[Polly] Generalize the pattern matching to the case of tensor contractions.

[gareevroman]

Summary

The pattern matching optimization of Polly detects and optimizes dense general matrix-matrix multiplication. The generated code is close to high performance implementations of matrix-matrix multiplications, which are contained in manually tuned libraries [1]. The described pattern matching optimization is a particular case of tensor contraction optimization, which was introduced in [2].

This patch generalizes the pattern matching to the case of tensor contractions using the algorithm described in [2]. Following the ideas introduced in [3], it logically represents tensor contraction operands as matrix multiplication operands and uses the approach presented in [1].

Optimization of tensor contractions will be added in the next patch. These modifications can be found in https://github.com/gareevroman/llvm-project.

[1] - Low T.M., Igual F.D., Smith T.M., Quintana-Orti E.S. Analytical Modeling Is Enough for High-Performance BLIS // ACM Transactions on Mathemat­ical Software. 2016. Vol. 43, no. 2. P. 12:1—12:18. DOI: 10.1145/2925987.

[2] - Gareev R., Grosser T., Kruse M. High-Performance Generalized Tensor Op­erations: A Compiler-Oriented Approach // ACM Transactions on Architec­ture and Code Optimization (TACO). 2018. Vol. 15, no. 3. P. 34:1–34:27. DOI: 10.1145/3235029.

[3] - Matthews D. High-Performance Tensor Contraction without BLAS // SIAM Journal on Scientific Computing. 2018. Vol. 40, no. 1. P. C 1—C 24. DOI: 110.1137/16m108968x.

[Meinersbur]

Thanks for upstreaming your tensor optimization!

The naming of polly-pattern-matching-based-opts suggests that it includes all pattern-based optimizations, yet this introduces another flag polly-pattern-matching-based-tc-opts. I'd prefer polly-pattern-matching-based-opts controlling both optimizations, and then additional flags for enabling matrix-multiplication and tensor optimizations. Alternatively, rename polly-pattern-matching-based-opts to e.g. polly-matmul-opt. Also, you are adding this functionality into a file called MatmulOptimizer and a function called tryOptimizeMatMulPattern. Since matrix-multiplication is a tensor contraction, is ts-opt supposed to supperseed matrix multiplication? In short, I would like to know what what the relation between those two optimizations should be. I'd prefer to not have to maintain two optimizations if one is strictly more powerful than the other.

I'd enjoy some occasional comments and grouping of statements (empty lines) inside the functions in addition to the doxygen comments. For instance isTCOperandAcc is just a wall of code. For such property-checking functions, ideally each return should mention what property is violated here and why this property is required to be compatible with tensor optimization.

polly/lib/Transform/MatmulOptimizer.cpp

172–173

Please add more details on what the members represent.

179–181

std::set is a high-overhead implementation. Consider using DenseSet or SmallDenseSet. See https://www.llvm.org/docs/ProgrammersManual.html#llvm-adt-denseset-h

182–188

Is there an argument to use 30 and small size? If not, consider using just SmallVector<int>.

1068

[style] No reason to make this a wide string literal, especially if just used as an assertion failed message.

Apples to other occurrences as well.

1087

1093

or introduce intFromIslSize.

1126–1127

SmallVectorImpl is not specific to what the vector's small size is.

1137

Consider using polly::rangeIslSize for iterating over dimensions.

1164

Although already in an anon namespace, the other methods add static as well. I found it helps the compiler to warn if a static function is unnused.

1166–1167

Do you know of #include <llvm/ADT/SetOperations.h>? Unfortunately, these modify one set rather than returning a new set.

1196

1206–1207

This computes whether two sets a disjoint, it should not be required to compute the intersection.

1223

1261

getAccessesInOrder requires Stmt to not be a RegionStmt. Please add a test for it.

1271

If any of the returns are executed, what causes the pattern to be rejected (it's not return false)?

1340–1341

1342–1344

The check should not depend on n_basic_set, which is fragile and depends on whether on eg. coalesce was successful. Consider using something like polly::getConstant.

1388–1389

Consider lexmin_pw_multi_aff/lexmax_pw_multi_aff

polly/lib/Transform/ScheduleOptimizer.cpp

459–463

Instead of modifying the idea of whether a node is tilable, consider introducing another constraint-checking function, as we should have done with prevectorization as well.

[gareevroman]

Thank you very much for the review!

I've tried to address all comments. Additionally, I've updated the optimization for the case of filter nodes (e.g., polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm.ll). I believe that the optimization of tensor contractions is strictly more powerful than the optimization of matrix-multiplications. So, I'd suggest to replace the optimization of matrix-multiplications with it eventually.

[gareevroman]

polly/lib/Transform/MatmulOptimizer.cpp

1093

As far I understand, Dimensions.size() returns a value of type size_t instead of a value of the type isl_size. So, in the new version I used the unsigned type to avoid the cast.

1206–1207

That check is redundant. Thanks.

1223

As far as I understand, we cannot do this here because of the assignment to TCI.ADimensions and TCI.BDimensions

1261

I’ve added a check to containsOnlyTCAcc. Could you clarify how the test case should look like? Should it be a region statement that contains a matrix multiplication with right order of memory accesses?

1342–1344

I think that this check was redundant. I’ve removed it.

[gareevroman]

ping

[Meinersbur]

I am extremely sorry for the late review, I hope you could use the time to work on/polish the follow-up patch. Some things are a bit hard to understand independently. I promise to review in a more timely manner from now on, although I might always add that many comments.

Please add to the tests cases what they are supposed to be testing and maybe give them more meaningful filenames.

The following is successfully detected as tensor contraction. Is this intended?

void foo(double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
     for (int i = 0; i < 1024; i++)
         for (int j = 0; j < 1024; j++)
           for (int l = 0; l < 64; ++l)
              if (l != 0)
                for (int w = 0; w < 64; ++w)
                  C[i][j] += A[i][l][w] * B[w][j][l];
}

It might be if the codegen part is able exclude the element 0. In contrast, this one is rejected:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
     for (int i = 0; i < 1024; i++)
         for (int j = 0; j < 1024; j++)
           for (int l = 0; l < 64; l++)
             for (int w = 0; w < 64; ++w)
                if (w != 0)
                  C[i][j] += A[i][l][w] * B[w][j][l];
}

or this:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
     for (int i = 0; i < 1024; i++)
         for (int j = 0; j < 1024; j++)
           for (int l = 0; l < 64; l+=2)
             for (int w = 0; w < 64; ++w)
                  C[i][j] += A[i][l][w] * B[w][j][l];
}

I do get an assertion failure with this one:

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {
     for (int i = 0; i < 32; i++)
         for (int j = 0; j < 32; j++)
           for (int l = 0; l < 32; l++)
             for (int w = 0; w < 32; ++w)
                  C[i][j] += A[i][l][w] * B[w][j][i+3];
}

Here, i occurs as indices for A, B, and C and detected as TC. Is this supported?

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {
    for (int i = 0; i < 32; i++)
    for (int j = 0; j < 32; j++)
    for (int l = 0; l < 32; l++)
    for (int w = 0; w < 32; w++)
        C[i][j] += A[i][l][w] * B[w][j][i];
}

Is it possible to make it work with -polly-reschedule=0 as well? AFAICS there is nothing that requires rescheduling and would allow us to test the detection independently. isTCPattern seems to already consider multiple bands; the hurdle at the moment seems to be that the bands are not marked as permutable.

polly/lib/Transform/MatmulOptimizer.cpp

1567–1568

5. should not be necessary; any permutation of the surrounding loops can be valid. Eg,

for (w = 0; w < 64; ++w)
  for (l = 0; l < 64; ++l)
    for (i = 0; i < 1024; i++)
      for (j = 0; j < 1024; j++)
           C[i][j] += A[i][l][w] * B[w][j][l];

yields the same result.

1579

Prefer Node.isa<isl::schedule_node_leaf>() (and then typed subclass: Node.as<isl_schedule_node_leaf>())

[Meinersbur]

polly/lib/Transform/MatmulOptimizer.cpp

175–178

Please use doxygen comments to describe class/struct members. @{ @} can be used to group them.

Also add an empty line before a new comment begins.

1093

rangeIslSize should make it easier.

1113

I like the idea of verifying the correctness by reconstructing and comparing to the original.

Maybe do it at the end to verify that the entire TCInfoTy is correct? On the other size, earlier fail would be better. What do you think?

1138–1141

This is a weird way to find out which indices map to what other index, I guess the equivalent of isMatMulOperandAcc. It requires that the dimension number if part of the AccMap's range, and if there is any expression will fail (eg Stmt[i][j] -> A[i-1][j+1] matches the wrong dimensions), but at least there is the verification afterwards.

I am not sure I like this sort of cleverness; I'd rather expect some sort of introspection into the map's coefficients, but I also think this should work in nearly all relevant cases and should be save due to the verification, so lets keep it.

However, please document this better, eg. add an example on what is expected to happen.

1148–1149

The "plain" function are unfortunately not very robust, eg its result is different depending on the internal representation. I'd suggestion getConstant (from ISLTools) but only takes an pw_aff.

Could you extract uses of plain_get_val_if_fixed into such a function, and mark it as TODO to cope with it later?

1196

I can use JScopImport to set a scalar memory access to a partial write without adding additional dependencies; that is, I don't think this can be just ignored.

I suggest to have a single function that calls getAccessesInOrder and sort out which MemAccess is read/write in there, then analyze them.

1227–1229

Introduce a is_superset (etc) call?

1247–1249

Could we add utility functions such that this becomes unite(J, P) == IndexSet?

1261

Test in containsOnlyTCAcc is exactly what I was looking for. A region statement could look like this:

c = C[i][j];
if (/*non-affine condition*/) {
  (void)A[i][k] + B[k][j];
} else {
  C[i][j] = c;
}

which has the correct order of accesses but is obviously not what we are looking for.

1351

1372

This seems to check setermine whether there is a reduction (contraction) carried by loop number Pos. The function name could be more meaningful. Suggestion: isDepCarryingReductionOverDim (not nice, but "TcDep" can mean anything)

1373

Consider passing by const-reference.

1376

plain_get_val_if_fixed is not really robust as it depends on the internal representation that can be different after eg simplify.

Since this just checks to a specific value, the best would be to create a new set where all the fixed dimensions are that value (here: 1), and check whether DepDelta is a subset of it.

1387–1389

Why not BoundDeltas.subtract() instead of deltas?

1428–1431

Should we also check whether WAW, RAW dependences are incompatible?

1437–1438

lexmin/lexmax can be expensive. Wrap into a IslMaxOperationsGuard?

1439

You seem to assume an functional relationship from here on. If that's the case, you can keep the type a pw_multi_aff which supports more functions that you may have missed such as pw_multi_aff::add

1441

Consider using reverse(rangeIslSize(0,DeltasDimNum)) (from ISLTools.h).

1442–1447

This is going to check whether each element out of Intersection is a contraction over dimension i. Don't we also need to check that every iteration out of the band i is contributing to that contraction?

1551

This is not part of the pattern detection, but the optimization. Could we move it to the patch that does the actual optimization?

1563–1564

Could you describe here what those 4 accesses are?

1569

Could you add a high-level description how the algorithm actually works? I.e. dependencies used to determine contraction dimensions, etc.

1584–1587

This condition is effectively identical to the next

1590

This constraint should not be intrinsic to the algorithm, but I agree it to be easier to handle for now.

1593–1594

[style] No Almost-Always-Auto in LLVM's coding style.

1599

This looks for the outermost node that is not a filter or band. Is it possible that while that outermost node is not a TC contraction, one of the inner ones might? What if the outermost node is a filter, looks like it would just return false in this case.

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm.ll

13 ↗

(On Diff #393734)

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

2–4

Since this is not FileCheck-ing the LLVM-IR output, suppress it with -disable-output

[gareevroman]

Thank you very much for the review! I am sorry for the late response. I will try to to address all your comments within the next few weeks.

[Meinersbur]

In D114336#3470843, @gareevroman wrote:

Thank you very much for the review! I am sorry for the late response. I will try to to address all your comments within the next few weeks.

No worries, I wasn't responsive either :-(

[gareevroman]

1

The following is successfully detected as tensor contraction. Is this intended?

void foo(double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {

for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; ++l)
         if (l != 0)
           for (int w = 0; w < 64; ++w)
             C[i][j] += A[i][l][w] * B[w][j][l];

}

Yes, it was intended. The transformation helps to optimize a class of programs, which is broader then a tensor contraction. However, it heavily depends on the codegen part. I think that the improvement of the detection can be the goal of the future work.

It might be if the codegen part is able exclude the element 0. In contrast, this one is rejected:

In this case, the codegen excludes the element 0 for i2. I added a test case for this.

domain: "{ Stmt4[i0, i1, i2, i3] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and 0 < i2 <= 63 and 0 <= i3 <= 63 }"
…

In contrast, this one is rejected:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {

for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; l++)
        for (int w = 0; w < 64; ++w)
           if (w != 0)
             C[i][j] += A[i][l][w] * B[w][j][l];

}

or this:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {

for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; l+=2)
        for (int w = 0; w < 64; ++w)
             C[i][j] += A[i][l][w] * B[w][j][l];

}

As far as I know, in these cases, the codegen modifies some memory accesses. Consequently, they are not correspond to the current pattern.

ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef0[i0, i2, 1 + i3] };
ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef1[1 + i3, i1, i2] };
ReadAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };
MustWriteAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };

2

I do get an assertion failure with this one:

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {

for (int i = 0; i < 32; i++)
    for (int j = 0; j < 32; j++)
      for (int l = 0; l < 32; l++)
        for (int w = 0; w < 32; ++w)
             C[i][j] += A[i][l][w] * B[w][j][i+3];

}

I fixed the isTCOperandAcc function and checked that all other asserts are used properly.

3

I do get an assertion failure with this one:

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {

for (int i = 0; i < 32; i++)
    for (int j = 0; j < 32; j++)
      for (int l = 0; l < 32; l++)
        for (int w = 0; w < 32; ++w)
             C[i][j] += A[i][l][w] * B[w][j][i+3];

}

I fixed the isTCOperandAcc function and checked that all other asserts are used properly.

4

Here, i occurs as indices for A, B, and C and detected as TC. Is this supported?

