i915g: Allow use of I915_DEBUG= options on non-DEBUG builds.

All the checks should be cheap, no need making things harder to debug.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11617>

i915g: Bake the decls and program together.

Simplifies program upload a bunch, and will let us disasm the program
independently of the whole cmd buffer.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11617>

i915g: Apply clang-format.

Add the format file copied from freedreno, and run clang-format
--style=file -i src/gallium/drivers/i915/*.[ch] (with clang-format 11).

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11617>

venus: properly support GPU_DATA_BUFFER for AHB

CrOS gralloc was missing support for GPU_DATA_BUFFER, which was only
fixed until recently. Now we can allcate properly.

This patch also removes a redundant TODO regarding image format list
support for AHB. vkGetPhysicalDeviceImageFormatProperties2 has already
checked the support for such though it only checks whether the optimal
tiling and the associated drm format modifier for the AHB image format
itself applies to the entire format list or not. Given the use case for
such combination is quite limited, we choose not to add new gralloc
support to force linear modifier as a fallback for wider coverage.

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11501>

intel: PAGE_SIZE used in allocators shouldn't be defined on FreeBSD

iris, i965, and anv define the PAGE_SIZE in anv_allocator and bufmgr
files. As on FreeBSD the page size is defined in machine/param.h that is
indirectly included by those files, we'd rather define it only when the
system is not FreeBSD to avoid compile errors.

v2: Changed the path in the comment to make clear that machine/params.h
is a FreeBSD system file.

Signed-off-by: Eleni Maria Stea <elene.mst@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11203>

intel: struct bitset is renamed to brw_bitset

Static struct bitset was renamed to brw_bitset as a struct bitset
is defined in sys/_bitset.h included by pthread_np.h on FreeBSD that
is indirectly included by src/intel/compiler/brw_nir_lower_shader_calls.c

Signed-off-by: Eleni Maria Stea <elene.mst@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11203>

util: Removed unused statement from FreeBSD build

Statement (void*)debug_name when FreeBSD is defined has no use. Removed
it to fix compiler warnings.

Signed-off-by: Eleni Maria Stea <elene.mst@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11203>

util: replaced ENODATA with ENOATTR for non-Linux systems

On Linux ENODATA is defined but on BSD, and MacOSX ENOATTR is used
instead. Defined ENODATA to be ENOATTR when the system is not Linux.

v2: Replaced ENODATA and ENOATTR with -EFAULT that is exists everywhere
and added a comment (Ian Romanick)

Signed-off-by: Eleni Maria Stea <elene.mst@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11203>

util/vbuf: flag fallback_always if any prim types are missing from restart modes

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11608>

util/vbuf: always claim support for PATCHES in restart modes

this avoids forcing rewrites for patch mode which break tessellation

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11608>

turnip: Short-circuit if ladder generation for constant index SSBO/UBOs.

The compiler *can* eventually chew through all the copy prop, constant
folding, and dead_cf necessary to use just our constant index, but we can
save a whole lot of hassle by chasing the MOVs up front and finding the
constant.

dEQP-VK.ubo.3_level_array.scalar.row_major_mat4.both goes from 2.0s to
1.6s on a release build (3.1s to 2.1s for a debug build like we use in CI).

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11613>

nir: Add a helper for chasing movs with nir_ssa_scalar().

Sometimes you might want to find a constant source without going through
all the copy prop and constant folding to make your source be a
load_const.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11613>

ir3/sched: Speed up live_effect

If we've identified another use that isn't scheduled yet, we can break
right away rather than iterating through all the other uses. While this
could be optimized further, this simple change makes
dEQP-VK.subgroups.ballot_broadcast.compute.subgroupbroadcast_ivec4 go
from 40 seconds to 1.9 seconds on a release build according to my
unscientific testing.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11613>

freedreno/computerator: Fix local_size typo

Fixes: cbc68c79a51 ("freedreno: Add local_size to ir3_shader_variant")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11622>

nir/opt_load_store_vectorize: fix check_for_robustness() with deref access

We could do better if we knew the nir_address_format to obtain
addition_bits, but the only affected driver (Turnip) probably won't
benefit because it doesn't vectorize across vec4.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 2e7bceb2204 ("nir/load_store_vectorizer: fix check_for_robustness() with indirect loads")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4922
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11382>

anv: expose primary node to VK_EXT_physical_device_drm even when VK_KHR_display is not enabled

Fixes: e9e1e0362b6c2d ("anv: implement VK_EXT_physical_device_drm")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11616>

radeonsi: move the accepting code into the bbox cull branch in prim discard CS

This reduces the number of jumps. No change in behavior.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11510>

radeonsi: add optimal multi draws and draw-level splitting for prim discard CS

This is a partial rewrite of some parts of the code.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11510>

radeonsi: fix issues with draw-level splitting for the prim discard CS

- move the unreachable statement up by inlining u_decomposed_prims_for_vertices
- the last draw was not submitted after splitting
- move "count" to the non-multi draw path

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11510>

radeonsi: change how the prim discard CS is enabled and splitting limits

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11510>

radeonsi: remove the GDS variants of compute-based primitive discard

The GDS ordered append variant is unstable due to kernel and firmware bugs.
The unordered GDS variant isn't faster than the memory-based variant.

