diff --git a/mace/core/runtime/hexagon/hexagon_control_wrapper.cc b/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
index 5443969510f2fc566d4763ba82ab95cd381fcd20..b90906610524c250062567bd606b4f46084e2369 100644
--- a/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
+++ b/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
@@ -21,6 +21,7 @@
 
 #include "mace/core/runtime/hexagon/hexagon_control_wrapper.h"
 #include "mace/core/runtime/hexagon/hexagon_nn_ops.h"
+#include "mace/core/types.h"
 
 namespace {
 inline int64_t NowMicros() {
@@ -58,9 +59,9 @@ bool HexagonControlWrapper::Config() {
 
 bool HexagonControlWrapper::Init() {
   LOG(INFO) << "Hexagon init";
-  nn_id_ = hexagon_nn_init();
+  MACE_CHECK(hexagon_nn_init(&nn_id_) == 0, "hexagon_nn_init failed");
   ResetPerfInfo();
-  return nn_id_ != 0;
+  return true;
 }
 
 bool HexagonControlWrapper::Finalize() {
@@ -138,9 +139,21 @@ bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
         inputs[i].src_id = node_id(op.node_input()[i].node_id());
         inputs[i].output_idx = op.node_input()[i].output_port();
       }
-      outputs.resize(op.out_max_byte_size().size());
-      for (int i = 0; i < op.out_max_byte_size().size(); ++i) {
-        outputs[i].max_size = op.out_max_byte_size()[i];
+      outputs.resize(op.output_shape().size());
+      for (int i = 0; i < op.output_shape().size(); ++i) {
+        outputs[i].rank = op.output_shape()[i].dims().size();
+        for (size_t j = 0; j < outputs[i].rank; ++j) {
+          outputs[i].max_sizes[j] = op.output_shape()[i].dims()[j];
+        }
+        if (outputs[i].rank == 0) {
+          outputs[i].rank = 1;
+          outputs[i].max_sizes[0] = 1;
+        }
+        outputs[i].max_sizes[outputs[i].rank] = 0;
+        outputs[i].elementsize = GetEnumTypeSize(
+            static_cast<DataType>(op.output_type()[i]));
+        outputs[i].zero_offset = 0;
+        outputs[i].stepsize = 0;
       }
       cached_inputs.push_back(inputs);
       cached_outputs.push_back(outputs);
@@ -215,13 +228,13 @@ bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
 
   int64_t t1 = NowMicros();
 
-  int res = hexagon_nn_prepare(nn_id_);
+  MACE_CHECK(hexagon_nn_prepare(nn_id_) == 0, "hexagon_nn_prepare failed");
 
   int64_t t2 = NowMicros();
 
   VLOG(1) << "Setup time: " << t1 - t0 << " " << t2 - t1;
 
-  return res == 0;
+  return true;
 }
 
 bool HexagonControlWrapper::TeardownGraph() {
@@ -289,6 +302,7 @@ void HexagonControlWrapper::GetPerfInfo() {
     std::string node_type(node_type_buf);
     LOG(INFO) << "node id: " << perf_info[i].node_id
               << ", node type: " << node_type
+              << ", node type id: " << node_type_id
               << ", executions: " << perf_info[i].executions
               << ", duration: " << node_id_counters[node_id];
 
diff --git a/third_party/nnlib/hexagon_nn.h b/third_party/nnlib/hexagon_nn.h
index 82308003e4f18fa187aed6aeec1898ad87e415ea..b8dbad353a059c2e55c10785b33c97880ca1dadb 100644
--- a/third_party/nnlib/hexagon_nn.h
+++ b/third_party/nnlib/hexagon_nn.h
@@ -77,8 +77,11 @@ struct hexagon_nn_input {
 };
 typedef struct hexagon_nn_output hexagon_nn_output;
 struct hexagon_nn_output {
-  unsigned int max_size;
-  unsigned int unused;
+  unsigned int rank;
+  unsigned int max_sizes[8];
+  unsigned int elementsize;
+  int zero_offset;
+  float stepsize;
 };
 typedef struct hexagon_nn_perfinfo hexagon_nn_perfinfo;
 struct hexagon_nn_perfinfo {
@@ -128,7 +131,7 @@ struct hexagon_nn_const_node {
 };
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_config)(void)
     __QAIC_HEADER_ATTRIBUTE;
-__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_init)(void)
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_init)(hexagon_nn_nn_id* g)
     __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_debug_level)(
     hexagon_nn_nn_id id, int level) __QAIC_HEADER_ATTRIBUTE;
diff --git a/third_party/nnlib/libhexagon_controller.so b/third_party/nnlib/libhexagon_controller.so
index 42dbfabb2cb90dfdcff463f4d46cac981de0dacf..7daedd621b12686032f952a0da6564a13b63981a 100755
Binary files a/third_party/nnlib/libhexagon_controller.so and b/third_party/nnlib/libhexagon_controller.so differ
diff --git a/third_party/nnlib/ops.h b/third_party/nnlib/ops.h
index a01e71acab6cdae2fd96020d844767554307fe3d..43670832fbd693ffe50d4767a71586e4400f9f2e 100644
--- a/third_party/nnlib/ops.h
+++ b/third_party/nnlib/ops.h
@@ -107,6 +107,7 @@ DEF_OP_WREF(QuantizedRelu_8)
 DEF_OP_WREF(QuantizedReluX_8)
 DEF_OP_WREF(QuantizedMaxPool_8)
 DEF_OP_WREF(QuantizedAvgPool_8)
+DEF_OP_WREF(QuantizedL2Pool_8)
 DEF_OP_WREF(QuantizedConcat_8)
 DEF_OP_WREF(QuantizedBiasAdd_8p8to32)
 DEF_OP_WREF(Min_f)
@@ -122,6 +123,7 @@ DEF_OP(MatMul_f)
 DEF_OP(Relu_f)
 DEF_OP(ReluX_f)
 DEF_OP(AvgPool_f)
+DEF_OP(L2Pool_f)
 DEF_OP(MaxPool_f)
 DEF_OP(Concat_f)
 DEF_OP(BiasAdd_f)
@@ -187,6 +189,7 @@ DEF_OP_WREF(QuantizedSigmoid_8)
 DEF_OP_WREF(QuantizedTanh_8)
 DEF_OP_WREF(QuantizedSoftmax_8)
 DEF_OP_WREF(QuantizedLRN_8)
+DEF_OP_WREF(Quantizedpad2d_frame_8p)
 DEF_OP_WREF(QuantizedSub_8p8to32)
 DEF_OP_WREF(QuantizedMaximum_8)
 DEF_OP_WREF(QuantizedMinimum_8)
@@ -194,22 +197,89 @@ DEF_OP_WREF(QuantizedMinimum_8)
 DEF_OP(Pad_f)
 DEF_OP(SpaceToBatchND_f)
 DEF_OP(BatchToSpaceND_f)
-DEF_OP(QuantizedSpaceToBatchND_8)
-DEF_OP(QuantizedBatchToSpaceND_8)
 DEF_OP(QuantizedPad_8)
 DEF_OP(ResizeBilinear_f)
-DEF_OP(QuantizedResizeBilinear_8)
 DEF_OP(ConcatV2_f)
 DEF_OP(ConcatV2_int32)
 DEF_OP(Prod_int32)
 DEF_OP(Slice_int32)
 
