When input to intrinsic is uniform value, reduced value is
same as input whereas if input value is divergent we need
to iterate over all active lanes of WaveFront to perform
the reduction.
The control flow for a `loop` has been set up, which
iterates over `only` active lanes to perform reduction.
Introduced WAVE_REDUCE_UMIN_PSEUDO_U32 and
WAVE_REDUCE_UMAX_PSEUDO_U32 Pseudos which
are lowered Post-ISel (in `EmitInstrWithCustomInserter `).
Reviewed By: arsenm, #amdgpu
Differential Revision: https://reviews.llvm.org/D154858
Extends the new frexp scaled reciprocal to the general case. The
reciprocal case is just the same thing when frexp of 1 is constant
folded. Could probably clean up the code to rely on that constant
folding.
Improves results for the IEEE path for the default OpenCL division. We
used to only emit the fdiv.fast intrinsic with a 2.5 ulp accuracy
threshold with DAZ, which uses explicit range checks. This gives us a
better fast option with the default IEEE behavior.
NFC-ish. Does trigger some reordering of the fdiv scalarization. Also
skips scalarizing in more cases where nothing was going to happen. We
can still scalarize in some no-op edge cases.
https://reviews.llvm.org/D155740
The highlight change is a new denormal safe 1ulp lowering which uses
rcp after using frexp to perform input scaling. This saves 2
instructions compared to other implementations which performed an
explicit denormal range change. This improves the OpenCL default, and
requires a flag for HIP. I don't believe there's any flag wired up for
OpenMP to emit the necessary fpmath metadata.
This provides several improvements and changes that were hard to
separate without regressing one case or another. Disturbingly the
OpenCL conformance test seems to have the reciprocal test commented
out. I locally hacked it back in to test this.
Starts introducing f32 rsq intrinsics in AMDGPUCodeGenPrepare. Like
the rcp case, we could do this in codegen if !fpmath were preserved
(although we would lose some computeKnownFPClass tricks). Start
requiring contract flags to form rsq. The rsq fusion actually improves
the result from ~2ulp to ~1ulp. We have some older fusion in codegen
which only keys off unsafe math which should be refined.
Expand rsq patterns by checking for denormal inputs and pre/post
multiplying like the current library code does. We also take advantage
of computeKnownFPClass to avoid the scaling when we can statically
prove the input cannot be a denormal. We could do the same for the rcp
case, but unlike rsq a large input can underflow to denormal. We need
additional upper bound exponent checks on the input in order to do the
same for rcp.
This rsq handling also now starts handling the negated case. We
introduce rsq with an fneg. In the case the fneg doesn't fold into its
user, it's a neutral change but provides improvement if it is foldable
as a source modifier.
Also starts respecting the arcp attribute properly, and more strictly
interprets afn. We were previously interpreting afn as implying you
could do the reciprocal expansion of an fdiv. The codegen handling of
these also needs to be revisited.
This also effectively introduces the optimization
combineRepeatedFPDivisors enables, just done in the IR instead (and
only for f32).
This is almost across the board better. The one minor regression is
for gfx6/buggy frexp case where for multiple reciprocals, we could
previously reuse rematerialized constants per instance (it's neutral
for a single rcp).
The fdiv.fast and sqrt handling need to be revisited next.
https://reviews.llvm.org/D155593
We kind of have to have multiple implementations of fdiv split between
the two selectors with some pre-processing. Add yet another test to
check for consistency of interpretation of flag combinations. We have
quite a bit of test redundancy here already, but there are so many
possible interesting permutations it's unwieldy to cover every detail
in any one of them. We have a number of overlapping fdiv tests but
it's hard to follow everything going on as it is.
This isn't always folded to fneg for a freestanding fsub depending on
the denormal mode. When matching source modifiers, we're implicitly
canonicalizing the input so we can fold it here.
Doesn't bother handling the VOP3P case since it's only relevant with
DAZ, which nobody really uses with f16.
