Summary:
Introduce a "hybrid" `-polly-target` option to optimise code for either the GPU or CPU.
When this target is selected, PPCGCodeGeneration will attempt first to optimise a Scop. If the Scop isn't modified, it is then sent to the passes that form the CPU pipeline, i.e. IslScheduleOptimizerPass, IslAstInfoWrapperPass and CodeGeneration.
In case the Scop is modified, it is marked to be skipped by the subsequent CPU optimisation passes.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: grosser
Subscribers: kbarton, nemanjai, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D34054
llvm-svn: 306863
Summary:
In case two arrays share base pointers in the same invariant load equivalence
class, we canonicalize all memory accesses to the first of these arrays
(according to their order in the equivalence class).
This enables us to optimize kernels such as boost::ublas by ensuring that
different references to the C array are interpreted as accesses to the same
array. Before this change the runtime alias check for ublas would fail, as it
would assume models of the C array with differing (but identically valued) base
pointers would reference distinct regions of memory whereas the referenced
memory regions were indeed identical.
As part of this change we remove most of the MemoryAccess::get*BaseAddr
interface. We removed already all references to get*BaseAddr in previous
commits to ensure that no code relies on matching base pointers between
memory accesses and scop arrays -- except for three remaining uses where we
need the original base pointer. We document for these situations that
MemoryAccess::getOriginalBaseAddr may return a base pointer that is distinct
to the base pointer of the scop array referenced by this memory access.
Reviewers: sebpop, Meinersbur, zinob, gareevroman, pollydev, huihuiz, efriedma, jdoerfert
Reviewed By: Meinersbur
Subscribers: etherzhhb
Tags: #polly
Differential Revision: https://reviews.llvm.org/D28518
llvm-svn: 302636
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
= Change of WAR, WAW generation: =
- `buildFlow(Sink, MustSource, MaySource, Sink)` treates any flow of the form
`sink <- may source <- must source` as a *may* dependence.
- we used to call:
```lang=cpp, name=old-flow-call.cpp
Flow = buildFlow(MustWrite, MustWrite, Read, Schedule);
WAW = isl_union_flow_get_must_dependence(Flow);
WAR = isl_union_flow_get_may_dependence(Flow);
```
- This caused some WAW dependences to be treated as WAR dependences.
- Incorrect semantics.
- Now, we call WAR and WAW correctly.
== Correct WAW: ==
```lang=cpp, name=new-waw-call.cpp
Flow = buildFlow(Write, MustWrite, MayWrite, Schedule);
WAW = isl_union_flow_get_may_dependence(Flow);
isl_union_flow_free(Flow);
```
== Correct WAR: ==
```lang=cpp, name=new-war-call.cpp
Flow = buildFlow(Write, Read, MustaWrite, Schedule);
WAR = isl_union_flow_get_must_dependence(Flow);
isl_union_flow_free(Flow);
```
- We want the "shortest" WAR possible (exact dependences).
- We mark all the *must-writes* as may-source, reads as must-souce.
- Then, we ask for *must* dependence.
- This removes all the reads that flow through a *must-write*
before reaching a sink.
- Note that we only block ealier writes with *must-writes*. This is
intuitively correct, as we do not want may-writes to block
must-writes.
- Leaves us with direct (R -> W).
- This affects reduction generation since RED is built using WAW and WAR.
= New StrictWAW for Reductions: =
- We used to call:
```lang=cpp,name=old-waw-war-call.cpp
Flow = buildFlow(MustWrite, MustWrite, Read, Schedule);
WAW = isl_union_flow_get_must_dependence(Flow);
WAR = isl_union_flow_get_may_dependence(Flow);
```
- This *is* the right model of WAW we need for reductions, just not in general.
- Reductions need to track only *strict* WAW, without any interfering reductions.
= Explanation: Why the new WAR dependences in tests are correct: =
- We no longer set WAR = WAR - WAW
- Hence, we will have WAR dependences that were originally removed.
