Cadence HiFi 5 Neural Network Library v1.6.0 (#276)

* Cadence HiFi 5 Neural Network Library v1.6.0
updated the download script to use hifi5 v1.6.0 archive.
depthwise_conv kernel is updated and patch is not required now.
quantize optimized kernel is renamed in the v1.6.0, updated with new
kernel name.

* Fix formatting.
Co-authored-by: Nbhanu prakash bandaru venkata <bhanup@cadence.com>
Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
Co-authored-by: NAdvait Jain <advaitjain@google.com>

tensorflow/lite/micro/kernels/xtensa/quantize.cc

-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -117,7 +117,7 @@ TfLiteStatus EvalXtensa(TfLiteContext* context, TfLiteNode* node) {
   auto* op_data = static_cast<OpData*>(node->user_data);
 #elif defined(FUSION_F1) || defined(HIFI5)
   auto* op_data = static_cast<OpDataQuantizeReference*>(node->user_data);
-#endif
+#endif  // defined(HIFIMINI) || defined(FUSION_F1) || defined(HIFI5)
 
   const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0);
   TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0);
@@ -129,7 +129,7 @@ TfLiteStatus EvalXtensa(TfLiteContext* context, TfLiteNode* node) {
                    tflite::micro::GetTensorData<int16_t>(input),
                    tflite::micro::GetTensorShape(output),
                    tflite::micro::GetTensorData<int8_t>(output));
-#elif defined(FUSION_F1) || defined(HIFI5)
+#elif defined(FUSION_F1)
     int size = ElementCount(*input->dims);
     TF_LITE_ENSURE_EQ(
         context,
@@ -140,9 +140,20 @@ TfLiteStatus EvalXtensa(TfLiteContext* context, TfLiteNode* node) {
             op_data->requantize_output_shift,
             op_data->requantize_output_multiplier, size),
         0);
+#elif defined(HIFI5)
+    int size = ElementCount(*input->dims);
+    TF_LITE_ENSURE_EQ(
+        context,
+        xa_nn_elm_requantize_asym16s_asym8s(
+            tflite::micro::GetTensorData<int8_t>(output),
+            tflite::micro::GetTensorData<int16_t>(input),
+            op_data->input_zero_point, op_data->quantization_params.zero_point,
+            op_data->requantize_output_shift,
+            op_data->requantize_output_multiplier, size),
+        0);
 #else
     static_assert(false, "Unsupported xtensa architecture.");
-#endif
+#endif  // defined(HIFIMINI) || defined(FUSION_F1) || defined(HIFI5)
   } else if (output->type == kTfLiteInt32 &&
              (input->type == kTfLiteInt16 || input->type == kTfLiteInt8)) {
     int size = ElementCount(*input->dims);
@@ -154,19 +165,47 @@ TfLiteStatus EvalXtensa(TfLiteContext* context, TfLiteNode* node) {
     int32_t zero_point = op_data->zero_point;
 #elif defined(FUSION_F1) || defined(HIFI5)
     int32_t zero_point = op_data->quantization_params.zero_point;
-#endif
+#endif  // defined(HIFIMINI) || defined(FUSION_F1) || defined(HIFI5)
     if (input->type == kTfLiteInt16) {
+#if defined(HIFI5)
+      int size = ElementCount(*input->dims);
+      TF_LITE_ENSURE_EQ(context,
+                        xa_nn_elm_requantize_asym16s_asym32s(
+                            tflite::micro::GetTensorData<int32_t>(output),
+                            tflite::micro::GetTensorData<int16_t>(input),
+                            op_data->input_zero_point,
+                            op_data->quantization_params.zero_point,
+                            op_data->requantize_output_shift,
+                            op_data->requantize_output_multiplier, size),
+                        0);
+#else
       reference_ops::Requantize(tflite::micro::GetTensorData<int16_t>(input),
                                 size, op_data->requantize_output_multiplier,
                                 op_data->requantize_output_shift,
                                 op_data->input_zero_point, zero_point,
                                 tflite::micro::GetTensorData<int32_t>(output));
+#endif  // defined(HIFI5)
     } else {
+#if defined(HIFI5)
+      const int8_t* input_data_ptr;
+      int32_t* output_data_ptr;
+      input_data_ptr = tflite::micro::GetTensorData<int8_t>(input);
+      output_data_ptr = tflite::micro::GetTensorData<int32_t>(output);
+
+      TF_LITE_ENSURE_EQ(
+          context,
+          xa_nn_elm_requantize_asym8s_asym32s(
+              output_data_ptr, input_data_ptr, op_data->input_zero_point,
+              zero_point, op_data->requantize_output_shift,
+              op_data->requantize_output_multiplier, size),
+          0);
+#else
       reference_ops::Requantize(tflite::micro::GetTensorData<int8_t>(input),
                                 size, op_data->requantize_output_multiplier,
                                 op_data->requantize_output_shift,
                                 op_data->input_zero_point, zero_point,
                                 tflite::micro::GetTensorData<int32_t>(output));
+#endif  // defined(HIFI5)
     }
   } else {
     TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.",
@@ -185,7 +224,7 @@ void* Init(TfLiteContext* context, const char* buffer, size_t length) {
 #else
   return context->AllocatePersistentBuffer(context,
                                            sizeof(OpDataQuantizeReference));
-#endif
+#endif  // defined(HIFIMINI)
 }
 
 TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
@@ -205,7 +244,7 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
   op_data->quantization_params.zero_point = output->params.zero_point;
   op_data->quantization_params.scale =
       static_cast<double>(output->params.scale);
-#endif
+#endif  // defined(HIFIMINI)
 
   op_data->input_zero_point = input->params.zero_point;
 
