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
This allows us to delinearize code such as the one below, where the array
sizes are A[][2 * n] as there are n times two elements in the innermost
dimension. Alternatively, we could try to generate another dimension for the
struct in the innermost dimension, but as the struct has constant size,
recovering this dimension is easy.
struct com {
double Real;
double Img;
};
void foo(long n, struct com A[][n]) {
for (long i = 0; i < 100; i++)
for (long j = 0; j < 1000; j++)
A[i][j].Real += A[i][j].Img;
}
int main() {
struct com A[100][1000];
foo(1000, A);
llvm-svn: 288489
Add an empty DeLICM pass, without any functional parts.
Extracting the boilerplate from the the functional part reduces the size of the
code to review (https://reviews.llvm.org/D24716)
Suggested-by: Tobias Grosser <tobias@grosser.es>
llvm-svn: 288160
We now collect:
Number of total loops
Number of loops in scops
Number of scops
Number of scops with maximal loop depth 1
Number of scops with maximal loop depth 2
Number of scops with maximal loop depth 3
Number of scops with maximal loop depth 4
Number of scops with maximal loop depth 5
Number of scops with maximal loop depth 6 and larger
Number of loops in scops (profitable scops only)
Number of scops (profitable scops only)
Number of scops with maximal loop depth 1 (profitable scops only)
Number of scops with maximal loop depth 2 (profitable scops only)
Number of scops with maximal loop depth 3 (profitable scops only)
Number of scops with maximal loop depth 4 (profitable scops only)
Number of scops with maximal loop depth 5 (profitable scops only)
Number of scops with maximal loop depth 6 and larger (profitable scops only)
These statistics are certainly completely accurate as we might drop scops
when building up their polyhedral representation, but they should give a good
indication of the number of scops we detect.
llvm-svn: 287973
Our original statistics were added before we introduced a more fine-grained
diagnostic system, but the granularity of our statistics has never been
increased accordingly. This change introduces now one statistic counter per
diagnostic to enable us to collect fine-grained statistics about who certain
scops are not detected. In case coarser grained statistics are needed, the
user is expected to combine counters manually.
llvm-svn: 287968
Introduce the new flag -polly-codegen-generate-expressions which forces Polly
to code generate AST expressions instead of using our SCEV based access
expression generation even for cases where the original memory access relation
was not changed and the SCEV based access expression could be code generated
without any issue.
This is an experimental option for better testing the isl ast expression
generation. The default behavior of Polly remains unchanged. We also exclude
a couple of cases for which the AST expression is not yet working.
llvm-svn: 287694
Drop instructions that do not influence the memory impact of a basic block.
They are not needed to reproduce the original bug (verified) and will cause
random test noise if we would decide to only model the instructions that
have visible side-effects.
llvm-svn: 287626
Add two store instructions at the end of basic blocks that are required to
reproduce the original bug to ensure we always process and model these basic
blocks. This makes this test case stable even in case we would decide to bail
out early of basic blocks which do not modify the global state. Also add
additional check lines to verify how we model the basic block.
llvm-svn: 287625
We add CHECK lines to this test case to make it easier to see the difference
between affine and non-affine memory accesses. We also change the test case to
use a parameteric index expression as otherwise our range analysis will
understand that the non-affine memory access can only access input[1],
which makes it difficult to see that the memory access is in-fact modeled as
non-affine access.
llvm-svn: 287623
Do not assume a load to be hoistable/invariant if the pointer is used by
another instruction in the SCoP that might write to memory and that is
always executed.
llvm-svn: 287272
The validity of a branch condition must be verified at the location of the
branch (the branch instruction), not the location of the icmp that is
used in the branch instruction. When verifying at the wrong location, we
may accept an icmp that is defined within a loop which itself dominates, but
does not contain the branch instruction. Such loops cannot be modeled as
we only introduce domain dimensions for surrounding loops. To address this
problem we change the scop detection to evaluate and verify SCEV expressions at
the right location.
This issue has been around since at least r179148 "scop detection: properly
instantiate SCEVs to the place where they are used", where we explicitly
set the scope to the wrong location. Before this commit the scope
was not explicitly set, which probably also resulted in the scope around the
ICmp to be choosen.
This resolves http://llvm.org/PR30989
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286769
Assumptions can either be added for a given basic block, in which case the set
describing the assumptions is expected to match the dimensions of its domain.
In case no basic block is provided a parameter-only set is expected to describe
the assumption.
The piecewise expressions that are generated by the SCEVAffinator sometimes
have a zero-dimensional domain (e.g., [p] -> { [] : p <= -129 or p >= 128 }),
which looks similar to a parameter-only domain, but is still a set domain.
