This simplifies the code that iterates over the per-component values
found in the matching copy_entry struct and checks whether the
register regions that were copied to each component are similar enough
to be treated as a single (reswizzled) value which can be propagated
into the current instruction.
Aside from being scattered between opt_copy_propagation(),
try_copy_propagate(), and try_constant_propagate(), what I found
terribly confusing about the preexisting logic was that
opt_copy_propagation() tried to reorder the array of values according
to the swizzle of the instruction source, which meant one would have
had to invert the reordering applied at the top level in order to find
out which component to take from each value (we were just taking the
i-th component from the i-th value, which is not correct in general).
The saturate mask was also being swizzled incorrectly.
This consolidates the logic for matching multiple components of a
copy_entry into a single function which returns the result as a
regular src_reg on success, as if the copy had been performed with a
single MOV instruction copying all components of the src_reg into the
destination.
Fixes several ARB_vertex_program MOV test-cases from:
https://cgit.freedesktop.org/~kwg/piglit/log/?h=arb_program
Acked-by: Matt Turner <mattst88@gmail.com>
Scalar immediates used to be handled correctly by swizzle() (as the
identity) but since commit 58fa9d47b5 it
will corrupt the contents of the immediate. Vector immediates were
never handled correctly, but we had ad-hoc code to swizzle VF
immediates in the vec4 copy propagation pass. This takes care of
swizzling V and UV in addition.
v2: Don't implement swizzling of V/UV immediates (Matt). If you need
to swizzle an integer vector immediate in the future apply the
following diff to go back to v1:
--- a/src/mesa/drivers/dri/i965/brw_eu.c
+++ b/src/mesa/drivers/dri/i965/brw_eu.c
@@ -119,11 +119,10 @@ brw_swap_cmod(uint32_t cmod)
static unsigned
imm_shift(enum brw_reg_type type, unsigned i)
{
- assert(type != BRW_REGISTER_TYPE_UV && type != BRW_REGISTER_TYPE_V &&
- "Not implemented.");
-
if (type == BRW_REGISTER_TYPE_VF)
return 8 * (i & 3);
+ else if (type == BRW_REGISTER_TYPE_UV || type == BRW_REGISTER_TYPE_V)
+ return 4 * (i & 7);
else
return 0;
}
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
And replace brw_swizzle1() with brw_swizzle(). Seems slightly cleaner
and will allow reusing brw_swizzle() in the vec4 back-end more easily.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This fixes some use-after-free situations in dEQP when an xfb state is
removed, and then a clear is triggered, which only does a partial
validation. It would attempt to read the no-longer-valid buffers,
resulting in crashes.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
The OES extensions clarify this behaviour to differentiate between
per-sample invocation and per-sample interpolation. Using sampleid/pos
will force per-sample invocation but not per-sample interpolation.
See https://www.khronos.org/bugzilla/show_bug.cgi?id=1462
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Since commit 922be4eab, the expectation is that the query result
contains the correct value. Unfortunately swrast does not distinguish
between GL_SAMPLES_PASSED and GL_ANY_SAMPLES_PASSED. As a result, we
must fix up the query result in a post-draw fixup.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94274
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Tested-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Cc: "11.2" <mesa-stable@lists.freedesktop.org>
compute.c: In function ‘launch_grid’:
compute.c:435:20: warning: assignment discards ‘const’ qualifier from pointer target type [enabled by default]
info.input = input;
^
Maybe the pipe_grid_info::input field should be const void *?
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
From Section 4.4.5 (Uniform and Shader Storage Block Layout
Qualifiers) of the OpenGL 4.50 spec:
"The align qualifier makes the start of each block member have a
minimum byte alignment. It does not affect the internal layout
within each member, which will still follow the std140 or std430
rules. The specified alignment must be a power of 2, or a
compile-time error results.
The actual alignment of a member will be the greater of the
specified align alignment and the standard (e.g., std140) base
alignment for the member's type. The actual offset of a member is
computed as follows: If offset was declared, start with that
offset, otherwise start with the next available offset. If the
resulting offset is not a multiple of the actual alignment,
increase it to the first offset that is a multiple of the actual
alignment. This results in the actual offset the member will have.
When align is applied to an array, it affects only the start of
the array, not the array's internal stride. Both an offset and an
align qualifier can be specified on a declaration.
