Even if the pixmap's storage has alpha, it may have been uploaded with
garbage in the alpha channel, so we need to force the shader to set
alpha to 1. This was broken way back in
355334fcd9.
Fixes rendercheck -t composite -f x8r8g8b8.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
The copy optimization in d37329cba4
replicated a bug from last time we did a copy optimization: CopyArea
is only defined for matching depths. This is only a problem at 15 vs
16 depth right now (24 vs 32 would also have matching Render formats,
but they should work) but be strict in case we store other depths
differently in the future.
Fixes rendercheck -t blend -o src -f x4r4g4b4,x3r4g4b4
v2: Drop excessive src->depth == dst->depth check that snuck in.
v3: Switch back to src->depth == dst->depth
v4: Touch up commit message (s/bpp/depth).
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
v2: Fix "orignal" too (review feedback by ajax, change by anholt)_
Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Reviewed-by: Eric Anholt <eric@anholt.net>
Commit b64108fa ("glamor: Check for composite operations which are
equivalent to copies") failed to copy conditions from exaComposite which
ensure that the composite operation doesn't access outside of the source
picture.
This fixes rendercheck regressions from the commit above.
Reviewed-by: Keith Packard <keithp@keithp.com>
Patch b64108fa30 added a short cut that
identifies composite operations that can be performed with a simple
copy instead.
glamor_copy works in absolute coordinates, so the dx and dy values
passed in need to be converted from drawable-relative to absolute by
adding the drawable x/y values.
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
A1 and A8 pixmaps are usually stored in the Red channel to conform
with more recent GL versions. When using these pixmaps as mask values,
that works great. When using these pixmaps as source values, then the
value we want depends on what the destination looks like.
For RGBA or RGB destinations, then we want to use the Red channel
for A values and leave RGB all set to zero.
For A destinations, then we want to leave the R values in the Red
channel so that they end up in the Red channel of the output.
This patch adds a helper function, glamor_bind_texture, which performs
the glBindTexture call along with setting the swizzle parameter
correctly for the Red channel. The swizzle parameter for the Alpha
channel doesn't depend on the destination as it's safe to leave it
always swizzled from the Red channel.
This fixes incorrect rendering in firefox for this page:
https://gfycat.com/HoarseCheapAmericankestrel
while not breaking rendering for this page:
https://feedly.com
v2: Add change accidentally left in patch for missing
glDisable(GL_COLOR_LOGIC_OP).
Found by Emil Velikov <emil.l.velikov@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=63397
Signed-off-by: Keith Packard <keithp@keithp.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
glamor_make_current is supposed to be called before any GL APIs.
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
If the logic op gets left enabled, it overrides the blending
operation, causing incorrect contents on the display.
v2: Disable only on non-ES2, but disable even for PictOpSrc
v3: Found another place this is needed in
glamor_composite_set_shader_blend
v4: Remove change dependent on new glamor_set_composite_texture
API. This belongs in a different patch.
Found by Emil Velikov <emil.l.velikov@gmail.com>
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Increases x11perf -compwinwin500 numbers by a factor of 10 for me with
radeonsi.
Conditions copied from exaComposite().
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
RENDER requires that sampling outside of any source/mask picture results
in alpha == 0.0.
The OpenGL border colour cannot set alpha = 0.0 if the texture format
doesn't have an alpha channel, so we have to use the RepeatFix handling
in that case.
Also, only force alpha = 1.0 when sampling inside of RGBx source/mask
pictures.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94514
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
glamor_composite_choose_shader() may upload our scratch pixmaps to get
a Render operation completed. We don't want to hang onto GL memory
for our scratch pixmaps, since we'll just have to reallocate them at a
new w/h next time around, and the contents will be updated as well.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
glamor_upload_sub_pixmap_to_texture() only had the one caller, so we
can merge it in, fix its silly return value, and propagate a bunch of
constants.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
For hardware that doesn't do actual jumps for conditionals (i915,
current vc4 driver), this reduces the number of texture fetches
performed (assuming the driver isn't really smart about noticing that
the same sampler is used on each side of an if just with different
coordinates).
No performance difference on i965 with x11perf -magpixwin100 (n=40).
