platform/upstream/mesa.git
5 years agofreedreno/a5xx: Fix indirect draw max_indices calculation
Eduardo Lima Mitev [Mon, 10 Jun 2019 21:27:22 +0000 (23:27 +0200)]
freedreno/a5xx: Fix indirect draw max_indices calculation

The number of elements to draw should not be affected by the offset.

A similar fix was submitted for a6xx at 79180a05.

Fixes these dEQP tests on a5xx:

dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawelements_separate_grid_500x500_drawcount_8
dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawelements_separate_grid_500x500_drawcount_2500
dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawarrays_separate_grid_500x500_drawcount_2500
dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawarrays_combined_grid_500x500_drawcount_2500
dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawelements_combined_grid_500x500_drawcount_8
dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawelements_combined_grid_500x500_drawcount_2500

Reviewed-by: Rob Clark <robdclark@gmail.com>
5 years agoradv: remove extra assignment in radv_decompress_resolve_subpass_src()
Samuel Pitoiset [Tue, 11 Jun 2019 06:17:22 +0000 (08:17 +0200)]
radv: remove extra assignment in radv_decompress_resolve_subpass_src()

baseArrayLayer is defined twice, trivial.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
5 years agoradv: add radv_get_resolve_pipeline() helper in the graphics path
Samuel Pitoiset [Mon, 10 Jun 2019 15:45:33 +0000 (17:45 +0200)]
radv: add radv_get_resolve_pipeline() helper in the graphics path

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: do not decompress all image layers before resolving inside a subpass
Samuel Pitoiset [Mon, 10 Jun 2019 15:45:32 +0000 (17:45 +0200)]
radv: do not decompress all image layers before resolving inside a subpass

When decompressing resolve source images, we should rely on the
framebuffer layer count instead of resolving all images layers.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: initialize the aspect mask when decompressing resolve source images
Samuel Pitoiset [Mon, 10 Jun 2019 15:45:31 +0000 (17:45 +0200)]
radv: initialize the aspect mask when decompressing resolve source images

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: perform proper layout transitions before resolving
Samuel Pitoiset [Mon, 10 Jun 2019 15:45:30 +0000 (17:45 +0200)]
radv: perform proper layout transitions before resolving

Use an explicit pipeline barrier for doing layout transitions
instead of duplicating some code.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: do not resolve all image layers with compute inside a subpass
Samuel Pitoiset [Mon, 10 Jun 2019 15:45:29 +0000 (17:45 +0200)]
radv: do not resolve all image layers with compute inside a subpass

When resolving inside a subpass, we should rely on the framebuffer
layer count instead of resolving all images layers. This should
improve performance of layered resolves a bit.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoiris: Bypass half-float pack/unpack lowering.
Kenneth Graunke [Mon, 10 Jun 2019 21:03:03 +0000 (14:03 -0700)]
iris: Bypass half-float pack/unpack lowering.

This skips GLSL IR lowering of pack/unpackHalf operations, allowing
the NIR optimizer to see them

Improves performance in Synmark2's OglCSDof by about 2x, by cutting
about 90% of the cycles from one of the compute shaders.

shader-db statistics on Skylake:

4 compute shaders went from SIMD8 to SIMD16.

total instructions in shared programs: 15598871 -> 15542568 (-0.36%)
instructions in affected programs: 143016 -> 86713 (-39.37%)
helped: 144
HURT: 0
helped stats (abs) min: 17 max: 4669 x̄: 390.99 x̃: 164
helped stats (rel) min: 7.48% max: 85.28% x̄: 30.17% x̃: 24.22%
95% mean confidence interval for instructions value: -510.50 -271.49
95% mean confidence interval for instructions %-change: -32.70% -27.65%
Instructions are helped.

total cycles in shared programs: 371973958 -> 368902103 (-0.83%)
cycles in affected programs: 5557722 -> 2485867 (-55.27%)
helped: 144
HURT: 0
helped stats (abs) min: 106 max: 1026600 x̄: 21332.33 x̃: 1697
helped stats (rel) min: 0.53% max: 88.98% x̄: 36.12% x̃: 34.67%
95% mean confidence interval for cycles value: -41570.02 -1094.64
95% mean confidence interval for cycles %-change: -38.44% -33.80%
Cycles are helped.

total spills in shared programs: 11936 -> 11903 (-0.28%)
spills in affected programs: 110 -> 77 (-30.00%)
helped: 3
HURT: 2

total fills in shared programs: 25644 -> 25178 (-1.82%)
fills in affected programs: 677 -> 211 (-68.83%)
helped: 5
HURT: 0

total loops in shared programs: 4830 -> 4829 (-0.02%)
loops in affected programs: 1 -> 0
helped: 1
HURT: 0

5 years agoradv: Handle UNDEFINED format in image format list.
Bas Nieuwenhuizen [Sat, 8 Jun 2019 21:51:16 +0000 (23:51 +0200)]
radv: Handle UNDEFINED format in image format list.

Was watching a presentation on YT where this was used and it turns
out it is not invalid.

The only case it is actually valid as format in the creation of an
image or image view is with Android Hardware Buffers which have
their format specified externally.

So we can just ignore all entries with VK_FORMAT_UNDEFINED.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
5 years agoradv: Prevent out of bound shift on 32-bit builds.
Bas Nieuwenhuizen [Mon, 10 Jun 2019 14:17:46 +0000 (16:17 +0200)]
radv: Prevent out of bound shift on 32-bit builds.

uintptr_t is 32-bits then and shifting it by 32 bits results in undefined
behavior IIRC.

Fixes: b3c8de1c55c "radv: save all descriptor pointers into the trace BO"
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
5 years agoglsl: Check order and uniqueness of interlock functions
Caio Marcelo de Oliveira Filho [Wed, 5 Jun 2019 08:25:24 +0000 (01:25 -0700)]
glsl: Check order and uniqueness of interlock functions

With this commit all remaining compilation tests in Piglit for
ARB_fragment_shader_interlock will pass.

Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Plamena Manolova <plamena.manolova@intel.com>
5 years agoglsl: Make interlock builtins follow same compiler rules as barriers
Caio Marcelo de Oliveira Filho [Wed, 5 Jun 2019 07:59:11 +0000 (00:59 -0700)]
glsl: Make interlock builtins follow same compiler rules as barriers

Generalize the barrier code to provide correct error messages for
other builtins.

Fixes most of piglit compilation tests for
ARB_fragment_shader_interlock.

Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Plamena Manolova <plamena.manolova@intel.com>
5 years agonir/opt_algebraic: Fix rules for imadsh_mix16
Eduardo Lima Mitev [Mon, 10 Jun 2019 19:38:39 +0000 (21:38 +0200)]
nir/opt_algebraic: Fix rules for imadsh_mix16

The rules added in patch 3addd7c are inverted:

It should be:

(al * bh) << 16 + c

instead of:

(ah * bl) << 16 + c

Fixes a number of regressions under
dEQP-GLES31.functional.draw_indirect.compute_interop.large.*
on Freedreno.

Reviewed-by: Rob Clark <robdclark@gmail.com>
5 years agopanfrost: Ignore discards in dead branch analysis
Alyssa Rosenzweig [Mon, 10 Jun 2019 15:21:24 +0000 (08:21 -0700)]
panfrost: Ignore discards in dead branch analysis

Fixes regressions in
dEQP-GLES2.functional.shaders.discard.dynamic_loop_*

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agoradv: fix setting CB_SHADER_MASK for dual source blending
Samuel Pitoiset [Thu, 6 Jun 2019 12:46:47 +0000 (14:46 +0200)]
radv: fix setting CB_SHADER_MASK for dual source blending

CB_SHADER_MASK was computed without the second color buffer
format which looks totally wrong to me.

While we are at it, copy a comment from RadeonSI.

Cc: 19.0 19.1 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agopanfrost/midgard: Disambiguate register mode
Alyssa Rosenzweig [Wed, 5 Jun 2019 22:43:52 +0000 (15:43 -0700)]
panfrost/midgard: Disambiguate register mode

We postfix instructions by their size if a destination override is in
place (a la AT&T assembly), disambiguating instruction sizes.
Previously, "16-bit instruction, 16-bit dest, 16-bit sources"
disassembled identically to "32-bit instruction, 16-bit dest, 16-bit
sources", which is semantically distinct due to the lessened opportunity
for parallelism but (potentially) greater precision. Adding a postfix
removes the ambiguity and relieves mental gymnastics reading weird
disassemblies even in some cases that are not ambiguous.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Expose vec8/vec16 modes
Alyssa Rosenzweig [Wed, 5 Jun 2019 22:41:03 +0000 (15:41 -0700)]
panfrost/midgard: Expose vec8/vec16 modes

Midgard ALUs can operate in one of four modes: vec2 64-bit, vec4 32-bit,
vec8 16-bit, or vec16 8-bit. Our compiler (and indeed, any OpenGL ES
shader) only uses 32-bit (and eventually vec4 16-bit) modes in normal
circumstances. Nevertheless, the other modes do exist and are easily
accessible through OpenCL; they also come up in cases like blend
shaders.

While we have had minimal support for decoding 8-bit/64-bit modes, we
did so pretending they were vec4 in each case; 16-bit registers had a
synthetically duplicated register file to separate lo/hi halves, etc.
This works for GL, but it doesn't map to what the hardware is -actually-
doing, which can cause some headscratchingly bizarre disassemblies from
OpenCL. So, we dive in the deep end and support these other modes
natively in the disassembler, using absurdly long masks/swizzles, since
the hardware is considerably more flexible than what was exposed before.

Outside of some fixed routines for blending, none of the above is
supported in the compiler yet. But it's better to have it in the ISA
definitions and disassembler than not, for future use if nothing else.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Add shifting int modifiers
Alyssa Rosenzweig [Wed, 5 Jun 2019 22:20:26 +0000 (15:20 -0700)]
panfrost/midgard: Add shifting int modifiers

As a source modifier, shift allows shifting a value left by the bit
size, useful in conjunction with a greater register mode, for instance
to implement `upsample`. As a concrete example, the following OpenCL:

   ushort hr0 = /* ... */, uint r1 = /* ... */;
   uint r2 = (convert_uint(hr0) << 16) ^ b;

compiles to the following Midgard assembly:

   ixor r, (hr0) << 16, b

In reverse, the ".hi" output modifier shifts the value right by the bit
size, leaving just the carry/overflow at the bottom. To implement *_hi
functions in OpenCL (for <64-bit), we do arithmetic in the 2x higher
mode with the .hi modifier. (For 64-bit, things are hairier, since there
is not an 128-bit int mode).

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Add integer outmods
Alyssa Rosenzweig [Wed, 5 Jun 2019 22:17:45 +0000 (15:17 -0700)]
panfrost/midgard: Add integer outmods

For floats, output modifiers determine clamping behaviour. For integers,
they determine wrapping/saturation behaviour (or shifting -- see next
commit). These are very different; they are conceptually two unrelated
enums union'ed together; the distinction is responsible for many-a-bug.
While clamping behaviour for floats was clear from GL, the int behaviour
is only known From OpenCL contortion with convert_*_sat() functions.

With the underlying functions known, clean up the codebase, likely
fixing outmod type related bugs in the process.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Note floating compares type convert
Alyssa Rosenzweig [Wed, 5 Jun 2019 22:02:20 +0000 (15:02 -0700)]
panfrost/midgard: Note floating compares type convert

OP_TYPE_CONVERTS denotes an opcode that returns a different type than is
source (going from int-domain to float-domain or vice versa), named
after the f2i/i2f family of opcodes it covers. We care because source
mods are determined by the source type (i/f) but output modifiers are
determined by the output type (equals the source type, unless the op
type converts, in which case it's the opposite).

