Otherwise we end up with funny things like:
mov.f32f32 r0.x, r1.y
mov.f32f32 r0.x, r1.y
(It doesn't happen as much after fixing the problem w/ CP into phi src,
but it can still happen since we aren't too clever about generating phi
sources in the first place.)
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The block defining a phi source might not have been executed. If we
allow copy propagation, we could end up pointing to a src instruction in
the wrong block.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Been on my TODO list for a while. If nothing else this will make gdb
properly grok the opc_t enum.
This first step preserves ir3_instruction::category (with an added
assert that category matches what is encoded in opc_t). Next step is
to drop the category field (and arg to ir3_instr_create()), but that
is split into next commit for bisectability and so that we can run
piglit in the intermediate state to flush out any problems.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
They are compute-shader only and that's where the code for doing atomics on
shared variables lives so it seemes to make sense.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Our hardware requires an LOD for all texelFetch commands even if they are
on buffer textures. GLSL IR gives us an LOD of 0 in that case, but the LOD
is really rather meaningless. This commit allows other NIR producers to be
more lazy and not provide one at all.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
%ld and %lu aren't the right format specifiers for int64_t and uint64_t
on 32-bit (x86) systems. They're %zu on Linux and %Iu on Windows.
Use the standard C99 macros in hopes that they work everywhere.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
There may not be a previous block. In this case, there's no real work
to do, so just continue on to the next one.
v2: Update for bblock->prev() API change.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The bblock_t::prev/prev_const/next/next_const API returns bblock_t
pointers, rather than exec_nodes. So it's a bit surprising.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
lower_variable_index_to_cond_assign() did not handle system values.
gl_SampleMaskIn[] is a system value, and also an array. Accessing it
with a variable index would trigger an unreachable() assert.
Rather than adding a new EmitNoIndirectSystemValues flag, we simply
lower unconditionally. There is exactly one case where this occurs,
and for all current drivers, lowering produces optimal code. Even
for future drivers with 32x MSAA, it produces reasonable code.
Fixes Piglit's new samplemaskin-indirect test. Also fixes many ES31-CTS
tests when OES_sample_variables is enabled.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
In the first pass of implementing exact handling, I made a mistake with
search-and-replace. In particular, we only reallly handled exact/inexact
on the root of the tree. Instead, we need to check every node in the tree
for an exact/inexact match. As an example of this, consider the following
GLSL code
precise float a = b + c;
if (a < 0) {
do_stuff();
}
In that case, only the add will be declared "exact" and an expression that
looks for "b + c < 0" will still match and replace it with "b < -c" which
may yield different results. The solution is to simply bail if any of the
values are exact when matching an inexact expression.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The SIN and COS instructions on Intel hardware can produce values
slightly outside of the [-1.0, 1.0] range for a small set of values.
Obviously, this can break everyone's expectations about trig functions.
According to an internal presentation, the COS instruction can produce
a value up to 1.000027 for inputs in the range (0.08296, 0.09888). One
suggested workaround is to multiply by 0.99997, scaling down the
amplitude slightly. Apparently this also minimizes the error function,
reducing the maximum error from 0.00006 to about 0.00003.
When enabled, fixes 16 dEQP precision tests
dEQP-GLES31.functional.shaders.builtin_functions.precision.
{cos,sin}.{highp,mediump}_compute.{scalar,vec2,vec4,vec4}.
at the cost of making every sin and cos call more expensive (about
twice the number of cycles on recent hardware). Enabling this
option has been shown to reduce GPUTest Volplosion performance by
about 10%.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
See commit 3b0279a69 - this restriction is documented in the "Surface
Format" field of RENDER_SURFACE_STATE.
Looking at newer documentation, this restriction appears to exist on
Haswell, but no longer applies on Gen8+.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
This extension is identical to ARB_base_instance. Reuse the same
entrypoints.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
The extension spec was extended to also support ES. This functionality
is provided all the way back to ES 1.0.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
For adding .v4f32 like suffixes to intrinsics, taking special care for
scalar case, which was being often neglected.
This fixes invalid IR when doing mipmap filtering on SSE2 (the only
case where we'd use intrinsics with scalars.)
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Also avoid double-adding the *sampler2DMS types when the array ext is
enabled.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
I can't tell whether this actually matters, but we're creating function
signatures with this predicate, so it should probably match when SSBO's
are available.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
As the relevant extensions get implemented, the lines should be
uncommented. I believe this is (almost) everything needed for those GL
versions though.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Oddly a bunch of the features it adds are actually from ESSL 3.20. But
the spec is quite clear, oh well.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
We could unconditionally use these instrinsics, but performance with SSE2
would suck, as LLVM falls back to calling libm.
lp_test_arit.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
LLVM often can't determine the mask elements are all ones/zeros, and
there doesn't seem to be a good way to hint that.
Thanks to Roland Scheidegger for spotting and analyzing the issue.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Only provide a fallback for LLVM 3.3.
One less dependency on LLVM C++ interface.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
