fix

lite/backends/nnadapter/nnadapter/driver/huawei_ascend_npu/converter/pad.cc

@@ -16,7 +16,7 @@
 #include "driver/huawei_ascend_npu/converter/converter.h"
 #include "utility/debug.h"
 #include "utility/logging.h"
-
+#include "utility/modeling.h"
 namespace nnadapter {
 namespace huawei_ascend_npu {
 
@@ -40,6 +40,39 @@ int ConvertPad(Converter* converter, hal::Operation* operation) {
   if (!input_operator) {
     input_operator = converter->ConvertOperand(input_operand);
   }
+  if ((mode == NNADAPTER_PAD_MODE_REFLECT || mode == NNADAPTER_PAD_MODE_EDGE) && 
+      input_operand->type.dimensions.count == 5 &&
+      IsConstantOperand(pads_operand) &&
+      !IsOperandWithDynamicShape(input_operand)) {
+    uint32_t pads_size = pads_operand->length / static_cast<uint32_t>(sizeof(int32_t));
+    auto pads_buffer = reinterpret_cast<int32_t*>(pads_operand->buffer);
+    if (pads_size == 10 && 
+        pads_buffer[6] == 0 && pads_buffer[7] == 0 && pads_buffer[8] == 0 && pads_buffer[9] == 0) {
+      // Reshape to 4-dimensions
+      std::vector<int32_t> shape_data(input_operand->type.dimensions.data, input_operand->type.dimensions.data + 3);
+      shape_data.push_back(input_operand->type.dimensions.data[3] * input_operand->type.dimensions.data[4]);
+      auto shape_operator = converter->AddInt32ConstantOperator(shape_data);
+      auto reshape_op = converter->AddOperator<ge::op::Reshape>(output_operand);
+      SET_INPUT(reshape_op, x, input_operator);
+      SET_INPUT(reshape_op, shape, shape_operator);
+      auto reshape_output_operator = MAP_OUTPUT(reshape_op, y, output_operand);
+      // Use NCHW data format
+      std::vector<int32_t> pad_data(pads_buffer, pads_buffer + 8);
+      auto pads_operator = converter->AddInt32ConstantOperator(pad_data);
+      auto pad_op = converter->AddOperator<ge::op::PadV3>(output_operand);
+      pad_op->set_attr_mode(pad_mode);
+      SET_INPUT(pad_op, x, reshape_output_operator);
+      SET_INPUT(pad_op, paddings, pads_operator);
+      auto pad2d_output_operator = MAP_OUTPUT(pad_op, y, output_operand);
+      // Reshape to 5-dimensions
+      auto shape_operator2 = converter->AddInt32ConstantOperator(std::vector<int32_t>(output_operand->type.dimensions.data, output_operand->type.dimensions.data + output_operand->type.dimensions.count));
+      auto reshape_op2 = converter->AddOperator<ge::op::Reshape>(output_operand);
+      SET_INPUT(reshape_op2, x, pad2d_output_operator);
+      SET_INPUT(reshape_op2, shape, shape_operator2);
+      MAP_OUTPUT(reshape_op2, y, output_operand);
+      return NNADAPTER_NO_ERROR;
+    }
+  }
   auto pads_operator = converter->GetMappedOperator(pads_operand);
   if (!pads_operator) {
     pads_operator = converter->ConvertOperand(pads_operand);

lite/backends/nnadapter/nnadapter/driver/huawei_ascend_npu/utility.cc

@@ -221,6 +221,7 @@ bool BuildOMModelToBuffer(
   options.insert(std::make_pair(ge::ir_option::LOG_LEVEL, "error"));
   options.insert(std::make_pair(ge::ir_option::OP_DEBUG_LEVEL, "0"));
   ATC_CALL(aclgrphBuildModel(ir_graph, options, om_buffer));
+  ATC_CALL(aclgrphSaveModel("ir_graph_model", om_buffer));  // For debug: save ascend offline model to local.
   // Copy from om model buffer
   model_buffer->resize(om_buffer.length);
   memcpy(reinterpret_cast<void*>(model_buffer->data()),

lite/core/optimizer/mir/fusion/unsqueeze2_pad3d_squeeze2_fuse.cc

@@ -17,6 +17,7 @@
 //
 
 #include "lite/core/optimizer/mir/fusion/unsqueeze2_pad3d_squeeze2_fuse.h"
+
 #include <memory>
 #include <vector>
 
@@ -26,6 +27,12 @@ namespace mir {
 namespace fusion {
 
 void Unsqueeze2Pad3dSqueeze2Fuser::BuildPattern() {
+  auto paddings_teller = [](const Node* node) -> bool {
+    auto op_desc = *const_cast<Node*>(node)->stmt()->op_info();
+    auto paddings = op_desc.GetAttr<std::vector<int>>("paddings");
+    return paddings.size() == 6 && paddings[4] == 0 && paddings[5] == 0;
+  };
+
   // create input nodes.
   auto* unsqu_input = VarNode("unsqu_input")
                           ->assert_is_op_input(unsqueeze2_type_, "X")
@@ -35,8 +42,10 @@ void Unsqueeze2Pad3dSqueeze2Fuser::BuildPattern() {
   auto* unsque = OpNode("unsqueeze2", unsqueeze2_type_)
                      ->assert_is_op(unsqueeze2_type_)
                      ->AsIntermediate();
-  auto* p3d =
-      OpNode("pad3d", pad3d_type_)->assert_is_op(pad3d_type_)->AsIntermediate();
+  auto* p3d = OpNode("pad3d", pad3d_type_)
+                  ->assert_is_op(pad3d_type_)
+                  ->assert_node_satisfied(paddings_teller)
+                  ->AsIntermediate();
   auto* sque = OpNode("squeeze2", squeeze2_type_)
                    ->assert_is_op(squeeze2_type_)
                    ->AsIntermediate();

lite/tests/kernels/pad2d_compute_test.cc

@@ -200,8 +200,7 @@ TEST(Pad2d, precision) {
 #if defined(NNADAPTER_WITH_HUAWEI_ASCEND_NPU)
               // Ascend does not support the following scenarios.
               if (std::abs(pad_value - 1) < 1e-6 ||
-                  (pad_top == 0 && pad_bottom == 1 && pad_left == 0 &&
-                   pad_right == 0))
+                  ((pad_top == 1 || pad_bottom == 1) && pad_left == 0 && pad_right == 0))
                 continue;
 #endif
               VLOG(5) << "pad param: " << pad_mode << " " << pad_value << " "