The upshot is that floating-point compares (feq/fne/etc) actually do
type-convert.  That is, that take in floating-points and output in
integer space (a boolean), so we mark them off this way to ensure the
correct output modifiers are used.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Align linear renderable resources
Alyssa Rosenzweig [Thu, 6 Jun 2019 21:36:41 +0000 (14:36 -0700)]
panfrost: Align linear renderable resources

It's just -easier- to render to aligned framebuffers. For winsys
targets, we already align, but even for an internal linear FBO we ought
to align everything nicely.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Fix stride check when mipmapping
Alyssa Rosenzweig [Sat, 8 Jun 2019 00:07:13 +0000 (17:07 -0700)]
panfrost: Fix stride check when mipmapping

Now that we support custom strides on mipmapped textures (theoretically,
at least), extend the stride check to support mipmaps.  Fixes incorrect
strides of linear windows in Weston.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
5 years agopanfrost: Refactor texture/sampler upload
Alyssa Rosenzweig [Fri, 7 Jun 2019 21:25:28 +0000 (14:25 -0700)]
panfrost: Refactor texture/sampler upload

We move some coding packing the texture/sampler descriptors into
dedicated functions (out of the terrifyingly long emit_for_draw
monolith), cleaning them up as we go.

The discovery triggering the cleanup is the format for including manual
strides in the presence of mipmaps/cubemaps. Rather than placed at the
end like previously assumed, they are interleaved after each address.
This difference is relevant when handling NPOT linear mipmaps.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Refactor blitting code
Alyssa Rosenzweig [Fri, 7 Jun 2019 17:32:17 +0000 (10:32 -0700)]
panfrost: Refactor blitting code

We refactor the wallpaper rendering code to separate the
wallpaper-specific bits from the general blitting capabilities. In the
(hopefully near) future, we'll turn this on to implement real Gallium
blits, e.g. for automatic mipmap generation.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Refactor AFBC code
Alyssa Rosenzweig [Fri, 7 Jun 2019 16:39:31 +0000 (09:39 -0700)]
panfrost: Refactor AFBC code

This patch does a substantial cleanup of the code for handling AFBC,
moving various disparate misplaced functions into a new central
pan_afbc.c file.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Move pan_screen() to pan_screen.h
Alyssa Rosenzweig [Fri, 7 Jun 2019 16:49:36 +0000 (09:49 -0700)]
panfrost: Move pan_screen() to pan_screen.h

Trivial.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost: Always align strides to cache line (64)
Alyssa Rosenzweig [Fri, 7 Jun 2019 15:58:16 +0000 (08:58 -0700)]
panfrost: Always align strides to cache line (64)

(Performance tweak.)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agodocs: fixup 19.0.5 <> 19.0.6 confusion
Emil Velikov [Fri, 7 Jun 2019 16:11:23 +0000 (17:11 +0100)]
docs: fixup 19.0.5 <> 19.0.6 confusion

The title of the release notes says 19.0.5 while the rest of the file
(correctly) says 19.0.6

Fixes: fe79d75ccf9 ("docs: Add relnotes for 19.0.6")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Dylan Baker <dylan at pnwbakers.com>
5 years agomapi: correctly handle the full offset table
Emil Velikov [Wed, 5 Jun 2019 15:45:03 +0000 (16:45 +0100)]
mapi: correctly handle the full offset table

Earlier commit converted ES1 and ES2 to a new, much simpler, dispatch
generator. At the same time, GL/glapi and the driver side are still
using the old code.

There is a hidden ABI between GL*.so and glapi.so, former referencing
entry-points by offset in the _glapi_table. Hence earlier commit added
the full table of entry-points, alongside a marker for other cases like
indirect GL(X) and driver-size remapping.

Yet the patches did not handle things fully, thus it was possible to
get different interpretations of the dispatch table after the marker.

This commit fixes that adding an indicative error message to catch
future bugs.

While here correct the marker (MAX_OFFSETS) comment.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110302
Fixes: cf317bf0937 ("mapi: add all _glapi_table entrypoints tostatic_data.py")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
5 years agomapi: add static_date offset to EXT_dsa
Emil Velikov [Fri, 7 Jun 2019 15:19:59 +0000 (16:19 +0100)]
mapi: add static_date offset to EXT_dsa

As elaborated in the next patch, there is some hidden ABI that
effectively require most entrypoints to be listed in the file.

Cc: Marek Olšák <marek.olsak@amd.com>
Fixes: d2906293c43 ("mesa: EXT_dsa add selectorless matrix stackfunctions")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
5 years agomapi: add static_date offset to MaxShaderCompilerThreadsKHR
Emil Velikov [Wed, 5 Jun 2019 16:14:20 +0000 (17:14 +0100)]
mapi: add static_date offset to MaxShaderCompilerThreadsKHR

As elaborated in the next patch, there is some hidden ABI that
effectively require most entrypoints to be listed in the file.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110302
Cc: Marek Olšák <maraeo@gmail.com>
Fixes: c5c38e831ee ("mesa: implement ARB/KHR_parallel_shader_compile")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
5 years agoegl: Let the caller of dri2_create_drawable decide about loaderPrivate.
Mathias Fröhlich [Fri, 7 Jun 2019 05:12:42 +0000 (07:12 +0200)]
egl: Let the caller of dri2_create_drawable decide about loaderPrivate.

In the call arguments to dri2_create_drawable decouple loaderPrivate
from dri2_surf. For all callers of dri2_create_drawable the two
pointers are the same with the exception of the gbm backed platform.
Let the calling code of dri2_create_drawable decide what
loaderPrivate shall be.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de>
5 years agoradv: fix alpha-to-coverage when there is unused color attachments
Samuel Pitoiset [Thu, 6 Jun 2019 14:31:01 +0000 (16:31 +0200)]
radv: fix alpha-to-coverage when there is unused color attachments

When alphaToCoverage is enabled, we should always write the alpha
channel of MRT0 if it's unused. This now matches RadeonSI.

