Slow DSP investigation with DFB1

[exxos]

I've established a baseline first.

GB6 was calibrated first of all. Then a second run.

IMG_5374.JPG (327.44 KiB) Viewed 62 times

IMG_5377.JPG (380.4 KiB) Viewed 62 times

Then with DFB1X plugged in , jumper set to DISABLED.

IMG_5375.JPG (355.32 KiB) Viewed 62 times

IMG_5376.JPG (352.16 KiB) Viewed 62 times

So as GB6 depends on timer ticks, 1-2% variation is easily normal anyway.

[exxos]

DSPBENCH & NEMBENCH


STOCK

Code: Select all

DSPBench v3.0b - precision DSP tests. dml/2026

memory integrity test:
 DSP-local IX:IY $0000-$01FF march OK
 SRAM EX:EY $0200-$3FFF march OK

HOST interrupt test:
 RXDF/CPU interrupts triggered OK
 TXDF/CPU interrupts triggered OK

hostport raw R/W datarate:
 write ( 24bit x:H:M:L )   -> 3.603 MB/sec (~100%)
 read  ( 24bit x:H:M:L )   -> 3.049 MB/sec (~100%)
 write ( 16bit -:-:M:L )   -> 4.410 MB/sec (~100%)
 read  ( 16bit -:-:M:L )   -> 3.326 MB/sec (~100%)

host transfer rate (STRam)
 TX ( 24b/packed )         -> 2.441 MB/sec (~100%)
 TX ( 24b/longs )          -> 2.107 MB/sec (~100%)
 RX ( 24b/longs )          -> 2.198 MB/sec (~100%)
 TX ( 16b/words )          -> 2.343 MB/sec (~100%)
 RX ( 16b/words )          -> 2.758 MB/sec (~100%)
 TX (  8b/bytes )          -> 1.761 MB/sec (~100%)
 RX (  8b/bytes )          -> 1.845 MB/sec (~100%)

ALU & SRAM speed:
 mac ( P:Int X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Int X:Ext Y:Ext ) -> 8.000 Mips (~100%)
 mac ( P:Ext X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Ext X:Ext Y:Ext ) -> 5.333 Mips (~100%)

Code: Select all

NemBench v2.2 - precision CPU/FPU profiler.

Integer multiply (16bit)      -> 0.616 Mips (~100%)
Integer divide (16bit)        -> 0.363 Mips (~100%)
Linear (stalled) integer      -> 7.992 Mips (~100%)
Interleaved (piped) integer   -> 8.000 Mips (~100%)

16bit read (100% cached)      -> 7.898 MByte/sec (~100%)
16bit write (100% cached)     -> 6.075 MByte/sec (~100%)
16bit copy (100% cached)      -> 5.096 MByte/sec (~84%)
32bit read (100% cached)      -> 15.785 MByte/sec (~100%)
32bit write (100% cached)     -> 6.720 MByte/sec (~100%)
32bit copy (100% cached)      -> 6.693 MByte/sec (~100%)

Linear 32bit read (ST-Ram)    -> 5.349 MByte/sec (~100%)
Linear 32bit write (ST-Ram)   -> 6.508 MByte/sec (~100%)
Linear 32bit copy (ST-Ram)    -> 3.260 MByte/sec (~101%)


Linear 16bit read (ST-Ram)    -> 4.563 MByte/sec (~85%)
Linear 16bit write (ST-Ram)   -> 5.872 MByte/sec (~91%)
Linear 16bit copy (ST-Ram)    -> 2.097 MByte/sec (~64%)


Linear 8bit read (ST-Ram)     -> 2.854 MByte/sec (~53%)
Linear 8bit write (ST-Ram)    -> 2.936 MByte/sec (~45%)
Linear 8bit copy (ST-Ram)     -> 1.047 MByte/sec (~32%)

DISABLED

Code: Select all

DSPBench v3.0b - precision DSP tests. dml/2026

memory integrity test:
 DSP-local IX:IY $0000-$01FF march OK
 SRAM EX:EY $0200-$3FFF march OK

