mesa.git
8 years agoradeonsi: increase MAX_CONTROL_FLOW_DEPTH AKA MaxIfDepth
Marek Olšák [Fri, 4 Nov 2016 11:31:53 +0000 (12:31 +0100)]
radeonsi: increase MAX_CONTROL_FLOW_DEPTH AKA MaxIfDepth

we don't want to lower deep IFs unconditionally

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoRevert "configure.ac: honour LLVM_LIBDIR when linking against LLVM"
Emil Velikov [Thu, 10 Nov 2016 15:10:34 +0000 (15:10 +0000)]
Revert "configure.ac: honour LLVM_LIBDIR when linking against LLVM"

This reverts commit a39ad185932eab4f25a0cb2b112c10d8700ef242.

The commit aims to address "missing" -L/foo/bar during linking stage. At
the same time it doesn't add the -L and yet the LLVM_LDFLAGS [which
provide -L/foo/bar] are already used throughout.

Seems like something pretty unique (broken?) on my end. Since the commit
introduces issues (due to the missing -L) revert until we get to the
root of it (PEBKAC or a genuine issue).

8 years agoradeonsi: fix/silence unused variable warnings in optimized builds
Nicolai Hähnle [Fri, 4 Nov 2016 09:39:45 +0000 (10:39 +0100)]
radeonsi: fix/silence unused variable warnings in optimized builds

I'm leaving num_out_sgpr around since it's not in a fast path, and besides
the compiler should be able to optimize it away easily. The alternative
with #if/#endif would be extremely ugly.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agogallivm: fix [IU]MUL_HI regression harder
Nicolai Hähnle [Wed, 9 Nov 2016 11:44:47 +0000 (12:44 +0100)]
gallivm: fix [IU]MUL_HI regression harder

The fix in commit 88f791db75e9f065bac8134e0937e1b76600aa36 was insufficient
for radeonsi because the vector case was not handled properly. It seems
piglit only covers the scalar case, unfortunately.

Fixes GL45-CTS.shader_bitfield_operation.[iu]mulExtended.*

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoRevert "wayland: Block for the frame callback in get_back_bo not dri2_swap_buffers"
Daniel Stone [Wed, 1 Jun 2016 08:59:06 +0000 (09:59 +0100)]
Revert "wayland: Block for the frame callback in get_back_bo not dri2_swap_buffers"

This reverts commit 25cc889004aad6d1cab9edd76db898658e347b97, though
since the code has changed, it was applied manually.

The intent of moving blocking from SwapBuffers to get_back_bo, was to
avoid unnecessary triple-buffering by ensuring that the compositor had
fully processed the previous frame before we started rendering. This
means that the only time we would have to resort to triple-buffering
would be when the buffer is directly scanned out, thus saving an extra
buffer for composition anyway.

The 'repaint window' changes introduced in Weston since then, however,
have narrowed the window of time between the frame event being sent and
the repaint loop needing to conclude, to 7ms by default, in order to
reduce latency. This means however that blocking in get_back_bo gives a
maximum of 7ms for the entire GL submission to begin and complete.

Not only this, but if a client is using buffer_age to avoid full
repaints, the buffer-age request will stall in get_back_bo until the
frame callback completes, meaning that the client cannot even calculate
the repaint area before the 7ms window.

The combination of the two meant that WebKit-GTK+ was failing to
achieve full framerate on a Minnowboard, due to spending a great deal of
its time attempting to query the age of the next buffer before redraw.

Revert to the previous behaviour of allowing rendering to begin but
delaying SwapBuffers, unless and until we can find a more gentle
behaviour.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Jonas Ådahl <jadahl@gmail.com>
Reviewed-by: Derek Foreman <derekf@osg.samsung.com>
Tested-by: Derek Foreman <derekf@osg.samsung.com>
Cc: Kristian Høgsberg <krh@bitplanet.net>
8 years agoglsl: validate output blocks against input blocks
Iago Toral Quiroga [Fri, 4 Nov 2016 12:37:58 +0000 (13:37 +0100)]
glsl: validate output blocks against input blocks

Until now were validating in/out blocks by listing the inputs in the
consumer stage and then, for each output of the producer, we checked that
it was a match if it was consumed. This method does not catch the case
where the consumer has an input that is not present as an output in the
producer stage, because it only generates link errors for outputs present
in the producer stage that don't match the inputs in the consumer stage.
The current method does catch the case were an output from the producer
stage is not consumed, which is irrelevant and is ignored.

By reversing the way we do this, we can detect this situation, so this
patch lists the outputs of the producer stage and then validates inputs
of the consumer stage against them. If we see an input in the consumer
for which there is no associated output in the producer, we produce a
link error.

The only exception to this is the special built-in input block gl_in[],
since this is implicitly generated for geometry and tessellation stages,
but we don't generate it if the producer stage does not write to any of
the pre-defined outputs (for example, if the vertex shader does not write
to gl_Position, etc). Since writing to these is not mandatory, do not
produce a link error in that case. There is a CTS tessellation test
(GL45-CTS.tessellation_shader.program_object_properties) that has an
empty vertex shader (so it does not produce gl_in[]) and would fail to
link if we don't do this.

This fixes the following dEQP test:
dEQP-GLES31.functional.shaders.linkage.io_block.missing_output_block

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98245
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradv: fixup botched llvm API changes.
Dave Airlie [Thu, 10 Nov 2016 04:12:32 +0000 (14:12 +1000)]
radv: fixup botched llvm API changes.

Reported-by: Jan Vesely <jan.vesely@rutgers.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoac/nir/llvm: adopt to new LLVM attribute API.
Dave Airlie [Thu, 10 Nov 2016 03:04:54 +0000 (13:04 +1000)]
ac/nir/llvm: adopt to new LLVM attribute API.

Ported from corresponding changes to gallivm.

tested build against 3.9 and master.

Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agovulkan/wsi/wayland: Clean up some error handling paths
Jason Ekstrand [Wed, 9 Nov 2016 18:21:03 +0000 (10:21 -0800)]
vulkan/wsi/wayland: Clean up some error handling paths

This gets rid of all the memory leaks reported by the WSI CTS tests.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agovulkan/wsi/wayland: Include pthread.h
Jason Ekstrand [Wed, 9 Nov 2016 18:20:31 +0000 (10:20 -0800)]
vulkan/wsi/wayland: Include pthread.h

We use pthreads and, for some reason, it wasn't getting included

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/device: Implicitly unmap memory objects in FreeMemory
Jason Ekstrand [Tue, 8 Nov 2016 01:25:07 +0000 (17:25 -0800)]
anv/device: Implicitly unmap memory objects in FreeMemory

From the Vulkan spec version 1.0.32 docs for vkFreeMemory:

   "If a memory object is mapped at the time it is freed, it is implicitly
   unmapped."

