| Age | Commit message (Collapse) | Author |
|---|
|
This patch implements nv04_surface_copy/fill using the new 2D engine module.
It supports falling back to the 3D engine using the u_blitter module, which will be
added in a later patch.
Also adds support for using the 3D engine, reusing the u_blitter module
created for r300.
This is used for unswizzling and copies between swizzled surfaces.
|
|
This patch add a brand new nv04-nv40 2D engine module.
It should correctly implement all operations involving swizzled, and 3D-swizzled surfaces.
This code is independent from the Gallium framework and can thus be reused in the DDX and classic Mesa drivers (it's only likely to be useful in the latter, though).
Currently, surface_copy and surface_fill are broken for 3D textures, for swizzled source textures and possibly for some misaligned cases
The code is based around the new nv04_region structure, which encapsulates the information from pipe_surface needed for the 2D engine and CPU copies.
The use of nv04_region makes the code independent of the Gallium framework and allows to transform the nv04_region without clobbering the nv04_region.
The existing M2MF, blitter, and SWIZZLED_SURFACE paths have been improved and a new CPU path has been added.
There is also support to tell the caller to use the 3D engine.
The main feature of the copy/fill setup algorithm is linearization/contiguous-linearization of swizzled surfaces.
The idea of linearization is that some swizzled surfaces are laid out like linear ones (1xN, 2xN, Nx1) and can thus be used as such (e.g. useful for copying single pixels).
Also, some rectangles (e.g. the whole surface) are contiguous in memory. If both the source and destination rectangles are swizzled but contiguous, then they can be regarded as both linear: this is the idea of "contiguous linearization".
This, for instance, allows to use the 2D engine to duplicate the content of a swizzled surface to another swizzled surface, by pretending they are actually linear.
After linearization, the result may not be 64-byte aligned. Another transformation is done to enlarge the linear surface so that it becomes 64-byte aligned.
This is also used to 64-byte align swizzled texture mipmaps.
The inner loop of the CPU path is as optimized as possible without using SSE/SSE2.
Future improvements could include SSE/SSE2 support, and possibly a faster coordinate swizzling algorithm (which is however not used in the inner loop).
It may be a good idea to autogenerate swizzling code at least for all possible POT 2D texture dimensions (less than 256), maybe for all 3D ones too (less than 4096).
Also, it woud be a very good idea to make a copy with the GPU first if the source surface is in uncached memory.
|
|
This greatly simplifies the code, and avoids ad-hoc copy code.
Also, these new transfers work for buffers too, even though they
are still used for miptrees only.
|
|
Changes:
- Disable swizzling on non-RGBA 2D textures, since the current 2D
code is mostly broken in those cases. A later patch will fix this.
Thanks to Andrew Randrianasulu who reported this.
- Fix compressed texture transfers and hack around the current 2D
code inability to copy compressed textures by using direct access.
Thanks to Andrew Randrianasulu who reported this.
This patch rewrites all the miptree layout and transfer code in the
nvfx driver.
The current code is broken in several ways:
1. 3D textures are laid out first by face, then by level, which is
incorrect
2. Cube maps should have 128-byte aligned faces
3. Swizzled textures have a strange alignment test that seems
unnecessary
4. We store the image_offsets for each face/slice but they can be
easily computed instead
5. "Swizzling" is not supported for compressed formats. They can be
"swizzled" but swizzling only means that there are no gaps (pitch is
level-dependant) and the layout is still linear
6. Swizzling is not supported for non-RGBA formats. All formats (except
possibly depth) can be swizzled according to my testing.
The miptree layout is rewritten based on my empirical testing, which I
posted in the "miptree findings" mail.
The image_offset array is removed, since it can be calculated with a
simple multiplication; the only array in the miptree structure is now
the one for mipmap level starts, which it seems cannot be easily
computed in constant time.
Also, we now directly store a nouveau_bo instead of a pipe_buffer in
the miptree structure, like nv50 does.
Support for render temporaries is removed, and will be readded in a
later patch.
Note that the current temporary code is broken, because it does not
copy the temporary back on render cache flushes.
|
|
Add a function to get whether a resource is likely on the GPU or not.
Currently always returns TRUE.
|
|
|
|
|
|
|
|
A source line was put in the wrong place.
|
|
The trace driver's implementation of sampler_view_destroy was calling
directly into the underlying pipe's sampler_view_destroy implementation.
This causes problems for pipes that keep references to sampler views
even after the state tracker has released them. Instead, we'll simply
drop the trace driver's reference to the pipe's sampler view.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
|
|
The trace driver was tracing the unwrapped version of the index buffer
when setting the index buffer. This caused an assert validating that
a resource belonged to the trace driver to fail. Instead, we'll log
the unmodified index buffer structure when setting the index buffer.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
|
|
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
|
|
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
|
|
Signed-off-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
|
|
|
|
|
|
|
|
Searched for them with:
git grep -E '[!=]=.*PIPE_TEXTURE_2D|PIPE_TEXTURE_2D.*[!=]=|case.*PIPE_TEXTURE_2D'
Behavior hasn't been changed.
|
|
|
|
|
|
|
|
To my knowledge, there is no way to flush zmask and thus write the clear
value.
This fixes zbuffer reads, among other things.
|
|
|
|
This pretty much ports the code from r600c, however it doesn't
always seem to work quite perfectly, but I can't find anything in this
code that is wrong. I'm guessing either literal input or constants
aren't working always.
|
|
|
|
|
|
the first two are straight op2's and the DDX/DDY are taken from r600c.
|
|
fixes fp-set-01 and glsl-fs-step
|
|
fixes biased texturing tests
|
|
Don't do perspective for TEX, and also copy input to a temporary for TEX
also add tex opcode names
|
|
makes glsl-fs-log2 and glsl1-integer division with uniform var pass
|
|
fixes piglit pointAtten and point-sprite tests
|
|
This still needs work, passes tex3d, fbo-scissor-bitmap, scissor-bitmap
|
|
h needs to be rounded up, this probably needs revisiting when we get
to tiling etc.
fixes fbo-generatemipmap-npot
|
|
8 more piglit tests pass,
fbo-clearmipmap, fbo-copyteximage, fbo-generatemipmap,
fbo-generatemipmap-nonsquare, fbo-generatemipmap-scissor,
fbo-generatemipmap-viewport, gen-teximage, gen-texsubimage
|
|
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
|
|
we were destroying the mm before unrefing all the objects, so segfault.
Signed-off-by: Dave Airlie <airlied@redhat.com>
|
|
Apart from the fact that the radeon.h/r600_states.h editing is a nightmare, this
wasn't so bad.
passes piglit user-clip test now also trivial tests.
Signed-off-by: Dave Airlie <airlied@redhat.com>
|
|
Reported-by: Niels Ole Salscheider <niels_ole@salscheider-online.de>
|
|
|
|
Signed-off-by: Marek Olšák <maraeo@gmail.com>
|
|
This changes r300_destroy_context() so it can be called
on a partially-initialized context, and uses it when
r300_create_context() hits a fatal error.
This makes sure r300_create_context() doesn't leak memory
or neglect to call r300_update_num_contexts() when it fails.
Signed-off-by: Marek Olšák <maraeo@gmail.com>
|
|
This fixes a potential bug if (has_hyperz) is false
(it would still init the atom as if has_hyperz were true).
Signed-off-by: Marek Olšák <maraeo@gmail.com>
|
|
|
|
|
|
Check for these and route them to a dedicated handler with one fewer
levels of recursive rasterization.
|
|
|
|
No need to calculate these values any longer, nor to store them in the
bin data. Improves isosurf a bit more, 115->123 fps.
|
|
For 16 and 64 pixel levels, calculate a mask which is linear in x and
y (ie not in the swizzle layout).
When iterating over full and partial masks, figure out position by
manipulating the bit number set in the mask, rather than relying on
postion arrays.
Similarly, calculate the lower-level c values from dcdx, dcdy and the
position rather than relying on the step array.
|
|
|
