Added UINT4 support (Sync from github.com/tensorflow/tensorflow) (#3283)

* Added UINT4 support

* Sync from upstream TF.

---------

Co-authored-by: TFLM-bot <tflm-github-bot@google.com>

Author: veblush

python/tflite_micro/numpy_utils.cc

@@ -55,6 +55,9 @@ int TfLiteTypeToPyArrayType(TfLiteType tf_lite_type) {
       return NPY_INT16;
     case kTfLiteUInt8:
       return NPY_UINT8;
+    case kTfLiteUInt4:
+      // TODO(b/246806634): NPY_UINT4 currently doesn't exist
+      return NPY_BYTE;
     case kTfLiteInt4:
       // TODO(b/246806634): NPY_INT4 currently doesn't exist
       return NPY_BYTE;

tensorflow/compiler/mlir/lite/core/c/tflite_types.h

@@ -65,6 +65,7 @@ typedef enum {
   kTfLiteInt4 = 18,
   kTfLiteBFloat16 = 19,
   kTfLiteInt2 = 20,
+  kTfLiteUInt4 = 21,
 } TfLiteType;
 // LINT.ThenChange(//tensorflow/lite/profiling/proto/model_runtime_info.proto:EdgeDataType)
 

tensorflow/compiler/mlir/lite/schema/schema.fbs

@@ -62,6 +62,7 @@ enum TensorType : byte {
   INT4 = 17,
   BFLOAT16 = 18,
   INT2 = 19,
+  UINT4 = 20,
 }
 
 // Custom quantization parameters for experimenting with new quantization

tensorflow/lite/core/api/flatbuffer_conversions.cc

@@ -1098,6 +1098,9 @@ TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type,
     case TensorType_INT2:
       *type = kTfLiteInt2;
       return kTfLiteOk;
+    case TensorType_UINT4:
+      *type = kTfLiteUInt4;
+      return kTfLiteOk;
     default:
       *type = kTfLiteNoType;
       TF_LITE_REPORT_ERROR(error_reporter,

tensorflow/lite/core/c/common.cc

@@ -511,6 +511,8 @@ const char* TfLiteTypeGetName(TfLiteType type) {
       return "INT4";
     case kTfLiteInt2:
       return "INT2";
+    case kTfLiteUInt4:
+      return "UINT4";
   }
   return "Unknown type";
 }

tensorflow/lite/micro/tools/layer_by_layer.cc

@@ -117,6 +117,9 @@ TfLiteStatus ConvertTensorType(TfLiteType type, TensorTypes& tensor_type) {
     case kTfLiteVariant:
       tensor_type = TensorTypes_VARIANT;
       return kTfLiteOk;
+    case kTfLiteUInt4:
+      tensor_type = TensorTypes_UINT4;
+      return kTfLiteOk;
     case kTfLiteInt4:
       tensor_type = TensorTypes_INT4;
       return kTfLiteOk;

tensorflow/lite/micro/tools/layer_by_layer_schema.fbs

@@ -36,6 +36,7 @@ enum TensorTypes : byte {
   INT4 = 17,
   BFLOAT16 = 18,
   INT2 = 19,
+  UINT4 = 20,
 }
 
 table TensorData {

tensorflow/lite/micro/tools/layer_by_layer_schema_generated.h

@@ -60,11 +60,12 @@ enum TensorTypes : int8_t {
   TensorTypes_INT4 = 17,
   TensorTypes_BFLOAT16 = 18,
   TensorTypes_INT2 = 19,
+  TensorTypes_UINT4 = 20,
   TensorTypes_MIN = TensorTypes_FLOAT32,
-  TensorTypes_MAX = TensorTypes_INT2
+  TensorTypes_MAX = TensorTypes_UINT4
 };
 
-inline const TensorTypes (&EnumValuesTensorTypes())[20] {
+inline const TensorTypes (&EnumValuesTensorTypes())[21] {
   static const TensorTypes values[] = {
     TensorTypes_FLOAT32,
     TensorTypes_FLOAT16,
@@ -85,13 +86,14 @@ inline const TensorTypes (&EnumValuesTensorTypes())[20] {
     TensorTypes_UINT16,
     TensorTypes_INT4,
     TensorTypes_BFLOAT16,
-    TensorTypes_INT2
+    TensorTypes_INT2,
+    TensorTypes_UINT4
   };
   return values;
 }
 
 inline const char * const *EnumNamesTensorTypes() {
-  static const char * const names[21] = {
+  static const char * const names[22] = {
     "FLOAT32",
     "FLOAT16",
     "INT32",
@@ -112,13 +114,14 @@ inline const char * const *EnumNamesTensorTypes() {
     "INT4",
     "BFLOAT16",
     "INT2",
+    "UINT4",
     nullptr
   };
   return names;
 }
 
 inline const char *EnumNameTensorTypes(TensorTypes e) {
-  if (::flatbuffers::IsOutRange(e, TensorTypes_FLOAT32, TensorTypes_INT2)) return "";
+  if (::flatbuffers::IsOutRange(e, TensorTypes_FLOAT32, TensorTypes_UINT4)) return "";
   const size_t index = static_cast<size_t>(e);
   return EnumNamesTensorTypes()[index];
 }

tensorflow/lite/python/schema_py_generated.py

@@ -28,6 +28,7 @@ class TensorType(object):
     INT4 = 17
     BFLOAT16 = 18
     INT2 = 19
+    UINT4 = 20
 
 
 class QuantizationDetails(object):

tensorflow/lite/schema/schema_generated.h

@@ -712,11 +712,12 @@ enum TensorType : int8_t {
   TensorType_INT4 = 17,
   TensorType_BFLOAT16 = 18,
   TensorType_INT2 = 19,
+  TensorType_UINT4 = 20,
   TensorType_MIN = TensorType_FLOAT32,
-  TensorType_MAX = TensorType_INT2
+  TensorType_MAX = TensorType_UINT4
 };
 
-inline const TensorType (&EnumValuesTensorType())[20] {
+inline const TensorType (&EnumValuesTensorType())[21] {
   static const TensorType values[] = {
     TensorType_FLOAT32,
     TensorType_FLOAT16,
@@ -737,13 +738,14 @@ inline const TensorType (&EnumValuesTensorType())[20] {
     TensorType_UINT16,
     TensorType_INT4,
     TensorType_BFLOAT16,
-    TensorType_INT2
+    TensorType_INT2,
+    TensorType_UINT4
   };
   return values;
 }
 
 inline const char * const *EnumNamesTensorType() {
-  static const char * const names[21] = {
+  static const char * const names[22] = {
     "FLOAT32",
     "FLOAT16",
     "INT32",
@@ -764,13 +766,14 @@ inline const char * const *EnumNamesTensorType() {
     "INT4",
     "BFLOAT16",
     "INT2",
+    "UINT4",
     nullptr
   };
   return names;
 }
 
 inline const char *EnumNameTensorType(TensorType e) {
-  if (::flatbuffers::IsOutRange(e, TensorType_FLOAT32, TensorType_INT2)) return "";
+  if (::flatbuffers::IsOutRange(e, TensorType_FLOAT32, TensorType_UINT4)) return "";
   const size_t index = static_cast<size_t>(e);
   return EnumNamesTensorType()[index];
 }