[onert-micro] Add Conv2D kernel (#12740)

This pr adds Conv2D refactored kernel into onert-micro. ONE-DCO-1.0-Signed-off-by: Artem Balyshev <[email protected]>
Samsung · Mar 12, 2024 · afd538e · afd538e
1 parent c783355
commit afd538e
Show file tree

Hide file tree

Showing 8 changed files with 763 additions and 0 deletions.
diff --git a/onert-micro/onert-micro/include/pal/common/PALConv2DCommon.h b/onert-micro/onert-micro/include/pal/common/PALConv2DCommon.h
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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.
+ */
+
+#ifndef ONERT_MICRO_PAL_CONV2D_COMMON_H
+#define ONERT_MICRO_PAL_CONV2D_COMMON_H
+
+#include "PALUtils.h"
+
+#include "OMStatus.h"
+
+namespace onert_micro
+{
+namespace execute
+{
+namespace pal
+{
+OMStatus ConvFloat(const core::FloatConv2D *params, const core::OMRuntimeShape &input_shape,
+                   const float *input_data, const core::OMRuntimeShape &filter_shape,
+                   const float *filter_data, const float *bias_data,
+                   const core::OMRuntimeShape &output_shape, float *output_data)
+{
+  const int stride_width = params->stride_w;
+  const int stride_height = params->stride_h;
+  const int dilation_width_factor = params->dilation_width_factor;
+  const int dilation_height_factor = params->dilation_height_factor;
+  const int pad_width = params->pad_w;
+  const int pad_height = params->pad_h;
+  const float output_activation_min = params->activation_min;
+  const float output_activation_max = params->activation_max;
+
+  const auto batches = input_shape.dims(0);
+  const int input_height = input_shape.dims(1);
+  const int input_width = input_shape.dims(2);
+  const int input_depth = input_shape.dims(3);
+  const int output_depth = filter_shape.dims(0);
+  const int filter_height = filter_shape.dims(1);
+  const int filter_width = filter_shape.dims(2);
+  const int output_height = output_shape.dims(1);
+  const int output_width = output_shape.dims(2);
+  for (int batch = 0; batch < batches; ++batch)
+  {
+    for (int out_y = 0; out_y < output_height; ++out_y)
+    {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x)
+      {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel)
+        {
+          float total = 0.f;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y)
+          {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x)
+            {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && (in_y < input_height);
+
+              if (!is_point_inside_image)
+              {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel)
+              {
+                const int input_data_offset =
+                  ((batch * input_height + in_y) * input_width + in_x) * input_depth + in_channel;
+
+                const int filter_data_offset =
+                  ((out_channel * filter_height + filter_y) * filter_width + filter_x) *
+                    input_depth +
+                  in_channel;
+
+                const float input_value = input_data[input_data_offset];
+                const float filter_value = filter_data[filter_data_offset];
+                total += (input_value * filter_value);
+              }
+            }
+          }
+          // float bias_value = 0.0f;
+          if (bias_data)
+          {
+            total += bias_data[out_channel];
+          }
+
+          const int output_data_offset =
+            ((batch * output_height + out_y) * output_width + out_x) * output_depth + out_channel;
+
+          output_data[output_data_offset] =
+            std::min(std::max(total, output_activation_min), output_activation_max);
+        }
+      }
+    }
+  }
+  return Ok;
+}
+
+} // namespace pal
+} // namespace execute
+} // namespace onert_micro
+
+#endif // ONERT_MICRO_PAL_CONV2D_COMMON_H
diff --git a/onert-micro/onert-micro/include/pal/mcu/PALConv2d.h b/onert-micro/onert-micro/include/pal/mcu/PALConv2d.h
@@ -0,0 +1,22 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_CONV2D_H
+#define LUCI_INTERPRETER_PAL_CONV2D_H
+#include "PALConv2DCommon.h"
+
+#endif // LUCI_INTERPRETER_PAL_CONV2D_H
diff --git a/onert-micro/onert-micro/include/test_models/conv2d/FloatConv2DKernel.h b/onert-micro/onert-micro/include/test_models/conv2d/FloatConv2DKernel.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2024 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 ONERT_MICRO_TEST_MODELS_CONV_2D_KERNEL_FLOAT_H
+#define ONERT_MICRO_TEST_MODELS_CONV_2D_KERNEL_FLOAT_H
+
+#include "TestDataConv2DBase.h"
+
