mesa.git
8 years agoradeonsi: convert to 64-bitness checks instead of doubles.
Dave Airlie [Fri, 10 Jun 2016 01:05:51 +0000 (11:05 +1000)]
radeonsi: convert to 64-bitness checks instead of doubles.

This converts to testing for 64-bit types and renames some things
in anticipation of 64-bit integer support.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agogallivm: make non-float return code bitcast consistent.
Dave Airlie [Fri, 10 Jun 2016 00:55:59 +0000 (10:55 +1000)]
gallivm: make non-float return code bitcast consistent.

This just uses the same form across the fetches.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agogallium/gallivm: use 64-bit test instead of doubles.
Dave Airlie [Fri, 10 Jun 2016 00:49:25 +0000 (10:49 +1000)]
gallium/gallivm: use 64-bit test instead of doubles.

This just makes some generic code that currently emits double
suitable for emitting 64-bit values.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agogallium/tgsi: add 64-bitness type check function.
Dave Airlie [Fri, 10 Jun 2016 00:48:35 +0000 (10:48 +1000)]
gallium/tgsi: add 64-bitness type check function.

Currently this just doubles, but we'll convert users to this
so making adding 64-bit integers easier.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoanv/entrypoints: Rework #if guards
Jason Ekstrand [Fri, 10 Jun 2016 19:42:52 +0000 (12:42 -0700)]
anv/entrypoints: Rework #if guards

This reworks the #if guards a bit.  When Emil originally wrote them, he
just guarded everything.  However, part of what anv_entrypoints_gen.py
generates is a hash table for looking up entrypoints based on their name.
This table *cannot* get out of sync between C and python regardless of
preprocessor flags.  In order to prevent this, this commit makes us use
void pointers in the dispatch table for those entrypoints which aren't
available.  This means that the dispatch table size and entry order is
constant and it should never get out-of-sync with the python.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Emil Velikov <emil.velikov@collabora.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/entrypoints: Use the function pointer types provided by vulkan.h
Jason Ekstrand [Fri, 10 Jun 2016 19:30:05 +0000 (12:30 -0700)]
anv/entrypoints: Use the function pointer types provided by vulkan.h

This is a bit cleaner than generating the types ourselves when making the
table.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Emil Velikov <emil.velikov@collabora.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agost/mesa: use base level size as "guess" when available
Nicolai Hähnle [Mon, 6 Jun 2016 21:15:10 +0000 (23:15 +0200)]
st/mesa: use base level size as "guess" when available

When an applications specifies mip levels _before_ setting a mipmap texture
filter, we will initially guess a single texture level. When the second level
image is created, we try to allocate the full texture -- however, we get the
base level size guess wrong if that size is odd. This leads to yet another
re-allocation of the texture later during st_finalize_texture.

Even worse, this re-allocation breaks a (reasonable) assumption made by
st_generate_mipmaps, because the re-allocation in the finalization call will
again allocate a single-level pipe texture (based on the non-mipmap texture
filter!). As a result, mipmap generation fails in interesting ways.

All of this can be avoided by just using the fact that we already know the
size of the base level.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95529
Cc: 12.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
8 years agoanv: Remove the PhysicalDeviceLimits FINISHME
Jason Ekstrand [Fri, 10 Jun 2016 16:43:45 +0000 (09:43 -0700)]
anv: Remove the PhysicalDeviceLimits FINISHME

At this point, the limits are probably more-or-less correct.  If there is
an invalid limit, that's a bug not a FINSHME.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/pipeline_cache: Allow for an zero-sized cache
Jason Ekstrand [Mon, 6 Jun 2016 18:20:44 +0000 (11:20 -0700)]
anv/pipeline_cache: Allow for an zero-sized cache

This gets ANV_ENABLE_PIPELINE_CACHE=false working again.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/pipeline: Store the (set, binding, index) tripple in the bind map
Jason Ekstrand [Mon, 6 Jun 2016 18:12:27 +0000 (11:12 -0700)]
anv/pipeline: Store the (set, binding, index) tripple in the bind map

This way the the bind map (which we're caching) is mostly independent of
the pipeline layout.  The only coupling remaining is that we pull the array
size of a binding out of the layout.  However, that size is also specified
in the shader and should always match so it's not really coupled.  This
rendering issues in Dota 2.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/descriptor_set: Ensure that bindings are always in increasing order
Jason Ekstrand [Mon, 6 Jun 2016 16:15:03 +0000 (09:15 -0700)]
anv/descriptor_set: Ensure that bindings are always in increasing order

Since applications are allowed to specify some set of bindings which need
not be dense they also need not be in order.  For most things, this doesn't
matter, but it could result getting the wrong dynamic offsets. This adds a
quick-and-dirty sort to ensure that everything is always in increasing
order of binding index.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/descriptor_set: Add a type field in debug builds
Jason Ekstrand [Mon, 6 Jun 2016 16:12:50 +0000 (09:12 -0700)]
anv/descriptor_set: Add a type field in debug builds