 DEF_OP(QuantizedAdd_8p8to8)
+DEF_OP(QuantizedResizeBilinear_8)
+DEF_OP(Supernode_8x8p8to8_d32)
+DEF_OP(Convert_to_d32)
+DEF_OP(Convert_from_d32)
+DEF_OP_WREF(QuantizedMaxPool_8_d32)
+DEF_OP_WREF(QuantizedConcat_8_d32)
+DEF_OP_WREF(QuantizedAvgPool_8_d32)
+
+DEF_OP(Sink)
 
+DEF_OP_WREF(QuantizedPRelu_8_d32)
 DEF_OP_WREF(AutoQuantize)
 DEF_OP_WREF(QuantizedDepthwiseConv2d_8x8to32)
-DEF_OP(DepthwiseConv2d_f)
-DEF_OP(QuantizedBiasAdd_8p8to8)
+DEF_OP_WREF(DepthwiseConv2d_f)
+DEF_OP(DepthwiseSupernode_8x8p8to8)
+DEF_OP(DepthwiseSupernode_8x8p8to8_d32)
+
+DEF_OP_WREF(QuantizedMul_8x8to8_d32)
+
+DEF_OP(FullyConnected_u8)
+#if 0
+    DEF_OP_WREF(QuantizedFC_8x8p8to8)
+#endif
+
+DEF_OP_WREF(QuantizedAdd_8p8to8_d32)
+
+DEF_OP_WREF(QuantizedClamp_8)
+DEF_OP(Clamp_f)
+DEF_OP(QuantizeForTest_d32)
+DEF_OP(Close_d32)
+DEF_OP_WREF(QuantizedSub_8p8to8_d32)
+
+DEF_OP(InputSupernode_8x8p8to8_outd32)
+DEF_OP(QuantizedLRN_8_d32)
+DEF_OP_WREF(QuantizedBiasAdd_32p32to32)
+DEF_OP_WREF(Quantize_int32)
+
+DEF_OP(Supernode_8x8p32to8)
+DEF_OP(DepthwiseSupernode_8x8p32to8)
+DEF_OP(Supernode_8x8p32to8_d32)
+DEF_OP(DepthwiseSupernode_8x8p32to8_d32)
+DEF_OP(InputSupernode_8x8p32to8_outd32)
+
+DEF_OP(PPrint_8_d32)
+DEF_OP(PPrintWithPadding_8_d32)
+DEF_OP_WREF(AutoQuantize_d32)
+
+DEF_OP_WREF(QuantizedTanh_8_d32)
+DEF_OP_WREF(QuantizedSigmoid_8_d32)
+DEF_OP_WREF(QuantizedSoftmax_8_d32)
+
+
+DEF_OP_WREF(QuantizedL2Pool_8_d32)
+
+DEF_OP(Gather_f)
+DEF_OP(Gather_int32)
+DEF_OP(Gather_8)
+DEF_OP(Table_f)
+DEF_OP(Table_int32)
+DEF_OP(Table_8)
+
+DEF_OP(FillPadding_8_d32)
+DEF_OP(QuantizedResizeBilinear_8_d32)
+
+DEF_OP(QuantizeINPUT_f_to_8)
+DEF_OP_WREF(DeconvBias_8x8to32)
+
+DEF_OP(SpaceToBatchND_8)
+DEF_OP(BatchToSpaceND_8)
+
+
+DEF_OP(SpaceToDepth_f)
+DEF_OP(DepthToSpace_f)
+DEF_OP(SpaceToDepth_8)
+DEF_OP(DepthToSpace_8)
 
 #ifdef __SELF_DEF_OP_WREF
 #undef __SELF_DEF_OP_WREF