Linear data copy to VRAM

[Loothor]

I have some data I'd like to store in VRAM, and of course I'd like to transfer it as quickly as possible. As far as I know, this entails using sceGuCopyImage(). However, sceGuCopyImage seems tailored to placing images in VRAM with a rectangular layout, and I'd really like to just copy various sizes of data to specified VRAM addresses in a linear fashion. My code currently works fine using:

memcpy(pVRAMAddr, pData, size);

but this is not as quick as the DMA copy triggered by sceGuCopyImage from what I understand... Can somebody give me some info on how to do a DMA copy of data to VRAM in a linear fashion? I tried using various manglings of sceGuCopyImage, but most of them just seem to crash the PSP or write data non-linearly.

[matkeupon]

I also had a performance problem with memcpy and the VRAM (it seems, copying the content of the whole screen with it is longer than 1/60s...).

There is an alignment requirement in the use of sceGuCopyImage() if I remember correctly. But using the Dmac functions would be better in your case.

http://forums.ps2dev.org/viewtopic.php?t=3890

This post talks about alignment requirement of the Dmac functions. I don't know if there is a way to know if the transfer is finished, must be a stub no one has found yet.

[Jim]

Is it possible just to set width = gfx_x * gfx_y and height = 1 in the transfer? ie, pretend you're transferring one very long thin bitmap.

Jim

[jsgf]

You could, though I'm not sure what the limit is; if you can factor your copy area into two approximately equal products, then you can still do it in a 2D way.

Also, there doesn't seem to be much cost in issuing lots of separate copies. In PSPGL, I flip images by copying each scanline as a separate copy operation, with no obvious performance hit compared to doing a single bulk copy.

[Loothor]

The sceDmacMemcpy() stub worked perfectly. Thanks so much for pointing out that thread. I didn't have any problems with alignment as discussed in that thread, but all of the data I am transferring is aligned on 16-byte boundaries.