HOST interrupt test:
 RXDF/CPU interrupts triggered OK
 TXDF/CPU interrupts triggered OK

hostport raw R/W datarate:
 write ( 24bit x:H:M:L )   -> 3.603 MB/sec (~100%)
 read  ( 24bit x:H:M:L )   -> 3.049 MB/sec (~100%)
 write ( 16bit -:-:M:L )   -> 4.404 MB/sec (~100%)
 read  ( 16bit -:-:M:L )   -> 3.326 MB/sec (~100%)

host transfer rate (STRam)
 TX ( 24b/packed )         -> 2.441 MB/sec (~100%)
 TX ( 24b/longs )          -> 2.107 MB/sec (~100%)
 RX ( 24b/longs )          -> 2.198 MB/sec (~100%)
 TX ( 16b/words )          -> 2.343 MB/sec (~100%)
 RX ( 16b/words )          -> 2.758 MB/sec (~100%)
 TX (  8b/bytes )          -> 1.763 MB/sec (~100%)
 RX (  8b/bytes )          -> 1.845 MB/sec (~100%)

ALU & SRAM speed:
 mac ( P:Int X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Int X:Ext Y:Ext ) -> 8.000 Mips (~100%)
 mac ( P:Ext X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Ext X:Ext Y:Ext ) -> 5.333 Mips (~100%)

(forgot to copy the NEMBENCH results, but assume the same as above)


ENABLED

Code: Select all

DSPBench v3.0b - precision DSP tests. dml/2026

memory integrity test:
 DSP-local IX:IY $0000-$01FF march OK
 SRAM EX:EY $0200-$3FFF march OK

HOST interrupt test:
 RXDF/CPU interrupts triggered OK
 TXDF/CPU interrupts triggered OK

hostport raw R/W datarate:
 write ( 24bit x:H:M:L )   -> 1.955 MB/sec (~54%)
 read  ( 24bit x:H:M:L )   -> 1.795 MB/sec (~59%)
 write ( 16bit -:-:M:L )   -> 2.598 MB/sec (~59%)
 read  ( 16bit -:-:M:L )   -> 2.212 MB/sec (~66%)

host transfer rate (STRam)
 TX ( 24b/packed )         -> 1.549 MB/sec (~63%)
 TX ( 24b/longs )          -> 1.412 MB/sec (~67%)
 RX ( 24b/longs )          -> 1.410 MB/sec (~64%)
 TX ( 16b/words )          -> 1.780 MB/sec (~76%)
 RX ( 16b/words )          -> 1.790 MB/sec (~65%)
 TX (  8b/bytes )          -> 1.575 MB/sec (~90%)
 RX (  8b/bytes )          -> 1.383 MB/sec (~75%)

ALU & SRAM speed:
 mac ( P:Int X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Int X:Ext Y:Ext ) -> 8.000 Mips (~100%)
 mac ( P:Ext X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Ext X:Ext Y:Ext ) -> 5.333 Mips (~100%)

Code: Select all

NemBench v2.2 - precision CPU/FPU profiler.

Integer multiply (16bit)      -> 0.616 Mips (~100%)
Integer divide (16bit)        -> 0.363 Mips (~100%)
Linear (stalled) integer      -> 7.992 Mips (~100%)
Interleaved (piped) integer   -> 7.992 Mips (~100%)

16bit read (100% cached)      -> 7.898 MByte/sec (~100%)
16bit write (100% cached)     -> 6.082 MByte/sec (~101%)
16bit copy (100% cached)      -> 5.102 MByte/sec (~84%)
32bit read (100% cached)      -> 15.785 MByte/sec (~100%)
32bit write (100% cached)     -> 6.729 MByte/sec (~100%)
32bit copy (100% cached)      -> 6.693 MByte/sec (~100%)