Improves -magpixwin100 by 12.9174% +/- 0.405272% (n=5) on vc4.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
All sorts of weird indentation, and some cuddled conditional
statements deep in the if tree.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
wh ratios are != 1.0 only when large, so with that we can simplify
down how we end up with RepeatFix being used.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This is a step toward using glamor_program.c for Render acceleration.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We can just hand in a constant mask and the driver will optimize away
the multiplication for us.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
GL_RED is supported by core profiles while GL_ALPHA is not; use GL_RED
for one channel objects (depth 1 to 8), and then swizzle them into the
alpha channel when used as a mask.
[airlied: updated to master, add swizzle to composited glyphs and xv paths]
v2: consolidate setting swizzle into the texture creation code, it
should work fine there. Handle swizzle when setting color as well.
v3: Fix drawing to a8 with Render (changes by anholt, reviewed by airlied).
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We only need it once at the top of the shader, so just put it
there.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
It's been on the list to add dual source blending support to avoid the
two pass componentAlpha code. Radeon has done this for a while in
EXA, so let's add support to bring glamor up to using it.
This adds dual blend to both render and composite glyphs paths.
Initial results show close to doubling of speed of x11perf -rgb10text.
v2: Fix breakage of all of CA acceleration for systems without
GL_ARB_blend_func_extended. Add CA support for all the ops we
support in non-CA mode when blend_func_extended is present. Clean
up some comments and formatting. (changes by anholt)
Signed-off-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
I originally inherited this from the EXA code, without determining
whether it was really needed. Regular composite should end up doing
the same thing, since it's all just shaders anyway. To the extent
that it doesn't, we should fix composite.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reading and writing to 16-depth pixmaps using PICT_x1r5g5b5 ends up
failing, unless you're doing a straight copy at the same bpp where the
misinterpretation matches on both sides.
Fixes rendercheck/blend/over and renderhceck/blend/src in piglit.
Please cherry-pick this to active stable branches.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
If the source/mask pixmap is a pixmap that doesn't have an FBO
attached, and it doesn't match the Render operation's size, then we'll
composite it to a CPU temporary (not sure why). We would take the
PictFormatShort from the source Picture, make a pixmap of that depth,
and try to look up the PictFormat description from the depth and the
PictFormatShort. However, the screen's PictFormats are only attached
to the screen's visuals' depths. So, with an x2r10g10b10 short format
(depth 30), we wouldn't find the screen's PictFormat for it
(associated with depth 32).
Instead of trying to look up from the screen, just use the pFormat
that came from our source picture. The only time we need to look up a
PictFormat when we're doing non-shader gradients, which we put in
a8r8g8b8.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
I'm amazed we've made it as far as we have without these checks: if
you made an unusual format picture that wasn't the normal a8r8g8b8 or
x8r8g8b8 or a8, we'd go ahead and try to render with it, ignoring that
the sampler would fetch totally wrong bits.
Fixes 260 tests in rendercheck -t blend -o src -f a8r8g8b8,x2r10g10b10
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
This gives the compiler a chance to optimize when the data is never
changed -- for example, with pict_format_combine_tab, the compiler
ends up inlining the 24 bytes of data into just 10 more bytes of code.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
We use this for all of our other performance-sensitive rendering, too.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
XRender defines this, GL really doesn't like it.
kwin 4.x and qt 4.x seem to make this happen for the
gradient in the titlebar, and on radeonsi/r600 hw
this draws all kinds of wrong.
v2: bump this up a level, and check it earlier.
(I assume the XXXX was for this case.)
v3: add same code to largepixmap paths (Keith)
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jasper St. Pierre <jstpierre@mecheye.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
New composite glyphs code uses the updated glamor program
infrastructure to create efficient shaders for drawing render text.
Glyphs are cached in two atlases (one 8-bit, one 32-bit) in a simple
linear fashion. When the atlas fills, it is discarded and a new one
constructed.
v2: Eric Anholt changed the non-GLSL 130 path to use quads instead of
two triangles for a significant performance improvement on hardware
with quads. Someone can fix the GLES quads emulation if they want to
make it faster there.
v3: Eric found more dead code to delete
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
By dropping the unconditional logic op disable at the end of
rendering, this fixes GL errors being thrown in GLES2 contexts (which
don't have logic ops). On desktop, this also means a little less
overhead per draw call from taking one less trip through the
glEnable/glDisable switch statement of doom in Mesa.
