[Mlir-commits] [mlir] [MLIR][Vector] Extend elementwise pattern to support unrolling from higher rank to lower rank (PR #162515)

[Nishant Patel]

================
@@ -440,3 +463,39 @@ func.func @vector_step() -> vector<32xindex> {
 // CHECK: %[[ADD3:.*]] = arith.addi %[[STEP]], %[[CST]] : vector<8xindex>
 // CHECK: %[[INS3:.*]] = vector.insert_strided_slice %[[ADD3]], %[[INS2]] {offsets = [24], strides = [1]} : vector<8xindex> into vector<32xindex>
 // CHECK: return %[[INS3]] : vector<32xindex>
+
+
+func.func @elementwise(%v1: vector<2x2x2xf32>, %v2: vector<2x2x2xf32>) -> vector<2x2x2xf32> {
+  %0 = arith.addf %v1, %v2 : vector<2x2x2xf32>
+  return %0 : vector<2x2x2xf32>
+}
+// CHECK-LABEL: func @elementwise
+//  CHECK-SAME: (%[[ARG0:.*]]: vector<2x2x2xf32>, %[[ARG1:.*]]: vector<2x2x2xf32>) -> vector<2x2x2xf32> {
+//       CHECK:   %[[CST:.*]] = arith.constant dense<0.000000e+00> : vector<2x2x2xf32>
+//       CHECK:   %[[E_LHS_0:.*]] = vector.extract_strided_slice %[[ARG0]] {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+//       CHECK:   %[[S_LHS_0:.*]] = vector.shape_cast %[[E_LHS_0]] : vector<1x2x2xf32> to vector<2x2xf32>
+//       CHECK:   %[[E_RHS_0:.*]] = vector.extract_strided_slice %[[ARG1]] {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+//       CHECK:   %[[S_RHS_0:.*]] = vector.shape_cast %[[E_RHS_0]] : vector<1x2x2xf32> to vector<2x2xf32>
+//       CHECK:   %[[ADD0:.*]] = arith.addf %[[S_LHS_0]], %[[S_RHS_0]] : vector<2x2xf32>
+//       CHECK:   %[[I0:.*]] = vector.insert_strided_slice %[[ADD0]], %[[CST]] {offsets = [0, 0, 0], strides = [1, 1]} : vector<2x2xf32> into vector<2x2x2xf32>
+//       CHECK:   %[[E_LHS_1:.*]] = vector.extract_strided_slice %[[ARG0]] {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+//       CHECK:   %[[S_LHS_1:.*]] = vector.shape_cast %[[E_LHS_1]] : vector<1x2x2xf32> to vector<2x2xf32>
+//       CHECK:   %[[E_RHS_1:.*]] = vector.extract_strided_slice %[[ARG1]] {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+//       CHECK:   %[[S_RHS_1:.*]] = vector.shape_cast %[[E_RHS_1]] : vector<1x2x2xf32> to vector<2x2xf32>
+//       CHECK:   %[[ADD1:.*]] = arith.addf %[[S_LHS_1]], %[[S_RHS_1]] : vector<2x2xf32>
+//       CHECK:   %[[I1:.*]] = vector.insert_strided_slice %[[ADD1]], %[[I0]] {offsets = [1, 0, 0], strides = [1, 1]} : vector<2x2xf32> into vector<2x2x2xf32>
+//       CHECK:   return %[[I1]] : vector<2x2x2xf32>
+
+
+func.func @elementwise_4D_to_2D(%v1: vector<2x2x2x2xf32>, %v2: vector<2x2x2x2xf32>) -> vector<2x2x2x2xf32> {
+  %0 = arith.addf %v1, %v2 : vector<2x2x2x2xf32>
+  return %0 : vector<2x2x2x2xf32>
+}
+
+//       CHECK-LABEL: func @elementwise_4D_to_2D
+//        CHECK-SAME: (%[[ARG0:.*]]: vector<2x2x2x2xf32>, %[[ARG1:.*]]: vector<2x2x2x2xf32>) -> vector<2x2x2x2xf32> {
+// CHECK-DAG-COUNT-4:   vector.extract_strided_slice %[[ARG0]] {offsets = [{{.*}}], sizes = [1, 1, 2, 2], strides = [1, 1, 1, 1]} : vector<2x2x2x2xf32> to vector<1x1x2x2xf32>
----------------
nbpatel wrote:

oh my bad its not a thing, I combined multiple and unordered CHECK and combined the two....I trimmed it to just checking adds for this test case, I think it should be enough since the other test case (3D to 2D) has all the instructions required to understand the approach

https://github.com/llvm/llvm-project/pull/162515