This is a rewrite of vc4_opt_qpu_schedule.c to operate on QIR. Texture
fetch can probably take as much as the rest of the cycles of the program,
so it's important to hide our other cycles during it (which is hard to do
after register allocation). Also, we can queue up multiple texture
requests before collecting the resulting samples, so that we keep the
texture unit busy more of the time.
High-settings openarena performance +2.35849% +/- 0.221154% (n=7). Also
about 2-3% on the multiarb demo. 8 piglit tests
(ext_framebuffer_multisample accuracy depthstencil) go from failing in
rendering to failing in register allocation, but hopefully I can fix that
up with some better register pressure handling here.
total instructions in shared programs: 87723 -> 88448 (0.83%)
instructions in affected programs: 78411 -> 79136 (0.92%)
total estimated cycles in shared programs: 276583 -> 246306 (-10.95%)
estimated cycles in affected programs: 265691 -> 235414 (-11.40%)
There's only high latency between a complete texture fetch setup and
collecting its result, not between each step of setting up the texture
fetch request.
So vertex shader input attributes are handled different than internal
varyings between shader stages, dvec3 and dvec4 only count as
one slot for vertex attributes, but for internal varyings, they
count as 2.
This patch comments all the uses of this API to clarify what we
pass in, except one which needs further investigation
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
The old function didn't work for matrices, and we need this
in other places to fix some other problems, so move to a helper
in glsl type and fix the one user so far.
A dual slot double is one that has 3 or 4 components in it's
base type.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Don't use a bool here, as for some 64-bit fixes we need
the stage.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Android builds with -Werror=pointer-to-int-cast causing an error on 32-bit
builds.
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Rob Herring <robh@kernel.org>
Signed-off-by: Rob Clark <robclark@freedesktop.org>
As in the previous patches, these can be implemented as
any(v) -> any_nequal(v, false)
all(v) -> all_equal(v, true)
and their removal simplifies the code in the next patch.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The GLSL IR to TGSI/Mesa IR paths for any_nequal have the same
optimizations the ir_unop_any paths had.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Unify the cleanup paths of the function rather than duplicating code.
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
NaNs mean it should be clipped, otherwise the NaNs might get passed to the
next stages (if clipping didn't happen for another reason already), which
might cause all kind of problems.
The llvm path got this right already (possibly by luck), but this isn't used
when there's a gs active.
Found by code inspection, verified with some hacked piglit test and some more
hacked debug output.
(Note the clipper can still itself incorrectly generate NaN and INF position
values in its output prims (at least after w divide / viewport transform) even
if the inputs weren't NaNs, if the position data of the vertices is
"sufficiently bad".)
Reviewed-by: Brian Paul <brianp@vmware.com>
Those stages only really work for OGL-style texturing (so number of samplers
and views mostly the same, certainly for the max values).
These get often set up all at once, thus there might be max number of both
even if all of them are just NULL. We must not set the max number of samplers
and views to the same value since that will lead to terrible things if a driver
supports more views than samplers (and the state tracker set up all the views).
(This will not make these stages magically work if a shader uses dx10-style
texturing, they might still replace an actually used sview in that case.)
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch modifies the SSE4.1 test in configure.ac to use a global
variable to initialize vector variables. In addition, we now return the
value of the computation instead of 0.
This is done so gcc 4.9 (and lower) won't optimize the SSE4.1 assembly
instructions (when using -O1 and higher), because then the configure test
might incorrectly pass even though the assembler doesn't support the
SSE4.1 instructions (the test will pass because the compiler does support the intrinsics).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91806
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Check for a 'python2.7' binary, 'python' and 'python2' are not
provided by the OpenBSD python 2.7.x packages.
Signed-off-by: Jonathan Gray <jsg@jsg.id.au>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Use PKG_CHECK_MODULES to get the flags to link libelf
v2: keep AC_CHECK_LIB as a fallback for elfutils provided
libelf that doesn't install a pkg-config file.