This change adds an assert that checks that we always pass parameter domains to
addAssumptions if BB is empty to make mismatches here fail early.
We also change visitTruncExpr to always convert to parameter sets, if BB is
null. This change resolves http://llvm.org/PR30941
Another alternative to this change would have been to inspect all code to make
sure we directly generate in the SCEV affinator parameter sets in case of empty
domains. However, this would likely complicate the code which combines parameter
and non-parameter domains when constructing a statement domain. We might still
consider doing this at some point, but as this likely requires several non-local
changes this should probably be done as a separate refactoring.
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286444
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
When extracting constant expressions out of SCEVs, new parameters may be
introduced, which have not been registered before. This change scans
SCEV expressions after constant extraction again to make sure newly
introduced parameters are registered.
We may for example extract the constant '8' from the expression '((8 * ((%a *
%b) + %c)) + (-8 * %a))' and obtain the expression '(((-1 + %b) * %a) + %c)'.
The new expression has a new parameter '(-1 + %b) * %a)', which was not
registered before, but must be registered to not crash.
This closes http://llvm.org/PR30953
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286430
In r248701 "Allow switch instructions in SCoPs" support for switch statements
has been introduced, but support for switch statements in loop latches was
incomplete. This change completely disables switch statements in loop latches.
The original commit changed addLoopBoundsToHeaderDomain to support non-branch
terminator instructions, but this change was incorrect: it added a check for
BI != null to the if-branch of a condition, but BI was used in the else branch
es well. As a result, when a non-branch terminator instruction is encounted a
nullptr dereference is triggered. Due to missing test coverage, this bug was
overlooked.
r249273 "[FIX] Approximate non-affine loops correctly" added code to disallow
switch statements for non-affine loops, if they appear in either a loop latch
or a loop exit. We adapt this code to now prohibit switch statements in
loop latches even if the control condition is affine.
We could possibly add support for switch statements in loop latches, but such
support should be evaluated and tested separately.
This fixes llvm.org/PR30952
Reported-by: Eli Friedman <efriedma@codeaurora.org>
llvm-svn: 286426
We don't actually check whether a MemoryAccess is affine in very many
places, but one important one is in checks for aliasing.
Differential Revision: https://reviews.llvm.org/D25706
llvm-svn: 285746
When adding an llvm.memcpy instruction to AliasSetTracker, it uses the raw
source and target pointers which preserve bitcasts.
MemAccInst::getPointerOperand() also returns the raw target pointers, but
Scop::buildAliasGroups() did not for the source pointer. This lead to mismatches
between AliasSetTracker and ScopInfo on which pointer to use.
Fixed by also using raw pointers in Scop::buildAliasGroups().
llvm-svn: 285071
Integer math in LLVM IR is modular. Integer math in isl is
arbitrary-precision. Modeling LLVM IR math correctly in isl requires
either adding assumptions that math doesn't actually overflow, or
explicitly wrapping the math. However, expressions with the "nsw" flag
are special; we can pretend they're arbitrary-precision because it's
undefined behavior if the result wraps. SCEV expressions based on IR
instructions with an nsw flag also carry an nsw flag (roughly; actually,
the real rule is a bit more complicated, but the details don't matter
here).
Before this patch, SCEV flags were also overloaded with an additional
function: the ZExt code was mutating SCEV expressions as a hack to
indicate to checkForWrapping that we don't need to add assumptions to
the operand of a ZExt; it'll add explicit wrapping itself. This kind of
works... the problem is that if anything else ever touches that SCEV
expression, it'll get confused by the incorrect flags.
Instead, with this patch, we make the decision about whether to
explicitly wrap the math a bit earlier, basing the decision purely on
the SCEV expression itself, and not its users.
Differential Revision: https://reviews.llvm.org/D25287
llvm-svn: 284848
Update test after commit r284501:
[SCEV] Make CompareValueComplexity a little bit smarter
Contributed-by: Sanjoy Das <sanjoy@playingwithpointers.com>
llvm-svn: 284543
lit recursively iterates through the test subdirectories and finds the ISL
unittest. For this test to work, the polly-isl-test executable needs to be
compiled.
Add the polly-isl-test dependency to POLLY_TEST_DEPS. This makes check-polly and
check-polly-tests work from a fresh build directory.
llvm-svn: 284339
The test non_affine_loop_used_later.ll also tests the profability heuristic. Add
the option -polly-unprofitable-scalar-accs explicitely to ensure that the test
succeeds if the default value is changed.
llvm-svn: 284338
Under some conditions MK_Value read accessed where converted to MK_ExitPHI read
accessed. This is unexpected because MK_ExitPHI read accesses are implicit after
the scop execution. This behaviour was introduced in r265261, which fixed a
failed assertion/crash in CodeGen.