Only the memory-based variant is kept.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11510>

intel/fs: use stack for temporary array

"regs" is an array of 2 ->
  "m" must be <= 2 ->
  "components" array can be allocated on the stack

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11575>

intel/blorp: initialize BLEND_STATE using braced initializer list

Coverity complains that memset has no effect, because of size 0.
Size of BLEND_STATE struct is 0 on gfx [6, 7.5], so memset has
nothing to do there. This is of course harmless, but we can make
code simpler by replacing memset with an empty initializer list
and at the same time avoid a warning from Coverity.

CID: 1486015

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

intel/tools: fix potential memory leaks

CID: 1444041
CID: 1444104

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

intel/tools: fix invalid type in argument to printf format specifier

Also mark parse_error as printf-like to catch such errors with gcc.

CID: 1473100
CID: 1473101

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

intel/tools: fix int-to-pointer/pointer-to-int cast warnings on 32-bit

gcc complains:
../src/intel/tools/aub_write.c: In function ‘populate_ppgtt_table’:
../src/intel/tools/aub_write.c:254:16: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
  254 |                (void *)(aub->phys_addrs_allocator++ << 12);
      |                ^
../src/intel/tools/aub_write.c:258:27: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
  258 |                        i, (uint64_t)table->subtables[i]);
      |                           ^
../src/intel/tools/aub_write.c:273:24: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
  273 |          (level == 1 ? (uint64_t)table->subtables[i] :
      |                        ^
../src/intel/tools/aub_write.c: In function ‘ppgtt_lookup’:
../src/intel/tools/aub_write.c:346:11: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
  346 |    return (uint64_t)L1_table(ppgtt_addr)->subtables[L1_index(ppgtt_addr)];
      |           ^
../src/intel/tools/intel_sanitize_gpu.c: In function ‘bo_size’:
../src/intel/tools/intel_sanitize_gpu.c:99:15: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
   99 |    return e ? (uint64_t)e->data : UINT64_MAX;
      |               ^

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

intel/tools: fix left shift overflow on 32-bit

gcc complains:
../src/intel/tools/aub_write.c:394:53: warning: left shift count >= width of type [-Wshift-count-overflow]
  394 |    assert((aub->phys_addrs_allocator + size) < (1UL << 32));
      |                                                     ^~

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

anv: fix potential integer overflows

In all cases both variables has a type of uint32_t, so multiplying
them will also generate uint32_t. The results of those multiplications
are used as uint64_t's, so Coverity thinks there might be integer
overflows here.
I don't think it's possible to hit them (query BOs should be relatively
small), but let's avoid those overflows.

CID: 1472820
CID: 1472821
CID: 1472822
CID: 1472824
CID: 1475934
CID: 1475927

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

intel/decoder: add assert for register size

Coverity complains about out-of-bounds access in
intel_field_iterator_init, because it doesn't know that the GT_MODE
register has a size of 4 bytes. Add an assertion to verify that.

CID: 1474552

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11574>

mesa/shaderapi: add an optional shader override mechanism

MESA_SHADER_READ_PATH is handy but it's not usable in
all cases.

This commit allows to implement an alternative mechanism
without assuming too much about how it's done, nor where/how
the shaders are stored.

When this is enabled MESA_SHADER_DUMP_PATH,
MESA_SHADER_CAPTURE_PATH and MESA_GLSL env var handling is
disabled.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11621>

mesa/shaderapi: change construct_name signature

Pass the sha1 as an argument.
This will be used by the next commit.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11621>

Update libva requirement

Bump libva requirement since mesa references API not present prior to
libva API version 1.1.0 (libva 2.1.0 release).

See fbf6511e35, which added VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2
requiring API version 1.1.0 or later.

Signed-off-by: Jeremy Newton <Jeremy.Newton@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11508>

mesa: lower GL_CLAMP in texture and sampler functions instead of st/mesa

This moves the GL_CLAMP translation out of st/mesa.
This should be faster assuming constant sampler states.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: add IsBorderColorNonZero to skip border color update for st/mesa faster

0.8% less overhead.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: add LodBias quantization from st/mesa

Apparently the quantization math isn't cheap.

