The GS has an interesting use for mul. Because the GS can emit multiple
vertices per input vertex, and it also has a unique count at the top of the URB
payload, the GS unit needs to be able to dynamically specify URB write offsets
(relative to the global offset). The documentation in the function has a very
good explanation from Paul on the mechanics.
This fixes around 2000 piglit tests on BSW.
v2:
Reworded commit message (Ben) no mention of CHV (Matt)
Change SHRT_MAX to USHRT_MAX (Ken, and Matt)
Update comment in code to reflect the use of UW (Ben)
Add Gen7+ assertion for the relevant GS code, since it won't work on Gen6- (Ken)
Drop the bogus hunk in emit_control_data_bits() (Ken)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=84777 (with many dupes)
Cc: "10.4 10.3 10.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
If an operation is the last one to read a register, the instruction
containing it can also include the op that has the next write to that
register.
total instructions in shared programs: 57486 -> 56995 (-0.85%)
instructions in affected programs: 43004 -> 42513 (-1.14%)
We were scheduling TLB operations as early as possible, and texture setup
as late as possible. When I introduced prioritization, I visually
inspected that an independent operation got moved above texture results
collection, which tricked me into thinking it was working (but it was just
because texture setup was being pushed late).
total instructions in shared programs: 57651 -> 57486 (-0.29%)
instructions in affected programs: 18532 -> 18367 (-0.89%)
Jason realized that we could fix the result of the CMP instruction on
Gen <= 5 by doing -(result & 1). Also do the resolves in the vec4
backend before use, rather than when the bool was created. The FS does
this and it saves some unnecessary resolves.
On Ironlake:
total instructions in shared programs: 4289762 -> 4287277 (-0.06%)
instructions in affected programs: 619430 -> 616945 (-0.40%)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This is a revert of commit 4656c14e ("i965/fs: Change the type of
booleans to UD and emit correct immediates") plus some small additional
fixes, like casting ctx->Const.UniformBooleanTrue to int and changing UD
to D in the ir_unop_b2f cases. Note that it's safe to leave 0x3f800000
as UD and as a literal it's more recognizable than 1065353216.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Three source instructions cannot directly source a packed vec4 (<0,4,1>
regioning) like vec4 uniforms, so we emit a MOV that expands the vec4 to
both halves of a register.
If these uniform values are used by multiple three-source instructions,
we'll emit multiple expansion moves, which we cannot combine in CSE
(because CSE emits moves itself).
So emit a virtual instruction that we can CSE.
Sometimes we demote a uniform to to a pull constant after emitting an
expansion move for it. In that case, recognize in opt_algebraic that if
the .file of the new instruction is GRF then it's just a real move that
we can copy propagate and such.
total instructions in shared programs: 5822418 -> 5812335 (-0.17%)
instructions in affected programs: 351841 -> 341758 (-2.87%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Cuts an instruction from two shaders in Tesseract, by allowing the
(x+y) cmp 0 -> x cmp -y optimization to take place.
instructions in affected programs: 1198 -> 1194 (-0.33%)
Reviewed-by: Eric Anholt <eric@anholt.net>
Nowadays GCC assumes stack pointer is 16-byte aligned even on 32-bits,
but that is an assumption OpenGL drivers (or any dynamic library for
that matter) can't afford to make as there are many closed- and open-
source application binaries out there that only assume 4-byte stack
alignment.
This fix uses force_align_arg_pointer GCC attribute, and is only a
stop-gap measure.
The right fix would be to pass -mstackrealign or
-mincoming-stack-boundary=2 to all source fails that use any -msse*
option, as there is no way to guarantee if/when GCC will decide to spill
SSE registers to the stack.
https://bugs.freedesktop.org/show_bug.cgi?id=86788
Reviewed-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4 10.3" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.3 10.4" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
The index_bias (aka base_vertex) applies to the downstream draw just as
much, since the actual index values are never modified.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.3 10.4" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
BRW_NEW_VERTICES is flagged every time we draw a primitive. Having
the brw_vs_prog atom depend on BRW_NEW_VERTICES meant that we had to
compute the VS program key and do a program cache lookup for every
single primitive. This is painfully expensive.
The workaround bit computation is almost entirely based on the vertex
attribute arrays (brw->vb.inputs[i]), which are set by brw_merge_inputs.
The only thing it uses the VS program for is to see which VS inputs are
actually read. brw_merge_inputs() happens once per primitive, and can
safely look at the currently bound vertex program, as it doesn't change
in the middle of a draw.
This patch moves the workaround bit computation to brw_merge_inputs(),
right after assigning brw->vb.inputs[i], and stores the previous WA bit
values in the context. If they've actually changed from the last draw
(which is uncommon), we signal that we need a new vertex program,
causing brw_vs_prog to compute a new key.
Improves performance in Gl32Batch7 by 13.6123% +/- 0.739652% (n=166)
on Haswell GT3e. I'm told Baytrail shows similar gains.
v2: Introduce a new BRW_NEW_VS_ATTRIB_WORKAROUNDS dirty bit, rather
than reusing BRW_NEW_VERTEX_PROGRAM (suggested by Chris Forbes).
This prevents unnecessary re-emission of surface/sampler related
atoms (and an SOL atom on Sandybridge).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
If you hit this, you didn't compile with --with-egl-platforms=...
