Update

[ghstack-poisoned]

Author: mcr229

backends/xnnpack/runtime/XNNCompiler.cpp

@@ -97,7 +97,10 @@ std::pair<float, float> getOutputMinMax(const NodePtr node) noexcept {
 }
 
 /*
-Converts flatbuffer xnn data type to xnnpack data type
+Converts flatbuffer xnn data type to xnnpack data type. 
+
+NOTE:
+Flatbuffer Enum Values are not the same as XNNPACK's datatype enum values.
 */
 xnn_datatype getDataType(const DataType& data_type) {
   switch (data_type) {
@@ -121,6 +124,14 @@ xnn_datatype getDataType(const DataType& data_type) {
       return xnn_datatype::xnn_datatype_qdint8;
     case DataType::xnn_datatype_qbint4:
       return xnn_datatype::xnn_datatype_qbint4;
+    case DataType::xnn_datatype_qpint8:
+      return xnn_datatype::xnn_datatype_qpint8;
+    case DataType::xnn_datatype_int32:
+      return xnn_datatype::xnn_datatype_int32;
+    case DataType::xnn_datatype_pfp32:
+      return xnn_datatype::xnn_datatype_pfp32;
+    case DataType::xnn_datatype_bf16:
+      return xnn_datatype::xnn_datatype_bf16;
     default:
       return xnn_datatype::xnn_datatype_invalid;
   }

backends/xnnpack/serialization/runtime_schema.fbs

@@ -29,6 +29,15 @@ enum XNNDatatype : short {
   xnn_datatype_qdint8 = 9,
   /// Quantized 4-bit signed integer with shared blockwise quantization parameters.
   xnn_datatype_qbint4 = 10,
+  /// Dynamically quantized 8-bit signed integers packed with their per-row
+  /// quantization parameters.
+  xnn_datatype_qpint8 = 11,
+  /// 32-bit signed integers.
+  xnn_datatype_int32 = 12,
+  /// IEEE754 single-precision packed floating-point.
+  xnn_datatype_pfp32 = 13,
+  /// BFloat16, i.e. the upper 16 bits of a float32.
+  xnn_datatype_bf16 = 14,
 }
 
 // type of quantization

backends/xnnpack/serialization/schema.fbs

@@ -29,6 +29,15 @@ enum XNNDatatype : short {
   xnn_datatype_qdint8 = 9,
   /// Quantized 4-bit signed integer with shared blockwise quantization parameters.
   xnn_datatype_qbint4 = 10,
+  /// Dynamically quantized 8-bit signed integers packed with their per-row
+  /// quantization parameters.
+  xnn_datatype_qpint8 = 11,
+  /// 32-bit signed integers.
+  xnn_datatype_int32 = 12,
+  /// IEEE754 single-precision packed floating-point.
+  xnn_datatype_pfp32 = 13,
+  /// BFloat16, i.e. the upper 16 bits of a float32.
+  xnn_datatype_bf16 = 14,
 }
 
 // type of quantization

backends/xnnpack/serialization/xnnpack_graph_schema.py

@@ -413,6 +413,10 @@ class XNNDatatype(IntEnum):
     xnn_datatype_qcint4 = 8
     xnn_datatype_qdint8 = 9
     xnn_datatype_qbint4 = 10
+    xnn_datatype_qpint8 = 11
+    xnn_datatype_int32 = 12
+    xnn_datatype_pfp32 = 13
+    xnn_datatype_bf16 = 14
 
 
 @dataclass