This fixes the new CTS:
dEQP-VK.pipeline.multisample.alpha_to_coverage_unused_attachment.samples_*.alpha_invisible

Cc: 19.0 19.1 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl
5 years agopanfrost: ci: Switch from direct Docker use to buildah
Tomeu Vizoso [Fri, 7 Jun 2019 08:20:28 +0000 (10:20 +0200)]
panfrost: ci: Switch from direct Docker use to buildah

Use the infrastructure in wayland/ci-templates to build the container
images.

This prevents from getting into some situations in which the images
wouldn't be rebuilt, and allows us to share some infrastructure with
other projects in freedesktop.org.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Suggested-by: Michel Dänzer <michel@daenzer.net>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agogallium/u_transfer_helper: Free the staging buffer on unmap.
Kenneth Graunke [Fri, 7 Jun 2019 08:16:16 +0000 (01:16 -0700)]
gallium/u_transfer_helper: Free the staging buffer on unmap.

u_transfer_helper sometimes mallocs a staging buffer, and leaked it.

Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
5 years agointel/gpu_dump: fix argument passing
Lionel Landwerlin [Sat, 8 Jun 2019 20:48:02 +0000 (23:48 +0300)]
intel/gpu_dump: fix argument passing

We were dropping "/' around arguments grouped together.
This was triggering failures with :

   $ ./framemetrics -g "Memory Writes Distribution Gen9" -o /tmp/output.csv -f ./my.trace 10 11

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Engestrom <eric@engestrom.ch>
5 years agoutil/os_file: suppress sign comparison warning
Eric Engestrom [Thu, 16 May 2019 14:37:28 +0000 (15:37 +0100)]
util/os_file: suppress sign comparison warning

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoutil/os_file: fix error being sign-cast back and forth
Eric Engestrom [Thu, 16 May 2019 12:08:53 +0000 (13:08 +0100)]
util/os_file: fix error being sign-cast back and forth

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoutil/os_file: avoid shadowing read() with a local variable
Eric Engestrom [Thu, 16 May 2019 14:02:45 +0000 (15:02 +0100)]
util/os_file: avoid shadowing read() with a local variable

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoutil/os_file: actually return the error read() gave us
Eric Engestrom [Thu, 16 May 2019 13:57:07 +0000 (14:57 +0100)]
util/os_file: actually return the error read() gave us

Fixes: 316964709e21286c2af5 "util: add os_read_file() helper"
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agovirgl: Work around possible memory exhaustion
Alexandros Frantzis [Wed, 5 Jun 2019 13:50:11 +0000 (16:50 +0300)]
virgl: Work around possible memory exhaustion

Since we don't normally flush before performing copy transfers, it's
possible in some scenarios to use too much memory for staging resources
and start failing. This can happen either because we exhaust the total
available memory (including system memory virtio-gpu swaps out to), or,
more commonly, because the total size of resources in a command buffer
doesn't fit in virtio-gpu video memory.

To reduce the chances of this happening, force a flush before a copy
transfer if the total size of queued staging resources exceeds a certain
limit. Since after a flush any queued staging resources will be
eventually released, this ensures both that each command buffer doesn't
require too much video memory, and that we don't end up consuming too
much memory for staging resources in total.

Fixes kernel errors reported when running texture_upload tests in glbench.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Remove incorrect resource wait condition
Alexandros Frantzis [Mon, 27 May 2019 21:06:03 +0000 (00:06 +0300)]
virgl: Remove incorrect resource wait condition

Now that we have copy transfers in place, we can remove the incorrect
resource wait condition. Copy transfers and other optimizations minimize
the performance impact of this removal, while providing the correct
behavior.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Use copy transfers for textures
Alexandros Frantzis [Fri, 24 May 2019 11:03:28 +0000 (14:03 +0300)]
virgl: Use copy transfers for textures

Extend copy transfers to also be used for busy textures.

Performance results:
Unigine Valley, qemu before: 22.7 FPS after: 23.1 FPS

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Use buffer copy transfers to avoid waiting when mapping
Alexandros Frantzis [Wed, 8 May 2019 09:10:21 +0000 (12:10 +0300)]
virgl: Use buffer copy transfers to avoid waiting when mapping

We typically need to wait for a buffer to become ready before mapping,
so that we don't write new contents while the host is still using the
old contents. However, if we are allowed to discard the contents of the
mapped buffer range, then we can avoid waiting by using a staging buffer
range which we guarantee to never be busy, copying from the staging
buffer range to the target buffer in the host.

This commit implements this optimization by utilizing a dedicated
u_upload_mgr for the staging buffer.

Performance results:
Twilight Struggle (Steam/Proton), qemu before: 7 FPS after: 25 FPS
glmark2 ubo, qemu before: 38 FPS after: 331 FPS

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Suggested-by: Gurchetan Singh <gurchetansingh@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Support copy transfers
Alexandros Frantzis [Wed, 8 May 2019 13:17:53 +0000 (16:17 +0300)]
virgl: Support copy transfers

Support transfers that use a different resource as the source of data to
transfer. This will be used in upcoming commits to send data to host
buffers through a transfer upload buffer, in order to avoid waiting
when the buffer resource is busy.