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Cc: "12.0 13.0" <mesa-dev@lists.freedesktop.org>
8 years agoanv/device: Return the right error for failed maps
Jason Ekstrand [Tue, 8 Nov 2016 01:24:24 +0000 (17:24 -0800)]
anv/device: Return the right error for failed maps

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Cc: "12.0 13.0" <mesa-dev@lists.freedesktop.org>
8 years agoanv/device: Add some asserts to MapMemory
Jason Ekstrand [Tue, 8 Nov 2016 01:23:44 +0000 (17:23 -0800)]
anv/device: Add some asserts to MapMemory

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoanv: Rework fences
Jason Ekstrand [Wed, 2 Nov 2016 16:11:11 +0000 (09:11 -0700)]
anv: Rework fences

Our previous fence implementation was very simple.  Fences had two states:
signaled and unsignaled.  However, this didn't properly handle all of the
edge-cases that we need to handle.  In order to handle the case where the
client calls vkGetFenceStatus on a fence that has not yet been submitted
via vkQueueSubmit, we need a three-status system.  In order to handle the
case where the client calls vkWaitForFences on fences which have not yet
been submitted, we need more complex logic and a condition variable.  It's
rather annoying but, so long as the client doesn't do that, we should still
hit the fast path and use i915_gem_wait to do all our waiting.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/wsi: Set the fence to signaled in AcquireNextImageKHR
Jason Ekstrand [Wed, 2 Nov 2016 21:45:37 +0000 (14:45 -0700)]
anv/wsi: Set the fence to signaled in AcquireNextImageKHR

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/gen8: Stall when needed in Cmd(Set|Reset)Event
Jason Ekstrand [Sat, 8 Oct 2016 06:02:25 +0000 (23:02 -0700)]
anv/gen8: Stall when needed in Cmd(Set|Reset)Event

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoglsl: record number of components used in each slot for varying packing
Ilia Mirkin [Sat, 5 Nov 2016 13:31:21 +0000 (09:31 -0400)]
glsl: record number of components used in each slot for varying packing

Instead of packing varyings into vec4's, keep track of how many
components each slot uses and create varyings with matching types. This
ensures that we don't end up using more components than the orginal
shader, which is especially important for geometry shader output limits.

This comes up for NVIDIA hw, where the limit is 1024 output components
for a GS, and the hardware complains *loudly* if you even think about
going over.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
8 years agoglsl: fix slot_end calculations and simplify reserved_slots check
Ilia Mirkin [Sat, 5 Nov 2016 01:01:18 +0000 (21:01 -0400)]
glsl: fix slot_end calculations and simplify reserved_slots check

The previous code was confused about whether slot_end was inclusive or
exclusive. Make it so that it is inclusive consistently, and use it for
setting the new location. This also avoids discrepancies in how
num_components is calculated vs the more manual approach taken for the
former reserved_slots check.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
8 years agoswr: correct setting of independentAlphaBlendEnable
Ilia Mirkin [Wed, 9 Nov 2016 19:31:25 +0000 (14:31 -0500)]
swr: correct setting of independentAlphaBlendEnable

This setting is for whether color and alpha have different blend
settings, not for whether blending is enabled on a per-RT basis.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: [rasterizer] add a .dir-locals.el to support 4-space indents
Ilia Mirkin [Wed, 9 Nov 2016 17:07:42 +0000 (12:07 -0500)]
swr: [rasterizer] add a .dir-locals.el to support 4-space indents

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: set halfz rasterizer setting
Ilia Mirkin [Wed, 9 Nov 2016 06:19:34 +0000 (01:19 -0500)]
swr: set halfz rasterizer setting

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: [rasterizer core] allow an OpenGL driver to specify halfz clipping
Ilia Mirkin [Wed, 9 Nov 2016 06:18:05 +0000 (01:18 -0500)]
swr: [rasterizer core] allow an OpenGL driver to specify halfz clipping

With ARB_clip_control, GL may also do 0..1 depth clipping, not just
-1..1. This removes clip's reliance on driver type. DX users will need
to be updated to set the new clipHalfZ flag to get proper clipping
functionality.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: fix support for inverted depth scales
Ilia Mirkin [Wed, 9 Nov 2016 06:16:32 +0000 (01:16 -0500)]
swr: fix support for inverted depth scales

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: [rasterizer jitter] fix logic op to work with unorm/snorm
Ilia Mirkin [Tue, 8 Nov 2016 00:12:23 +0000 (19:12 -0500)]
swr: [rasterizer jitter] fix logic op to work with unorm/snorm

Most logic op usage is probably going to end up with normalized
textures. Scale the floating point values and convert to integer before
performing the logic operations.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agovc4: Clamp the shadow comparison value.
Eric Anholt [Mon, 7 Nov 2016 20:25:47 +0000 (12:25 -0800)]
vc4: Clamp the shadow comparison value.

Fixes piglit glsl-fs-shadow2D-clamp-z.

Cc: <mesa-stable@lists.freedesktop.org>
8 years agovc4: Don't pair up TLB scoreboard locking instructions early in QPU sched.
Eric Anholt [Mon, 7 Nov 2016 18:52:32 +0000 (10:52 -0800)]
vc4: Don't pair up TLB scoreboard locking instructions early in QPU sched.

Jonas Pfeil noticed that we were putting passthrough tlb_z writes early in
the shader, despite QIR and QPU scheduling both trying to delay scoreboard
locking for as long as possible.

The problem was that when trying to pair up QPU instructions, at some
point the passthrough tlb_z would be the last one available and it would
get paired, even if the other half would open up other instructions to be
scheduled and we could have paired tlb_z with something later in the
program.  Also, since passthrough z is just a mov, it pairs up really
easily.

The proper fix would probably be to flip the order of scheduling
instructions so we went from bottom to top (also relevant for branch delay
slot scheduling).

However, we can do a quick fix here to just not schedule a TLB lock until
there's nothing but TLB left in the program, at a slight instruction cost
(est .61% cycle count in shader-db) but a major fragment shader
parallelism win.

glmark2 results:
  texture:texture-filter=linear: +1.24481% +/- 0.626117% (n=15)
  bump:bump-render=height: 1.24991% +/- 0.154793% (n=136,133 -- screensaver
    outliers removed)

8 years agovc4: Print a reg pressure estimate in our reg allocation failure dump.
Eric Anholt [Mon, 7 Nov 2016 18:16:45 +0000 (10:16 -0800)]
vc4: Print a reg pressure estimate in our reg allocation failure dump.

