Extension nodes make schedule trees are less flexible: Many operations,
such as rescheduling, do not work on such schedule trees with extension.
As such, some functionality such as determining parallel loops in isl's
AST are disabled.
Currently, only the pattern-matching generalized matrix-matrix
multiplication optimization adds extension nodes (to add copy-in
statements).
This patch removes all extension nodes as the last step of the schedule
optimization by hoisting the extension node's added domain up to the
root domain node. All following passes can assume that schedule trees
work without restrictions, including the parallelism test. Mark the
outermost loop of the optimized matrix-matrix multiplication as parallel
such that -polly-parallel is able to parallelize that loop.
Differential Revision: https://reviews.llvm.org/D58202
llvm-svn: 362257
IslAst could mark two nested outer loops as "OutermostParallel". It
caused that the code generator tried to OpenMP-parallelize both loops,
which it is not prepared loop.
It was because the recursive AST build algorithm managed a flag
"InParallelFor" to ensure that no nested loop is also marked as
"OutermostParallel". Unfortunatetly the same flag was used by nodes
marked as SIMD, and reset to false after the SIMD node. Since loops can
be marked as SIMD inside "OutermostParallel" loops, the recursive
algorithm again tried to mark loops as "OutermostParellel" although
still nested inside another "OutermostParallel" loop.
The fix exposed another bug: The function "astScheduleDimIsParallel" was
only called when a loop was potentially "OutermostParallel" or
"InnermostParallel", but as a side-effect also determines the minimum
dependence distance. Hence, changing when we need to know whether a loop
is "OutermostParallel" also changed which loop was annotated with
"#pragma minimal dependence distance".
Moreover, some complex condition linked with "InParallelFor" determined
whether a loop should be an "InnermostParallel" loop. It missed some
situations where it would not use mark as such although being inside an
SIMD mark node, and therefore not be annotated using "#pragma simd".
The changes in particular:
1. Split the "InParallelFor" flag into an "InParallelFor" and an
"InSIMD" flag.
2. Unconditionally call "astScheduleDimIsParallel" for its side-effects
and store the result in "InParallel" for later use.
3. Simplify the condition when a loop is "InnermostParallel".
Fixes llvm.org/PR33153 and llvm.org/PR38073.
llvm-svn: 343212
This update:
- Removes several deprecated functions (e.g., isl_band).
- Improves the pretty-printing of sets by detecting modulos and "false"
equalities.
- Minor improvements to coalescing and increased robustness of the isl
scheduler.
This update does not yet include isl commit isl-0.18-90-gd00cb45
(isl_pw_*_alloc: add missing check for compatible spaces, Wed Sep 6 12:18:04
2017 +0200), as this additional check is too tight and unfortunately causes
two test case failures in Polly. A patch has been submitted to isl and will be
included in the next isl update for Polly.
llvm-svn: 325557
Summary:
Most changes are mechanical, but in one place I changed the program semantics
by fixing a likely bug:
In `Scop::hasFeasibleRuntimeContext()`, I'm now explicitely handling the
error-case. Before, when the call to `addNonEmptyDomainConstraints()`
returned a null set, this (probably) accidentally worked because
isl_bool_error converts to true. I'm checking for nullptr now.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: Meinersbur
Subscribers: nemanjai, kbarton, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D39971
llvm-svn: 318632
Fix walking over the schedule tree to collect its properties
(Number of permutable bands etc.).
Also add regression tests for these statistics.
llvm-svn: 313750
The remaining parts produced by the full partial tile isolation can contain
hot spots that are worth to be optimized. Currently, we rely on the simple
loop unrolling pass, LiCM and the SLP vectorizer to optimize such parts.
However, the approach can suffer from the lack of the information about
aliasing that Polly provides using additional alias metadata or/and the lack
of the information required by simple loop unrolling pass.
This patch is the first step to optimize the remaining parts. To do it, we
unroll and separate them. In case of, for instance, Intel Kaby Lake, it helps
to increase the performance of the generated code from 39.87 GFlop/s to
49.23 GFlop/s.
The next possible step is to avoid unrolling performed by Polly in case of
isolated and remaining parts and rely only on simple loop unrolling pass and
the Loop vectorizer.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D37692
llvm-svn: 312929
Summary:
There is no need to emit alias metadata for scalars, as basicaa will easily
distinguish them from arrays. This reduces the size of the metadata we generate.