Note that we don't support queueing copy transfers in the transfer
queue. Copy transfers should be emitted directly in the command queue,
allowing us to avoid flushes before them and leads to better
performance.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Add copy_transfer3d definitions
Alexandros Frantzis [Tue, 14 May 2019 09:38:24 +0000 (12:38 +0300)]
virgl: Add copy_transfer3d definitions

Introduce definitions for the copy_transfer3d protocol command and virgl
capability. This command transfers data to the host by copying through
another resource, and will be used in upcoming commits to avoid waiting
when transferring data for busy resources.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Make VIRGL_BIND_STAGING resources cacheable
Alexandros Frantzis [Tue, 4 Jun 2019 13:43:31 +0000 (16:43 +0300)]
virgl: Make VIRGL_BIND_STAGING resources cacheable

This could help performance when trying to recreate such resources for
copy transfers.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Support VIRGL_BIND_STAGING
Alexandros Frantzis [Mon, 20 May 2019 10:00:38 +0000 (13:00 +0300)]
virgl: Support VIRGL_BIND_STAGING

Support a new virgl bind type for staging buffers which don't require
dedicated host-side storage. These will be used to implement copy
transfers.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Avoid unfinished transfer_get with PIPE_TRANSFER_DONTBLOCK
Alexandros Frantzis [Thu, 23 May 2019 18:16:48 +0000 (21:16 +0300)]
virgl: Avoid unfinished transfer_get with PIPE_TRANSFER_DONTBLOCK

If we are not allowed to block, and we know that we will have to wait,
either because the resource is busy, or because it will become busy due
to a readback, return early to avoid performing an incomplete
transfer_get. Such an incomplete transfer_get may finish at any time,
during which another unsynchronized map could write to the resource
contents, leaving the contents in an undefined state.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Suggested-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Deduplicate checks for resource caching
Alexandros Frantzis [Tue, 4 Jun 2019 13:40:33 +0000 (16:40 +0300)]
virgl: Deduplicate checks for resource caching

Also fixes a missed check for VIRGL_BIND_CUSTOM in one of the duplicate
code snippets.

Note that legacy fences also use VIRGL_BIND_CUSTOM, but we ensured they
don't go through the cache in the previous commit.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: Don't try to use cached resources for legacy fences
Alexandros Frantzis [Wed, 5 Jun 2019 07:32:01 +0000 (10:32 +0300)]
virgl: Don't try to use cached resources for legacy fences

Resources for fences should not be from the cache, since we are basing
the fence status on the resource creation busy status.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: More info about chosen alignment value
Alexandros Frantzis [Thu, 23 May 2019 11:58:46 +0000 (14:58 +0300)]
virgl: More info about chosen alignment value

Add more info about why the value of VIRGL_MAP_BUFFER_ALIGNMENT.

Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
5 years agovirgl: store all info about atomic buffers
Chia-I Wu [Thu, 16 May 2019 22:01:36 +0000 (15:01 -0700)]
virgl: store all info about atomic buffers

We will need the full info.  This also speeds up
virgl_attach_res_atomic_buffers and fixes resource leaks when the
context is destroyed.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
5 years agovirgl: add shader images to virgl_shader_binding_state
Chia-I Wu [Thu, 16 May 2019 21:33:15 +0000 (14:33 -0700)]
virgl: add shader images to virgl_shader_binding_state

It replaces virgl_context::images.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
5 years agovirgl: add SSBOs to virgl_shader_binding_state
Chia-I Wu [Thu, 16 May 2019 21:33:15 +0000 (14:33 -0700)]
virgl: add SSBOs to virgl_shader_binding_state

It replaces virgl_context::ssbos.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
5 years agovirgl: add UBOs to virgl_shader_binding_state
Chia-I Wu [Thu, 16 May 2019 21:00:54 +0000 (14:00 -0700)]
virgl: add UBOs to virgl_shader_binding_state

It replaces virgl_context::ubos.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
5 years agovirgl: add virgl_shader_binding_state
Chia-I Wu [Thu, 16 May 2019 20:32:18 +0000 (13:32 -0700)]
virgl: add virgl_shader_binding_state

virgl_shader_binding_state will be used to manage all per-stage
shader bindings.  For now, it manages only sampler views.

This replaces virgl_textures_info and fixes some issues

 - start_slot is now honored
 - views outside of [start_slot, slart_slot+count) are unmodified
 - views are released when the context is destroyed

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
5 years agoiris: Zero shs->cbuf0 when binding a passthrough TCS
Kenneth Graunke [Fri, 7 Jun 2019 19:41:28 +0000 (12:41 -0700)]
iris: Zero shs->cbuf0 when binding a passthrough TCS

Fixes valgrind errors when running two CTS tests back to back:
- KHR-GL45.shader_image_load_store.basic-allTargets-loadStoreT*
(The first test has an actual TCS, the second uses passthrough.)

5 years agointel/blorp: Only double the fast-clear rect alignment on HSW
Jason Ekstrand [Fri, 7 Jun 2019 20:13:30 +0000 (15:13 -0500)]
intel/blorp: Only double the fast-clear rect alignment on HSW

This restriction was accidentally added to the BSpec/PRM as an
unrestricted restriction starting with the HSW docs and it was never
removed.  However, it only ever applied to HSW and actually potentially
causes problems on BDW and above where we have mipmapped fast-clears.

Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
5 years agofreedreno/a6xx: re-arrange program stageobj/group
Rob Clark [Thu, 6 Jun 2019 16:55:33 +0000 (09:55 -0700)]
freedreno/a6xx: re-arrange program stageobj/group

Split out a separate program config state group to run early before the
other groups.

This seems to help w/ intermittent "missed tiles" (although I had
assumed that was a mem2gmem issue), or at least I can't reproduce that
issue with this patch, but can without.

It has the benefit of HLSQ_VS_CNTL.CONSTLEN matching for VS and BS.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
5 years agofreedreno/a6xx: fix hangs with newer sqe fw
Rob Clark [Thu, 6 Jun 2019 17:22:04 +0000 (10:22 -0700)]
freedreno/a6xx: fix hangs with newer sqe fw

With the newer (v1.76) fw, we were getting hangs (compared to older
v1.66 fw).  Re-work the GMEM code to structure things a bit closer to
the blob.  This moves some PKT7 packets from IB2 to IB1, which I think
is what was confusing SQE and causing it to get stuck in an infinite
loop.  But in general structuring things at least closer to the same way
blob does makes it easier to compare cmdstream.