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {

for (int i = 0; i < 32; i++)
for (int j = 0; j < 32; j++)
for (int l = 0; l < 32; l++)
for (int w = 0; w < 32; w++)
    C[i][j] += A[i][l][w] * B[w][j][i];

}

For some reason, I cannot reproduce that. I have added a corresponding test case. As far as I understand, that should be detected because of the line 1365.

1347 static bool containsOnlyTCAcc(isl::set Domain, isl::map PartialSchedule,

                            TCInfoTy &TCI) {
...

1365 if (intersect(IandJIndexSet, TCI.P).size() != 0)
1366 return false;

5

I think that that it is redundant to require that bands are marked as permutable, since we check the form of dependencies and memory accesses. I propose to remove such checks for pattern matching optimizations.

polly/lib/Transform/MatmulOptimizer.cpp

1093

I think rangeIslSize can’t be used in this case. However, I’ve tried to use rangeIslSize to improve the patch.

1113

Other parts of TCInfoTy are verified in isTCOperandAcc too. I think that it would be better to verify the related information in one place as much and as early as possible.

Probably, the earlier fail would simplify the debugging, since we exactly know the form of memory accesses and can rely on it. Additionally, the performance can be improved, since the earlier fail helps to avoid additional operations with sets.

1196

I am not sure whether modifications of implementations of tensor contractions, which contain read and write scalar memory accesses, are useful in practice.

Moreover, since bundles of induction variables I, J, P can contain an unlimited number of dimensions, we possibly cannot follow the algorithm from the containsOnlyMatrMultAcc function, which permutes dimensions and checks that additional memory accesses have stride 0 in terms of dimensions MMM.i, MMM.j, and MMM.k. Consequently, such memory accesses can be treated as scalar memory accesses. I have not come up with an effective alternative yet.

That is why I do not consider scalar memory accesses in getWriteAccess and setReadAccesses functions. Could we mark it as TODO and do it future?

1261

Thanks for the example! I have added a corresponding test case. If I am not mistaken, it requires DeLICM.

1372

Could we name it isReductionCarriedOverDim? I think, in this case, we should rename the parameter Pos to Dim to make it more readable.

1387–1389

As far as I understand, these operations are not equal.

deltas computes a set containing the differences between image elements and corresponding domain elements in the input. subtract computes a subtraction of sets.

For example, in the case of the following sets they compute the following:

BoundDeltas : {Stmt_for_body15[31, 31, 31, 31, 31, 31] }
isl::manage(isl_set_neg(DepDelta.copy())): {Stmt_for_body15[0, 0, 0, 0, 0, -1]}

BoundDeltas.subtract(isl::manage(isl_set_neg(DepDelta.copy()))) : {Stmt_for_body15[31, 31, 31, 31, 31, 31]}
deltas: {Stmt_for_body15[31, 31, 31, 31, 31, 32]}

BoundDeltas : {Stmt_for_body15[31, 31, 31, 31, 31, 31]}
isl::manage(isl_set_neg(DepDelta.copy())): {Stmt_for_body15[0, 0, 0, -1, 0, 31]}

BoundDeltas.subtract(isl::manage(isl_set_neg(DepDelta.copy()))) : {Stmt_for_body15[31, 31, 31, 31, 31, 31]}
deltas: {Stmt_for_body15[31, 31, 31, 32, 31, 0]}

These comment interferes with the comment about pw_multi_aff. Consequently, I replaced the usage of isl_map_deltas with operations on pw_multi_aff.

1428–1431

As far as I understand, that is not necessary, because subsequently we check that the statement has the form C(shuffle(I, J)) = E(A(shuffle(I, P)),B(shuffle(P, J))C(shuffle(I, J))), where E is an expression that contains reads from the tensors A, B, C, and an arbitrary number of reads from constants with respect to bundles I, J , and P.

I have added a comment that describes this.

"The form of anti and output dependencies is determined specified by the form the SCoP statement, which is checked by subsequent analysis."

1437–1438

What is the maximal amount of computational steps we should use by default? I set it to 500000 according to DependenceInfo.cpp.

1442–1447

Could you clarify what do you mean by the band i? Are these indexes ki, which describe the dependencies?

isTcDep checks that the dependency has the form

/// S(..., ki, max(k(i + 1)), ..., max(kn), ...) -> S(..., ki + 1, min(k(i + 1)), ..., min(kn), …)

1579

Could we factor out this condition into ScheduleTreeOptimizer::isPMOptimizableBandNode, since it is common for isTCPattern and isMatrMultPattern functions? A new version of the patch shows how it could look like.

1590

Could we add a TODO comment for this?

1599

If I am not mistaken, this only checks that all band nodes, which represent the statement, are not split by filter nodes. These accepts a straightforward implementation of TC with/without delicm. For example,

domain: "{ Stmt_for_body8[i0, i1, i2] : 0 <= i0 <= 1599 and

                              0 <= i1 <= 1799 and
                              0 <= i2 <= 2199;
Stmt_for_body3[i0, i1] :      0 <= i0 <= 1599 and
                              0 <= i1 <= 1799;
Stmt_for_body3_last[i0, i1] : 0 <= i0 <= 1599 and
                              0 <= i1 <= 1799 }"

child:

sequence:
- filter: "{ Stmt_for_body3[i0, i1] }"
  child:
    schedule: "[{ Stmt_for_body3[i0, i1] -> [(i0)] }, { Stmt_for_body3[i0, i1] -> [(i1)] }]"
    permutable: 1
    coincident: [ 1, 1 ]
- filter: "{ Stmt_for_body3_last[i0, i1] }"
  child:
    schedule: "[{ Stmt_for_body3_last[i0, i1] -> [(i0)] }, { Stmt_for_body3_last[i0, i1] -> [(i1)] }]"
    permutable: 1
    coincident: [ 1, 1 ]
- filter: "{ Stmt_for_body8[i0, i1, i2] }"
  child:
    schedule: "[{ Stmt_for_body8[i0, i1, i2] -> [(i0)] },
                { Stmt_for_body8[i0, i1, i2] -> [(i1)] },
                { Stmt_for_body8[i0, i1, i2] -> [(i2)] }]"
    permutable: 1
    coincident: [ 1, 1, 0 ]

domain: "{ Stmt2[i0, i1, i2] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and 0 <= i2 <= 31 }"
child:

schedule: "[{ Stmt2[i0, i1, i2] -> [(i0)] }, { Stmt2[i0, i1, i2] -> [(i1)] }, { Stmt2[i0, i1, i2] -> [(i2)] }]"
permutable: 1
coincident: [ 1, 1, 0 ]

Sorry, I have not committed an updated version of the optimization of TC to my github repo. However, I believe that, if this is that case, we can safely replace all such nodes.

+ auto NodeType = isl_schedule_node_get_type(Node.get());
+ while ((NodeType != isl_schedule_node_domain) &&
+ (NodeType != isl_schedule_node_filter)) {
+ assert((NodeType != isl_schedule_node_sequence) &&
+ L"Prevent the undefined behavior");
+ Node = Node.parent();
+ NodeType = isl_schedule_node_get_type(Node.get());
+ }
+ Node = Node.child(0);
+ Node = isl::manage(isl_schedule_node_cut(Node.release()));
+ return Node.insert_partial_schedule(Dimensions);

I think taht the detection of a more sophisticated implementations of TC is a possible goal of a future research.

I have described this in the comment.

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

2–4

Could we fix the existing test cases in a separate patch?

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s
; REQUIRES: asserts

polly/test/ScheduleOptimizer/pattern-matching-based-opts_16.ll

; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s

polly/test/ScheduleOptimizer/pattern-matching-based-opts_17.ll

; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s

polly/test/ScheduleOptimizer/pattern-matching-based-opts_18.ll

; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s

polly/test/ScheduleOptimizer/pattern-matching-based-opts_19.ll

; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s

polly/test/ScheduleOptimizer/pattern-matching-based-opts_20.ll

; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s

[Meinersbur]

In D114336#3500858, @gareevroman wrote:

1
Yes, it was intended. The transformation helps to optimize a class of programs, which is broader then a tensor contraction. However, it heavily depends on the codegen part. I think that the improvement of the detection can be the goal of the future work.

Please document what pattern is intended to be recognized. I don't think the doc for isTCPattern is sufficient, it only mentioned what is checked. Documenting the intended pattern would help identifying if a check has been forgotten. E.g. for the statement domain.

As far as I know, in these cases, the codegen modifies some memory accesses. Consequently, they are not correspond to the current pattern.

ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef0[i0, i2, 1 + i3] };
ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef1[1 + i3, i1, i2] };
ReadAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };
MustWriteAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };

What do you mean by "codegen modifies some memory accesses"? Polly's Codegen? What is the check to exclude this? Is the 1 + i3 memory access expression? Where does it come from?

4

Here, i occurs as indices for A, B, and C and detected as TC. Is this supported?

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {

for (int i = 0; i < 32; i++)
for (int j = 0; j < 32; j++)
for (int l = 0; l < 32; l++)
for (int w = 0; w < 32; w++)
    C[i][j] += A[i][l][w] * B[w][j][i];

}

For some reason, I cannot reproduce that. I have added a corresponding test case. As far as I understand, that should be detected because of the line 1365.

Should or should not be detected?

While debugging, it is now rejected because of this:
``

if (!TCI.B) {
  // IndexSet should be a union of J and P sets.
  if (unite(TCI.P, TCI.J) != IndexSet)
    return false;

``

Could you choose more meaningful identifiers than I, J, P and J, or use them in the pattern described in isTCPattern? I think of something like:

for (...) {
  ...
  for (...) {
    if (c)
      auto acc = C[P]; // ReadFromC
      auto a = A[Pa, I];
      auto b = B[Pb,J];
      auto arg = f(a,b);
      acc = acc op arg;
      C[P] = acc; // WriteToC
    }
  }
}

where P, I, J are sets of indices of the surrounding loops, Pa and Pb are subsets of P, I are the indices only occurring in the subscript of reading from A, J are the indices only occurring in the subscript for reading from B. There must be no indices not occuring in either P, I or J. `op=` is a commutative operation ...., c is an affine condition usually just `true`. `f` is a side-effect free operation.

(I don't whether this is correct, I want to understand whether the checked conditions are sufficient).

5

I think that that it is redundant to require that bands are marked as permutable, since we check the form of dependencies and memory accesses. I propose to remove such checks for pattern matching optimizations.

Ok.

polly/lib/Transform/MatmulOptimizer.cpp

196

@{ is not needed when documenting just a single member.

1223

The assignments should just make a copy of the array . With Dimensions being passed by-value, the caller has to make the copy which it should not need to.

SmallVector has no overload for being assigned an ArrayRef, but you could use llvm::append_range to insert all the values.

1268

Compiler warning:

/home/meinersbur/src/llvm-project/polly/lib/Transform/MatmulOptimizer.cpp:1310:13: warning: moving a temporary object prevents copy elision [-Wpessimizing-move]
    TCI.I = std::move(set_difference(IndexSet, TCI.P));

The result of set_difference is already an r-value, no need to cast it to an r-value.

1273

Same compiler warning.

1372

Sounds ok.

[gareevroman]

In D114336#3517540, @Meinersbur wrote:

In D114336#3500858, @gareevroman wrote:

1
Yes, it was intended. The transformation helps to optimize a class of programs, which is broader then a tensor contraction. However, it heavily depends on the codegen part. I think that the improvement of the detection can be the goal of the future work.

Please document what pattern is intended to be recognized. I don't think the doc for isTCPattern is sufficient, it only mentioned what is checked. Documenting the intended pattern would help identifying if a check has been forgotten. E.g. for the statement domain.

I've added a description of the TC-like kernel, which is the intended pattern, to the doc for isTCPattern function. I've added additional remarks according to your comments and restrictions of the current implementation.

As far as I know, in these cases, the codegen modifies some memory accesses. Consequently, they are not correspond to the current pattern.

ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef0[i0, i2, 1 + i3] };
ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef1[1 + i3, i1, i2] };
ReadAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };
MustWriteAccess :=	[Reduction Type: +] [Scalar: 0]
    { Stmt3[i0, i1, i2, i3] -> MemRef2[i0, i1] };

What do you mean by "codegen modifies some memory accesses"? Polly's Codegen?

Sorry, I meant ScopBuilder. In the following case

void foo(double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; ++l)
         if (l != 0)
           for (int w = 0; w < 64; ++w)
             C[i][j] += A[i][l][w] * B[w][j][l];
}

ScopBuilder generates the following memory accesses, which correspond to the pattern:

{ Stmt4[i0, i1, i2, i3] -> MemRef0[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt4[i0, i1, i2, i3] -> MemRef3[o0, o1, o2] : o0 = i3 and o1 = i1 and o2 = i2 }
{ Stmt4[i0, i1, i2, i3] -> MemRef2[o0, o1, o2] : o0 = i0 and o1 = i2 and o2 = i3 }
{ Stmt4[i0, i1, i2, i3] -> MemRef0[o0, o1] : o0 = i0 and o1 = i1 }

If we changes that code a bit

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; l++)
        for (int w = 0; w < 64; ++w)
           if (w != 0)
             C[i][j] += A[i][l][w] * B[w][j][l];
}

ScopBuilder generates the following memory accesses:

{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef1[o0, o1, o2] : o0 = 1 + i3 and o1 = i1 and o2 = i2 }
{ Stmt3[i0, i1, i2, i3] -> MemRef0[o0, o1, o2] : o0 = i0 and o1 = i2 and o2 = 1 + i3 }

In the context of the previous discussion, I meant that memory accesses are modified in comparison to the previous considered case.

What is the check to exclude this?

They will be rejected by isTCOperandAcc. If we fix the output dimensions, values of output dimensions will not form a permutation of a subset of values of input dimensions. Please see comments inside this function.

Is the 1 + i3 memory access expression?

Yes, I believe so.

Where does it come from?