This allows for some extra validation and makes it easier to see what's
going on when poking around in gdb.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoanv/descriptor_set: Set array_size to zero for non-existant descriptors
Jason Ekstrand [Mon, 6 Jun 2016 16:12:20 +0000 (09:12 -0700)]
anv/descriptor_set: Set array_size to zero for non-existant descriptors

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agovl/dri3: support receiving new pixmap for front buffer
Leo Liu [Thu, 9 Jun 2016 16:53:54 +0000 (12:53 -0400)]
vl/dri3: support receiving new pixmap for front buffer

With glx of gstreamer-vaapi, the temporary pixmap for front buffer gets
renewed in each frame, so when we receive a new pixmap, should get a new
front buffer for it.

This also fixes Totem player playback corruption.

Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agovl/dri3: get Makefile properly
Leo Liu [Thu, 9 Jun 2016 17:11:52 +0000 (13:11 -0400)]
vl/dri3: get Makefile properly

From original commit, the macro "if HAVE_DRI3" was in Makefile.sources,
this file is shared with SCons, SCons is not able to parse this marco,
the SCons build failed. Jose quickly gave two approaches and quick fix
with his second approach, thanks Jose for the solutions and fixes.

This patch is Jose's first approach, and it's more proper, because the
dri3 c file should not be included to build when DRI3 is not enabled.

Signed-off-by: Leo Liu <leo.liu@amd.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agogallivm: Never emit llvm.fmuladd on LLVM 3.3.
Jose Fonseca [Fri, 10 Jun 2016 14:42:55 +0000 (15:42 +0100)]
gallivm: Never emit llvm.fmuladd on LLVM 3.3.

Besides the old JIT bug, it seems the X86 backend on LLVM 3.3 doesn't
handle llvm.fmuladd and instead it fall backs to a C function.  Which in
turn causes a segfault on Windows.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agogallivm: Use llvm.fmuladd.*.
Jose Fonseca [Sun, 3 Apr 2016 23:05:33 +0000 (00:05 +0100)]
gallivm: Use llvm.fmuladd.*.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoutil,gallivm: Explicitly enable/disable fma attribute.
Jose Fonseca [Tue, 24 May 2016 13:36:28 +0000 (22:36 +0900)]
util,gallivm: Explicitly enable/disable fma attribute.

As suggested by Roland Scheidegger.

Use the same logic as f16c, since fma requires VEX encoding.

But disable FMA on LLVM 3.3 without MCJIT.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoradeonsi: Reinitialize all descriptors in CE preamble.
Bas Nieuwenhuizen [Mon, 6 Jun 2016 20:36:35 +0000 (22:36 +0200)]
radeonsi: Reinitialize all descriptors in CE preamble.

This fixes a problem with the CE preamble and restoring only stuff in the
preamble when needed.

To illustrate suppose we have two graphics IB's 1 and 2, which  are submitted in
that order. Furthermore suppose IB 1 does not use CE ram, but IB 2 does, and we
have a context switch at the start of IB 1, but not between IB 1 and IB 2.

The old code put the CE RAM loads in the preamble of IB 2. As the preamble of
IB 1 does not have the loads and the preamble of IB 2 does not get executed, the
old values are not load into CE RAM.

Fix this by always restoring the entire CE RAM.

v2: - Just load all descriptor set buffers instead of load and store the entire
      CE RAM.
    - Leave the ce_ram_dirty tracking in place for the non-preamble case.

v3: - Fixed parameter alignment.
    - Rebased to master (Nicolai's descriptor series).

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agomesa: Wrap extensions.h declarations with extern "C".
Jose Fonseca [Fri, 10 Jun 2016 10:00:42 +0000 (11:00 +0100)]
mesa: Wrap extensions.h declarations with extern "C".

This should fix the MSVC linker failures that arose with commit
5e2d25894b962aae9158261897e13843377e3b95.

Trivial.

8 years agost/mesa: fix type confusion with reladdrs
Ilia Mirkin [Wed, 8 Jun 2016 15:32:54 +0000 (11:32 -0400)]
st/mesa: fix type confusion with reladdrs

The reality is that this doesn't matter, because we manually emit the
ARL to the sampler reladdr, and those arguments don't get an extra load
later, so it's effectively just a boolean. However having the types be
wrong is confusing and could trigger very odd bugs should usage change
down the line.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Brian Paul <brianp@vmware.com>
8 years agoglsl/ir: remove TABs in ir_constant_expression.cpp
Dave Airlie [Fri, 10 Jun 2016 00:00:14 +0000 (10:00 +1000)]
glsl/ir: remove TABs in ir_constant_expression.cpp

Adding 64-bit integers support was going to make this file worse,
just remove the tabs from it now.

Acked-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoi965/gen9: Don't change halign and valign to fit in fast copy blit
Anuj Phogat [Wed, 8 Jun 2016 18:28:12 +0000 (11:28 -0700)]
i965/gen9: Don't change halign and valign to fit in fast copy blit

An update in graphics specs has deleted the halign and valign fields
from XY_FAST_COPY_BLT command. See mesa commit 97f0f91.

