIA-A PRA bit 1 inverted (minimig.v: pwr_led = ~_led), i.e. pwr_led = '1' means "LED bright" means "filter engaged" — the classic Amiga power-LED/audio-filter coupling.

MiSTer's stereo mix for Paula's hard-panned channels is a third piece: the userio command 0xF2 stores a 2-bit aud_mix value that minimig merely re-exports; the actual blending happens in MiSTer's framework (sys/audio_out.sv, module aud_mix_top).

2.2 The generic MiSTer final audio_out filter¶

After the core has produced its final audio stream, MiSTer routes it through audio_out (wired in sys/sys_top.v). This final stage synchronizes audio into the output clock domain, can apply a configurable IIR filter, applies a DC blocker, and supports attenuation and stereo-to-mono mixing.

The default coefficients in MiSTer Minimig are:

aflt_rate = 7056000
acx       = 4258969
acx0      = 3
acx1      = 3
acx2      = 1
acy0      = -6216759
acy1      =  6143386
acy2      = -2023767

Those are generic MiSTer output-filter defaults, not Amiga-specific: the same values appear in several MiSTer cores because this is part of the shared output path.

3. What AExp implements¶

Minimig Paula (15-bit signed L/R)
  -> audio_filters.vhd: widen to 16-bit signed ({data, 1'b0})
  -> audio_filters.vhd: lpf4400 A500 fixed low-pass   [OSM "A500 Filter", default ON]
  -> audio_filters.vhd: lpf3275 LED filter            [OSM "LED Filter" arms it; engaged = armed AND pwr_led]
  -> audio_filters.vhd: stereo crossfeed              [OSM "Stereo: %s", default Full Stereo]
  -> main.vhd audio_volume_proc: OSM master volume    [default 100% = bit-transparent]
  -> M2M av_pipeline (raw path, qnice_audio_filter_o = '0')
  -> HDMI audio and MEGA65 analog audio jack

CORE/vhdl/audio_filters.vhd is a faithful port of the Minimig.sv output stage between Paula and the MiSTer framework:

The two IIR instances reuse M2M's copy of MiSTer's 3-tap stereo IIR (M2M/vhdl/controllers/MiSTer/iir_filter.v — functionally identical to the submodule's sys/iir_filter.v, only declaration order differs) with the Minimig.sv coefficients verbatim; coefficient ports that Minimig.sv leaves unconnected are tied to zero, which is what synthesis makes of an unconnected input port.
ce is clk7_en or clk7n_en (14.19 MHz). The IIR time-multiplexes both channels on it, so each channel updates at the 7.09 MHz rate the coefficients are designed for — the same clocking as MiSTer after their "Move filters to system clock" change.
The lpf4400 runs unconditionally so its state is warm when switched into the path; the A500 Filter toggle only selects it. The lpf3275 is chained behind the model select exactly as in Minimig.sv (with the A500 Filter off it operates on the raw Paula mix).
LED filter engagement is led_filter_i and pwr_led_i: the OSM toggle arms the mechanism, the simulated software controls it live (MiSTer "Auto(LED)" semantics; there is no "force always on" mode).
MiSTer's old_l0/old_l1 double-latch after the filter mux is their CDC into CLK_AUDIO; AExp does not need it because the M2M framework's cdc_stable handles the domain crossing downstream.

3.1 Filter properties worth knowing¶

Both filters have an intrinsic DC gain slightly above unity — a property of the MiSTer coefficient sets, verified in simulation and present on MiSTer hardware as well:

lpf4400: gain 17/16 = +0.53 dB
lpf3275: about x1.137 = +1.11 dB

Full-scale material can therefore hit the IIR's internal 16-bit saturating clamp (iir_filter.v) with filters engaged. That is faithful MiSTer behavior: graceful saturation, same clamp, same coefficients. Apart from the gain, the measured responses match the analytic RC prototypes (4400 Hz 1st-order; 3000 Hz + 3400 Hz cascade) to within about 2% across the audio band.

3.2 The stereo crossfeed¶

The crossfeed replicates MiSTer's aud_mix_top blends with identical floor-shift arithmetic. In aud_mix_top, pre_in is the halved opposite channel (pre_out <= a2[16:1]), so the effective blends are:

OSM setting	aud_mix	own side	opposite side
Full Stereo	00	100%	0%
Wide Stereo	01	87.5%	12.5%
Narrow Stereo	10	75%	25%
Mono	11	50%	50%

All three blends are energy-preserving (the weights sum to 1.0), so the crossfeed itself cannot clip; the MiSTer-style 17-bit clamp is kept as a guard anyway. AExp computes both channels in the same cycle instead of MiSTer's registered pre_out ping-pong — value-identical per sample, minus one sample of opposite-channel delay that only exists at 48 kHz in MiSTer.

