The transfer staging texture is always freshly allocated, so for write-only
transfers we don't need to explicitly wait for the BO to become idle.
Squeezes a few hundered MB/s more out of x11perf -shmput500 with glamor.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This manifested as rendering failures or sometimes GPU hangs in
compositors when they accidentally got MSAA visuals due to a bug in the X
Server. Today we decided that the problem in compositors was equivalent
to a corruption bug we'd noticed recently in resizing MSAA-visual
glxgears, and debugging got a lot easier.
When we allocate our MCS MT, libdrm takes the size we request, aligns it
to Y tile size (blowing it up from 300x300=900000 bytes to 384*320=122880
bytes, 30 pages), then puts it into a power-of-two-sized BO (131072 bytes,
32 pages). Because it's Y tiled, we attach a 384-byte-stride fence to it.
When we memset by the BO size in Mesa, between bytes 122880 and 131072 the
data gets stored to the first 20 or so scanlines of each of the 3 tiled
pages in that row, even though only 2 of those pages were allocated by
libdrm. In the glxgears case, the missing 3rd page happened to
consistently be the static VBO that got mapped right after the first MCS
allocation, so corruption only appeared once window resize made us throw
out the old MCS and then allocate the same BO to back the new MCS.
Instead, just memset the amount of data we actually asked libdrm to
allocate for, which will be smaller (more efficient) and not overrun.
Thanks go to Kenneth for doing most of the hard debugging to eliminate a
lot of the search space for the bug.
Cc: "10.0 10.1" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77207
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We don't have any piglit tests for this currently.
v2: Use vec3s for the texcoords so it has some hope of working.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
You'll note from the previous commits that there's something of a loop
here: You call CTSI, which calls BlitFB, then if things go wrong that
falls back to CTSI. As a result, meta CTSI reaches over into blitfb to
tell it "no, don't try that fallback".
v2: Drop the _mesa_update_state(), which was only necessary due to use of
_mesa_clip_blit() in _mesa_meta_BlitFramebuffer() in another patch
series.
v3: Drop an _EXT suffix I copy-and-pasted.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v2)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I added support to bind_fbo_image in the process of building meta
CopyTexSubImage, and found that it broke generatemipmap because previously
we would just throw a GL error there and then end up with an incomplete
FBO and fallback.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I need to do the same code again for CopyTexSubImage().
v2: Drop incorrect, not-terribly-useful comment (review by Ken)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This avoids a ReadPixels() if there's accelerated CopyTexImage present.
It now requires GLSL as opposed to just fragment programs, but we don't
have any drivers that do ARB_fp but not GLSL.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
There shouldn't be anything special about copying out a subset of the src
rb to a temp before texturing from it, so just do it when we're figuring
out our src texture binding.
This drops Anuj's change to copy an extra border of 1 pixel around the src
area. I can't see how that change could be valid, and presumably if
there's some filtering problem at edges we just need to set the right
wrap mode.
v2: Don't fall back to swrast on non-2D/RECT/2D_MS textures when we can
still CopyTexSubImage. Fixes a segfault regression on i965 with
gl-3.2-layered-rendering-blit.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
Tested-by: Anuj Phogat <anuj.phogat@gmail.com>
I think we can assert that renderbuffer size is <= maximum 2D texture
size. Our source coordinates should have already been clipped to the src
renderbuffer size, but haven't actually (so we could potentially have
trouble if there's scaling, and we're in the CopyTexImage path that tries
to use src size). However, this texture size dependency was blocking the
next refactors, so I'm not sure if we want to go ahead with this series
before we get the clipping sorted out or not.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Putting NoDDClr and NoDDChk dependency control on instruction
sequences that include math opcodes can cause corruption of channels.
Treat math opcodes like send opcodes and suppress dependency hinting.
Signed-off-by: Mike Stroyan <mike@LunarG.com>
Tested-by: Tony Bertapelli <anthony.p.bertapelli@intel.com>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Matt Turner <mattst88@gmail.com>
total instructions in shared programs: 1653399 -> 1651790 (-0.10%)
instructions in affected programs: 92157 -> 90548 (-1.75%)
GAINED: 2
LOST: 2
Also significantly reduces the number of optimization loop iterations:
total loop iterations in shared programs: 39724 -> 31651 (-20.32%)
loop iterations in affected programs: 21617 -> 13544 (-37.35%)
Including some great pathological cases, like 29 -> 3 in Strike Suit
Zero and 24 -> 3 in Dota2.
Reviewed-by: Eric Anholt <eric@anholt.net>
We previously stopped searching for unread writes after encountering
control flow, but we can instead just search backwards until we hit
control flow.
instructions in affected programs: 22854 -> 22194 (-2.89%)
The generator uses its destination as a source implicitly, which breaks
some assumptions in dead code elimination. Giving the instruction a
source allows us to reason about it better.
We originally thought that GL 3.0 required GL_DEPTH_COMPONENT16 to map
exactly to Z16. However, we misread the specification, thanks in part
to LaTeX reordering the tables in the PDF.
Page 180 of the GL 3.0 specification (glspec30.20080923.pdf) says:
"[...] memory allocation per texture component is assigned by the GL to
match the allocations listed in tables 3.16-3.18 as closely as possible.