Cc: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
8 years agomesa: Add a helper function for shared code in get_tex_rgba_{un}compressed
Anuj Phogat [Mon, 28 Dec 2015 18:46:51 +0000 (10:46 -0800)]
mesa: Add a helper function for shared code in get_tex_rgba_{un}compressed

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
8 years agomesa: Let compute shaders work in compatibility profiles
Samuel Pitoiset [Wed, 25 May 2016 19:17:24 +0000 (21:17 +0200)]
mesa: Let compute shaders work in compatibility profiles

The extension is already advertised in compatibility profile, but
the _mesa_has_compute_shaders only returns true in core profile.
If we advertise it, we should allow it to work.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoswr: implement clipPlanes/clipVertex/clipDistance/cullDistance
Tim Rowley [Wed, 25 May 2016 23:49:34 +0000 (18:49 -0500)]
swr: implement clipPlanes/clipVertex/clipDistance/cullDistance

v2: only load the clip vertex once

v3: fix clip enable logic, add cullDistance

v4: remove duplicate fields in vs jit key, fix test of clip fixup needed

v5: fix clipdistance linkage for slot!=0,4

v6: support clip+cull; passes most piglit clip (failures understood)

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
8 years agoglx: fix crash with bad fbconfig
Daniel Czarnowski [Wed, 10 Feb 2016 17:36:05 +0000 (09:36 -0800)]
glx: fix crash with bad fbconfig

GLX documentation states:
glXCreateNewContext can generate the following errors: (...)
GLXBadFBConfig if config is not a valid GLXFBConfig

Function checks if the given config is a valid config and sets proper
error code.

Fixes currently crashing glx-fbconfig-bad Piglit test.

v2: coding style cleanups (Emil, Topi)
    use DefaultScreen macro (Emil)

Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Cc: "11.2" <mesa-stable@lists.freedesktop.org>
8 years agost/vdpau: implement luma keying
Nayan Deshmukh [Wed, 8 Jun 2016 09:36:25 +0000 (15:06 +0530)]
st/vdpau: implement luma keying

Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
8 years agovl: Apply luma key filter before CSC conversion
Nayan Deshmukh [Wed, 8 Jun 2016 09:22:48 +0000 (14:52 +0530)]
vl: Apply luma key filter before CSC conversion

    Apply the luma key filter to the YCbCr values during the CSC conversion
    in video buffer shader. The initial values of max and min luma are set
    to opposite values to disable the filter initially and will be set when
    enabling it.

    Add extra parmeters min and max luma for the luma key filter in
    vl_compositor_set_csc_matrix in va, xvmc. Setting them
    to opposite value 1.f and 0.f respectively won't effect the CSC
    conversion

    v2: -Squash 1,2 and 3 into one patch to avoid breaking build of
        other components. (Christian)
        -use ureg_swizzle. (Christian)
        -change name of the variables. (Christian)

    v3: -Squash all patches in one to avoid breaking of build. (Emil)
        -wrap functions properly. (Emil)
        -use 0.0f and 1.0f instead of 0.f and 1.f respectively. (Emil)

    v4: -Divide it in two patches one which introduces the functionality
 and assigs dummy values to the changed functions and second which
 implements the lumakey filter. (Christian)
-use ureg_scalar instead ureg_swizzle. (Christian)

Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
8 years agoi965: Emit surface states for extra planes prior to gen8
Jason Ekstrand [Thu, 9 Jun 2016 02:56:46 +0000 (19:56 -0700)]
i965: Emit surface states for extra planes prior to gen8

When Kristian implemented GL_TEXTURE_EXTERNAL_OES, he hooked it up for gen8
but not for gen7 or earlier.  It all works, we just need to emit the states
for the extra planes.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agovirgl: fix checking fences
Marc-André Lureau [Tue, 7 Jun 2016 12:54:34 +0000 (14:54 +0200)]
virgl: fix checking fences

When calling virgl_fence_wait() with timeout=0,
virgl_{drm,vtest}_resource_is_busy() is called. However, it returns TRUE
for a busy resource, whereace virgl_fence_wait() should return TRUE for
a completed (non-busy) resource.

This fixes running supertuxkart in a VM (I could not reproduce locally
with vtest though there is a similar fix)

Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Cc: "11.1 11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoglsl/types: rename is_dual_slot_double to is_dual_slot_64bit.
Dave Airlie [Wed, 8 Jun 2016 21:50:43 +0000 (07:50 +1000)]
glsl/types: rename is_dual_slot_double to is_dual_slot_64bit.