If we look at the domain for the i3 variable, we see that the value 0 from the domain of w-loop is excluded and the loop bounds are modified to start from 0. Memory accesses correspond to this.

domain: "{ Stmt3[i0, i1, i2, i3] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and 0 <= i2 <= 63 and 0 <= i3 <= 62 }"

4

Here, i occurs as indices for A, B, and C and detected as TC. Is this supported?

void foo(double C[64][64], double A[64][64][64], double B[64][64][64]) {

for (int i = 0; i < 32; i++)
for (int j = 0; j < 32; j++)
for (int l = 0; l < 32; l++)
for (int w = 0; w < 32; w++)
    C[i][j] += A[i][l][w] * B[w][j][i];

}

For some reason, I cannot reproduce that. I have added a corresponding test case. As far as I understand, that should be detected because of the line 1365.

Should or should not be detected?

That should not be detected, because the intersection of free and contracted indices should always be empty. We check this at the line "if (intersect(IandJIndexSet, TCI.P).size() != 0)".

While debugging, it is now rejected because of this:
``

if (!TCI.B) {
  // IndexSet should be a union of J and P sets.
  if (unite(TCI.P, TCI.J) != IndexSet)
    return false;

``

You’re right. Thank you. "if (intersect(IandJIndexSet, TCI.P).size() != 0)" doesn’t help in this case. In that example, we have dependencies of the form:

{ Stmt3[i0, i1, i2, i3] -> Stmt3[o0, o1, o2, o3] : (o0 = i0 and o1 = i1 and o2 = i2 and o3 = 1 + i3 and i0 >= 0 and i0 <= 31 and i1 >= 0 and i1 <= 31 and i2 >= 0 and i2 <= 31 and i3 >= 0 and i3 <= 30) or (i3 = 31 and o0 = i0 and o1 = i1 and o2 = 1 + i2 and o3 = 0 and i0 >= 0 and i0 <= 31 and i1 >= 0 and i1 <= 31 and i2 >= 0 and i2 <= 30) }

the isl ast has the form:

{ domain: "{ Stmt3[i0, i1, i2, i3] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and 0 <= i2 <= 31 and 0 <= i3 <= 31 }", child: { mark: "Loop with Metadata", child: { schedule: "[{ Stmt3[i0, i1, i2, i3] -> [(i0)] }]", child: { mark: "Loop with Metadata", child: { schedule: "[{ Stmt3[i0, i1, i2, i3] -> [(i1)] }]", child: { mark: "Loop with Metadata", child: { schedule: "[{ Stmt3[i0, i1, i2, i3] -> [(i2)] }]", child: { mark: "Loop with Metadata", child: { schedule: "[{ Stmt3[i0, i1, i2, i3] -> [(i3)] }]" } } } } } } } } }
if (1 && (&MemRef5[31][31][32] <= &MemRef0[0][0] || &MemRef0[31][32] <= &MemRef5[0][0][0]) && (&MemRef4[31][31][32] <= &MemRef0[0][0] || &MemRef0[31][32] <= &MemRef4[0][0][0]))

    // Loop with Metadata
    for (int c0 = 0; c0 <= 31; c0 += 1) {
      // Loop with Metadata
      for (int c1 = 0; c1 <= 31; c1 += 1) {
        // Loop with Metadata
        for (int c2 = 0; c2 <= 31; c2 += 1) {
          // Loop with Metadata
          for (int c3 = 0; c3 <= 31; c3 += 1)
            Stmt3(c0, c1, c2, c3);
        }
      }
    }

else
    {  /* original code */ }

Consequently, only "l" and "w" are treated as "contracted indices", which are stored in TCI.P. Sorry, I missed that.

If indexes of an operand of the tensor contraction don’t contain TCI.P, we don't accept the program. We check this in lines

…
if (!isSuperset(IndexSet, TCI.P))
      return false;
…

…
if (unite(TCI.P, TCI.J) != IndexSet)
      return false;
…

unite(TCI.P, TCI.J) still isn't equal to IndexSet in the considered case, because IndexSet doesn't contain the "l".

If we didn't have the "l-loop", unite(TCI.P, TCI.J) still would not be equal to IndexSet, which would contain i, j and w, because TCI.I would contain only i and TCI.J would contain only j.

test/ScheduleOptimizer/pattern-matching-based-opts_21.ll is the corresponding test case. Additionally, I've added test/ScheduleOptimizer/pattern-matching-based-opts_25.ll.

P.S.: I had to use the -fno-unroll-loops option of clang. Otherwise, one of the loops is optimized out on my machine.

Could you choose more meaningful identifiers than I, J, P and J, or use them in the pattern described in isTCPattern? I think of something like:

I’ve added a description of bundles I, J, and P to the description of the pattern. I hope it clarifies their purpose. I propose to apply the terminology, which is used in the paper [1] and it predecessors (e.g., [2]), to simplify the understanding of the code for their readers.

[1] - Gareev R., Grosser T., Kruse M. High-Performance Generalized Tensor Operations: A Compiler-Oriented Approach ACM Transactions Architecture and Code Optimization (TACO). 2018. Vol. 15, no. 3. P. 34:1–34:27. DOI: 10.1145/3235029.
[2] - Matthews D. High-Performance Tensor Contraction without BLAS SIAM Journal on Scientific Computing. 2018. Vol. 40, no. 1. P. C 1—C 24. DOI: 110.1137/16m108968x.

for (...) {
  ...
  for (...) {
    if (c)
      auto acc = C[P]; // ReadFromC
      auto a = A[Pa, I];
      auto b = B[Pb,J];
      auto arg = f(a,b);
      acc = acc op arg;
      C[P] = acc; // WriteToC
    }
  }
}

where P, I, J are sets of indices of the surrounding loops, Pa and Pb are subsets of P, I are the indices only occurring in the subscript of reading from A, J are the indices only occurring in the subscript for reading from B. There must be no indices not occuring in either P, I or J.

I think the definition of the TC-like corresponds to the information about sets P, I, J. Probably, it’s redundant to introduce the terminology for subsets at this point. Could we do this in the description of the optimization, if it’d be needed?

op= is a commutative operation ....,

I think this a redundant condition. However, you can find it in the paper [1]. I believe that, if preserve the order of loops with indexes from the bundle P during the optimization, there would not be any violation.

c is an affine condition usually just true.

Could you elaborate on that? I’m not sure that I understand where such a condition is used.

f is a side-effect free operation.

If I’m not mistaken, according to, for example, ScopDetection.cpp, only side effect free functions calls can be located inside a Scop.

(I don't whether this is correct, I want to understand whether the checked conditions are sufficient).

> 5
> 
> I think that that it is redundant to require that bands are marked as permutable, since we check the form of dependencies and memory accesses. I propose to remove such checks for pattern matching optimizations.

Ok.

polly/lib/Transform/MatmulOptimizer.cpp

1223

I think we can use llvm::replace to avoid clearing the vector and preserve the logic.

[Meinersbur]

In D114336#3576199, @gareevroman wrote:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; l++)
        for (int w = 0; w < 64; ++w)
           if (w != 0)
             C[i][j] += A[i][l][w] * B[w][j][l];
}

ScopBuilder generates the following memory accesses:

{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef1[o0, o1, o2] : o0 = 1 + i3 and o1 = i1 and o2 = i2 }
{ Stmt3[i0, i1, i2, i3] -> MemRef0[o0, o1, o2] : o0 = i0 and o1 = i2 and o2 = 1 + i3 }

In the context of the previous discussion, I meant that memory accesses are modified in comparison to the previous considered case.

Where does it come from?

If we look at the domain for the i3 variable, we see that the value 0 from the domain of w-loop is excluded and the loop bounds are modified to start from 0. Memory accesses correspond to this.

Looks like some other optimization (maybe JumpThreading?) modifies the loop range. Ideally, the detection would be robust enough to not depend on the whether the domain space has an offset.

Could you choose more meaningful identifiers than I, J, P and J, or use them in the pattern described in isTCPattern? I think of something like:

I’ve added a description of bundles I, J, and P to the description of the pattern. I hope it clarifies their purpose. I propose to apply the terminology, which is used in the paper [1] and it predecessors (e.g., [2]), to simplify the understanding of the code for their readers.

In a paper the names must be shorter to fit on the page and a described close to the figure. It also should not be necessary to have access to the paper to understand the algorithm. For narrowly scoped variables such as loop counters single letters might be ok because the definition is likely on the same screen (or for a paper: the same page), but globals should be more identifiable. paper [1] also does not make the connection to the symbols it uses and what they correspond to in Polly's data structures.

However, I don't request such a change atm.

for (...) {
  ...
  for (...) {
    if (c)
      auto acc = C[P]; // ReadFromC
      auto a = A[Pa, I];
      auto b = B[Pb,J];
      auto arg = f(a,b);
      acc = acc op arg;
      C[P] = acc; // WriteToC
    }
  }
}

where P, I, J are sets of indices of the surrounding loops, Pa and Pb are subsets of P, I are the indices only occurring in the subscript of reading from A, J are the indices only occurring in the subscript for reading from B. There must be no indices not occuring in either P, I or J.

I think the definition of the TC-like corresponds to the information about sets P, I, J. Probably, it’s redundant to introduce the terminology for subsets at this point. Could we do this in the description of the optimization, if it’d be needed?

It is also 'redundant' with the code itself, but it helps understanding it.

The description of I and P are "Input dimensions of the schedule space, which represent free indices of tensors." One has to know what "free indices" are, indices of what, etc.

The "definition" of isTCPattern is declarative, does not even explain what all those symbols are, and refers to our matmul paper which does not apply because it only is a special case of the TC pattern.

op= is a commutative operation ....,

I think this a redundant condition. However, you can find it in the paper [1]. I believe that, if preserve the order of loops with indexes from the bundle P during the optimization, there would not be any violation.

This is exactly the sort of thin I would want to clarify.

c is an affine condition usually just true.

Could you elaborate on that? I’m not sure that I understand where such a condition is used.

It corresponds to a filter node in the TC body. You have used the if (w != 0) to illustrate where the access function deviates from the usually pattern which implied that it would be something you would like to support. If not, that would not be part of the pattern.

f is a side-effect free operation.

If I’m not mistaken, according to, for example, ScopDetection.cpp, only side effect free functions calls can be located inside a Scop.

f is not necessarily a function call, but as mentioned a "operation" representing the calculation done in the the TC body.

Side-effect here means something different than ScopDetection. A write to an unrelated array D would be a side-effect for the TC, but accurately represented by a Scop.
We could allow unknown side-effects in polly in the future with a general "memory" dependency, an extension to what Polly already does with -polly-allow-modref-calls. These could just not be reordered relative to each other.

I really think the documentation should be better. I had a hard time fixing bugs in the matmul optimization just with understanding what the code is supposed to be doing after a long time and would prefer to no repeat that again. See rGcad9f98a2ad98fecf663e9ce39502b8e43676fc9 and rGa56bd7dec8da4348d847d53c96d8a30f4a821d36.

polly/lib/Support/ISLTools.cpp

266 ↗

(On Diff #436206)

nice

polly/lib/Transform/MatmulOptimizer.cpp

1196

The concern is that I can modify what isLatestArrayKind() returns by simply importing a JScop. The continue just ignores such weirdness but I think it is safer to fail in this case.

You yourself mention that scalar accesses are likely not useful, so why not fail when one is found instead (return null instead of continue)? Some exceptions may be possible, such as read-only scalars (VirtualUse::ReadOnly, VirtualUse::Synthesizable)

1442–1447

There is a check Intersection.is_empty() which is going to detect if a dependency is completely missing. But what detects that only some of the dependencies are present. Such as:

[p] -> { Stmt3[i0, i1, i2, i3] -> Stmt3[o0, o1, o2, o3] : .... and p != 0 }

or

{ Stmt3[i0, i1, i2, i3] -> Stmt3[o0, o1, o2, o3] : .... and i0 < 42 }

(assuming contracting over i=0)

isReductionCarriedOverDim doesn't seem to check whether the dependency is over the complete domain either.

1590

Yes, that would be great.

1596

1599

I think some info in the comment like "all surrounding band nodes are assumed to be part of the TC and must not be interleaved by filter nodes."

Since it is not checking for it, it seems to imply that all other nodes types are OK? (sequence, set, expansion, extension, marker). Maybe reject them too? (I think ignoring marker nodes might still be ok)

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

2–4

👍

[gareevroman]

In D114336#3621094, @Meinersbur wrote:

In D114336#3576199, @gareevroman wrote:

void foo(int n, double C[1024][1024], double A[1024][64][64], double B[64][1024][64]) {
for (int i = 0; i < 1024; i++)
    for (int j = 0; j < 1024; j++)
      for (int l = 0; l < 64; l++)
        for (int w = 0; w < 64; ++w)
           if (w != 0)
             C[i][j] += A[i][l][w] * B[w][j][l];
}

ScopBuilder generates the following memory accesses:

{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef2[o0, o1] : o0 = i0 and o1 = i1 }
{ Stmt3[i0, i1, i2, i3] -> MemRef1[o0, o1, o2] : o0 = 1 + i3 and o1 = i1 and o2 = i2 }
{ Stmt3[i0, i1, i2, i3] -> MemRef0[o0, o1, o2] : o0 = i0 and o1 = i2 and o2 = 1 + i3 }

In the context of the previous discussion, I meant that memory accesses are modified in comparison to the previous considered case.

Where does it come from?

If we look at the domain for the i3 variable, we see that the value 0 from the domain of w-loop is excluded and the loop bounds are modified to start from 0. Memory accesses correspond to this.

Looks like some other optimization (maybe JumpThreading?) modifies the loop range. Ideally, the detection would be robust enough to not depend on the whether the domain space has an offset.

I agree. Improving the detection is a possible goal of a future work.

Could you choose more meaningful identifiers than I, J, P and J, or use them in the pattern described in isTCPattern? I think of something like:

I’ve added a description of bundles I, J, and P to the description of the pattern. I hope it clarifies their purpose. I propose to apply the terminology, which is used in the paper [1] and it predecessors (e.g., [2]), to simplify the understanding of the code for their readers.

In a paper the names must be shorter to fit on the page and a described close to the figure. It also should not be necessary to have access to the paper to understand the algorithm. For narrowly scoped variables such as loop counters single letters might be ok because the definition is likely on the same screen (or for a paper: the same page), but globals should be more identifiable. paper [1] also does not make the connection to the symbols it uses and what they correspond to in Polly's data structures.

However, I don't request such a change atm.

