Here we skip the recreation of uniform storage if we are relinking
after a cache miss. This is improtant because uniform values may
have already been set by the application and we don't want to reset
them.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This will allow us to skip certain things when falling back to
a full recompile on a cache miss such as avoiding reinitialising
uniforms.
In this change we use it to avoid reading the program metadata
from the cache and skipping linking during a fallback.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
These may be lowered constant arrays or uniform values that we set before linking
so we need to cache the actual uniform values.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
For now this disables the shader cache when transform feedback is
enabled via the GL API as we don't currently allow for it when
generating the sha for the shader.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
V2: don't store pointers use an enum instead to flag what should be
restored. Also do the work in a helper that we will later use for
the subroutine remap table.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The three additional tables are AttributeBindings, FragDataBindings,
and FragDataIndexBindings.
The first table (AttributeBindings) was identified as missing by
trying to test the shader cache with a program that called
glGetAttribLocation.
Many thanks to Tapani Pälli <tapani.palli@intel.com>, as it was review
of related work that he had done previously that pointed me to the
necessity to also save and restore FragDataBindings and
FragDataIndexBindings.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The scenario is:
glShaderSource
glCompileShader <-- deferred due to cache hit of shader
glShaderSource <-- with new source code
glAttachShader
glLinkProgram <-- no cache hit for program
At this point we need to compile the original source when we
fallback.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The hash key for glsl metadata is a hash of the hashes of each GLSL
source string.
This commit uses the put_key/get_key support in the cache put the SHA-1
hash of the source string for each successfully compiled shader into the
cache. This allows for early, optimistic returns from glCompileShader
(if the identical source string had been successfully compiled in the past),
in the hope that the final, linked shader will be found in the cache.
This is based on the intial patch by Carl.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This uses disk_cache.c to write out a serialization of various
state that's required in order to successfully load and use a
binary written out by a drivers backend, this state is referred to as
"metadata" throughout the implementation.
This initial version is intended to work with all stages beside
compute.
This patch is based on the initial work done by Carl.
V2: extend the file's doxygen comment to cover some of the
design decisions.
V3:
- skip cache for fixed function shaders
- add int64 support
- fix glsl IR program parameter caching/restore and cache the
parameter values which are used by gallium backends.
- use new link status enum
V4:
- add compute program support
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
These are enough for the spir-v generator to handle UConvert
and SConvert operations, and fix the 4 tests in CTS.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This just adds the support at the spirv->nir level for the Int64
cap.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
For GS input arrays, we may turn a packed_type of ivec4 into an
array of ivec4s. We still want flat qualification.
Found by inspection. Not known to help anything.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Allow that capability if the driver indicates that it is supported, and
flag whether images are read-only/write-only in the nir_variable (based
on the NonReadable and NonWritable decorations), which drivers may need
to implement this.
Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
As soon as we support shaderStorageImageWriteWithoutFormat we can see
write-only images (sampled == 2) that don't have a format specified.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Passes the newly added piglit test for this extension on i965.
V2: Fix comments by Ilia.
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
TCS and TES inputs without an array size are implicitly sized to
gl_MaxPatchVertices. But TCS outputs are apparently not:
"If no size is specified, it will be taken from the output patch size
(gl_VerticesOut) declared in the shader."
Fixes dEQP-GLES31.functional.program_interface_query.program_output.
array_size.separable_tess_ctrl.var.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
OpenGL ES actually has spec text to prohibit this. It's just OpenGL
that's confusing.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
From GLSL ES 3.10 spec, section 4.1.9 "Arrays":
"If an array is declared as the last member of a shader storage block
and the size is not specified at compile-time, it is sized at run-time.
In all other cases, arrays are sized only at compile-time."
In desktop GLSL it is allowed to have unsized-arrays that are
not last, as long as we can determine that they are implicitly
sized, which is detected at link-time.