MiSTer configures the mix via userio command 0xF2; in AExp that command is a no-op (minimig's aud_mix output is unconnected in minimig_m65.v) and amiga_config.vhd keeps sending 0xF2 = 0. The OSM radio drives the crossfeed directly instead.

3.3 The OSM master volume¶

The "Volume" control is a 21-position radio from 100% down to 0% in 5% steps. mega65.vhd decodes the one-hot selection into a step index (0 = mute, 20 = 100%, the default), and main.vhd turns the index into an amplitude gain via C_VOL_LUT and multiplies the filtered stereo mix with it (registered, so Vivado maps it to two DSP48 slices).

Design properties:

The taper is perceptual, not linear in amplitude: each 5% step is a 5 percentage-point change in perceived loudness, so 50% sounds half as loud as 100% (-10 dB) and 25% a quarter (-20 dB). The gain follows (percent/100)^1.661, stored as unsigned Q15. The LUT values are identical to C64MEGA65, so both cores sound alike at the same slider position.
100% multiplies by Q15 0x8000, which is exactly transparent.
The gain never exceeds 1.0, so the multiply can never clip; 0% is a true digital mute.
The volume is the last stage in main.vhd, after the filters and the crossfeed: it models the volume knob on the monitor or amplifier, not part of the simulated machine. Everything Amiga-side stays untouched upstream: Paula's per-channel 6-bit volume registers, the 4-channel mix, and the filters.
It is applied at the single point ahead of the M2M split into the HDMI and analog paths, so both outputs track exactly.

config.vhd: the "Stereo: %s" submenu (radio group OPTM_G_STEREO), "A500 Filter" (OPTM_G_A500FILT) and "LED Filter" (OPTM_G_LEDFILT), the latter two single-select with OPTM_G_STDSEL = default ON. All three are HDL-read; no firmware logic is involved.
mega65.vhd: the C_MENU_STEREO range is decoded into MiSTer's 2-bit aud_mix encoding; C_MENU_A500FILT/C_MENU_LEDFILT are wired straight into main.vhd. All in the core clock domain from the static main_osm_control_i vector, like the keyboard and VGA bits.
The line numbers behind these constants follow the menu layout in config.vhd; a change there needs the menu checks and the heap budget of architecture.md.

4. The generic M2M "audio improvements" filter stays off¶

M2M always instantiates its audio_out block (M2M/vhdl/av_pipeline/av_pipeline.vhd), a port of MiSTer's generic output conditioning: an IIR low-pass with its -3 dB point around 18-20 kHz plus a DC blocker. A mux selects raw versus filtered audio for both the analog and the HDMI path, controlled by qnice_audio_filter_o — which AExp ties to '0' in mega65.vhd.

The C64 and Game Boy MEGA65 cores expose exactly this switch as their "Audio improvements" menu item. For AExp it stays off:

Its audible effect on Amiga material is marginal (about -0.4 dB at 15 kHz, -2.6 dB at 20 kHz, plus DC removal).
With the real A500 filters in place it is redundant, and a second, overlapping filter system in the menu would only confuse.

The globals.vhd coefficients for it are nevertheless correct and MiSTer-Minimig-faithful (audio_cx1 = 3; the M2M template and C64MEGA65 carry a cx1 = 2 typo worth roughly 0.4 dB), so enabling it later would be a one-line experiment.

5. Verification¶

The testbench CORE/sim/audio/tb_iir_amiga.v (iverilog) drives both IIR instances exactly as audio_filters.vhd instantiates them and checks the measured gains at DC, 1 kHz, 3.2 kHz, 4.4 kHz and 10 kHz against the analytic RC prototypes scaled by the intrinsic DC gains, plus channel-separation of the time-multiplexed stereo core (right channel silent while the left plays).
The testbench CORE/sim/audio/tb_audio_filters.vhd (nvc, with the +100-offset IIR stub CORE/sim/stubs/iir_stub_sim.vhd) proves the glue: bit-transparent bypass with everything off, the A500/LED mux decisions, LED gating (armed AND live pwr_led), and all crossfeed blends against golden values computed by an independent Python implementation of the aud_mix_top arithmetic.
CORE/sim/audio/run.sh runs both in a few seconds.

6. Practical conclusion¶

The default sound of AExp is the sound of a real A500: fixed filter in, LED filter under software control, hard-panned stereo. The brightest possible configuration (both filters off) remains available in the menu and is bit-transparent raw Paula. The stereo crossfeed is a pure listening-comfort option with no authenticity cost while it is off. The generic M2M output filter remains a deliberate non-feature.