radv: set the LayerId output usage mask if FS needs it

When the stage preceding FS doesn't export it the fragment shader
might read it, even if it's 0.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

panfrost: Update supported formats

Much of the format selection code was inherited from softpipe (!) of all
places, and a lot of it is accordingly cruft. Later if-elses were added
in random places to workaround missing formats at various points in
history. Clean up some of this.

Theoretically, any format we can texture from we can also render to. In
practice, there are a few corner cases that we need to disable
explicitly.

For one, we do have to restrict SCANOUT formats to workaround
buggy apps (in particular, dEQP which with --deqp-surface-type=window
under Weston will end up with RGB10_A2 and complain about low alpha
precision). Just be clearer about how/why.

Also, RGB5_A1 support is still broken; let's not worry about that quite
yet.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/mfbd: Cleanup format code selection

Rather than have random variables flying around and a long if-else
chain, use a switch. They're literally *designed* for this.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Cleanup blend switch

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/mfbd: Handle PIPE_FORMAT_B10G10R10A2_UNORM

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Handle PIPE_FORMAT_B10G10R10A2_UNORM

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Implement ES3-format writeout

We add support for writing out (via a blend shader):

   - RGBA4
   - RGB10_A2_UNORM
   - RGB10_A2_UINT
   - RGB5_A1_UNORM
   - R11G11B10_FLOAT

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Refactor blend infrastructure

We would like to permit keying blend shaders against the framebuffer
format, which requires some new blending abstractions.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Use unsigned blend patch offset

We would like the offset field to be unsigned, letting 0 represent "no
offset" and positive represent an offset.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Handle pure int formats

I'm not sure I'm totally comfortable with this, but conceptually neither
float nor pure-int formats require any format conversion, except size
conversion. Going from a shaderable format (fp32 or i16, for instance)
into a blendable format (fp16) is a separate question, one we can defer
momentarily while we're not interested in actually blending.

As an aside, I'd be fascinated by an integer-based blending
implementation.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/mfbd: Handle pure int formats

We see that the render target itself turns out to be typeless
(surprise!)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Set rt_count_2 for bpp>4 formats

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Implement preliminary float converters

We'll need some careful handling, but for now, get some baseline code
out for handling float formats in a blend shader.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Skip blend for REPLACE (shader)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Handle "blend disabled" blend shaders

Normally, disabled blend can definitely be fixed-function'd away, but
if a blend shader is used merely for format conversion rather than
blending, this code path can be nevertheless hit.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Route format through fixed-function blending

Not all framebuffer formats are supported by the fixed-function blender.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Pipe framebuffer format around

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Use Gallium framebuffer formats

Ideally, we would keep Galliumisms far away from the compiler;
unfortunately, Mesa hasn't standardized on system of format codes to be
shared across APIs and across drivers, so using Gallium formats is our
best bet in the short run.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Use fp16 exclusively while blending

We now have some preliminary fp16 support available. We're not able to
expose this for GLSL quite yet, but for internal blend shaders, we're
able to do control bitness ourselves just fine. So let's fp16 that
stuff!

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Remove opt_copy_prop_tex

Eventually this should be replaced by proper tex RA / not emitting so
many silly moves to begin with / better general copy prop. For now
remove it since it breaks things.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Fix scalarification

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Handle fp16 in embedded_to_inline_constants

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Eliminate redundant type convert

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Fix fp16 embedded constants

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Hoist mask field

Share a single mask field in midgard_instruction with a unified format,
rather than using separate masks for each instruction tag with
hardware-specific formats. Eliminates quite a bit of duplicated code and
will enable vec8/vec16 masks as well (which don't map as cleanly to the
hardware as we might like).

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Allow fp16 in scalar ALU

The packing is a little different, so implement that.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Implement f2u16 and friends

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Implement f2f16/f2f32

These conversions handle half-floats within the shader.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Verify src_bitsize == dst_bitsize

We can handle differing, but we'd prefer not to because there are
restrictions on sizing which aren't accounted for yet.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Simplify blend read

It's not clear where the extra indirection was from (older hardware or
just older blobs?)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: NIRify blend load scale/convert

The scale and type-convert can now be expressed in NIR, rather than MIR,
which is significantly more maintainable and demonstrates correctness of
the type conversion patches.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Fix blend constant scheduling bug

Blend constant conflicts run in two directions.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Implement upscaling type converts

Rather than using a dest_override, we upscale integers by using a half
field with a sign-extend bit. A variant of this trick should also work
for floats, but one step at a time!

