If the only dirty state is mesa's _NEW_PROGRAM_CONSTANTS flag, we can
skip state validation before drawing a bitmap since that state doesn't
effect bitmap rendering.
This further increases the performance of the ipers demo on llvmpipe
to about what it was before commit 36c93a6fae.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
We were checking the dirty->st flags but not the dirty->mesa flags.
When we took the early return, we didn't clear the dirty->mesa flags
so the next time we called st_validate_state() we'd often flush the
glBitmap cache. And since st_validate_state() is called from
st_Bitmap(), it meant we flushed the bitmap cache for every glBitmap()
call.
This change seems to recover most of the performance loss observed
with the ipers demo on llvmpipe since commit commit 36c93a6fae.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Previously each member was being counted as using a single slot,
count_attribute_slots() fixes the count for array and struct members.
Also don't assign a negitive to the unsigned expl_location variable.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Previously we were only reserving a single location for arrays and
structs.
We also didn't take into account implicit locations clashing with
explicit locations when assigning locations for their arrays or
structs.
This patch fixes both issues.
V5: fix regression for patch inputs/outputs in tessellation shaders
V4: just use count_attribute_slots() to get the number of slots,
also calculate the correct number of slots to reserve for gs and
tess stages by making use of the new get_varying_type() helper.
V3: handle arrays of structs
V2: also fix for arrays of arrays and structs.
Acked-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
This will be used in the following patch for calculating array sizes correctly
when reserving explicit varying locations.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Previously we would pack varyings before trying to remove them, this
relied on the packing pass not packing varyings with a location of -1
to avoid packing varyings that should be removed.
However this meant unused varyings with an explicit location would be
packed before they could be removed when we enable packing of them in a
later patch.
V2: fix regression in V1 removing unused varyings in multi-stage SSO,
fix regression with single stage programs.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
ALIGN_DIVUP is a driver specific(r600g) macro that duplicates DIV_ROUND_UP functionality.
Replacing it with DIV_ROUND_UP eliminates this problems.
Signed-off-by: Krzysztof A. Sobiecki <sobkas@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
I had the driver all tested for the last series, and in my last build I
noticed that get_swizzled_channel was unused now, and removed
it... apparently without testing to find that I removed the wrong channel
swizzle function.
We routinely have code like:
vec1 ssa_220 = fge ssa_104, ssa_61
vec1 ssa_199 = bcsel ssa_220, ssa_106, ssa_105
and we would compare fge's args and choose between ~0 and 0 to generate
ssa_220, then compare ssa_220 to 0 and choose between bcsel's args.
Instead, try to notice the pattern and compare between fge's args to
select between bcsel's args.
total instructions in shared programs: 88019 -> 87574 (-0.51%)
instructions in affected programs: 9985 -> 9540 (-4.46%)
total estimated cycles in shared programs: 245752 -> 245237 (-0.21%)
estimated cycles in affected programs: 17232 -> 16717 (-2.99%)
We can't use its other features currently (mostly because we don't want
Newton-Raphson on rcps for texture coordinates), but it gets us started.
This eliminates some comparisons with constants in GLB2.7 and ETQW traces
at the QIR level by moving the comparisons into NIR, where they get
constant-folded out.
instructions in affected programs: 165 -> 156 (-5.45%)
total uniforms in shared programs: 32087 -> 32085 (-0.01%)
total estimated cycles in shared programs: 245762 -> 245752 (-0.00%)
estimated cycles in affected programs: 461 -> 451 (-2.17%)
I'm moving away from QIR being SSA (since NIR is doing lots of SSA
optimization for us now) and instead having QIR just be QPU operations
with virtual registers. By making our SELs be composed of two MOVs, we
could potentially coalesce the registers for the MOV's src and dst and
eliminate the MOV.
total instructions in shared programs: 88448 -> 88028 (-0.47%)
instructions in affected programs: 39845 -> 39425 (-1.05%)
total estimated cycles in shared programs: 246306 -> 245762 (-0.22%)
estimated cycles in affected programs: 162887 -> 162343 (-0.33%)
If you want the SF of the value of a register produced from a series of
packing MOVs or conditional MOVs, we can't just SF on the last MOV into
the register.
QPitch is usually expressed as rows of surface elements (where a surface
element is an compression block or a single surface sample. Skylake 1D
is an outlier; there QPitch is expressed as individual surface
elements.
For 1D surfaces and for surfaces with Yf or Ys tiling, the hardware
ignores SurfaceVerticalAlignment and SurfaceHorizontalAlignment.