[...]
Required Texture Formats
[...]
In addition, implementations are required to support the following sized
internal formats. Requesting one of these internal formats for any
texture type will allocate exactly the internal component sizes and
types shown for that format in tables 3.16-3.17:"
Notably, however, GL_DEPTH_COMPONENT16 does /not/ appear in table 3.16
or table 3.17. It appears in table 3.18, where the "exact" rule doesn't
apply, and it falls back to the "closely as possible" rule.
The confusing part is that the ordering of the tables in the PDF is:
Table 3.16 (pages 182-184)
Table 3.18 (bottom of page 184 to top of 185)
Table 3.17 (page 185)
Presumably, people saw table 3.16, then saw the table immediately
following with DEPTH_COMPONENT* formats, and assumed it was 3.17.
Based on a patch by Chia-I Wu, but without the driconf option to force
Z16 to be used. It's not required, and there's apparently no benefit
to actually using it.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Chia-I Wu <olv@lunarg.com>
We've learned a few things since we originally disabled Z16; this attempts
to summarize the issue. I am no expert on this subject, though, so the
comment may not be totally accurate.
I did some benchmarking on GM45 and Ironlake, and discovered that for
GLBenchmark 2.7 EgyptHD, using Z16 was 3% slower on GM45 (n=15), and
4.5% slower on Ironlake (n=95). So, we can drop the "on Ivybridge"
aspect of the comment - it's always slower.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Chia-I Wu <olv@lunarg.com>
Significantly reduces BO allocation / destruction overhead for transfers,
e.g. measurable via x11perf -shm{ge,pu}t* with glamor.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Copied from similar code in gen8_vec4_generator.cpp.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This patch fixes gtest build errors on Mac OS X 10.9.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=73106
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
The rules are about to get a bit more complex to account for
gl_InstanceID and gl_VertexID, which are system values.
Extracting this first avoids introducing duplication.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
glClearBuffer() is currently clearing all active draw color buffers (all
buffers that have not been set to GL_NONE when calling glDrawBuffers) instead
of only clearing the one it receives as parameter. Altough brw_clear()
receives a bit mask indicating the color buffers that should be cleared,
this mask is ignored when calling brw_blorp_clear_color().
This was breaking the 'fbo-drawbuffers-none glClearBuffer' piglit test.
The patch provides the bit mask to brw_blorp_clear_color() so it can limit
clearing to the color buffers present in the mask.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=76832
Reviewed-by: Eric Anholt <eric@anholt.net>
This always looks crazy when I stumble across it, until I remember
what the hardware is doing. Describing it ought to short-circuit
that process next time :)
V2: Fix indents to 6 spaces, not 7.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Many shaders use a pattern such as:
for (int i = 0; i < NUM_LIGHTS; i++) {
...access a uniform array, or shader input/output array...
}
where NUM_LIGHTS is a small constant (such as 2, 4, or 8).
The expectation is that the compiler will unroll those loops, turning
the array access into constant indexing, which is more efficient, and
which may enable array splitting and other optimizations.
In many cases, our heuristic fails - either there's another tiny nested
loop inside, or the estimated number of instructions is just barely
beyond the threshold. So, we fail to unroll the loop, leaving the
variable indexing in place.
Drivers which don't support the particular flavor of variable indexing
will call lower_variable_index_to_cond_assign(), which generates piles
and piles of immensely inefficient code. We'd like to avoid generating
that.
This patch detects unsupported forms of variable-indexing in loops, where
the array index is a loop induction variable. In that case, it bypasses
the loop-too-large heuristic and forces unrolling.
Improves performance in various microbenchmarks: Gl32PSBump8 by 47%,
Gl32ShMapVsm by 80%, and Gl32ShMapPcf by 27%. No changes in shader-db.
v2: Check ir->array for being an array or matrix, rather than the
ir_dereference_array itself.
v3: Fix and expand statistics in commit message.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The "fail" flag is set if loop_unroll_count encounters a nested loop;
calling the flag "nested_loop" is a bit clearer.
The original reasoning was that count is inaccurate (too small) if there
are nested loops, as we don't do any sort of analysis on the inner loop.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Loop unrolling will need to know a few more options in the future.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Now that we pass in gl_shader_compiler_options, it makes sense to just
use options->MaxUnrollIterations, rather than passing a separate
parameter.
Half of the invocations already passed options->MaxUnrollIterations,
while the other half passed in a hardcoded value of 32.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
This will prevent the two from getting out of sync again.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
These were out of sync with the flags used to control
lower_variable_index_to_cond_assign in brw_shader.cpp.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Not setting this would prevented coalescing after a failed attempt if
the sources for both MOVs were the same.
total instructions in shared programs: 1654531 -> 1650224 (-0.26%)
instructions in affected programs: 423167 -> 418860 (-1.02%)
GAINED: 2
LOST: 0
Reviewed-by: Eric Anholt <eric@anholt.net>
It's just the array index, so we can just go look at the array and see
which element we are.
No significant performance difference (n=140)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Michel Dänzer
Matt Turner
Eric Anholt
Mike Stroyan
Kenneth Graunke
Jordan Justen
Chia-I Wu
Vinson Lee
Chris Forbes
Simone Scanzoni
Iago Toral Quiroga