In the future int64 support will have the same requirements.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agost/glsl_to_tgsi: move to checking 64-bitness instead of double
Dave Airlie [Wed, 8 Jun 2016 20:57:35 +0000 (06:57 +1000)]
st/glsl_to_tgsi: move to checking 64-bitness instead of double

This uses the new types interfaces to check for 64-bit types,
as futureproofing against int64 support.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agost/glsl_to_tgsi: use enum glsl_base_type instead of unsigned
Dave Airlie [Wed, 8 Jun 2016 20:48:33 +0000 (06:48 +1000)]
st/glsl_to_tgsi: use enum glsl_base_type instead of unsigned

This is just some better type safety that I noticed while working
on 64-bit integer support.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agomesa: use new 64-bit checks instead of explicit double checks.
Dave Airlie [Wed, 8 Jun 2016 21:07:42 +0000 (07:07 +1000)]
mesa: use new 64-bit checks instead of explicit double checks.

This just moves to the new interfaces in advance of int64.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoglsl/link_varyings: switch to 64bit check instead of double.
Dave Airlie [Wed, 8 Jun 2016 21:02:36 +0000 (07:02 +1000)]
glsl/link_varyings: switch to 64bit check instead of double.

This is prep work for int64 support.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoglsl: use new interfaces for 64-bit checks.
Dave Airlie [Wed, 8 Jun 2016 21:02:07 +0000 (07:02 +1000)]
glsl: use new interfaces for 64-bit checks.

This is just prep work for int64 support, changing
places where 64-bit matters no doubles.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agocompiler: use 64bit check for sizing instead of double check.
Dave Airlie [Wed, 8 Jun 2016 21:07:06 +0000 (07:07 +1000)]
compiler: use 64bit check for sizing instead of double check.

This just moves code to the new check in advance of int64 support.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agocompiler/types: add 64-bitness queries.
Dave Airlie [Wed, 8 Jun 2016 20:52:15 +0000 (06:52 +1000)]
compiler/types: add 64-bitness queries.

This adds an inline and type query for if a type is 64-bit.

Fow now this is equivalent to double, but int64 will change
this.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoglapi/glx: Add overflow checks to the client-side indirect code
Adam Jackson [Tue, 24 May 2016 19:45:11 +0000 (15:45 -0400)]
glapi/glx: Add overflow checks to the client-side indirect code

Coverity complains that the computed sizes can lead to negative lengths
passed to memcpy. If that happens we've been handed invalid arguments
anyway, so just bomb out.

The funky "0%s" is because the size string for the variable-length part
of the request is of the form "+ safe_pad() ...", and a unary + would
coerce the result to always be positive, defeating the overflow check.

Signed-off-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoradeonsi: improve the computation and comment of scratch_waves
Marek Olšák [Wed, 8 Jun 2016 12:34:11 +0000 (14:34 +0200)]
radeonsi: improve the computation and comment of scratch_waves

2% isn't much. If you think the number should be decreased, please speak up.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: print the number of spilled VGPRs
Marek Olšák [Wed, 8 Jun 2016 11:51:04 +0000 (13:51 +0200)]
radeonsi: print the number of spilled VGPRs

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agogallium/radeon: remove dead code creating LLVMTargetMachine
Marek Olšák [Tue, 7 Jun 2016 22:42:01 +0000 (00:42 +0200)]
gallium/radeon: remove dead code creating LLVMTargetMachine

This was for some old unsupported LLVM version.
Only si_create_context creates the target machine now.
r600g doesn't use this function.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: don't enable scratch just for SGPR spills
Marek Olšák [Tue, 7 Jun 2016 22:14:58 +0000 (00:14 +0200)]
radeonsi: don't enable scratch just for SGPR spills

Diff from shader-db:
  Scratch: 3221504 -> 17408 (-99.46 %) bytes per wave

v2: add "break;"

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agost/mesa: try not to compile compute shader on the first use
Marek Olšák [Tue, 7 Jun 2016 21:03:38 +0000 (23:03 +0200)]
st/mesa: try not to compile compute shader on the first use

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agoRevert "radeonsi: allow direct hw MSAA resolve for scanout surfaces"
Marek Olšák [Wed, 8 Jun 2016 16:52:25 +0000 (18:52 +0200)]
Revert "radeonsi: allow direct hw MSAA resolve for scanout surfaces"

This reverts commit ffd54d1936fcd07424265b780e1d049222a01e94.

No, it doesn't work. The test case is "glxgears -samples 2".

8 years agost/mesa: directly compute level=0 texture size in st_finalize_texture
Nicolai Hähnle [Tue, 7 Jun 2016 20:40:49 +0000 (22:40 +0200)]
st/mesa: directly compute level=0 texture size in st_finalize_texture

The width0/height0/depth0 on stObj may not have been set at this point.
Observed in a trace that set up levels 2..9 of a 2d texture, and set the base
level to 2, with height 1. This made the guess logic always bail.

Originally investigated by Ilia Mirkin, this patch gets rid of the somewhat
redundant storage of width0/height0/depth0 and makes sure we always compute
pipe texture sizes that are compatible with the base level image of the
GL texture.

