When generating a sub-determinate matrix, a 3-element swizzle array was
indexed with clever inline boolean logic. Unfortunately, when i and j
are both 3, the index overruns the array, smashing the next variable on
the stack.
For 64 bit builds, the alignment of the 3-element unsigned array leaves
32 bits of spacing before the next local variable, hiding this bug. On
i386, a subcolumn pointer was smashed then dereferenced.
The adjusted polynomial coefficients come from the numerical
minimization of the L2 norm of the relative error. The old
coefficients would give a maximum relative error of about 15000 ULP in
the neighborhood around acos(x) = 0, the new ones give a relative
error bounded by less than 2000 ULP in the same neighborhood.
The dEQP "precision" test tries to verify that the reference functions
float sinh(float a) { return ((exp(a) - exp(-a)) / 2); }
float cosh(float a) { return ((exp(a) + exp(-a)) / 2); }
float tanh(float a) { return (sinh(a) / cosh(a)); }
produce the same values as the built-ins. We simplified away the
multiplication by 0.5 in the numerator and denominator, and apparently
this causes them not to match for exactly 1 out of 13,632 values.
So, put it back in, fixing the test, but making our code generation
(and precision?) worse.
The if/then/else block was bogus, as it can only take a scalar
condition, and we need to select component-wise. The GLSL IR
implementation of atan2 handles this by looping over components,
but I decided to try and do it vector-wise, and messed up.
For now, just bcsel. It means that we do the atan1 math even if
all components hit the quick case, but it works, and presumably
at least one component will hit the expensive path anyway.
We were botching this for negative numbers - floor of a negative rounds
the wrong way. Additionally, both results are supposed to retain the
sign of the original.
To fix this, just take the abs of both values, then put the sign back.
There's probably a better way to do this, but this works for now.
At a later stage we might want to split out the NIR specific [XXX:
which one was it], as to make things move obvious and rename the files
appropriately. This patch aims to split it out of nir.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jose Fonseca <jfonseca@vmware.com>
Allows us to remove the SCons workaround :-)
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jose Fonseca <jfonseca@vmware.com>
This way one can reuse it in glsl, nir or other infrastructure without
pulling nir as dependency.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jose Fonseca <jfonseca@vmware.com>
Currently it's an empty library, although it'll be used to store common
code between GLSL and NIR that is compiler specific (rather than generic
as the one in src/util).
XXX: strictly speaking we could add a python/mako parser to generate the
relevant files instead including builtin_type_macros.h in such a manner.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jose Fonseca <jfonseca@vmware.com>
Before we were asuming that a deref would either be something in a block or
something that we could pass off to NIR directly. However, it is possible
that someone would choose to load/store/copy a split structure all in one
go. We need to be able to handle that.
Previously, we were creating nir_deref's immediately. Now, instead, we
have an intermediate vtn_access_chain structure. While a little more
awkward initially, this will allow us to more easily do structure splitting
on-the-fly.
This currently sets the base and size of all push constants to the
entire push constant block. The idea is that we'll use the base and size
to eventually optimize the amount we actually push, but for now we don't
do that.
This allows us to first generate atomic operations for shared
variables using these opcodes, and then later we can lower those to
the shared atomics intrinsics with nir_lower_io.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Previously we were receiving shared variable accesses via a lowered
intrinsic function from glsl. This change allows us to send in
variables instead. For example, when converting from SPIR-V.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
