Add UInt16 support to Cadence kernels

Summary:
In preparation for using uint16 on Cadence, add support to the quant kernels.

Same as: pytorch#6724, but added UInt16 as well
as Bits16.

Differential Revision: D66016288

Author: dulinriley

backends/cadence/hifi/operators/dequantize_per_tensor.cpp

@@ -41,7 +41,7 @@ void dequantize_per_tensor_out(
   } else if (input.scalar_type() == ScalarType::Short) {
     const int16_t* input_data = input.const_data_ptr<int16_t>();
     dequantize<int16_t>(out_data, input_data, scale, zero_point, numel);
-  } else if (input.scalar_type() == ScalarType::Bits16) {
+  } else if (input.scalar_type() == ScalarType::Bits16 || input.scalar_type() == ScalarType::UInt16) {
     const uint16_t* input_data = input.const_data_ptr<uint16_t>();
     dequantize<uint16_t>(out_data, input_data, scale, zero_point, numel);
   } else if (input.scalar_type() == ScalarType::Int) {

backends/cadence/hifi/operators/quantize_per_tensor.cpp

@@ -44,7 +44,7 @@ void quantize_per_tensor_out(
     int16_t* out_data = out.mutable_data_ptr<int16_t>();
     cadence::impl::HiFi::kernels::quantize<int16_t>(
         out_data, input_data, 1. / scale, zero_point, numel);
-  } else if (out.scalar_type() == ScalarType::Bits16) {
+  } else if (out.scalar_type() == ScalarType::Bits16 || out.scalar_type() == ScalarType::UInt16) {
     uint16_t* out_data = out.mutable_data_ptr<uint16_t>();
     cadence::impl::HiFi::kernels::quantize<uint16_t>(
         out_data, input_data, 1. / scale, zero_point, numel);

backends/cadence/reference/operators/dequantize_per_tensor.cpp

@@ -37,7 +37,7 @@ void dequantize_per_tensor_out(
     const int8_t* input_data = input.const_data_ptr<int8_t>();
     impl::reference::kernels::dequantize<int8_t>(
         out_data, input_data, scale, zero_point, numel);
-  } else if (input.scalar_type() == ScalarType::Bits16) {
+  } else if (input.scalar_type() == ScalarType::Bits16 || input.scalar_type() == ScalarType::UInt16) {
     const uint16_t* input_data = input.const_data_ptr<uint16_t>();
     impl::reference::kernels::dequantize<uint16_t>(
         out_data, input_data, scale, zero_point, numel);

backends/cadence/reference/operators/quantize_per_tensor.cpp

@@ -39,7 +39,7 @@ void quantize_per_tensor_out(
     int8_t* out_data = out.mutable_data_ptr<int8_t>();
     impl::reference::kernels::quantize<int8_t>(
         out_data, input_data, 1. / scale, zero_point, numel);
-  } else if (out.scalar_type() == ScalarType::Bits16) {
+  } else if (out.scalar_type() == ScalarType::Bits16 || out.scalar_type() == ScalarType::UInt16) {
     uint16_t* out_data = out.mutable_data_ptr<uint16_t>();
     impl::reference::kernels::quantize<uint16_t>(
         out_data, input_data, 1. / scale, zero_point, numel);