Fixes the gl-1.2-texture-base-level piglit test provided by Brian Paul.

v2:
- try to re-use an existing pipe texture when possible
- handle a corner case where the base level is not level 0 and it is of
  size 1x1x1

v3:
- ptHeight = ptWidth in cube map 1x1 case (suggested by Brian)

Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
8 years agoglsl: stop allocating memory for SSBOs and builtins
Timothy Arceri [Thu, 2 Jun 2016 05:32:14 +0000 (15:32 +1000)]
glsl: stop allocating memory for SSBOs and builtins

This just stops counting and assigning a storage location for
these uniforms, the count is only used to create the uniform storage.

These uniform types don't use this storage.

Reviewed-by: Dave Airlie <airlied@redhat.com>
8 years agost/mesa: use buffer usage history to set dirty flags for revalidation
Ilia Mirkin [Sat, 4 Jun 2016 17:26:46 +0000 (13:26 -0400)]
st/mesa: use buffer usage history to set dirty flags for revalidation

We were previously unconditionally doing this for arrays and ubo's, and
ignoring texture/storage/atomic buffers. Instead use the usage history
to determine which atoms need to be revalidated.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agoi965: Integrate precise trig into configuration infrastructure
Gurchetan Singh [Wed, 11 May 2016 20:32:09 +0000 (13:32 -0700)]
i965: Integrate precise trig into configuration infrastructure

With this change, to enable precise SIN and COS instructions
on Intel hardware, one can put

<option name="precise_trig" value="true"/>

in the proper drirc file.

V2: Make option name more generic

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Stephane Marchesin <stephane.marchesin@gmail.com>
8 years agoradeonsi: re-enable PBO ReadPixels acceleration
Marek Olšák [Tue, 7 Jun 2016 19:34:31 +0000 (21:34 +0200)]
radeonsi: re-enable PBO ReadPixels acceleration

disabled by 4f1cccf570112f93265a4cace504eb763fa8f73e

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: allow MSAA resolving into a texture that has DCC enabled
Marek Olšák [Mon, 6 Jun 2016 00:01:36 +0000 (02:01 +0200)]
radeonsi: allow MSAA resolving into a texture that has DCC enabled

Since DCC is enabled almost everywhere now, it's important not to disable
this fast path.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agogallium/radeon: move DCC clearing into a separate function
Marek Olšák [Sun, 5 Jun 2016 23:54:20 +0000 (01:54 +0200)]
gallium/radeon: move DCC clearing into a separate function

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: allow direct hw MSAA resolve for scanout surfaces
Marek Olšák [Sun, 5 Jun 2016 15:43:43 +0000 (17:43 +0200)]
radeonsi: allow direct hw MSAA resolve for scanout surfaces

No idea why this was disabled, but it works fine.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: don't allocate DCC for the temporary MSAA resolve surface
Marek Olšák [Sun, 5 Jun 2016 23:42:46 +0000 (01:42 +0200)]
radeonsi: don't allocate DCC for the temporary MSAA resolve surface

Allocating it has no effect, but it adds overhead (useless DCC clear).

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: don't enable DCC in the sampler if first_level doesn't have it
Marek Olšák [Mon, 6 Jun 2016 15:33:42 +0000 (17:33 +0200)]
radeonsi: don't enable DCC in the sampler if first_level doesn't have it

If first_level > 0 and DCC is disabled for that level, let's skip DCC
reads entirely.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agowinsys/amdgpu: enable DCC for mipmapped textures
Marek Olšák [Fri, 3 Jun 2016 18:48:01 +0000 (20:48 +0200)]
winsys/amdgpu: enable DCC for mipmapped textures

Also add dcc_fast_clear_size for clearing only the necessary subset
of DCC. For no AA, it's equal to the size of the whole DCC level.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agogallium/radeon: don't disable DCC because of SDMA
Marek Olšák [Sun, 5 Jun 2016 13:45:30 +0000 (15:45 +0200)]
gallium/radeon: don't disable DCC because of SDMA

We want to keep DCC enabled to save bandwidth. It was a bad idea to disable
it here.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: don't flag renderbuffer feedback loop if DCC has just been disabled
Marek Olšák [Fri, 3 Jun 2016 18:51:47 +0000 (20:51 +0200)]
radeonsi: don't flag renderbuffer feedback loop if DCC has just been disabled

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: add per-level dcc_enabled flags
Marek Olšák [Fri, 3 Jun 2016 18:40:30 +0000 (20:40 +0200)]
radeonsi: add per-level dcc_enabled flags

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: compute DCC register parameters in si_emit_framebuffer_state
Marek Olšák [Fri, 3 Jun 2016 17:56:38 +0000 (19:56 +0200)]
radeonsi: compute DCC register parameters in si_emit_framebuffer_state

This will get more complicated with mipmapped DCC or when DCC is enabled
after allocation.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agogallium/radeon: add an assertion checking the validity of PIPE_BIND_SCANOUT
Marek Olšák [Mon, 6 Jun 2016 09:28:16 +0000 (11:28 +0200)]
gallium/radeon: add an assertion checking the validity of PIPE_BIND_SCANOUT

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agogallium/radeon: don't allocate DCC for non-renderable texture formats
Marek Olšák [Sun, 5 Jun 2016 23:29:14 +0000 (01:29 +0200)]
gallium/radeon: don't allocate DCC for non-renderable texture formats

R9G9B9E5 is the only uncompressed one hopefully.

