Slow DSP investigation with DFB1

[exxos]

So, how do you write the firmware on the DFB1X?

[stephen_usher]

Ah. I'll have to get the old laptop out of hiding then.

[Badwolf]

Ah. I'll have to get the old laptop out of hiding then.

xc3sprog will do it with appropriate cable. Or the Raspberry Pi trick that's demo'd around various places -- usually in connection with the Terriblefire boards.

Code: Select all

xc3sprog -c jtaghs2 -p 0 -v <firmware.jed>

for my cable.

BW

[mikro]

Well, it didn't take long for someone to find something to complain about. Was expecting something from you like this in fact already.

Not sure what triggered such a hostile reply, I just wanted to point out that Centek had very likely found out the same thing and decided to take another path. In fact I don't own the card anymore so I have no reason to complain.

The implementation can be done surely in a more elegant way than this:

ct2a_patch_board.jpg (335.33 KiB) Viewed 24 times

but the point is that custom GALs bring much more than just a fixed DSP delay.

Nembench results when accessing ST RAM (MByte/s):

Code: Select all

  ┌───────────────────────────────────┬────────┬────────┬──────┐
  │               Test                │ DFB1X  │  CT2   │  Δ   │                                                                                                                        
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 32bit read ST-RAM          │  5.297 │ 10.240 │ +93% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 32bit write ST-RAM         │  6.406 │ 11.477 │ +79% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 32bit copy ST-RAM          │  3.209 │  5.738 │ +79% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 16bit read ST-RAM          │  4.890 │  9.256 │ +89% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 16bit write ST-RAM         │  5.779 │ 11.497 │ +99% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 16bit copy ST-RAM          │  2.066 │  3.828 │ +85% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 8bit read ST-RAM           │  4.190 │  6.660 │ +59% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 8bit write ST-RAM          │  2.894 │  5.748 │ +99% │
  ├───────────────────────────────────┼────────┼────────┼──────┤
  │ Linear 8bit copy ST-RAM           │  1.034 │  1.914 │ +85% │

DSPBench host port tests:

Code: Select all

  ┌─────────────────────────┬───────┬───────┬───────┐
  │          Test           │ DFB1X │  CT2  │   Δ   │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ Raw write 24bit x:H:M:L │ 1.676 │ 3.561 │ +112% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ Raw read 24bit x:H:M:L  │ 1.611 │ 3.366 │ +109% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ Raw write 16bit -:-:M:L │ 2.232 │ 4.628 │ +107% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ Raw read 16bit -:-:M:L  │ 2.067 │ 4.158 │ +101% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ TX 24b/packed (ST-RAM)  │ 1.373 │ 2.763 │ +101% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ TX 24b/longs            │ 1.250 │ 2.497 │ +100% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ RX 24b/longs            │ 1.250 │ 2.597 │ +108% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ TX 16b/words            │ 1.630 │ 3.051 │  +87% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ RX 16b/words            │ 1.630 │ 3.434 │ +111% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ TX 8b/bytes             │ 1.536 │ 2.781 │  +81% │
  ├─────────────────────────┼───────┼───────┼───────┤
  │ RX 8b/bytes             │ 1.280 │ 2.668 │ +108% │
  └─────────────────────────┴───────┴───────┴───────┘

TL;DR version:
- ST-RAM: about 1.6x - 2.0x faster on CT2
- FastRAM: 11 - 30% faster on CT2.
- DSP host port: roughly 2x faster on CT2 in every mode but something is surely gained also by the fact that the DSP runs at 50 MHz

I'd argue that those results are worth putting in "a board and bodge wires".

[exxos]

@mikro where did you get the DFB1X DSPBENCH figures from? They don't match what I posted, and they look like figures from before the GAL fix.

With the new U68 and the 1.06 test firmware, DSPBENCH reads 99% to 102% of a stock Falcon on every host port test. That's with the DSP and the bus both running at stock speed, no overclocking anywhere.

[mikro]

Of course they are figures from a stock DFB1X. I don't have spare GALs to program.

[exxos]

Got AI to build a corrected tablet. Someone else can verify if they are bothered..

Note these speeds are with a stock DSP clock and bus.

Code: Select all

DSPBENCH v3.0b host port (MB/sec)

Test                       Old DFB1   Fixed DFB1X   CT2     CT2 vs fixed
Raw write 24bit x:H:M:L    1.676      3.603         3.561   -1%
Raw read  24bit x:H:M:L    1.611      3.015         3.366   +12%
Raw write 16bit -:-:M:L    2.232      4.380         4.628   +6%
Raw read  16bit -:-:M:L    2.067      3.330         4.158   +25%
TX 24b/packed (ST-RAM)     1.373      2.441         2.763   +13%
TX 24b/longs               1.250      2.107         2.497   +19%
RX 24b/longs               1.250      2.198         2.597   +18%
TX 16b/words               1.630      2.360         3.051   +29%
RX 16b/words               1.630      2.767         3.434   +24%
TX 8b/bytes                1.536      1.792         2.781   +55%
RX 8b/bytes                1.280      1.872         2.668   +43%

Fixed DFB1X = stock 16MHz bus, stock 32MHz DSP (99% to 102% of a stock Falcon)
CT2         = faster bus, 50MHz DSP

NOTE: This is NOT a fair comparison. The fixed DFB1X runs a stock 16MHz bus
and a stock 32MHz DSP. The CT2 runs a faster bus and a 50MHz DSP, so the
difference above comes from overclocking, not from the GAL fix. For a like
for like comparison, the CT2 would need to be run at a 16MHz bus and a
32MHz DSP.

[exxos]

Of course they are figures from a stock DFB1X. I don't have spare GALs to program.

That's not really the point though. The fixed results were already posted in this thread, in code tags, from the same DSPBENCH build you used. You chose to ignore them and quote old figures from before the fix instead, then put them in a table to show the CT2 being "roughly 2x faster".

You don't need spare GALs to copy numbers from a post. So I have to ask, why is that? Because from where I'm sitting, it looks like the table was put together to make the DFB1X look as bad as possible.

[mikro]

I sold my card to a friend of mine. We tested the card on his machine, I tested my CT2 on my machine. I posted the results gathered from both.

[exxos]

I sold my card to a friend of mine. We tested the card on his machine, I tested my CT2 on my machine. I posted the results gathered from both

Fair enough, that explains where they came from. The problem is the DFB1X figures are from before the fix, in a thread where the fixed results were already posted from the same DSPBENCH build. With the new U68 and 1.06 firmware the DFB1X is at 99% to 102% of a stock Falcon on every host port test.

It's also not a like for like comparison. The CT2 runs a faster bus and a 50MHz DSP, while the DFB1X runs everything at stock. To compare the two fairly, the CT2 would need to be run at a 16MHz bus and a 32MHz DSP.