Ok. I've tried to make the description of the algorithm self-consistent.

for (...) {
  ...
  for (...) {
    if (c)
      auto acc = C[P]; // ReadFromC
      auto a = A[Pa, I];
      auto b = B[Pb,J];
      auto arg = f(a,b);
      acc = acc op arg;
      C[P] = acc; // WriteToC
    }
  }
}

where P, I, J are sets of indices of the surrounding loops, Pa and Pb are subsets of P, I are the indices only occurring in the subscript of reading from A, J are the indices only occurring in the subscript for reading from B. There must be no indices not occuring in either P, I or J.

I think the definition of the TC-like corresponds to the information about sets P, I, J. Probably, it’s redundant to introduce the terminology for subsets at this point. Could we do this in the description of the optimization, if it’d be needed?

It is also 'redundant' with the code itself, but it helps understanding it.

The description of I and P are "Input dimensions of the schedule space, which represent free indices of tensors." One has to know what "free indices" are, indices of what, etc.

The "definition" of isTCPattern is declarative, does not even explain what all those symbols are, and refers to our matmul paper which does not apply because it only is a special case of the TC pattern.

I've tried to improve the description.

op= is a commutative operation ....,

I think this a redundant condition. However, you can find it in the paper [1]. I believe that, if preserve the order of loops with indexes from the bundle P during the optimization, there would not be any violation.

This is exactly the sort of thin I would want to clarify.

We don't check for associativity, because it's difficult and not necessary for the optimization. The optimization doesn't change the order of loops with indexes from the bundle P during the optimization, even if you parallize the outermost loop. Hence, it doesn't violate anything. For the same reason, we don't check for associativity in the case of the optimization of the generalization of matrix-matrix multiplication, which is currently used in Polly.

c is an affine condition usually just true.

Could you elaborate on that? I’m not sure that I understand where such a condition is used.

It corresponds to a filter node in the TC body. You have used the if (w != 0) to illustrate where the access function deviates from the usually pattern which implied that it would be something you would like to support. If not, that would not be part of the pattern.

Ok. Unfortunately, the current approach does't support this. So, it's not the part of the pattern.

f is a side-effect free operation.

If I’m not mistaken, according to, for example, ScopDetection.cpp, only side effect free functions calls can be located inside a Scop.

f is not necessarily a function call, but as mentioned a "operation" representing the calculation done in the the TC body.

Side-effect here means something different than ScopDetection. A write to an unrelated array D would be a side-effect for the TC, but accurately represented by a Scop.
We could allow unknown side-effects in polly in the future with a general "memory" dependency, an extension to what Polly already does with -polly-allow-modref-calls. These could just not be reordered relative to each other.

I agree that only side-effect free operations are considered in the pattern. Nevertheless, I propose not to use terms that may require an additional specification. I've tried to improve the description of the pattern.

I really think the documentation should be better. I had a hard time fixing bugs in the matmul optimization just with understanding what the code is supposed to be doing after a long time and would prefer to no repeat that again. See rGcad9f98a2ad98fecf663e9ce39502b8e43676fc9 and rGa56bd7dec8da4348d847d53c96d8a30f4a821d36.

Sure. Let's continue improving it.

[gareevroman]

polly/lib/Transform/MatmulOptimizer.cpp

1196

I've added such a return statement to avoid scalar write memory accesses. Sorry, I was wrong. We need scalar read memory accesses. For example, in the case of the following matrix-matrix multiplication, a SCoP statement, which represents the body of the loop, contains the constant alpha.

C = alpha*A*B

Could we accept non-partial scalar read memory accesses? I think this is legal.

1442–1447

Are dependencies determined by the form of memory accesses? In the isCorrectAccessMap function, we check that memory accesses aren't partial. Isn't it sufficient?

I've tried to check whether the dependency is over the complete domain though.

1590

Ok. I've left that TODO comment.

1596

Looks like I missed that. Sorry. I will fix it in the next patch.

1599

I think some info in the comment like "all surrounding band nodes are assumed to be part of the TC and must not be interleaved by filter nodes."

I've added it it.

Since it is not checking for it, it seems to imply that all other nodes types are OK? (sequence, set, expansion, extension, marker). Maybe reject them too? (I think ignoring marker nodes might still be ok)

Sequence nodes could be necessary, if DeLICM was applied. Please, see the example inside the isTCPattern. Yes, I think other types except for marker nodes should be rejected. Additionally, as a precaution, I propose to check that a filter node has only a sequence and a domain nodes as its predecessors. I've updated the patch.

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

2–4

Ok. I've added this to my TODO list.

[Meinersbur]

Thank you Gareev. I think the description can still be improved, I but we should also move forward and can improve iteratively.

Looking forward for the actual TC optimization.

polly/lib/Transform/MatmulOptimizer.cpp

185

AFAIU multiplication by β is not part of this detection, but required to be loop-distributed by the isl scheduler.

1118

1261

It does not require DeLICM, but -polly-allow-nonaffine-branches (which is enabled by default)

1591

[typo]

1672

What is Goto here? GotoBLAS?

[gareevroman]

In D114336#3694323, @Meinersbur wrote:

Thank you Gareev. I think the description can still be improved, I but we should also move forward and can improve iteratively.

Looking forward for the actual TC optimization.

Thanks! I've tried to address new comments in the committed patch.

polly/lib/Transform/MatmulOptimizer.cpp

185

Yes, it's not. I've added a comment about this.

1261

If I'm not mistaken, in your example the form of the dependencies doesn't correspond to the pattern.

c = C[i][j];
if (/*non-affine condition*/) {
  A[i][k] + B[k][j];
} else {
  C[i][j] = c;
}

MayWrite: { Stmt_for_body8TOfor_inc[i0, i1, i2] -> MemRef_C[i0, i1] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and 0 <= i2 <= 31 }

I've added a slightly modified version of it to polly/test/ScheduleOptimizer/pattern-matching-based-opts_23.ll. It produces a region statement too.

for (int i = 0; i < 32; i++)
  for (int j = 0; j < 32; j++)
    for (int k = 0; k < 32; k++) {
      int c = C[i][j];
      if (i*j*k < 10) {
        C[i][j] = A[i][k] + B[k][j];
      } else {
        C[i][j] = c;
      } 
}

However, it introduces store merge phi nodes. It makes DeLICM necessary.

Statements {
	Stmt_for_body8__TO__if_end
        Domain :=
            { Stmt_for_body8__TO__if_end[i0, i1, i2] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and 0 <= i2 <= 31 };
        Schedule :=
            { Stmt_for_body8__TO__if_end[i0, i1, i2] -> [i0, i1, i2, 0] };
        ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
            { Stmt_for_body8__TO__if_end[i0, i1, i2] -> MemRef_A[i0, i2] };
        ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
            { Stmt_for_body8__TO__if_end[i0, i1, i2] -> MemRef_B[i2, i1] };
        ReadAccess :=	[Reduction Type: NONE] [Scalar: 0]
            { Stmt_for_body8__TO__if_end[i0, i1, i2] -> MemRef_C[i0, i1] };
        MustWriteAccess :=	[Reduction Type: NONE] [Scalar: 1]
            { Stmt_for_body8__TO__if_end[i0, i1, i2] -> MemRef_storemerge__phi[] };
       new: { Stmt_for_body8__TO__if_end[i0, i1, i2] -> MemRef_C[i0, i1] };
	Stmt_if_end
        Domain :=
            { Stmt_if_end[i0, i1, i2] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and 0 <= i2 <= 31 };
        Schedule :=
            { Stmt_if_end[i0, i1, i2] -> [i0, i1, i2, 1] };
        ReadAccess :=	[Reduction Type: NONE] [Scalar: 1]
            { Stmt_if_end[i0, i1, i2] -> MemRef_storemerge__phi[] };
       new: { Stmt_if_end[i0, i1, i2] -> MemRef_C[i0, i1] };
        MustWriteAccess :=	[Reduction Type: NONE] [Scalar: 0]
            { Stmt_if_end[i0, i1, i2] -> MemRef_C[i0, i1] };

[Meinersbur]

polly/lib/Transform/MatmulOptimizer.cpp

1261

The storemerge PHI node is introduced by one of the canonicalization passes (in this case: InstCombine). It it possible to not run that pass, disable the matching of this particular pattern, or use some other trick to not be matched. In any case, we cannot rely on the InstCombine to happen. It might be safer to bail out if any RegionStmt is encountered.

[gareevroman]

polly/lib/Transform/MatmulOptimizer.cpp

1261

I see. I haven't managed to fix that test case. So, I've decided to remove it. I've left the check that bails out if any RegionStmt is encountered in containsOnlyTCAcc.

polly/lib/Transform/MatmulOptimizer.cpp

 //===- MatmulOptimizer.cpp -----------------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "polly/MatmulOptimizer.h"
 #include "polly/DependenceInfo.h"
 #include "polly/Options.h"
 #include "polly/ScheduleTreeTransform.h"
 #include "polly/ScopInfo.h"
 #include "polly/ScopPass.h"
 #include "polly/Simplify.h"
+#include "polly/Support/GICHelper.h"
 #include "polly/Support/ISLTools.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/Sequence.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/iterator_range.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
static cl::opt<int> VectorRegisterBitwidth(
     cl::Hidden, cl::init(-1), cl::ZeroOrMore, cl::cat(PollyCategory));
 
 static cl::opt<int> PollyPatternMatchingNcQuotient(
     "polly-pattern-matching-nc-quotient",
     cl::desc("Quotient that is obtained by dividing Nc, the parameter of the"
              "macro-kernel, by Nr, the parameter of the micro-kernel"),
     cl::Hidden, cl::init(256), cl::ZeroOrMore, cl::cat(PollyCategory));
 
+static cl::opt<bool>
+    PMTCBasedOpts("polly-pattern-matching-based-tc-opts",
+                  cl::desc("Perform optimizations of tensor contractions based "
+                           "on pattern matching"),
+                  cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
+
 namespace {
 /// Parameters of the micro kernel.
 ///
 /// Parameters, which determine sizes of rank-1 (i.e., outer product) update
 /// used in the optimized matrix multiplication.
 struct MicroKernelParamsTy {
   int Mr;
   int Nr;
struct MatMulInfoTy {
   MemoryAccess *B = nullptr;
   MemoryAccess *ReadFromC = nullptr;
   MemoryAccess *WriteToC = nullptr;
   int i = -1;
   int j = -1;
   int k = -1;
 };
 
+/// Parameters of the tensor contraction operands.
+///
+/// Parameters, which describe access relations that represent operands of the
+/// tensor contraction.
+struct TCInfoTy {
+  MemoryAccess *A = nullptr;
+  MemoryAccess *B = nullptr;
+  MemoryAccess *ReadFromC = nullptr;
+  MemoryAccess *WriteToC = nullptr;
+  std::set<int> I;
+  std::set<int> J;
+  std::set<int> P;
+  SmallVector<int, 30> DimensionSizes;
+  SmallVector<int, 30> ADimensions;
+  SmallVector<int, 30> BDimensions;
+  SmallVector<int, 30> CDimensions;
+  SmallVector<int, 30> OrderedI;
+  SmallVector<int, 30> OrderedJ;
+  SmallVector<int, 30> OrderedP;
+};
+
 /// Create an isl::union_set, which describes the option of the form
 /// [isolate[] -> unroll[x]].
 ///
 /// @param Ctx An isl::ctx, which is used to create the isl::union_set.
 static isl::union_set getUnrollIsolatedSetOptions(isl::ctx Ctx) {
   isl::space Space = isl::space(Ctx, 0, 0, 1);
   isl::map UnrollIsolatedSetOption = isl::map::universe(Space);
   isl::id DimInId = isl::id::alloc(Ctx, "isolate", nullptr);
   isl::id DimOutId = isl::id::alloc(Ctx, "unroll", nullptr);
   UnrollIsolatedSetOption =
       UnrollIsolatedSetOption.set_tuple_id(isl::dim::in, DimInId);
   UnrollIsolatedSetOption =
       UnrollIsolatedSetOption.set_tuple_id(isl::dim::out, DimOutId);
   return UnrollIsolatedSetOption.wrap();
if (LeafType != isl_schedule_node_leaf ||
       isl_union_map_n_map(PartialSchedule.get()) != 1)
     return false;
   auto NewPartialSchedule = isl::map::from_union_map(PartialSchedule);
   if (containsMatrMult(NewPartialSchedule, D, MMI))
     return true;
   return false;
 }
 