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Move blend load/store into NIR

We have dedicated intrinsics to access the raw contents of the tile
buffer so we can use a dedicated NIR pass to lower appropriately for
blend shaders, rather than introducing a bizarre hardcoded blend
epilogue that only works for RGBA8_UNORM.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Use nir_dest_num_components

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Implement integer downsize ops

Oh, dear. No turning back now.

We begin implementing non-32-bit types, using downsizing integer type
conversions as the initial instructions. We implement them naively as
type-converting moves; substantially more efficient operation is
possible by copypropping the type conversion modifier, but this
optimization is not implemented here.

Size converting modifiers on Midgard allow an instruction to write to a
destination 1/2 the size, or to read from a source 1/2 the size. If we
need an extreme conversion (32-bit to 8-bit, for instance), multiple
type converting ops are chained together, which here is handled via an
algebraic pass.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Move scale from MIR to NIR

This begins the process of removing blend shader specific MIR into a
more general NIR lowering pass for formats.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/midgard: Passthrough nir_lower_framebuffer

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Extend clear colour packing

Eventually, this will allow packing clear colours for all formats,
including floating-point framebuffers, pure integer buffers, and special
formats. Currently, a few of these formats are supported, and many more
are handled through a generic Gallium colour packing path (which is not
a perfect fit for the hardware, but works for many formats and is a sane
default for the moment.)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost/mfbd: Include codes for float framebuffers

We see the hardware doesn't actually support float framebuffers in the
native sense -- rather, it just allows higher bpp framebuffers and lets
a blend shader / additional clear_color fields sort out the formats.
This will be.. interesting.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Prepare some code for MRT

Full MRT support is a while away, but in the mean time, we can remove
code that explicitly assumes nr_cbufs <= 0, to minimize the obstacles
we'll face later when we add the whole thing.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Use standard ALIGN_POT/INFINITY macros

We had vendored duplicates from pre-Mesa days; clean that up.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

egl: add glvnd symbols check

According to the spec [1], `__egl_Main` is the only symbol that needs to
be exported. We don't want applications directly linking against
libEGL_mesa.so (apps should always go through libEGL.so, regardless of
who is providing it), so we shouldn't export any other symbols either.

[1] https://github.com/NVIDIA/libglvnd/blob/master/include/glvnd/libeglabi.h
    (this header is the closest there is to a spec)

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

egl: rewrite entrypoints check

Part of the effort to replace shell scripts with portable python scripts.
I could've used a trivial `assert lines == sorted(lines)`, but this way
the caller is shown which entrypoint is out of order.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

mapi: add shared glapi symbols check

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

tu: add exported symbols check

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

vulkan: add symbols file

According to the Vulkan ICD spec [1], these two symbols must be exposed:
- vk_icdGetInstanceProcAddr
- vk_icdNegotiateLoaderICDInterfaceVersion

and this one is optional:
- vk_icdGetPhysicalDeviceProcAddr

[1] https://github.com/KhronosGroup/Vulkan-Loader/blob/master/loader/LoaderAndLayerInterface.md

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

meson: remove unused env_test

No longer used as of last commit :)

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

gles: use new symbols check script

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

gbm: sort symbols

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>

gbm: use new symbols check script

Note: the list in gbm-symbols.txt is the same as the one that was in
gbm-symbols-check, I just took the opportunity to sort it.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

egl: use new symbols check script

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

symbols-check: introduce new python script

I've re-written this in bash a couple times over the years, and then
I realised python is much more portable and already required by Mesa, so
we might as well make use of it.

I decided to still use the build system's NM instead of re-implementing
symbols extraction, to offload the complexity of keeping it compatible
with many systems (Linux, Unix, BSD, MacOS, etc.), especially when
cross-building.