Note: this is a bit on the large side for what I'd normally consider for
stable.. but right now it is looking  like it is the newer fw that is
headed for linux-firmware.  This should defn have some soak time on
master, but probably a good idea for this patch to end up in distro mesa
builds by the time a630_sqe.fw hits linux-firmware.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
5 years agofreedreno/a6xx: WFI before RB_CCU_CNTL writes
Rob Clark [Thu, 6 Jun 2019 17:19:07 +0000 (10:19 -0700)]
freedreno/a6xx: WFI before RB_CCU_CNTL writes

This seems to be in a block of non buffered/context regs.  Blob always
WFIs before write, so probably a good idea.

Annoyingly, compared to ealier gens, it is a bit harder to tell from the
register offset whether it is a buffered reg, it isn't as simple as
everything below 0x2000, it seems.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
5 years agofreedreno/a6xx: don't pre-dispatch texture fetch on accident
Rob Clark [Thu, 6 Jun 2019 16:53:15 +0000 (09:53 -0700)]
freedreno/a6xx: don't pre-dispatch texture fetch on accident

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
5 years agofreedreno/a6xx: fix issues with gallium HUD
Rob Clark [Thu, 6 Jun 2019 16:45:25 +0000 (09:45 -0700)]
freedreno/a6xx: fix issues with gallium HUD

In some cases the draw for the text wasn't working.  This seems to be
fixed by resyncing some of the "golded registers" from blob (initial
values were based on somewhat older blob version).

Perhaps good to have a bit of soak time on master, but would be good
to eventually land in 19.x stable branches.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
5 years agoanv/cmd_buffer: Initalize the clear color struct for CNL+
Nanley Chery [Wed, 24 Oct 2018 21:50:32 +0000 (14:50 -0700)]
anv/cmd_buffer: Initalize the clear color struct for CNL+

On CNL+, the clear color struct is composed of RGBA channel values and
fields which are either reserved by the HW or used to control
fast-clears. Currently anv initializes the channel values to zero and
allows the other fields to be undefined.

Satisfy the MBZ field requirements by removing an optimization that
doesn't hold true for CNL+ and pulling in the number of dwords to
initialize from ISL.

Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
5 years agoglx/windows: Fix compilation with -Werror-format
Jon Turney [Thu, 6 Jun 2019 15:44:08 +0000 (16:44 +0100)]
glx/windows: Fix compilation with -Werror-format

Fix compilation where the DWORD type is used with a format, after
-Werror-format added by c9c1e261.

Some Win32 API types are different fundamental types in the 32-bit and
64-bit versions. This problem is then further compounded by the fact
that whilst both 32-bit Cygwin and 32-bit MinGW use the ILP32 data
model, 64-bit MinGW uses the LLP64 data model, but 64-bit Cygwin uses
the LP64 data model. This makes it near impossible to write printf
format specifiers which are correct for all those targets.

In the Win32 API, DWORD is an unsigned, 32-bit type.  So, it is defined
in terms of an unsigned long, except in the LP64 data model used by
64-bit Cygwin, where it is an unsigned int.

It should always be safe to cast it to unsigned int and use %u or %x.

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoiris: Rename bind_state to bind_shader_state.
Kenneth Graunke [Fri, 7 Jun 2019 18:26:20 +0000 (11:26 -0700)]
iris: Rename bind_state to bind_shader_state.

bind_state is possibly the worst name ever.  For create, we used
create_shader_state, which is more descriptive.  Put shader in the name.

5 years agoisl: Mark enum isl_channel_select packed so it becomes 1 byte.
Kenneth Graunke [Fri, 7 Jun 2019 00:36:09 +0000 (17:36 -0700)]
isl: Mark enum isl_channel_select packed so it becomes 1 byte.

I recently discovered that the following code lead to valgrind errors:

   struct isl_swizzle swizzle = ISL_SWIZZLE_IDENTITY;
   VALGRIND_CHECK_MEM_IS_DEFINED(&swizzle, sizeof(swizzle));

which is surprising, because struct isl_swizzle is simply:

   struct isl_swizzle {
      enum isl_channel_select r:4;
      enum isl_channel_select g:4;
      enum isl_channel_select b:4;
      enum isl_channel_select a:4;
   };

and the above code initializes all of them with a C99 initializer.
Iván Briano reminded me that C99 initializers don't necessarily zero
padding.  A quick inspection revealed that sizeof(struct isl_swizzle)
was 4 (rather than the expected 2).  Ian Romanick suggested changing
it to uint16_t, since this is essentially dicing up an unsigned, and
that worked.

This patch marks enum isl_channel_select packed, changing its size
from 4 bytes to 1 byte.  This then makes struct isl_swizzle 2 bytes,
with no bogus padding fields.  This eliminates valgrind undefined
memory warnings.

These isl_swizzle values become part of our BLORP blit program keys,
which are then hashed.  This undefined padding was being included in
the hashing, possibly leading to issues.  I originally saw this error
when running KHR-GL45.texture_size_promotion.functional in iris under
valgrind.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
5 years agopanfrost/ci: Texture wrap tests are legitimately fixed
Alyssa Rosenzweig [Thu, 6 Jun 2019 19:04:16 +0000 (12:04 -0700)]
panfrost/ci: Texture wrap tests are legitimately fixed

These depended on the wallpaper reload.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Lower inot to inor with 0
Alyssa Rosenzweig [Thu, 6 Jun 2019 18:58:57 +0000 (11:58 -0700)]
panfrost/midgard: Lower inot to inor with 0

We were previously lowering to inand, but the second arg was not
duplicated so inot would always return ~0. Oops.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Cleanup tag fetch in disassembler
Alyssa Rosenzweig [Thu, 6 Jun 2019 18:20:21 +0000 (11:20 -0700)]
panfrost/midgard: Cleanup tag fetch in disassembler

