!13572 fusion pass adjust for transpose as output

From: @zhaozhenlong
Reviewed-by: @zhanghaibo5,@zhang_xue_tong
Signed-off-by: @zhang_xue_tong

mindspore/lite/src/runtime/agent/npu/optimizer/npu_fusion_pass.cc

@@ -141,6 +141,23 @@ void UpdatePreTensors(kernel::LiteKernel *cur_kernel) {
 
 void UpdatePostTensors(kernel::LiteKernel *cur_kernel) {
   auto tensor = cur_kernel->out_tensors()[0];
+
+  // in case: node->nh2nc->nc2nh(graph output) --->>> node->nc2nh, node out_tensor should be put to nnc2nh out tensors
+  auto out_kernels = cur_kernel->out_kernels();
+  if (out_kernels.size() == 1 && out_kernels[0]->out_kernels().size() == 1 &&
+      out_kernels[0]->out_kernels()[0]->out_kernels().empty() &&
+      out_kernels[0]->out_kernels()[0]->type_str() == "Transpose") {
+    auto nc_tensor = out_kernels[0]->out_tensors()[0];  // nh2nc's out tensor
+    cur_kernel->set_out_tensors({nc_tensor});
+    auto post_post_kernel = out_kernels[0]->out_kernels()[0];
+    // nc2nh kernel set in_tensor out_tensor
+    auto post_post_k_in_tensors = post_post_kernel->in_tensors();
+    post_post_k_in_tensors[0] = nc_tensor;
+    post_post_kernel->set_in_tensors(post_post_k_in_tensors);
+    post_post_kernel->set_out_tensors({tensor});
+    return;
+  }
+
   tensor->set_format(schema::Format_NCHW);
   auto nhwc_shape = tensor->shape();
   tensor->set_shape({nhwc_shape[0], nhwc_shape[3], nhwc_shape[1], nhwc_shape[2]});

mindspore/lite/src/runtime/agent/npu/optimizer/npu_insert_transform_pass.cc

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 #include "src/runtime/agent/npu/optimizer/npu_insert_transform_pass.h"
+#include <algorithm>
 #include <set>
 #include <string>
 #include "src/runtime/agent/npu/optimizer/npu_pass_utils.h"
@@ -51,7 +52,8 @@ int GetInsertState(kernel::LiteKernel *kernel) {
 
   // current kernel is target kernel
   // use out kernels to count how many out lines from current kernel
-  size_t in_out_tensor_num = kernel->in_tensors().size() + kernel->out_kernels().size();
+  size_t in_out_tensor_num =
+    kernel->in_tensors().size() + std::max(kernel->out_kernels().size(), static_cast<size_t>(1));
   size_t transpose_input_num = 0;
   size_t transpose_output_num = 0;
   bool need_pre_insert = false;
@@ -65,6 +67,9 @@ int GetInsertState(kernel::LiteKernel *kernel) {
       need_pre_insert = true;
     }
   }
+  if (kernel->out_kernels().empty()) {
+    need_post_insert = true;
+  }
   for (const auto out_kernel : kernel->out_kernels()) {
     if (NPUPassUtils::IsNhwc2Nchw(out_kernel)) {
       transpose_output_num++;

mindspore/lite/src/runtime/agent/npu/optimizer/npu_pass_utils.cc

@@ -106,6 +106,9 @@ void NPUPassUtils::UpdateNH2NCTransNodePreKernel(kernel::LiteKernel *pre_kernel,
       break;
     }
   }
+  if (out_kernels.empty()) {
+    out_kernels.push_back(trans_kernel);
+  }
   pre_kernel->set_out_kernels(out_kernels);
 }
 

mindspore/lite/src/runtime/agent/npu/optimizer/npu_transform_pass.cc

@@ -156,8 +156,8 @@ int NPUTransformPass::InsertPostNodes(kernel::LiteKernel *kernel, std::vector<ke
       nc2nh_out_tensors[0] = out_tensor;
 
       // Create post transform kernel: Nchw2Nhwc
-      auto *post_trans_kernel =
-        NPUPassUtils::CreateNchw2NhwcKernel({nc2nh_tensor, nc2nh_perm_tensor}, nc2nh_out_tensors, context_, name);
+      auto *post_trans_kernel = NPUPassUtils::CreateNchw2NhwcKernel(
+        {nc2nh_tensor, nc2nh_perm_tensor}, nc2nh_out_tensors, context_, name + "_" + std::to_string(i));
       // Set in_kernels, out_kernels, in_tensors, out_tensors for transform kernel
       NPUPassUtils::UpdateKernel(post_trans_kernel, {kernel}, {post_insert_kernel}, post_trans_kernel->in_tensors(),
                                  post_trans_kernel->out_tensors());

mindspore/lite/test/models_npu.cfg

@@ -70,3 +70,4 @@ ml_video_edit_v10_best_model_nomean_20200723 8
 ml_edu_kit_hand_key_position.onnx 2
 #ml_video_edit_oneclick_adaptis.pb #too many subgraphs
 densenet.tflite 3
+resnet_v2_101_299.tflite 1