Invoke optimized depthwise convolution.

Change-Id: I099ec7a9345e279b91fdd6de4c0ae1d33ddd0bc2
diff --git a/tensorflow/lite/micro/kernels/kelvin/depthwise_conv.cc b/tensorflow/lite/micro/kernels/kelvin/depthwise_conv.cc
new file mode 100644
index 0000000..5dc2b6d
--- /dev/null
+++ b/tensorflow/lite/micro/kernels/kelvin/depthwise_conv.cc
@@ -0,0 +1,176 @@
+/* Copyright 2017 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.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/micro/kernels/depthwise_conv.h"
+
+#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
+#include "tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h"
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h"
+#include "tensorflow/lite/kernels/kernel_util.h"
+#include "tensorflow/lite/micro/kernels/kernel_util.h"
+#include "tensorflow/lite/micro/micro_log.h"
+#include "tflm/opt/opt.h"
+
+namespace tflite {
+namespace {
+
+void* Init(TfLiteContext* context, const char* buffer, size_t length) {
+  TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr);
+  return context->AllocatePersistentBuffer(context, sizeof(OpDataConv));
+}
+
+TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
+  TFLITE_DCHECK(node->user_data != nullptr);
+  TFLITE_DCHECK(node->builtin_data != nullptr);
+
+  auto& params =
+      *(reinterpret_cast<TfLiteDepthwiseConvParams*>(node->builtin_data));
+  const OpDataConv& data = *(static_cast<const OpDataConv*>(node->user_data));
+
+  TfLiteEvalTensor* output =
+      tflite::micro::GetEvalOutput(context, node, kDepthwiseConvOutputTensor);
+  const TfLiteEvalTensor* input =
+      tflite::micro::GetEvalInput(context, node, kDepthwiseConvInputTensor);
+  const TfLiteEvalTensor* filter =
+      tflite::micro::GetEvalInput(context, node, kDepthwiseConvWeightsTensor);
+  const TfLiteEvalTensor* bias =
+      (NumInputs(node) == 3)
+          ? tflite::micro::GetEvalInput(context, node, kDepthwiseConvBiasTensor)
+          : nullptr;
+
+  switch (input->type) {  // Already know in/out types are same.
+    case kTfLiteFloat32: {
+      tflite::reference_ops::DepthwiseConv(
+          DepthwiseConvParamsFloat(params, data),
+          tflite::micro::GetTensorShape(input),
+          tflite::micro::GetTensorData<float>(input),
+          tflite::micro::GetTensorShape(filter),
+          tflite::micro::GetTensorData<float>(filter),
+          tflite::micro::GetTensorShape(bias),
+          tflite::micro::GetOptionalTensorData<float>(bias),
+          tflite::micro::GetTensorShape(output),
+          tflite::micro::GetTensorData<float>(output));
+      break;
+    }
+    case kTfLiteInt8: {
+      switch (filter->type) {
+        case kTfLiteInt4: {
+          int8_t* unpacked_filter_data = static_cast<int8_t*>(
+              context->GetScratchBuffer(context, data.filter_buffer_index));
+          tflite::tensor_utils::UnpackDenseInt4IntoInt8(
+              tflite::micro::GetTensorData<int8_t>(filter),
+              tflite::micro::GetTensorShape(filter).FlatSize(),
+              unpacked_filter_data);
+          reference_integer_ops::DepthwiseConvPerChannel(
+              DepthwiseConvParamsQuantized(params, data),
+              data.per_channel_output_multiplier, data.per_channel_output_shift,
+              tflite::micro::GetTensorShape(input),
+              tflite::micro::GetTensorData<int8_t>(input),
+              tflite::micro::GetTensorShape(filter), unpacked_filter_data,
+              tflite::micro::GetTensorShape(bias),
+              tflite::micro::GetOptionalTensorData<int32_t>(bias),
+              tflite::micro::GetTensorShape(output),
+              tflite::micro::GetTensorData<int8_t>(output));
+          break;
+        }