This fixes incorrect rendering not discovered (due to a lack of tests)
until DCC mipmapping was enabled.

Cc: 11.1 11.2 12.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
8 years agoradeonsi: enable WQM in PS prolog when needed
Nicolai Hähnle [Wed, 1 Jun 2016 11:24:19 +0000 (13:24 +0200)]
radeonsi: enable WQM in PS prolog when needed

WQM is needed when the PS prolog computes a VGPR that is consumed by a shader
with (implicit or explicit) derivatives.

Depends on http://reviews.llvm.org/D20839 / LLVM r272063 for this to be
effective (otherwise it's just a no-op).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95130
Cc: 12.0 <mesa-dev@lists.freedesktop.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agotgsi/scan: add uses_derivatives (v2)
Nicolai Hähnle [Wed, 1 Jun 2016 11:17:29 +0000 (13:17 +0200)]
tgsi/scan: add uses_derivatives (v2)

v2:
- TG4 does not calculate derivatives (Ilia)
- also handle SAMPLE* instructions (Roland)

Cc: 12.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
Reviewed-by: Brian Paul <brianp@vmware.com> (v1)
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agodocs/devinfo: Expound on helpful extension tips
Nanley Chery [Fri, 3 Jun 2016 17:59:18 +0000 (10:59 -0700)]
docs/devinfo: Expound on helpful extension tips

Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agodocs/devinfo: Update bullet in stale extension guide
Nanley Chery [Fri, 3 Jun 2016 17:58:05 +0000 (10:58 -0700)]
docs/devinfo: Update bullet in stale extension guide

Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agodocs/devinfo: Add closing paragraph tag
Nanley Chery [Fri, 3 Jun 2016 17:56:46 +0000 (10:56 -0700)]
docs/devinfo: Add closing paragraph tag

Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoswr: fix provoking vertex
Tim Rowley [Wed, 1 Jun 2016 14:56:26 +0000 (09:56 -0500)]
swr: fix provoking vertex

Use rasterizer provoking vertex API.

Fix rasterizer provoking vertex for tristrips and quad list/strips.

v2: make provoking vertex tables static const

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
8 years agost/mesa: revalidate image atoms when a texture is updated
Ilia Mirkin [Sat, 4 Jun 2016 17:25:35 +0000 (13:25 -0400)]
st/mesa: revalidate image atoms when a texture is updated

A texture may be redefined with _NEW_TEXTURE, which might have been
bound to a shader image slot. We have to revalidate the image atoms to
pick up on the new resource.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
8 years agogk104/ir: fix conditions for adding a texbar
Ilia Mirkin [Tue, 7 Jun 2016 01:25:05 +0000 (21:25 -0400)]
gk104/ir: fix conditions for adding a texbar

Sometimes a register source can actually be double- or even quad-wide.
We must make sure that the inserted texbars take that width into
account.

Based on an earlier patch by Samuel Pitoiset.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Cc: "12.0 11.2" <mesa-stable@lists.freedesktop.org>
8 years agoradeonsi: keep track of dirty descriptor sets
Nicolai Hähnle [Fri, 3 Jun 2016 15:40:12 +0000 (17:40 +0200)]
radeonsi: keep track of dirty descriptor sets

Reduces CPU load for draw calls that change none or few of the descriptors.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: move si_descriptors into a per-context array
Nicolai Hähnle [Fri, 3 Jun 2016 13:56:39 +0000 (15:56 +0200)]
radeonsi: move si_descriptors into a per-context array

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: pass shader stage to si_disable_shader_image
Nicolai Hähnle [Fri, 3 Jun 2016 13:36:45 +0000 (15:36 +0200)]
radeonsi: pass shader stage to si_disable_shader_image

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: access descriptor sets via local variables
Nicolai Hähnle [Fri, 3 Jun 2016 13:14:39 +0000 (15:14 +0200)]
radeonsi: access descriptor sets via local variables

This will simplify moving them to a per-context array.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: add si_set_rw_buffer to be used for internal descriptors
Nicolai Hähnle [Fri, 3 Jun 2016 13:27:09 +0000 (15:27 +0200)]
radeonsi: add si_set_rw_buffer to be used for internal descriptors

So that callers outside of si_descriptors.c need to worry less about the
details of descriptor handling.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: pass shader stage to si_set_shader_image
Nicolai Hähnle [Fri, 3 Jun 2016 13:04:40 +0000 (15:04 +0200)]
radeonsi: pass shader stage to si_set_shader_image