With this patch Mesa reports a compilation error as glslang does with
the following shader:
buffer SSBO { vec4 data[]; vec4 moreData;};
void main (void)
{
}
Fixes:
dEQP-GLES31.functional.debug.negative_coverage.log.shader.compile_compute_shader
dEQP-GLES31.functional.debug.negative_coverage.callbacks.shader.compile_compute_shader
dEQP-GLES31.functional.debug.negative_coverage.get_error.shader.compile_compute_shader
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously if you used MESA_GL_VERSION_OVERRIDE=3.3COMPAT, Mesa exposed
an OpenGL 3.3 compatibility profile context (with various unimplemented
features and bugs), but still refused to compile shaders with
#version 330 compatibility
This patch simply adds a small bit of plumbing to let that through.
Of course the same caveats apply: compatibility profile is still not
supported (and will not be supported), so there are no guarantees that
anything will work.
Tested-by: Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
For the on-disk shader cache we want to be able to differentiate
between a program that was linked and one that was loaded from cache.
V2:
- don't return the new enum directly to the application when queried,
instead return GL_TRUE or GL_FALSE as required. Fixes google-chrome
corruptions when using cache.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
As per the spec -
"The functions memoryBarrierShared() and groupMemoryBarrier() are
available only in compute shaders; the other functions are available
in all shader types."
Conform to this by adding another delegate to check for compute
shader support instead of only whether the current stage is compute
This allows some fragment shaders in Dirt Rally to compile
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
See "glsl: Rewrite atan2 implementation to fix accuracy and handling
of zero/infinity." for the rationale, but note that the instruction
count benefit discussed there is somewhat less important for the SPIRV
implementation, because the current code already emitted no control
flow instructions -- Still this saves us one hardware instruction per
scalar component on Intel SKL hardware.
Fixes the following Vulkan CTS tests on Intel hardware:
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.scalar
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec2
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec3
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec4
dEQP-VK.glsl.builtin.precision.atan2.mediump_compute.vec2
dEQP-VK.glsl.builtin.precision.atan2.mediump_compute.vec4
Note that most of the test-cases above expect IEEE-compliant handling
of atan2(±∞, ±∞), which this patch doesn't explicitly handle, so
except for the last two the test-cases above weren't expected to pass
yet. The reason they do is that the i965 back-end implementation of
the NIR fmin and fmax instructions is not quite GLSL-compliant (it
complies with IEEE 754 recommendations though), because fmin/fmax of a
NaN and a non-NaN argument currently always return the non-NaN
argument, which causes atan() to flush NaN to one and return the
expected value. The front-end should probably not be relying on this
behavior for correctness though because other back-ends are likely to
behave differently -- A follow-up patch will handle the atan2(±∞, ±∞)
corner cases explicitly.
v2: Fix up argument scaling to take into account the range and
precision of exotic FP24 hardware. Flip coordinate system for
arguments along the vertical line as if they were on the left
half-plane in order to avoid division by zero which may give
unspecified results on non-GLSL 4.1-capable hardware. Sprinkle in
some more comments.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This addresses several issues of the current atan2 implementation:
- Negative zero (and negative denorms which end up getting flushed to
zero) isn't handled correctly by the current implementation. The
reason is that it does 'y >= 0' and 'x < 0' comparisons to decide
on which side of the branch cut the argument is, which causes us to
return incorrect results (off by up to 2π) for very small negative
values.
- There is a serious precision problem for x values of large enough
magnitude introduced by the floating point division operation being
implemented as a mul+rcp sequence. This can lead to the quotient
getting flushed to zero in some cases introducing an error of over
8e6 ULP in the result -- Or in the most catastrophic case will
cause us to return NaN instead of the correct value ±π/2 for y=±∞
and x very large. We can fix this easily by scaling down both
arguments when the absolute value of the denominator goes above
certain threshold. The error of this atan2 implementation remains
below 25 ULP in most of its domain except for a neighborhood of y=0
where it reaches a maximum error of about 180 ULP.
- It emits a bunch of instructions including no less than three
if-else branches per scalar component that don't seem to get
optimized out later on. This implementation uses about 13% less
instructions on Intel SKL hardware and doesn't emit any control
flow instructions.
v2: Fix up argument scaling to take into account the range and
precision of exotic FP24 hardware. Flip coordinate system for
arguments along the vertical line as if they were on the left
half-plane in order to avoid division by zero which may give
unspecified results on non-GLSL 4.1-capable hardware. Sprinkle in
some more comments.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Timothy Arceri