Linear 32bit read (ST-Ram)    -> 5.358 MByte/sec (~100%)
Linear 32bit write (ST-Ram)   -> 6.508 MByte/sec (~100%)
Linear 32bit copy (ST-Ram)    -> 3.263 MByte/sec (~101%)


Linear 16bit read (ST-Ram)    -> 4.570 MByte/sec (~85%)
Linear 16bit write (ST-Ram)   -> 5.872 MByte/sec (~91%)
Linear 16bit copy (ST-Ram)    -> 2.097 MByte/sec (~64%)


Linear 8bit read (ST-Ram)     -> 2.879 MByte/sec (~54%)
Linear 8bit write (ST-Ram)    -> 2.941 MByte/sec (~45%)
Linear 8bit copy (ST-Ram)     -> 1.047 MByte/sec (~32%)

OPTION 2 JUMPERED

Code: Select all

DSPBench v3.0b - precision DSP tests. dml/2026

memory integrity test:
 DSP-local IX:IY $0000-$01FF march OK
 SRAM EX:EY $0200-$3FFF march OK

HOST interrupt test:
 RXDF/CPU interrupts triggered OK
 TXDF/CPU interrupts triggered OK

hostport raw R/W datarate:
 write ( 24bit x:H:M:L )   -> 1.955 MB/sec (~54%)
 read  ( 24bit x:H:M:L )   -> 1.777 MB/sec (~58%)
 write ( 16bit -:-:M:L )   -> 2.598 MB/sec (~59%)
 read  ( 16bit -:-:M:L )   -> 2.171 MB/sec (~65%)

host transfer rate (STRam)
 TX ( 24b/packed )         -> 1.549 MB/sec (~63%)
 TX ( 24b/longs )          -> 1.410 MB/sec (~67%)
 RX ( 24b/longs )          -> 1.410 MB/sec (~64%)
 TX ( 16b/words )          -> 1.735 MB/sec (~74%)
 RX ( 16b/words )          -> 1.790 MB/sec (~65%)
 TX (  8b/bytes )          -> 1.460 MB/sec (~83%)
 RX (  8b/bytes )          -> 1.383 MB/sec (~75%)

ALU & SRAM speed:
 mac ( P:Int X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Int X:Ext Y:Ext ) -> 8.000 Mips (~100%)
 mac ( P:Ext X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Ext X:Ext Y:Ext ) -> 5.333 Mips (~100%)

[exxos]

New firmware and new U68 U68turbo.jed.

Code: Select all

DSPBench v3.0b - precision DSP tests. dml/2026

memory integrity test:
 DSP-local IX:IY $0000-$01FF march OK
 SRAM EX:EY $0200-$3FFF march OK

HOST interrupt test:
 RXDF/CPU interrupts triggered OK
 TXDF/CPU interrupts triggered OK

hostport raw R/W datarate:
 write ( 24bit x:H:M:L )   -> 3.603 MB/sec (~100%)
 read  ( 24bit x:H:M:L )   -> 3.015 MB/sec (~99%)
 write ( 16bit -:-:M:L )   -> 4.380 MB/sec (~99%)
 read  ( 16bit -:-:M:L )   -> 3.330 MB/sec (~100%)

host transfer rate (STRam)
 TX ( 24b/packed )         -> 2.441 MB/sec (~100%)
 TX ( 24b/longs )          -> 2.107 MB/sec (~100%)
 RX ( 24b/longs )          -> 2.198 MB/sec (~100%)
 TX ( 16b/words )          -> 2.360 MB/sec (~100%)
 RX ( 16b/words )          -> 2.767 MB/sec (~101%)
 TX (  8b/bytes )          -> 1.792 MB/sec (~102%)
 RX (  8b/bytes )          -> 1.872 MB/sec (~101%)

ALU & SRAM speed:
 mac ( P:Int X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Int X:Ext Y:Ext ) -> 8.000 Mips (~100%)
 mac ( P:Ext X:Int Y:Int ) -> 16.000 Mips (~100%)
 mac ( P:Ext X:Ext Y:Ext ) -> 5.333 Mips (~100%)

Code: Select all

NemBench v2.2 - precision CPU/FPU profiler.

