The hardware only supports 4 MRTs. It should be possible to emulate
support for 8, but doesn't seem worth the trouble.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This complication is unnecessary and makes MRTs more complicated and
likely to generate tons of variants.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
The thing we want to avoid is int/float comparisons, but int/unsigned
comparisons with 0 are equivalent.
total instructions in shared programs: 6194829 -> 6193996 (-0.01%)
instructions in affected programs: 117192 -> 116359 (-0.71%)
helped: 471
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
No shader-db changes, probably because they're all removed by the GLSL
compiler optimization added in commit 69ad5fd4.
Reviewed-by: Eric Anholt <eric@anholt.net>
Doesn't work for analogous && cases, because of NaNs.
total instructions in shared programs: 6195712 -> 6194829 (-0.01%)
instructions in affected programs: 42000 -> 41117 (-2.10%)
helped: 403
Reviewed-by: Eric Anholt <eric@anholt.net>
InputsRead is a 64-bit bitfield. Using _mesa_fls would silently
truncate off the high bits, claiming inputs 32..56 (VARYING_SLOT_MAX)
were never read.
Using <= here was a hack I threw in at the last minute to fix programs
which happened to use input slot 32. Switch back to using < now that
the underlying problem is fixed.
Fixes crashes in "Euro Truck Simulator 2" when using prog->nir, which
uses input slot 33.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
ptn_move_dest and nir_fadd already take care of replicating the last
channel out, so we can just use a scalar and skip splatting it.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
We run lowering and optimization passes that might leave garbage lying
around. This keeps the FS cse from having to clean it up.
Reviewed-by: Matt Turner <mattst88@gmail.com>
The idea here is that fusing multiply-add combinations too early can reduce
our ability to perform CSE and value-numbering. Instead, we split ffma
opcodes up-front, hope CSE cleans up, and then fuse after-the-fact.
Unless an algebraic pass does something silly where it inserts something
between the multiply and the add, splitting and re-fusing should never
cause a problem. We run the late algebraic optimizations after this so
that things like compare-with-zero don't hurt our ability to fuse things.
shader-db results for fragment shaders on Haswell:
total instructions in shared programs: 4390538 -> 4379236 (-0.26%)
instructions in affected programs: 989359 -> 978057 (-1.14%)
helped: 5308
HURT: 97
GAINED: 78
LOST: 5
This does, unfortunately, cause some substantial hurt to a shader in Kerbal
Space Program. However, the damage is caused by changing a single
instruction from a ffma to an add. This, in turn, *decreases* register
pressure in one part of the program causing it to fail to register allocate
and spill. Given the overwhelmingly positive results in other shaders and
the fact that the NIR for the Kerbal shaders is actually better, this
should be considered a positive.
Reviewed-by: Matt Turner <mattst88@gmail.com>
total instructions in shared programs: 4422307 -> 4422363 (0.00%)
instructions in affected programs: 4230 -> 4286 (1.32%)
helped: 0
HURT: 12
While this does hurt some things, the losses are minor and it prevents the
compare-with-zero optimization from fighting with ffma which is much more
important.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, we couldn't generate two algebraic passes in the same file
because of multiple structure definitions. To solve this, we play the
age-old header file trick and just #define around it.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, NIR would just print 4 swizzle components if the swizzle was
anything other than foo.xyzw. This creates lots of noise if, for example,
you have a one-component element with a swizzle of foo.xxxx.
Reviewed-by: Kenneth Grunke <kenneth@whitecape.org>
Unused as of commit 630ab0d27ba(mesa: remove last of MAX_WIDTH,
MAX_HEIGHT). Update all the remaining references to the defines.
v2: Use the correct variable name in the comments
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
TGSI's conditional discards take float arg and negate it, so GLSL to TGSI
generates a b2f and negates that value. Only, in NIR we want a proper
bool once again, so we compare with 0. This is a lot of pointless extra
instructions.
total instructions in shared programs: 39735 -> 39702 (-0.08%)
instructions in affected programs: 1342 -> 1309 (-2.46%)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Since we have patterns based on b2f, generate them if we see the b2f
equivalent using an iand. This is common when generating NIR from TGSI.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
I was previously using temporary disables of VC4 optimization to show the
benefits of improved NIR optimization, but this can get me quick and dirty
numbers for NIR-only improvements without having to add hacks to disable
VC4's code (disabling of which might hide ways that the NIR changes would
hurt actual VC4 codegen).
