The pass only looks at var load/store intrinsics, not input load/store
intrinsics, so assert that we don't see the other type.
v2: Adjust comment indentation.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
v2:
- Use helper to add a new source to the texture instruction.
v3:
- Use nir_tex_instr_src_index() to simplify the patch (Jason).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
We only need to add a check to validate output locations here. For
inputs with invalid locations we will fail to link when we can't
find a matching output in the same (invalid) location.
v2: compute location slots properly depending on shader stage and
variable type / direction
Fixes:
KHR-GL45.enhanced_layouts.varying_location_limit
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
We won't split varyings marked as always active because there
is no point in doing so. This means we need to mark both
sides of the interface as always active otherwise we will have
a mismatch and start removing things we shouldn't.
Reviewed-by: Eric Anholt <eric@anholt.net>
This is intended to be called before nir_lower_io() so that we
can do some linking optimisations with the results. It can also
be used with drivers that don't use nir_lower_io() at all such
as RADV.
v2: pass mode mask rather than first and last stage integer.
Reviewed-by: Eric Anholt <eric@anholt.net>
ssize_t is a GNU extension and is not available on Windows or MacOS.
Instead, we use intptr_t which should be effectively equivalent and is
part of the C standard. This should fix the Windows and Mac OS builds.
Fixes: 3af1c82989
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103253
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Tested-by: Vinson Lee <vlee@freedesktop.org>
Since blob.h moved up to src/compiler the test should include that
instead of src/compiler/glsl
fixes: 0e3bd56c6e ("compiler: Move blob up a level")
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This looks like a copy+paste error. They don't actually write into that
variable as would be implied by putting the return there.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable@lists.freedesktop.org
We didn't fold correctly in the case of 0x1 because we never let the
loop counter hit 0. Switching it to bit >= 0 solves this problem.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
There are certain advantages to using uint8_t internally such as
well-defined arithmetic on all platforms. However, interfaces that
work in terms of raw data should use a void* type.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
These helpers not only call blob_reserve_bytes but also make sure that
the blob is properly aligned as if blob_write_* were called.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Despite the name, it could only be used if you immediately wrote to the
pointer. Noboby was using it outside of one test, so clearly this
behavior wasn't that useful. Instead, make it return an offset into the
data buffer so that the result isn't invalidated if you later write to
the blob. In conjunction with blob_overwrite_bytes(), this will be
useful for leaving a placeholder and then filling it in later, which
we'll need to do for handling phi nodes when serializing NIR.
v2 (Jason Ekstrand):
- Detect overflow in the offset + to_write computation
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
These can be used to easily count up the number of bytes that will be
required by "writing" it into the NULL blob.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
There's no reason why that tiny bit of memory needs to be on the heap.
We always put blob_reader on the stack, so why not do the same with the
writable blob.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
We're going to want to use the blob for Vulkan pipeline caching so it
makes sense to have it in libcompiler not libglsl.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Otherwise we could have a failure followed by a smaller write that
succeeds and get a corrupted blob. If we ever OOM, we should stop.
v2 (Jason Ekstrand):
- Initialize the new boolean member in create_blob
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Cc: mesa-stable@lists.freedesktop.org
Otherwise, if you have a large read fail and then try to do a small
read, the small read may succeed even though it's at the wrong offset.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Cc: mesa-stable@lists.freedesktop.org
For TGSI-based drivers, st_glsl_to_tgsi records this information.
For NIR-based drivers, nir_shader_gather_info() will do so.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I'd like to put this sort of metadata in the shader_info structure,
rather than adding more things to gl_program.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
With the ssao demo from Vulkan demos:
radv/rx480: 440->440fps
anv/haswell: 24->34 fps
The demo does a 0->32 loop across a ubo with 32 members.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
I've been doing this inside of vc4, but vc5 wants it as well and it may be
useful for other drivers (Intel has a related path for pre-gen6 with MRT,
and freedreno had a TGSI path for it at one point).
This required defining a common enum for the standard comparison
functions, but other lowering passes are likely to also want that enum.
v2: Add to meson.build as well.
Acked-by: Rob Clark <robdclark@gmail.com>
Unlike uniforms, the limit on shared memory size is not called out
explicitly in the list of things that cause linker errors, but presumably
that's just an oversight in the spec.
Fixes dEQP-GLES31.functional.debug.negative_coverage.{callbacks,get_error,log}.compute.exceed_shared_memory_size_limit
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
c7affbf687 enabled GLSLOptimizeConservatively on some
drivers. The idea was to speed up compile times by running
the GLSL IR passes only once each time do_common_optimization()
is called. However loop unrolling can create a big mess and
with large loops can actually case compile times to increase
significantly due to a bunch of redundant if statements being
propagated to other IRs.
Here we make sure to clean things up before moving on.
There was no measureable difference in shader-db compile times,
but it makes compile times of some piglit tests go from a couple
of seconds to basically instant.
