[Lang] MatrixNdarray refactor part8: Add scalarization for BinaryOpStmt with TensorType-operands

taichi/ir/frontend_ir.cpp

@@ -350,12 +350,14 @@ void BinaryOpExpression::flatten(FlattenContext *ctx) {
     auto ret = ctx->push_back<LocalLoadStmt>(result);
     ret->tb = tb;
     stmt = ret;
+    stmt->ret_type = ret_type;
     return;
   }
   flatten_rvalue(rhs, ctx);
   ctx->push_back(std::make_unique<BinaryOpStmt>(type, lhs->stmt, rhs->stmt));
   ctx->stmts.back()->tb = tb;
   stmt = ctx->back_stmt();
+  stmt->ret_type = ret_type;
 }
 
 void make_ifte(Expression::FlattenContext *ctx,

taichi/transforms/scalarize.cpp

@@ -116,7 +116,6 @@ class Scalarize : public IRVisitor {
 
       auto matrix_init_stmt =
           std::make_unique<MatrixInitStmt>(matrix_init_values);
-
       matrix_init_stmt->ret_type = src_dtype;
 
       stmt->replace_usages_with(matrix_init_stmt.get());
@@ -178,6 +177,73 @@ class Scalarize : public IRVisitor {
     }
   }
 
+  /*
+    Before:
+      TensorType<4 x i32> val = BinaryStmt(TensorType<4 x i32> lhs,
+                                           TensorType<4 x i32> rhs)
+
+      * Note that "lhs" and "rhs" should have already been scalarized to
+    MatrixInitStmt
+
+    After:
+      i32 calc_val0 = BinaryStmt(lhs->cast<MatrixInitStmt>()->val[0],
+                                 rhs->cast<MatrixInitStmt>()->val[0])
+      i32 calc_val1 = BinaryStmt(lhs->cast<MatrixInitStmt>()->val[1],
+                                 rhs->cast<MatrixInitStmt>()->val[1])
+      i32 calc_val2 = BinaryStmt(lhs->cast<MatrixInitStmt>()->val[2],
+                                 rhs->cast<MatrixInitStmt>()->val[2])
+      i32 calc_val3 = BinaryStmt(lhs->cast<MatrixInitStmt>()->val[3],
+                                 rhs->cast<MatrixInitStmt>()->val[3])
+
+      tmp = MatrixInitStmt(calc_val0, calc_val1,
+                           calc_val2, calc_val3)
+
+      stmt->replace_all_usages_with(tmp)
+  */
+  void visit(BinaryOpStmt *stmt) override {
+    auto lhs_dtype = stmt->lhs->ret_type;
+    auto rhs_dtype = stmt->rhs->ret_type;
+
+    // BinaryOpExpression::type_check() should have taken care of the
+    // broadcasting and neccessary conversions. So we simply add an assertion
+    // here to make sure that the operands are of the same shape and dtype
+    TI_ASSERT(lhs_dtype == rhs_dtype);
+
+    if (lhs_dtype->is<TensorType>() && rhs_dtype->is<TensorType>()) {
+      // Scalarization for LoadStmt should have already replaced both operands
+      // to MatrixInitStmt
+      TI_ASSERT(stmt->lhs->is<MatrixInitStmt>());
+      TI_ASSERT(stmt->rhs->is<MatrixInitStmt>());
+
+      auto lhs_matrix_init_stmt = stmt->lhs->cast<MatrixInitStmt>();
+      std::vector<Stmt *> lhs_vals = lhs_matrix_init_stmt->values;
+
+      auto rhs_matrix_init_stmt = stmt->rhs->cast<MatrixInitStmt>();
+      std::vector<Stmt *> rhs_vals = rhs_matrix_init_stmt->values;
+
+      TI_ASSERT(rhs_vals.size() == lhs_vals.size());
+
+      size_t num_elements = lhs_vals.size();
+      std::vector<Stmt *> matrix_init_values;
+      for (size_t i = 0; i < num_elements; i++) {
+        auto binary_stmt = std::make_unique<BinaryOpStmt>(
+            stmt->op_type, lhs_vals[i], rhs_vals[i]);
+        matrix_init_values.push_back(binary_stmt.get());
+
+        modifier_.insert_before(stmt, std::move(binary_stmt));
+      }
+
+      auto matrix_init_stmt =
+          std::make_unique<MatrixInitStmt>(matrix_init_values);
+      matrix_init_stmt->ret_type = stmt->ret_type;
+
+      stmt->replace_usages_with(matrix_init_stmt.get());
+      modifier_.insert_before(stmt, std::move(matrix_init_stmt));
+
+      modifier_.erase(stmt);
+    }
+  }
+
   void visit(Block *stmt_list) override {
     for (auto &stmt : stmt_list->statements) {
       stmt->accept(this);

tests/python/test_matrix.py

@@ -970,3 +970,25 @@ def verify(x):
     field = ti.Matrix.field(2, 2, ti.f32, shape=5)
     ndarray = ti.Matrix.ndarray(2, 2, ti.f32, shape=5)
     _test_field_and_ndarray(field, ndarray, func, verify)
+
+
+@test_utils.test(arch=[ti.cuda, ti.cpu],
+                 real_matrix=True,
+                 real_matrix_scalarize=True)
+def test_binary_op_scalarize():
+    @ti.func
+    def func(a: ti.template()):
+        a[0] = [[0., 1.], [2., 3.]]
+        a[1] = [[3., 4.], [5., 6.]]
+        a[2] = a[0] + a[0]
+        a[3] = a[1] * a[1]
+        a[4] = ti.max(a[2], a[3])
+
+    def verify(x):
+        assert (x[2] == [[0., 2.], [4., 6.]]).all()
+        assert (x[3] == [[9., 16.], [25., 36.]]).all()
+        assert (x[4] == [[9., 16.], [25., 36.]]).all()
+
+    field = ti.Matrix.field(2, 2, ti.f32, shape=5)
+    ndarray = ti.Matrix.ndarray(2, 2, ti.f32, shape=5)
+    _test_field_and_ndarray(field, ndarray, func, verify)