follow comments

paddle/operators/math/matmul.h

@@ -45,14 +45,15 @@ class MatMulFunctor {
     std::vector<int64_t> out_dim;
     int64_t batch_count = 1;
     if (dim_a.size() > 3) {
-      PADDLE_ENFORCE(dim_b.size() > 3,
+      PADDLE_ENFORCE(dim_b.size() == dim_a.size(),
                      "The dimensions of X and Y must be the same, and both of "
                      "them should be %d-dimensional.",
                      dim_b.size());
-      // The previous Rank-2 dimensions are accumulated on the batch_count.
+      // The front rank-2 dimensions are accumulated on the batch_count, and the
+      // last two dimensions are used for matrix multiplication.
       for (int j = 0; j < dim_a.size() - 2; ++j) {
-        PADDLE_ENFORCE(dim_b[j] == dim_a[j],
-                       "The dimensions of X[%d] and Y[%d] must be the same.", j,
+        PADDLE_ENFORCE_EQ(dim_b[j], dim_a[j],
+                          "The %d-th dimension of X and Y must be the same.",
                           j);
         out_dim.push_back(dim_a[j]);
         batch_count *= dim_a[j];

paddle/operators/matmul_op.cc

@@ -45,15 +45,17 @@ class MatMulOp : public framework::OperatorWithKernel {
     std::vector<int64_t> out_dim;
     int64_t batch_count = 1;
     if (dim_x.size() > 3) {
-      PADDLE_ENFORCE(dim_y.size() == dim_x.size(),
+      PADDLE_ENFORCE_EQ(
+          dim_y.size(), dim_x.size(),
           "The dimensions of X and Y must be the same, and both of "
           "them should be %d-dimensional.",
           dim_x.size());
 
-      // The previous Rank-2 dimensions are accumulated on the batch_count.
+      // The front rank-2 dimensions are accumulated on the batch_count, and the
+      // last two dimensions are used for matrix multiplication.
       for (int j = 0; j < dim_x.size() - 2; ++j) {
-        PADDLE_ENFORCE(dim_y[j] == dim_x[j],
-                       "The dimensions of X[%d] and Y[%d] must be the same.", j,
+        PADDLE_ENFORCE_EQ(dim_y[j], dim_x[j],
+                          "The %d-th dimension of X and Y must be the same.",
                           j);
         out_dim.push_back(dim_x[j]);
         batch_count *= dim_x[j];
@@ -191,10 +193,10 @@ Examples without transpose:
 
 The behavior is designed to be similar to the `numpy.matmul` function.
 The differences are:
-- When the rank of the input is greater than 3, the rank of X and
-  Y must be equal, and the former rank-2 dimensions are equal.
 - When the rank of the input data is less than or equal to 3, it
   is similar to the `numpy.matmul` function.
+- When the rank of the input is greater than 3, the rank of X and
+  Y must be equal, and the front `rank - 2` dimensions must be equal.
 - We add `transpose_X` and `transpose_Y` flags.
 
 Both the input `X` and `Y` can carry the LoD (Level of Details) information,

paddle/operators/matmul_op.h

@@ -138,7 +138,8 @@ class MatMulGradKernel : public framework::OpKernel<T> {
     }
 
     int batch_count = 0;
-    // The previous Rank-2 dimensions are accumulated on the batch_count.
+    // The front rank-2 dimensions are accumulated on the batch_count, and the
+    // last two dimensions are used for matrix multiplication.
     if (x_dims.size() > 3) {
       batch_count = accumulate(x_dims.begin(), x_dims.end() - 2, 1,
                                std::multiplies<int>());