@@ -222,7 +261,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   return EvalXtensa(context, node);
 #else
   return EvalQuantizeReference(context, node);
-#endif
+#endif  // defined(HIFIMINI) || defined(FUSION_F1) || defined(HIFI5)
 }
 
 }  // namespace

tensorflow/lite/micro/tools/make/ext_libs/xtensa_depthwise_patch_hifi5.patch

-diff --git a/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise.c b/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise.c
-index 48fccdc..3c10040 100644
---- a/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise.c
-+++ b/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise.c
-@@ -348,8 +348,6 @@ WORD32 xa_nn_conv2d_depthwise_getsize
-   XA_NNLIB_CHK_COND((kernel_height <= 0), -1);
-   XA_NNLIB_CHK_COND((kernel_width <= 0), -1);
-   XA_NNLIB_CHK_COND((channels_multiplier <= 0), -1);
--  XA_NNLIB_CHK_COND((x_stride <= 0 || x_stride > kernel_width), -1); //TODO: x_stride > kernel_width is supported ?
--  XA_NNLIB_CHK_COND((y_stride <= 0 || y_stride > kernel_height), -1);
-   XA_NNLIB_CHK_COND((x_padding < 0), -1);
-   XA_NNLIB_CHK_COND((y_padding < 0), -1);
-   XA_NNLIB_CHK_COND((output_height <= 0), -1);
-diff --git a/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise_sym8sxasym8s.c b/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-index 71611f4..0fd19ae 100644
---- a/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-+++ b/algo/kernels/cnn/hifi5/xa_nn_conv2d_depthwise_sym8sxasym8s.c
-@@ -2122,10 +2122,7 @@ WORD32 xa_nn_conv2d_depthwise_nhwc_per_chan_sym8sxasym8s_k3x3
-   /* Implementation dependent checks */
-   //TOOD: support y_stride 2
-   XA_NNLIB_ARG_CHK_COND((y_stride != 1) && (y_stride != 2), -1);
--  XA_NNLIB_ARG_CHK_COND((y_stride > kernel_height), -1);
--  XA_NNLIB_ARG_CHK_COND((x_stride > kernel_width), -1);
-   XA_NNLIB_ARG_CHK_COND((kernel_height > input_height), -1);
--  XA_NNLIB_ARG_CHK_COND((kernel_width > input_width), -1);
- 
- #ifndef DISABLE_DEPTHWISE_CONV2D_K3X3_SPECIAL_CASE
-   WORD32 input_zero_bias_neg = -input_zero_bias;
-@@ -2271,7 +2268,6 @@ WORD32 xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s_generic
-   XA_NNLIB_ARG_CHK_COND((input_channels <= 0), -1);
-   XA_NNLIB_ARG_CHK_COND((kernel_height <= 0 || kernel_width <= 0), -1);
-   XA_NNLIB_ARG_CHK_COND((kernel_height > input_height), -1);
--  XA_NNLIB_ARG_CHK_COND((kernel_width > input_width), -1);
-   XA_NNLIB_ARG_CHK_COND((channels_multiplier <= 0), -1);
-   XA_NNLIB_ARG_CHK_COND((y_stride <= 0 || x_stride <= 0), -1);
-   XA_NNLIB_ARG_CHK_COND((y_padding < 0 || x_padding < 0), -1);
-@@ -2282,8 +2278,6 @@ WORD32 xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s_generic
-   XA_NNLIB_ARG_CHK_COND((inp_data_format != 0 && inp_data_format != 1), -1);
-   XA_NNLIB_ARG_CHK_COND((out_data_format != 0), -1);
-   /* Implementation dependent checks */
--  XA_NNLIB_ARG_CHK_COND((y_stride > kernel_height), -1);
--  XA_NNLIB_ARG_CHK_COND((x_stride > kernel_width), -1);
- 
-   if(inp_data_format == 0)
-   {

tensorflow/lite/micro/tools/make/ext_libs/xtensa_download.sh

@@ -51,9 +51,9 @@ if [[ ${2} == "hifi4" ]]; then
   LIBRARY_DIRNAME="xa_nnlib_hifi4"
   LIBRARY_MD5="8b934f61ffe0a966644849602810fb1b"
 elif [[ ${2} == "hifi5" ]]; then
-  LIBRARY_URL="http://github.com/foss-xtensa/nnlib-hifi5/raw/master/archive/xa_nnlib_hifi5_02_22.zip"
+  LIBRARY_URL="http://github.com/foss-xtensa/nnlib-hifi5/raw/master/archive/xa_nnlib_hifi5_06_30.zip"
   LIBRARY_DIRNAME="xa_nnlib_hifi5"
-  LIBRARY_MD5="08cd4d446b3e0b7d180f9ef0dec9ad0a"
+  LIBRARY_MD5="0c832b15d27ac557fa5453c902c5662a"
 else
   echo "Attempting to download an unsupported xtensa variant: ${2}"
   exit 1
@@ -80,10 +80,6 @@ else
     apply_patch_to_folder ${DOWNLOADS_DIR}/xa_nnlib_hifi4/ ../../ext_libs/xtensa_patch.patch
   fi
 
-  if [[ ${2} == "hifi5" ]]; then
-    apply_patch_to_folder ${DOWNLOADS_DIR}/xa_nnlib_hifi5/ ../../ext_libs/xtensa_depthwise_patch_hifi5.patch
-  fi
-
 fi
 
 echo "SUCCESS"