+/// Get the dimension size.
+///
+/// Return the size of the dimension @p Pos, which is obtained from @p SAI.
+/// Return -1 in the case of first dimension of a multi-dimensional array, since
+/// the ScopArrayInfo class does not carry any size information.
+///
+/// @param SAI The information about the array.
+/// @param Pos The position of the dimension.
+/// @return The size of the dimension.
+static int getDimSize(const ScopArrayInfo *SAI, unsigned Pos) {
+  if (Pos == 0)
+    return -1;
+  const llvm::SCEV *SCEVDimSize = SAI->getDimensionSize(Pos);
+  assert(SCEVDimSize && L"Prevent the undefined behavior");
+  auto *ConstantDimSize = dyn_cast<const SCEVConstant>(SCEVDimSize);
+  assert(ConstantDimSize && L"Prevent the undefined behavior");
+  auto *IntDimSize = dyn_cast<ConstantInt>(ConstantDimSize->getValue());
+  assert(IntDimSize && L"Prevent the undefined behavior");
+  return IntDimSize->getSExtValue();
+}
+
+/// Check whether the access relation has the specified form.
+///
+/// Check that the access relation @p AccMap has the form T[i1, …, in], where
+/// indexes i1, …, in are specified by @p Dimensions.
+///
+/// @param Domain     The domain of the access relation.
+/// @param AccMap     The access relation to be checked.
+/// @param Dimensions The permutation of the subset of the input dimensions.
+/// @return True if @p AccMap has the expected form and false,
+///         otherwise.
+static bool isCorrectAccessMap(isl::set Domain, isl::map AccMap,
+                               const SmallVector<int, 30> &Dimensions) {
+  isl::space Space = AccMap.get_space();
+  if (unsignedFromIslSize(Space.dim(isl::dim::out)) != Dimensions.size())
+    return false;
+  isl::map Universe = isl::map::universe(Space);
+  isl::map PossibleTensor = isl::manage(Universe.copy());
+  for (int i = 0; i < static_cast<int>(Dimensions.size()); i++) {
+    const int InPos = Dimensions[i];
+    if (InPos == -1)
+      return false;
+    PossibleTensor =
+        PossibleTensor.equate(isl::dim::in, InPos, isl::dim::out, i);
+  }
+  AccMap = AccMap.intersect_domain(Domain);
+  PossibleTensor = PossibleTensor.intersect_domain(Domain);
+
+  // If AccMap spans entire domain (Non-partial write),
+  // compute FirstPos and SecondPos.
+  // If AccMap != PossibleTensor here (the two maps have been gisted at
+  // this point), it means that the writes are not complete, or in other
+  // words, it is a Partial write and Partial writes must be rejected.
+  return AccMap.is_equal(PossibleTensor);
+}
+
+/// Check whether the access can represent the tensor contraction operand.
+///
+/// Check that the access relation @p AccMap has the form T[i1, …, in].
+/// Obtained indexes i1, …, in, their sizes and permutation are stored into
+/// @p  IndexSet, @p DimensionSizes, and @p Dimensions, respectively.
+///
+/// @param Domain         The domain of the access relation.
+/// @param AccMap         The access relation to be checked.
+/// @param IndexSet       The subset of the input dimensions.
+/// @param DimensionSizes Sizes of the input dimensions of @p Dimensions.
+/// @param Dimensions     The permutation of the subset of the input dimensions.
+/// @return True if @p AccMap has the expected form and false,
+///         otherwise.
+static bool isTCOperandAcc(isl::set Domain, isl::map AccMap,
+                           std::set<int> &IndexSet,
+                           SmallVector<int, 30> &DimensionSizes,
+                           SmallVector<int, 30> &Dimensions) {
+  isl::id Id = AccMap.get_tuple_id(isl::dim::out);
+  const ScopArrayInfo *SAI = ScopArrayInfo::getFromId(Id);
+  assert(SAI && "AccMap should represent memory access");
+  isl::map CheckMap = isl::manage(AccMap.copy());
+  unsigned OutDimNum = unsignedFromIslSize(CheckMap.dim(isl::dim::out));
+  for (unsigned i = 0; i < OutDimNum; i++)
+    CheckMap = CheckMap.fix_si(isl::dim::out, i, i);
+  Dimensions.assign(OutDimNum, -1);
+  unsigned InDimNum = unsignedFromIslSize(CheckMap.dim(isl::dim::in));
+  for (unsigned i = 0; i < InDimNum; i++) {
+    isl::val Val = isl::manage(
+        isl_map_plain_get_val_if_fixed(CheckMap.get(), isl_dim_in, i));
+    if (!Val.is_int())
+      continue;
+    int OutPos = -1;
+    llvm::APInt ValAPInt = APIntFromVal(Val);
+    if (ValAPInt.isSignedIntN(32))
+      OutPos = ValAPInt.getSExtValue();
+    assert((OutPos != -1) && (OutPos < static_cast<int>(OutDimNum)) &&
+           L"Prevent the undefined behavior");
+    if (IndexSet.count(i))
+      return false;
+    IndexSet.insert(i);
+    Dimensions[OutPos] = i;
+    if (DimensionSizes[i] <= 0)
+      DimensionSizes[i] = getDimSize(SAI, OutPos);
+  }
+  return isCorrectAccessMap(Domain, AccMap, Dimensions);
+}
+
+/// Find the intersection of two sets.
+///
+/// Find the intersection of the set @p A and the set @p B.
+///
+/// @param A, B Sets to intersect.
+/// @return The set intersection.
+std::set<int> intersect(const std::set<int> &A, const std::set<int> &B) {
+  std::set<int> Intersection;
+  set_intersection(A.begin(), A.end(), B.begin(), B.end(),
+                   std::inserter(Intersection, Intersection.begin()));
+  return Intersection;
+}
+
+/// Find the difference of two sets.
+///
+/// Find the difference of the set @p A and the set @p B.
+///
+/// @param A The first operand of set difference.
+/// @param B The second operand of set difference.
+/// @return The set difference.
+std::set<int> difference(const std::set<int> &A, const std::set<int> &B) {
+  std::set<int> Difference;
+  set_difference(A.begin(), A.end(), B.begin(), B.end(),
+                 std::inserter(Difference, Difference.begin()));
+  return Difference;
+}
+
+/// Determine the access, which writes to the tensor C.
+///
+/// @param Domain        The domain of the statement.
+/// @param Stmt          The statement, which writes to memory.
+/// @param TCI           The information about the tensor contraction.
+/// @param IandJIndexSet The set, which contains free indexes of tensors.
+void setWriteAccess(isl::set Domain, ScopStmt *Stmt, TCInfoTy &TCI,
+                    std::set<int> &IandJIndexSet) {
+  TCI.WriteToC = nullptr;
+  SmallVector<MemoryAccess *, 32> Accesses = getAccessesInOrder(*Stmt);
+  auto *MemA = Accesses.end() - 1;
+  for (; MemA != Accesses.begin(); MemA--) {
+    MemoryAccess *MemAccessPtr = *MemA;
+    if (!MemAccessPtr->isLatestArrayKind())
+      continue;
+    if (!MemAccessPtr->isWrite())
+      return;
+    isl::map AccMap = MemAccessPtr->getLatestAccessRelation();
+    if (!isTCOperandAcc(Domain, AccMap, IandJIndexSet, TCI.DimensionSizes,
+                        TCI.CDimensions))
+      return;
+    if (intersect(IandJIndexSet, TCI.P).size())
+      return;
+    TCI.WriteToC = MemAccessPtr;
+    break;
+  }
+}
+
+/// Determine an access, which reads elements of either an operand of the tensor
+/// contraction or its result
+///
+/// @param MemAccessPtr  The access, which reads elements of the tensor.
+/// @param IndexSet      The set, which contains indexes of the tensors.
+/// @param IandJIndexSet The set, which contains free indexes of tensors.
+/// @param Dimensions    The permutation of the subset of the input dimensions.
+/// @param TCI           The information about the tensor contraction.
+bool setReadAccess(MemoryAccess *MemAccessPtr, const std::set<int> &IndexSet,
+                   const std::set<int> &IandJIndexSet,
+                   SmallVector<int, 30> Dimensions, TCInfoTy &TCI) {
+  if (!TCI.A) {
+    // Probably IndexSet is a union of I and P sets
+    if (intersect(IndexSet, TCI.P).size() != TCI.P.size())
+      return false;
+    TCI.I = std::move(difference(IndexSet, TCI.P));
+    if (intersect(IandJIndexSet, TCI.I).size() != TCI.I.size())
+      return false;
+    TCI.J = std::move(difference(IandJIndexSet, TCI.I));
+    TCI.A = MemAccessPtr;
+    TCI.ADimensions = std::move(Dimensions);
+    return true;
+  }
+  if (!TCI.B) {
+    // Probably IndexSet is a union of J and P sets
+    if (!(intersect(IndexSet, TCI.J).size() == TCI.J.size()) &&
+        (intersect(IndexSet, TCI.P).size() == TCI.P.size()) &&
+        (IndexSet.size() == TCI.P.size() + TCI.J.size()))
+      return false;
+    TCI.B = MemAccessPtr;
+    TCI.BDimensions = std::move(Dimensions);
+    return true;
+  }
+  return false;
+}
+
+/// Determine accesses, which read elements of operands of the tensor
+/// contraction and its result
+///
+/// @param Domain        The domain of the statement.
+/// @param Stmt          The statement, which writes to memory.
+/// @param TCI           The information about the tensor contraction.
+/// @param IandJIndexSet The set, which contains free indexes of tensors.
+void setReadAccesses(isl::set Domain, ScopStmt *Stmt, TCInfoTy &TCI,
+                     std::set<int> &IandJIndexSet) {
+  TCI.A = nullptr;
+  TCI.B = nullptr;
+  TCI.ReadFromC = nullptr;
+  SmallVector<MemoryAccess *, 32> Accesses = getAccessesInOrder(*Stmt);
+  for (auto *MemA = Accesses.begin(); *MemA != TCI.WriteToC; MemA++) {
+    MemoryAccess *MemAccessPtr = *MemA;
+    if (!MemAccessPtr->isLatestArrayKind())
+      continue;
+    if (MemAccessPtr->isWrite())
+      return;
+    isl::map AccMap = MemAccessPtr->getLatestAccessRelation();
+    if (AccMap.is_equal(TCI.WriteToC->getLatestAccessRelation())) {
+      if (TCI.ReadFromC)
+        return;
+      TCI.ReadFromC = MemAccessPtr;
+      continue;
+    }
+    SmallVector<int, 30> Dimensions;
+    std::set<int> IndexSet;
+    if (!isTCOperandAcc(Domain, AccMap, IndexSet, TCI.DimensionSizes,
+                        Dimensions))
+      return;
+    if (!setReadAccess(MemAccessPtr, IndexSet, IandJIndexSet,
+                       std::move(Dimensions), TCI))
+      return;
+  }
+}
+
+/// Check accesses to operands of the tensor contraction.
+///
+/// Check that accesses of the SCoP statement, which corresponds to
+/// the partial schedule @p PartialSchedule, represents accesses
+/// to the non-scalar operands of the tensor contraction.
+///
+/// @param  Domain          The domain of the SCoP statement.
+/// @param  PartialSchedule The partial schedule of the SCoP statement.
+/// @param  TCI             Parameters of the tensor contraction operands.
+/// @return                 True if the corresponding SCoP statement
+///                         represents tensor contraction and false,
+///                         otherwise.
+static bool containsOnlyTCAcc(isl::set Domain, isl::map PartialSchedule,
+                              TCInfoTy &TCI) {
+  isl::id InputDimsId = PartialSchedule.get_tuple_id(isl::dim::in);
+  ScopStmt *Stmt = static_cast<ScopStmt *>(InputDimsId.get_user());
+  if (Stmt->size() <= 1)
+    return false;
+
+  unsigned DimNum = unsignedFromIslSize(PartialSchedule.dim(isl::dim::in));
+  TCI.DimensionSizes.resize(DimNum);
+  std::set<int> IandJIndexSet;
+  setWriteAccess(Domain, Stmt, TCI, IandJIndexSet);
+  if (!TCI.WriteToC)
+    return false;
+
+  if (intersect(IandJIndexSet, TCI.P).size() != 0)
+    return false;
+
+  setReadAccesses(Domain, Stmt, TCI, IandJIndexSet);
+  return TCI.ReadFromC && TCI.A && TCI.B;
+}
+
+/// Check that dependency corresponds to the tensor contraction.
+///
+/// Check that the dependency has the form
+/// S(..., ki, max(k(i + 1)), ..., max(kn), ...) ->
+/// S(..., ki + 1, min(k(i + 1)), ..., min(kn), ...), where S is the SCoP
+/// statement. For this purpose, we analyze the set @p DepDelta, which
+/// represents the differences between image elements and domain elements of
+/// the corresponding map.
+///
+/// @param  DepDelta    The set contains the differences between image elements
+///                     and corresponding domain elements of the map, which
+///                     represents the dependency.
+/// @param  Pos         The position of the index ki.
+/// @param  BoundDeltas In the case of indexes of ki, the difference between
+///                     image elements and corresponding domain elements
+///                     corresponds to the difference between lexicographic
+///                     minimum and lexicographic maximum of the corresponding
+///                     dimension of the domain of the statement.
+/// @param  IndexSet    Obtained indexes ki, which describe the dependency.
+/// @return True if dependencies correspond to the tensor contraction
+///         and false, otherwise.
+static bool isTcDep(isl::set DepDelta, unsigned Pos, isl::set BoundDeltas,
+                    std::set<int> *IndexSet) {
+  if ((unsignedFromIslSize(DepDelta.n_basic_set()) != 1) ||
+      !DepDelta.plain_get_val_if_fixed(isl::dim::set, Pos).is_one())
+    return false;
+  for (unsigned j = 0; j < Pos; j++)
+    if (!DepDelta.plain_get_val_if_fixed(isl::dim::set, j).is_zero())
+      return false;
+  isl::map DepDeltaNegToBoundDeltas = isl::map::from_domain_and_range(
+      isl::manage(isl_set_neg(DepDelta.copy())), BoundDeltas);
+  isl::set Complement = DepDeltaNegToBoundDeltas.deltas();
+  const unsigned DimNum = unsignedFromIslSize(DepDelta.dim(isl::dim::set));
+  for (unsigned j = Pos + 1; j < DimNum; j++) {
+    if (!IndexSet->count(j)) {
+      if (DepDelta.plain_get_val_if_fixed(isl::dim::set, j).is_zero())
+        continue;
+      return false;
+    }
+    if (!Complement.plain_get_val_if_fixed(isl::dim::set, j).is_zero())
+      return false;
+  }
+  return true;
+}
+
+/// Check that dependencies correspond to the tensor contraction.
+///
+/// Check that there are only true dependencies of the form
+/// S(..., ki, max(k(i + 1)), ..., max(kn), ...) ->
+/// S(..., ki + 1, min(k(i + 1)), ..., min(kn), ...), where S is the SCoP
+/// statement represented by @p Schedule. Such dependencies are produced by
+/// the tensor contraction. Obtained indexes ki  are stored into @p IndexSet.
+///
+/// @param  Schedule The schedule of the SCoP statement.
+/// @param  D        The SCoP dependencies.
+/// @param  Domain   The domain of the statement.
+/// @param  IndexSet Obtained indexes ki, which describe the dependencies.
+/// @return True if dependencies correspond to the tensor contraction
+///         and false, otherwise.
+static bool containsOnlyTcDeps(isl::map Schedule, const Dependences *D,
+                               std::set<int> *IndexSet, isl::set Domain) {
+  isl::union_map Dep = D->getDependences(Dependences::TYPE_RAW);
+  isl::union_map Red = D->getDependences(Dependences::TYPE_RED);
+  if (!Red.is_null())
+    Dep = Dep.unite(Red);
+  isl::space DomainSpace = Schedule.get_space().domain();
+  isl::space Space = DomainSpace.map_from_domain_and_range(DomainSpace);
+  isl::set DepDeltas = Dep.extract_map(Space).deltas();
+  unsigned DeltasDimNum = unsignedFromIslSize(DepDeltas.dim(isl::dim::set));
+  isl::set UpperBound = Domain.lexmax();
+  isl::set LowerBound = Domain.lexmin();
+  isl::map BoundMap = isl::map::from_domain_and_range(LowerBound, UpperBound);
+  isl::set BoundDeltas = BoundMap.deltas();
+  for (int i = static_cast<int>(DeltasDimNum) - 1; i >= 0; i--) {
+    isl::set Intersection = DepDeltas.fix_si(isl::dim::set, i, 1);
+    if (Intersection.is_empty())
+      continue;
+    if (!isTcDep(Intersection, i, BoundDeltas, IndexSet))
+      return false;
+    IndexSet->insert(i);
+    DepDeltas = DepDeltas.subtract(Intersection);
+  }
+  if ((DeltasDimNum == 0) || !DepDeltas.is_empty())
+    return false;
+  return true;
+}
+
+/// Order the dimensions of operands of the tensor contraction.
+///
+/// To improve the spatial locality of the data, we use a heuristic, which is
+/// introduced in [1]. The heuristic logically reorders the tensor dimensions
+/// and consists of the following steps:
+///
+/// 1. Sort the dimensions in I and J by increasing stride in the tensor C.
+/// 2. If tensor A contains the last dimension of C, swap A with B and I with J.
+/// 3. Sort the dimensions in P by increasing stride in the tensor A.
+///
+/// [1] - Devin Matthews. High-Performance Tensor Contraction without BLAS.
+/// SIAM Journal on Scientific Computing, 40(1). 2016. DOI: 10.1137/16M108968X.
+///
+/// @param TCI The information about tensors.
+static void orderDimensions(TCInfoTy &TCI) {
+  for (const int &Dimension : TCI.CDimensions) {
+    assert(!(TCI.I.count(Dimension) && TCI.J.count(Dimension)) &&
+           L"Prevent the undefined behavior");
+    if (TCI.I.count(Dimension)) {
+      TCI.OrderedI.push_back(Dimension);
+      continue;
+    }
+    if (TCI.J.count(Dimension))
+      TCI.OrderedJ.push_back(Dimension);
+  }
+  if (TCI.I.count(TCI.CDimensions.back())) {
+    MemoryAccess *Access = TCI.A;
+    TCI.A = TCI.B;
+    TCI.B = Access;
+    TCI.I.swap(TCI.J);
+    TCI.ADimensions.swap(TCI.BDimensions);
+    TCI.OrderedI.swap(TCI.OrderedJ);
+  }
+  for (const int &Dimension : TCI.ADimensions)
+    if (TCI.P.count(Dimension))
+      TCI.OrderedP.push_back(Dimension);
+}
+
+/// Check if the SCoP statement could probably be optimized with analytical
+/// modeling.
+///
+/// containsTC tries to determine whether the following conditions
+/// are true:
+/// 1. The last memory access modeling an array, MA1, represents writing to
+///    memory and has the form S(..., I, ..., J, ...) -> M(shuffle(I, J)),
+///    where S is the SCoP statement under consideration and shuffle(I, J)
+///    is a permutation of indexes of sets I and J.
+/// 2. There are only true dependencies of the form
+///    S(..., ki, max(k(i + 1)), ..., max(kn), ...) ->
+///    S(..., ki + 1, min(k(i + 1)), ..., min(kn), ...), where S is the SCoP
+///    statement represented by @p Schedule and ki are indexes of the set P.
+/// 3. SCoP contains only three access relations, MA2, MA3, and MA4 that
+///    represent reading from memory and have the form
+///    S(..., I, ..., P, ...) -> M(shuffle(I, P)),
+///    S(..., P, ..., J, ...) -> M(shuffle(J, P)),
+///    S(...) -> M(shuffle(I, J)), respectively.
+///
+/// @param PartialSchedule The PartialSchedule that contains a SCoP statement
+///                        to check.
+/// @param D               The SCoP dependencies.
+/// @param TCI             Parameters of the tensor contraction operands.
+/// @param Domain          The domain of the statement.
+static bool containsTCInfoTy(isl::map PartialSchedule, const Dependences *D,
+                             TCInfoTy &TCI, isl::set Domain) {
+  if (!containsOnlyTcDeps(PartialSchedule, D, &TCI.P, Domain))
+    return false;
+
+  // TODO: handle this case if needed
+  if (TCI.P.size() == 0)
+    return false;
+
+  if (!containsOnlyTCAcc(Domain, PartialSchedule, TCI))
+    return false;
+
+  // TODO: handle cases of GEMV if needed
+  if ((TCI.I.size() == 0) || (TCI.J.size() == 0))
+    return false;
+
+  orderDimensions(TCI);
+
+  return true;
+}
+
+/// Check if this node contains a partial schedule that could
+/// probably be optimized with analytical modeling.
+///
+/// isTCPattern tries to determine whether the following conditions
+/// are true:
+/// 1. the partial schedule contains only one statement.
+/// 2. there are exactly three input dimensions.
+/// 3. there are four memory accesses that represent accesses to tensor
+///    contraction operand.
+/// 4. all memory accesses of the statement except from the last one, are
+///    read memory accesses and the last one is a write memory access.
+/// 5. all subscripts of the last memory access of the statement don't
+///    contain the variable used in the inner loop.
+/// If this is the case, we could use an approach that is similar to
+/// the one used to get close-to-peak performance of matrix multiplications.
+///
+/// @param Node The node to check.
+/// @param D    The SCoP dependencies.
+/// @param TCI  Parameters of the tensor contraction operands.
+static bool isTCPattern(isl::schedule_node Node, const Dependences *D,
+                        TCInfoTy &TCI) {
+  Node = Node.child(0);
+  auto LeafType = isl_schedule_node_get_type(Node.get());
+  isl::union_map PartialSchedule = Node.get_prefix_schedule_union_map();
+  Node = Node.parent();
+  if (LeafType != isl_schedule_node_leaf ||
+      isl_union_map_n_map(PartialSchedule.get()) != 1)
+    return false;
+  isl::map PartialScheduleMap = isl::map::from_union_map(PartialSchedule);
+  auto NodeType = isl_schedule_node_get_type(Node.get());
+  while (NodeType != isl_schedule_node_domain) {
+    Node = Node.parent();
+    NodeType = isl_schedule_node_get_type(Node.get());
+  }
+  isl::union_set Domain = Node.as<isl::schedule_node_domain>().domain();
+  if (isl_union_set_n_set(Domain.get()) != 1)
+    return false;
+  if (containsTCInfoTy(PartialScheduleMap, D, TCI, isl::set(Domain)))
+    return true;
+  return false;
+}
+
 } // namespace
 
