[Mlir-commits] [mlir] [MLIR][SCF] Add an API to fuse consumer to a producer within scf loop (PR #88712)

Quinn Dawkins llvmlistbot at llvm.org
Sat May 18 10:47:31 PDT 2024


================
@@ -0,0 +1,322 @@
+// RUN: mlir-opt --transform-interpreter --cse --split-input-file %s | FileCheck %s
+
+#map = affine_map<(d0) -> (d0)>
+module {
+  func.func @fuse_tileable_consumer_scf_for(%arg0: tensor<32xf32>, %arg1: tensor<32xf32>, %arg2: tensor<64xf32>) -> tensor<64xf32> {
+    %c4 = arith.constant 4 : index
+    %c64 = arith.constant 64 : index
+    %c0 = arith.constant 0 : index
+    %1:2 = scf.for %arg3 = %c0 to %c64 step %c4 iter_args(%arg4 = %arg2, %arg5 = %arg2) -> (tensor<64xf32>, tensor<64xf32>) {
+      %extracted_slice = tensor.extract_slice %arg4[%arg3] [32] [1] : tensor<64xf32> to tensor<32xf32>
+      %3 = linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel"]} ins(%arg0, %arg1 : tensor<32xf32>, tensor<32xf32>) outs(%extracted_slice : tensor<32xf32>) {
+        ^bb0(%in: f32, %in_16: f32, %out: f32):
+          %13 = arith.mulf %in, %in_16 : f32
+          %14 = arith.addf %out, %13 : f32
+          linalg.yield %14 : f32
+        } -> tensor<32xf32>
+      %4 = tensor.insert_slice %3 into %arg4[%arg3] [32] [1] : tensor<32xf32> into tensor<64xf32>
+      scf.yield %arg5, %4 : tensor<64xf32>, tensor<64xf32>
+    }
+    %in_operand_2 = tensor.empty() : tensor<64xf32>
+    %out_operand_3 = tensor.empty() : tensor<64xf32>
+    %2 = linalg.elemwise_binary {fun = #linalg.binary_fn<add>} ins(%1#1, %in_operand_2 : tensor<64xf32>, tensor<64xf32>) outs(%out_operand_3 : tensor<64xf32>) -> tensor<64xf32>
+    return %2 : tensor<64xf32>
+  }
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
+    %yield = transform.structured.match ops{["tensor.insert_slice"]} in %arg1
+      : (!transform.any_op) -> !transform.any_op
+    %a, %b = transform.test.fuse_consumer %yield
+      : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+//      CHECK: func.func @fuse_tileable_consumer_scf_for(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<32xf32>
+// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<32xf32>
+// CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: tensor<64xf32>)
+//      CHECK:   %[[C0:.*]] = arith.constant 0 : index
+//      CHECK:   %0 = tensor.empty() : tensor<64xf32>
+//      CHECK:   %[[FINAL_RESULT:.*]]:3 = scf.for %[[IV:.*]] = %[[C0]]
+// CHECK-SAME:      iter_args(%[[FIRST_OUT_ARG:.*]] = %[[ARG2]], %[[SECOND_OUT_ARG:.*]] = %[[ARG2]], %[[ELEM_OUT_ARG:.*]] = %0)
+// CHECK-SAME:   {
+//      CHECK:      %[[MAT_OUT_SLICE:.*]] = tensor.extract_slice %[[FIRST_OUT_ARG]][%[[IV]]] [32] [1]
+//      CHECK:      %[[MAT_OUT:.*]] = linalg.generic
+// CHECK-SAME:              outs(%[[MAT_OUT_SLICE]] : tensor<32xf32>)
+//      CHECK:      %[[INSERT_MAT:.*]] = tensor.insert_slice %[[MAT_OUT]] into %[[FIRST_OUT_ARG]][%[[IV]]] [32] [1]
+//      CHECK:      %[[SLICE_OPERAND1:.*]] = tensor.extract_slice %[[INSERT_MAT]][%[[IV]]] [32] [1]
+//      CHECK:      %[[SLICE_OPERAND2:.*]] = tensor.extract_slice %0[%[[IV]]] [32] [1]
+//      CHECK:      %[[SLICE_OUT:.*]] = tensor.extract_slice %[[ELEM_OUT_ARG]][%[[IV]]] [32] [1]
+//      CHECK:      %[[ELEM_OUT:.*]] = linalg.elemwise_binary {fun = #linalg.binary_fn<add>}
+// CHECK-SAME:              ins(%[[SLICE_OPERAND1]], %[[SLICE_OPERAND2]] :
+// CHECK-SAME:              outs(%[[SLICE_OUT]] :
+//      CHECK:      %[[INSERT_ELEM:.*]] = tensor.insert_slice %[[ELEM_OUT]] into %[[ELEM_OUT_ARG]][%[[IV]]] [32] [1]
+//      CHECK:      scf.yield %[[SECOND_OUT_ARG]], %[[INSERT_MAT]], %[[INSERT_ELEM]] :
+//      CHECK:   }
