[Mlir-commits] [mlir] [mlir][sparse] fix bug in custom reduction scalarization code (PR #74898)

Aart Bik llvmlistbot at llvm.org
Fri Dec 8 15:43:15 PST 2023


https://github.com/aartbik created https://github.com/llvm/llvm-project/pull/74898

Bug found with BSR of "spy" SDDMM method

>From 802b106f4fb16da1396bc9c925a2be3686f3d6c4 Mon Sep 17 00:00:00 2001
From: Aart Bik <ajcbik at google.com>
Date: Fri, 8 Dec 2023 15:40:35 -0800
Subject: [PATCH] [mlir][sparse] fix bug in custom reduction scalarization code

Bug found with BSR of "spy" SDDMM method
---
 .../Transforms/Sparsification.cpp             |  24 ++--
 .../Dialect/SparseTensor/spy_sddmm_bsr.mlir   | 103 ++++++++++++++++++
 2 files changed, 120 insertions(+), 7 deletions(-)
 create mode 100755 mlir/test/Dialect/SparseTensor/spy_sddmm_bsr.mlir

diff --git a/mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp b/mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp
index 6c9adf9fa21a0..992be434fc623 100644
--- a/mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp
+++ b/mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp
@@ -673,25 +673,35 @@ static void genInvariants(CodegenEnv &env, OpBuilder &builder, ExprId exp,
     // All exhausted at current level.
     if (!isCurrentLoop)
       return;
+    // Generate code for a scalarized reduction or invariant. Note that
+    // because custom reduction lhs may occur several times in the IR,
+    // we have a built-in safety for only initializing and wrapping-up
+    // the scalarized reduction once.
     OpOperand *lhs = op.getDpsInitOperand(0);
     if (lhs == &t) {
       // Start or end a scalarized reduction.
       if (isStart) {
-        Value load = env.isCustomReduc() ? env.getCustomRedId()
-                                         : genTensorLoad(env, builder, exp);
-        env.startReduc(exp, load);
+        if (env.isCustomReduc()) {
+          if (!env.isReduc())
+            env.startReduc(exp, env.getCustomRedId());
+        } else {
+          env.startReduc(exp, genTensorLoad(env, builder, exp));
+        }
         if (env.hasSparseOutput())
           env.setValidLexInsert(constantI1(builder, env.op().getLoc(), false));
       } else {
-        genTensorStore(env, builder, exp, env.endReduc());
-        env.clearValidLexInsert();
+        if (!env.isCustomReduc() || env.isReduc())
+          genTensorStore(env, builder, exp, env.endReduc());
+        if (env.hasSparseOutput())
+          env.clearValidLexInsert();
       }
     } else {
       // Start or end loop invariant hoisting of a tensor load.
-      if (isStart)
+      if (isStart) {
         env.merger().setExprValue(exp, genTensorLoad(env, builder, exp));
-      else
+      } else {
         env.merger().clearExprValue(exp);
+      }
     }
   } else if (env.exp(exp).kind != TensorExp::Kind::kInvariant &&
              env.exp(exp).kind != TensorExp::Kind::kLoopVar &&
diff --git a/mlir/test/Dialect/SparseTensor/spy_sddmm_bsr.mlir b/mlir/test/Dialect/SparseTensor/spy_sddmm_bsr.mlir
new file mode 100755
index 0000000000000..ed8d639878967
--- /dev/null
+++ b/mlir/test/Dialect/SparseTensor/spy_sddmm_bsr.mlir
@@ -0,0 +1,103 @@