 isl::schedule_node
 polly::tryOptimizeMatMulPattern(isl::schedule_node Node,
                                 const llvm::TargetTransformInfo *TTI,
                                 const Dependences *D) {
+  TCInfoTy TCI;
+  if (PMTCBasedOpts && isTCPattern(Node, D, TCI))
+    LLVM_DEBUG(dbgs() << "The tensor contraction pattern was detected\n");
   MatMulInfoTy MMI;
   if (isMatrMultPattern(Node, D, MMI)) {
     LLVM_DEBUG(dbgs() << "The matrix multiplication pattern was detected\n");
     return optimizeMatMulPattern(Node, TTI, MMI);
   }
   return {};
 }

polly/lib/Transform/ScheduleOptimizer.cpp

if (isl_schedule_node_get_type(Node.get()) != isl_schedule_node_band)
     return false;
 
   if (isl_schedule_node_n_children(Node.get()) != 1)
     return false;
 
   if (!isl_schedule_node_band_get_permutable(Node.get()))
     return false;
 
-  auto Space = isl::manage(isl_schedule_node_band_get_space(Node.get()));
-
-  if (unsignedFromIslSize(Space.dim(isl::dim::set)) <= 1u)
-    return false;
-
   return isSimpleInnermostBand(Node);
 }
 
 __isl_give isl::schedule_node
 ScheduleTreeOptimizer::applyTileBandOpt(isl::schedule_node Node) {
   if (FirstLevelTiling) {
     Node = tileNode(Node, "1st level tiling", FirstLevelTileSizes,
                     FirstLevelDefaultTileSize);
if (OAI->PatternOpts) {
     isl::schedule_node PatternOptimizedSchedule =
         tryOptimizeMatMulPattern(Node, OAI->TTI, OAI->D);
     if (!PatternOptimizedSchedule.is_null()) {
       MatMulOpts++;
       return PatternOptimizedSchedule.release();
     }
   }
 
+  if (isl_schedule_node_band_n_member(Node.get()) <= 1)
+    return Node.release();
+
   if (OAI->Postopts)
     Node = applyTileBandOpt(Node);
 
   if (OAI->Prevect) {
     // FIXME: Prevectorization requirements are different from those checked by
     // isTileableBandNode.
     Node = applyPrevectBandOpt(Node);
   }

polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm_2.ll

+; RUN: opt %loadPolly -polly-delicm -polly-simplify -polly-opt-isl \
+; RUN: -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1 \
+; RUN: | FileCheck %s
+; REQUIRES: asserts
+;
+; Check that the pattern matching detects the tensor contraction pattern
+; after a full run of -polly-delicm. This test case generates the following
+; schedule, which contans two band nodes. Without DeLICM two statement are
+; generated.
+;
+; domain: "{ Stmt5[i0, i1, i2, i3, i4, i5] : 0 <= i0 <= 31 and 0 <= i1 <= 31 and
+;                                            0 <= i2 <= 31 and 0 <= i3 <= 31 and
+;                                            0 <= i4 <= 31 and 0 <= i5 <= 31 }"
+; child:
+;   schedule: "[{ Stmt5[i0, i1, i2, i3, i4, i5] -> [(i0)] },
+;               { Stmt5[i0, i1, i2, i3, i4, i5] -> [(i1)] },
+;               { Stmt5[i0, i1, i2, i3, i4, i5] -> [(i2)] },
+;               { Stmt5[i0, i1, i2, i3, i4, i5] -> [(i4)] },
+;               { Stmt5[i0, i1, i2, i3, i4, i5] -> [(i3)] }]"
+;   permutable: 1
+;   coincident: [ 1, 1, 1, 1, 0 ]
+;   child:
+;     schedule: "[{ Stmt5[i0, i1, i2, i3, i4, i5] -> [(i5)] }]"
+;     permutable: 1
+;     child:
+;       leaf
+;
+;   for (i = 0; i < 32; i++)
+;     for (j = 0; j < 32; j++)
+;       for (k = 0; k < 32; ++k)
+;         for (l = 0; l < 32; ++l)
+;           for (w = 0; w < 32; ++w)
+;             for (q = 0; q < 32; ++q)
+;               C[i][j][k][w] += A[i][l][j][q] * B[q][w][l][k];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal fastcc void @kernel_tc([32 x [32 x [32 x double]]]* nocapture %C, [32 x [32 x [32 x double]]]* nocapture readonly %A, [32 x [32 x [32 x double]]]* nocapture readonly %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc50, %entry
+  %indvars.iv19 = phi i64 [ 0, %entry ], [ %indvars.iv.next20, %for.inc50 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc47, %for.cond1.preheader
+  %indvars.iv16 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next17, %for.inc47 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc44, %for.cond4.preheader
+  %indvars.iv13 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next14, %for.inc44 ]
+  br label %for.cond10.preheader
+
+for.cond10.preheader:                             ; preds = %for.inc41, %for.cond7.preheader
+  %indvars.iv10 = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next11, %for.inc41 ]
+  br label %for.cond13.preheader
+
+for.cond13.preheader:                             ; preds = %for.inc38, %for.cond10.preheader
+  %indvars.iv7 = phi i64 [ 0, %for.cond10.preheader ], [ %indvars.iv.next8, %for.inc38 ]
+  %arrayidx37 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %C, i64 %indvars.iv19, i64 %indvars.iv16, i64 %indvars.iv13, i64 %indvars.iv7
+  %.pre = load double, double* %arrayidx37, align 8
+  br label %for.body15
+
+for.body15:                                       ; preds = %for.body15, %for.cond13.preheader
+  %i = phi double [ %.pre, %for.cond13.preheader ], [ %add, %for.body15 ]
+  %indvars.iv = phi i64 [ 0, %for.cond13.preheader ], [ %indvars.iv.next, %for.body15 ]
+  %arrayidx21 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %A, i64 %indvars.iv19, i64 %indvars.iv10, i64 %indvars.iv16, i64 %indvars.iv
+  %i1 = load double, double* %arrayidx21, align 8
+  %arrayidx29 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %B, i64 %indvars.iv, i64 %indvars.iv7, i64 %indvars.iv10, i64 %indvars.iv13
+  %i2 = load double, double* %arrayidx29, align 8
+  %mul = fmul fast double %i2, %i1
+  %add = fadd fast double %i, %mul
+  store double %add, double* %arrayidx37, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, 32
+  br i1 %exitcond.not, label %for.inc38, label %for.body15
+
+for.inc38:                                        ; preds = %for.body15
+  %indvars.iv.next8 = add nuw nsw i64 %indvars.iv7, 1
+  %exitcond9.not = icmp eq i64 %indvars.iv.next8, 32
+  br i1 %exitcond9.not, label %for.inc41, label %for.cond13.preheader
+
+for.inc41:                                        ; preds = %for.inc38
+  %indvars.iv.next11 = add nuw nsw i64 %indvars.iv10, 1
+  %exitcond12.not = icmp eq i64 %indvars.iv.next11, 32
+  br i1 %exitcond12.not, label %for.inc44, label %for.cond10.preheader
+
+for.inc44:                                        ; preds = %for.inc41
+  %indvars.iv.next14 = add nuw nsw i64 %indvars.iv13, 1
+  %exitcond15.not = icmp eq i64 %indvars.iv.next14, 32
+  br i1 %exitcond15.not, label %for.inc47, label %for.cond7.preheader
+
+for.inc47:                                        ; preds = %for.inc44
+  %indvars.iv.next17 = add nuw nsw i64 %indvars.iv16, 1
+  %exitcond18.not = icmp eq i64 %indvars.iv.next17, 32
+  br i1 %exitcond18.not, label %for.inc50, label %for.cond4.preheader
+
+for.inc50:                                        ; preds = %for.inc47
+  %indvars.iv.next20 = add nuw nsw i64 %indvars.iv19, 1
+  %exitcond21.not = icmp eq i64 %indvars.iv.next20, 32
+  br i1 %exitcond21.not, label %for.end52, label %for.cond1.preheader
+
+for.end52:                                        ; preds = %for.inc50
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts.ll

