On 11/09/2024 16:39, Ilia Mirkin wrote:
On Wed, Sep 11, 2024 at 10:19 AM Jocelyn Falempe <jfalempe@xxxxxxxxxx
<mailto:jfalempe@xxxxxxxxxx>> wrote:
On 06/09/2024 21:36, James Jones wrote:
> Right, there are 3 iterations of block linear tiling actually.
NV50 does
> support scanout of block linear surfaces. All block-linear-
capable GPUs
> do. The 3 generations are:
>
> NV5x/G8x/GTXXX line: Original block size.
> GFXXX(nvc0 I believe in nouveau terms)-GV100: double the block
height I
> believe.
> GTXXX+: Same block size, but the layout within a block is subtly
> different, at least as visible in CPU mappings.
>
I'm looking at how to check for specific chip in nouveau, and fix the
tiling for other cards than Turing.
It looks like in most case nouveau uses device->info.chipset, with
hardcoded hex value. so for nvc0+ I should check device-
>info.chipset >=
0xc0 ?
chipset < c0 : block_height 4, "old layout"
chipset >= c0 : block_height 8, "old layout"
chispet >= ?? : block_height 8, "new layout"
For testing, I have at hand a GTX1650 (Turing) and an old Geforce
8800GTS (Tesla?), so it's a NV92, and still uses this nv50+ code ?
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/
drivers/gpu/drm/nouveau/nvkm/engine/device/base.c#n2399 <https://
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/
gpu/drm/nouveau/nvkm/engine/device/base.c#n2399>
So probably >= 0x160
There should also be a device->card_type which is an enum that you can
use too, depending on what you have available? Set here:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/
drivers/gpu/drm/nouveau/nvkm/engine/device/base.c#n3177 <https://
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/
gpu/drm/nouveau/nvkm/engine/device/base.c#n3177>
Once you have a patch tested, ideally someone with the hardware should
test on a >= fermi && < turing card.
I'm testing on my 8800 GTS.
The problem I have is that fb->modifier is 0, but the buffer is still
tiled. I found there is a nouveau_framebuffer_get_layout() which gives
tile_mode and kind, and works on my Tesla.
So if I understand correctly, if kind != 0, there is tiling, and the
block size is 8 * (1 << tile_mode). (and half this on Tesla).
At least tiling is much easier on Tesla, there is no tiling inside the
block. In my test, block size is 4K, 16x64 pixels, tile_mode is 4, and
kind is 122.
I will send a v3 which works on Tesla and Turing soon.
Thanks for your help.
--
Jocelyn
Hope this helps,
-ilia
