SNB has 64k urb space, we only use piece of them.
The more urb space we alloc,
the more concurrent vs threads we can run.
push the urb space usage to the limit.
Signed-off-by: Zou Nan hai <nanhai.zou@intel.com>
this fixes fbo-generatemipmap-formats rgtc and s3tc in NPOT mode
with softpipe.
r600g fails to even get level 0 correct so have to look into that
a bit further.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This was always converting to 8-bit per channel unsigned formats,
which isn't suitable for RGTC signed formats, this special cases
those two formats and converts to floats for those.
Signed-off-by: Dave Airlie <airlied@redhat.com>
SNORM needs a bit of work in the state tracker in order for mipmap
generation to work I believe.
I'm also not sure that having unorm fetches for an snorm format is
sane.
With signed types we weren't hitting this test however the comment
stating this doesn't happen often doesn't apply when using signed
types since an all 0 block is quite common which isn't abs min or max.
this fixes the limits correctly again also.
Signed-off-by: Dave Airlie <airlied@redhat.com>
if the values are all in the last dword, the high bits can be 0,
This fixes a valgrind warning I saw when playing with mipmaps.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Drivers can call this function as needed. It tells the VBO module to
always unmap the current glBegin/glEnd VBO when we flush. Otherwise
it's possible to be in a flushed state but still have the VBO mapped.
Among other benefits, parallel makes now work. Since many people have
parallel builds by default (via MAKEFLAGS environment variable), this
sames some irritation at release time...when there's usually not any
other irritation already.
In my last commit I introduced a build dependency upon a new libdrm.
Add the associated autoconf checks. As the headers are part of the core
libdrm, we need to bump that version and so may as well bump the chipset
specific versions simultaneously.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With relaxed relocation checking in the kernel, we can specify a
negative delta (i.e. pointing outside of the target bo) in order to fake
a range in a large buffer. We only then need to upload the elements used
and adjust the buffer offset such that they correspond with the indices
used in the DrawArrays.
(Depends on libdrm 0209428b3918c4336018da9293cdcbf7f8fedfb6)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
... and take advantage of start_vertex_bias to trim to [min_index,
max_index] where possible (i.e. when we need to upload all arrays).
Fixes half_float_vertex(misc.fillmode.wireframe)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=34595
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When doing copy swapbuffers using drm, throttle on outstanding copy operations.
Introduces a new environment variable, EGL_THROTTLE_FENCES that the
user can use to indicate the desired number of outstanding swapbuffers, or
disable throttling using EGL_THROTTLE_FENCES=0.
This can and perhaps should be extended to the pageflip case as well, since
with some hardware pageflips can be pipelined. In case the pageflip syncs, the
throttle operation will be a no-op anyway.
Update copyright notices.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Use the pageflip ioctl when available.
Otherwise, or when the backbuffer contents need to be preserved,
fall back to a copy operation.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The copy swap can be used when we need to preserve the contents of
the back buffer or when there is no way to do native page-flipping.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
This makes it usable also for native helpers.
Also add inline functions to access the context and to
uninit the native display structure.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Vinson Lee