8 years agovc4: Don't abort when a shader compile fails.
Eric Anholt [Mon, 7 Nov 2016 17:05:59 +0000 (09:05 -0800)]
vc4: Don't abort when a shader compile fails.

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>
8 years agomesa: Fix pixel shader scratch space allocation on Gen9+ platforms.
Kenneth Graunke [Tue, 8 Nov 2016 01:12:54 +0000 (17:12 -0800)]
mesa: Fix pixel shader scratch space allocation on Gen9+ platforms.

We had missed a bit of errata - PS scratch needs to be computed as if
there were 4 subslices per slice, rather than 3.

                          Skylake      Broxton        Kabylake
                      GT1 GT2 GT3 GT4  2x6 3x6  GT1 GT1.5 GT2 GT3 GT4
Actual Slices          1   1   2   3    1   1    1    1    1   2   3
Total Subslices        3   3   6   9    2   3    2    3    3   6   9
Subsl. for PS Scratch  4   4   8   12   4   4    4    4    4   8   12

Note that Skylake GT1-3 already worked because we allocated 64 * 9
(trying to use a value that would work on GT4, with 9 subslices),
and the actual required values were 64 * 4 or 64 * 8.  However, all
others (Skylake GT4, Broxton, and Kabylake GT1-4) underallocated,
which can lead to scratch writes trashing random process memory,
and rendering corruption or GPU hangs.

Fixes GPU hangs and rendering corruption on Skylake GT4 in shaders that
spill.  Particularly, dEQP-GLES31.functional.ubo.all_per_block_buffers.*
now runs successfully with no hangs and renders correctly.  This may
fix problems on Broxton and Kabylake as well.

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agomesa/extensions: expose OES_vertex_half_float for ES2
Kevin Strasser [Mon, 10 Oct 2016 21:29:58 +0000 (14:29 -0700)]
mesa/extensions: expose OES_vertex_half_float for ES2

Half float support already exists for desktop GL. Reuse the
ARB_half_float_vertex enable bit and account for the different enum to
enable the extension for ES2.

Signed-off-by: Kevin Strasser <kevin.strasser@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoRevert "egl: remove explicit config_id management from dri2_add_config()"
Emil Velikov [Wed, 9 Nov 2016 21:45:07 +0000 (21:45 +0000)]
Revert "egl: remove explicit config_id management from dri2_add_config()"

This reverts commit 3652d1d5942a857f225700d67ce2c900396982f2.

Self nack/reject on this one. The base.ConfigID is overwritten
immediately after we store the current value, thus one memcpy [further
down] the wrong value will be copied.

8 years agoutil: add MSVC HAS_TRIVIAL_DESTRUCTOR implementation
Brian Paul [Wed, 9 Nov 2016 21:48:45 +0000 (14:48 -0700)]
util: add MSVC HAS_TRIVIAL_DESTRUCTOR implementation

Based on a patch by George Kyriazis but changed to test for
_MSC_VER >= 1800 (Visual Studio 2015).

This fixes the failed CANARY assertion in src/util/ralloc.c:get_header()
on Windows.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98595
Tested-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
8 years agoradv: Suffix the radeon_icd file with the host CPU
Emil Velikov [Wed, 9 Nov 2016 18:11:27 +0000 (18:11 +0000)]
radv: Suffix the radeon_icd file with the host CPU

Port of the anv commit d96345de989 ("anv: Suffix the intel_icd file with
the host CPU").

v2: s/intel_icd/radeon_icd/ in commit summary (Gražvydas)

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com> (IRC)
8 years agoradv: use correct .specVersion for extensions
Emil Velikov [Wed, 9 Nov 2016 18:10:47 +0000 (18:10 +0000)]
radv: use correct .specVersion for extensions

Analogous to previous commit.

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Dave Airlie <airlied@redhat.com> (IRC)
8 years agoanv: use correct .specVersion for extensions
Emil Velikov [Wed, 9 Nov 2016 18:10:46 +0000 (18:10 +0000)]
anv: use correct .specVersion for extensions

Vulkan has introduced the consept of .specVersion which can be used to
attribute changes of the said extension.

The current loader does not check the value, thus it have gone unnoticed
that the driver exposes an old version of the following extensions:

VK_KHR_xcb_surface        (Rev 6)
VK_KHR_xlib_surface       (Rev 6)
VK_KHR_wayland_surface    (Rev 5)
- Updated the surface create function to take a pCreateInfo structure

VK_KHR_swapchain          (Rev 68)
- Moved the "validity" include for vkAcquireNextImage to be in its proper
  place, after the prototype and list of parameters.
...

According to the documentation:

  * pname:specVersion is the version of this extension.
    It is an integer, incremented with backward compatible changes.

Based on the history of vk.xml the above (latest) revision has been
available since Vulkan 1.0 so even if they were any backwards
incompatible change(s) [as hinted by the revision log] those should be
safe.

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoamd/addrlib: limit fastcall/regparm to GCC i386
Emil Velikov [Wed, 2 Nov 2016 15:32:22 +0000 (15:32 +0000)]
amd/addrlib: limit fastcall/regparm to GCC i386

The use of regparm causes an error on arm/arm64 builds with clang.
fastcall is allowed, but still throws a warning. As both options only
have effect on 32-bit x86 builds, limit them to that case.

v2: keep the __i386__ within GCC (Nicolai)

Cc: 13.0 <mesa-stable@lists.freedesktop.org>
Cc: Rob Herring <robh@kernel.org>
Cc: Nicolai Hähnle <nhaehnle@gmail.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Rob Herring <robh@kernel.org>
8 years agoconfigure.ac: honour LLVM_LIBDIR when linking against LLVM
Emil Velikov [Mon, 24 Oct 2016 17:11:33 +0000 (18:11 +0100)]
configure.ac: honour LLVM_LIBDIR when linking against LLVM

Currently if one uses a non-default prefix, the path won't get
propagated and we'll fail at link-time.

A very quick and easy example is to install to /usr/local.
At this point, llvm-config will be picked even without the
--with-llvm-prefix, but regardless of the latter linking will fail.

Currently people can workaround that via LD_LIBRARY_PATH.

Cc: "12.0 13.0" <mesa-stable@lists.freedesktop.org>
Cc: Tom Stellard <thomas.stellard@amd.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoegl: remove explicit config_id management from dri2_add_config()
Emil Velikov [Fri, 30 Sep 2016 12:45:55 +0000 (13:45 +0100)]
egl: remove explicit config_id management from dri2_add_config()

Currently we only saved the id to memcpy the whole _EGLConfig to write
back the exact same id value.