This is especially useful after we moved to -polly-position=before-vectorizer,
where a lot more scalar dependences are introduced, which increased the size of
the alias analysis metadata and made us commonly reach the limits after which
we do not emit alias metadata that have been introduced to prevent quadratic
growth of this alias metadata.
This improves 2mm performance from 1.5 seconds to 0.17 seconds.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: Meinersbur
Subscribers: pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D37028
llvm-svn: 311498
Currently, in case of GEMM and the pattern matching based optimizations, we
use only the SLP Vectorizer out of two LLVM vectorizers. Since the Loop
Vectorizer can get in the way of optimal code generation, we disable the Loop
Vectorizer for the innermost loop using mark nodes and emitting the
corresponding metadata.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D36928
llvm-svn: 311473
The pattern recognition for MatMul is restrictive.
The number of "disjuncts" in the isl_map containing constraint
information was previously required to be 1
(as per isl_*_coalesce - which should ideally produce a domain map with
a single disjunct, but does not under some circumstances).
This was changed and made more flexible.
Contributed-by: Annanay Agarwal <cs14btech11001@iith.ac.in>
Differential Revision: https://reviews.llvm.org/D36460
llvm-svn: 311302
We introduce another level of alias metadata to distinguish the individual
non-aliasing accesses that have inter iteration alias-free base pointers
marked with "Inter iteration alias-free" mark nodes. To distinguish two
accesses, the comparison of raw pointers representing base pointers is used.
In case of, for example, ublas's prod function that implements GEMM, and
DeLiCM we can get accesses to same location represented by different raw
pointers. Consequently, we create different alias sets that can prevent
accesses from, for example, being sinked or hoisted.
To avoid the issue, we compare the corresponding SCEV information instead
of the corresponding raw pointers.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D35761
llvm-svn: 310380
Currently, only convex isolation sets can be efficiently processed by isl.
Consequently, as a temporary solution, we use a different algorithm for partial
tile isolation that helps to build convex isolation sets in some cases.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D36278
llvm-svn: 310374
In certain cases delicm might decide to not leave the original array write in
the loop body, but to remove it and instead leave a transformed phi node as
write access. This commit teached the matmul pattern detection to order the
memory accesses according to when the access actually happens and use this
information to detect the new pattern. This makes pattern based matmul
optimization work for 2mm and 3mm in polybench 4 after
polly-position=before-vectorizer has been enabled.
llvm-svn: 310338
Some optimizations (e.g., DeLICM) can modify memory accesses (e.g., change
their MemoryKind). Consequently, the pattern matching should take it into
the account.
Reviewed-by: Tobias Grosser <tobias@grosser.es>,
Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D33138
llvm-svn: 308494
Summary:
We do not keep domain constraints on access functions when building the
scop. Hence, for consistency reasons, it makes also sense to not include
them when storing a new access function. This change results in simpler
access functions that make output easier to read.
This patch also helps to make DeLICMed memory accesses to be understood by
our matrix multiplication pattern matching pass. Further changes to the
matrix multiplication pattern matching are needed for this to work, so the
corresponding test case will be added in a future commit.
Reviewers: Meinersbur, bollu, gareevroman, efriedma, huihuiz, sebpop, simbuerg
Subscribers: pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D35237
llvm-svn: 308215
Summary: This is a general maintenance update
Reviewers: grosser
Subscribers: srhines, fedor.sergeev, pollydev, llvm-commits
Contributed-by: Maximilian Falkenstein <falkensm@student.ethz.ch>
Differential Revision: https://reviews.llvm.org/D34903
llvm-svn: 307090
Ensure that all array base pointers are assigned before generating
aliasing metadata by allocating new arrays beforehand.
Before this patch, getBasePtr() returned nullptr for new arrays because
the arrays were created at a later point. Nullptr did not match to any
array after the created array base pointers have been assigned and when
the loads/stores are generated.
llvm-svn: 305675
Today Polly generates induction variable in this way:
polly.indvar = phi 0, polly.indvar.next
...
polly.indvar.next = polly.indvar + stide
polly.loop_cond = predicate polly.indvar, (UB - stride)
Instead of:
polly.indvar = phi 0, polly.indvar.next
...
polly.indvar.next = polly.indvar + stide
polly.loop_cond = predicate polly.indvar.next, UB
The way Polly generate induction variable cause some problem in the indvar simplify pass.
