Transform
sqrt a, b
rcp c, a
into
sqrt a, b
rsq c, b
The improvement here is that we've broken a dependency between these
instructions. Leads to 330 fewer INV instructions and 330 more RSQ.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Transform
sqrt a, b
rcp c, a
into
sqrt a, b
rsq c, b
In most cases the sqrt's result is still used, so the improvement here
is that we've broken a dependency between these instructions. Leads to
80 fewer INV instructions and 80 more RSQ.
Occasionally the sqrt's result is no longer used, leading to:
instructions in affected programs: 5005 -> 4949 (-1.12%)
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The next patch adds an algebraic optimization for the pattern
sqrt a, b
rcp c, a
and turns it into
sqrt a, b
rsq c, b
but many vertex shaders do
a = sqrt(b);
var1 /= a;
var2 /= a;
which generates
sqrt a, b
rcp c, a
rcp d, a
If we apply the algebraic optimization before CSE, we'll end up with
sqrt a, b
rsq c, b
rcp d, a
Applying CSE combines the RCP instructions, preventing this from
happening.
No shader-db changes.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Helps a handful of programs in Serious Sam 3 that use do-while loops.
instructions in affected programs: 16114 -> 16075 (-0.24%)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Do not rely on LLVMMCJITMemoryManagerRef being available.
The c binding to the memory manager objects only appeared
on llvm-3.4.
The change is based on an initial patch of Brian Paul.
Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
Blitter can still have transfers hanging around which it frees in
util_blitter_destroy(). So let it clean up before we yank the
transfer_pool from under it.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Indirect registers consume an additional token. Try to clean up the
token calculation math a bit, and fix it at the same time.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
If the name is just going to get dropped, don't bother making it. If
the name is made, release it sooner (rather than later).
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
If the name is just going to get dropped, don't bother making it. If
the name is made, release it sooner (rather than later).
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 74 40,578,719,715 67,762,208 62,263,404 5,498,804 0
After (32-bit): 52 40,565,579,466 66,359,800 61,187,818 5,171,982 0
Before (64-bit): 74 37,129,541,061 95,195,160 87,369,671 7,825,489 0
After (64-bit): 76 37,134,691,404 93,271,352 85,900,223 7,371,129 0
A real savings of 1.0MiB on 32-bit and 1.4MiB on 64-bit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
These few places were using ir_var_auto for seemingly no reason. The
names were not added to the symbol table.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
No change Valgrind massif results for a trimmed apitrace of dota2.
v2: Minor rebase on _mesa_init_constants changes.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Later patches will give every ir_var_temporary the same name in release
builds. Adding a bunch of variables named "compiler_temp" to the symbol
table can only cause problems.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Specifically, ir_var_temporary variables constructed with a NULL name
will all have the name "compiler_temp" in static storage.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 44 40,577,049,140 68,118,608 62,441,063 5,677,545 0
After (32-bit): 71 40,583,408,411 67,761,528 62,263,519 5,498,009 0
Before (64-bit): 63 37,122,829,194 95,153,008 87,333,600 7,819,408 0
After (64-bit): 67 37,123,303,706 95,150,544 87,333,600 7,816,944 0
A real savings of 173KiB on 32-bit and no change on 64-bit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
At least one of these pointers must be NULL, and we can determine which
will be NULL by looking at other fields. Use this information to store
both pointers in the same location.
If anyone can think of a better name for the union than "u", I'm all
ears.
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 63 40,574,239,515 68,117,280 62,618,607 5,498,673 0
After (32-bit): 44 40,577,049,140 68,118,608 62,441,063 5,677,545 0
Before (64-bit): 53 37,126,451,468 95,150,256 87,711,304 7,438,952 0
After (64-bit): 63 37,122,829,194 95,153,008 87,333,600 7,819,408 0
A real savings of 173KiB on 32-bit and 368KiB on 64-bit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Also move num_state_slots inside ir_variable_data for better packing.
The payoff for this will come in a few more patches.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
The payoff for this will come in a few more patches.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
warn_extension_index was moved to improve packing.
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 73 40,580,476,304 68,488,400 62,796,151 5,692,249 0
After (32-bit): 73 40,575,751,558 68,116,528 62,618,607 5,497,921 0
Before (64-bit): 71 37,124,890,613 95,889,584 88,089,008 7,800,576 0
After (64-bit): 62 37,123,578,526 95,150,784 87,711,304 7,439,480 0
A real savings of 173KiB on 32-bit and 368KiB on 64-bit.
v2: Use the enum name with the bit-field and remove the extra casts.
Suggested by Ken.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Tapani Pälli <tapani.palli@intel.com> [v1]
Also move the new warn_extension_index into ir_variable::data. This
enables slightly better packing.