NIR brings us better optimization than I would have bothered to write
within the driver, developers sharing future optimization work, and the
ability to share device-specific lowering code that we and other
GLES2-level drivers need.
total uniforms in shared programs: 13421 -> 13422 (0.01%)
uniforms in affected programs: 62 -> 63 (1.61%)
total instructions in shared programs: 39961 -> 39707 (-0.64%)
instructions in affected programs: 15494 -> 15240 (-1.64%)
v2: Add missing imov support, and assert that there are no dest saturates.
v3: Rebase on the target-specific algebraic series.
v4: Rebase on gallium-includes-from-NIR changes in mater.
v5: Rebase on variables being in lists instead of hash tables.
v6: Squash in intermediate changes that used the NIR-to-TGSI pass (which
I'm not committing)
This will be used by the VC4 driver for doing device-independent
optimization, and hopefully eventually replacing its whole IR. It also
may be useful to other drivers for the same reason.
v2: Add all of the instructions I was relying on tgsi_lowering to remove,
and more.
v3: Rebase on SSA rework of the builder.
v4: Use the NIR ineg operation instead of doing a src modifier.
v5: Don't use ineg for fnegs. (infer_src_type on MOV doesn't do what I
expect, again).
v6: Fix handling of multi-channel KILL_IF sources.
v7: Make ttn_get_f() return a swizzle of a scalar load_const, rather than
a vector load_const. CSE doesn't recognize that srcs out of those
channels are actually all the same.
v8: Rebase on nir_builder auto-sizing, make the scalar arguments to
non-ALU instructions actually be scalars.
v9: Add support for if/loop instructions, additional texture targets, and
untested support for indirect addressing on temps.
v10: Rebase on master, drop bad comment about control flow and just choose
the X channel, use int comparison opcodes in LIT for now, drop unused
pipe_context argument..
v11: Fix translation of LRP (previously missed because I mis-translated
back out), use nir_builder init helpers.
v12: Rebase on master, adding explicit include of mtypes.h to get
INTERP_QUALIFIER_*
v13: Rebase on variables being in lists instead of hash tables, drop use
of mtypes.h in favor of util/pipeline.h. Use Ken's nir_builder
swizzle and fmov/imov_alu helpers, drop "struct" in front of
nir_builder, use nir_builder directly as the function arg in a lot of
cases, drop redundant members of ttn_compile that are also in
nir_builder, drop some half-baked malloc failure handling.
v14: The indirect uniform src0 should be scalar, not vector (noticed as
odd by robclark, confirmed by cwabbott). Apply Ken's review to
initialize s->num_uniforms and friends, skip ttn_channel for dot
products, and use the simpler discard_if intrinsic.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v13)
Acked-by: Rob Clark <robclark@freedesktop.org>
NIR uses these enums/#defines in nir_variables and associated intrinsics,
but I want to be able to use them from TGSI->NIR and NIR->TGSI.
Otherwise, we had to pull in all of mtypes.h.
This doesn't cover all of the enums we might want from a shared compiler
core (like varying slots or vert attribs), but it at least covers what I
need at the moment (system values and interp qualifiers).
v2: Move to src/glsl since util/ is really vague. Include in Makefile.am
list. Use plain bitshifts and stdint types instead of undefined
BITFIELD64_BIT.
v3: Rename to shader_enums.h. Move it into Makefile.sources.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v2, with
recommendation to rename)
The header was added with commit 2a135c470e3(nir: Add an ALU op builder
kind of like ir_builder.h) but did not made it into to the sources list.
Fortunately it remained unused until a recent commit faf6106c6f6(nir:
Implement a Mesa IR -> NIR translator.)
v2: Remove the bogus dependency. Tweak commit message.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
