We were allocating too much memory for linear layouts. The block_size
factor is already included in the row_stride and should not be used in
the img_stride calculation. This is typically a 4x savings!
If debug build, keep a linked list of all allocated resources (textures).
The llvmipe_print_resources() function can be called from a debugger to
print a list of all resources, their sizes, total size, etc.
When we have DRI2 protocol at least 2.3, we get an event from the
server when the back buffers get invalidated. When that's the case
let the driver know that it can rely on invalidate instead of the
glViewport polling.
The presence of this extension indicates to the DRI driver that the
loader will call invalidate in the __DRI2_FLUSH extension, whenever
the needs to query for new buffers. This means that the DRI driver
can drop the polling in glViewport().
Now that transfers are context operations it is the driver's
responsibility to ensure that transfers happen in order with all other
context operations, so flushes and finishes inside Mesa should be no
longer necessary. The attached patch implements that.
This should proportionate significant improvements for hardware drivers
which are able to stream transfers in the command buffers.
You can use the softpipe/llvmpipe_flush_resource() as reference
implementation of the worst case scenario, where the driver is not able
to streamline transfers. But the expectation is that driver
implementators will want to avoid flushing as much as possible.
Without this patch, any old intel_flush() call will cause a round trip to
the server and do a copy from fake to real front. We only actually
guarantee that frontbuffer results show up when glFlush() ia called, so
move the flushing to intel_glFlush().
We also need to flush fake to front before getting new buffers, but
we just handle that manually.
When we call intel_prepare_render() from intelReadPixels(), we'll mark
the front buffer dirty. That's silly, since we're only reading from it
and marking it dirty will cause us to copy from fake front to front
eventually.
Just clear the dirty flag after doing the read.
Now that we have intel_prepare_render() in place, we can use it to mark
the front buffer dirty if we're rendering to the front buffer once we
get there.
We only need this when the server may have swapped the buffers or
when we receive an invalidate event from the server. The default
behaviour is still that the DRI driver will invalidate its own buffers
when glViewport is called.
https://bugs.freedesktop.org/show_bug.cgi?id=27277
