radeonsi: raise number of samplers per shader to 32

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94835
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: expand the compressed color and depth texture masks to 64 bits

This is in preparation of raising the number of exposed sampler views to 32
bits, which will raise the total number of sampler views to 33 for the
polygon stipple texture. That texture should never be compressed (and it's
certainly not a depth texture), but this approach seems cleaner to me than
special-casing the last slot in all affected code paths.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: replace magic 16 by SI_NUM_USER_SAMPLERS

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: raise PIPE_MAX_SAMPLERS to 32

The previous value of 18 was motivated by having drivers that want to expose
16 samplers but also use some additional samplers for internal use. Raising
the value even higher isn't going to hurt that case.

On the other hand, some drivers actually use PIPE_MAX_SAMPLERS as the number
of samplers they expose externally, so raising this number above 32 is fragile
(because several places in the code use bitfields, and tracking down and
widening all of them is prone to miss some case).

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/glsl_to_tgsi: make samplers_used an uint32_t (v2)

It is used as a bitfield, so it seems cleaner to keep it unsigned.

The literal 1 is a (signed) int, and shifting into the sign bit is undefined
in C, so change occurences of 1 to 1u.

v2: add an assert for bitfield size and use 1u << idx

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com> (v1)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)

tgsi/scan: add an assert for the size of the samplers_declared bitfield

The literal 1 is a (signed) int, and shifting into the sign bit is undefined
in C, so change occurences of 1 to 1u.

Reviewed-by: Brian Paul <brianp@vmware.com>

draw/aaline: stronger guard against no free samplers (v2)

Line anti-aliasing will fail when there is no free sampler available. Make
the corresponding guard more robust in preparation of raising
PIPE_MAX_SAMPLERS to 32.

The literal 1 is a (signed) int, and shifting into the sign bit is undefined
in C, so change occurences of 1 to 1u.

v2: add an assert for bitfield size and use 1u << idx

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com> (v1)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> (v1)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)

util/pstipple: stronger guard against no free samplers (v2)

When hasFixedUnit is false, polygon stippling will fail when there is no free
sampler available. Make the corresponding guard more robust in preparation
of raising PIPE_MAX_SAMPLERS to 32.

The literal 1 is a (signed) int, and shifting into the sign bit is undefined
in C, so change occurences of 1 to 1u.

v2: add an assert for bitfield size and use 1u << idx

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com> (v1)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> (v1)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)

svga: new SVGA_MSAA env var to disable/enable MSAA pixel formats

On by default.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

svga: add some trivial null pointer checks

These small mallocs will probably never fail, but static analysis tools
may complain about the missing checks.

Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

trace: add missing set_shader_images()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: disable perfect ZPASS counts for PIPE_QUERY_OCCLUSION_PREDICATE

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: don't use the real barrier instruction in tess ctrl shaders

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

Revert "clover: Fix build against clang SVN >= r265359"

This reverts commit 0daab9878d2b96356cf667591a2c877d912be52d.

The corresponding clang change was reverted.

Trivial.

nir/types: Add a wrapper for count_attribute_slots

Reviewed-by: Rob Clark <robdclark@gmail.com>

r600: use radeon_emit in a few more places in evergreen_compute

This is just a cleanup of the code.

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: make compute global buffer functions static.

This moves things around so that the global buffer handling
functions in evergreen_compute.c are static.

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: make two compute functions static.

These aren't used outside evergreen_compute.c

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: using pipe_grid_info more in evergreen_compute.

No reason to pull the pieces apart here, also make
one of the functions static as it's unused outside this.

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: in evergreen_compute use ctx consistently instead of ctx_

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: use rctx consistently in evergreen_compute.c

Another step towards cleaning this up.

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: cleanup whitespace in evergreen_compute.c

This aligns the code with the style of the rest of the driver.

Makes editing it a lot less painful.