This patch make polly generate the later form, by assuming the induction variable never overflow
Differential Revision: https://reviews.llvm.org/D33089
llvm-svn: 302866
Dimensions of band nodes can be implicitly permuted by the algorithm applied
during the schedule generation.
For example, in case of the following matrix-matrix multiplication,
for (i = 0; i < 1024; i++)
for (k = 0; k < 1024; k++)
for (j = 0; j < 1024; j++)
C[i][j] += A[i][k] * B[k][j];
it can produce the following schedule tree
domain: "{ Stmt_for_body6[i0, i1, i2] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and
0 <= i2 <= 1023 }"
child:
schedule: "[{ Stmt_for_body6[i0, i1, i2] -> [(i0)] },
{ Stmt_for_body6[i0, i1, i2] -> [(i1)] },
{ Stmt_for_body6[i0, i1, i2] -> [(i2)] }]"
permutable: 1
coincident: [ 1, 1, 0 ]
The current implementation of the pattern matching optimizations relies on the
initial ordering of dimensions. Otherwise, it can produce the miscompilation
(e.g., [1]).
This patch helps to restore the initial ordering of dimensions by recreating
the band node when the corresponding conditions are satisfied.
Refs.:
[1] - https://bugs.llvm.org/show_bug.cgi?id=32500
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D31741
llvm-svn: 299662
Because Polly exposes parameters that directly influence tile size
calculations, one can setup situations like divide-by-zero.
Check against a possible divide-by-zero in getMacroKernelParams
and return early.
Also assert at the end of getMacroKernelParams that the block sizes
computed for matrices are positive (>= 1).
Tags: #polly
Differential Revision: https://reviews.llvm.org/D31708
llvm-svn: 299633
Introduce another level of alias metadata to distinguish the individual
non-aliasing accesses that have inter iteration alias-free base pointers
marked with "Inter iteration alias-free" mark nodes. It can be used to,
for example, distinguish different stores (loads) produced by unrolling of
the innermost loops and, subsequently, sink (hoist) them by LICM.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D30606
llvm-svn: 298510
In ScheduleOptimizer::isTileableBand(), allow the case in which
the band node's child is an isl_schedule_sequence_node and its
grandchildren isl_schedule_leaf_nodes. This case can arise when
two or more statements are fused by the isl scheduler.
The tile_after_fusion.ll test has two statements in separate
loop nests and checks whether they are tiled after being fused
when polly-opt-fusion equals "max".
Reviewers: grosser
Subscribers: gareevroman, pollydev
Tags: #polly
Contributed-by: Theodoros Theodoridis <theodort@student.ethz.ch>
Differential Revision: https://reviews.llvm.org/D30815
llvm-svn: 297587
Currently, pattern based optimizations of Polly can identify matrix
multiplication and optimize it according to BLIS matmul optimization pattern
(see ScheduleTreeOptimizer for details). This patch makes optimizations
based on pattern matching be enabled by default.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D30293
llvm-svn: 295958
To determine parameters of the matrix multiplication, we check RAW dependencies
that can be expressed using only reduction dependencies. Consequently, we
should check the reduction dependencies, if this is the case.
Reviewed-by: Tobias Grosser <tobias@grosser.es>,
Sven Verdoolaege <skimo-polly@kotnet.org>
Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29814
llvm-svn: 294836
The size of the operands type is the one of the parameters required
to determine the BLIS micro-kernel. We get the size of the widest type
of the matrix multiplication operands in case there are several
different types.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29269
llvm-svn: 294828
There are problems with using the machine information to derive the precise
vector size on polly-amd64-linux and polly-arm-linux. We temporarily disable
the problematic run lines.
llvm-svn: 294571
optimization
Isolate a set of partial tile prefixes to allow hoisting and sinking out of
the unrolled innermost loops produced by the optimization of the matrix
multiplication.