+// RUN: mlir-opt %s --sparse-reinterpret-map -sparsification | FileCheck %s
+
+//
+// A SDDMM implementation with "spy" function and
+// in-place update of the sampling sparse matrix.
+//
+
+#BSR = #sparse_tensor.encoding<{
+  map = (i, j) -> (
+    i floordiv 2 : dense,
+    j floordiv 2 : compressed,
+    i mod 2 : dense,
+    j mod 2 : dense)
+}>
+
+#trait_SDDMM = {
+  indexing_maps = [
+    affine_map<(i,j,k) -> (i,k)>,  // A
+    affine_map<(i,j,k) -> (k,j)>,  // B
+    affine_map<(i,j,k) -> (i,j)>   // S (in/out)
+  ],
+  iterator_types = ["parallel", "parallel", "reduction"],
+  doc = "S(i,j) += spy[S(i,j)] x SUM_k A(i,k) B(k,j)"
+}
+
+//
+// CHECK: #[[$BSR:.+]] = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 floordiv 2 : dense, d1 floordiv 2 : compressed, d0 mod 2 : dense, d1 mod 2 : dense) }>
+// CHECK: #[[$MAP:.+]] = #sparse_tensor.encoding<{ map = (d0, d1, d2, d3) -> (d0 : dense, d1 : compressed, d2 : dense, d3 : dense) }>
+//
+// CHECK-LABEL:   func.func @SDDMM_block(
+// CHECK-SAME:      %[[VAL_0:.*]]: tensor<?x?xf32, #[[$BSR]]>,
+// CHECK-SAME:      %[[VAL_1:.*]]: tensor<?x?xf32>,
+// CHECK-SAME:      %[[VAL_2:.*]]: tensor<?x?xf32>) -> tensor<?x?xf32, #[[$BSR]]> {
+// CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 1 : index
+// CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
+// CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 2 : index
+// CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 0.000000e+00 : f32
+// CHECK:           %[[VAL_7:.*]] = sparse_tensor.reinterpret_map %[[VAL_0]] : tensor<?x?xf32, #[[$BSR]]> to tensor<?x?x2x2xf32, #[[$MAP]]>
+// CHECK:           %[[VAL_8:.*]] = tensor.dim %[[VAL_1]], %[[VAL_3]] : tensor<?x?xf32>
+// CHECK:           %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<?x?xf32>
+// CHECK:           %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]] : memref<?x?xf32>
+// CHECK:           %[[VAL_11:.*]] = sparse_tensor.lvl %[[VAL_7]], %[[VAL_4]] : tensor<?x?x2x2xf32, #[[$MAP]]>
+// CHECK:           %[[VAL_12:.*]] = sparse_tensor.positions %[[VAL_7]] {level = 1 : index} : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xindex>
+// CHECK:           %[[VAL_13:.*]] = sparse_tensor.coordinates %[[VAL_7]] {level = 1 : index} : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xindex>
+// CHECK:           %[[VAL_14:.*]] = sparse_tensor.values %[[VAL_7]] : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xf32>
+// CHECK:           scf.for %[[VAL_15:.*]] = %[[VAL_4]] to %[[VAL_11]] step %[[VAL_3]] {
+// CHECK:             %[[VAL_16:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_15]]] : memref<?xindex>
+// CHECK:             %[[VAL_17:.*]] = arith.addi %[[VAL_15]], %[[VAL_3]] : index
+// CHECK:             %[[VAL_18:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_17]]] : memref<?xindex>
+// CHECK:             scf.for %[[VAL_19:.*]] = %[[VAL_16]] to %[[VAL_18]] step %[[VAL_3]] {
+// CHECK:               %[[VAL_20:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_19]]] : memref<?xindex>
+// CHECK:               scf.for %[[VAL_21:.*]] = %[[VAL_4]] to %[[VAL_5]] step %[[VAL_3]] {
+// CHECK:                 %[[VAL_22:.*]] = arith.muli %[[VAL_19]], %[[VAL_5]] : index
