Adding quantized conv operators kernel optimization

backends/cadence/aot/functions_hifi.yaml

@@ -125,6 +125,11 @@
     - arg_meta: null
       kernel_name: impl::HiFi::dequantize_per_tensor_out
 
+- func: cadence::quantized_conv.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::quantized_conv_out
 
 - func: cadence::quantized_layer_norm.out(Tensor input, Tensor in_scale, Tensor in_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor(a!)
   kernels:

backends/cadence/hifi/kernels/CMakeLists.txt

@@ -9,6 +9,9 @@ add_library(
   cadence_kernels
   kernels.cpp
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/matmul_asym8uxasym8u_asym8u.cpp
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_conv2d_asym8xasym8.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_conv2d_std_circ_buf.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_matXvec_asym8xasym8_asym8_circ.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_add_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_mode_f32_broadcast.c
@@ -28,6 +31,7 @@ target_include_directories(
     ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/nnlib-hifi4/xa_nnlib/include/nnlib
     ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/nnlib-hifi4/xa_nnlib/include
     ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/nnlib-hifi4/xa_nnlib/algo/ndsp/hifi4/include/
+    ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/nnlib-hifi4/xa_nnlib/algo/kernels/cnn/hifi4/
 	${_common_include_directories}
 )
 

backends/cadence/hifi/kernels/kernels.h

@@ -16,6 +16,32 @@
 #include "xa_nnlib_kernels_api.h"
 
 /* Potential NNLIB function/APIs */
+extern "C" WORD32 xa_nn_conv2d_per_chan_asym8xasym8(UWORD8* __restrict__ p_out,
+                                const UWORD8* __restrict__ p_inp,
+                                const UWORD8* __restrict__ p_kernel,
+                                const WORD32* __restrict__ p_bias,
+                                WORD32 input_height,
+                                WORD32 input_width,
+                                WORD32 input_channels,
+                                WORD32 kernel_height,
+                                WORD32 kernel_width,
+                                WORD32 kernel_channels,
+                                WORD32 dilation_height,
+                                WORD32 dilation_width,
+                                WORD32 out_channels,
+                                WORD32 x_stride,
+                                WORD32 y_stride,
+                                WORD32 x_padding,
+                                WORD32 y_padding,
+                                WORD32 out_height,
+                                WORD32 out_width,
+                                WORD32 input_zero_bias,
+                                WORD32 * p_out_multiplier,
+                                WORD32 * p_out_shift,
+                                WORD32 out_zero_bias,
+                                WORD32 out_data_format,
+                                VOID *p_scratch);
+
 extern "C" WORD32 xa_nn_elm_add_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
                                 const WORD32 *const p_out_shape,
                                 const FLOAT32 * __restrict__ p_inp1,

backends/cadence/hifi/operators/CMakeLists.txt

@@ -66,6 +66,7 @@ target_include_directories(
 add_library(
   custom_ops "quantized_linear_out.cpp" "quantized_layer_norm.cpp"
              "quantize_per_tensor.cpp" "dequantize_per_tensor.cpp"
+             "quantized_conv_out.cpp"
 )
 target_include_directories(
   custom_ops PUBLIC ${ROOT_DIR}/.. ${CMAKE_BINARY_DIR}