When the earlier block ended with control flow, we'd mistakenly remove
some of its links to its children. The same happened with the later
block.
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
A pattern in certain shaders is:
uniform vec4 colors[NUM_LIGHTS];
for (int i = 0; i < NUM_LIGHTS; i++) {
...use colors[i]...
}
In this case, the application author expects the shader compiler to
unroll the loop. By doing so, it replaces variable indexing of the
array with constant indexing, which is more efficient.
This patch extends the heuristic to see if arrays accessed within the
loop are indexed by an induction variable, and if the array size exactly
matches the number of loop iterations. If so, the application author
probably intended us to unroll it. If not, we rely on the existing
loop-too-large heuristic.
Improves performance in a phong shading microbenchmark by 2.88x, and a
shadow mapping microbenchmark by 1.63x. Without variable indexing, we
can upload the small uniform arrays as push constants instead of pull
constants, avoiding shader memory access. Affects several games, but
doesn't appear to impact their performance.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Kristian Høgsberg <krh@bitplanet.net>
Consider GLSL code such as:
const ivec2 offsets[] =
ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
ivec2(0, -1), ivec2(0, 0), ivec2(0, 1),
ivec2(1, -1), ivec2(1, 0), ivec2(1, 1));
ivec2 offset = offsets[<non-constant expression>];
Both i965 and nv50 currently handle this very poorly. On i965, this
becomes a pile of MOVs to load the immediate constants into registers,
a pile of scratch writes to move the whole array to memory, and one
scratch read to actually access the value - effectively the same as if
it were a non-constant array.
We'd much rather upload large blocks of constant data as uniform data,
so drivers can simply upload the data via constbufs, and not have to
populate it via shader instructions.
This is currently non-optional because both i965 and nouveau benefit
from it, and according to Marek radeonsi would benefit today as well.
(According to Tom, radeonsi may want to handle this itself in the long
term, but we can always add a flag when it becomes useful.)
Improves performance in a terrain rendering microbenchmark by about 2x,
and cuts the number of instructions in about half. Helps a lot of
"Natural Selection 2" shaders, as well as one "HOARD" shader.
total instructions in shared programs: 5473459 -> 5471765 (-0.03%)
instructions in affected programs: 5880 -> 4186 (-28.81%)
v2: Use ir_var_hidden to avoid exposing the new uniform via the GL
uniform introspection API.
v3: Alphabetize Makefile.sources properly.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77957
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
In the compiler, we'd like to generate implicit uniforms for internal
use. These should not be visible via the GL uniform introspection API.
To support that, we add a new ir_variable::how_declared value of
ir_var_hidden, and plumb that through to gl_uniform_storage.
v2 (idr): Fix some memory management issues in
move_hidden_uniforms_to_end. The comment block on the function has more
details.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Makes use of SSE 4.1 to speed up compute of min and max elements.
Callgrind cpu usage results from pts benchmarks:
Openarena 0.8.8: 3.67% -> 1.03%
UrbanTerror: 2.36% -> 0.81%
V5:
- actually make use of the optimisation in android (Emil Velikov)
- set a better array size limit for using SSE and added TODO
V4:
- fixed bugs with incrementing pointer and updating counters
V3:
- Removed sse_minmax.c from Makefile.sources
- handle the first few values without SSE until the pointer is aligned
and use _mm_load_si128 rather than _mm_loadu_si128
- guard the call to the SSE code better at build time
V2:
- removed GL* types
- use _mm_store_si128() rather than _mm_store_ps()
- add runtime check for SSE
- use aligned attribute for local mix/max
- bunch of tidyups
Reviewed-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Timothy Arceri <t_arceri@yahoo.com.au>
Walk through the list and free each config, and finally free the list
itself. Freeing approx 20KiB of memory, according to valgrind.