 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=false \
-; RUN: -debug < %s 2>&1| FileCheck %s
+; RUN: -debug -polly-pattern-matching-based-tc-opts < %s 2>&1| FileCheck %s
 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true -debug < %s 2>&1| FileCheck %s --check-prefix=PATTERN-MATCHING-OPTS
 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true -polly-ast-detect-parallel -polly-ast -analyze < %s | FileCheck %s --check-prefix=PARALLEL-AST
 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true -stats -disable-output < %s 2>&1| FileCheck %s --check-prefix=STATS -match-full-lines
 ; REQUIRES: asserts
 ;
 ;    /* C := alpha*A*B + beta*C */
 ;    for (i = 0; i < _PB_NI; i++)
 ;      for (j = 0; j < _PB_NJ; j++)
 ;        {
 ;    C[i][j] *= beta;
 ;    for (k = 0; k < _PB_NK; ++k)
 ;      C[i][j] += alpha * A[i][k] * B[k][j];
 ;        }
 ;
 ; CHECK-NOT: The matrix multiplication pattern was detected
+; CHECK-NOT: The tensor contraction pattern was detected
 ; PATTERN-MATCHING-OPTS: The matrix multiplication pattern was detected
 ; PARALLEL-AST: #pragma known-parallel
 ; PARALLEL-AST: #pragma known-parallel
 ; STATS:  1 polly-opt-isl    - Number of matrix multiplication patterns detected and optimized
 ;
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-unknown"
 

polly/test/ScheduleOptimizer/pattern-matching-based-opts_11.ll

 ; RUN: opt %loadPolly -polly-import-jscop \
 ; RUN: -polly-import-jscop-postfix=transformed \
 ; RUN: -polly-pattern-matching-based-opts=true \
 ; RUN: -polly-target-throughput-vector-fma=1 \
 ; RUN: -polly-target-latency-vector-fma=8 \
 ; RUN: -polly-target-1st-cache-level-associativity=8 \
 ; RUN: -polly-target-2nd-cache-level-associativity=8 \
 ; RUN: -polly-target-1st-cache-level-size=32768 \
 ; RUN: -polly-target-vector-register-bitwidth=256 \
 ; RUN: -polly-target-2nd-cache-level-size=262144 \
-; RUN: -polly-opt-isl -debug < %s 2>&1 \
+; RUN: -polly-opt-isl -debug \
+; RUN: -polly-pattern-matching-based-tc-opts=true < %s 2>&1 \
 ; RUN: | FileCheck %s
 ; REQUIRES: asserts
 ;
 ; Check that the pattern matching detects the matrix multiplication pattern
 ; in case scalar memory accesses were replaced by accesses to newly created
 ; arrays.
 ;
+; CHECK: The tensor contraction pattern was detected
 ; CHECK: The matrix multiplication pattern was detected
 ;
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-unknown"
 
 define void @kernel_gemm(i32 %ni, i32 %nj, i32 %nk, double %A, [1024 x double]* %B, [1024 x double]* %C) {
 entry:
   br label %entry.split

polly/test/ScheduleOptimizer/pattern-matching-based-opts_15.ll

 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
-; RUN: -debug-only=polly-opt-isl -disable-output < %s 2>&1 | FileCheck %s
+; RUN: -debug-only=polly-opt-isl -disable-output \
+; RUN: -polly-pattern-matching-based-tc-opts=true < %s 2>&1 | FileCheck %s
 ; REQUIRES: asserts
 ;
 ;    for (i = 0; i < _PB_NI; i++)
 ;      for (j = 0; j < _PB_NJ; j++)
 ;      {
 ;        for (k = 0; k < _PB_NK; k++)
 ;        {
 ;          double Mul = A[i][k] * B[k][j];
 ;          D[i][j][k] += Mul;
 ;          C[i][j] += Mul;
 ;        }
 ;      }
 ;
 ; CHECK-NOT: The matrix multiplication pattern was detected
+; CHECK-NOT: The tensor contraction pattern was detected
 
 target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 define void @kernel_gemm([1024 x double]* %C, [1024 x double]* %A, [1024 x double]* %B, [1024 x [1024 x double]]* %D) {
 entry:
   br label %for.cond1.preheader
 

polly/test/ScheduleOptimizer/pattern-matching-based-opts_16.ll

+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1| FileCheck %s
+; REQUIRES: asserts
+;
+;      for (i = 0; i < 1024; i++)
+;        for (j = 0; j < 1024; j++)
+;          for (l = 0; l < 64; ++l)
+;            for (w = 0; w < 64; ++w)
+;              C[i][j] += A[i][l][w] * B[w][j][l];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal void @kernel_tc(i32 %ni, i32 %nj, i32 %nl, i32 %nq, i32 %nw, double %alpha, double %beta, [1024 x double]* %C, [64 x [64 x double]]* %A, [1024 x [64 x double]]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc30, %entry
+  %indvars.iv43 = phi i64 [ 0, %entry ], [ %indvars.iv.next44, %for.inc30 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc27, %for.cond1.preheader
+  %indvars.iv40 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next41, %for.inc27 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc24, %for.cond4.preheader
+  %indvars.iv37 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next38, %for.inc24 ]
+  br label %for.body9
+
+for.body9:                                        ; preds = %for.body9, %for.cond7.preheader
+  %indvars.iv = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next, %for.body9 ]
+  %arrayidx13 = getelementptr inbounds [64 x [64 x double]], [64 x [64 x double]]* %A, i64 %indvars.iv43, i64 %indvars.iv37, i64 %indvars.iv
+  %i = load double, double* %arrayidx13, align 8
+  %arrayidx19 = getelementptr inbounds [1024 x [64 x double]], [1024 x [64 x double]]* %B, i64 %indvars.iv, i64 %indvars.iv40, i64 %indvars.iv37
+  %i1 = load double, double* %arrayidx19, align 8
+  %mul = fmul fast double %i1, %i
+  %arrayidx23 = getelementptr inbounds [1024 x double], [1024 x double]* %C, i64 %indvars.iv43, i64 %indvars.iv40
+  %i2 = load double, double* %arrayidx23, align 8
+  %add = fadd fast double %i2, %mul
+  store double %add, double* %arrayidx23, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, 64
+  br i1 %exitcond, label %for.body9, label %for.inc24
+
+for.inc24:                                        ; preds = %for.body9
+  %indvars.iv.next38 = add nuw nsw i64 %indvars.iv37, 1
+  %exitcond39 = icmp ne i64 %indvars.iv.next38, 64
+  br i1 %exitcond39, label %for.cond7.preheader, label %for.inc27
+
+for.inc27:                                        ; preds = %for.inc24
+  %indvars.iv.next41 = add nuw nsw i64 %indvars.iv40, 1
+  %exitcond42 = icmp ne i64 %indvars.iv.next41, 1024
+  br i1 %exitcond42, label %for.cond4.preheader, label %for.inc30
+
+for.inc30:                                        ; preds = %for.inc27
+  %indvars.iv.next44 = add nuw nsw i64 %indvars.iv43, 1
+  %exitcond45 = icmp ne i64 %indvars.iv.next44, 1024
+  br i1 %exitcond45, label %for.cond1.preheader, label %for.end32
+
+for.end32:                                        ; preds = %for.inc30
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts_17.ll

+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1| FileCheck %s
+; REQUIRES: asserts
+;
+;      for (i = 0; i < 32; i++)
+;        for (j = 0; j < 1024; j++)
+;          for (k = 0; k < 32; ++k)
+;            for (l = 0; l < 1024; ++l)
+;              C[i][j][k] += A[i][k][l] * B[l][j];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal void @kernel_tc(i32 %ni, i32 %nj, i32 %nk, i32 %nl, double %alpha, double %beta, [1024 x [32 x double]]* %C, [32 x [1024 x double]]* %A, [1024 x double]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc30, %entry
+  %indvars.iv43 = phi i64 [ 0, %entry ], [ %indvars.iv.next44, %for.inc30 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc27, %for.cond1.preheader
+  %indvars.iv40 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next41, %for.inc27 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc24, %for.cond4.preheader
+  %indvars.iv37 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next38, %for.inc24 ]
+  br label %for.body9
+
+for.body9:                                        ; preds = %for.body9, %for.cond7.preheader
+  %indvars.iv = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next, %for.body9 ]
+  %arrayidx13 = getelementptr inbounds [32 x [1024 x double]], [32 x [1024 x double]]* %A, i64 %indvars.iv43, i64 %indvars.iv37, i64 %indvars.iv
+  %i = load double, double* %arrayidx13, align 8
+  %arrayidx17 = getelementptr inbounds [1024 x double], [1024 x double]* %B, i64 %indvars.iv, i64 %indvars.iv40
+  %i1 = load double, double* %arrayidx17, align 8
+  %mul = fmul fast double %i1, %i
+  %arrayidx23 = getelementptr inbounds [1024 x [32 x double]], [1024 x [32 x double]]* %C, i64 %indvars.iv43, i64 %indvars.iv40, i64 %indvars.iv37
+  %i2 = load double, double* %arrayidx23, align 8
+  %add = fadd fast double %i2, %mul
+  store double %add, double* %arrayidx23, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, 1024
+  br i1 %exitcond, label %for.body9, label %for.inc24
+
+for.inc24:                                        ; preds = %for.body9
+  %indvars.iv.next38 = add nuw nsw i64 %indvars.iv37, 1
+  %exitcond39 = icmp ne i64 %indvars.iv.next38, 32
+  br i1 %exitcond39, label %for.cond7.preheader, label %for.inc27
+
+for.inc27:                                        ; preds = %for.inc24
+  %indvars.iv.next41 = add nuw nsw i64 %indvars.iv40, 1
+  %exitcond42 = icmp ne i64 %indvars.iv.next41, 1024
+  br i1 %exitcond42, label %for.cond4.preheader, label %for.inc30
+
+for.inc30:                                        ; preds = %for.inc27
+  %indvars.iv.next44 = add nuw nsw i64 %indvars.iv43, 1
+  %exitcond45 = icmp ne i64 %indvars.iv.next44, 32
+  br i1 %exitcond45, label %for.cond1.preheader, label %for.end32
+
+for.end32:                                        ; preds = %for.inc30
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts_18.ll

+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1| FileCheck %s
+; REQUIRES: asserts
+;
+;      for (i = 0; i < 32; i++)
+;        for (j = 0; j < 32; j++)
+;          for (k = 0; k < 32; ++k)
+;            for (l = 0; l < 32; ++l)
+;              for (w = 0; w < 32; ++w)
+;                for (q = 0; q < 32; ++q)
+;                  C[i][j][k][w] += A[i][l][j][q] * B[q][w][l][k];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal void @kernel_tc(i32 %ni, i32 %nj, i32 %nk, i32 %nl, i32 %nq, i32 %nw, double %alpha, double %beta, [32 x [32 x [32 x double]]]* %C, [32 x [32 x [32 x double]]]* %A, [32 x [32 x [32 x double]]]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc50, %entry
+  %indvars.iv71 = phi i64 [ 0, %entry ], [ %indvars.iv.next72, %for.inc50 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc47, %for.cond1.preheader
+  %indvars.iv68 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next69, %for.inc47 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc44, %for.cond4.preheader
+  %indvars.iv65 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next66, %for.inc44 ]
+  br label %for.cond10.preheader
+
+for.cond10.preheader:                             ; preds = %for.inc41, %for.cond7.preheader
+  %indvars.iv62 = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next63, %for.inc41 ]
+  br label %for.cond13.preheader
+
+for.cond13.preheader:                             ; preds = %for.inc38, %for.cond10.preheader
+  %indvars.iv59 = phi i64 [ 0, %for.cond10.preheader ], [ %indvars.iv.next60, %for.inc38 ]
+  br label %for.body15
+
+for.body15:                                       ; preds = %for.body15, %for.cond13.preheader
+  %indvars.iv = phi i64 [ 0, %for.cond13.preheader ], [ %indvars.iv.next, %for.body15 ]
+  %arrayidx21 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %A, i64 %indvars.iv71, i64 %indvars.iv62, i64 %indvars.iv68, i64 %indvars.iv
+  %i = load double, double* %arrayidx21, align 8
+  %arrayidx29 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %B, i64 %indvars.iv, i64 %indvars.iv59, i64 %indvars.iv62, i64 %indvars.iv65
+  %i1 = load double, double* %arrayidx29, align 8
+  %mul = fmul fast double %i1, %i
+  %arrayidx37 = getelementptr inbounds [32 x [32 x [32 x double]]], [32 x [32 x [32 x double]]]* %C, i64 %indvars.iv71, i64 %indvars.iv68, i64 %indvars.iv65, i64 %indvars.iv59
+  %i2 = load double, double* %arrayidx37, align 8
+  %add = fadd fast double %i2, %mul
+  store double %add, double* %arrayidx37, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, 32
+  br i1 %exitcond, label %for.body15, label %for.inc38
+
+for.inc38:                                        ; preds = %for.body15
+  %indvars.iv.next60 = add nuw nsw i64 %indvars.iv59, 1
+  %exitcond61 = icmp ne i64 %indvars.iv.next60, 32
+  br i1 %exitcond61, label %for.cond13.preheader, label %for.inc41
+
+for.inc41:                                        ; preds = %for.inc38
+  %indvars.iv.next63 = add nuw nsw i64 %indvars.iv62, 1
+  %exitcond64 = icmp ne i64 %indvars.iv.next63, 32
+  br i1 %exitcond64, label %for.cond10.preheader, label %for.inc44
+
+for.inc44:                                        ; preds = %for.inc41
+  %indvars.iv.next66 = add nuw nsw i64 %indvars.iv65, 1
+  %exitcond67 = icmp ne i64 %indvars.iv.next66, 32
+  br i1 %exitcond67, label %for.cond7.preheader, label %for.inc47
+
+for.inc47:                                        ; preds = %for.inc44
+  %indvars.iv.next69 = add nuw nsw i64 %indvars.iv68, 1
+  %exitcond70 = icmp ne i64 %indvars.iv.next69, 32
+  br i1 %exitcond70, label %for.cond4.preheader, label %for.inc50
+
+for.inc50:                                        ; preds = %for.inc47
+  %indvars.iv.next72 = add nuw nsw i64 %indvars.iv71, 1
+  %exitcond73 = icmp ne i64 %indvars.iv.next72, 32
+  br i1 %exitcond73, label %for.cond1.preheader, label %for.end52
+
+for.end52:                                        ; preds = %for.inc50
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts_19.ll