The shader-db results seemed positive also:
Totals:
SGPRS: 2829456 -> 2828376 (-0.04 %)
VGPRS: 1720793 -> 1721457 (0.04 %)
Spilled SGPRs: 7707 -> 7707 (0.00 %)
Spilled VGPRs: 33 -> 33 (0.00 %)
Private memory VGPRs: 3140 -> 2060 (-34.39 %)
Scratch size: 3308 -> 2180 (-34.10 %) dwords per thread
Code Size: 79441464 -> 79214616 (-0.29 %) bytes
LDS: 436 -> 436 (0.00 %) blocks
Max Waves: 558670 -> 558571 (-0.02 %)
Wait states: 0 -> 0 (0.00 %)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
The old code assumed that loop terminators will always be at
the start of the loop, resulting in otherwise unrollable
loops not being unrolled at all. For example the current
code would unroll:
int j = 0;
do {
if (j > 5)
break;
... do stuff ...
j++;
} while (j < 4);
But would fail to unroll the following as no iteration limit was
calculated because it failed to find the terminator:
int j = 0;
do {
... do stuff ...
j++;
} while (j < 4);
Also we would fail to unroll the following as we ended up
calculating the iteration limit as 6 rather than 4. The unroll
code then assumed we had 3 terminators rather the 2 as it
wasn't able to determine that "if (j > 5)" was redundant.
int j = 0;
do {
if (j > 5)
break;
... do stuff ...
if (bool(i))
break;
j++;
} while (j < 4);
This patch changes this pass to be more like the NIR unrolling pass.
With this change we handle loop terminators correctly and also
handle cases where the terminators have instructions in their
branches other than a break.
V2:
- fixed regression where loops with a break in else were never
unrolled in v1.
- fixed confusing/wrong naming of bools in complex unrolling.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
do-while loops can increment the starting value before the
condition is checked. e.g.
do {
ndx++;
} while (ndx < 3);
This commit changes the code to detect this and reduces the
iteration count by 1 if found.
V2: fix terminator spelling
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
These instructions will be executed on every iteration of the loop
we cannot drop them.
V2:
- move removal of unreachable terminators from the terminator list
to the same place they are removed from the IR as suggested by
Nicolai.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
This gets pretty much the entire classic tree building, as well as
i965, including the various glapis. There are some workarounds for bugs
that are fixed in meson 0.43.0, which is due out on October 8th.
I have tested this with piglit using glx.
v2: - fix typo "vaule" -> "value"
- use gtest dep instead of linking to libgtest (rebase error)
- use gtest dep instead of linking against libgtest (rebase error)
- copy the megadriver, then create hard links from that, then delete
the megadriver. This matches the behavior of the autotools build.
(Eric A)
- Use host_machine instead of target_machine (Eric A)
- Put a comment in the right place (Eric A)
- Don't have two variables for the same information (Eric A)
- Put pre_args at top of file in this patch (Eric A)
- Fix glx generators in this patch instead of next (Eric A)
- Remove -DMESON hack (Eric A)
- add sha1_h to mesa in this patch (Eric A)
- Put generators in loops when possible to reduce code in
mapi/glapi/gen (Eric A)
v3: - put HAVE_X11_PLATFORM in this patch
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This was missed in a rebase, and doesn't affect radv or anv, only i965.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
In truth gtest is an external dependency that upstream expects you to
"vendor" into your own tree. As such, it makes sense to treat it more
like a dependency than an internal library, and collect it's
requirements together in a dependency object.
v2: - include with -isystem instead of setting compiler args (Eric)
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
It's inside an if-statement that already checks that the variables are
not NULL.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
The optimization as done in opt_copy_propagation would have to be
removed in the next patch. If we just eliminate that optimization
altogether, shader-db results, even on platforms that use NIR, are hurt
quite substantially. I have not investigated why NIR isn't picking up
the slack here.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Instead of generating a sequence like:
run_default = true;
if (i == 3) // some label that appears after default
run_default = false;
if (i == 4) // some label that appears after default
run_default = false;
...
if (run_default) {
...
}
generate something like:
run_default = !((i == 3) || (i == 4) || ...);
if (run_default) {
...
}
This eliminates one use of conditional assignment, and it enables the
elimination of another.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Previously the instruction stream was walked looking for comparisons
with case-label values. This should generate nearly identical code.
For at least fs-default-notlast-fallthrough.shader_test, the code is
identical.
This change will make later changes possible.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
The values being compared are scalars, so these are the same. While
I'm here, simplify the run_default condition to just deref the flag
(instead of comparing a scalar bool with true).
There is a bit of extra change in this patch. When constructing an
ir_binop_equal ir_expression, there is an assertion that the types are
the same. There is no such assertion for ir_binop_all_equal, so
passing glsl_type::uint_type with glsl_type::int_type was previously
fine. A bunch of the code motion is to deal with that.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
This happens to work now because ir_binop_all_equal is used. This
causes vector typed init-expressions to produce scalar Boolean values
after comparison.
The next commit changes ir_binop_all_equal to ir_binop_equal. Vector
typed init-expressions will then produce vector Boolean values, and, in
debug builds, the ir_assignment constructor will fail an assertion.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Mostly tabs-before-spaces, but there was some other trivium too.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