The exchange here is that we end up taking a trip through it in the
XV, Render, and gradient-generation paths. If the glEnable() is
actually costly, we should probably cache our logic op state in our
screen, since there's no way the GL could make that switch statement
as cheap as the caller caching it would be.
v2: Don't forget to set the logic op in Xephyr's drawing.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Keith Packard <keithp@keithp.com>
This will let us eliminate the pixmap types shortly
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Just embed the large elements in the regular pixmap private and
collapse the union to a single struct.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
There's no reason to waste memory storing redundant copies of the same
pointer all over the system; just pass in pointers as necessary to
each function.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
The core rendering code already requires that FBOs be allocated at
exactly the pixmap size so that tiling and stippling work
correctly. Remove the allocation support for that, along with the
render code.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
These were used by the non-standard glamor implementation in the intel
driver.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
It's unused since keithp's copy acceleration code completely replaced
glamor_copyarea.c and removed the blit path.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Keith Packard <keithp@keithp.com>
All users of glamor had the same value set, and it complicated things
for no reason.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Keith Packard <keithp@keithp.com>
To understand this patch, let's start at the protocol interface where
the relationship between the coordinate spaces is documented:
static Bool
_glamor_composite(CARD8 op,
PicturePtr source,
PicturePtr mask,
PicturePtr dest,
INT16 x_source,
INT16 y_source,
INT16 x_mask,
INT16 y_mask,
INT16 x_dest, INT16 y_dest,
CARD16 width, CARD16 height, Bool fallback)
The coordinates are passed to this function directly off the wire and
are all relative to their respective drawables. For Windows, this means
that they are relative to the upper left corner of the window, in
whatever pixmap that window is getting drawn to.
_glamor_composite calls miComputeCompositeRegion to construct a clipped
region to actually render to. In reality, miComputeCompositeRegion clips
only to the destination these days; source clip region based clipping
would have to respect the transform, which isn't really possible. The
returned region is relative to the screen in which dest lives; offset by
dest->drawable.x and dest->drawable.y.
What is important to realize here is that, because of clipping, the
composite region may not have the same position within the destination
drawable as x_dest, y_dest. The protocol coordinates now exist solely to
'pin' the three objects together.
extents->x1,y1 Screen origin of clipped operation
width,height Extents of the clipped operation
x_dest,y_dest Unclipped destination-relative operation coordinate
x_source,y_source Unclipped source-relative operation coordinate
x_mask,y_mask Unclipped mask-relative operation coordinate
One thing we want to know is what the offset is from the original
operation origin to the clipped origin
Destination drawable relative coordinates of the clipped operation:
x_dest_clipped = extents->x1 - dest->drawable.x
y_dest_clipped = extents->y1 - dest->drawable.y
Offset from the original operation origin:
x_off_clipped = x_dest_clipped - x_dest
y_off_clipped = y_dest_clipped - y_dest
Source drawable relative coordinates of the clipped operation:
x_source_clipped = x_source + x_off_clipped;
y_source_clipped = y_source + y_off_clipped;
Mask drawable relative coordinates of the clipped operation:
x_mask_clipped = x_source + x_off_clipped;
y_mask_clipped = y_source + y_off_clipped;
This is where the original code fails -- it doesn't subtract the
destination drawable location when computing the distance that the
operation has been moved by clipping. Here's what it does when
constructing a temporary source picture:
temp_src =
glamor_convert_gradient_picture(screen, source,
extent->x1 + x_source - x_dest,
extent->y1 + y_source - y_dest,
width, height);
...