+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1| FileCheck %s
+; REQUIRES: asserts
+;
+;   for (i = 0; i < 8; i++)
+;     for (j = 0; j < 8; j++)
+;       for (k = 0; k < 4; ++k)
+;         for (l = 0; l < 1024; ++l)
+;           for (w = 0; w < 1024; ++w)
+;             for (q = 0; q < 4; ++q)
+;               C[i][j][k][w][q] += A[q][k][j][l][i] * B[l][w];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal void @kernel_tc([8 x [4 x [1024 x [4 x double]]]]* %C, [4 x [8 x [1024 x [8 x double]]]]* %A, [1024 x double]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc50, %entry
+  %indvars.iv71 = phi i64 [ 0, %entry ], [ %indvars.iv.next72, %for.inc50 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc47, %for.cond1.preheader
+  %indvars.iv68 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next69, %for.inc47 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc44, %for.cond4.preheader
+  %indvars.iv65 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next66, %for.inc44 ]
+  br label %for.cond10.preheader
+
+for.cond10.preheader:                             ; preds = %for.inc41, %for.cond7.preheader
+  %indvars.iv62 = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next63, %for.inc41 ]
+  br label %for.cond13.preheader
+
+for.cond13.preheader:                             ; preds = %for.inc38, %for.cond10.preheader
+  %indvars.iv59 = phi i64 [ 0, %for.cond10.preheader ], [ %indvars.iv.next60, %for.inc38 ]
+  br label %for.body15
+
+for.body15:                                       ; preds = %for.body15, %for.cond13.preheader
+  %indvars.iv = phi i64 [ 0, %for.cond13.preheader ], [ %indvars.iv.next, %for.body15 ]
+  %arrayidx23 = getelementptr inbounds [4 x [8 x [1024 x [8 x double]]]], [4 x [8 x [1024 x [8 x double]]]]* %A, i64 %indvars.iv, i64 %indvars.iv65, i64 %indvars.iv68, i64 %indvars.iv62, i64 %indvars.iv71
+  %i = load double, double* %arrayidx23, align 8
+  %arrayidx27 = getelementptr inbounds [1024 x double], [1024 x double]* %B, i64 %indvars.iv62, i64 %indvars.iv59
+  %i1 = load double, double* %arrayidx27, align 8
+  %mul = fmul fast double %i1, %i
+  %arrayidx37 = getelementptr inbounds [8 x [4 x [1024 x [4 x double]]]], [8 x [4 x [1024 x [4 x double]]]]* %C, i64 %indvars.iv71, i64 %indvars.iv68, i64 %indvars.iv65, i64 %indvars.iv59, i64 %indvars.iv
+  %i2 = load double, double* %arrayidx37, align 8
+  %add = fadd fast double %i2, %mul
+  store double %add, double* %arrayidx37, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, 4
+  br i1 %exitcond, label %for.body15, label %for.inc38
+
+for.inc38:                                        ; preds = %for.body15
+  %indvars.iv.next60 = add nuw nsw i64 %indvars.iv59, 1
+  %exitcond61 = icmp ne i64 %indvars.iv.next60, 1024
+  br i1 %exitcond61, label %for.cond13.preheader, label %for.inc41
+
+for.inc41:                                        ; preds = %for.inc38
+  %indvars.iv.next63 = add nuw nsw i64 %indvars.iv62, 1
+  %exitcond64 = icmp ne i64 %indvars.iv.next63, 1024
+  br i1 %exitcond64, label %for.cond10.preheader, label %for.inc44
+
+for.inc44:                                        ; preds = %for.inc41
+  %indvars.iv.next66 = add nuw nsw i64 %indvars.iv65, 1
+  %exitcond67 = icmp ne i64 %indvars.iv.next66, 4
+  br i1 %exitcond67, label %for.cond7.preheader, label %for.inc47
+
+for.inc47:                                        ; preds = %for.inc44
+  %indvars.iv.next69 = add nuw nsw i64 %indvars.iv68, 1
+  %exitcond70 = icmp ne i64 %indvars.iv.next69, 8
+  br i1 %exitcond70, label %for.cond4.preheader, label %for.inc50
+
+for.inc50:                                        ; preds = %for.inc47
+  %indvars.iv.next72 = add nuw nsw i64 %indvars.iv71, 1
+  %exitcond73 = icmp ne i64 %indvars.iv.next72, 8
+  br i1 %exitcond73, label %for.cond1.preheader, label %for.end52
+
+for.end52:                                        ; preds = %for.inc50
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts_2.ll

-; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true -debug < %s 2>&1 | FileCheck %s
+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1 | FileCheck %s
 ; REQUIRES: asserts
 ;
 ;    /* C := alpha*A*B + beta*C */
 ;    for (i = 0; i < _PB_NI; i++)
 ;      for (j = 0; j < _PB_NJ; j += 2)
 ;        {
 ;    C[i][j] *= beta;
 ;    for (k = 0; k < _PB_NK; ++k)
 ;      C[i][j] += alpha * A[i][k] * B[k][j];
 ;        }
 ;
 ; Check that we won’t detect the matrix multiplication pattern,
 ; if, for example, there are memory accesses that have stride 2
 ; after the interchanging of loops.
 ;
 ; CHECK-NOT: The matrix multiplication pattern was detected
+; CHECK-NOT: The tensor contraction pattern was detected
 ;
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-unknown"
 
 define internal void @kernel_gemm(i32 %arg, i32 %arg1, i32 %arg2, double %arg3, double %arg4, [1056 x double]* %arg5, [1024 x double]* %arg6, [1056 x double]* %arg7) #0 {
 bb:
   br label %bb8
 

polly/test/ScheduleOptimizer/pattern-matching-based-opts_20.ll

+; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
+; RUN: -polly-pattern-matching-based-tc-opts=true -debug < %s 2>&1| FileCheck %s
+; REQUIRES: asserts
+;
+;   for (i = 0; i < 16; i++)
+;     for (j = 0; j < 16; j++)
+;       for (k = 0; k < 8; ++k)
+;         for (l = 0; l < 1024; ++l)
+;           for (w = 0; w < 8; ++w)
+;             for (q = 0; q < 8; ++q)
+;               for (x = 0; x < 8; ++x)
+;                 C[i][j][k][w][q][x] += A[l][x][j][k] * B[w][q][l][i];
+;
+; CHECK: The tensor contraction pattern was detected
+;
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define internal void @kernel_tc([16 x [8 x [8 x [8 x [8 x double]]]]]* %C, [8 x [16 x [8 x double]]]* %A, [8 x [1024 x [16 x double]]]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.inc60, %entry
+  %indvars.iv85 = phi i64 [ 0, %entry ], [ %indvars.iv.next86, %for.inc60 ]
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.inc57, %for.cond1.preheader
+  %indvars.iv82 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next83, %for.inc57 ]
+  br label %for.cond7.preheader
+
+for.cond7.preheader:                              ; preds = %for.inc54, %for.cond4.preheader
+  %indvars.iv79 = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next80, %for.inc54 ]
+  br label %for.cond10.preheader
+
+for.cond10.preheader:                             ; preds = %for.inc51, %for.cond7.preheader
+  %indvars.iv76 = phi i64 [ 0, %for.cond7.preheader ], [ %indvars.iv.next77, %for.inc51 ]
+  br label %for.cond13.preheader
+
+for.cond13.preheader:                             ; preds = %for.inc48, %for.cond10.preheader
+  %indvars.iv73 = phi i64 [ 0, %for.cond10.preheader ], [ %indvars.iv.next74, %for.inc48 ]
+  br label %for.cond16.preheader
+
+for.cond16.preheader:                             ; preds = %for.inc45, %for.cond13.preheader
+  %indvars.iv70 = phi i64 [ 0, %for.cond13.preheader ], [ %indvars.iv.next71, %for.inc45 ]
+  br label %for.body18
+
+for.body18:                                       ; preds = %for.body18, %for.cond16.preheader
+  %indvars.iv = phi i64 [ 0, %for.cond16.preheader ], [ %indvars.iv.next, %for.body18 ]
+  %arrayidx24 = getelementptr inbounds [8 x [16 x [8 x double]]], [8 x [16 x [8 x double]]]* %A, i64 %indvars.iv76, i64 %indvars.iv, i64 %indvars.iv82, i64 %indvars.iv79
+  %i = load double, double* %arrayidx24, align 8
+  %arrayidx32 = getelementptr inbounds [8 x [1024 x [16 x double]]], [8 x [1024 x [16 x double]]]* %B, i64 %indvars.iv73, i64 %indvars.iv70, i64 %indvars.iv76, i64 %indvars.iv85
+  %i1 = load double, double* %arrayidx32, align 8
+  %mul = fmul fast double %i1, %i
+  %arrayidx44 = getelementptr inbounds [16 x [8 x [8 x [8 x [8 x double]]]]], [16 x [8 x [8 x [8 x [8 x double]]]]]* %C, i64 %indvars.iv85, i64 %indvars.iv82, i64 %indvars.iv79, i64 %indvars.iv73, i64 %indvars.iv70, i64 %indvars.iv
+  %i2 = load double, double* %arrayidx44, align 8
+  %add = fadd fast double %i2, %mul
+  store double %add, double* %arrayidx44, align 8
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, 8
+  br i1 %exitcond, label %for.body18, label %for.inc45
+
+for.inc45:                                        ; preds = %for.body18
+  %indvars.iv.next71 = add nuw nsw i64 %indvars.iv70, 1
+  %exitcond72 = icmp ne i64 %indvars.iv.next71, 8
+  br i1 %exitcond72, label %for.cond16.preheader, label %for.inc48
+
+for.inc48:                                        ; preds = %for.inc45
+  %indvars.iv.next74 = add nuw nsw i64 %indvars.iv73, 1
+  %exitcond75 = icmp ne i64 %indvars.iv.next74, 8
+  br i1 %exitcond75, label %for.cond13.preheader, label %for.inc51
+
+for.inc51:                                        ; preds = %for.inc48
+  %indvars.iv.next77 = add nuw nsw i64 %indvars.iv76, 1
+  %exitcond78 = icmp ne i64 %indvars.iv.next77, 1024
+  br i1 %exitcond78, label %for.cond10.preheader, label %for.inc54
+
+for.inc54:                                        ; preds = %for.inc51
+  %indvars.iv.next80 = add nuw nsw i64 %indvars.iv79, 1
+  %exitcond81 = icmp ne i64 %indvars.iv.next80, 8
+  br i1 %exitcond81, label %for.cond7.preheader, label %for.inc57
+
+for.inc57:                                        ; preds = %for.inc54
+  %indvars.iv.next83 = add nuw nsw i64 %indvars.iv82, 1
+  %exitcond84 = icmp ne i64 %indvars.iv.next83, 16
+  br i1 %exitcond84, label %for.cond4.preheader, label %for.inc60
+
+for.inc60:                                        ; preds = %for.inc57
+  %indvars.iv.next86 = add nuw nsw i64 %indvars.iv85, 1
+  %exitcond87 = icmp ne i64 %indvars.iv.next86, 16
+  br i1 %exitcond87, label %for.cond1.preheader, label %for.end62
+
+for.end62:                                        ; preds = %for.inc60
+  ret void
+}

polly/test/ScheduleOptimizer/pattern-matching-based-opts_4.ll

 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
-; RUN: -debug < %s 2>&1| FileCheck %s
+; RUN: -debug -polly-pattern-matching-based-tc-opts=true < %s 2>&1| FileCheck %s
 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
 ; RUN: -polly-target-throughput-vector-fma=1 \
 ; RUN: -polly-target-latency-vector-fma=8 \
 ; RUN: -polly-target-1st-cache-level-size=32768 \
 ; RUN: -polly-target-vector-register-bitwidth=256 \
 ; RUN: -polly-target-2nd-cache-level-size=262144 -polly-ast \
-; RUN: -analyze < %s | FileCheck %s --check-prefix=PATTERN-MATCHING-OPTS
+; RUN: -analyze -polly-pattern-matching-based-tc-opts=true < %s |  \
+; RUN: FileCheck %s --check-prefix=PATTERN-MATCHING-OPTS
 ; REQUIRES: asserts
 ;
 ;    C := A * B + C
 ;    Check that the pattern matching optimizations can detect different
 ;    permutations of GEMM loop and produce the correct ISL AST. In this case,
 ;    dimensions of band nodes can be implicitly permuted by the algorithm
 ;    applied during the schedule generation. It should be taken into the
 ;    account during the pattern matching optimizations.
 ;    for (i = 0; i < _PB_NI; i++)
 ;      for (k = 0; k < _PB_NK; ++k)
 ;        for (j = 0; j < _PB_NJ; j++)
 ;    C[i][j] += A[i][k] * B[k][j];
 ;
+; CHECK: The tensor contraction pattern was detected
 ; CHECK: The matrix multiplication pattern was detected
 ;
 ; PATTERN-MATCHING-OPTS:    // 1st level tiling - Tiles
 ; PATTERN-MATCHING-OPTS-NEXT:    for (int c1 = 0; c1 <= 3; c1 += 1) {
 ; PATTERN-MATCHING-OPTS-NEXT:      for (int c3 = 256 * c1; c3 <= 256 * c1 + 255; c3 += 1)
 ; PATTERN-MATCHING-OPTS-NEXT:        for (int c4 = 0; c4 <= 1023; c4 += 1)
 ; PATTERN-MATCHING-OPTS-NEXT:          CopyStmt_0(0, c3, c4);
 ; PATTERN-MATCHING-OPTS-NEXT:      for (int c2 = 0; c2 <= 10; c2 += 1) {