Implement SQRT_EX op in CPU (#3760)

This patch implements SQRT_EX op in CPU.
Related issue: #3459

Signed-off-by: prasannar <prasanna.r@samsung.com>

diff --git a/runtimes/pure_arm_compute/src/compilation.cc b/runtimes/pure_arm_compute/src/compilation.cc
index 133f075..4f4d9da 100644 (file)
--- a/runtimes/pure_arm_compute/src/compilation.cc
+++ b/runtimes/pure_arm_compute/src/compilation.cc
@@ -104,6 +104,7 @@
 #include "internal/layers/SimplePackLayer.h"
 #include "internal/layers/SimpleSpaceToBatchND.h"
 #include "internal/layers/SimpleNeg.h"
+#include "internal/layers/SimpleSQRT.h"
 #include "internal/layers/ArgMinMax.h"
 
 #include "util/matrix/IndexIterator.h"
@@ -3633,21 +3634,33 @@ void Planner::visit(const ::internal::tflite::op::SQRT::Node &node)
     const ::arm_compute::ActivationLayerInfo act_info{
         ::arm_compute::ActivationLayerInfo::ActivationFunction::SQRT};
 
-    if (::internal::arm_compute::isGpuMode())
+    if (from_env<bool>(std::getenv("USE_SIMPLE_SQRT")))
     {
-      auto fn = nnfw::make_unique<::arm_compute::CLActivationLayer>();
+      // USE CPU VERSION OF SQRT
+      auto fn = nnfw::make_unique<SimpleSQRT>();
 
-      fn->configure(CAST_CL(input_alloc), CAST_CL(output_alloc), act_info);
+      fn->configure(input_alloc, output_alloc);
 
       builder.append("SQRT", std::move(fn));
     }
     else
     {
-      auto fn = nnfw::make_unique<::arm_compute::NEActivationLayer>();
+      if (::internal::arm_compute::isGpuMode())
+      {
+        auto fn = nnfw::make_unique<::arm_compute::CLActivationLayer>();
 
-      fn->configure(input_alloc, output_alloc, act_info);
+        fn->configure(CAST_CL(input_alloc), CAST_CL(output_alloc), act_info);
 
-      builder.append("SQRT", std::move(fn));
+        builder.append("SQRT", std::move(fn));
+      }
+      else
+      {
+        auto fn = nnfw::make_unique<::arm_compute::NEActivationLayer>();
+
+        fn->configure(input_alloc, output_alloc, act_info);
+
+        builder.append("SQRT", std::move(fn));
+      }
     }
   };
 

diff --git a/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.cc b/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.cc
new file mode 100644 (file)
index 0000000..32d2828
--- /dev/null
+++ b/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.cc
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2018 Samsung Electronics Co., Ltd. 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 "internal/layers/SimpleSQRT.h"
+
+#include <math.h>
+#include <arm_compute/runtime/CL/CLScheduler.h>
+
+void SimpleSQRT::configure(::arm_compute::ITensor *input, ::arm_compute::ITensor *output)
+{
+  _input = input;
+  _output = output;
+}
+
+void SimpleSQRT::run()
+{
+  auto &queue = ::arm_compute::CLScheduler::get().queue();
+  if (::internal::arm_compute::isGpuMode())
+  {
+    CAST_CL(_input)->map(queue);
+    CAST_CL(_output)->map(queue);
+  }
+
+  arm_compute::Window window;
+  window.use_tensor_dimensions(_output->info()->tensor_shape());
+
+  execute_window_loop(window, [this](const arm_compute::Coordinates &id) {
+    // NOTE Must be two input tensors of identical type
+    //      Must be output tensor of the same type as input0.
+    assert(_input->info()->data_type() == _output->info()->data_type());
+
+    const auto input_value = *reinterpret_cast<float *>(_input->ptr_to_element(id));
+    *reinterpret_cast<float *>(_output->ptr_to_element(id)) = sqrt(input_value);
+  });
+
+  if (::internal::arm_compute::isGpuMode())
+  {
+    CAST_CL(_input)->unmap(queue);
+    CAST_CL(_output)->unmap(queue);
+  }
+}

diff --git a/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.h b/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.h
new file mode 100644 (file)
index 0000000..ec2e7b0
--- /dev/null
+++ b/runtimes/pure_arm_compute/src/internal/layers/SimpleSQRT.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2018 Samsung Electronics Co., Ltd. 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.
+ */
+
+#ifndef __SIMPLE_SQRT_H__
+#define __SIMPLE_SQRT_H__
+
+#include "internal/arm_compute.h"
+#include <arm_compute/core/ITensor.h>
+#include <arm_compute/runtime/IFunction.h>
+
+class SimpleSQRT : public ::arm_compute::IFunction
+{
+public:
+  SimpleSQRT(void) : _input(nullptr), _output(nullptr)
+  {
+    // DO NOTHING
+  }
+
+  void configure(::arm_compute::ITensor *input, ::arm_compute::ITensor *output);
+
+  void run() override;
+
+private:
+  ::arm_compute::ITensor *_input;
+  ::arm_compute::ITensor *_output;
+};
+
+#endif /*__SIMPLE_SQRT_H__ */