Recompile with something like --with-egl-platforms=x11,drm and make
clean and make again.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
These stopped being necessary in commit ab973403e4.
v2: Update commit message with a better explanation (thanks to Eric
Anholt for doing the git archaeology).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We don't access brw->vertex_program or ctx->_Shader since the previous
commit, so we don't need this dirty bit.
I think it's still necessary on Gen6 because it still conflates
constant uploading with unit state uploading. We can fix that later.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We use IEEE mode for GLSL programs, but need to use ALT mode for ARB
programs so that 0^0 == 1. The choice is based entirely on the shader
source language.
Previously, our code to determine which mode we wanted was duplicated
in 8 different places (VS and FS for Gen4-5, Gen6, Gen7, and Gen8).
The ctx->_Shader->CurrentProgram[stage] == NULL check was confusing
as well - we use CurrentProgram (non-derived state), but _Shader
(derived state). It also relies on knowing that ARB programs don't
use gl_shader_program structures today. The compiler already makes
this assumption in a few places, but I'd rather keep that assumption
out of the state upload code.
With this patch, we select the mode at compile time, and store that
choice in prog_data. The state upload code simply uses that decision.
This eliminates a BRW_NEW_*_PROGRAM dependency in the state upload code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Commit c0347705 changed the Gen6-7 code to use ctx->_Shader rather than
ctx->Shader, but neglected to change the Gen4-5 or Gen8+ code.
This might fix SSO related bugs, but ALT mode is only used for ARB
programs, so if there's an actual problem, it's likely no one would
run into it.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The "Pixel Shader Computed Depth Mode" value is entirely based on the
shader program, so we can easily do it at compile time. This avoids the
if+switch on every 3DSTATE_WM (Gen7)/3DSTATE_PS_EXTRA (Gen8+) upload,
and shares a bit more code.
This also simplifies the PMA stall code, making it match the formula
more closely, and drops a BRW_NEW_FRAGMENT_PROGRAM dependency. (Note
that the previous comment was wrong - the code and the documentation
have != PSCDEPTH_OFF, not ==.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We shouldn't receive variables with invalid locations set - adding these
assertions should help catch problems before they cause crashes later.
Inspired by similar code in st_glsl_to_tgsi.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Half gives you the second half of a SIMD16 register, but if the register
is a uniform it would incorrectly give you the next register.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Previously only geometry shader outputs would be assigned locations if
the geometry shader was the only stage in the linked program.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: pavol@klacansky.com
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=82585
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
producer_var could be NULL if consumer_var is not NULL and
consumer_is_fs is false. This will occur when the producer is NULL and
the consumer is the geometry shader for a program that contains only a
geometry shader. This will occur starting with the next patch.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: pavol@klacansky.com
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=82585
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Nine code to match vertex declaration to vs inputs was limiting
the number of possible combinations.
Some sm3 games have issues with that, because arbitrary (usage/index)
can be used.
This patch does the following changes to fix the problem:
. Change the numbers given to (usage/index) combinations to uint16
. Do not put limits on the indices when it doesn't make sense
. change the conversion rule (usage/index) -> number to fit all combinations
. Instead of having a table usage_map mapping a (usage/index) number to
an input index, usage_map maps input indices to their (usage/index)
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
Tested-by: Yaroslav Andrusyak <pontostroy@gmail.com>
Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
With sm3, you can declare an input/output with an usage and an usage index.
Nine code hardcodes the translation usage/index to a corresponding TGSI code.
The translation was limited to a few usage/index combinations that were corresponding
to most of the needs of games, but some games did not work.
This patch rewrites that Nine code to map all possible usage/index combination
to TGSI code. The index associated to TGSI_SEMANTIC_GENERIC doesn't need to be low
for good performance, as the old code was supposing, and is not particularly bounded
(it's UINT16). Given the index is BYTE, we can map all combinations.
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
Tested-by: Yaroslav Andrusyak <pontostroy@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
This is the behaviour that Wine tests.
Reviewed-by: David Heidelberg <david@ixit.cz>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
This is the behaviour that Wine tests
Tested-by: David Heidelberg <david@ixit.cz>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
Nine was allowing that behaviour, but was not filling the result.
Tested-by: David Heidelberg <david@ixit.cz>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
Issuing D3DISSUE_END should:
. reset previous queries if possible
. end the query
Previous behaviour wasn't calling end_query for
queries not needing D3DISSUE_BEGIN, nor resetting
previous queries.
This fixes several applications not launching properly.
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
Tested-by: David Heidelberg <david@ixit.cz>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
It is the same behaviour as wine has.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: David Heidelberg <david@ixit.cz>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
Some queries need the driver to advertise a cap to be supported.
For example r300 doesn't support them.
v2 (David): check also for PIPE_CAP_QUERY_PIPELINE_STATISTICS, fix wine
tests on r300g
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: David Heidelberg <david@ixit.cz>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
get_query_result flushes automatically, we don't need to flush.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Axel Davy <axel.davy@ens.fr>
Applications are supposed to call CreateQuery with a NULL
ppQuery to know if the query is supported. We supported that.
However when ppQuery was not NULL, we were accepting to create the
query and were creating a dummy query even when the query is not
supported.
Wine has different behaviour. This patch drops the dummy queries
support and matches wine behaviour.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: David Heidelberg <david@ixit.cz>
Signed-off-by: Axel Davy <axel.davy@ens.fr>