TOSA: concat: fix canonicalization that would result in concat with no operands

The fold() of the concat had two issues:
a) It would fold `concat (tensor<0x2>, tensor<0x4>), axis=1 -> tensor<0x6>` into `concat ()` (no operands)
based on the observation that both operands have zero elements. This is invalid as concat needs to have
at least one operand.

b) After that fix, it would fold `concat (tensor<0x2>, tensor<0x4>), axis=1 -> tensor<0x6>`
int `concat (tensor<0x4>), axis=1 -> tensor<0x6>`. This is also invalid; even though the concat
still produces the same number of elements (= none), shape relations are not corret (0x4 input, but 0x6 output).

I fixed that by only removing operands from the concat when
- the operand has zero dim on the concatenation axis (i.e. it doesn't contribute to the result shape)
- there is at least one operand left after removing

Author: mgehre-amd

mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp

@@ -1349,19 +1349,13 @@ OpFoldResult tosa::AbsOp::fold(FoldAdaptor adaptor) {
   return {};
 }
 
-static bool hasZeroSize(Type ty) {
-  auto ranked = dyn_cast<RankedTensorType>(ty);
-  if (!ranked)
-    return false;
-  return any_of(ranked.getShape(), [](auto d) { return d == 0; });
-}
-
 OpFoldResult ConcatOp::fold(FoldAdaptor adaptor) {
   /// Remove operands that have zero elements.
   bool changed = false;
   for (size_t i = 0; i < getInput1().size(); ) {
-    auto input = getInput1()[i];
-    if (hasZeroSize(input.getType())) {
+    auto input = cast<RankedTensorType>(getInput1()[i].getType());
+    // Ensure that we have at least one operand left.
+    if (input.getDimSize(getAxis()) == 0 && getInput1().size() > 1) {
       getInput1Mutable().erase(i);
       changed = true;
     } else {

mlir/test/Dialect/Tosa/canonicalize.mlir

@@ -204,6 +204,32 @@ func.func @concat_fold_zero(%arg0: tensor<?x0xf32>, %arg1: tensor<?x1xf32>, %arg
   %0 = tosa.concat %arg0, %arg1, %arg2 {axis = 1 : i32}: (tensor<?x0xf32>, tensor<?x1xf32>, tensor<?x2xf32>) -> tensor<?x3xf32>
   return %0 : tensor<?x3xf32>
 }
+// -----
+
+// CHECK-LABEL: @concat_fold_zero
+func.func @concat_fold_zero_all(%arg0: tensor<?x0xf32>, %arg1: tensor<?x0xf32>) -> tensor<?x0xf32> {
+  // CHECK: return %arg1
+  %0 = tosa.concat %arg0, %arg1 {axis = 1 : i32}: (tensor<?x0xf32>, tensor<?x0xf32>) -> tensor<?x0xf32>
+  return %0 : tensor<?x0xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @concat_fold_zero
+func.func @concat_fold_zero_different_axis(%arg0: tensor<0x2xf32>, %arg1: tensor<0x4xf32>) -> tensor<0x6xf32> {
+  // CHECK: tosa.concat %arg0, %arg1
+  %0 = tosa.concat %arg0, %arg1 {axis = 1 : i32}: (tensor<0x2xf32>, tensor<0x4xf32>) -> tensor<0x6xf32>
+  return %0 : tensor<0x6xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @concat_fold_zero_size
+func.func @concat_fold_zero_size(%arg0: tensor<?x0xf32>, %arg1: tensor<?x1xf32>, %arg2: tensor<?x2xf32>) -> tensor<?x3xf32> {
+  // CHECK: tosa.concat %arg1, %arg2 {axis = 1 : i32}
+  %0 = tosa.concat %arg0, %arg1, %arg2 {axis = 1 : i32}: (tensor<?x0xf32>, tensor<?x1xf32>, tensor<?x2xf32>) -> tensor<?x3xf32>
+  return %0 : tensor<?x3xf32>
+}
 
 // -----