Same as commit 3654b6d4 to the fs backend.
total instructions in shared programs: 5945788 -> 5945787 (-0.00%)
instructions in affected programs: 36 -> 35 (-2.78%)
helped: 1
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Same as commit c4fab711 to the fs backend.
total instructions in shared programs: 5945998 -> 5945788 (-0.00%)
instructions in affected programs: 74665 -> 74455 (-0.28%)
helped: 399
HURT: 180
It hurts some programs because we make no attempts in the vec4 backend
to avoid MADs if they have constant (or vector uniform) arguments.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Skylake+ doesn't support setting a depth buffer to a 1D surface but it
does allow pretending it's a 2D texture with a height of 1 instead.
This fixes the GL_DEPTH_COMPONENT_* tests of the copyteximage piglit
test (and also seems to avoid a subsequent GPU hang).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89037
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Null surfaces are going to be useful to have something to point
unbound image units to, as the ARB_shader_image_load_store extension
requires us to behave deterministically in cases where some shader
tries to access an unbound image unit: Invalid stores and atomics are
supposed to be discarded and invalid loads are supposed to return
zero, which is precisely what the null surface does.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Bump the number of supported image uniforms to 32 (Ken).
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It doesn't really improve locality of texture fetches, quite the
opposite it's a waste of memory bandwidth and space due to tile
alignment.
v2: Check mt->logical_height0 instead of mt->target (Ken). Add short
comment explaining why they shouldn't be tiled.
Reviewed-by: Neil Roberts <neil@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Negation of UD/UW sources behaves the same as for D/W sources, taking
the two's complement of the source, except for bitwise logical
operations on Gen8 and up which take the one's complement. Fixes
crash in a GLSL shader with subtraction of two unsigned values.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Scalar registers are required to have zero stride, fix the
regs_written calculation not to assume that the instruction writes
zero registers in that case.
v2: Rename CEILING() to DIV_ROUND_UP(). (Matt, Ken)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Using 'ralloc*(this, ...)' is wrong if the object has automatic
storage or was allocated through any other means. Use normal dynamic
memory instead.
Reviewed-by: Matt Turner <mattst88@gmail.com>
The only reason why you need a vec4_visitor to construct a
vec4_instruction is to initialize vec4_instruction::ir and
::annotation. Instead set them from vec4_visitor::emit() just like
fs_visitor does.
Reviewed-by: Matt Turner <mattst88@gmail.com>
One should be able to manipulate i965 IR without pulling the whole
FS/VEC4 visitor classes -- Optimization passes and other
transformations would ideally be visitor-agnostic. Among other issues
this avoids a circular dependency between the header file where such
visitor-agnostic code will be defined and the main FS/VEC4 header
where both IR (layer below) and visitor (layer above) happen to be
defined.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Right now virtual GRF book-keeping and allocation is performed in each
visitor class separately (among other hundred different things),
leading to duplicated logic in each visitor and preventing layering as
it forces any code that manipulates i965 IR and needs to allocate
virtual registers to depend on the specific visitor that happens to be
used to translate from GLSL IR.
v2: Use realloc()/free() to allocate VGRF book-keeping arrays (Connor).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cubemap array images are unlike cubemap array samplers in that they don't need
an additional coordinate to index individual cubemaps in the array, instead
they behave like a 2D array of 6n layers, with n the number of cubemaps in the
array. Take this exception into account.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Some people have complained that code using the CEILING() macro is
difficult to understand because it's not immediately obvious what it
is supposed to do until you go and look up its definition. Use a more
descriptive name that matches the similar utility macro in the Linux
kernel.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Without this when an application issues that query, it would try to
wait the result from the gpu, and since no query has been actually
issued, it will wait forever.
Signed-off-by: Tiziano Bacocco <tizbac2@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Add a specific optimisation pass for NV50 to check whether SRC0 or SRC1 is
a MOV dst, IMM. If so: fold the IMM in and try to drop the MOV. Must be
done post-RA because it requires that SDST == SSRC2.
V2: improve readability and add comments to clarify decisions
V3: Remove redundant code... compiler already attempts to put the IMM in
SSRC1
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
But don't enable generation of it in the opProperties, because we can't
guarantee the SDST==SRC2 constraint until after register assignment. We'll
add a post-RA folding pass to utilise this.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Add emission rules for negative and saturate flags for MAD 4-byte opcodes,
and get rid of some of the constraints. Obviously tested with a wide variety
of shaders.
V2: Document MAD as supported short form
V3: Split up IMM from short-form modifiers
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
The old way made it impossible for the optimizer to reason about what
was going on. The new way is the same number of instructions (the neg
gets folded into the cvt) but enables the optimizer to be cleverer if
comparing to a constant (most common case). [The optimizer is presently
not sufficiently clever to work this out, but it could relatively easily
be made to be. The old way would have required significant complexity to
work out.]
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Printing instructions doesn't modify them, so we can mark the parameter
const.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
+82 Piglits - 100% of border color tests now pass on Haswell.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Cc: mesa-stable@lists.freedesktop.org
This should have no effect, but will make it easier to implement other
bug fixes.
v2: Eliminate "unsigned one" local; just use the value where necessary.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: mesa-stable@lists.freedesktop.org
The hardware's integer luminance formats are completely unusable;
currently we fall back to RGBA. This means we need to override
the texture swizzle to obtain the XXX1 values expected for luminance
formats.
Fixes spec/EXT_texture_integer/texwrap formats bordercolor [swizzled]
on Broadwell - 100% of border color tests now pass on Broadwell.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Cc: mesa-stable@lists.freedesktop.org
This provides for atomic addition, which will be used by an upcoming
shader-cache patch. A simple test is added to "make check" as well.
Note: The various O/S functions differ on whether they return the
original value or the value after the addition, so I did not provide
an add_return() macro which would be sensitive to that difference.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Previously, the set_insert function would bail early if it found a deleted
slot that it could re-use. However, this is a problem if the key being
inserted is already in the set but further down the list. If this happens,
the element ends up getting inserted in the set twice. This commit makes
it so that we walk over all of the possible entries for the given key and
then, if we don't find the key, place it in the available free entry we
found.
Reviewed-by: Eric Anholt <eric@anholt.net>
Previously, the hash_table_insert function would bail early if it found a
deleted slot that it could re-use. However, this is a problem if the key
being inserted is already in the hash table but further down the list. If
this happens, the element ends up getting inserted in the hash table twice.
This commit makes it so that we walk over all of the possible entries for
the given key and then, if we don't find the key, place it in the available
free entry we found.
Reviewed-by: Eric Anholt <eric@anholt.net>
For framebuffer completeness checks, consider renderbuffers as not
layered. Previously, they would have counted as layered if a layered
textured had previously been bound to the same attachment point. This
could cause framebuffer completeness checks to incorrectly fail with
GL_FRAMEBUFFER_INCOMPLETE_LAYER_TARGETS, even if no layered attachments
were present.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89026
Blits to or from a y-tiled surface must always be a multiple of the tile size.
From page 16 of the HSW PRM
(https://01.org/linuxgraphics/sites/default/files/documentation/intel-gfx-prm-osrc-hsw-memory-views.pdf#16)
"The pitch of a tiled enclosing region must be an integral number of tile
widths"
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