x_temp_src = -extent->x1 + x_dest;
y_temp_src = -extent->y1 + y_dest;
glamor_convert_gradient_picture needs source drawable relative
coordinates, but that is not what it's getting; it's getting
screen-relative coordinates for the destination, adjusted by the
distance between the provided source and destination operation
coordinates. We want x_source_clipped and y_source_clipped:
x_source_clipped = x_source + x_off_clipped
= x_source + x_dest_clipped - x_dest
= x_source + extents->x1 - dest->drawable.x - x_dest
x_temp_src/y_temp_src are supposed to be the coordinates of the original
operation translated to the temporary picture:
x_temp_src = x_source - x_source_clipped;
y_temp_src = y_source - y_source_clipped;
Note that x_source_clipped/y_source_clipped will never be less than
x_source/y_source because all we're doing is clipping. This means that
x_temp_src/y_temp_src will always be non-positive; the original source
coordinate can never be strictly *inside* the temporary image or we
could have made the temporary image smaller.
x_temp_src = x_source - x_source_clipped
= x_source - (x_source + x_off_clipped)
= -x_off_clipped
= x_dest - x_dest_clipped
= x_dest - (extents->x1 - dest->drawable.x)
Again, this is off by the destination origin within the screen
coordinate space.
The code should look like:
temp_src =
glamor_convert_gradient_picture(screen, source,
extent->x1 + x_source - x_dest - dest->pDrawable->x,
extent->y1 + y_source - y_dest - dest->pDrawable->y,
width, height);
x_temp_src = -extent->x1 + x_dest + dest->pDrawable->x;
y_temp_src = -extent->y1 + y_dest + dest->pDrawable->y;
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Markus Wick <markus@selfnet.de>
This reverts commit 4e9aabb6fc.
It broke kwin decorations with XRender compositing.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Keith Packard <keithp@keithp.com>
Paints with textures, using a temporary buffer for overlapping copies
Performs CPU to GPU transfers for pixmaps in memory. Accelerates copy
plane when both objects are in the GPU. Includes copy_window
acceleration too.
v2: Use NV_texture_barrier for non-overlapping copies within the same
drawable
v3: Switch to glamor_make_current
v4: Do overlap check on the bounding box of the region rather than
on individual boxes
v5: Use Eric Anholt's re-written comments which provide a more accurate
description of the code
v6: Use floating point uniform for copy plane bit multiplier. This
avoids an int to float conversion in the copy plane fragment shader.
Use round() instead of adding 0.5 in copy plane. round() and +0.5
end up generating equivalent code, and performance measurements
confirm that they are the same speed. Round() is a bit clearer
though, so we'll use it.
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Markus Wick <markus@selfnet.de>
These offer a simpler and more efficient means for temporarily
transitioning to CPU-accessible memory for fallback implementations.
v2: Do not attempt fallbacks with GLAMOR_DRM_ONLY pixmaps
glamor cannot transfer pixels for GLAMOR_DRM_ONLY pixmaps using
glReadPixels and glTexSubImage2D, and so there's no way to perform
fallback operations with these pixmaps.
v3: Clear ->pbo field when deleting the PBO. Otherwise, we'd reuse
the old name next time we fall back on the pixmap, which would
potentially conflict with some other pixmap that genned a new
name, or just do a lazy allocation of the name (compat GL context,
like we currently use) or error out (core GL context, like we hope
to use some day). Also, style fixes. Changes by anholt, acked by
keithp.
Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
There were actually two issues with the original code I believe, the
first is that the call to glamor_convert_gradient_picture wasn't
properly referencing the coordinates of the source/mask pictures. The
second, was that the updated references (x_temp/y_temp) were also
improperly set, they should always be 0 because the temp pictures are
new ones that start at (0, 0). The reason it worked in certain cases
and it didn't in others (notably the tray icons) was due to the
numbers working out based on the call to glamor_composite. In the
cases that it did work extent->x1 would equal x_dest and extent->y1
would equal y_dest, making it so what was actually passed into
glamor_convert_gradient_picture and the settings for x_temp/y_temp
were correct. However, for the case when extent->x1 wouldn't equal
x_dest and extent->y1 wouldn't equal y_dest (for example with the tray
icons) then the wrong parameters get passed into
glamor_convert_gradient_picture and x_temp/y_temp are set improperly.
Fixes issues with tray icons not appearing properly in certain cases.
Bug:
https://bugs.freedesktop.org/show_bug.cgi?id=64738
Signed-Off-by: Anthony Waters <awaters1@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
