Merge branch 'hexagon-2.1' into 'master'

Support hexagon_nn version 2.1 in Hexagon SDK 3.4.0

See merge request !719

mace/core/runtime/hexagon/hexagon_control_wrapper.cc

@@ -51,10 +51,9 @@ int HexagonControlWrapper::GetVersion() {
 
 bool HexagonControlWrapper::Config() {
   LOG(INFO) << "Hexagon config";
-  if (hexagon_nn_set_powersave_level(0) != 0) {
-    return false;
-  }
-  return hexagon_nn_config() == 0;
+  MACE_CHECK(hexagon_nn_set_powersave_level(0) == 0, "hexagon power error");
+  MACE_CHECK(hexagon_nn_config() == 0, "hexagon config error");
+  return true;
 }
 
 bool HexagonControlWrapper::Init() {
@@ -80,7 +79,10 @@ bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
   int64_t t0 = NowMicros();
 
   // const node
-  std::thread const_thread([&]() {
+#if defined(MACE_USE_NNLIB_CAF) || defined(MACE_USE_NNLIB_OLD)
+  std::thread const_thread([&]()
+#endif
+  {
     std::vector<hexagon_nn_const_node> const_node_list;
     for (const ConstTensor &const_tensor : net_def.tensors()) {
       std::vector<int> tensor_shape(const_tensor.dims().begin(),
@@ -124,10 +126,16 @@ bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
           "append const node error");
     }
     const_node_list.clear();
-  });
+  }
+#if defined(MACE_USE_NNLIB_CAF) || defined(MACE_USE_NNLIB_OLD)
+  );  // NOLINT
+#endif
 
   // op node
-  std::thread op_thread([&]() {
+#if defined(MACE_USE_NNLIB_CAF) || defined(MACE_USE_NNLIB_OLD)
+  std::thread op_thread([&]()
+#endif
+  {
     OpMap op_map;
     op_map.Init();
     std::vector<hexagon_nn_op_node> op_node_list;
@@ -197,10 +205,12 @@ bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
     op_node_list.clear();
     cached_inputs.clear();
     cached_outputs.clear();
-  });
-
+  }
+#if defined(MACE_USE_NNLIB_CAF) || defined(MACE_USE_NNLIB_OLD)
+  );  // NOLINT
   const_thread.join();
   op_thread.join();
+#endif
 
   // input info
   num_inputs_ = 0;

third_party/nnlib/hexagon_nn.h

@@ -225,7 +225,152 @@ __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_execute_new)(
 }
 #endif
 