Trivial.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Use fancy iterator
Alyssa Rosenzweig [Thu, 6 Jun 2019 18:19:44 +0000 (11:19 -0700)]
panfrost/midgard: Use fancy iterator

Trivial cleanup.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Cull dead branches
Alyssa Rosenzweig [Thu, 6 Jun 2019 18:19:13 +0000 (11:19 -0700)]
panfrost/midgard: Cull dead branches

This fixes bugs with complex control flow.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Add mir_print_bundle helper
Alyssa Rosenzweig [Thu, 6 Jun 2019 18:18:30 +0000 (11:18 -0700)]
panfrost/midgard: Add mir_print_bundle helper

This helps with debugging scheduling/emission.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard/disasm: Pretty-print branch tags
Alyssa Rosenzweig [Thu, 6 Jun 2019 17:21:57 +0000 (10:21 -0700)]
panfrost/midgard/disasm: Pretty-print branch tags

Just makes it a little more obvious what's going on.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/ci: Note some since-fixed tests
Alyssa Rosenzweig [Thu, 6 Jun 2019 16:42:23 +0000 (09:42 -0700)]
panfrost/ci: Note some since-fixed tests

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Vectorize I/O
Alyssa Rosenzweig [Thu, 6 Jun 2019 16:15:26 +0000 (09:15 -0700)]
panfrost/midgard: Vectorize I/O

This uses the new mesa/st functionality for NIR I/O vectorization, which
eliminates a number of corner cases (resulting in assorted dEQP
failures and regressions) and should improve performance substantial due
to lessened pressure on the load/store pipe.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Remove varyings delay pass
Alyssa Rosenzweig [Thu, 6 Jun 2019 15:21:27 +0000 (08:21 -0700)]
panfrost/midgard: Remove varyings delay pass

This pass interfered with the more delicate path required for
non-vectorized I/O. It's also ugly and duplicating the job of an actual
honest-to-goodness scheduler.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agopanfrost/midgard: Apply component to load_input
Alyssa Rosenzweig [Thu, 6 Jun 2019 15:16:04 +0000 (08:16 -0700)]
panfrost/midgard: Apply component to load_input

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
5 years agonir: fix s/&&/||/ typo
Eric Engestrom [Fri, 7 Jun 2019 15:04:25 +0000 (16:04 +0100)]
nir: fix s/&&/||/ typo

Fixes: cd73b6174b093b75f581 "nir/lower_to_source_mods: Stop turning add, sat, and neg into mov"
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
5 years agofreedreno/a6xx: Drop struct stage array
Kristian H. Kristensen [Fri, 7 Jun 2019 03:29:38 +0000 (20:29 -0700)]
freedreno/a6xx: Drop struct stage array

This now boils down to just picking between binning or vertex shader
and dummy_fs or real fs, which we can do in a couple of lines of code
instead.  The constlen logic isn't doing what it thinks it's doing,
both constlens at this point

  MAX2(s[VS].constlen, align(state->bs->constlen, 4));

are binning shader constlens.  We'll have to revisit the constlen
logic, but this commit doesn't change how it works.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agofreedreno/a6xx: Drop support for SS6_DIRECT shader upload
Kristian H. Kristensen [Tue, 4 Jun 2019 20:44:48 +0000 (13:44 -0700)]
freedreno/a6xx: Drop support for SS6_DIRECT shader upload

a6xx only supports indirect shaders.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agofreedreno/a6xx: Share shader_t_to_opcode
Kristian H. Kristensen [Thu, 6 Jun 2019 04:35:35 +0000 (21:35 -0700)]
freedreno/a6xx: Share shader_t_to_opcode

We have a similar function in fd6_program.c. Move to fd6_emit.h and
share.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agofreedreno/a6xx: Consolidate more of dword 0 building in fd6_draw_vbo
Kristian H. Kristensen [Mon, 3 Jun 2019 21:12:59 +0000 (14:12 -0700)]
freedreno/a6xx: Consolidate more of dword 0 building in fd6_draw_vbo

There's already a bit of duplicated logic here and tessellation will
add more. Build up dword 0 in fd6_draw_vbo() and drop the a4xx in the
process.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agofreedreno: Move fd4_size2indextype() helper to freedreno_util.h
Kristian H. Kristensen [Mon, 3 Jun 2019 21:01:14 +0000 (14:01 -0700)]
freedreno: Move fd4_size2indextype() helper to freedreno_util.h

In preparation for refactoring fd6_draw.c a bit.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoradv: enable VK_EXT_sample_locations
Samuel Pitoiset [Thu, 30 May 2019 07:58:01 +0000 (09:58 +0200)]
radv: enable VK_EXT_sample_locations

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: enable HTILE for images that might need variable sample locations
Samuel Pitoiset [Thu, 30 May 2019 08:26:43 +0000 (10:26 +0200)]
radv: enable HTILE for images that might need variable sample locations

This is now supported.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: handle sample locations during automatic layout transitions
Samuel Pitoiset [Thu, 30 May 2019 10:27:29 +0000 (12:27 +0200)]
radv: handle sample locations during automatic layout transitions

From the Vulkan spec 1.1.109:

   "Some implementations may need to evaluate depth image values
    while performing image layout transitions. To accommodate this,
    instances of the VkSampleLocationsInfoEXT structure can be
    specified for each situation where an explicit or automatic
    layout transition has to take place. [...] and
    VkRenderPassSampleLocationsBeginInfoEXT can be chained from
    VkRenderPassBeginInfo to provide sample locations for layout
    transitions performed implicitly by a render pass instance."

