Fix issue with transposing shape in CMSIS-NN batch matmul (#2741)

BUG=#2740
tensorflow · Nov 6, 2024 · 9245002 · 9245002
1 parent 4bb78c7
commit 9245002
Showing 1 changed file with 6 additions and 1 deletion.
diff --git a/tensorflow/lite/micro/kernels/cmsis_nn/batch_matmul.cc b/tensorflow/lite/micro/kernels/cmsis_nn/batch_matmul.cc
@@ -121,7 +121,12 @@ inline TfLiteStatus PopulateEvalData(
     RuntimeShape tmp_r = SwapRowColumnDims(*rhs_shape);
     rhs_shape->ReplaceWith(tmp_r.DimensionsCount(), tmp_r.DimsData());
   }
-  if (!params->adj_x) {
+  // ReferenceOps and CMSIS-NN have different requirements for when the
+  // lhs shape should be transposed, so we have to treat float differently.
+  if (!params->adj_x && original_lhs_input->type == kTfLiteFloat32) {
+    RuntimeShape tmp_l = SwapRowColumnDims(*lhs_shape);
+    lhs_shape->ReplaceWith(tmp_l.DimensionsCount(), tmp_l.DimsData());
+  } else if (params->adj_x && original_lhs_input->type != kTfLiteFloat32) {
     RuntimeShape tmp_l = SwapRowColumnDims(*lhs_shape);
     lhs_shape->ReplaceWith(tmp_l.DimensionsCount(), tmp_l.DimsData());
   }