This further reduces overhead by 2% for drawoverhead/8 textures.
The improvement is measured by looking at the sysprof percentage delta and
multiplying by 2 (because we have the frontend and gallium threads with
equal overhead, so the benefit is doubled compared to 1 thread).

Both per-sampler and per-unit lod bias values are quantized.

The difference in behavior is that both values are quantized separately
and then added up, instead of first added up and then quantized.
The worst case error is +- 1/256 in the reduced precision, i.e. off by one
in a fixed-point representation, which should be fine.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: translate into pipe_sampler_state in GL functions

- pipe_sampler_state is added into gl_sampler_attrib.
- The border color field is redundant, so the non-gallium field is removed.
- Other GL fields are still tracked because of glGet.
- Some fields are not set by GL functions because they are always derived
  from multiple states.
- The behavior in GL functions is the same as st_convert_sampler.

This reduces overhead by 2% for drawoverhead/8 textures, and more is going
to be done.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: remove mutex locking from a glBindTexture early out path

The locking doesn't seem to be needed. This removes large locking
overhead.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: use atomics instead of mutexes for refcounting renderbuffers

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: use atomics instead of mutexes for refcounting sampler objects

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

mesa: use atomics instead of mutexes for refcounting texture objects

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11339>

cso: don't look up a sampler CSO if the last one is identical

This is benefical when sampler states are identical often, and detrimental
if they are not. The average case seems to be in favor of this.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

cso: update max_sampler_seen only once in cso_set_samplers

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

cso: disallow NULL sampler state templates in cso_single_sampler

No code passes NULL into this except the single place that is updated.
Let's remove the if. The next commit depends on this.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: remove the sampler min_lod/max_lod value swap

It might not be needed at all.

v2: fix v3d assertion failure

Acked-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: add a mechanism to bypass atomics when binding sampler views

This is the same mechanism we already use for buffers. The code is mostly
copied from there. See the big comment for explanation.

This will be very effective when take_ownership is added into pipe_context
::set_sampler_views because that and this commit together will eliminate
atomics for sample views almost entirely.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: sink refcounting from st_get_sampler_views into st_sampler_view.c

Do it deeper in the sampler view return path. It minimizes the deltas
for the next commit.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: remove the const qualifier for a few st_sampler_view instances

It will be needed by following commits.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: return sview from st_update_single_texture via return value, not param

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: read Target only once in st_update_single_texture

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: sink _mesa_get_samplerobj into st_update_single_texture

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: don't memset the sampler view array, don't init trailing slots to NULL

No functional change.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: don't track FS sampler views for bitmap/drawpix in st_context

Just query the sampler views from the update function. This will help
optimize sampler view updates.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

st/mesa: don't track VS sampler views for st_draw_feedback in st_context

Just query the sampler views from the update function. This will help
optimize sampler view updates.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11428>

turnip: Add CrOS Gralloc support

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11612>

turnip: Fix AcquireImageANDROID() handle type

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11612>

turnip: Use drmIoctl()

Replace open-coded ioctl with drmIoctl() to get restart on interrupted
system calls.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11612>

anv: allocate bigger batches as we grow command buffers

This is the first time we see an application running out of mmap().

We essentially allocate too many batches (+65k) and end up not being
able to mmap them, at which point we can't mmap anything anymore and
things go sideways.

This change allocates bigger batch BOs as we grow an existing command
buffer. This drastically reduces the number of BOs we need to allocate
(the benchmark that reported the issue now reaches a max of ~630 BOs,
instead of reaching 65k and failing previously).

v2: Track the total batch size of command buffers (Jason)
    Just give 0 for batch_len to i915 (Jason)

v3: Fix indentation (Jason)

v4: Drop uncessary reshuffling of error labels (Jason)

v5: Remove empty lines (Marcin)

v6: Limit BO growing to chunks of 16Mb (Jason)

v7: Add assert on initial size (Jason)

v8: Add define for max size (Jason)

v9: Fixup v7 assert for non softpin platforms (Lionel)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4956
Cc: mesa-stable
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11482>

anv: Advertise VK_KHR_shader_subgroup_uniform_control_flow

This is always valid for Intel, so we don't need to do anything
different if the Execution Mode is being used.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11476>

spirv: Support SPV_KHR_subgroup_uniform_control_flow

There's no SPIR-V Capability associated, so check in the Execution
Mode.  For now, don't keep track of whether a shader uses uniform
control flow in the shader_info, we can add that when/if a driver
actually need that information.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11476>

spirv: Update headers and metadata from latest Khronos commit

This corresponds to f95c3b3761ee1b1903f54ae69b526ed6f0edc3b9
("Merge pull request #219 from cmarcelo/SPV_EXT_shader_atomic_float16_add")
in https://github.com/KhronosGroup/SPIRV-Headers.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11476>

nir: Add test to check edge case in Split ALU optimization

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11476>

nir/opt_if: Don't split ALU for single block infinite loops

Some infinite loop cases were already covered by other
restrictions (e.g. if the loop had a body), but the case with a single
block in the loop body wasn't yet.

