While working on the Phoenix BLUE (the FPGA Blitter for the ST with @ijor's core) I started wondering: why doesn't the TT have a Blitter? Could I take the Phoenix BLUE as a base and make a TT version? That idea became the Caladrius Blitter.
Caladrius (another mythical bird).
Why didn't Atari fit one, does anyone know?
My best guess is that a 32 MHz 68030 with caches was already quick at doing blits in software. The ST Blitter is a 16-bit chip with a 24-bit address range, built for the ST bus. Fitting one would have meant a new 32-bit part, more cost and more board space.
What's already in TOS 3.06
Digging through the reconstructed TOS 3.06 sources (thanks to the late Thorsten Otto for those), a surprising amount of Blitter support is still there:
- The boot-time probe is still there. At start-up the BIOS still does a tst.w $FFFF8A00 behind a bus-error trap, exactly like ST TOS. On a stock TT nothing answers, so TOS decides there's no Blitter.
- XBIOS 64 Blitmode() is still there. On a stock TT it returns 0. Fit a Blitter and it returns 2, "present but off", which is what SYSINFO reports if there were a TT Blitter.
- The VDI still has the plumbing. The line-A tables for switching between software and hardware drawing routines are all present. In 3.0x, though, the hardware routines are compiled out: vdidefs.h enables them for TOS 2.x and 4.x only. In their place are ten stubs that do nothing but rts. Switch the Blitter on and those stubs replace the drawing code, and that's when I got two bombs.
- The desktop still has the menu item. On a 68030 the desktop just relabels it: the Options menu's "Blitter" object is the same one that shows "Cache", and the "Blitter" text is still in the ROM.
- The routines exist for the 68030. Falcon TOS 4.0x has its Blitter code back, and that code already flushes the 68030 data cache around each blit. So putting it back into 3.06 was mostly a build switch rather than a rewrite.
So, is it even worth it?
Before spending money on hardware I wanted an answer, so:
- I added a Caladrius Blitter to Hatari based on the hardware research: a 32-bit, 68030 bus-master version of the ST Blitter that starts hidden, so stock TOS 3.06 stays safe.
- I patched TOS 3.06 (as a build option on the reconstructed sources) to put the Blitter routines back in.
- I added a Blitter item to the desktop Options menu next to Cache, as on the ST.
SysInfo showing the TT has a Blitter present but turned off:
The test
GEMBench 6.31 on the emulated TT (TOS 3.06, 640x200x4, caches on, run from ST-RAM). The reference is a standard TT with Cache on; the second run has Cache and Blitter on.
The results
They're interesting. Everything the screen does gets faster: blitting is 3.6x, text roughly 2x, scrolling 1.8x, and the display average is 170%. The CPU, maths and memory tests stay at 100%, as you'd expect. So even next to a cached 68030, a Blitter makes a real difference to GEM.
A caveat: this is Hatari, using its STE Blitter timing on an emulated TT. It shows the gains are real and how much of GEM benefits, but the real numbers will come from real hardware. That's the next step for the Caladrius Blitter.
Showing the two options side by side in the patched TOS:
