Once upon a time this was used for creating fake vertex shader
attribute semantics for d3d8 shaders. We don't need this anymore since
device_stream_info_from_declaration() will use the vertex
declaration's output slot to load the data, if present. That also
avoids the potentially expensive matching of attribute semantics
between vertex shader and declaration for d3d8.
This moves the GLSL and ARB specific reserved constants out of directx.c into
the get_caps methods of the shader backends. That way the number of reserved
constants remains in the backend.
GL_LIMITS({v/p}shader_constantsF) now contains the real number of constants as
advertised by GL instead of some mixture of GL info and backend implementation
specifics. This makes it easier for backends to decide how many constants to
use.
DrawPrimitive now accepts the number of vertices, not the number of
primitives. This code was forgotten in the patch that changed this.
We're drawing triangles, so we are drawing num_primitives * 3 vertices.
In most cases we're fine with the vbo and glMapBuffer and never use the actual
heap memory copy. Try to stick to just the vbo copy and avoid allocating the
extra heap memory. In case it is needed(emulation or vertex conversion), fall
back to the old double buffering mode.
This fixes a regression introduced by d7c7c76850
and exposed by ef2d704183. Unfortunately having
EXT_vertex_array_bgra support masks the problem, which is why the regression
tests didn't catch it.
Internally we want to store some extra data, like ffp_valid. One of the later
patches also stores the format desc in the declaration elements, instead of
the current WINED3DDECLTYPE.
Add a new wined3d-internal PreLoad function to textures and surfaces
that takes a parameter specifying wether the rgb or srgb texture
should be loaded.
This reduces the number of srgb switching reloads quite a lot. The only
situation in which a reload is needed is if the rgb copy is modified on the GL
side and the srgb copy is needed.
Other than being a bit nicer than passing function pointers all over the
place, this helps dxgi/d3d10. While the swapchain itself is created in dxgi,
its surfaces are constructed in d3d10core, which makes it impractical for dxgi
to pass the appropriate function pointers.
Doing the levels calculation before the object is created prevents leaking
resources and avoids changing baseTexture.levels after it's been initialized
by basetexture_init().
startIdx should be the first index to draw, either from the vertex
array or the index array, depending on if the draw is indexed or
not. Having both at the same time wouldn't make sense.
Most callers work on a stateblock rather than a device, and the main fields
we check (vertexShader and pixelShader) are part of the stateblock as well.
Based on a patch by Stefan Dösinger. This is more flexible, and allows
the shader backend implementation to be simpler, since it doesn't have
to know about specific formats. The next patch makes use of this.
Note that minMipLookup and magLookup aren't particularly safe to use,
they're global arrays initialized from IWineD3DImpl_FillGLCaps(). The same
goes for the other global dynamic lookup tables.
A number of considerations contribute to this:
1) The shader backend knows best which shader(s) it needs. GLSL needs
both, arb only one
2) The shader backend may pass some parameters to the compilation
code(e.g. which pixel format fixup to use)
3) The structures used in (2) are different in vs and ps, so a
baseshader::Compile won't work
4) The structures in (2) are wined3d-private structures, so
having a public method in the vtable won't work(its a bad idea
anyway).
This prevents the target from changing during the first PreLoad() call
on a surface, which would be inconvenient when attaching a surface to
a FBO for example.
This gives a small performance improvement for applications that are
smart enough to set the D3DPRESENTFLAG_DISCARD_DEPTHSTENCIL flag, or
to create depth stencils with Discard set to TRUE.
Since those surfaces are stored in blocks, the 4 pixel step doesn't only apply to surfaces < 4, but
also to surfaces bigger than that, with a non-multiple-of-4 size.
Clear all attachments before deleting FBOs. It should be valid to
delete FBOs that still have attachments, but for some reason the
nvidia drivers don't like it. The resulting memory corruption can be
pretty nasty, and this workaround seems clean enough.
This is the prefered format of many codecs, and for some codecs this
is the only supported output format. As usual I try to handle all the
conversion in the GPU and keep the CPU involvement minimal to gain the
full performance of PBO transfers.
Although sharing FBOs across contexts is allowed by EXT_framebuffer_object
(issue 76), it causes issues with nVidia drivers. Considering the GL 3 spec
explicitly disallows sharing of FBOs accross contexts (Appendix D), this
patch is probably the right thing to do.
This is a long-needed cleanup aimed at removing the ddraw_primary,
ddraw_window, ddraw_width and ddraw_height members from
IWineD3DDeviceImpl, which just do not belong there. Destination
window and screen handling is supposed to be done by swapchains.
This is an ATI specific format designed for compressed normal maps,
and quite a few games check for its existence. While it is an
ATI-specific "extension" in d3d9, it is a core part of
D3D10(DXGI_FORMAT_BC5), and supported on Geforce 8 cards.
This happened to work because most cards have the same amount of
pshader and vshader constants, but for some reason this doesn't hold
true on this macbook pro here, which lead to a crash due to heap
corruption
This is cleaner than the if statements in the code. Also np2 textures
should in theory support linear filtering, but fglrx doesn't seem to
like it. This needs further investigation. So far we've never used
linear filtering on np2 textures, so there should not be a
regression. Furthermore I think shader support is more important than
filtering, since NP2 textures are mostly used for 1:1 copying to the
screen.
ATI cards prior to the radeon HD series did not have unconditional non
power of two support. So far we've used texture_rectangle for that, or
created a bigger power of two texture with padding. This had the
disadvantage that we had to correct the coordinates, which causes
extreme problems with shaders(doesn't work, pretty much).
Both the MacOS and the fglrx driver have support for
GL_ARB_texture_non_power_of_two, and run it on the hardware as long as
we stay within the texture_rectangle limitations. This allows us to
have conditional non power of two textures with normalized
coordinates. This patch adds an internal extension, and the code
creates a regular GL_TEXTURE_2D texture with NP2 size, but refuses
mipmapping, filtering and texture_rectangle incompatible
operations. This makes np2 textures work with shaders on fglrx and
macos.
Since atifs is only doing the fragment pipeline replacement right now
there is no need for the shader backend structure any longer. The ffp
private data is stored in new fragment pipeline private data(which
could potentially be set to equal the shader private data if needed).
Destroying the stateblock potentially references the shader backend.
If the stateblock has active shaders when it is released, the shader's
destructor will tell the shader backend to destroy the corresponding
resources. This was exposed by my patch that moved the glsl program
lookup table into the backend's private data.
The idea of this patchset is to split the monolithic state set into 3
parts, vertex processing, fragment processing and other states(depth,
stencil, scissor, ...). The states will be provided in templates which
can be (mostly) independently combined, and are merged into a single
state table at device creation time. This way we retain the advantages
of the single state table and having the advantage of separated
pipeline implementations which can be combined without any manually
written glue code.
Constant numbers start at 0, and the loading loop has a for(i; i <
dirtyconsts; i++). This means that the highest dirty constant isn't
loaded correctly. Rather than replacing the < with <=, which would
make it impossible to have no dirty constant, add 1 to the dirty
constant counter.
This gets rid of depth_copy_state in the device, and instead tracks
the most up to date location per-surface. This makes things a lot
easier to follow, and allows us to make a copy when switching depth
stencils in SetDepthStencilSurface().
This makes the depth copy independent of the currently attached render
targets. This is important for the next patch because it might do a
depth copy when the render targets aren't in a valid configuration
(SetDepthStencilSurface()).
The idea is to make setting depth attachments a bit more consistent
with set_render_target_fbo()/attach_surface_fbo(). I've also got an
upcoming patch in my tree that needs this.
Henri Verbeet