Acked-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

GL3.txt: Mark ARB_framebuffer_no_attachments as done

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

r600g: Enable ARB_framebuffer_no_attachments

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Enable ARB_framebuffer_no_attachments

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Improve assert info out of si_set_framebuffer_state()

Lets give the developer a little hand if we are going to assert
on a zero literal at the end of a branch.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Allow 16 samples MSAA mode for PIPE_FORMAT_NONE

For ARB_framebuffer_no_attachment; A is_format_supported() query
with 'PIPE_FORMAT_NONE' passed implies a query of the number of
samples supported from the framebuffer with no attachment.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

softpipe: Set samples and layers in set_framebuffer_state() cb

Carries across the number of samples and layers state in the
'softpipe_set_framebuffer_state()' callback. This state is
part of 'ARB_framebuffer_no_attachments' support.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

mesa/st: Update framebuffer state with no.of samples,layers

Handle the case of ARB_framebuffer_no_attachment.
Also, kill off a dead debug printf() call while we are here.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium/trace: Dump no.of samples and layers in fb state

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: Put no.of {samples,layers} into pipe_framebuffer_state

Here we store the number of samples and layers directly in the
pipe_framebuffer_state so that in the case of
ARB_framebuffer_no_attachment we may make use of them directly.

Further, we adjust various gallium/auxiliary helper functions
accordingly.

V2:
  Convert branches in util_framebuffer_get_num_layers() and
  util_framebuffer_get_num_samples() to their canonical form.

V3:
  'git stash pop' the typo fix of 'cbufs' which should be
  'nr_cbufs' that was missing in V2, woops! Thanks Marek for
  pointing this out yet again.

V4:
  Squash in the following patch:

  'gallium/util: Ensure util_framebuffer_get_num_samples() is valid'

   Upon context creation, internal driver structures are malloc()'ed
   and memset() to zero them. This results in a invalid number of
   samples 'by default'. Handle this in the simplest way to avoid
   elaborate and probably equally sub-optimial solutions.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

mesa/st: Set _NumSamples in update_framebuffer_state()

Using PIPE_FORMAT_NONE to indicate what MSAA modes are supported
with a framebuffer using no attachment.

V.2:
 Rewrite MSAA mode loop to be more general.
V.3:
 Move comment to right place after loop was rewritten.
V.4: [airlied]
 remove unneeded variable, and assert, and unneeded pipe assignment

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

gallium: Obtain ARB_framebuffer_no_attachment constants

Set default values for the constants required in
ARB_framebuffer_no_attachments and obtained the number
of layers from ``PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS``.

We also obtain the MaxFramebufferSamples value using
a query back to the driver for PIPE_FORMAT_NONE.

V.1:
 Merge if branch predicates into one branch.
 Move const init into st_init_limits()

[airlied: whitespace fixup]
Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

gallium: Add PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT

Add PIPE_CAP to determine if the GL extension
'GL_ARB_framebuffer_no_attachments' shall be
supported.

The driver is required to support 'PIPE_FORMAT_NONE'
via its 'is_format_supported()' callback in order
to determine the MSAA modes the hardware supports so
that values requested from the application using
'GL_ARB_framebuffer_no_attachments' may be quantized
to what the hardware expects.

V.2:
 Fix doc for a more detailed description of the PIPE_CAP
 and the corresponding GL constant.

V.3:
 Renamed and repurposed once again.

V.4:
 Remove CAP from cap_mapping array.

[airlied: fix damaged whitespace]

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

mesa/st: Use _mesa_geometric_ functions appropriately

Change references to gl_framebuffer::Width, Height, MaxNumLayers
and Visual::samples to use the _mesa_geometric_ convenience functions
for those places where the geometry of the gl_framebuffer is needed.
This is in contrast to the geometry of the intersection of the
attachments of the gl_framebuffer.

This patch paves the way to enable GL_ARB_framebuffer_no_attachements
for all gallium drivers.

V.2:
 Remove itermeditate variable state.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

mesa: Add comment to framebuffer_parameteri()

V.2:
 Change 'N.B.,' to 'NOTE:'.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

i965/sf_state: Pull flat_enables out of prog_data

Previously, we were walking over the shader source to figure out which
inputs should be marked flat.  Now, we can just pull it out of prog_data.
This is needed for properly setting up 3DSTATE_SF/SBE for Vulkan and it
also means that it will get properly cached.