+// CHECK:                 %[[VAL_23:.*]] = arith.addi %[[VAL_22]], %[[VAL_21]] : index
+// CHECK:                 scf.for %[[VAL_24:.*]] = %[[VAL_4]] to %[[VAL_5]] step %[[VAL_3]] {
+// CHECK:                   %[[VAL_25:.*]] = arith.muli %[[VAL_23]], %[[VAL_5]] : index
+// CHECK:                   %[[VAL_26:.*]] = arith.addi %[[VAL_25]], %[[VAL_24]] : index
+// CHECK:                   %[[VAL_27:.*]] = scf.for %[[VAL_28:.*]] = %[[VAL_4]] to %[[VAL_8]] step %[[VAL_3]] iter_args(%[[VAL_29:.*]] = %[[VAL_6]]) -> (f32) {
+// CHECK:                     %[[VAL_30:.*]] = arith.muli %[[VAL_15]], %[[VAL_5]] : index
+// CHECK:                     %[[VAL_31:.*]] = arith.addi %[[VAL_30]], %[[VAL_21]] : index
+// CHECK:                     %[[VAL_32:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_31]], %[[VAL_28]]] : memref<?x?xf32>
+// CHECK:                     %[[VAL_33:.*]] = arith.muli %[[VAL_20]], %[[VAL_5]] : index
+// CHECK:                     %[[VAL_34:.*]] = arith.addi %[[VAL_33]], %[[VAL_24]] : index
+// CHECK:                     %[[VAL_35:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_28]], %[[VAL_34]]] : memref<?x?xf32>
+// CHECK:                     %[[VAL_36:.*]] = arith.mulf %[[VAL_32]], %[[VAL_35]] : f32
+// CHECK:                     %[[VAL_37:.*]] = arith.addf %[[VAL_29]], %[[VAL_36]] : f32
+// CHECK:                     scf.yield %[[VAL_37]] : f32
+// CHECK:                   } {"Emitted from" = "linalg.generic"}
+// CHECK:                   memref.store %[[VAL_27]], %[[VAL_14]]{{\[}}%[[VAL_26]]] : memref<?xf32>
+// CHECK:                 } {"Emitted from" = "linalg.generic"}
+// CHECK:               } {"Emitted from" = "linalg.generic"}
+// CHECK:             } {"Emitted from" = "linalg.generic"}
+// CHECK:           } {"Emitted from" = "linalg.generic"}
+// CHECK:           %[[VAL_38:.*]] = sparse_tensor.load %[[VAL_7]] : tensor<?x?x2x2xf32, #[[$MAP]]>
+// CHECK:           %[[VAL_39:.*]] = sparse_tensor.reinterpret_map %[[VAL_38]] : tensor<?x?x2x2xf32, #[[$MAP]]> to tensor<?x?xf32, #[[$BSR]]>
+// CHECK:           return %[[VAL_39]] : tensor<?x?xf32, #[[$BSR]]>
+// CHECK:         }
+module {
+  func.func @SDDMM_block(%args: tensor<?x?xf32, #BSR>,
+                         %arga: tensor<?x?xf32>,
+                         %argb: tensor<?x?xf32>) -> tensor<?x?xf32, #BSR> {
+    %result = linalg.generic #trait_SDDMM
+      ins(%arga, %argb: tensor<?x?xf32>, tensor<?x?xf32>)
+      outs(%args: tensor<?x?xf32, #BSR>) {
+        ^bb(%a: f32, %b: f32, %s: f32):
+           %f0 = arith.constant 0.0 : f32
+           %u = sparse_tensor.unary %s : f32 to f32
+             present={
+                ^bb0(%p: f32):
+                  %mul = arith.mulf %a, %b : f32
+                  sparse_tensor.yield %mul : f32
+             }
+             absent={}
+           %r = sparse_tensor.reduce %s, %u, %f0 : f32 {
+              ^bb0(%p: f32, %q: f32):
+                %add = arith.addf %p, %q : f32
+                sparse_tensor.yield %add : f32
+            }
+           linalg.yield %r : f32
+      } -> tensor<?x?xf32, #BSR>
+    return %result : tensor<?x?xf32, #BSR>
+  }
+}



More information about the Mlir-commits mailing list