| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
| |
It's much better to just skip the draw call entirely. Getting this
information out of register allocation will also be useful for
implementing threaded fragment shaders, which will need to retry
non-threaded if RA fails.
Cc: <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 4d019bd703e7c20d56d5b858577607115b4926a3)
|
|
|
|
|
|
| |
We would assertion fail in setting up the simulator the second time
around. This at least postpones the assertion failure until we've closed
all of the first set of screens and started opening a new set.
|
|
|
|
| |
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Track rendering to each FBO independently and flush rendering only when
necessary. This lets us avoid the overhead of storing and loading the
frame when an application momentarily switches to rendering to some other
texture in order to continue rendering the main scene.
Improves glmark -b desktop:effect=shadow:windows=4 by 27%
Improves glmark -b
desktop:blur-radius=5:effect=blur:passes=1:separable=true:windows=4
by 17%
While I haven't tested other apps, this should help X rendering a lot, and
I've heard GLBenchmark needed it too.
|
|
|
|
|
| |
This is a preparation step for having multiple jobs being queued up at the
same time.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The clear of Z or stencil will end up clearing the other as well, instead
of masking. There's no way around this that I know of, so if we are
clearing just one then we need to draw a quad.
Fixes a regression in the job-shuffling code, where the clear values move
to the job and don't just have the last clear's value laying around when
you do glClear(DEPTH) and then glClear(STENCIL) separately
(ext_framebuffer_multisample-clear 4 depth)).
This causes regressions in ext_framebuffer_multisample/multisample-blit
depth and ext_framebuffer_multisample/no-color depth, but these were
formerly false positives due to the reference image also being black. Now
the reference and test images are both being drawn, and it looks like
there's an incorrect resolve of depth during blitting to an MSAA FBO.
|
| |
|
|
|
|
|
|
|
| |
We need the source to be in r0-r3, so make a new register class for it.
It will be up to the surrounding passes to make sure that the r0-r3
allocation of its source won't conflict with anything other class
requirements on that temp.
|
|
|
|
|
|
|
|
|
|
|
|
| |
We don't want to bake the whole array into the FS key, because of the
hashing overhead. But we can keep a set of the arrays seen, and use a
pointer to the copy in as the array's proxy.
Between this and the previous patch, gl-1.0-blend-func now passes on
hardware, where previously it was filling the 256MB CMA area with shaders
and OOMing.
Drops 712 shaders from shader-db.
|
|
|
|
|
| |
We don't tell the hardware whether we're computing depth, so we need
to manage early Z state manually. Fixes piglit early-z.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
To support general GL_TEXTURE_BASE_LEVEL we have to copy to a temporary
miptree. However, if a single level is being selected, we can use the
existing miptree and force all the sampling to be from that particular
level.
This avoids a ton of software fallbacks in glGenerateMipmaps(), which uses
base levels in the blit implementation in gallium. Improves "glmark2 -b
terrain" from 2 fps to 3 (perhaps some more precision would be useful?),
and cuts its CPU usage during the benchmarking from ~30% to ~10% (total
CPU time from 8.8s to 7.6s).
|
|
|
|
|
|
| |
The compiler uses the per-stage flags already, so it didn't need this.
vc4_uniforms was using it, so just replace it with both of the stage flags
for now.
|
|
|
|
| |
We use a big VC4_DIRTY_FRAGTEX/VC4_DIRTY_VERTEX on the stage, instead.
|
|
|
|
| |
This came from deriving from freedreno.
|
|
|
|
|
|
| |
Fixes piglit mixed-immediate-and-vbo, and may significantly improve
performance of applications that store a 4-byte IB in the same VBO as
vertex data.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We can't do this all the time, because you want blending to be done in
linear space, and sRGB would lose too much precision being done in 4x8.
The win on instructions is pretty huge when you can, though.
total uniforms in shared programs: 32065 -> 32168 (0.32%)
uniforms in affected programs: 327 -> 430 (31.50%)
total instructions in shared programs: 92644 -> 89830 (-3.04%)
instructions in affected programs: 15580 -> 12766 (-18.06%)
Improves openarena performance at 1920x1080 from 10.7fps to 11.2fps.
|
|
|
|
|
|
| |
I haven't proven that this happens (I've got other GPU hangs in the
way), but the closed driver also does this and it's documented as an
errata.
|
|
|
|
|
|
|
|
| |
Similar to 9ffc1049ca (freedreno/ir3: use nir two-sided-color lowering).
No piglit regression.
Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
| |
(originally part of previous patch, split out to separate patch by Rob)
v2: squash in some fixes from Eric
v3: Another fix from Eric for point coords.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
|
|
|
|
|
|
|
|
| |
This allows creating compute-only and debug contexts.
Reviewed-by: Brian Paul <brianp@vmware.com>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
|
|
|
|
|
|
|
|
|
|
| |
I switched us to tracking whether the results *could* go to r4, but then
didn't make a separate register class for the class bits that included r4.