-#else  // nnlib version
+#elif defined(MACE_USE_NNLIB_2_1)  // nnlib version
+
+#ifndef __QAIC_HEADER
+#define __QAIC_HEADER(ff) ff
+#endif //__QAIC_HEADER
+
+#ifndef __QAIC_HEADER_EXPORT
+#define __QAIC_HEADER_EXPORT
+#endif // __QAIC_HEADER_EXPORT
+
+#ifndef __QAIC_HEADER_ATTRIBUTE
+#define __QAIC_HEADER_ATTRIBUTE
+#endif // __QAIC_HEADER_ATTRIBUTE
+
+#ifndef __QAIC_IMPL
+#define __QAIC_IMPL(ff) ff
+#endif //__QAIC_IMPL
+
+#ifndef __QAIC_IMPL_EXPORT
+#define __QAIC_IMPL_EXPORT
+#endif // __QAIC_IMPL_EXPORT
+
+#ifndef __QAIC_IMPL_ATTRIBUTE
+#define __QAIC_IMPL_ATTRIBUTE
+#endif // __QAIC_IMPL_ATTRIBUTE
+#ifdef __cplusplus
+extern "C" {
+#endif
+#if !defined(__QAIC_STRING1_OBJECT_DEFINED__) && !defined(__STRING1_OBJECT__)
+#define __QAIC_STRING1_OBJECT_DEFINED__
+#define __STRING1_OBJECT__
+typedef struct _cstring1_s {
+   char* data;
+   int dataLen;
+} _cstring1_t;
+
+#endif /* __QAIC_STRING1_OBJECT_DEFINED__ */
+typedef struct hexagon_nn_input hexagon_nn_input;
+struct hexagon_nn_input {
+   unsigned int src_id;
+   unsigned int output_idx;
+};
+typedef struct hexagon_nn_output hexagon_nn_output;
+struct hexagon_nn_output {
+   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 {
+   unsigned int node_id;
+   unsigned int node_type;
+   unsigned int executions;
+   unsigned int unused;
+   unsigned int counter_lo;
+   unsigned int counter_hi;
+};
+typedef int hexagon_nn_nn_id;
+enum hexagon_nn_padding_type {
+   NN_PAD_NA,
+   NN_PAD_SAME,
+   NN_PAD_VALID,
+   NN_PAD_MIRROR_REFLECT,
+   NN_PAD_MIRROR_SYMMETRIC,
+   NN_PAD_SAME_CAFFE,
+   _32BIT_PLACEHOLDER_hexagon_nn_padding_type = 0x7fffffff
+};
+typedef enum hexagon_nn_padding_type hexagon_nn_padding_type;
+enum hexagon_nn_corner_type {
+   NN_CORNER_RELEASE,
+   NN_CORNER_TURBO,
+   NN_CORNER_NOMPLUS,
+   NN_CORNER_NOMINAL,
+   NN_CORNER_SVSPLUS,
+   NN_CORNER_SVS,
+   NN_CORNER_SVS2,
+   _32BIT_PLACEHOLDER_hexagon_nn_corner_type = 0x7fffffff
+};
+typedef enum hexagon_nn_corner_type hexagon_nn_corner_type;
+enum hexagon_nn_dcvs_type {
+   NN_DCVS_DEFAULT,
+   NN_DCVS_ENABLE,
+   NN_DCVS_DISABLE,
+   _32BIT_PLACEHOLDER_hexagon_nn_dcvs_type = 0x7fffffff
+};
+typedef enum hexagon_nn_dcvs_type hexagon_nn_dcvs_type;
+typedef struct hexagon_nn_tensordef hexagon_nn_tensordef;
+struct hexagon_nn_tensordef {
+   unsigned int batches;
+   unsigned int height;
+   unsigned int width;
+   unsigned int depth;
+   unsigned char* data;
+   int dataLen;
+   unsigned int data_valid_len;
+   unsigned int unused;
+};
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_config)(void) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_get_dsp_offset)(unsigned int* libhexagon_addr, unsigned int* fastrpc_shell_addr) __QAIC_HEADER_ATTRIBUTE;
+__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;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_snpprint)(hexagon_nn_nn_id id, unsigned char* buf, int bufLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_getlog)(hexagon_nn_nn_id id, unsigned char* buf, int bufLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_append_node)(hexagon_nn_nn_id id, unsigned int node_id, unsigned int operation, hexagon_nn_padding_type padding, const hexagon_nn_input* inputs, int inputsLen, const hexagon_nn_output* outputs, int outputsLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_append_const_node)(hexagon_nn_nn_id