This new script checks not only that nothing is exported when it
shouldn't be, but also that everything that should be exported is.
Sometimes, some symbols _can_ be exported but don't have to be, in which
case they can be prefixed with `(optional)`.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

nv50/ir/nir: implement load/store_global

required by OpenCL

v2: fix setting globalAccess

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Pierre Moreau <pierre.morrow@free.fr>

nv50/ir/nir: handle kernel inputs

required by OpenCL

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Pierre Moreau <pierre.morrow@free.fr>

nv50/ir/nir: don't assert on !main

required for OpenCL

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Pierre Moreau <pierre.morrow@free.fr>

nv50/ir/nir: parse system values first and stop for compute shaders

required by OpenCL

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Pierre Moreau <pierre.morrow@free.fr>

nir/lower_io: Don't use variable to get deref mode

Drivers only use lower_io for modes where pointers don't have a
meaningful value, and dereferences can always be traced back to a
variable. But there can be other modes, like global mode with
VK_EXT_buffer_device_address, where pointers cannot be traced back to a
variable, and lower_io would segfault on loads/stores of these since
nir_deref_instr_get_variable() would return NULL.

Just use the mode on the deref itself to filter out these modes before
we try to get the variable.

Fixes: 118a66df990 ("radv: Use NIR barycentric coordinates")
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: Don't optimize after lowering FS inputs

Currently this is done rather late in radv, after lowering booleans, so
it isn't safe to run additional optimizations that may add e.g. 1-bit
booleans. We could move the lowering parts earlier, but since right now
we only lower FS inputs and by this point all indirects have been
lowered away, there's no reason we should need to optimize anything.

One shader from Devil May Cry 5 was getting optimized, but only because
the optimization loop was working on 32-bit booleans which revealed an
opportunity that was hidden with 1-bit booleans, and we generated a
1-bit boolean which is invalid.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111092
Fixes: 118a66df9907772bb9e5503b736c95d7bb62d52c
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

android: amd/addrlib: add gfx10 support

Fix the following building error:

external/mesa/src/amd/addrlib/src/gfx10/gfx10addrlib.cpp:35:10:
fatal error: 'gfx10_gb_reg.h' file not found
         ^~~~~~~~~~~~~~~~
1 error generated.

Fixes: 78cdf9a ("amd/addrlib: add gfx10 support")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>
Acked-by: Marek Olšák <marek.olsak@amd.com>

android: amd/common/gfx10: add register JSON

The necessary Android makefile building rules are added
and the generation rules are simplified for readability

Fixes the following building errors:

external/mesa/src/amd/common/ac_llvm_build.c:1496:45:
error: use of undeclared identifier 'V_008F0C_IMG_FORMAT_8_UINT'
   case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_IMG_FORMAT_8_UINT; break;
                                             ^
Fixes: 74a26af ("amd/common/gfx10: add register JSON")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>
Acked-by: Marek Olšák <marek.olsak@amd.com>

android: radeonsi/gfx10: generate gfx10_format_table.h (v2)

Fix Android building rules for gfx10_format_table.h generated header

(v2) Add LOCAL_C_INCLUDES += $(intermediates)/radeonsi to fix error:

external/mesa/src/gallium/drivers/radeonsi/si_state.c:46:10:
fatal error: 'gfx10_format_table.h' file not found
         ^~~~~~~~~~~~~~~~~~~~~~
1 error generated.

Fixes: 0ffa229 ("radeonsi/gfx10: generate gfx10_format_table.h")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>
Acked-by: Marek Olšák <marek.olsak@amd.com>

android: virgl: remove unnecessary LOCAL_C_INCLUDES

The path could be imported automatically.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Reviewed-by: Mauro Rossi <issor.oruam@gmail.com>

android: vulkan/util: fix generating vk_enum_to_str.*

The gen_enum_to_str.py generates vk_enum_to_str.c and its header at once.
However, the makefiles incorrectly list both files parallel with the same
recipes. That means both two files may be generated simultaneously by two
processes. The generating files may be truncated by another process, as
shown below:

$ cd $OUT/obj/STATIC_LIBRARIES/libmesa_vulkan_util_intermediates/util
$ ls -l

-rw-rw-r-- 1 lh lh 193713 Jul  5 13:31 vk_enum_to_str.c
-rw-rw-r-- 1 lh lh   4609 Jul  5 13:31 vk_enum_to_str.d
-rw-rw-r-- 1 lh lh      0 Jul  5 16:21 vk_enum_to_str.h

Let one file depends on the other with empty recipe to avoid the issue.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

android: radv: import include paths from used libraries

It's unnecessary to manually add these include paths since they could
be imported automatically.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: anv: import include path of libmesa_nir

Add libmesa_nir to a common LOCAL_STATIC_LIBRARIES defined by
ANV_STATIC_LIBRARIES so that its include path can be imported
automatically. Then ANV_INCLUDES is unnecessary and could be
eliminated.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: anv: eliminate libmesa_anv_entrypoints

The dummy library libmesa_anv_entrypoints is totally unnecessary.
The four VULKAN_GENERATED_FILES could be generated and built in
libmesa_vulkan_common directly. The libraries using the generated
headers should get it via the exported include path.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: vulkan/util: fix export path

Export the correct include path so that the libraries use it can
get it automatically.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: radv: fix improper use of LOCAL_WHOLE_STATIC_LIBRARIES

The libmesa_git_sha1 is a dummy library. There is no reason to put
it into LOCAL_WHOLE_STATIC_LIBRARIES.

Move libmesa_vulkan_util to the vulkan.radv which really needs it.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: anv: fix improper use of LOCAL_WHOLE_STATIC_LIBRARIES

The libmesa_anv_entrypoints and libmesa_genxml are dummy libraries.
There is no reason to put them into LOCAL_WHOLE_STATIC_LIBRARIES.

Move libmesa_vulkan_util to the vulkan HAL which really needs it.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: radv: remove unused LOCAL_EXPORT_C_INCLUDE_DIRS

The vulkan module is the final HAL. No need to export its headers
since none will import it.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

android: anv: remove unused LOCAL_EXPORT_C_INCLUDE_DIRS

The vulkan module is the final HAL. No need to export its headers
since none will import it.

Signed-off-by: Chih-Wei Huang <cwhuang@linux.org.tw>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Emil Velikov <emil.velikov@collabora.com>

nir/loop_analyze: Pass nir_const_values directly to helpers

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Properly handle swizzles in loop conditions

This commit re-plumbs all of nir_loop_analyze to use nir_ssa_scalar for
all intermediate values so that we can properly handle swizzles.  Even
though if conditions are required to be scalars, they may still consume
swizzles so you could have ((a.yzw < b.zzx).xz && c.xx).y == 0 as your
loop termination condition.  The old code would just bail the moment it
saw its first non-zero swizzle but we can now properly chase the scalar
from the if condition to all the way to a, b, and c.

Shader-db results on Kaby Lake:

    total loops in shared programs: 4388 -> 4364 (-0.55%)
    loops in affected programs: 29 -> 5 (-82.76%)
    helped: 29
    HURT: 5

Shader-db results on Haswell:

    total loops in shared programs: 4370 -> 4373 (0.07%)
    loops in affected programs: 2 -> 5 (150.00%)
    helped: 2
    HURT: 5

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Refactor detection of limit vars

This commit reworks both get_induction_and_limit_vars() and
try_find_trip_count_vars_in_iand to return true on success and not
modify their output parameters on failure.  This makes their callers
significantly simpler.

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir: Add some helpers for chasing SSA values properly

There are various cases in which we want to chase SSA values through ALU
ops ranging from hand-written optimizations to back-end translation
code.  In all these cases, it can be very tricky to do properly because
of swizzles.  This set of helpers lets you easily work with a single
component of an SSA def and chase through ALU ops safely.

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Bail if we encounter swizzles

None of the current code knows what to do with swizzles.  Take the safe
option for now and bail if we see one.  This does have a small shader-db
impact but it is at least safe.