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Add a flat_inputs field to prog_data

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

brw/device_info: Add a helper for getting a device name

This is needed by the Vulkan driver

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs_surface_builder: Mask signed integers after conversion

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

i965/fs: Make the repclear shader support either a uniform or a flat input

In the Vulkan driver we use a single flat input instead of a uniform
because setting up push constants is more disruptive to the pipeline than
setting up another vertex input.  This uses the number of uniforms as a key
to keep it working for the GL driver.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Move get_hw_prim_for_gl_prim to brw_util.c

It's used by brw_compile_gs in brw_vec4_gs_visitor.cpp so it needs to be in
a file that's linked into libi965_compiler.la.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

radeonsi: set shader calling conventions

Note that old mesa + new LLVM or new mesa + old LLVM breaks
with this change and the corresponding LLVM change (D18559).

For LLVM version <= 3.8 we use the old method, but we can't detect
people using a post 3.8 svn version that is still too old.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

drirc: add a workaround for blackness in Warsow

Cc: 11.1 11.2 <mesa-stable@lists.freedesktop.org>

glsl: use has_shader_storage_buffer_objects helper

Replaces open-coded logic with existing helper.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>

glsl: remove remaining tabs in link_uniform_blocks.cpp

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>

mesa: remove unused IsShaderStorage field

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>

glsl: fully split apart buffer block arrays

With this change we create the UBO and SSBO arrays separately from the
beginning rather than putting them into a combined array and splitting
it apart later.

A bug is with UBO and SSBO stage reference querying is also fixed as
we now use the block index to lookup the references in the separate arrays
not the combined buffer block array.

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>

freedreno/ir3: insert extra move into phi

We had an implicit assumption that the phi src was assigned in it's
source (pred) block leading into the phi.  But this is not true with
NIR, so we can't just ignore the source block specified in the
nir_phi_src.  Insert an extra mov in the source block.  If it is not
required the CP pass will take it back out again.

Fixes:

  ./tests/spec/glsl-1.10/execution/vs-call-in-nested-loop.shader_test
  ./tests/spec/glsl-1.10/execution/vs-inner-loop-modifies-outer-loop-var.shader_test

and probably others.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: eliminate unnecessary absneg's

The frontend inserts (abs) and (neg)'s to convert between NIR boolean
(~0/0) and native boolean (1/0).  So we'd end up with things like:

   cmps.s.ge r1.x, ...
   absneg.s r1.x, (neg)r1.x
   absneg.s r1.x, (abs)r1.x
   sel.b32 r2.x, r0.x, r1.x, r0.y

The (neg) already gets collapsed due to the following (abs).  Now by
realizing that r1.x comes from a cmps.s instruction, we can drop the
(abs) as well.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

clover: Fix build against clang SVN >= r265359

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

radeonsi: use bounded indexing for samplers

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: use bounded indexing for constant buffers

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: allow multiple exports of the same texture with different usage

Instead of failing an assertion, disable DCC and CMASK on the first export
that needs it, and merge the external usage flags.

v2: clear the EXPLICIT_FLUSH flag if it's not set; whitespace fixes

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

docs/relnotes: document EGL_KHR_reusable_sync

egl: add EGL_KHR_reusable_sync to egl_dri

This patch enables an EGL extension, EGL_KHR_reusable_sync.
This new extension basically provides a way for multiple APIs or
threads to be excuted synchronously via a "reusable sync"
primitive shared by those threads/API calls.

This was implemented based on the specification at

https://www.khronos.org/registry/egl/extensions/KHR/EGL_KHR_reusable_sync.txt

v2
- use thread functions defined in C11/threads.h instead of
  using direct pthread calls
- make the timeout set with reference to CLOCK_MONOTONIC
- cleaned up the way expiration time is calculated
- (bug fix) in dri2_client_wait_sync, case EGL_SYNC_CL_EVENT_KHR
  has been added.
- (bug fix) in dri2_destroy_sync, return from cond_broadcast
  call is now stored in 'err' intead of 'ret' to prevent 'ret'
  from being reset to 'EGL_FALSE' even in successful case
- corrected minor syntax problems

v3
- dri2_egl_unref_sync now became 'void' type. No more error check
  is needed for this function call as a result.
- (bug fix) resolved issue with duplicated unlocking of display in
  eglClientWaitSync when type of sync is "EGL_KHR_REUSABLE_SYNC"

Signed-off-by: Dongwon Kim <dongwon.kim@intel.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

freedreno/ir3: deal with duplicate phi sources

Otherwise we end up with funny things like:

  mov.f32f32 r0.x, r1.y
  mov.f32f32 r0.x, r1.y

(It doesn't happen as much after fixing the problem w/ CP into phi src,
but it can still happen since we aren't too clever about generating phi
sources in the first place.)