id, unsigned int node_id, unsigned int batches, unsigned int height, unsigned int width, unsigned int depth, const unsigned char* data, int dataLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_append_empty_const_node)(hexagon_nn_nn_id id, unsigned int node_id, unsigned int batches, unsigned int height, unsigned int width, unsigned int depth, unsigned int size) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_populate_const_node)(hexagon_nn_nn_id id, unsigned int node_id, const unsigned char* data, int dataLen, unsigned int target_offset) __QAIC_HEADER_ATTRIBUTE;
+typedef struct hexagon_nn_op_node hexagon_nn_op_node;
+struct hexagon_nn_op_node {
+   unsigned int node_id;
+   unsigned int operation;
+   hexagon_nn_padding_type padding;
+   hexagon_nn_input* inputs;
+   int inputsLen;
+   hexagon_nn_output* outputs;
+   int outputsLen;
+};
+typedef struct hexagon_nn_const_node hexagon_nn_const_node;
+struct hexagon_nn_const_node {
+   unsigned int node_id;
+   hexagon_nn_tensordef tensor;
+};
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_append_node_list)(hexagon_nn_nn_id id, const hexagon_nn_op_node* ops, int opsLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_append_const_node_list)(hexagon_nn_nn_id id, const hexagon_nn_const_node* consts, int constsLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_prepare)(hexagon_nn_nn_id id) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_execute)(hexagon_nn_nn_id id, unsigned int batches_in, unsigned int height_in, unsigned int width_in, unsigned int depth_in, const unsigned char* data_in, int data_inLen, unsigned int* batches_out, unsigned int* height_out, unsigned int* width_out, unsigned int* depth_out, unsigned char* data_out, int data_outLen, unsigned int* data_len_out) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_teardown)(hexagon_nn_nn_id id) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_powersave_level)(unsigned int level) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_powersave_details)(hexagon_nn_corner_type corner, hexagon_nn_dcvs_type dcvs, unsigned int latency) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_get_perfinfo)(hexagon_nn_nn_id id, hexagon_nn_perfinfo* info_out, int info_outLen, unsigned int* n_items) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_reset_perfinfo)(hexagon_nn_nn_id id, unsigned int event) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_last_execution_cycles)(hexagon_nn_nn_id id, unsigned int* cycles_lo, unsigned int* cycles_hi) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_version)(int* ver) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_op_name_to_id)(const char* name, unsigned int* node_id) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_op_id_to_name)(unsigned int node_id, char* name, int nameLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_disable_dcvs)(void) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_GetHexagonBinaryVersion)(int* ver) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_PrintLog)(const unsigned char* buf, int bufLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_execute_new)(hexagon_nn_nn_id id, const hexagon_nn_tensordef* inputs, int inputsLen, hexagon_nn_tensordef* outputs, int outputsLen) __QAIC_HEADER_ATTRIBUTE;
+#ifdef __cplusplus
+}
+#endif
+
+#else  // nnlib version : MACE_USE_NNLIB_CAF
 