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: pass shader stage to si_set_sampler_view
Nicolai Hähnle [Fri, 3 Jun 2016 13:03:59 +0000 (15:03 +0200)]
radeonsi: pass shader stage to si_set_sampler_view

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: move descriptor set begin_new_cs handling into a separate function
Nicolai Hähnle [Fri, 3 Jun 2016 12:50:42 +0000 (14:50 +0200)]
radeonsi: move descriptor set begin_new_cs handling into a separate function

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: move enabled_mask out of si_descriptors
Nicolai Hähnle [Fri, 3 Jun 2016 12:47:10 +0000 (14:47 +0200)]
radeonsi: move enabled_mask out of si_descriptors

This mask is irrelevant for the generic descriptor set handling, and having it
outside simplifies subsequent changes slightly.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoanv/entrypoints: Stop using the C preprocessor
Jason Ekstrand [Mon, 6 Jun 2016 21:29:19 +0000 (14:29 -0700)]
anv/entrypoints: Stop using the C preprocessor

Now that we emit guards for everything, we can just generate the files and
trust build flags to keep us safe.  This should also fix the tarball
problems.

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoanv/entrypoints: Emit #if guards for all platforms
Jason Ekstrand [Mon, 6 Jun 2016 21:29:18 +0000 (14:29 -0700)]
anv/entrypoints: Emit #if guards for all platforms

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoplatform_android: prevent deadlock in droid_swap_buffers
Haixia Shi [Thu, 2 Jun 2016 19:48:23 +0000 (12:48 -0700)]
platform_android: prevent deadlock in droid_swap_buffers

To avoid blocking other EGL calls, release the display mutex before
we enqueue buffer to android frameworks and re-acquire the mutex
upon return.

v2: moved lock/unlock inside droid_window_enqueue_buffer().

TEST=verify pinch zoom in Photos app no longer causes hangs

Signed-off-by: Haixia Shi <hshi@chromium.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agomesa: automake: distclean git_sha1.h when building OOT
Emil Velikov [Mon, 6 Jun 2016 18:39:40 +0000 (19:39 +0100)]
mesa: automake: distclean git_sha1.h when building OOT

In the case of out-of-tree (OOT) builds, in particular when building
from tarball, we'll end up with the file in both srcdir and builddir.

We want the former to remain intact (since we need it on rebuild) while
the latter should be removed otherwise `make distclean' gets angry at
us.

Ideally there'll be a solution that feels a bit less of a hack. Until
then this does the job exactly as expected.

Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agomesa: automake: ensure that git_sha1.h.tmp has the right attributes
Emil Velikov [Mon, 6 Jun 2016 16:31:05 +0000 (17:31 +0100)]
mesa: automake: ensure that git_sha1.h.tmp has the right attributes

... when copied from git_sha1.h.

As the latter file can we lacking the write attribute, one should set it
explicitly. Otherwise we'll get a warning/failure at cleanup stage.

Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agomesa: automake: add directory prefix for git_sha1.h
Emil Velikov [Mon, 6 Jun 2016 15:50:14 +0000 (16:50 +0100)]
mesa: automake: add directory prefix for git_sha1.h

Otherwise the build will assume that we've talking about builddir, which
is not the case in the else statement.

Here the file is already generated and is part of the tarball.

Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: android: don't add the image loader extension for !render_node
Emil Velikov [Sat, 4 Jun 2016 00:09:14 +0000 (01:09 +0100)]
egl: android: don't add the image loader extension for !render_node

With earlier commit we introduced support for render_node devices, which
was couples with the use of the image loader extension.

As the work was inspired by egl/wayland we (erroneously) added the
extension for the !render_node path as well.

That works for wayland, as the implementations of the DRI2 and IMAGE
loader extensions converge behind the scenes. As that is not yet
the case for Android we shouldn't expose the extension.

Fixes: 34ddef39cef ("egl: android: add dma-buf fd support")
Cc: <mesa-stable@lists.freedesktop.org>
Reported-by: Mauro Rossi <issor.oruam@gmail.com>
Tested-by: Mauro Rossi <issor.oruam@gmail.com>
Acked-by: Rob Herring <robh@kernel.org>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
8 years agogallium/radeon: add support for sharing textures with DCC between processes
Marek Olšák [Thu, 2 Jun 2016 21:36:43 +0000 (23:36 +0200)]
gallium/radeon: add support for sharing textures with DCC between processes

v2: use a function for calculating WORD1 of bo metadata

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agogallium/radeon: don't discard DCC if an external user can write to it
Marek Olšák [Thu, 2 Jun 2016 21:30:01 +0000 (23:30 +0200)]
gallium/radeon: don't discard DCC if an external user can write to it

We don't import textures with DCC now, but soon we will.

v2: if we can't disable DCC for image writes, at least decompress DCC
    at bind time

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoi915: fix typo CAP.
Dave Airlie [Tue, 7 Jun 2016 08:30:54 +0000 (18:30 +1000)]
i915: fix typo CAP.

Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agoglsl: initialise pointer to NULL
Jakob Sinclair [Fri, 3 Jun 2016 23:09:52 +0000 (01:09 +0200)]
glsl: initialise pointer to NULL

Could cause issues if you tried to read from an uninitialised pointer.
This just initalises the pointer to null to avoid that being a problem.
Discovered by Coverity.

CID: 1343616

Signed-off-by: Jakob Sinclair <sinclair.jakob@openmailbox.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/gen8: fix cull distance emission for tessellation shaders.
Dave Airlie [Tue, 7 Jun 2016 00:27:44 +0000 (10:27 +1000)]
i965/gen8: fix cull distance emission for tessellation shaders.

This fixes some cases of:
GL45-CTS.cull_distance.functional
on Skylake.

Reviewed-by: Chris Forbes <chrisforbes@google.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
8 years agonvc0: add support for VOTE tgsi opcodes
Ilia Mirkin [Sun, 29 May 2016 16:42:49 +0000 (12:42 -0400)]
nvc0: add support for VOTE tgsi opcodes

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
8 years agost/mesa: expose GL_ARB_shader_group_vote when supported by backend
Ilia Mirkin [Sun, 29 May 2016 15:43:05 +0000 (11:43 -0400)]
st/mesa: expose GL_ARB_shader_group_vote when supported by backend

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
8 years agogallium: add PIPE_CAP_TGSI_VOTE for when the VOTE ops are allowed
Ilia Mirkin [Sun, 29 May 2016 15:39:52 +0000 (11:39 -0400)]
gallium: add PIPE_CAP_TGSI_VOTE for when the VOTE ops are allowed

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
8 years agogallium: add VOTE_* opcodes to implement GL_ARB_shader_group_vote
Ilia Mirkin [Sun, 29 May 2016 15:01:05 +0000 (11:01 -0400)]
gallium: add VOTE_* opcodes to implement GL_ARB_shader_group_vote

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
8 years agomesa: hook up core bits of GL_ARB_shader_group_vote
Ilia Mirkin [Sun, 29 May 2016 14:49:03 +0000 (10:49 -0400)]
mesa: hook up core bits of GL_ARB_shader_group_vote

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
8 years agoglsl: Make opt_copy_propagation_elements actually propagate into loops.
Kenneth Graunke [Sat, 30 Apr 2016 05:06:37 +0000 (22:06 -0700)]
glsl: Make opt_copy_propagation_elements actually propagate into loops.

We've had a FINISHME here since Eric originally wrote the code in 2011.
This patch implements his suggested approach, which makes us actually
able to copy propagate into the loops, at the unfortunate cost of making
this pass even more expensive.

The shader-db statistics are basically a wash:

   No change in instruction counts.

   total cycles in shared programs: 78685980 -> 78680730 (-0.01%)
   cycles in affected programs: 2102646 -> 2097396 (-0.25%)
   helped: 48
   HURT: 83

I figured if we're going to do this for one copy propagation pass,
we may as well do it in both.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoglsl: Make opt_copy_propagation actually propagate into loops.
Kenneth Graunke [Sat, 30 Apr 2016 05:06:37 +0000 (22:06 -0700)]
glsl: Make opt_copy_propagation actually propagate into loops.

We've had a FINISHME here since Eric originally wrote the code in 2010.
This patch implements his suggested approach, which makes us actually
able to copy propagate into the loops, at the unfortunate cost of making
this pass even more expensive.

The shader-db statistics are not terribly impressive:

   total instructions in shared programs: 9008589 -> 9008613 (0.00%)
   instructions in affected programs: 4293 -> 4317 (0.56%)
   helped: 0
   HURT: 10

   total cycles in shared programs: 78550978 -> 78575760 (0.03%)
   cycles in affected programs: 655426 -> 680208 (3.78%)
   helped: 75
   HURT: 88

   GAINED: 2

Most of the "regressions" appear to be us successfully copy propagating
uniforms, which i965 is loading as pull constants instead of push, so we
occasionally have two pulls instead of one.  That doesn't seem like this
pass's job - it's propagating correctly, and we should be smarter about
pull loads in the backend.

This patch is also useful for a couple of reasons:

1. It can clean up copies created by varying packing (previously, we
   couldn't if the uses were inside a loop).

   This fixes a bug when interpolateAt*() is used on a packed varying
   inside a loop: glsl_to_nir struggles to see through the extra copy
   and mistakenly believed the variable was not an input.

2. It will help propagate uniform array access created by
   lower_const_array_to_uniforms().

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agonv50/ir: use round toward 0 when converting doubles to integers
Samuel Pitoiset [Mon, 6 Jun 2016 19:12:15 +0000 (21:12 +0200)]
nv50/ir: use round toward 0 when converting doubles to integers

Like floats, we should use the round toward 0 mode instead of the
nearest one (which is the default) for doubles to integers.

This fixes all arb_gpu_shader_fp64 piglits which convert doubles to
integers (16 tests).

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>