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: fix silly brain-fart in RA

We want to consider all the vars, not 1/32nd of them, when extending
live-ranges.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: don't cp into phi's

The block defining a phi source might not have been executed.  If we
allow copy propagation, we could end up pointing to a src instruction in
the wrong block.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: we can't store immediate values

Fixes some transform-feedback piglits, like:

bin/ext_transform_feedback-nonflat-integral

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: add dumping for use/def/live-in/live-out

Turned out to be useful to debug an issue in RA.  Let's keep it.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: drop unused instr category arg

No longer used, so drop the extra arg to ir3_instr_create()

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: remove ir3_instruction::category

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/ir3: encode instruction category in opc_t

Been on my TODO list for a while.  If nothing else this will make gdb
properly grok the opc_t enum.

This first step preserves ir3_instruction::category (with an added
assert that category matches what is encoded in opc_t).  Next step is
to drop the category field (and arg to ir3_instr_create()), but that
is split into next commit for bisectability and so that we can run
piglit in the intermediate state to flush out any problems.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

i965/fs: Move the code for load/store_shared to emit_cs_intrinsic

They are compute-shader only and that's where the code for doing atomics on
shared variables lives so it seemes to make sense.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

i965/nir: Provide a default LOD for buffer textures

Our hardware requires an LOD for all texelFetch commands even if they are
on buffer textures.  GLSL IR gives us an LOD of 0 in that case, but the LOD
is really rather meaningless.  This commit allows other NIR producers to be
more lazy and not provide one at all.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

nir: Use PRIi64 and PRIu64 instead of %ld and %lu.

%ld and %lu aren't the right format specifiers for int64_t and uint64_t
on 32-bit (x86) systems.  They're %zu on Linux and %Iu on Windows.

Use the standard C99 macros in hopes that they work everywhere.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Fix invalid pointer read in dead_control_flow_eliminate().

There may not be a previous block.  In this case, there's no real work
to do, so just continue on to the next one.

v2: Update for bblock->prev() API change.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>

i965: Make bblock_t::next and friends return NULL at sentinels.

The bblock_t::prev/prev_const/next/next_const API returns bblock_t
pointers, rather than exec_nodes.  So it's a bit surprising.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>

glsl: Lower variable indexing of system value arrays unconditionally.

lower_variable_index_to_cond_assign() did not handle system values.
gl_SampleMaskIn[] is a system value, and also an array.  Accessing it
with a variable index would trigger an unreachable() assert.

Rather than adding a new EmitNoIndirectSystemValues flag, we simply
lower unconditionally.  There is exactly one case where this occurs,
and for all current drivers, lowering produces optimal code.  Even
for future drivers with 32x MSAA, it produces reasonable code.

Fixes Piglit's new samplemaskin-indirect test.  Also fixes many ES31-CTS
tests when OES_sample_variables is enabled.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965/peephole_ffma: Only match a mul+add if none of the ops are exact

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

nir/search: Don't match inexact expressions with exact subexpressions

In the first pass of implementing exact handling, I made a mistake with
search-and-replace.  In particular, we only reallly handled exact/inexact
on the root of the tree.  Instead, we need to check every node in the tree
for an exact/inexact match.  As an example of this, consider the following
GLSL code

precise float a = b + c;
if (a < 0) {
   do_stuff();
}

In that case, only the add will be declared "exact" and an expression that
looks for "b + c < 0" will still match and replace it with "b < -c" which
may yield different results.  The solution is to simply bail if any of the
values are exact when matching an inexact expression.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Add an INTEL_PRECISE_TRIG=1 option to fix SIN/COS output range.

The SIN and COS instructions on Intel hardware can produce values
slightly outside of the [-1.0, 1.0] range for a small set of values.
Obviously, this can break everyone's expectations about trig functions.