- These may look incorrect, but in fact make sense.
== Code: ==
```lang=llvm, name=new-war-dependence.ll
; void manyreductions(long *A) {
; for (long i = 0; i < 1024; i++)
; for (long j = 0; j < 1024; j++)
; S0: *A += 42;
;
; for (long i = 0; i < 1024; i++)
; for (long j = 0; j < 1024; j++)
; S1: *A += 42;
;
```
=== WAR dependence: ===
{ S0[1023, 1023] -> S1[0, 0] }
- Between `S0[1023, 1023]` and `S1[0, 0]`, we will have the dependences:
```lang=cpp, name=dependence-incorrect, counterexample
S0[1023, 1023]:
*-- tmp = *A (load0)--*
WAR 2 add = tmp + 42 |
*-> *A = add (store0) |
WAR 1
S1[0, 0]: |
tmp = *A (load1) |
add = tmp + 42 |
A = add (store1)<-*
```
- One may assume that WAR2 *hides* WAR1 (since store0 happens before
store1). However, within a statement, Polly has no idea about the
ordering of loads and stores.
- Hence, according to Polly, the code may have looked like this:
```lang=cpp, name=dependence-correct
S0[1023, 1023]:
A = add (store0)
tmp = A (load0) ---*
add = A + 42 |
WAR 1
S1[0, 0]: |
tmp = A (load1) |
add = A + 42 |
A = add (store1) <-*
```
- So, Polly generates (correct) WAR dependences. It does not make sense
to remove these dependences, since they are correct with respect to
Polly's model.
Reviewers: grosser, Meinersbur
tags: #polly
Differential revision: https://reviews.llvm.org/D31386
llvm-svn: 299429
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
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
Add a simple example to update the documentation on how the packing
transformation is implemented.
Reviewed-by: Tobias Grosser <tobias@grosser.es>,
Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D28021
llvm-svn: 293429
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 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
LLVM's coding guideline suggests to not use @brief for one-sentence doxygen
comments to improve readability. Switch this once and for all to ensure people
do not copy @brief comments from other parts of Polly, when writing new code.
llvm-svn: 280468
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 third 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 replacement of
the access relations and create empty arrays, which are steps to implement
the packing transformation. In subsequent changes we will implement copying
to created arrays.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: http://reviews.llvm.org/D22187
llvm-svn: 278666
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 ensures that the error status set with -polly-on-isl-error-abort is
maintained even after running DependenceInfo and ScheduleOptimizer. Both
passes temporarily set the error status to CONTINUE as the dependence
analysis uses a compute-out and the scheduler may not be able to derive
a schedule. In both cases we want to not abort, but to handle the error
gracefully. Before this commit, we always set the error reporting to ABORT
after these passes. After this commit, we use the error reporting mode that was
active earlier.
This comes without a test case as this would require us to introduce (memory)
errors which would trigger the isl errors.
llvm-svn: 274272
llvm commonly adds a comment to the closing brace of a namespace to indicate
which namespace is closed. clang-tidy provides with llvm-namespace-comment
a handy tool to check for this habit. We use it to ensure we consitently use
namespace comments in Polly.
There are slightly different styles in how namespaces are closed in LLVM. As
there is no large difference between the different comment styles we go for the
style clang-tidy suggests by default.
To reproduce this fix run:
for i in `ls tools/polly/lib/*/*.cpp`; \
clang-tidy -checks='-*,llvm-namespace-comment' -p build $i -fix \
-header-filter=".*"; \
done
This cleanup was suggested by Eugene Zelenko <eugene.zelenko@gmail.com> in
http://reviews.llvm.org/D21488 and was split out to increase readability.
llvm-svn: 273621
Instead of using 0 or NULL use the C++11 nullptr symbol when referencing null
pointers.
This cleanup was suggested by Eugene Zelenko <eugene.zelenko@gmail.com> in
http://reviews.llvm.org/D21488 and was split out to increase readability.
llvm-svn: 273435
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