Moreover, the anv_halign[] and anv_valign[] lookup tables may not even
contain the surface's actual alignment values. So don't do the lookup
for those surfaces.
Fix a 's/unsigned int/unsigned/' consistency case while here.
Found-by: Coccinelle
Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Fix silly issue with MSVC case fall-though support to need
a extra 'break;'
Found-by: Coccinelle
Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
This patch converts the SSE-optimized lp_rast_triangle_32_3_16()
to VMX/VSX.
I measured the results on POWER8 machine with 32 cores at 3.4GHz and
16GB of RAM.
FPS/Score
Name Before After Delta
------------------------------------------------
openarena 16.35 16.7 2.14%
xonotic 4.707 4.97 5.57%
glmark2 didn't show a significant (more than 1%) difference.
v2: Make sure code is build only on POWER8 LE machine
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This patch converts the SSE-optimized build_mask_32() and
build_mask_linear_32() to VMX/VSX.
I measured the results on POWER8 machine with 32 cores at 3.4GHz and
16GB of RAM.
FPS/Score
Name Before After Delta
------------------------------------------------
glmark2 (score) 139.8 142.7 2.07%
openarena and xonotic didn't show a significant (more than 1%)
difference.
v2: Make sure code is build only on POWER8 LE machine
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This patch converts the SSE optimization done in do_triangle_ccw to
VMX/VSX.
I measured the results on POWER8 machine with 32 cores at 3.4GHz and
16GB of RAM.
FPS/Score
Name Before After Delta
------------------------------------------------
glmark2 (score) 136.6 139.8 2.34%
openarena 16.14 16.35 1.30%
xonotic 4.655 4.707 1.11%
v2:
- Convert loads to use aligned loads
- Make sure code is build only on POWER8 LE machine
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This file provides a portability layer that will make it easier to convert
SSE-based functions to VMX/VSX-based functions.
All the functions implemented in this file are prefixed using "vec_".
Therefore, when converting from SSE-based function, one needs to simply
replace the "_mm_" prefix of the SSE function being called to "vec_".
Having said that, not all functions could be converted as such, due to the
differences between the architectures. So, when doing such
conversion hurt the performance, I preferred to implement a more ad-hoc
solution. For example, converting the _mm_shuffle_epi32 needed to be done
using ad-hoc masks instead of a generic function.
All the functions in this file support both little-endian and big-endian
but currently the file is build only on POWER8 LE machine.
All of the functions are implemented using the Altivec/VMX intrinsics,
except one where I needed to use inline assembly (due to missing
intrinsic).
v2:
- Use vec_vgbbd instead of __builtin_vec_vgbbd
- Add an aligned load function
- Don't use typeof()
- Make file build only on POWER8 LE machine
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
To determine if we could use special POWER8 assembly directives, we first
need to detect whether we are running on POWER8 architecture. This patch
adds this detection to configure.ac and adds the necessary compilation
flags accordingly.
v2:
- Add option to disable POWER8 instructions generation
- Detect whether building on BE or LE machine and build with
-mpower8-vector only on LE machine
- Make the printed messages more standard
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The nir_opt_algebraic rule
(('fadd', ('flog2', a), ('fneg', ('flog2', b))), ('flog2', ('fdiv', a, b))),
can produce new fdiv operations, which need to be lowered on i965,
as we don't actually implement fdiv. (Normally, we handle this in
GLSL IR's lower_instructions pass, but in the above case we introduce
an fdiv after that point. So, make NIR do it for us.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
Compute shaders require reconfiguring the L3 for shared local memory
support. We have to be able to write the L3 registers to do that.
This effectively turns off compute shaders prior to Kernel 4.2.
(Previously, the extension enable was in an API_OPENGL_CORE conditional.
However, that isn't necessary - core Mesa extension handling already
restricts it properly. I've moved it out in this patch.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This fixes some piglit subtests for ARB_program_interface_query.
V3: remove some of the unnecessary parentheses
V2: fix alignment
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
There is a function dedicated to demoting unused varyings lets
trust it to do its job.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
After lowering the matching flag is_unmatched_generic_inout is lost so
we need to move this validation before lowering.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
An SSO program can have multiple stages and we only want to add the externally
facing varyings. The current code was adding both the packed inputs and outputs
for the first and last stage of each program.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Instead of trying to crawl through predecessor chains and build phi nodes,
we just do a poor-man's out-of-ssa on the spot. The into-SSA pass will
deal with putting the actual phi nodes in for us.
Conditions modified allow skl+ to use blitter:
- for all tiling formats
- to write data to YF/YS tiled surfaces
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