+        case kTfLiteInt8: {
+          reference_integer_ops::DepthwiseConvPerChannel(
+              DepthwiseConvParamsQuantized(params, data),
+              data.per_channel_output_multiplier, data.per_channel_output_shift,
+              tflite::micro::GetTensorShape(input),
+              tflite::micro::GetTensorData<int8_t>(input),
+              tflite::micro::GetTensorShape(filter),
+              tflite::micro::GetTensorData<int8_t>(filter),
+              tflite::micro::GetTensorShape(bias),
+              tflite::micro::GetOptionalTensorData<int32_t>(bias),
+              tflite::micro::GetTensorShape(output),
+              tflite::micro::GetTensorData<int8_t>(output));
+          break;
+        }
+        default:
+          MicroPrintf("Filter type %s (%d) for input type %s not supported.",
+                      TfLiteTypeGetName(filter->type), filter->type,
+                      TfLiteTypeGetName(input->type));
+          return kTfLiteError;
+      }
+      break;
+    }
+    case kTfLiteInt16: {
+      switch (filter->type) {
+        case kTfLiteInt8: {
+          tflite::DepthwiseParams dw_params =
+              DepthwiseConvParamsQuantized(params, data);
+          tflite::RuntimeShape filter_shape =
+              tflite::micro::GetTensorShape(filter);
+          tflite::RuntimeShape input_shape =
+              tflite::micro::GetTensorShape(input);
+          if (dw_params.padding_type == PaddingType::kValid &&
+              dw_params.stride_width == 1 && dw_params.stride_height == 1 &&
+              filter_shape.Dims(1) == 1 && filter_shape.Dims(2) == 3 &&
+              dw_params.dilation_width_factor == 1 &&
+              dw_params.dilation_height_factor == 1) {
+            kelvin::opt::DepthwiseConv2DKelvinS16K3x1(
+                tflite::micro::GetTensorData<int16_t>(input),
+                tflite::micro::GetTensorData<int8_t>(filter),
+                tflite::micro::GetOptionalTensorData<int64_t>(bias),
+                filter_shape.Dims(3) /*channels*/,
+                input_shape.Dims(2) /*frames*/,
+                data.per_channel_output_multiplier,
+                data.per_channel_output_shift,
+                dw_params.quantized_activation_min,
+                dw_params.quantized_activation_max,
+                tflite::micro::GetTensorData<int16_t>(output)
+            );
+            break;
+          }
+
+          reference_integer_ops::DepthwiseConvPerChannel(
+              dw_params,
+              data.per_channel_output_multiplier, data.per_channel_output_shift,
+              tflite::micro::GetTensorShape(input),
+              tflite::micro::GetTensorData<int16_t>(input),
+              tflite::micro::GetTensorShape(filter),
+              tflite::micro::GetTensorData<int8_t>(filter),
+              tflite::micro::GetTensorShape(bias),
+              tflite::micro::GetOptionalTensorData<int64_t>(bias),
+              tflite::micro::GetTensorShape(output),
+              tflite::micro::GetTensorData<int16_t>(output));
+          break;
+        }
+        default:
+          MicroPrintf("Filter type %s (%d) for input type %s not supported.",
+                      TfLiteTypeGetName(filter->type), filter->type,
+                      TfLiteTypeGetName(input->type));
+          return kTfLiteError;
+      }
+      break;
+    }
+    default:
+      MicroPrintf("Input type %s (%d) not supported.",
+                  TfLiteTypeGetName(input->type), input->type);
+      return kTfLiteError;
+  }
+  return kTfLiteOk;
+}
+
+}  // namespace
+
+TFLMRegistration Register_DEPTHWISE_CONV_2D() {
+  return tflite::micro::RegisterOp(Init, DepthwiseConvPrepare, Eval);
+}
+
+}  // namespace tflite