[MLIR] Fix affine fusion bug/efficiency issue / enable more fusion

[bondhugula]

Summary

The list of destination load ops while evaluating producer-consumer
fusion wasn't being maintained as a set, and as such, duplicate load ops
were being added to it. Although this is harmless correctness-wise, it's
a killer efficiency-wise and it prevents interesting/useful fusions
(including for eg. reshapes into a matmul). The reason the latter
fusions would be missed is that a slice union would be unnecessarily
needed due to the duplicate load ops on a memref added to the 'dst
loads' list. Since slice union is unimplemented for the local var case,
a single destination load op that leads to local vars (like a floordiv /
mod producing fusion), a common case, would not get fused due to an
unnecessary union being tried with itself. (The union would actually be
the same thing but we would bail out.)

Besides the above, this would also significantly speed up fusion as all
the unnecessary slice computations / unions, checks, etc. due to the
duplicates go away.

[andydavis1]

Thanks Uday! Looks good...

mlir/lib/Transforms/LoopFusion.cpp

1628

Is this a TODO: consider using a hash set if performance becomes an issue?

[bondhugula]

mlir/lib/Transforms/LoopFusion.cpp

1628

It's not really a TODO at this point. A vector should be fine unless there are use cases where this is shown to be a bottleneck. We are only going to have a handful of ops, and if we have hundreds for some use case, there'd also other places to address.

[dcaballe]

LGTM, Thanks!

mlir/lib/Transforms/LoopFusion.cpp

while (!worklist.empty()) {
             // Collect dst loop stats after memref privatization transformation.
             LoopNestStateCollector dstLoopCollector;
             dstLoopCollector.collect(dstAffineForOp.getOperation());
 
             // Add new load ops to current Node load op list 'loads' to
             // continue fusing based on new operands.
             for (auto *loadOpInst : dstLoopCollector.loadOpInsts) {
               auto loadMemRef = cast<AffineLoadOp>(loadOpInst).getMemRef();
-              if (visitedMemrefs.count(loadMemRef) == 0)
+              // NOTE: Change 'loads' to a hash set in case efficiency is an
+              // issue. We still use a vector since it's expected to be small.
+              if (visitedMemrefs.count(loadMemRef) == 0 &&
+                  !llvm::is_contained(loads, loadOpInst))
                 loads.push_back(loadOpInst);
             }
 
             // Clear and add back loads and stores.
             mdg->clearNodeLoadAndStores(dstNode->id);
             mdg->addToNode(dstId, dstLoopCollector.loadOpInsts,
                            dstLoopCollector.storeOpInsts);
             // Remove old src loop nest if it no longer has outgoing dependence

mlir/test/Transforms/loop-fusion.mlir

affine.for %arg0 = 0 to 1 {
     %1 = affine.load %b[%arg0] : memref<1xf32>
   }
   // CHECK: affine.for %{{.*}} = 0 to 1
   // CHECK-NEXT: affine.load %[[A:.*]][{{.*}}]
   // CHECK-NEXT: affine.store %{{.*}}, %[[B:.*]][{{.*}}]
   // CHECK-NEXT: affine.store %{{.*}}, %[[A]]
   // CHECK-NEXT: affine.load %[[B]]
   // CHECK-NOT: affine.for %{{.*}}
+  // CHECK: return
   return
 }
+
+// -----
+
+// MAXIMAL-LABEL: func @reshape_into_matmul
+func @reshape_into_matmul(%lhs : memref<1024x1024xf32>,
+              %R: memref<16x64x1024xf32>, %out: memref<1024x1024xf32>) {
+  %rhs = alloc() :  memref<1024x1024xf32>
+
+  // Reshape from 3-d to 2-d.
+  affine.for %i0 = 0 to 16 {
+    affine.for %i1 = 0 to 64 {
+      affine.for %k = 0 to 1024 {
+        %v = affine.load %R[%i0, %i1, %k] : memref<16x64x1024xf32>
+        affine.store %v, %rhs[64*%i0 + %i1, %k] : memref<1024x1024xf32>
+      }
+    }
+  }
+
+  // Matmul.
+  affine.for %i = 0 to 1024 {
+    affine.for %j = 0 to 1024 {
+      affine.for %k = 0 to 1024 {
+        %0 = affine.load %rhs[%k, %j] : memref<1024x1024xf32>
+        %1 = affine.load %lhs[%i, %k] : memref<1024x1024xf32>
+        %2 = mulf %1, %0 : f32
+        %3 = affine.load %out[%i, %j] : memref<1024x1024xf32>
+        %4 = addf %3, %2 : f32
+        affine.store %4, %out[%i, %j] : memref<1024x1024xf32>
+      }
+    }
+  }
+  return
+}
+// MAXIMAL-NEXT: alloc
+// MAXIMAL-NEXT: affine.for
+// MAXIMAL-NEXT:   affine.for
+// MAXIMAL-NEXT:     affine.for
+// MAXIMAL-NOT:      affine.for
+// MAXIMAL:      return