In case it cannot be proved that the number of loop iterations can be evenly
divided by tile sizes and we tile and unroll the point loop, the isl generates
conditional expressions. Subsequently, the conditional expressions can prevent
stores and loads of the unrolled loops from being sunk and hoisted.
The patch isolates a set of partial tile prefixes, which have exactly Mr x Nr
iterations of the two innermost loops, the result of the loop tiling performed
by the matrix multiplication optimization, where Mr and Mr are parameters of
the micro-kernel. This helps to get rid of the conditional expressions of
the unrolled innermost loops. Probably this approach can be replaced with
padding in future.
In case of, for example, the gemm from Polybench/C 3.2 and parametric loop
bounds, it helps to increase the performance from 7.98 GFlops (27.71% of
theoretical peak) to 21.47 GFlops (74.57% of theoretical peak). Hence, we
get the same performance as in case of scalar loops bounds.
It also cause compile time regression. The compile-time is increased from
0.795 seconds to 0.837 seconds in case of scalar loops bounds and from 1.222
seconds to 1.490 seconds in case of parametric loops bounds.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D29244
llvm-svn: 294564
with optimizeMatMulPattern
This patch makes ScheduleTreeOptimizer::optimizeBand return a schedule node
optimized with optimizeMatMulPattern. Otherwise, it could not use the isolate
option, because standardBandOpts could try to tile a band node with anchored
subtree and get the error, since the use of the isolate option causes any tree
containing the node to be considered anchored. Furthermore, it is not intended
to apply standard optimizations, when the matrix multiplication has been
detected.
llvm-svn: 294444
multiplication
The current identification of a SCoP statement that implement a matrix
multiplication does not help to identify different permutations of loops that
contain it and check for dependencies, which can prevent it from being
optimized. It also requires external determination of the operands of
the matrix multiplication. This patch contains the implementation of a new
algorithm that helps to avoid these issues. It also modifies the test cases
that generate matrix multiplications with linearized accesses, because
the new algorithm does not support them.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>,
Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28357
llvm-svn: 293890
If the parameters of the target cache (i.e., cache level sizes, cache level
associativities) are not specified or have wrong values, we use ones for
parameters of the macro-kernel and do not perform data-layout optimizations of
the matrix multiplication. In this patch we specify the default values of the
cache parameters to be able to apply the pattern matching optimizations even in
this case. Since there is no typical values of this parameters, we use the
parameters of Intel Core i7-3820 SandyBridge that also help to attain the
high-performance on IBM POWER System S822 and IBM Power 730 Express server.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28090
llvm-svn: 290518
Typically processor architectures do not include an L3 cache, which means that
Nc, the parameter of the micro-kernel, is, for all practical purposes,
redundant ([1]). However, its small values can cause the redundant packing of
the same elements of the matrix A, the first operand of the matrix
multiplication. At the same time, big values of the parameter Nc can cause
segmentation faults in case the available stack is exceeded.
This patch adds an option to specify the parameter Nc as a multiple of
the parameter of the micro-kernel Nr.
In case of Intel Core i7-3820 SandyBridge and the following options,
clang -O3 gemm.c -I utilities/ utilities/polybench.c -DPOLYBENCH_TIME
-march=native -mllvm -polly -mllvm -polly-pattern-matching-based-opts=true
-DPOLYBENCH_USE_SCALAR_LB -mllvm -polly-target-cache-level-associativity=8,8
-mllvm -polly-target-cache-level-sizes=32768,262144 -mllvm
-polly-target-latency-vector-fma=8
it helps to improve the performance from 11.303 GFlops/sec (39,247% of
theoretical peak) to 17.896 GFlops/sec (62,14% of theoretical peak).
Refs.:
[1] - http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28019
llvm-svn: 290256
multiplication
Previously we had two-dimensional accesses to store packed operands of
the matrix multiplication for the sake of simplicity of the packed arrays.
