Merge branch 'matcopy_batch_pr' into omatadd_batch_pr

include/operations/extension/transpose.h

@@ -161,7 +161,11 @@ make_transpose(in_t &A, index_t inc_a, index_t &stride_a, out_t &At,
  * while remaining customizable Tiling-size wise.
  *
  * @tparam both_trans Whether both A & B matrices are transposed (or just the
- * first one)
+ * first one A). In fact, this kernel is implemented in such a way that if
+ * only one matrix is transposed in the OmatAdd operator, it should be placed
+ * as the first operand to this kernel (A & alpha). This reduces the original 4
+ * possible combinations of A and B transpose cases into only two cases depicted
+ * by the template parameter both_trans.
  * @tparam Tile_size Tiling size used explicitly in the local memory kernel, and
  * used to compute work-group size in the non-local memory case.
  * @tparam wg_size work group size

src/operations/extension/transpose.hpp

@@ -445,9 +445,9 @@ TransposeAdd<both_trans, Tile_size, wg_size, cl_size, local_memory, in1_t,
                                             cl::sycl::nd_item<1> id) {
   value_t *local = local_mem.localAcc.get_pointer();
 
-  auto A = A_.get_data().get_pointer();
-  auto B = B_.get_data().get_pointer();
-  auto C = C_.get_data().get_pointer();
+  auto A = A_.get_pointer();
+  auto B = B_.get_pointer();
+  auto C = C_.get_pointer();
 
   index_t in_a_idx, in_b_idx, in_local_id, out_idx, out_local_id;
   index_t i_block_start, j_block_start;