Integer multiply (16bit)      -> 1.921 Mips (~313%)
Integer divide (16bit)        -> 1.135 Mips (~313%)
Linear (stalled) integer      -> 24.899 Mips (~312%)
Interleaved (piped) integer   -> 24.899 Mips (~312%)

Float multiply (64bit)        -> 0.839 MegaFlops (~316%)
Float divide (64bit)          -> 0.543 MegaFlops (~313%)
Linear (stalled) float        -> 1.625 MegaFlops (~304%)
Interleaved (piped) float     -> 1.742 MegaFlops (~327%)

16bit read (100% cached)      -> 24.600 MByte/sec (~313%)
16bit write (100% cached)     -> 6.060 MByte/sec (~100%)
16bit copy (100% cached)      -> 4.995 MByte/sec (~83%)
32bit read (100% cached)      -> 49.140 MByte/sec (~313%)
32bit write (100% cached)     -> 6.711 MByte/sec (~100%)
32bit copy (100% cached)      -> 6.702 MByte/sec (~100%)

Linear 32bit read (ST-Ram)    -> 5.371 MByte/sec (~101%)
Linear 32bit write (ST-Ram)   -> 6.495 MByte/sec (~100%)
Linear 32bit copy (ST-Ram)    -> 3.257 MByte/sec (~100%)


Linear 16bit read (ST-Ram)    -> 5.064 MByte/sec (~95%)
Linear 16bit write (ST-Ram)   -> 5.851 MByte/sec (~90%)
Linear 16bit copy (ST-Ram)    -> 2.095 MByte/sec (~64%)


Linear 8bit read (ST-Ram)     -> 4.233 MByte/sec (~79%)
Linear 8bit write (ST-Ram)    -> 2.925 MByte/sec (~45%)
Linear 8bit copy (ST-Ram)     -> 1.044 MByte/sec (~32%)

Feed it all into AI for the round up. Someone else can check it

1.PNG (37.34 KiB) Viewed 44 times

2.PNG (31.85 KiB) Viewed 44 times

[exxos]

ACE TRACKER running.





CUBASE loads, but won't play anything unless I connect up my SCSI drive, and can't be bothered tonight

[viking272]

New firmware and new U68 U68turbo.jed.

Wow, that's a great result, all looks really positive.
Is this the work of Pakman, BW and yourself? Thanks to you all!

[exxos]

Wow, that's a great result, all looks really positive.
Is this the work of Pakman, BW and yourself? Thanks to you all!

Its @Badwolf and @Pacman work. I only tested it.

[Badwolf]

ACE TRACKER running.

Cool! I wonder why you don't have the black screen of doom I get on Acetracker when switching resolutions?

Its @Badwolf and @Pacman work. I only tested it.

The GAL is Holger's work (@pakman). Chapeau.

BW

[exxos]

Cool! I wonder why you don't have the black screen of doom I get on Acetracker when switching resolutions?

Don't think heard of that one.. Never had any troubles really. but hardly use it either. Bad install?

I'd presume the new U68 would fix the DSP issue on the CT60 as well..?

[viking272]

I'd presume the new U68 would fix the DSP issue on the CT60 as well..?

So is the original U68 on the Falcon at fault then? Can it be re-programmed easily or does it need to be replaced?

[exxos]

So is the original U68 on the Falcon at fault then? Can it be re-programmed easily or does it need to be replaced?

Yes..

When any booster (or DMA master) owns the bus, Atari’s U68 GAL waits two bus clocks after AS before selecting the DSP, against one for the Falcon’s own CPU, roughly halving host port speed. U68turbo selects the DSP immediately when the booster is master, and the 1.06 firmware ends DSP cycles itself one bus clock after XAS instead of waiting for the Falcon’s late DTACK

I reprogrammed the chip in my falcon.. Mostly out of desperation as I couldn't find any other chips anywhere when I've got a tube somewhere