According to an internal presentation, the COS instruction can produce
a value up to 1.000027 for inputs in the range (0.08296, 0.09888).  One
suggested workaround is to multiply by 0.99997, scaling down the
amplitude slightly.  Apparently this also minimizes the error function,
reducing the maximum error from 0.00006 to about 0.00003.

When enabled, fixes 16 dEQP precision tests

   dEQP-GLES31.functional.shaders.builtin_functions.precision.
   {cos,sin}.{highp,mediump}_compute.{scalar,vec2,vec4,vec4}.

at the cost of making every sin and cos call more expensive (about
twice the number of cycles on recent hardware).  Enabling this
option has been shown to reduce GPUTest Volplosion performance by
about 10%.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

i965: Allow 8x MSAA on >= 64bpp formats on Gen8+.

See commit 3b0279a69 - this restriction is documented in the "Surface
Format" field of RENDER_SURFACE_STATE.

Looking at newer documentation, this restriction appears to exist on
Haswell, but no longer applies on Gen8+.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>

docs: remove stray 'TBD' in 11.2.0 relnotes file

docs: add news item and link release notes for 11.2.0

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

docs: add sha256 checksums for 11.2.0

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
(cherry picked from commit e7fb889dcc002f87c316f3cdc6e7907a88c12697)

docs: Update 11.2.0 release notes

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
(cherry picked from commit ff9ddb9eb1b3b25f40e71a95bb48421abfcb11d9)

mesa/get: fix MAX_GEOMETRY_SHADER_STORAGE_BLOCKS

this was returning the fragment shader value.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

nvc0: add hardware ETC2 and ASTC support on GK20A and GM107+

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

docs: add note about GL_EXT_base_instance, sort entries

Trivial.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

mesa: expose EXT_base_instance in ES3 contexts

This extension is identical to ARB_base_instance. Reuse the same
entrypoints.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

mesa: expose EXT_polygon_offset_clamp in ES contexts

The extension spec was extended to also support ES. This functionality
is provided all the way back to ES 1.0.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

glsl: Print "precise" on ir_variable nodes.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>

gallivm: Introduce lp_format_intrinsic.

For adding .v4f32 like suffixes to intrinsics, taking special care for
scalar case, which was being often neglected.

This fixes invalid IR when doing mipmap filtering on SSE2 (the only
case where we'd use intrinsics with scalars.)

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

glsl: make *sampler2DMSArray available in ESSL 3.20

Also avoid double-adding the *sampler2DMS types when the array ext is
enabled.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

glsl: make ssbo predicate return true when in a GLSL 430 or ESSL 310 shader

I can't tell whether this actually matters, but we're creating function
signatures with this predicate, so it should probably match when SSBO's
are available.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

glsl: allow conservative depth qualifiers in GLSL 420

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

mesa: add always-false-for-now enables for GL 4.3, 4.4, 4.5.

As the relevant extensions get implemented, the lines should be
uncommented. I believe this is (almost) everything needed for those GL
versions though.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

glsl: add ARB_ES3_1_compatibility support

Oddly a bunch of the features it adds are actually from ESSL 3.20. But
the spec is quite clear, oh well.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

mesa: add ES3_1_compatibility extension enable

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>

gallivm: Use llvm.fabs.

Exactly the same code.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

gallivm: Prefer backend agnostic intrinsic for rounding.

We could unconditionally use these instrinsics, but performance with SSE2
would suck, as LLVM falls back to calling libm.

lp_test_arit.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

gallivm: Add debug option to force SSE2.

For simulating less capable machines.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

llvmpipe: Test abs.

Trivial.

llvmpipe: Build lp_test_arit on MSVC too.

It builds fine now.  Probably due to C99 support.

Trivial.

gallivm: Fix performance regressions due to vector selects.

LLVM often can't determine the mask elements are all ones/zeros, and
there doesn't seem to be a good way to hint that.

Thanks to Roland Scheidegger for spotting and analyzing the issue.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

gallivm: Remove lp_build_load_volatile.

No longer needed.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

gallivm: Use standard LLVMSetAlignment from LLVM 3.4 onwards.

Only provide a fallback for LLVM 3.3.

One less dependency on LLVM C++ interface.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

mesa: remove unrequired else

The if always returns so no need for an else.

Reviewed-by: Brian Paul <brianp@vmware.com>