Switch the "any" class to actually be "any", and name the "any but r4"
class more appropriately.
total instructions in shared programs: 96798 -> 94680 (-2.19%)
instructions in affected programs: 62736 -> 60618 (-3.38%)
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, SFU values always moved to a temporary, and TLB color reads
and texture reads always lived in r4. Instead, we can have these results
just be normal temporaries, and the register allocator can leave the
values in r4 when they don't interfere with anything else using r4.
shader-db results:
total instructions in shared programs: 100809 -> 100040 (-0.76%)
instructions in affected programs: 42383 -> 41614 (-1.81%)
|
|
|
|
|
|
|
|
| |
It's a bunch of work for us to emit it (and its uniforms), more work for
the kernel to validate it, and additional work for the CLE to read
it. Improves es2gears framerate by about 50%.
Signed-off-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
| |
Cuts another 12% of vc4_uniforms.o, in exchange for computing it at
CSO creation time.
|
|
|
|
|
| |
In exchange for a bit of space and computation in CSO setup, we cut
vc4_uniform.c (draw time) code size by 4.8%.
|
|
|
|
|
|
|
| |
This avoids a security issue where userspace could have written the tile
state/tile alloc behind the GPU's back, and will apparently be necessary
for fixing stability bugs (tile state buffers are missing some top bits
for the tile alloc's address).
|
|
|
|
|
| |
There weren't that many variations of RCL generation, and this lets us
skip all the in-kernel validation for what we generated.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The idea I had when I wrote the original shadow code was that you'd see a
set_index_buffer to the IB, then a bunch of draws out of it. What's
actually happening in openarena is that set_index_buffer occurs at every
draw, so we end up making a new shadow BO every time, and converting more
of the BO than is actually used in the draw.
While I could maybe come up with a better caching scheme, for now just
do the simple thing that doesn't result in a new shadow IB allocation
per draw.
Improves performance of isosurf in drawelements mode by 58.7967% +/-
3.86152% (n=8).
|
| |
|
|
|
|
|
| |
There will be other blit code showing up, and it seems like the place
you'd look.
|
|
|
|
|
|
| |
I want to be able to have multiple jobs being set up at the same time (for
example, a render job to do a little fixup blit in the course of doing a
render to the main FBO).
|
|
|
|
|
|
| |
I'd like to compile as much of the device-specific code as possible
when building for simulator, and using if (using_simulator) instead of
ifdefs helps.
|
|
|
|
|
|
|
|
| |
No need to recheck the FS compile when the VS source has changed, but
there *is* a need to recheck the VS compile when the compiled VS has
changed (since the live inputs may change).
Fixes es3conform's blend test.
|
|
|
|
|
|
| |
This will give the compiler the chance to dead-code eliminate unused VPM
reads. This is particularly a big deal in the CS where a bunch of vattrs
are just not going to be used.
|
|
|
|
|
| |
This gives a 2.7x improvement in x11perf -rect100, since we only end up
load/storing the x11perf window, not the whole screen.
|
| |
|
|
|
|
|
| |
No performance difference on a microbenchmark with norast that should hit it
enough to have mattered, n=220.
|
|
|
|
|
| |
Our submits now return immediately and you have to manually wait for
things to complete if you want to (like a normal driver).
|
|
|
|
|
| |
Fixes 14 ARB_vp tests (which had no lowering done), and should improve
performance of indirect uniform array access in GLSL.
|
|
|
|
| |
I'm going to want to make some other decisions here before flushing.
|
|
|
|
|
|
| |
I'm going to be using VC4_DEBUG=shaderdb,norast to do shaderdb stats, but
when debugging regressions, I want to match shaderdb output to shader
disassembly.
|
|
|
|
| |
Fixes about 15 piglit tests about interpolation and clipping.
|
|
|
|
|
|
|
|
|
| |
If the VS doesn't output a value that the FS needs, we still need to read
the right contents for the remaining FS inputs, by emitting padding. And
if the VS outputs something the FS doesn't need, we shouldn't put it in
the VPM at all (so the code producing it can get DCEed).
Fixes 77 piglit tests.
|
|
|
|
|
| |
Improves simulated norast performance on a little benchmark by 38.0965%
+/- 3.27534% (n=11).
|
|
|
|
|
|
|
|
|
|
|
| |
Merging VS and CS into the same struct wasn't winning us anything except
for not allocating a separate BO (but if we want to pack programs into
BOs, we should pack not just those 2 programs together). What it was
getting us was a bunch of code duplication about hash table lookups and
propagating vc4_compile contents into a vc4_compiled_shader.
I was about to make the situation worse with indirect uniform buffer
access.
|
|
|
|
|
|
|
|
|
|
| |
It's fairly easy, thanks to Rob Clark's lowering code. Fixes
two-sided-lighting and 4 vertex-program-two-side testcases, while
regressing 8 testcases that involve enabling two-sided color while only
initializing one of the two colors in the VS. If you're enabling two
sided color, it's of course expected that you really do set up both
colors, so this is still an improvement (and when we set up a linker for
TGSI, we'll hopefully fix those 8 fails).
|
|
|
|
|
|
| |
We have to expose them for GL 2.0, but we just always return a value of 0.
We should be advertising 0 query bits instead of 64, but gallium doesn't
have plumbing for that yet. At least this stops the segfaults.
|
|
|
|
|
| |
Setting the bit without setting the offset values is kind of useless.
Fixes piglit polygon-offset (but not polygon-mode-offset).
|
|
|
|
|
|
|
| |
This is just the GL 1.1 flat shading of colors -- we don't need to support
TGSI constant interpolation bits, because we don't do GLSL 1.30.
Fixes 7 piglit tests.
|