The align qualifier, when used on a block, has the same effect as
qualifying each member with the same align value as declared on
the block, and gets the same compile-time results and errors as if
this had been done. As described in general earlier, an individual
member can specify its own align, which overrides the block-level
align, but just for that member.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
The old comment was for the location not the offset, we now use
the field for block members so mention that also.
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
In this patch we also copy the offset value from the ast and
implement offset linking rules by adding it to the record_compare()
function.
From Section 4.4.5 (Uniform and Shader Storage Block Layout Qualifiers)
of the GLSL 4.50 spec:
"Two blocks linked together in the same program with the same block
name must have the exact same set of members qualified with
offset and their integral-constant-expression values must be the
same, or a link-time error results."
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
This implements the rules for the offset qualifier on block members.
From Section 4.4.5 (Uniform and Shader Storage Block Layout Qualifiers)
of the GLSL 4.50 spec:
"The offset qualifier can only be used on block members of blocks
declared with std140 or std430 layouts."
...
"It is a compile-time error to specify an offset that is smaller than
the offset of the previous member in the block or that lies within the
previous member of the block."
...
"The specified offset must be a multiple of the base alignment of the
type of the block member it qualifies, or a compile-time error results."
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Global in validation is already handled, this will do the validation
for variables, blocks and block members.
This fixes some CTS tests for the new enhanced layouts transform
feedback qualifiers.
V2: add some more valid input flags
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Previously interface blocks were giving the global default flags of
uniform blocks. This meant we could not check for invalid qualifiers
on interface blocks because they always contained invalid flags.
This changes parsing so that interface blocks now get an empty
set of layouts.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
If the following patch we will stop setting these layouts by default
on interface blocks, so we need to do this to avoid hitting the
assert.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
When BaseLevel > 0, we magnify the dimensions to fill out the size of
miplevels [0..BaseLevel). In particular, this was magnifying depth,
thinking that the depth doubles at each level. This is perfectly
reasonable for 3D textures, but dead wrong for array textures.
Changing the depth != 1 condition to a target == GL_TEXTURE_3D check
should make this only happen in the appropriate cases.
Fixes about 32 dEQP tests:
- dEQP-GLES31.functional.texture.gather.*.level_{1,2}
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
opt_vector_float() transforms several scalar MOV operations to a single
vectorial MOV.
This is done when those MOV covers all the components of the destination
register. So something like:
mov vgrf3.0.xy:D, 0D
mov vgrf3.0.w:D, 1065353216D
mov vgrf3.0.z:D, 0D
is transformed in:
mov vgrf3.0:F, [0F, 0F, 0F, 1F]
But there are cases where not all the components are written. For
example, in:
mov vgrf2.0.x:D, 1073741824D
mov vgrf3.0.xy:D, 0D
mov vgrf3.0.w:D, 1065353216D
mov vgrf4.0.xy:D, 1065353216D
mov vgrf4.0.w:D, 0D
mov vgrf6.0:UD, u4.xyzw:UD
Nor vgrf3 nor vgrf4 .z components are written, so the optimization is
not applied.
But it could be applied anyway with the components covered, using a
writemask to select the ones written. So we could transform it in:
mov vgrf2.0.x:D, 1073741824D
mov vgrf3.0.xyw:F, [0F, 0F, 0F, 1F]
mov vgrf4.0.xyw:F, [1F, 1F, 0F, 0F]
mov vgrf6.0:UD, u4.xyzw:UD
This commit does precisely that: opportunistically apply
opt_vector_float() when possible.
total instructions in shared programs: 7124660 -> 7114784 (-0.14%)
instructions in affected programs: 443078 -> 433202 (-2.23%)
helped: 4998
HURT: 0
total cycles in shared programs: 64757760 -> 64728016 (-0.05%)
cycles in affected programs: 1401686 -> 1371942 (-2.12%)
helped: 3243
HURT: 38
v2: change vectorize_mov() signature (Matt).
v3: take in account predicates (Juan).
v4 [mattst88]: Update shader-db numbers. Fix some whitespace issues.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
screen may still be used by other resources that are not yet freed.
To correctly fix this there will be a need to account for resources
differently, but this quick fix is not any worse than the original
code that leaked screens anyway.