Instead, we fix this failure in CodeGen itself. createExitPHINodeMerges(),
despite its name, also handles accesses of kind MK_Value, only to skip them
because they access values that are usually not PHI nodes in the SCoP region's
exit block. Except in the situation observed in r265261.
Do not convert value accessed to ExitPHI accesses and do not handle
value accesses like ExitPHI accessed in CodeGen anymore.
llvm-svn: 284023
Folders in Visual Studio solutions help organize the build artifacts from all
LLVM projects. There is a folder to keep Polly-built files in.
llvm-svn: 283546
Running isl tests is important to gain confidence that the isl build we created
works as expected. Besides the actual isl tests, there are also isl AST
generation tests shipped with isl. This change only adds support for the isl
unit tests. AST generation test support is left for a later commit.
There is a choice to run tests directly through the build system or in the
context of lit. We choose to run tests as part of lit to as this allows us to
easily set environment variables, print output only on error and generally run
the tests directly from the lit command.
Reviewers: brad.king, Meinersbur
Subscribers: modocache, brad.king, pollydev, beanz, llvm-commits, mgorny
Differential Revision: https://reviews.llvm.org/D25155
llvm-svn: 283245
With this option one can disable the heuristic that assumes that statements with
a scalar write access cannot be profitably optimized. Such a statement instances
necessarily have WAW-dependences to itself. With DeLICM scalar accesses can be
changed to array accesses, which can avoid these WAW-dependence.
llvm-svn: 283233
ScopArrayInfo used to determine base pointer origins by looking up whether the
base pointer is a load. The "base pointer" for scalar accesses is the
llvm::Value being accessed. This is only a symbolic base pointer, it
represents the alloca variable (.s2a or .phiops) generated for it at code
generation.
This patch disables determining base pointer origin for scalars.
A test case where this caused a crash will be added in the next commit. In that
test SAI tried to get the origin base pointer that was only declared later,
therefore not existing. This is probably only possible for scalars used in
PHINode incoming blocks.
llvm-svn: 283232
Summary:
Both `canUseISLTripCount()` and `addOverApproximatedRegion()` contained checks
to reject endless loops which are now removed and replaced by a single check
in `isValidLoop()`.
For reporting such loops the `ReportLoopOverlapWithNonAffineSubRegion` is
renamed to `ReportLoopHasNoExit`. The test case
`ReportLoopOverlapWithNonAffineSubRegion.ll` is adapted and renamed as well.
The schedule generation in `buildSchedule()` is based on the following
assumption:
Given some block B that is contained in a loop L and a SESE region R,
we assume that L is contained in R or the other way around.
However, this assumption is broken in the presence of endless loops that are
nested inside other loops. Therefore, in order to prevent erroneous behavior
in `buildSchedule()`, r265280 introduced a corresponding check in
`canUseISLTripCount()` to reject endless loops. Unfortunately, it was possible
to bypass this check with -polly-allow-nonaffine-loops which was fixed by adding
another check to reject endless loops in `allowOverApproximatedRegion()` in
r273905. Hence there existed two separate locations that handled this case.
Thank you Johannes Doerfert for helping to provide the above background
information.
Reviewers: Meinersbur, grosser
Subscribers: _jdoerfert, pollydev
Differential Revision: https://reviews.llvm.org/D24560
Contributed-by: Matthias Reisinger <d412vv1n@gmail.com>
llvm-svn: 281987
In case sequential kernels are found deeper in the loop tree than any parallel
kernel, the overall scop is probably mostly sequential. Hence, run it on the
CPU.
llvm-svn: 281849
Offloading to a GPU is only beneficial if there is a sufficient amount of
compute that can be accelerated. Many kernels just have a very small number
of dynamic compute, which means GPU acceleration is not beneficial. We
compute at run-time an approximation of how many dynamic instructions will be
executed and fall back to CPU code in case this number is not sufficiently
large. To keep the run-time checking code simple, we over-approximate the
number of instructions executed in each statement by computing the volume of
the rectangular hull of its iteration space.
llvm-svn: 281848
We may generate GPU kernels that store into scalars in case we run some
sequential code on the GPU because the remaining data is expected to already be
on the GPU. For these kernels it is important to not keep the scalar values
in thread-local registers, but to store them back to the corresponding device
memory objects that backs them up.
We currently only store scalars back at the end of a kernel. This is only
correct if precisely one thread is executed. In case more than one thread may
be run, we currently invalidate the scop. To support such cases correctly,
we would need to always load and store back from a corresponding global
memory slot instead of a thread-local alloca slot.
llvm-svn: 281838
Tobias Grosser