migrate jarvis quant-per-tensor hifi ops to oss (pytorch#6293)

Summary:

- only migrated quant-per-tensor in this diff. will do the rest of 3 in the stack
- solved the --start-lib --end-lib option not recognized in here - need to import libs from //executorch. Nothing do with the cxx wrapper in buck
- aligned namespace to cadence::impl:HIFI::native
- kernel to be removed after all ops are migrated

Reviewed By: skrtskrtfb

Differential Revision: D64194227

Author: zonglinpeng

backends/cadence/hifi/kernels/kernels.cpp

@@ -10,6 +10,7 @@
 #include <xa_nnlib_common.h>
 #include <xa_nnlib_common_macros.h>
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace kernels {
@@ -231,3 +232,4 @@ typed_requantize_vec(uint8_t, int8_t);
 }; // namespace kernels
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/kernels/kernels.h

@@ -12,6 +12,7 @@
 #include <stddef.h>
 #include <xa_type_def.h>
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace kernels {
@@ -63,3 +64,4 @@ void dequantize(
 }; // namespace kernels
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/operators/dequantize_per_tensor.cpp

@@ -10,6 +10,7 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <xa_nnlib_kernels_api.h>
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace native {
@@ -50,3 +51,4 @@ void dequantize_per_tensor_out(
 }; // namespace native
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/operators/quantize_per_tensor.cpp

@@ -10,6 +10,7 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <xa_nnlib_kernels_api.h>
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace native {
@@ -21,16 +22,16 @@ using executorch::runtime::KernelRuntimeContext;
 // Quantize the input tensor (PT2 version). Note that quant_<min,max> are not
 // used in any computation.
 void quantize_per_tensor_out(
-    KernelRuntimeContext& context,
+    KernelRuntimeContext& ctx,
     const Tensor& input,
     double scale,
     int64_t zero_point,
-    int64_t quant_min,
-    int64_t quant_max,
+    __ET_UNUSED int64_t quant_min,
+    __ET_UNUSED int64_t quant_max,
     ScalarType dtype,
     Tensor& out) {
   const float* input_data = input.const_data_ptr<float>();
-  size_t numel = out.numel();
+  const size_t numel = out.numel();
 
   if (out.scalar_type() == ScalarType::Byte) {
     uint8_t* out_data = out.mutable_data_ptr<uint8_t>();
@@ -40,6 +41,10 @@ void quantize_per_tensor_out(
     int8_t* out_data = out.mutable_data_ptr<int8_t>();
     xa_nn_elm_quantize_f32_asym8s(
         out_data, input_data, scale, zero_point, numel);
+  } else if (out.scalar_type() == ScalarType::Short) {
+    int16_t* out_data = out.mutable_data_ptr<int16_t>();
+    impl::HiFi::kernels::quantize<int16_t>(
+        out_data, input_data, 1. / scale, zero_point, numel);
   } else if (out.scalar_type() == ScalarType::Int) {
     int32_t* out_data = out.mutable_data_ptr<int32_t>();
     impl::HiFi::kernels::quantize<int32_t>(
@@ -52,3 +57,4 @@ void quantize_per_tensor_out(
 }; // namespace native
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/operators/quantized_layer_norm.cpp

@@ -16,6 +16,7 @@ using executorch::aten::Tensor;
 using executorch::runtime::getLeadingDims;
 using executorch::runtime::KernelRuntimeContext;
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace native {
@@ -157,3 +158,4 @@ void quantized_layer_norm_out(
 }; // namespace native
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/operators/quantized_linear_out.cpp

@@ -11,6 +11,7 @@
 #include <algorithm>
 #include <cmath>
 
+namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace native {
@@ -69,3 +70,4 @@ void quantized_linear_out(
 }; // namespace native
 }; // namespace HiFi
 }; // namespace impl
+}; // namespace cadence

backends/cadence/hifi/operators/targets.bzl

@@ -26,5 +26,6 @@ def define_common_targets():
         ],
         visibility = [
             "//executorch/backends/cadence/...",
+            "@EXECUTORCH_CLIENTS",
         ],
     )