Remove the unneeded and confusing/misleading code.

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
8 years agolinker: Accurately track gl_uniform_block::stageref
Ian Romanick [Mon, 7 Nov 2016 23:59:35 +0000 (15:59 -0800)]
linker: Accurately track gl_uniform_block::stageref

As the linked per-stage shaders are processed, mark any block that has a
field that is accessed as referenced.  When combining all the linked
shaders, combine the per-stage stageref masks.

This fixes a number of GLES CTS tests:

    ES31-CTS.core.geometry_shader.program_resource.program_resource
    ES32-CTS.core.geometry_shader.program_resource.program_resource
    ESEXT-CTS.geometry_shader.program_resource.program_resource
    piglit.gl45-cts.geometry_shader.program_resource.program_resource

However, it makes quite a few more fail:

    ES31-CTS.functional.program_interface_query.buffer_variable.random.6
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.compute.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.separable_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_fragment_only_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_geo_fragment_only_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_geo_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_fragment_only_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_geo_fragment_only_fragment.unnamed_block.float
    ES31-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_geo_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.random.6
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.compute.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.separable_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_fragment_only_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_geo_fragment_only_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_geo_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_fragment_only_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_geo_fragment_only_fragment.unnamed_block.float
    ES32-CTS.functional.program_interface_query.buffer_variable.referenced_by.vertex_tess_geo_fragment.unnamed_block.float

I have diagnosed the failures, but I'm not sure whether we or the
tests are wrong.  After optimizations are applied, all of the tests
are of the form:

    buffer X {
        float f;
    } x;

    void main()
    {
        x.f = x.f;
    }

The test then queries that x is referenced by that shader stage.  We
eliminate the assignment of x.f to itself, and that removes the last
reference to x.  We report that x is not referenced, and the test fails.
I do not know whether or not we are allowed to eliminate that assignment
of x.f to itself.

After discussions with the OpenGL ES group in Khronos, we believe that
Mesa's behavior is correct.  I will provide patches to the CTS tests
to Khronos.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agolinker: Slight code rearrange to prevent duplication in the next commit
Ian Romanick [Mon, 7 Nov 2016 23:54:46 +0000 (15:54 -0800)]
linker: Slight code rearrange to prevent duplication in the next commit

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agolinker: Trivial coding standards fixes
Ian Romanick [Fri, 4 Nov 2016 19:36:08 +0000 (12:36 -0700)]
linker: Trivial coding standards fixes

v2: Revert the unreachable to assert in
parcel_out_uniform_storage::visit_field.  Suggested by Ilia.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoglsl: Add some comments to methods of ir_variable_refcount_visitor
Ian Romanick [Mon, 7 Nov 2016 22:24:39 +0000 (14:24 -0800)]
glsl: Add some comments to methods of ir_variable_refcount_visitor

It was not obvious from the just the .h file what the hash table
contained.  It was also not obvious that get_variable_entry would create
a new entry in the hash table.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agollvmpipe: Fix build after removal of deprecated attribute API v2
Aaron Watry [Tue, 8 Nov 2016 03:55:14 +0000 (21:55 -0600)]
llvmpipe: Fix build after removal of deprecated attribute API v2

Applies on top of v3 of Tom's gallivm change.

v2:
  - Tom Stellard: Use enums instread of strings.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Aaron Watry <awatry@gmail.com>
CC: Tom Stellard <thomas.stellard@amd.com>
CC: Jan Vesely <jan.vesely@rutgers.edu>
8 years agogallivm: Fix build after removal of deprecated attribute API v3
Tom Stellard [Mon, 7 Nov 2016 18:35:09 +0000 (18:35 +0000)]
gallivm: Fix build after removal of deprecated attribute API v3

v2:
  Fix adding parameter attributes with LLVM < 4.0.

v3:
  Fix typo.
  Fix parameter index.
  Add a gallivm enum for function attributes.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradv: fix GetFenceStatus for signaled fences
Dave Airlie [Wed, 9 Nov 2016 01:21:30 +0000 (01:21 +0000)]
radv: fix GetFenceStatus for signaled fences

if a fence is created pre-signaled we should return that
in GetFenceStatus even if it hasn't been submitted.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Gustaw Smolarczyk <wielkiegie@gmail.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoradv: enable conditional discard optimisation on radv.
Dave Airlie [Wed, 2 Nov 2016 01:23:11 +0000 (01:23 +0000)]
radv: enable conditional discard optimisation on radv.

This fixes a bunch of GPU hangs introduced in some CTS
tests like
dEQP-VK.memory.pipeline_barrier.host_write_uniform_buffer.65536

It works around an issue seen in the LLVM backend, but
also makes the radv code work more like the radeonsi stack.

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agonir: add conditional discard optimisation (v4)
Dave Airlie [Wed, 2 Nov 2016 01:22:07 +0000 (01:22 +0000)]
nir: add conditional discard optimisation (v4)

This is ported from GLSL and converts

if (cond)
discard;

into
discard_if(cond);

This removes a block, but also is needed by radv
to workaround a bug in the LLVM backend.

v2: handle if (a) discard_if(b) (nha)
cleanup and drop pointless loop (Matt)
make sure there are no dependent phis (Eric)
v3: make sure only one instruction in the then block.
v4: remove sneaky tabs, add cursor init (Eric)

Reviewed-by: Eric Anholt <eric@anholt.net>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoac/nir: add support for discard_if intrinsic (v2)
Dave Airlie [Wed, 2 Nov 2016 01:21:15 +0000 (01:21 +0000)]
ac/nir: add support for discard_if intrinsic (v2)

We are going to start lowering to this in NIR code,
so prepare radv for it.

v2: handle conversion to kilp properly (nha)

Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoanv: Do relocations in userspace before execbuf ioctl
Kristian Høgsberg Kristensen [Tue, 8 Mar 2016 23:31:47 +0000 (15:31 -0800)]
anv: Do relocations in userspace before execbuf ioctl

Since our surface state buffer is shared by all batches, the kernel does a
full stall and sync with the CPU between batches every time we call
execbuf2 because it refuses to do relocations on an active buffer.  Doing
them in userspace and passing the NO_RELOC flag to the kernel allows us to
perform the relocations without stalling.