Inspired by a similar patch by enpeng xu.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
driUnbindContext() checks for valid drawables before calling the driver
unbind function. In case of Surfaceless contexts, the drawables are always
Null and we end up not releasing the underlying DRI context. Moving the
call to the driver function before the drawable validity checks fixes things.
Steps to trigger this bug are following:
- create surfaceless context and make it current
- make some other context current
- {another thread} destroy surfaceless context
- make another context current
Signed-off-by: Alexandros Frantzis <Alexandros.Frantzis@canonical.com>
Signed-off-by: Kalyan Kondapally <kalyan.kondapally@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=74563
The dummy shader sends an EOT message to end itself. There are many more
works need to be done on the compiler side before we can advertise
PIPE_CAP_COMPUTE.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Use util_dynarray in ilo_set_global_binding() to allow for unlimited number of
global bindings. Add a comment for global bindings.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
We need to know the local/input/private sizes and others. This is not
complete. We need many others for CURBE setup.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
They will be used to report compute params or program compute states.
thread_count can also be used for 3DSTATE_VS.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
This started hitting an assertion recently. Only affects Haswell
(Ivybridge doesn't support this meddling with the sampler state pointer,
and ARB_gpu_shader5 is not enabled yet on Broadwell)
14 Piglits crash->pass.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Improves performance in GLBenchmark 2.7 TRex by 3.88889% +/- 0.336383%
(n=80) at 1280x720 on Broadwell GT3. Together with the previous patch,
it improves performance by 5.42738% +/- 0.541971% (n=10) at 1920x1080.
Note that without the PMA stall fix, this would instead decrease
performance by 22%.
v2: Update comment (noticed by Kristian Høgsberg).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Certain non-promoted depth cases typically incur stalls. In very
specific cases, we can enable a workaround which improves performance.
Improves performance in GLBenchmark 2.7 TRex by 1.17762% +/- 0.448765%
(n=75) at 1280x720 on Broadwell GT3.
Haswell has this feature as well, but we can't currently write registers
from userspace batches (and we'd incur additional software batch
scanning overhead as well), so we haven't enabled it. Broadwell allows
us to write CACHE_MODE_1. Backporters beware: the formula and flushing
incantation differs between Haswell and Broadwell.
v2: Move pma_stall_bits from brw->state to brw itself (requested by
Kristian Høgsberg).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Matt requested this in review feedback on the original patch, which I
completely missed when pushing this series. Kristian also made this
change, but I grabbed the wrong version of the patch.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Otherwise, calling glPopAttrib on drivers that don't support
ARB_clip_control gives you a GL error, which is surprising at best.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
We're not programming the clear values yet, so this won't work.
This patch should be (effectively) reverted eventually.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
On Skylake, the MOCS bits are an index into a table of 63 different,
configurable cache configurations. As for previous GENs, we only care about
WB and WT, which are available in the documented default set. Define
SKL_MOCS_WB and SKL_MOCS_WT to the indices for those configucations and use
those for the Skylake MOCS values.
Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Skylake has separate controls for enabling the Z Clip Test for the near
and far planes. For now, maintain the legacy behavior by setting both.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Skylake's 3DSTATE_DS packet has a few more fields; we don't support
domain shaders yet though.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
They are the same as for BDW, so just add a case for SKL to the init switch.
Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
On SKL, 3DSTATE_CONSTANT_* command is not committed until we give
the corresponding 3DSTATE_BINDING_TABLE_POINTERS_* command. If we
fail to do so, the constant buffers wont be read and push constants
will be wrong.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Otherwise they overlap and horrible things happen. All the new DWords
are for fast color clear values, which we don't do yet.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
We will need to allocate more DWords on Skylake.
v2: Don't mark brw_context parameter const. It's modified.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Skylake introduces a new base address for a feature we don't yet expose.
Setting these to 0 should be safe.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Skylake uploads the stencil reference values in DW3 of the
3DSTATE_WM_DEPTH_STENCIL packet, rather than in COLOR_CALC_STATE.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
