[GPU] Dont do large tensor tiling for linalg ops with no payload

Signed-off-by: Nirvedh Meshram <[email protected]>

compiler/src/iree/compiler/Codegen/Common/TileLargeTensors.cpp

@@ -184,7 +184,11 @@ static void processRegion(RewriterBase &rewriter, Region *region,
         // significant computation anyway. Equivalent generics are still tiled
         // as they typically arise organically. Fills in particular are almost
         // never found on their own and will be fused when tiling if need be.
-        if (isa<linalg::TransposeOp, linalg::CopyOp, linalg::FillOp>(op)) {
+        // Additonally, if linalg.yeild is the only op in the linalg payload then it is
+        // a reshape/ broadcast type op and same expectations of them being
+        // carefully introduced holds. E.g, they could be transpose
+        // ops which got generalized by a reshape fusion pattern.
+        if (linalgOp.getBlock()->getOperations().size() == 1) {
           continue;
         }
         tileToMaxVectorSize(rewriter, linalgOp, maxVectorSize);

compiler/src/iree/compiler/Codegen/Common/test/tile_large_tensors.mlir

@@ -128,3 +128,18 @@ func.func @no_tile_fill(%arg0: f32) -> tensor<64x256xf32> {
 //   CHECK-NOT:   scf.for
 //       CHECK:   %[[FILL:.+]] = linalg.fill
 //       CHECK:   return %[[FILL]]
+
+
+func.func @no_tile_reshape(%arg0: tensor<1x4x4x1x4x4xi32>) -> tensor<1x4x1x4x4x4xi32> {
+  %empty = tensor.empty() : tensor<1x4x1x4x4x4xi32>
+  %0 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d3, d2, d4, d5)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d3, d4, d5)>], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x4x4x1x4x4xi32>) outs(%empty : tensor<1x4x1x4x4x4xi32>) {
+        ^bb0(%in: i32, %out: i32):
+          linalg.yield %in : i32
+    } -> tensor<1x4x1x4x4x4xi32>
+    return %0 : tensor<1x4x1x4x4x4xi32>
+}
+
+// CHECK-LABEL: func.func @no_tile_reshape
+//   CHECK-NOT:   scf.for
+//       CHECK:   %[[RESHAPE:.+]] = linalg.generic
+//       CHECK:   return %[[RESHAPE]]