For f64, tests show an existing bug where DAGCombiner tries to respect
the denormal mode for fsub -0, x, but not after it's lowered to fadd
-0, (fneg x). Either the fold is wrong or we shouldn't restrict the
fsub case based on the denormal mode.
https://reviews.llvm.org/D155652
We were dropping the flags and thus blocking contract into potential
fadd users. GlobalISel was already preserving the flags here.
https://reviews.llvm.org/D155443
This fixes sinking a VGPR def out of a loop past the reconvergence
point at the SI_END_CF. There was a prior fix which introduced
blockPrologueInterferes (D121277) to fix the same basic problem for
the post RA sink. This also had the special case isIgnorableUse case
which was incorrect, because in some contexts the exec use is not
ignorable.
I'm thinking about a new way to represent this which will avoid
needing hasIgnorableUse and isBasicBlockPrologue, which would function
more like the exception handling.
Fixes: SWDEV-407790
https://reviews.llvm.org/D155343
CMP(A,C)||CMP(B,C) => CMP(MIN/MAX(A,B), C)
CMP(A,C)&&CMP(B,C) => CMP(MIN/MAX(A,B), C)
This first patch handles integer types.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D153502
Inspired by some of the cases from D145468
Let SimplifyDemandedBits handle the narrowing of lshr to half-width if we don't require the upper bits, the narrowed shift is profitable and the zext/trunc are free.
A future patch will propose the equivalent shl narrowing combine.
Differential Revision: https://reviews.llvm.org/D146121
- Switch to generated checks
- Use a different run line per denormal mode to reduce test duplication
- Add test coverage for rsqrt cases
- Add test coverage for repeated arcp denominator
- Fix the optnone test
The previous heuristic rejected a PHI if one of its user was an unbreakable PHI, no matter what the other users were.
This worked well in most cases, but there's one case in rocRAND where
it doesn't work. In that case, a PHI node has 2 PHI users where one is
breakable but not the other. When that PHI node isn't broken performance falls by 35%.
Relaxing the restriction to "require that half of the PHI node users are breakable" fixes the issue, and seems like a sensible change.
Solves SWDEV-409648, SWDEV-398393
Reviewed By: #amdgpu, arsenm
Differential Revision: https://reviews.llvm.org/D155184
Do the LDS frame calculation once, in the IR pass, instead of repeating the work in the backend.
Prior to this patch:
The IR lowering pass sets up a per-kernel LDS frame and annotates the variables with absolute_symbol
metadata so that the assembler can build lookup tables out of it. There is a fragile association between
kernel functions and named structs which is used to recompute the frame layout in the backend, with
fatal_errors catching inconsistencies in the second calculation.
After this patch:
The IR lowering pass additionally sets a frame size attribute on kernels. The backend uses the same
absolute_symbol metadata that the assembler uses to place objects within that frame size.
Deleted the now dead allocation code from the backend. Left for a later cleanup:
- enabling lowering for anonymous functions
- removing the elide-module-lds attribute (test churn, it's not used by llc any more)
- adjusting the dynamic alignment check to not use symbol names
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D155190
This adds the IGLP strategy for single-wave gemms. The SchedGroup pipeline is laid out in multiple phases, with each phase corresponding to a distinct pattern present in gemm kernels. The resilience of the optimization is dependent upon IR (as seen by pre-RA scheduling) continuing to have these patterns (as defined by instruction class and dependencies) in their current relative ordering.
The kernels of interest have these specific phases:
NT: 1, 2a, 2c
NN: 1, 2a, 2b
TT: 1, 2b, 2c
TN: 1, 2b
The general approach taken was to have a long SchedGroup pipeline. In this way the scheduler will have less capability of doing the wrong thing. In order to resolve the challenge of correctly fitting these long pipelines, we leverage the rules infrastructure to help the solver.
Differential Revision: https://reviews.llvm.org/D149773
Change-Id: I1a35962a95b4bdf740602b8f110d3297c6fb9d96
This time without the extra `->dump()`
A recent addition to the device libs, `__ockl_dm_trim`, caused a series of
failures at O0 due to a i64 ballot intrinsic being inlined into a wave32 function.
The quick fix for this is to support codegen for this rare case.
A proper long-term fix for this type of issue is still being discussed.
Fixes SWDEV-408929, SWDEV-408957, SWDEV-409885, SWDEV-410193
Reviewed By: #amdgpu, arsenm
Differential Revision: https://reviews.llvm.org/D155050