+namespace onert_micro
+{
+namespace test_model
+{
+namespace conv2d_float
+{
+
+/*
+ * Conv2D Kernel:
+ *
+ * Input(1, 4, 3, 2)   Weight(1, 2, 2, 2)   Bias(2)
+ *              \            |             /
+ *               \           |            /
+ *                    FullyConnected
+ *                           |
+ *                    Output(1, 2, 2, 2)
+ */
+
+const unsigned char test_kernel_model_circle[] = {
+  0x18, 0x00, 0x00, 0x00, 0x43, 0x49, 0x52, 0x30, 0x00, 0x00, 0x0e, 0x00, 0x14, 0x00, 0x00, 0x00,
+  0x0c, 0x00, 0x08, 0x00, 0x10, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
+  0x9c, 0x00, 0x00, 0x00, 0x10, 0x02, 0x00, 0x00, 0x2c, 0x02, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00,
+  0x88, 0x00, 0x00, 0x00, 0x7c, 0x00, 0x00, 0x00, 0x74, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,
+  0x04, 0x00, 0x00, 0x00, 0xea, 0xff, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x04, 0x00,
+  0x06, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f,
+  0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x40, 0xc0, 0x00, 0x00, 0x80, 0xc0, 0x00, 0x00, 0xa0, 0xc0,
+  0x00, 0x00, 0xc0, 0x40, 0x00, 0x00, 0xe0, 0xc0, 0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x80, 0x40,
+  0x00, 0x00, 0x00, 0xc0, 0x00, 0x00, 0x40, 0x40, 0x00, 0x00, 0x80, 0xbf, 0x00, 0x00, 0x00, 0xc1,
+  0x00, 0x00, 0xc0, 0xc0, 0x00, 0x00, 0xe0, 0x40, 0x00, 0x00, 0xa0, 0x40, 0xf8, 0xff, 0xff, 0xff,
+  0xfc, 0xff, 0xff, 0xff, 0x04, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+  0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x14, 0x00, 0x10, 0x00, 0x0c, 0x00,
+  0x08, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
+  0x7c, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00, 0x84, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+  0x6d, 0x61, 0x69, 0x6e, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x0e, 0x00, 0x14, 0x00, 0x00, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x07, 0x00, 0x08, 0x00,
+  0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x18, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00,
+  0x2c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x14, 0x00, 0x13, 0x00, 0x0c, 0x00, 0x08, 0x00, 0x07, 0x00,
+  0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+  0x03, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+  0x98, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+  0x84, 0xff, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
+  0x03, 0x00, 0x00, 0x00, 0x6f, 0x66, 0x6d, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+  0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0xb0, 0xff, 0xff, 0xff,
+  0x0c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+  0x62, 0x69, 0x61, 0x73, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00,
+  0xd4, 0xff, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
+  0x03, 0x00, 0x00, 0x00, 0x6b, 0x65, 0x72, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00,
+  0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x10, 0x00,
+  0x0c, 0x00, 0x00, 0x00, 0x08, 0x00, 0x04, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
+  0x01, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x69, 0x66, 0x6d, 0x00,
+  0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+  0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x0c, 0x00,
+  0x0b, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x03, 0x11, 0x00, 0x00, 0x00, 0x4f, 0x4e, 0x45, 0x2d, 0x74, 0x66, 0x6c, 0x69,
+  0x74, 0x65, 0x32, 0x63, 0x69, 0x72, 0x63, 0x6c, 0x65, 0x00, 0x00, 0x00};
+
+const std::vector<float> input_data = {
+  18.776451, 25.97969,   -9.277071, -3.5493946, 12.334248, 5.50226,   -2.224743,  -7.2292213,
+  10.259663, -1.0846977, 15.823856, 3.3193378,  4.9413986, 4.3529205, -10.353054, 3.7166824,
+  27.324902, -6.2231064, 10.370632, 22.661959,  20.206001, 8.245907,  9.984943,   21.379955};
+
+const std::vector<float> reference_output_data = {1.0177879, 128.43202, 0.0, 55.28556,
+                                                  39.483513, 0.0,       0.0, 7.0231743};
+
+} // namespace conv2d_float
+
+class TestDataFloatConv2D : public TestDataConv2DBase<float>
+{
+public:
+  TestDataFloatConv2D()
+  {
+    _input_data = conv2d_float::input_data;
+    _reference_output_data = conv2d_float::reference_output_data;
+    _test_kernel_model_circle = conv2d_float::test_kernel_model_circle;
+  }
+
+  ~TestDataFloatConv2D() override = default;
+};
+
+} // namespace test_model
+} // namespace onert_micro
+
+#endif // ONERT_MICRO_TEST_MODELS_CONV_2D_KERNEL_FLOAT_H