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: determine the first subpass id for every attachments
Samuel Pitoiset [Thu, 30 May 2019 12:10:42 +0000 (14:10 +0200)]
radv: determine the first subpass id for every attachments

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: handle sample locations during explicit depth/stencil transitions
Samuel Pitoiset [Thu, 30 May 2019 10:23:21 +0000 (12:23 +0200)]
radv: handle sample locations during explicit depth/stencil transitions

From the Vulkan spec 1.1.109,

   "Some implementations may need to evaluate depth image values
    while performing image layout transitions. To accommodate this,
    instances of the VkSampleLocationsInfoEXT structure can be
    specified for each situation where an explicit or automatic
    layout transition has to take place. VkSampleLocationsInfoEXT
    can be chained from VkImageMemoryBarrier structures to provide
    sample locations for layout transitions performed by
    vkCmdWaitEvents and vkCmdPipelineBarrier calls."

This handles explicit depth/stencil layout transitions performed
with CmdWaitEvents() or CmdPipelineBarrier().

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: allow the depth decompress pass to emit dynamic sample locations
Samuel Pitoiset [Thu, 30 May 2019 10:20:12 +0000 (12:20 +0200)]
radv: allow the depth decompress pass to emit dynamic sample locations

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: allow to set dynamic sample locations to the depth decompress pass
Samuel Pitoiset [Thu, 30 May 2019 09:52:56 +0000 (11:52 +0200)]
radv: allow to set dynamic sample locations to the depth decompress pass

If VK_EXT_sample_locations is used, the driver might need to emit
the sample locations specified during layout transitions.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoradv: allow to save/restore sample locations during meta operations
Samuel Pitoiset [Thu, 30 May 2019 09:50:22 +0000 (11:50 +0200)]
radv: allow to save/restore sample locations during meta operations

This will be used for the depth decompress pass that might need
to emit variable sample locations during layout transitions.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
5 years agoiris: Sweep the NIR in iris_create_uncompiled_shader().
Kenneth Graunke [Fri, 7 Jun 2019 07:57:25 +0000 (00:57 -0700)]
iris: Sweep the NIR in iris_create_uncompiled_shader().

We run a ton of backend specific passes here (mostly brw_preprocess_nir)
and ought to sweep up any unused memory at this point, since we're going
to hang on to this NIR for as long as the linked program lives.

5 years agoir3: Use the new NIR lowering pass for integer multiplication
Eduardo Lima Mitev [Sun, 12 May 2019 22:33:57 +0000 (00:33 +0200)]
ir3: Use the new NIR lowering pass for integer multiplication

Shader-db stats courtesy of Eric Anholt:

total instructions in shared programs: 6480215 -> 6475457 (-0.07%)
instructions in affected programs: 662105 -> 657347 (-0.72%)
helped: 1209
HURT: 13
total constlen in shared programs: 1432704 -> 1427769 (-0.34%)
constlen in affected programs: 100063 -> 95128 (-4.93%)
helped: 512
HURT: 0
total max_sun in shared programs: 875561 -> 873387 (-0.25%)
max_sun in affected programs: 46179 -> 44005 (-4.71%)
helped: 1087
HURT: 0

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoir3/nir: Add new NIR AlgebraicPass for lowering imul
Eduardo Lima Mitev [Sun, 12 May 2019 22:23:58 +0000 (00:23 +0200)]
ir3/nir: Add new NIR AlgebraicPass for lowering imul

Currently, ir3 backend compiler is lowering integer multiplication from:

dst = a * b

to:

dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)

by emitting this code:

mull.u tmp0, a, b           ; mul low, i.e. al * bl
madsh.m16 tmp1, a, b, tmp0  ; mul-add shift high mix, i.e. ah * bl << 16
madsh.m16 dst, b, a, tmp1   ; i.e. al * bh << 16

which at that point has very low chances of being optimized.

This patch adds a new nir_algebraic.AlgebraicPass to performs this
lowering during NIR algebraic optimization passes, giving it a better
chance for optimizing the resulting code.

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agonir_algebraic: Add basic optimizations for umul_low and imadsh_mix16
Eduardo Lima Mitev [Sun, 12 May 2019 22:09:38 +0000 (00:09 +0200)]
nir_algebraic: Add basic optimizations for umul_low and imadsh_mix16

For umul_low (al * bl), zero is returned if the low 16-bits word of either
source is zero.

for imadsh_mix16 (ah * bl << 16 + c), c is returned if either 'ah' or 'bl'
is zero.

A couple of nir_search_helpers are added:

is_upper_half_zero() returns true if the highest word of all components of
an integer NIR alu src are zero.

is_lower_half_zero() returns true if the lowest word of all components of
an integer nir alu src are zero.

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agoir3/compiler: Handle new alu opcodes 'umul_low' and 'imadsh_mix16'
Eduardo Lima Mitev [Sun, 12 May 2019 19:12:59 +0000 (21:12 +0200)]
ir3/compiler: Handle new alu opcodes 'umul_low' and 'imadsh_mix16'

They directly emit ir3_MULL_U and ir3_MADSH_M16 respectively.

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agonir/opcodes: Add new 'umul_low' and 'imadsh_mix16' opcodes
Eduardo Lima Mitev [Fri, 29 Mar 2019 09:49:12 +0000 (10:49 +0100)]
nir/opcodes: Add new 'umul_low' and 'imadsh_mix16' opcodes

'umul_low' is the low 32-bits of unsigned integer multiply. It maps
directly to ir3's MULL_U.

'imadsh_mix16' is multiply add with shift and mix, an ir3 specific
instruction that maps directly to ir3's IMADSH_M16.

Both are necessary for the lowering of integer multiplication on
Freedreno, which will be introduced later in this series.

Reviewed-by: Eric Anholt <eric@anholt.net>
5 years agov3d: don't emit point coordinates varyings if the FS doesn't read them
Iago Toral Quiroga [Thu, 6 Jun 2019 08:04:27 +0000 (10:04 +0200)]
v3d: don't emit point coordinates varyings if the FS doesn't read them

We still need to emit them in V3D 3.x since there there is no mechanism to
disable them.

Reviewed-by: Eric Anholt <eric@anholt.net>