Shader-db results on Kaby Lake:

    total loops in shared programs: 4364 -> 4388 (0.55%)
    loops in affected programs: 5 -> 29 (480.00%)
    helped: 5
    HURT: 29

Shader-db results on Haswell:

    total loops in shared programs: 4373 -> 4370 (-0.07%)
    loops in affected programs: 5 -> 2 (-60.00%)
    helped: 5
    HURT: 2

Fixes: 6772a17acc8ee "nir: Add a loop analysis pass"
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Use new eval_const_* helpers in test_iterations

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Handle bit sizes correctly in calculate_iterations

The current code assumes everything is 32-bit which is very likely true
but not guaranteed by any means.  Instead, use nir_eval_const_opcode to
do the calculations in a bit-size-agnostic way.  We also use the new
constant constructors to build the correct size constants.

Fixes: 6772a17acc8ee "nir: Add a loop analysis pass"
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/loop_analyze: Fix phi-of-identical-alu detection

One issue was that the original version didn't check that swizzles
matched when comparing ALU instructions so it could end up matching
very different instructions.  Using the nir_instrs_equal function from
nir_instr_set.c which we use for CSE should be much more reliable.
Another was that the loop assumes it will only run two iterations which
may not be true.  If there's something which guarantees that this case
only happens for phis after ifs, it wasn't documented.

Fixes: 9e6b39e1d521 "nir: detect more induction variables"
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/instr_set: Expose nir_instrs_equal()

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir/builder: Use nir_const_value_for_* for constructing immediates

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir: Refactor nir_src_as_* constant functions

Now that we have the nir_const_value_as_* helpers, every one of these
functions is effectively the same except for the suffix they use so we
can easily define them with a repeated macro.  This also means that
they're inline and the fact that the nir_src is being passed by-value
should no longer really hurt anything.

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

nir: Add more helpers for working with const values

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

virgl: remove virgl_transfer_queue_lists

COMPLETED_LIST is always empty.  We only need one list.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

virgl: simplify virgl_transfer_queue_extend

We can reuse virgl_transfer_queue_find_pending.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

virgl: remove transfer after transfer_write

Now that virgl_transfer_queue_is_queued does not search
COMPLETED_LIST, we don't need to move transfers to that list.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

virgl: improve virgl_transfer_queue_is_queued

Search only the pending list and return immediately on the first
hit.

When the transfer queue was introduced, the function was used to
deal with

  write transfer -> draw -> write transfer

sequence.  It was used to tell if the second transfer intersects
with the first transfer. If yes, the transfer queue avoided
reordering the second transfer to before the draw (by flushing) in
case the draw uses the transferred data.

With the recent changes to the transfer code, the function is used
to deal with

  write transfer -> readback transfer

We want to avoid reordering the readback transfer to before the
first transfer (also by flushing).

In the old code, we needed to track the compeleted transfers as well
to avoid reordering.  But in the new code, a readback transfer is
guaranteed to see the data from the completed transfers (in other
words, it cannot be reoderered to before the already completed
transfers).  We don't need to search the COMPLETED_LIST.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

virgl: fix transfers_intersect for mipmaps

We never use transfers_intersect with textures, but fix it anyway to
avoid confusion.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

virgl: fix some false positives in transfers_overlap

Rewrite the function and check z/depth more carefully.  We
intentionally avoid u_box_test_intersection_2d because it returns
true when two boxes touch but do not intersect and can be confusing.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>

radeonsi/gfx10: enable primitive binning by default

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi/gfx10: implement primitive binning

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi: simplify primitive binning enablement

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi: set primitive binning tunables for dGPUs

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi: set FLUSH_ON_BINNING_TRANSITION when needed

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi/gfx10: use the new scan converter when binning is disabled

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi/gfx9: fix an oversight in primitive binning code

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi: use BREAK_BATCH instead of FLUSH_DFSM when CB_TARGET_MASK changes

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi/gfx10: don't expose unimplemented PIPE_CAP_QUERY_SO_OVERFLOW

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>

radeonsi/gfx10: launch 2 compute waves per CU before going onto the next CU

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Dave Airlie <airlied@redhat.com>