[Mlir-commits] [mlir] [mlir][linalg] Emit a warning when tile_using_forall generates non thread-safe code (PR #80813)

[Oleksandr Alex Zinenko]

================
@@ -586,3 +586,144 @@ module attributes {transform.with_named_sequence} {
     transform.yield
   }
 }
+
+// -----
+
+#map = affine_map<(d0, d1) -> (d0, d1)>
+#map1 = affine_map<(d0, d1) -> (d0)>
+
+func.func @tile_thread_safety1(%arg0: tensor<100x300xf32>, %arg1: tensor<100xf32>) -> tensor<100xf32> {
+  // expected-warning at below {{tiling is not thread safe at axis #1}}
+  %0 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "reduction"]} ins(%arg0 : tensor<100x300xf32>) outs(%arg1 : tensor<100xf32>) {
+  ^bb0(%in: f32, %out: f32):
+    %1 = arith.addf %in, %out : f32
+    linalg.yield %1 : f32
+  } -> tensor<100xf32>
+  return %0 : tensor<100xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 num_threads [4, 2]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+
+// -----
+
+#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
+#map1 = affine_map<(d0, d1, d2) -> (d1, d2)>
+
+func.func @tile_thread_safety2(%arg0: tensor<100x300x8xf32>, %arg1: tensor<300x8xf32>) -> tensor<300x8xf32> {
+  // expected-warning at below {{tiling is not thread safe at axis #0}}
+  %0 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["reduction", "parallel", "parallel"]} ins(%arg0 : tensor<100x300x8xf32>) outs(%arg1 : tensor<300x8xf32>) {
+  ^bb0(%in: f32, %out: f32):
+    %1 = arith.addf %in, %out : f32
+    linalg.yield %1 : f32
+  } -> tensor<300x8xf32>
+  return %0 : tensor<300x8xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 num_threads [8]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+
+// -----
+
+#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
+#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
+
+func.func @tile_thread_safety3(%arg0: tensor<100x300x8xf32>, %arg1: tensor<100x8xf32>) -> tensor<100x8xf32> {
+  // expected-warning at below {{tiling is not thread safe at axis #1}}
+  %0 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "reduction", "parallel"]} ins(%arg0 : tensor<100x300x8xf32>) outs(%arg1 : tensor<100x8xf32>) {
+  ^bb0(%in: f32, %out: f32):
+    %1 = arith.addf %in, %out : f32
+    linalg.yield %1 : f32
+  } -> tensor<100x8xf32>
+  return %0 : tensor<100x8xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 num_threads [8, 4, 2]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+
+// -----
+
+#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
+#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
+#map2 = affine_map<(d0, d1, d2) -> (d2)>
+
+func.func @tile_thread_safety4(%arg0: tensor<100x300x8xf32>, %arg1: tensor<100x8xf32>, %arg2 : tensor<8xf32>) -> (tensor<100x8xf32>, tensor<8xf32>) {
+  // expected-warning at +2 {{tiling is not thread safe at axis #0}}
+  // expected-warning at below {{tiling is not thread safe at axis #1}}
+  %0:2 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["reduction", "reduction", "parallel"]} ins(%arg0 : tensor<100x300x8xf32>) outs(%arg1, %arg2 : tensor<100x8xf32>, tensor<8xf32>) {
+  ^bb0(%in: f32, %out1: f32, %out2: f32):
+    %1 = arith.addf %in, %out1 : f32
+    %2 = arith.addf %in, %out2 : f32
+    linalg.yield %1, %2 : f32, f32
+  } -> (tensor<100x8xf32>, tensor<8xf32>)
+  return %0#0, %0#1 : tensor<100x8xf32>, tensor<8xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 num_threads [8, 4, 2]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+
+// -----
+
+#map = affine_map<(d0, d1) -> (d0, d1)>
+#map1 = affine_map<(d0, d1) -> (d0)>
+
+func.func @tile_thread_safety5(%arg0: tensor<100x300xf32>, %arg1: tensor<100xf32>) -> tensor<100xf32> {
+  // expected-warning at below {{tiling is not thread safe at axis #1}}
+  %0 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "reduction"]} ins(%arg0 : tensor<100x300xf32>) outs(%arg1 : tensor<100xf32>) {
+  ^bb0(%in: f32, %out: f32):
+    %1 = arith.addf %in, %out : f32
+    linalg.yield %1 : f32
+  } -> tensor<100xf32>
+  return %0 : tensor<100xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 tile_sizes [10, 1]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+
+// -----
+
+func.func @tile_thread_safety6(%A: tensor<?x?xf32>, %B: tensor<?x?xf32>, %C: tensor<?x?xf32>) -> tensor<?x?xf32> {
+  // expected-warning at below {{tiling is not thread safe at axis #2}}
+  %0 = linalg.matmul ins(%A, %B : tensor<?x?xf32>, tensor<?x?xf32>)
+                    outs(%C : tensor<?x?xf32>) -> (tensor<?x?xf32>)
+  return %0 : tensor<?x?xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %forall, %tiled_generic = transform.structured.tile_using_forall %0 num_threads [2, 4, 8]
+          : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
----------------
ftynse wrote:

Nit: please keep the trailing newline.

https://github.com/llvm/llvm-project/pull/80813