This prevents an infinite loop when optimizing the shader in
dEQP-VK.reconvergence.subgroup_uniform_control_flow_ballot.compute.nesting2.3.2
and various others reconvergence tests.

Fixes: 0881e90c099 ("nir: Split ALU instructions in loops that read phis")
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> [v1]
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11476>

i965: Prevent invalid framebuffer usage

When a surfaceless context is in use, driDrawablePriv might be NULL, so needs to
be checked before calling dri2InvalidateDrawable. Same for read calls.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Eleni Maria Stea <estea@igalia.com>
See https://gitlab.freedesktop.org/mesa/mesa/-/issues/778

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11459>

i915: Prevent invalid framebuffer usage

When a surfaceless context is in use, driDrawablePriv might be NULL, so needs to
be checked before calling dri2InvalidateDrawable. Same for read calls.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Eleni Maria Stea <estea@igalia.com>
Closes https://gitlab.freedesktop.org/mesa/mesa/-/issues/778

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11459>

amd/ci: Use MESA_VK_IGNORE_CONFORMANCE_WARNING to reduce warnings

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11563>

ci: Unify on MESA_VK_IGNORE_CONFORMANCE_WARNING

Move and rename warn_non_conformant_implementation() to common location
of src/vulkan/util/vk_util.c as vk_warn_non_conformant_implementation().

In freedreno/ci,  move MESA_VK_IGNORE_CONFORMANCE_WARNING to common
location of .baremetal-deqp-test-freedreno-vk.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11563>

anv: VK_EXT_multi_draw implementation

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11531>

anv: unify some draw state vertex constant emission

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11531>

zink: repack zink_context struct a bit

Reviewed-by: Joshua Ashton <joshua@froggi.es>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11593>

freedreno, tu: Set SP_XS_PVT_MEM_HW_STACK_OFFSET

Theoretically this register should only be used when function calls in
the shader are used, which we don't support. But with the default value
of 0 it seems like pvtmem doesn't work on a650. Just set it to the total
per-SP size, effectively leaving no space for the return-address stack,
like the blob does.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4949
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11581>

freedreno/a6xx: Make SP_XS_PVT_MEM_HW_STACK_OFFSET non-inline

Otherwise we can't use the helper to pack it as it collides with the
function in a6xx-pack.xml.h.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11581>

ir3: add ldg.a,stg.a which allow complex in-place offset calculation

The full form for ldg.a/stg.a offset is:
 g[reg_address + reg_offset << (imm_shift + 2) + imm_offset << 2]

where imm_shift is in [0, 3] and imm_offset is in [0, 3]

a6xx blob was found to produce a bit simplier offset calculations
for TES/TCS shaders in GTA V:

 [c002000a_03c14215] ldg.a.f32 r2.z, g[r1.y+((r2.z+1)<<2)], 3;
 [c0020004_01c14609] ldg.a.f32 r1.x, g[r1.y+((r1.x+3)<<2)], 1;

Our new syntax:
 stg.a.u32 g[r2.x+(r1.x+1)<<2], r5.x, 1
 stg.a.u32 g[r2.x+r1.x<<4+3<<2], r5.x, 1
 ldg.a.f32 r1.w, g[r1.y+(r1.w+1)<<2], 3
 ldg.a.f32 r1.w, g[r1.y+r1.w<<5+2<<2], 3

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11431>

freedreno/isa: add uoffset type to print positive-only offsets

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11431>

freedreno/computerator: pass iova of buffer to const register

The syntax is:
  @buf 32 (c2.x)
The "(c2.x)" is optional.

This makes possible to test stg, ldg, and global atomics.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11431>

anv: do not dereference VkPipelineMultisampleStateCreateInfo always

VkGraphicsPipelineCreateInfo.pMultisampleState is a pointer to a
VkPipelineMultisampleStateCreateInfo structure, and is ignored if the
pipeline has rasterization disabled.

Fixes a crash in one CTS tests that checks this.

Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11601>

turnip: early exit in tu6_draw_common to save cpu cycles

Improves Zink + drawoverhead perf up to 4%

Before:
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 3981
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 3977

After:
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 4136
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 4163

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11556>

turnip: do not re-emit same vs params

Improves drawoverhead perf through Zink up to 260%

Before:
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 1518
After:
  1, DrawElements ( 1 VBO| 0 UBO|  0    ) w/ no state change, 3981

This brings it close to Freedreno, which has around 4300.

In vkQuake vs params re-emission now occurs in 0.23% of draw calls.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11556>

intel/genxml: Remove old scratch fields on GFX version 12.5

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

anv: Add support for scratch on XeHP

Rework:
 * Jordan: Handle per_thread_scratch==0 in anv_scratch_pool_get_surf
 * Jordan: Update subslices in anv_scratch_pool_alloc
 * Jason: Clean up the patch a bit

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

iris: Add support for scratch on XeHP

Rework:
 * Jordan: Handle prog_data->total_scratch==0 in iris_upload_compute_walker
 * Jordan: Resolve iris_get_scratch_space conflict with e2c5ef6cd65
 * Jordan: Rebase on 4256f7ed584. broken
 * Ken: Mostly fixed the rebase
 * Jordan: Fix two small compilation issues
 * Jordan: Rebase on Ken's ("iris: Make a pin_scratch_space() helper")
 * Lionel: Fix a few bugs with scratch handles
 * Jason: Tidy the patch up a bit

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

iris: Add a MEMZONE_BINDLESS and uploader

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

intel/genxml: Add new ScratchSpaceBuffer fields on GFX version 12.5

Rework:
 * Jordan: Fix MEMZONE_BINDER_START detection
 * Jordan: Bump the IRIS_BINDLESS_SIZE to 8M

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

intel/fs: Implement load/store_scratch on XeHP

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

intel/fs: Implement spilling on XeHP

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

intel/isl: Add support for scratch buffers

XeHP adds support for a new surface type for scratch.  It's similar to
SURFTYPE_STRBUF in that it's a 2D array-of-struct format but the one
key difference is that the U coordinate is computed automatically based
on the thread ID and only the V coordinate is provided in the dataport
message.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

intel/genxml: Add SURFTYPE_SCRATCH on GFX version 12.5

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>

Revert "freedreno: Cooperate with tc to stop checking the BC for resource_busy()."

This reverts commit 5cb043cf82af4af88dbfe83ce1eff4cc41e2a766.

While the numbers were impressive for drawoverhead, it comes at the cost
of additional flushes, which for gmem access (what we actually care about
most!) would greatly increase the actual cost to render.  Also, gl_driver2
overhead is increased, probably due to spending time in the kernel for the
flushes.

drawoverhead's win came from the increased flushing causing the GPU to
start processing the buffers sooner on everything but test 1, which
already had some incremental flushing happening.  That was certainly not
intended by the change.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11599>

zink: flush pending clears if a resource is bound as a descriptor

ensure that resource data is in sync with the expected command stream

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: only queue deferred descriptor layout change on first bind or change

otherwise, the layout is either already correct or already queued

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: improve automatic layout transitions for sampler+image descriptors

binding new images may require auto-transitions for the other pipeline, and
unbinding an image requires that any persistent sampler binds have their layouts
updated

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: use bind counts to more accurately determine image descriptor's exact layout

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: only do deferred image barriers if layout changes

otherwise these should be handled by apps using glMemoryBarrier

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: split deferring of barriers to image and buffer functions

make this more explicit

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

zink: break out image descriptor layout into util function

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11590>

ci/windows: fix zink msvc build-rules

When I added the build-rules for zink to the windows build, I
accidentally added it to .windows-test-rules instead of
.windows-build-rules. This seems to trigger a build-error if we trigger
*just* a test due to a zink-change, but not a build.

Hopefully this fixes the problems Mike has had with Zink CI recently.

Fixes: a426d7c2643 ("ci/windows: enable msvc builds of zink")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11597>

microsoft/spirv_to_dxil: Add drive_location assignment

Acked-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10989>

microsoft/compiler: Switch io sort to use nir_sort_variables_with_modes

Acked-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10989>

nir: Add modes filter to nir_sort_variables

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10989>

nir: Add a function for sorting variables

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10989>

util: Add qsort_r/s args adapter for MSVC and BSD/macOS

MSVC's qsort_s behaves similarly to sort_r. Unfortunately, qsort_s's
compare function has the "context"/"args" as its first argument. BSD's
qsort_r has a different order than GNU's qsort_r. Finally, C11 added
qsort_s's which look like GNU's gsort_r.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10989>