However, addition of the third dimension helps to simplify the corresponding
memory access, reduce the execution time of isl operations applied to it, and
consequently reduce the compile-time of Polly. For example, in case of
Intel Core i7-3820 SandyBridge and the following options,
clang -O3 gemm.c -I utilities/ utilities/polybench.c -DPOLYBENCH_TIME
-march=native -mllvm -polly -mllvm -polly-pattern-matching-based-opts=true
-DPOLYBENCH_USE_SCALAR_LB -mllvm -polly-target-cache-level-associativity=8,8
-mllvm -polly-target-cache-level-sizes=32768,262144 -mllvm
-polly-target-latency-vector-fma=7
it helps to reduce the compile-time from about 361.456 seconds to about 0.816
seconds.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>,
Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D27878
llvm-svn: 290251
To prevent copy statements from accessing arrays out of bounds, ranges of their
extension maps are restricted, according to the constraints of domains.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D25655
llvm-svn: 289815
gemm ([1]). In particular, elements of the matrix B, the second operand of
matrix multiplication, are reused between iterations of the innermost loop.
To keep the reused data in cache, only elements of matrix A, the first operand
of matrix multiplication, should be evicted during an iteration of the
innermost loop. To provide such a cache replacement policy, elements of the
matrix A can, in particular, be loaded first and, consequently, be
least-recently-used.
In our case matrices are stored in row-major order instead of column-major
order used in the BLIS implementation ([1]). One of the ways to address it is
to accordingly change the order of the loops of the loop nest. However, it
makes elements of the matrix A to be reused in the innermost loop and,
consequently, requires to load elements of the matrix B first. Since the LLVM
vectorizer always generates loads from the matrix A before loads from the
matrix B and we can not provide it. Consequently, we only change the BLIS micro
kernel and the computation of its parameters instead. In particular, reused
elements of the matrix B are successively multiplied by specific elements of
the matrix A .
Refs.:
[1] - http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D25653
llvm-svn: 289806
Providing the context to the ast generator allows for additional simplifcations
and -- more importantly -- allows to generate loops with only partially bounded
domains, assuming the domains are bounded for all parameter configurations
that are valid as defined by the context.
This change fixes the crash reported in http://llvm.org/PR30956
The original reason why we did not include the context when generating an
AST was that CLooG and later isl used to sometimes transfer some of the
constraints that bound the size of parameters from the context into the
generated AST. This resulted in operations with very large constants, which
sometimes introduced problematic integer overflows. The latest versions of
the isl AST generator are careful to not introduce such constants.
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286442
This is the fourth patch to apply the BLIS matmul optimization pattern on matmul
kernels (http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel, plus two
packing routines. The macro-kernel is implemented in terms of two additional
loops around a micro-kernel. The micro-kernel is a loop around a rank-1
(i.e., outer product) update. In this change we perform copying to created
arrays, which is the last step to implement the packing transformation.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D23260
llvm-svn: 281441
We do not need the size of the outermost dimension in most cases, but if we
allocate memory for newly created arrays, that size is needed.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D23991
llvm-svn: 281234
Dump polyhedral descriptions of Scops optimized with the isl scheduling
optimizer and the set of post-scheduling transformations applied
on the schedule tree to be able to check the work of the IslScheduleOptimizer
pass at the polyhedral level.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D23740
llvm-svn: 279395
This is the second patch to apply the BLIS matmul optimization pattern
on matmul kernels
(http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel, plus
two packing routines. The macro-kernel is implemented in terms
of two additional loops around a micro-kernel. The micro-kernel
is a loop around a rank-1 (i.e., outer product) update. In this change
we create the BLIS macro-kernel by applying a combination of tiling
and interchanging. In subsequent changes we will implement the packing
transformation.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D21491
llvm-svn: 276627
This is the first patch to apply the BLIS matmul optimization pattern
on matmul kernels
(http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf).
BLIS implements gemm as three nested loops around a macro-kernel,
plus two packing routines. The macro-kernel is implemented in terms
of two additional loops around a micro-kernel. The micro-kernel
is a loop around a rank-1 (i.e., outer product) update.
In this change we create the BLIS micro-kernel by applying
a combination of tiling and unrolling. In subsequent changes
we will add the extraction of the BLIS macro-kernel
and implement the packing transformation.
Contributed-by: Roman Gareev <gareevroman@gmail.com>
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D21140
llvm-svn: 273397
multiplication
Fix small issues related to characters, operators and descriptions of tests.
Differential Revision: http://reviews.llvm.org/D20806
llvm-svn: 271264
Fangrui Song