This improves the performance of Dota 2 by around 30% on a Sky Lake GT2.

v2 (Jason Ekstrand):
 - Better comments (Chris Wilson)
 - Fixed write_reloc for correct canonical form (Chris Wilson)

v3 (Jason Ekstrand):
 - Skip relocations which aren't needed
 - Provide an environment variable to always use the kernel
 - More comments about correctness (Chris Wilson)

v4 (Jason Ekstrand):
 - More comments (Chris Wilson)

v5 (Jason Ekstrand):
 - Rebase on top of moving execbuf2 setup go QueueSubmit

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Move relocation handling from EndCommandBuffer to QueueSubmit
Jason Ekstrand [Sun, 6 Nov 2016 02:47:33 +0000 (19:47 -0700)]
anv: Move relocation handling from EndCommandBuffer to QueueSubmit

Ever since the early days of the Vulkan driver, we've been setting up the
lists of relocations at EndCommandBuffer time.  The idea behind this was to
move some of the CPU load out of QueueSubmit which the client is required
to lock around and into command buffer building which could be done in
parallel.  Then QueueSubmit basically just becomes a bunch of execbuf2
calls.

Technically, this works.  However, when you start to do more in QueueSubmit
than just execbuf2, you start to run into problems.  In particular, if a
block pool is resized between EndCommandBuffer and QueueSubmit, the list of
anv_bo's and the execbuf2 object list can get out of sync.  This can cause
problems if, for instance, you wanted to do relocations in userspace.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/batch: Move last_ss_pool_bo_offset to the command buffer
Jason Ekstrand [Mon, 7 Nov 2016 16:07:43 +0000 (08:07 -0800)]
anv/batch: Move last_ss_pool_bo_offset to the command buffer

The original reason for putting it in the batch_bo was to allow primaries
to share it across secondaries or something like that.  However, the
relocation lists in secondary command buffers are are always left alone and
copied into the primary command buffer's relocation list.  This means that
the offset really applies at the command buffer level and putting it in the
batch_bo doesn't make sense.  This fixes a couple of potential bugs around
re-submission of command buffers that are not likely to be hit but are bugs
none the less.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Add an anv_execbuf helper struct
Jason Ekstrand [Sun, 6 Nov 2016 02:01:44 +0000 (19:01 -0700)]
anv: Add an anv_execbuf helper struct

This commit adds a little helper struct for storing everything we use to
build an execbuf2 call.  Since the add_bo function really has nothing to do
with a command buffer, it makes sense to break it out a bit.  This also
reduces some of the churn in the next commit.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/batch_chain: Improve write_reloc
Jason Ekstrand [Wed, 2 Nov 2016 17:42:45 +0000 (10:42 -0700)]
anv/batch_chain: Improve write_reloc

The old version wasn't properly handling large addresses where we have to
sign-extend to get it into the "canonical form" expected by the hardware.
Also, the new version is capable of doing a clflush of the newly written
reloc if requested.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Initialize anv_bo::offset to -1
Jason Ekstrand [Tue, 1 Nov 2016 03:25:08 +0000 (20:25 -0700)]
anv: Initialize anv_bo::offset to -1

Since -1 is an invalid GPU address, this lets us know whether or not we
have a valid address for a buffer.  We don't get a valid address until the
first time that buffer is used in an execbuf2 ioctl.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/allocator: Simplify anv_scratch_pool
Jason Ekstrand [Tue, 1 Nov 2016 20:10:11 +0000 (13:10 -0700)]
anv/allocator: Simplify anv_scratch_pool

The previous implementation was being overly clever and using the
anv_bo::size field as its mutex.  Scratch pool allocations don't happen
often, will happen at most a fixed number of times, and never happen in the
critical path (they only happen in shader compilation).  We can make this
much simpler by just using the device mutex.  This also means that we can
start using anv_bo_init_new directly on the bo and avoid setting fields
one-at-a-time.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Add a new bo_pool_init helper
Jason Ekstrand [Tue, 1 Nov 2016 20:09:36 +0000 (13:09 -0700)]
anv: Add a new bo_pool_init helper

This ensures that we're always setting all of the fields in anv_bo

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Don't presume to know what address is in a surface relocation
Jason Ekstrand [Tue, 1 Nov 2016 14:21:00 +0000 (07:21 -0700)]
anv: Don't presume to know what address is in a surface relocation

Because our relocation processing happens at EndCommandBuffer time and
because RENDER_SURFACE_STATE objects may be shared by batches, we really
have no clue whatsoever what address is actually written to the relocation
offset in the BO.  We need to stop making such claims to the kernel and
just let it relocate for us.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Add a cmd_buffer_execbuf helper
Jason Ekstrand [Wed, 2 Nov 2016 17:33:54 +0000 (10:33 -0700)]
anv: Add a cmd_buffer_execbuf helper

This puts the actual execbuf2 call in anv_batch_chain.c along with the
other relocation stuff.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/device: Add an execbuf wrapper
Jason Ekstrand [Tue, 1 Nov 2016 03:36:26 +0000 (20:36 -0700)]
anv/device: Add an execbuf wrapper

This wrapper ensures that we always update all anv_bo::offset fields based
on the offsets returned by the kernel.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv: Make anv_finishme only warn once per call-site
Jason Ekstrand [Wed, 9 Nov 2016 04:43:09 +0000 (20:43 -0800)]
anv: Make anv_finishme only warn once per call-site

When you fire up Dota2 on Haswell you get spammed with thousands of
"Implement Gen7 HZ ops" finishme's.  The point of anv_finishme is to act as
a reminder that there is something left to implement.  Printing it once
should be sufficient.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoi965/compute: Allow ARB_compute_shader in compat profile
Jordan Justen [Thu, 3 Nov 2016 22:22:11 +0000 (15:22 -0700)]
i965/compute: Allow ARB_compute_shader in compat profile

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97447
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Tested-by: Evan Odabashian <eodabash@gmail.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoRevert "draw: use vectorized calculations for fetch"
Roland Scheidegger [Wed, 9 Nov 2016 04:46:12 +0000 (05:46 +0100)]
Revert "draw: use vectorized calculations for fetch"

Trivial. There's some regressions internally, related to overflow
behavior. I'll have to look at it at another time, some interactions
with vsplit/vcache are actually mind-blowing.

This reverts commit 3fa10ffb496cc4e6d1003891cf0381bb5bec2a74.

