If the pixel pipes have a different number of subslices, emit a slice
hashing table that will ensure proper workload distribution.
v2: Don't need to set the mask - it's mbo (Ken).
v3: Don't keep a reference to the resource used for emitting the table
(Ken).
Add these fields and the 3DSTATE_SLICE_TABLE_STATE_POINTERS instruction
so we can properly configure the slice and subslice hashing on ICL+
v2: Make 'Mask' field a mbo (Ken).
We don't need it for state setup but it's a useful statistic we want to
pass on to developers.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This commit is all annoying plumbing work which just adds support for a
new brw_compile_stats struct. This struct provides a binary driver
readable form of the same statistics we dump out to stderr when we
INTEL_DEBUG is set with a shader stage.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
While NIR's lower_imul64() solves the case of 64 bit integer multiplications
generated early, we don't have a way to lower such instructions when they are
generated by our own backend, such as the scan/reduce intrinsics. We'll need
this soon, so implement it now.
An easy way to test this is to simply disable nir_lower_imul64 to let
those operations reach the backend.
v2:
- Fix Q/UQ copy/paste errors (Caio).
- Transform an 'if' into 'else if' (Caio).
- Add an extra comment to clarify the need for 64b = 32b * 32b
(Caio).
- Make private functions private (Caio).
v3:
- Remove ambiguity with 'b' and 'd' variables (Caio).
- Allocate potentially less regs for the dwords (Caio).
Cc: Jason Ekstrand <jason.ekstrand@intel.com>
Cc: Matt Turner <matt.turner@intel.com>
Cc: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Invert the logic of how progress is handled: remove the continue
statements and mark progress inside the places where it actually
happens.
We're going to add a new lowering that also looks for BRW_OPCODE_MUL,
so inverting the logic here makes the resulting code much easier to
follow.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Don't instantiate a builder for each instruction during
lower_integer_multiplication(). Instantiate one only when needed.
On the other hand, these unneeded builders don't seem to cost much to
init, so I don't expect any significant difference in performance:
this is mostly about code organization.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
The lower_integer_multiplication() function is already a little too
big. I want to add more to it, so let's reorganize the existing code
first. Let's start with just extracting the current code to
subfunctions. Later we'll change them a little more.
v2: Make private functions private (Caio).
v3: Fix typo (Caio).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
v2: add to series
v3: update Makefile.sources
v4: don't remove a comment and break statement
v4: use nir_can_move_instr
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This is mostly the same as nir_move_load_const() but can also move
undef instructions, comparisons and some intrinsics (being careful with
loops).
v2: actually delete nir_move_load_const.c
v3: fix nir_opt_sink() usage in freedreno
v3: update Makefile.sources
v4: replace get_move_def with nir_can_move_instr and nir_instr_ssa_def
v4: handle if uses
v4: fix handling of nested loops
v5: re-write adjust_block_for_loops
v5: re-write setting of use_block for if uses
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Co-authored-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Eric Anholt <eric@anholt.net>
See "i965/gen9: Optimize slice and subslice load balancing behavior."
for the rationale. According to Jason, improves Aztec Ruins
performance by 2.7%.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
v2: Undo CPU performance micro-optimization done in i965 and iris due
to lack of data justifying it on anv. Use
cmd_buffer_apply_pipe_flushes wrapper instead of emitting pipe
control command directly. (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Lower fddx and fddy and set the right bits in codegen.
Signed-off-by: Andreas Baierl <ichgeh@imkreisrum.de>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Reviewed-by: Erico Nunes <nunes.erico@gmail.com>
This allows enabling the shader info keeping on a per shader basis.
Also disables the cache on a per shader basis.
Reviewed-by: Dave Airlie <airlied@redhat.com>
The default pixel hashing mode settings used for slice and subslice
load balancing are far from optimal under certain conditions (see the
comments below for the gory details). The top-of-the-line GT4 parts
suffer from a particularly severe performance problem currently due to
a subslice load balancing issue. Fixing this seems to improve
graphics performance across the board for most of the benchmarks in my
test set, up to ~20% in some cases, e.g. from SKL GT4:
unigine/valley: 3.44% ±0.11%
gfxbench/gl_manhattan31: 3.99% ±0.13%
gputest/pixmark_piano: 7.95% ±0.33%
synmark/OglTexFilterAniso: 15.22% ±0.07%
synmark/OglTexMem128: 22.26% ±0.06%
Lower-end platforms are also affected by some subslice load imbalance
to a lesser degree, especially during CCS resolve and fast clear
operations, which are handled specially here due to rasterization
ocurring in reduced CCS coordinates, which changes the semantics of
the pixel hashing mode settings.
No regressions seen during my tests on some SKL, KBL and BXT
configurations. Additional benchmark reports welcome on any Gen9
platforms (that includes anything with Skylake, Broxton, Kabylake,
Geminilake, Coffeelake, Whiskey Lake, Comet Lake or Amber Lake in your
renderer string).
P.S.: A similar problem is likely to be present on other non-Gen9
platforms, especially for CCS resolve and fast clear operations.
Will follow-up with additional patches fixing the hashing mode
for those once I have enough performance data to justify it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This is a bit of a hack, but it'll hold us over until we have 64-bit
support wired through.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This used a delicate hack to try to find indirect inputs and skip them
as candidates for pairing. Let's use a better criterion -- no sources --
and pair based on that.
We could do better, but that would require more complex data flow
analysis than we're interested in doing here.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We implement gl_WorkGroupID and gl_LocalInvocationID, which map to
ld_compute_id with special sources.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This allows liveness analysis within a loop to be more fine grained,
fixing RA failures with partial spilled movs within a loop, as well as
enabling a slight reduction of register pressure more generally:
total registers in shared programs: 350 -> 347 (-0.86%)
registers in affected programs: 12 -> 9 (-25.00%)
helped: 3
HURT: 0
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 25.00% max: 25.00% x̄: 25.00% x̃: 25.00%
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