Reviewed-by: Brian Paul <brianp@vmware.com>
No shader-db changes, but does recognize some extract_u16 which enables
the next patch to optimize some code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Two shaders that appear in Unigine benchmarks (Heaven and Valley) unpack
three bytes from an integer and convert each into a float:
float((val >> 16u) & 0xffu)
float((val >> 8u) & 0xffu)
float((val >> 0u) & 0xffu)
Instead of shifting, masking, and type converting like this:
shr(8) g15<1>UD g25<8,8,1>UD 0x00000010UD
and(8) g16<1>UD g15<8,8,1>UD 0x000000ffUD
mov(8) g17<1>F g16<8,8,1>UD
shr(8) g18<1>UD g25<8,8,1>UD 0x00000008UD
and(8) g19<1>UD g18<8,8,1>UD 0x000000ffUD
mov(8) g20<1>F g19<8,8,1>UD
and(8) g21<1>UD g25<8,8,1>UD 0x000000ffUD
mov(8) g22<1>F g21<8,8,1>UD
i965 can simply extract a byte and convert to float in a single
instruction:
mov(8) g17<1>F g25.2<32,8,4>UB
mov(8) g20<1>F g25.1<32,8,4>UB
mov(8) g22<1>F g25.0<32,8,4>UB
This patch implements the first step: recognizing byte extraction. A
later patch will optimize out the conversion to float.
instructions in affected programs: 28568 -> 27450 (-3.91%)
helped: 7
cycles in affected programs: 210076 -> 203144 (-3.30%)
helped: 7
This patch decreases the number of instructions in the two Unigine
programs by:
#1721: 4520 -> 4374 instructions (-3.23%)
#1706: 3752 -> 3582 instructions (-4.53%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This resolves some order dependencies between the already existing
callback the newly created one.
Tested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
There is a limit of 10 display connections, which was a
problem for apps/tests that were continuously opening/closing display
connections.
This fix uses XAddExtension() and XESetCloseDisplay() to keep track
of the status of the display connections from the X server, freeing
mesa-related data as X displays get destroyed by the X server.
Poster child is the VTK "TimingTests"
Tested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
glXCreatePixmap() may specify a GLX_TEXTURE_FORMAT_RGB_EXT format
for an RGBA resource, causing us to create an RGBX view for an
RGBA resource, a combination vgpu10 does not support.
When this is detected, change the request to create an RGBA view
instead.
Reviewed-by: Brian Paul <brianp@vmware.com>
With this patch, make sure the shader resource view is properly created
before referencing it in the generate mipmap command.
Reviewed-by: Brian Paul <brianp@vmware.com>
If running with a software renderer backend, the timeout may be
insufficient, and we don't want to release busy buffers too early.
In practice, SVGA gpu lockups are extremely rare.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Reported-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviwed-by: Brian Paul <brianp@vmware.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
dEQP-GLES31.functional.fbo.no_attachments.maximums.{all,height,size,width}
started hitting assertion failures when emitting SURFACE_STATE, after
commit e8fd60e789 where Samuel increased the maximum viewport size to
32768, from 16384.
MaxFramebufferWidth/Height were being set to the maximum viewport size,
but are actually limited by the SURFACE_STATE width/height field range,
which is 16384 on Gen7+ (where ARB_framebuffer_no_attachments is
exposed). So, reduce these to 16384 explicitly.
Fixes assert fails in the above mentioned dEQP tests. (Those tests
still fail, however.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
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.
Fixes four dEQP subtests:
dEQP-GLES31.functional.shaders.builtin_functions.precision.acos.
highp_compute.{scalar,vec2,vec3,vec4}
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This will allow us to share the implementation while using different
polynomials for asin() and acos().
Francisco Jerez did this in the SPIR-V front-end; I'm merely porting
his idea to the GLSL world.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
From section 6.2 ("State Tables") of the GL 2.1 specification
(the text also appears in the GL 3.0 and ES 3.1 specifications):
"However, state variables for which IsEnabled is listed as the query
command can also be obtained using GetBooleanv, GetIntegerv, GetFloatv,
and GetDoublev."
GL_DEBUG_OUTPUT, GL_DEBUG_OUTPUT_SYNCHRONOUS, and GL_FRAGMENT_SHADER_ATI
were missing from the glGet*() functions. All other IsEnabled() pnames
look to be present, as far as I can tell.
Fixes 8 dEQP-GLES31.functional.debug.state_query subtests:
debug_output[_synchronous]_get{boolean,float,integer,integer64}.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>