Signed-off-by: Jonathan Gray <jsg@jsg.id.au>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
OpenGLES 3.1 cannot be enabled for gen 7 (Ivy Bridge, Haswell) since
they are still missing ARB_stencil_texturing.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Marta Lofstedt <marta.lofstedt@intel.com>
Previously we were checking the desktop OpenGL ARB_compute_shader
requirements, but for OpenGLES 3.1, the requirements are lower.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Marta Lofstedt <marta.lofstedt@intel.com>
The OpenGL ARB_compute_shader extension specfication requires at least
1024 for GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, whereas OpenGLES 3.1
only required 128.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Previous OpenGLES 3.1 testing had been done when ARB_compute_shader
was overridden to enabled.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Marta Lofstedt <marta.lofstedt@intel.com>
I apparently regressed this when rewriting the built-ins using
ir_builder, in 76d2f73643.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93387
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This fixes a memory leak introduced in 6a9c151
("nv50: add compute-related MP perf counters on G84+")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1" <mesa-stable@lists.freedesktop.org>
The destroy_query() helper was actually never called. This fixes
a memory leak while monitoring performance metrics.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1" <mesa-stable@lists.freedesktop.org>
This fixes a long time ago memory leak (even before all my query
related changes).
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
The maximum number of resident warps per multiprocessor is 64 on
Kepler instead of 48 on Fermi.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
It's nonsense to drain the pipeline like this.
v2: keep the drain for DMA-buf exports.
v3: flush before the export and after compositing and add TODO comment.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Julien Isorce <j.isorce@samsung.com>
Tested-by: Julien Isorce <j.isorce@samsung.com>
This reverts commit 839793680f.
The patch was breaking DRI3 because driGLFormatToImageFormat does not
handle MESA_FORMAT_B8G8R8X8_SRGB which ended up making it fail to
create the renderbuffer and it would later crash. It's not trivial to
add this format because there is no __DRI_IMAGE_FORMAT nor
__DRI_IMAGE_FOURCC define for the format either. I'm not sure how
difficult adding this would be and whether adding a new format would
require some sort of new version for DRI. Seeing as this might take a
while to fix I think it makes sense to just revert the patch in the
meantime in order to avoid regressing master.
It is also not handled in intel_gles3_srgb_workaround and there may be
other cases where it breaks.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93388
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
If EGL_KHR_surfaceless_context is used then glViewport can be called
with NULL for the draw and read surfaces. This was previously causing
a crash because the i965 driver tries to use this point to invalidate
the surfaces and it was derferencing the NULL pointer.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93257
Cc: Nanley Chery <nanley.g.chery@intel.com>
Cc: "11.1" <mesa-stable@lists.freedesktop.org>
Tested-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Previously the GL spec required that whenever glBlitFramebuffer is
used with either buffer being multisampled, the internal formats must
match. However the GL 4.4 spec was later changed to remove this
restriction. In the section entitled “Changes in the released
Specification of July 22, 2013” it says:
“Relax BlitFramebuffer in section 18.3.1 so that format conversion can
take place during multisample blits, since drivers already allow this
and some apps depend on it.”
If most drivers already allowed this in earlier versions I think it's
safe to assume that this is a spec bug and it should also be allowed
in all versions.
This patch just removes the restriction on desktop GL. For GLES there
are conformance tests that assert the previous behaviour so it is
probably safer to leave it in.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92706
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
"image" is not ready yet since it will be set at
the end of the function by: *image = *img;
Signed-off-by: Julien Isorce <j.isorce@samsung.com>
Reviewed-by: Christian K<C3><B6>nig <christian.koenig@amd.com>
We just ignored them altogether. While this feature is rather old-fashioned
supporting it is actually rather trivial.
This fixes the associated piglit tests (2 gl-1.0-edgeflag, 2 gl-2.0-edgeflag
and all (7) of point-vertex-id).
v2: comment fixes, and make the use of the edgeflag in clipmask consistent
with when it's actually there (should be impossible to hit a case where the
difference would actually matter but still...)
Reviewed-by: Brian Paul <brianp@vmware.com>
This just adds confusion, these parameters are used when fetching vertices
by translate, but certainly not when emitting hw vertices for drivers, they
make no sense there (setting them has no consequences otherwise since there
won't be any elements with instance_divisor set). So just set them to 0 (the
draw_pipe_vbuf code for emitting vertices when the draw pipeline is run
already does exactly that).
Also while here do some whitespace cleanup.
Reviewed-by: Brian Paul <brianp@vmware.com>
Now that we have a helper in the builder for system values and a helper in
core NIR to get the intrinsic opcode, there's really no point in having
things split out into a helper function. This commit "modernizes" this
pass to use helpers better and look more like newer passes.
Reviewed-by: Eric Anholt <eric@anholt.net>
The VC4_DEBUG=cl,qpu is nice and all, but I want to be able to get more
detailed dumps, and to replay the same exact commands in simulation. For
that I need a dump with all of the VBOs, shaders, shader recs, etc. This
dump can be parsed by vc4-gpu-tools.
For now this is only doable from simulator mode, because otherwise we
don't have access to the RCL contents generated by the kernel.
