Blitter screen copy where the DMA chip plays the sound

[diegoparrilla]

What

On a machine with a blitter the ST can copy each frame to the screen with the blitter instead of the 68000's MOVEM loop. By default it does so only where the DMA sound chip plays the sound (an STE, a Mega STE), and keeps the MOVEM loop on the YM path.

• userfw asks TOS for a blitter at boot (XBIOS Blitmode(-1)) and reports it after the hello ($FB8Exx, st_session_features(), stSessionFeatures over SWD).
• The blitter writes each 48-byte chunk where the predecrement MOVEM would (destination Y increment −94, then the tail), so the RP's cart framebuffer layout is unchanged. It owns the bus 40 chunks at a time (about 1 ms, under an IKBD byte's 1.28 ms); interrupts run between pieces.
• Shared slot 7 ($FA402C, CART_BLIT_MODE_OFFSET), read every VBL, picks the copy: fb_set_copy_mode() with auto (the default), CPU or blitter, plus the chunks per piece.
• FB_SLACK_REPORT now measures with MFP Timer-A (no interrupt, 81 µs a count), on both sound paths; one byte at $FB8Bxx ($FB8Cxx is free again). Still off by default.

Why only on the DMA path

Both copies are bound by the bus (8 cycles a word against MOVEM's 9). While the blitter owns the bus, Timer-B's YM samples wait. Measured on a Mega ST (YM sound, blitter fitted):

Copy	Slack after the blit	Sound (440 Hz tone)
CPU	0.2–0.3 ms	reference
Blitter sharing the bus	none, 25 fps	–
Blitter, 4 / 7 chunks a piece	0.4 / 1.0 ms	audibly rougher
Blitter, 40 chunks a piece	3.4 ms	breaks (Timer-B ticks dropped)

On a Mega STE (TOS 2.06, 8 MHz, DMA sound): 1.9–2.0 ms with the CPU, 3.3–3.4 ms with the blitter, the tone unchanged, and 50 fps with the mouse moved flat out.

Tests

• make -C tests/host test: test_chunk_layout.c models the blitter's copy (in one run and in pieces of 1, 7, 40, 255 and 666 chunks) against every pixel drawn; test_layout.py checks the new shared constants and the Timer-A count.
• Mega ST, debug build: blitter detected, every copy mode from the host, the screen correct, ESC to GEM (desktop and a window drawn by TOS) and a reset back into the app, the mouse with 40-chunk pieces (no IKBD overrun or resync).
• Mega STE, debug build: the measurements above, the tone A/B, the mouse flat out with music.
• Mega STE, release build: the menu, demos 1–4, the input test with the mouse, ESC to GEM and a reset: 50 fps, no late audio slice or underrun, no IKBD error.