The home-grown heap scheme (which is ultra-simple but probably not good
to always allocate and memset such a chunk of memory up front) was a
remnant of fdre (where the ir originally came from). But since we have
ralloc in mesa, lets just use that instead.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Normally this would never happen (constant-propagation in NIR would
eliminate the instruction), except it does happen for 'undef' which
we turn into immed 0.0 for bookkeeping purposes.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
See a7eb12d0.. but that wasn't restrictive enough. Fixes
dEQP-GLES3.functional.rasterization.primitives.line_strip_wide, and
similar
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We seem to need range reduction to get sane results. Fixes glmark2
jellyfish bench, and a whole bunch of
dEQP-GLES3.functional.shaders.builtin_functions.precision.{sin,cos,tan}.*
v2: squashed in android build fixes from Rob Herring
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Currently if the texture binding is changed, emit_fs_consts()
is triggered to update texture scaling factor for
rectangle texture or texture buffer size in the constant buffer.
But the update is only relevant if the texture binding includes
a rectangle texture or a texture buffer.
To eliminate the unnecessary constant buffer updates due to other texture
binding changes, a new flag SVGA_NEW_TEXTURE_CONSTS will be used
to trigger fragment shader constant buffer update when a rectangle texture
or a texture buffer is bound.
With this patch, the number of constant buffer updates in Lightsmark2008
reduces from hundreds per frame to about 28 per frame.
Reviewed-by: Brian Paul <brianp@vmware.com>
Instead of unconditionally mark the texture subresource dirty at transfer map,
we'll set the dirty bit for write transfer only.
Tested with lightsmark2008 and glretrace.
Reviewed-by: Brian Paul <brianp@vmware.com>
With this patch, when running in hardware version 11, we'll use
SVGA3D_QUERYTYPE_OCCLUSION query type for PIPE_QUERY_OCCLUSION_PREDICATE
and return TRUE if samples-passed count is greater than 0.
Fixes glretrace/solidworks2012_viewport running in hardware version 11.
Reviewed-by: Brian Paul <brianp@vmware.com>
Currently, we always do a surface flush when we try to establish
a synchronized write transfer map. But if the subresource has not
been modified, we can skip the surface flush. In other words,
we only need to do a surface flush if the to-be-mapped subresource
has been modified in this command buffer.
With this patch, lightsmark2008 shows about 15% performance improvement.
Reviewed-by: Brian Paul <brianp@vmware.com>
Count depth/stencil, blend, sampler, etc. state objects separately
but just report the sum for the HUD. This change lets us use gdb to
see the breakdown of state objects in more detail.
Also, count sampler views too.
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Avoid leaking buffer allocated for task if an error has occured.
Coverity id: 1213929
Signed-off-by: Robert Foss <robert.foss@collabora.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Mark variables MAYBE_UNUSED to avoid unused-but-set-variable warnings
in release build.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
combineLd/St would combine, i.e. :
st u32 # g[$r2+0x0] $r2
st u32 # g[$r2+0x4] $r3
into:
st u64 # g[$r2+0x0] $r2d
But this is only valid if r2 contains an 8 byte aligned address,
which is not guaranteed for compute shaders
This commit checks for src0 dim 0 not being indirect when combining
loads / stores as combining indirect loads / stores may break alignment
rules.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Currently we were two restrictive, and would insert an output move in
cases like: MOV OUT[0], IN[0].xyzw
Loosen the restriction to allow the current instruction to appear in the
neighbor list but only at it's current possition.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Normally the offset in the group would be the same, but not always. For
example, in a sam(w) which only writes the 4th component.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
This *seems* like a hw bug, and maybe only applies to certain a4xx
variants/revisions. But setting the SRGB bit in sampler view state
(texconst0) causes invalid alpha for ASTC textures. Work around this
setting up a second texture state and using that to sample alpha
separately.
This way, srgb->linear conversion happens in hw *prior* to
interpolation.
This fixes 546 dEQP tests: dEQP-GLES3.functional.texture.*astc*srgb*
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We're trying to move to more of the new style intrinsics with include
the correct target name, and map directly to ISA instructions.
v2:
- Only do this with LLVM 3.8 and newer.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The range metadata tells LLVM the range of expected values for this intrinsic,
so it can do some additional optimizations on the result.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Switch boolean template arguments to typename template arguments of type
std::integral_constant<bool, VALUE>.
This allows the template argument unroller to easily be extended to enums.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