8 years agoswr: disable logic op when the rt format is float or srgb
Ilia Mirkin [Tue, 8 Nov 2016 22:30:03 +0000 (17:30 -0500)]
swr: disable logic op when the rt format is float or srgb

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: fix AND_INVERTED logic op conversion
Ilia Mirkin [Tue, 8 Nov 2016 00:18:49 +0000 (19:18 -0500)]
swr: fix AND_INVERTED logic op conversion

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: add support for EXT_depth_bounds_test
Ilia Mirkin [Tue, 1 Nov 2016 20:45:13 +0000 (16:45 -0400)]
swr: add support for EXT_depth_bounds_test

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoswr: [rasterizer core] set depth hottile when depth bounds test enabled
Ilia Mirkin [Tue, 1 Nov 2016 20:45:12 +0000 (16:45 -0400)]
swr: [rasterizer core] set depth hottile when depth bounds test enabled

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
8 years agoi965: Fix GPU hang related to multiple render targets and alpha testing
Anuj Phogat [Mon, 24 Oct 2016 23:03:00 +0000 (16:03 -0700)]
i965: Fix GPU hang related to multiple render targets and alpha testing

This patch should have been the part of commit e592f7df.
In a situation when there are multiple render targets with alpha testing
enabled, if fragment shader doesn't write to draw buffer zero, it causes
the GPU hang on SKL. No GPU hang is seen on HSW. Simulator gives a
warning for all gen6+ h/w:
"Illegal render target write message length 0xa expected 0xc"

This patch fixes the GPU hang as well as the simulator warning with
new piglit test fbo-mrt-alphatest-no-buffer-zero-write:
https://patchwork.freedesktop.org/patch/118212

No regressions in Jenkins CI system.

Cc: "12.0 13.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agoswr: allow alphatest without blend or logicop
Tim Rowley [Fri, 4 Nov 2016 18:10:56 +0000 (13:10 -0500)]
swr: allow alphatest without blend or logicop

We need to compile a blend function when alphatest is enabled.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
8 years agoradv: emit correct last export when Z/stencil export is enabled
Dave Airlie [Tue, 8 Nov 2016 06:22:39 +0000 (16:22 +1000)]
radv: emit correct last export when Z/stencil export is enabled

I was getting a random GPU hang in the renderpass simple tests,
it turns out sometimes radv emitted the wrong thing "last".

This fixes the logic to emit Z/stencil last if they occur,
and not mark a color output as last. Also this relies on the
Z/STENCIL being the first two fragment outputs, which they are
so yay.

Fixes: dEQP-VK.renderpass.simple.color_depth (random hangs)
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agotgsi/scan: turn a huge if-else-if.. chain into a switch statement
Marek Olšák [Sat, 5 Nov 2016 17:21:57 +0000 (18:21 +0100)]
tgsi/scan: turn a huge if-else-if.. chain into a switch statement

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agotgsi/scan: fix images_buffers regression
Marek Olšák [Sat, 5 Nov 2016 17:16:16 +0000 (18:16 +0100)]
tgsi/scan: fix images_buffers regression

The first IF statement disabled the second one.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98599

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoanv: Document cmd_buffer_alloc_binding_table
Jason Ekstrand [Mon, 7 Nov 2016 20:32:28 +0000 (12:32 -0800)]
anv: Document cmd_buffer_alloc_binding_table

Some of the details of this function are very confusing and have a long
history.  We should document that history and this seems like the best
place to do it.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agointel/blorp: Emit all the binding tables
Jason Ekstrand [Sun, 23 Oct 2016 05:27:23 +0000 (22:27 -0700)]
intel/blorp: Emit all the binding tables

At least on Sky Lake, after emitting 3DSTATE_CONSTANT_*, you are required
to re-emit the 3DSTATE_BINDING_TABLE_POINTERS packet for the corresponding
stage.  If you don't, double-buffering may fail and you may get the wrong
constants.  It turns out that you need to do this even if you have no push
constants to speak of or else the next 3DSTATE_CONSTANT packet you emit for
that stage may not work correctly.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
8 years agoi965/gen9: Allow sampling with hiz when supported
Jordan Justen [Fri, 21 Oct 2016 15:50:50 +0000 (16:50 +0100)]
i965/gen9: Allow sampling with hiz when supported

For gen9+ this will indicate when we should allow hiz based sampling
during rendering.

Improves performance in :
  - Synmark's OglDeferred by 2.2% (n=20)
  - Synmark's OglShMapPcf by 0.44% (n=20)

v2 by Ben: Add spec reference, and make it fix with some of the changes made on
the previous patches
Change the check from mt->aux_buf to mt->num_samples. The presence of an aux_buf
isn't enough to determine there isn't a HiZ buffer to use.

v3: It seems all depth surface end up with num_samples = 0 by default,
    so allow sampling from depth HiZ if num_samples <= 1. (Lionel)
    Allow sampling from HiZ only if all LOD are available from the HiZ
    buffer. (Lionel)

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com> (v1)
Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v2)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v3)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/gen9: Add HiZ auxiliary buffer support
Ben Widawsky [Fri, 21 Oct 2016 15:47:23 +0000 (16:47 +0100)]
i965/gen9: Add HiZ auxiliary buffer support

The original functionality this patch introduces was authored by a patch from
Ken (the commit subject was the same). Since I ended up changing so many patches
in the code before this one, I had some non-trivial decisions to make, and I
didn't feel it was appropriate to keeps Ken's name as author (mostly because he
might not like what I've done). Ken's original patch was like 2 LOC :-)

In either case, some credit needs to go to Ken, and to Jordan for a few small
other changes in that original patch.

v2: Back to a smaller diff now that ISL handles most of the actual
    programming (Lionel)

Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v1)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v2)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Add function to indicate when sampling with hiz is supported
Jordan Justen [Fri, 21 Oct 2016 14:46:37 +0000 (15:46 +0100)]
i965: Add function to indicate when sampling with hiz is supported

Currently it indicates that this is never supported, but soon it will
be supported for gen8+^w gen9+

v2 by Ben:
- Explicitly disable aux_hiz for gen < 9 (with comment)
- squashed in next patch to avoid unused and useless functions

   i965: Support sampling with hiz during rendering

   For gen8, we can sample from depth while using the hiz buffer. This
   allows us to sample depth without resolving from hiz to the depth
   texture.

   To do this we must resolve to hiz before drawing so we can use the hiz
   buffer to sample while rendering. Hopefully the hiz buffer will
   already be resolved in most cases because it was previously rendered,
   meaning the hiz resolve is a no-op.

   Note that this is still controlled by the
   intel_miptree_sample_with_hiz function, and we will enable hiz
   sampling for gen8 in a separate patch.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com> (v1)
Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v2)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/miptree: Create a hiz mcs type
Ben Widawsky [Fri, 21 Oct 2016 14:10:56 +0000 (15:10 +0100)]
i965/miptree: Create a hiz mcs type

This seems counter to the goal of consolidating hiz, mcs, and later ccs buffers.
Unfortunately, hiz on gen6 is a thing the code supports, and this wart will be
helpful to achieve that. Overall, I believe it does help unify AUX buffers on
gen7+.