+//      CHECK:   return %[[FINAL_RESULT]]#2 :
+
+// -----
+
+module {
+  func.func @fuse_tileable_consumer_scf_forall(%arg0: tensor<32x32xf32>, %arg1: tensor<32x32xf32>, %arg2: tensor<64x64xf32>) -> tensor<64x64xf32> {
+    %c4 = arith.constant 4 : index
+    %c64 = arith.constant 64 : index
+    %c0 = arith.constant 0 : index
+    %1:2 = scf.forall (%arg3, %arg4) in (2, 2) shared_outs(%arg5 = %arg2, %arg6 = %arg2) -> (tensor<64x64xf32>, tensor<64x64xf32>) {
+      %extracted_slice = tensor.extract_slice %arg5[%arg3, %arg4] [32, 32] [1, 1] : tensor<64x64xf32> to tensor<32x32xf32>
+      %extracted_slice_1 = tensor.extract_slice %arg6[%arg3, %arg4] [32, 32] [1, 1] : tensor<64x64xf32> to tensor<32x32xf32>
+      %3 = linalg.matmul ins(%arg0, %arg1 : tensor<32x32xf32>, tensor<32x32xf32>) outs(%extracted_slice : tensor<32x32xf32>) -> tensor<32x32xf32>
+      scf.forall.in_parallel {
+         tensor.parallel_insert_slice %3 into %arg6[%arg3, %arg4] [32, 32] [1, 1] : tensor<32x32xf32> into tensor<64x64xf32>
+         tensor.parallel_insert_slice %extracted_slice_1 into %arg5[%arg3, %arg4] [32, 32] [1, 1] : tensor<32x32xf32> into tensor<64x64xf32>
+      }
+    }
+    %in_operand_2 = tensor.empty() : tensor<64x64xf32>
+    %out_operand_3 = tensor.empty() : tensor<64x64xf32>
+    %2 = linalg.elemwise_binary {fun = #linalg.binary_fn<add>} ins(%1#1, %in_operand_2 : tensor<64x64xf32>, tensor<64x64xf32>) outs(%out_operand_3 : tensor<64x64xf32>) -> tensor<64x64xf32>
+    return %2 : tensor<64x64xf32>
+  }
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
+    %slice_ops = transform.structured.match ops{["tensor.parallel_insert_slice"]} in %arg1
+      : (!transform.any_op) -> !transform.any_op
+    %first_slice_op, %second_slice_op = transform.split_handle %slice_ops
+        : (!transform.any_op)
+        -> (!transform.any_op, !transform.any_op)
+    %a, %b = transform.test.fuse_consumer %first_slice_op
+      : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+//      CHECK: func.func @fuse_tileable_consumer_scf_forall(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<32x32xf32>
+// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<32x32xf32>
+// CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: tensor<64x64xf32>)
+//      CHECK:   %[[OUT_INIT:.*]] = tensor.empty() : tensor<64x64xf32>
+//      CHECK:   %[[FINAL_RESULT:.*]]:3 = scf.forall (%[[IV1:.*]], %[[IV2:.*]]) in (2, 2)
+// CHECK-SAME:      shared_outs(%[[FIRST_OUT_ARG:.*]] = %[[ARG2]], %[[SECOND_OUT_ARG:.*]] = %[[ARG2]], %[[ELEM_OUT_ARG:.*]] = %[[OUT_INIT]])
+// CHECK-SAME:   {
+//      CHECK:      %[[MAT_OUT_SLICE:.*]] = tensor.extract_slice %[[FIRST_OUT_ARG]][%[[IV1]], %[[IV2]]] [32, 32] [1, 1]
+//      CHECK:      %[[SECOND_ARG_SLICE:.*]] = tensor.extract_slice %[[SECOND_OUT_ARG]][%[[IV1]], %[[IV2]]] [32, 32] [1, 1]
+//      CHECK:      %[[MAT_OUT:.*]] = linalg.matmul
+// CHECK-SAME:              outs(%[[MAT_OUT_SLICE]] :
+//      CHECK:      %[[SLICE_OPERAND1:.*]] = tensor.extract_slice %[[MAT_OUT]][%[[IV1]], %[[IV2]]] [32, 32] [1, 1]
----------------
qedawkins wrote:

This looks wrong to me. After fusing the consumer, the operand to the fused op is this `extract_slice` of the matmul result with offsets equal to the ones used by the parallel insert. The full operation is:
```
tensor.extract_slice %matmul[%arg3, %arg4] [32, 32] [1, 1] : tensor<32x32xf32> to tensor<32x32xf32>
```
which is out of bounds unless the offsets are zero. Instead the fused consumer should just be operating directly on the result of the matmul.

https://github.com/llvm/llvm-project/pull/88712


More information about the Mlir-commits mailing list