On Tue, 2010-02-23 at 12:27 -0800, Brian Paul wrote: > José Fonseca wrote: > > On Mon, 2010-02-22 at 06:34 -0800, José Fonseca wrote: > >> On Sun, 2010-02-21 at 06:40 -0800, Marek Olšák wrote: > >>> Hi José, > >>> > >>> the attached patch fixes incorrect swizzles in u_format.csv. There are > >>> basically just 2 drivers which depend on the swizzles in this table: > >>> llvmpipe and r300g. Since r300g started supporting pretty much every > >>> texture format except SCALED ones, a few regressions have showed up. > >>> This patch resolves all issues I had, especially with the SRGB formats > >>> but I decided to clean it up all. git log: > >>> > >>> util: fix swizzles in the format table for 8-bits-per-channel > >>> formats > >>> > >>> The 8-bits-per-channel formats having the same channel order had > >>> completely > >>> different swizzles, and at the same time, the same swizzles were > >>> used for > >>> completely different channel orders of 8bpc formats. > >>> > >>> This made the whole table self-contradicting, caused headaches, > >>> and last > >>> but not least, incorrent rendering for the drivers relying on > >>> these swizzles. > >>> > >>> I hope I got it right. I didn't make a special distinction between the > >>> array and arith layouts. All I did was to make sure that if I grep > >>> e.g. A8R8G8B8, I'll get the same swizzles and not entirely different > >>> ones. > >> Hi Marek, > >> > >> I'll need a bit more time to investigate this. > >> > >> One problem is that the interpretation of the swizzle varies with array > >> vs arith. The ordering for array is the lowest significant word to the > >> highest significant word (where word for 8bit formats is a byte), where > >> for arith it goes from least significant bit to the highest significant > >> bit. This is the same difference as array indexation and bit shifts. > >> > >> There is also the problem of byte order which affects the bit shift > >> interpretation... > >> > >> I admit thet the current format description table is terribly > >> underdocumented/confusing and likely broken in several ways. I wrote it > >> to be able to code generate pixel format translation (which is wroking > >> reasonably) and never expected people to use it for hardware drivers as > >> well, although it is perfectly legitimate use. > >> > >> I have my own interpretation of these concepts, you and others hw driver > >> writers have their own different interpretation. Furthermore in > >> draw/translate/util modules there are some inconsistencies in these > >> interpretations too. So I need to read the GL and DX10 specs very well, > >> see how current drivers are using the descriptions, and come up with > >> something that makes sense for everybody. > >> > >> So please hold on to your patch for a couple of days. > >> > >> I'd appreciate if the interested parties could take a good look to > >> u_format.h comments, and summarize what they think the format semantics > >> should be. > >> > >> Jose > > > > > > There are two inconsistencies in formats ATM: > > a) the notation used in PIPE_FORMAT_xxx, and > > b) the values in util_format_description::swizzles . > > > > > > There's a D3D9 <-> D3D10 format conversion table in > > http://msdn.microsoft.com/en-us/library/ee415668(VS.85).aspx#Porting_Content > > and D3D9 <-> GL format table in > > http://source.winehq.org/git/wine.git/?a=blob;f=dlls/wined3d/utils.c;hb=HEAD > > . > > > > D3D10 dropped all arithmetically encoded formats, and inverted the > > swizzle notation (e.g., D3DFMT_A2B10G10R10 and DXGI_FORMAT_R10G10B10A2 > > are equivalent). > > > > Gallium has to represent both kinds and mixes both notations (the > > MSB->LSB notation traditionally used for texture formats, the LSB->MSB > > for vertex declarations). So instead of the current inconsistencies, > > both on p_format.h and u_format.csv, I suggest we all normalize on one > > notation, lets say MSB->LSB for pixel/vertex formats. > > > > For example, the vertex format > > > > struct vertex { > > float x; > > float y; > > }; > > > > should be described the format PIPE_FORMAT_G32R32_FLOAT (not the current > > PIPE_FORMAT_R32G32_FLOAT), which is equivalent: > > - D3D9's D3DFMT_G32R32F texture format > > - D3D9's D3DDECLTYPE_FLOAT2 vertex declaration type > > - D3D10's DXGI_FORMAT_R32G32_FLOAT > > - OpenGL's GL_RG32F format > > - OpenGL's glVertexAttrib2f attribute > > - OpenGL's glVertexPointer(2, GL_FLOAT, 0, 0); > > - etc. > > > > > > For the util_format_description::swizzles I suggest we always refer the > > swizzles from LSB->MSB. > > > > > > Leaving the interface change aside for now (Keith is away this week), > > unless somebody has a better suggestion I'll update at least the meaning > > of util_format_description::swizzles and u_format.csv to be consistent > > with this. > > > > Does everybody agree this is a sensible thing to do? > > It feels a little unnatural that an XY vertex format would be > described by PIPE_FORMAT_G32R32_FLOAT, but if that makes things more > consistant I can live with it. > > I think the priority is to have consistency.
FWIW, I'm perfectly fine standardizing in PIPE_FORMAT_R32G23_FLOAT and LSB->MSB everywhere. Jose ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev _______________________________________________ Mesa3d-dev mailing list Mesa3d-dev@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mesa3d-dev