 #ifndef __QAIC_HEADER
 #define __QAIC_HEADER(ff) ff

third_party/nnlib/ops.h

@@ -218,7 +218,231 @@ DEF_OP(QuantizedBiasAdd_8p8to8)
 #undef DEF_OP_WREF
 #endif
 
-#else  // nnlib version
+#elif defined(MACE_USE_NNLIB_2_1)  // nnlib version
+
+DEF_OP(INPUT)
+DEF_OP(OUTPUT)
+DEF_OP(Nop)
+DEF_OP(Const)
+DEF_OP(Check)
+DEF_OP(Close_f)
+DEF_OP(Close_quint8)
+DEF_OP(Close_q_quint8)
+DEF_OP(Close_int32)
+DEF_OP(Close_qint32)
+DEF_OP(PPrint_8)
+DEF_OP(PPrint_32)
+DEF_OP(PPrint_f)
+DEF_OP(PreFree)
+DEF_OP(Flatten)
+
+#ifndef DEF_OP_WREF
+#define DEF_OP_WREF(NAME) DEF_OP(NAME) DEF_OP(NAME##_ref)
+#define __SELF_DEF_OP_WREF
+#endif
+
+DEF_OP_WREF(QuantizedConv2d_8x8to32)
+DEF_OP_WREF(QuantizedMatMul_8x8to32)
+DEF_OP_WREF(QuantizeDownAndShrinkRange_32to8)
+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)
+DEF_OP_WREF(Max_f)
+DEF_OP_WREF(Quantize)
+DEF_OP_WREF(Dequantize)
+DEF_OP_WREF(Supernode_8x8p8to8)
+
+DEF_OP(QuantizedFlatten)
+DEF_OP(Softmax_f)
+DEF_OP(Conv2d_f)
+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)
+DEF_OP(LRN_f)
+
+DEF_OP(Variable)
+DEF_OP(Assign)
+DEF_OP(Reshape)
+DEF_OP(QuantizedReshape)
+DEF_OP(Tanh_f)
+DEF_OP(Sigmoid_f)
+DEF_OP(Slice_8)
+DEF_OP(Slice_f)
+DEF_OP(QuantizedSlice_8)
+DEF_OP(Add_f)
+DEF_OP(Mul_f)
+DEF_OP(Minimum_f)
+DEF_OP(Maximum_f)
+
+DEF_OP_WREF(Requantize_32to8)
+DEF_OP_WREF(RequantizationRange_32)
+
+DEF_OP(Neg_f)
+DEF_OP(Sub_f)
+DEF_OP(AddN_f)
+DEF_OP(Range_int32)
+DEF_OP(Rank_int32)
+DEF_OP(Transpose_int32)
+DEF_OP(Transpose_f)
+DEF_OP(InstanceNorm_f)
+DEF_OP_WREF(QuantizedInstanceNorm_8)
+DEF_OP(Sub_int32)
+DEF_OP(Add_int32)
+DEF_OP(Split_f)
+DEF_OP(Dequantize_qint32_f)
+DEF_OP(PRelu_f)
+DEF_OP_WREF(QuantizedPRelu_8)
+DEF_OP(Sum_f)
+DEF_OP(Prod_f)
+DEF_OP(Mul_int32)
+DEF_OP(LogicalAnd_int32)
+DEF_OP(LogicalOr_int32)
+DEF_OP(LogicalXor_int32)
+DEF_OP(Shape_int32)
+DEF_OP(Pack_int32)
+DEF_OP(MirrorPad_f)
+DEF_OP(ResizeNearestNeighbor_f)
+DEF_OP(StridedSlice_int32)
+DEF_OP(StridedSlice_f)
+DEF_OP(ExpandDims_int32)
+DEF_OP(ExpandDims_f)
+
+DEF_OP(LogSoftmax_f)
+DEF_OP(Split_int32)
+DEF_OP(QuantizedSplit_8)
+
+DEF_OP(Deconv_f)
+DEF_OP_WREF(QuantizedDeconv_8x8to32)
+
+DEF_OP_WREF(QuantizedMul_8x8to32)
+DEF_OP_WREF(QuantizedAdd_8p8to32)
+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)
+
+DEF_OP(Pad_f)
+DEF_OP(SpaceToBatchND_f)
+DEF_OP(BatchToSpaceND_f)
+DEF_OP(QuantizedPad_8)
+DEF_OP(ResizeBilinear_f)
+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_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)
+
+DEF_OP(DequantizeOUTPUT_8tof)
+DEF_OP(QuantizedBatchNorm_8x8p8to8)
+DEF_OP(QuantizedBatchNorm_8x8p32to8)
+DEF_OP(QuantizedBatchNorm_8x8p8to8_d32)
+DEF_OP(QuantizedBatchNorm_8x8p32to8_d32)
+
+DEF_OP_WREF(QuantizedInstanceNorm_8_d32)
+DEF_OP_WREF(QuantizedInstanceNormBG_8)
+DEF_OP_WREF(QuantizedInstanceNormBG_8_d32)
+
+
+DEF_OP(ChannelShuffle_f)
+DEF_OP(ChannelShuffle_int32)
+DEF_OP(QuantizedChannelShuffle_8)
+
+#ifdef __SELF_DEF_OP_WREF
+#undef __SELF_DEF_OP_WREF
+#undef DEF_OP_WREF
+#endif
+
+#else  // nnlib version : MACE_USE_NNLIB_CAF
 
 DEF_OP(INPUT)
 DEF_OP(OUTPUT)

tools/bazel.rc

@@ -10,6 +10,7 @@ build --copt=-std=c++11
 build --copt=-D_GLIBCXX_USE_C99_MATH_TR1
 build --copt=-DMACE_OBFUSCATE_LITERALS
 build --copt=-DGEMMLOWP_USE_OPENMP
+build --copt=-DMACE_USE_NNLIB_CAF
 
 # Usage example: bazel build --config android
 build:android --crosstool_top=//external:android/crosstool