I updated the size field which I introduced in the previous patch, even though
we have no use for it.

XXX: As I mentioned in the last patch, the height given to the MCS buffer
allocation in intel_miptree_alloc_mcs() looks wrong, but I don't claim to fully
understand how the MCS buffer is laid out.

v2: rebase on master (Lionel)

Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v1)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v2)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Drop the aux mt when not used
Ben Widawsky [Fri, 21 Oct 2016 13:20:39 +0000 (14:20 +0100)]
i965: Drop the aux mt when not used

This patch will preserve the BO & offset, and not the miptree for the
aux_mcs buffer. Eventually it might make sense to pull put the sizing
function in miptree creation, but for now this should be sufficient
and not too hideous.

v2: Save BO's offset too (Lionel)

v3: Squash previous patch storing the size of the allocated aux buffer
    (Lionel)
    Fix memory leak with mcs_buf->bo (Lionel)

Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v1)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v2)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/miptree: Directly gtt map the mcs buffer
Ben Widawsky [Fri, 21 Oct 2016 13:17:21 +0000 (14:17 +0100)]
i965/miptree: Directly gtt map the mcs buffer

The next patch will change the map type, and this will make sure there are no
regressions as a result of the other stuff. Since the miptree is newly created,
I believe it is always safe to just map.

It is possible to CPU map this buffer on LLC platforms (it additionally requires
rounding up to tile size). I did experiment with that patch, and found no
performance gains to be had.

I've added in error handling while here. Generally GTT mapping is an operation
which is highly unlikely to fail, but we may as well handle it when it does.

v2: rebase on master (Lionel)

v3: print out error if gtt mapping fails (Topi)

Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com> (v1)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v2)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Wrap MCS miptree in intel_miptree_aux_buffer
Jordan Justen [Fri, 21 Oct 2016 11:56:49 +0000 (12:56 +0100)]
i965: Wrap MCS miptree in intel_miptree_aux_buffer

This will allow us to treat HiZ and MCS the same when using as an
auxiliary surface buffer.

v2: (Ben) Minor rebase conflict resolution.
   Rename mcs_buf to aux_buf to address upcoming change for hiz specific buffers.
   That second thing is essentially a squash of:
   i965/gen8: Use intel_miptree_aux_buffer for auxiliary buffer - which didn't need
   to be separate in my opinion.

v3: rebase on master (Lionel)

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com> (v1)
Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com>a (v2)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> (v3)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agogallivm: fix [IU]MUL_HI regression
Nicolai Hähnle [Tue, 8 Nov 2016 09:14:00 +0000 (10:14 +0100)]
gallivm: fix [IU]MUL_HI regression

This patch does two things:

1. It separates the host-CPU code generation from the generic code
   generation. This guards against accidently breaking things for
   radeonsi in the future.

2. It makes sure we actually use both arguments and don't just compute
   a square :-p

Fixes a regression introduced by commit 29279f44b3172ef3b84d470e70fc7684695ced4b

Cc: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agodraw: use vectorized calculations for fetch
Roland Scheidegger [Fri, 4 Nov 2016 04:13:03 +0000 (05:13 +0100)]
draw: use vectorized calculations for fetch

Instead of doing all the math with scalars, use vectors. This means the
overflow math needs to be done manually, albeit that's only really
problematic for the stride/index mul, the rest has been pretty much
moved outside the shader loop (albeit the mul could actually be optimized
away too), where things are still scalar. Because llvm is complete fail
with the zero-extend widening mul, roll our own even...
To eliminate control flow in the main shader loop fetch, provide fake
buffers (so index 0 is always valid to fetch).
Still uses aos fetch though in the end - mostly because some more code
would be needed to handle unaligned fetches in that path, and because for
most formats it won't make a difference anyway (we generate some truly
horrendous code for things like R16G16_something for instance).

Instanced fetch however stays roughly the same as before, except that
no longer the same element is fetched multiple times (I've seen a reduction
of ~3 times in main shader loop size due to apparently llvm not being able
to deduce it's really all the same with a couple instanced elements).

Also, for elts gathering, use vectorized code as well - provide a fake
elt buffer if there's no valid one bound.

The generated shaders are smaller and faster to compile (not entirely sure
about execution speed, but generally unless there's just single vertices
to handle I would expect it to be faster - there's more opportunities
for future improvements by using soa fetch).

No piglit change.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agogallivm: introduce 32x32->64bit lp_build_mul_32_lohi function
Roland Scheidegger [Fri, 4 Nov 2016 03:55:09 +0000 (04:55 +0100)]
gallivm: introduce 32x32->64bit lp_build_mul_32_lohi function

This is used by shader umul_hi/imul_hi functions (and soon by draw).
It's actually useful separating this out on its own, however the real
reason for doing it is because we're using an optimized sse2 version,
since the code llvm generates is atrocious (since there's no widening
mul in llvm, and it does not recognize the widening mul pattern, so
it generates code for real 64x64->64bit mul, which the cpu can't do
natively, in contrast to 32x32->64bit mul which it could do).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agoi965: Add space before paren
Anuj Phogat [Fri, 28 Oct 2016 18:01:42 +0000 (11:01 -0700)]
i965: Add space before paren

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoi965: Remove unnecessary white space
Anuj Phogat [Fri, 28 Oct 2016 17:58:44 +0000 (10:58 -0700)]
i965: Remove unnecessary white space

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoi965: Fix alpha-to-coverage and alpha test enabled checks
Anuj Phogat [Thu, 20 Oct 2016 18:40:40 +0000 (11:40 -0700)]
i965: Fix alpha-to-coverage and alpha test enabled checks

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agomesa: Add helper function _mesa_is_alpha_to_coverage_enabled()
Anuj Phogat [Tue, 25 Oct 2016 18:56:07 +0000 (11:56 -0700)]
mesa: Add helper function _mesa_is_alpha_to_coverage_enabled()

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agomesa: Add helper function _mesa_is_alpha_test_enabled()
Anuj Phogat [Tue, 25 Oct 2016 18:55:44 +0000 (11:55 -0700)]
mesa: Add helper function _mesa_is_alpha_test_enabled()

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agomesa: Use separate line for function return type
Anuj Phogat [Tue, 25 Oct 2016 18:54:36 +0000 (11:54 -0700)]
mesa: Use separate line for function return type

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agonvc0: simplify draw parameters upload for vertex shaders
Samuel Pitoiset [Tue, 25 Oct 2016 19:41:12 +0000 (21:41 +0200)]
nvc0: simplify draw parameters upload for vertex shaders

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agogallium/hud: protect against and initialization race
Steven Toth [Mon, 24 Oct 2016 14:10:51 +0000 (10:10 -0400)]
gallium/hud: protect against and initialization race

In the event that multiple threads attempt to install a graph
concurrently, protect the shared list.

Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agogallium/hud: close a previously opened handle
Steven Toth [Mon, 24 Oct 2016 14:10:50 +0000 (10:10 -0400)]
gallium/hud: close a previously opened handle

We're missing the closedir() to the matching opendir().

Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agogallium/hud: fix a problem where objects are free'd while in use.
Steven Toth [Mon, 24 Oct 2016 14:10:49 +0000 (10:10 -0400)]
gallium/hud: fix a problem where objects are free'd while in use.

Instead of trying to maintain a reference counted list of valid HUD
objects, and freeing them accordingly, creating race conditions
between unanticipated multiple threads, simply accept they're
allocated once and never released until the process terminates.

They're a shared resource between multiple threads, so accept
they're always available for use.

Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agomesa: drop current draw/read buffer when ctx is released
Rob Clark [Wed, 26 Oct 2016 20:52:52 +0000 (16:52 -0400)]
mesa: drop current draw/read buffer when ctx is released

This fixes a problem seen with gallium drivers vs android wallpaper.
Basically, what happens is:

   EGLSurface tmpSurface = mEgl.eglCreatePbufferSurface(mEglDisplay, mEglConfig, attribs);
   mEgl.eglMakeCurrent(mEglDisplay, tmpSurface, tmpSurface, mEglContext);

   int[] maxSize = new int[1];
   Rect frame = surfaceHolder.getSurfaceFrame();
   glGetIntegerv(GL_MAX_TEXTURE_SIZE, maxSize, 0);

   mEgl.eglMakeCurrent(mEglDisplay, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
   mEgl.eglDestroySurface(mEglDisplay, tmpSurface);

   ... check maxSize vs frame size and bail if needed ...

   mEglSurface = mEgl.eglCreateWindowSurface(mEglDisplay, mEglConfig, surfaceHolder, null);
   ... error checking ...
   mEgl.eglMakeCurrent(mEglDisplay, mEglSurface, mEglSurface, mEglContext);

When the window-surface is created, it ends up with the same ptr address
as the recently freed tmpSurface pbuffer surface.  Which after many
levels of indirection, results in st_framebuffer_validate() ending up with
the same/old framebuffer object, and in the end never calling the
DRIimageLoaderExtension::getBuffers().  Then in droid_swap_buffers(), the
dri2_surf is still the old pbuffer surface (with dri2_surf->buffer being
NULL, obviously, so when wallpaper app calls eglSwapBuffers() nothing
gets enqueued to the compositor).  Resulting in a black/blank background
layer.

Note that at the EGL layer, when the context is unbound, EGL drops it's
references to the draw and read buffer as well.

Signed-off-by: Rob Clark <robdclark@gmail.com>
Tested-by: Robert Foss <robert.foss@collabora.com>
Acked-by: Tapani Pälli <tapani.palli@intel.com>
8 years agoclover: Add CL_PROGRAM_BINARY_TYPE support (CL1.2).
Serge Martin [Mon, 31 Oct 2016 00:21:15 +0000 (17:21 -0700)]
clover: Add CL_PROGRAM_BINARY_TYPE support (CL1.2).

v3 [Francisco Jerez]: Loosely based on Serge's v1 of this patch in
   order to avoid CL-specific enums in the clover module binary
   format.  In addition to other changes made in v2: Represent the CL
   program binary type as the section type instead of adding a CL
   API-specific enum, check that the binary types of the input objects
   are valid during clLinkProgram(), pass section type as argument to
   build_module_library() instead of using separate function.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoclover: add missing clGetDeviceInfo CL1.2 queries
Serge Martin [Sat, 1 Oct 2016 16:51:11 +0000 (18:51 +0200)]
clover: add missing clGetDeviceInfo CL1.2 queries

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Vedran Miletić <vedran@miletic.net>
8 years agonvc0: get rid of NVE4_COMPUTE_MP_PM_{A,B}_SIGSEL_XXX
Samuel Pitoiset [Sat, 5 Nov 2016 16:56:02 +0000 (17:56 +0100)]
nvc0: get rid of NVE4_COMPUTE_MP_PM_{A,B}_SIGSEL_XXX

Instead, hardcode group sigsel because there are a bunch of unknown
groups, especially on SM50/SM52.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
8 years agogm107/ir: emit RED instead of ATOM when no dst
Samuel Pitoiset [Fri, 4 Nov 2016 19:08:57 +0000 (20:08 +0100)]
gm107/ir: emit RED instead of ATOM when no dst

This is similar to NVC0 and GK110 emitters where we emit
reduction operations instead of atomic operations when the
destination is not used.

Found after writing some tests which check if performance counters
return the expected value. In that case, gred_count returned 0
on gm107 while at least gk106 returned the correct value.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agost/mesa: initialize members of glsl_to_tgsi_instruction in emit_asm()
Brian Paul [Sat, 5 Nov 2016 17:00:10 +0000 (11:00 -0600)]
st/mesa: initialize members of glsl_to_tgsi_instruction in emit_asm()

This fixes random crashes with MSVC release builds.  It seems the
members are implicitly initialized to zero with gcc, but not MSVC.
In particular, the tex_offset_num_offset field was non-zero causing
a loop over the NULL tex_offsets array to crash.

Zero-init those fields and a few others to be safe.

The regression began with acc23b04cfd64e "ralloc: remove memset from
ralloc_size".

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoandroid: amd/common: add support for libmesa_amd_common
Mauro Rossi [Fri, 4 Nov 2016 23:00:29 +0000 (00:00 +0100)]
android: amd/common: add support for libmesa_amd_common

Fixes the following building error introduced with commit 7115e56
and related amd/common dependencies:

external/mesa/src/gallium/drivers/radeonsi/si_shader.c:6861: error: undefined reference to 'ac_is_sgpr_param'
external/mesa/src/gallium/drivers/radeonsi/si_shader.c:6951: error: undefined reference to 'ac_is_sgpr_param'
clang++: error: linker command failed with exit code 1 (use -v to see invocation)

ninja: build stopped: subcommand failed.
build/core/ninja.mk:148: recipe for target 'ninja_wrapper' failed
make: *** [ninja_wrapper] Error 1

Signed-off-by: Marek Olšák <marek.olsak@amd.com>