Conversation
src/pi/pi_controller.c
Outdated
| } | ||
| if (((dram_address | rom_address | longueur) & 3) != 0) | ||
| { | ||
| for(i = 0; i < longueur; ++i) |
There was a problem hiding this comment.
Can I suggest to add a tab in from of the for loop to keep code readable?
From the manual:
So, to keep code clean, I suggest to remove old code and assume DMA are always 16 bits aligned. What do you think? |
|
Depends on what the console actually does. Manuals are known to be "best practice for stability/speed" instead of "the hardware requires this". I would keep the former behavior until someone pops in with a hardware test. |
|
Thinking about it, you're surely right, if it's technically possible we can't assume it will not happen... |
|
I'm a little hesitant to merge this because it increases the code complexity a fair amount, for a performance gain on the raspberry pi / arm architectures. Do you have any quantitative data on the performance improvement from using this aligned memcpy in real world scenarios? If it's just moving a few hundred or thousand bytes per VI then it's probably not worth it. But if these loops are run for millions of iterations per VI then it would be. |
|
The performance increase is 1 to 2 % faster (timing to frame 501), depending on ROM on a Pi B. It would be interesting to know if any ROMS require byte-wise copy as if none do then the word-wise only implementation would be simpler than the current byte-wise copy. |
|
There is merit to this. RSP DMA transfers must be 8-byte aligned. The N64 ignores the lower 3 bits of the source and destination address. It also ignores the lower 3 bits of the length (or rather, it assumes the lower 3 bits are all 1's, then it adds 1). As long as we properly mask the source and destination addresses, the transfers should always be 8-byte aligned. @ricrpi I don't know if you're still around. I'm far less confident in the alignment of the PI DMA's, but I am very confident in the 8-byte alignment of the RSP DMA's. If you update this PR to just to 8-byte aligned RSP DMA's (with no fallback) I'd be willing to do a fair bit of testing on it |
|
I am trying to keep up with all the commits/issues but have very limited time to do any coding at the moment :-( I will try to take a look at this though as it shouldn't take too long to patch in. |
5 participants
If DMA transfers are word aligned then a more efficient copy can be done. I have not seen a dma transfer that is not word aligned however I have not extensively checked all roms so have left the safer byte-wise implementation in.