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 82 40,580,040,531 68,488,992 62,973,695 5,515,297 0
After (32-bit): 73 40,580,476,304 68,488,400 62,796,151 5,692,249 0
Before (64-bit): 65 37,124,013,542 95,892,768 88,466,712 7,426,056 0
After (64-bit): 71 37,124,890,613 95,889,584 88,089,008 7,800,576 0
A real savings of 173KiB on 32-bit and 368KiB on 64-bit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
The payoff for this will come in the next patch.
No change Valgrind massif results for a trimmed apitrace of dota2.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
After compilation (and before linking) we can eliminate quite a few
built-in variables. Basically, any uniform or constant (e.g.,
gl_MaxVertexTextureImageUnits) that isn't used (with one exception) can
be eliminated. System values, vertex shader inputs (with one
exception), and fragment shader outputs that are not used and not
re-declared in the shader text can also be removed.
gl_ModelViewProjectMatrix and gl_Vertex are used by the built-in
function ftransform. There are some complications with eliminating
these variables (see the comment in the patch), so they are not
eliminated.
Valgrind massif results for a trimmed apitrace of dota2:
n time(i) total(B) useful-heap(B) extra-heap(B) stacks(B)
Before (32-bit): 46 40,661,487,174 75,116,800 68,854,065 6,262,735 0
After (32-bit): 50 40,564,927,443 69,185,408 63,683,871 5,501,537 0
Before (64-bit): 64 37,200,329,700 104,872,672 96,514,546 8,358,126 0
After (64-bit): 59 36,822,048,449 96,526,888 89,113,000 7,413,888 0
A real savings of 4.9MiB on 32-bit and 7.0MiB on 64-bit.
v2: Don't remove any built-in with Transpose in the name.
v3: Fix comment typo noticed by Anuj.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Anuj Phogat <anuj.phogat@gmail.com>
Cc: Eric Anholt <eric@anholt.net>
All built-in uniforms are supposed to be backed by some GL state. The
state_slots field describes this backing state.
This helped me track down a bug in a later patch.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Anuj Phogat <anuj.phogat@gmail.com>
Reuse the LLVMContext already allocated in llvmpipe_context
for draw_llvm if ppossible. This should decrease the memory
footprint of an llvmpipe context.
v2: Fix compile with llvm disabled.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
This fixes the remaining problem with the recently introduced
global jit memory manager. This change again uses a memory manager
that is local to gallivm_state. This implementation still frees
the majority of the memory immediately after compilation.
Only the generated code is deferred until this code is no longer used.
This change and the previous one using private LLVMContext instances
I can now safely run several independent OpenGL contexts driven
by llvmpipe from different threads.
v3: Rebase on llvm-3.6 compile fixes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
This is one step to make llvmpipe thread safe as mandated by the OpenGL
standard. Using the global LLVMContext is obviously a problem for
that kind of use pattern. The patch introduces two LLVMContext
instances that are private to an OpenGL context and used for all
compiles. One is put into struct draw_llvm and the other
one into struct llvmpipe_context.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
The big pile of patches I just pushed regresses about 25 piglit tests on
SNB. This fixes the regressions.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
The screen argument isn't actually used by lp_jit_screen_init() at this
time, but let's move the call so that we pass a valid pointer.
v2: don't leak screen if lp_jit_screen_init() fails.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
For cube resources, the array_size value should be 6. So handle
that case as we do for array texture resources. But assert that
array_size==6 just to be safe.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
layers (aka array_size) should be 6 for cube textures so we don't need
to special-case it. But add an assertion just to be safe.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Earlier in the function we assert layers==6 for PIPE_TEXTURE_CUBE so
there's no reason to special-case the pt.array_size = layers assignment.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
The core sw primitive restart code is still around, because i965 uses it
in some cases, but there are no drivers that want it on all the time.
Reviewed-by: Rob Clark <robdclark@gmail.com>
The drivers not flagging primitive restart support are r300 swtcl, svga,
nv30, and vc4.
The point of primitive restart is to slightly reduce draw call overhead
for apps by batching multiple draws. If we do an extra pass to read the
index buffer and split back into multiple draws, we've entirely missed the
point. This is particularly bad for drivers that otherwise have hardware
IB reads, where the readback is probably uncached.
Reviewed-by: Rob Clark <robdclark@gmail.com>
On gen 7, the MRF was removed and we gained the ability to do send
instructions directly from the GRF. This commit enables that
functinoality for FB writes.
v2: Make handling of components more sane.
i965/fs: Force a high register for the final FB write
v2: Renamed the array for the range mappings and added a comment
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
If we are going to use LOAD_PAYLOAD operations to fill MRF registers, then
we will need this.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, we were use the base_mrf parameter of fs_inst to store the MRF
location. In preparation for doing FB writes from the GRF, we now also
allow you to set inst->base_mrf to -1 and provide a source register.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt Turner
Mathias Fröhlich
Rob Clark
Ian Romanick
Eric Anholt
Jason Ekstrand
Brian Paul