Add Op MaxpoolWithArgmax

Signed-off-by: Kainan Cha <kainan.zha@verisilicon.com>

include/tim/vx/ops/maxpoolwithargmax.h

@@ -0,0 +1,66 @@
+/****************************************************************************
+*
+*    Copyright (c) 2021 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+#ifndef TIM_VX_OPS_MAXPOOLWITHARGMAX_H_
+#define TIM_VX_OPS_MAXPOOLWITHARGMAX_H_
+
+#include <array>
+
+#include "tim/vx/operation.h"
+#include "tim/vx/types.h"
+
+namespace tim {
+namespace vx {
+namespace ops {
+
+/**
+ * ## MaxpoolWithArgmax
+ *
+ * Performs an 2-D Max pooling operation and return indices
+ *
+ * - padding : AUTO, VALID or SAME.
+ * - ksize : filter size.
+ * - stride : stride along each spatial axis.
+ * - round_type : CEILING or FLOOR.
+ */
+
+class MaxpoolWithArgmax : public Operation {
+ public:
+  MaxpoolWithArgmax(Graph* graph, PadType padding,
+         const std::array<uint32_t, 2>& ksize,
+         const std::array<uint32_t, 2>& stride,
+         RoundType round_type = RoundType::FLOOR,
+         DataLayout layout = DataLayout::WHCN);
+
+ protected:
+  const PadType padding_;
+  const std::array<uint32_t, 2> ksize_;
+  const std::array<uint32_t, 2> stride_;
+  const RoundType round_type_;
+};
+
+}  // namespace ops
+}  // namespace vx
+}  // namespace tim
+
+#endif /* TIM_VX_OPS_MAXPOOLWITHARGMAX_H_ */

src/tim/vx/internal/src/ops/vsi_nn_op_poolwithargmax.c

@@ -244,6 +244,14 @@ static vsi_bool op_setup
     )
 {
     vsi_bool ret = TRUE;
+    vsi_nn_compute_padding(
+        inputs[0]->attr.size,
+        self->nn_param.pool.ksize,
+        self->nn_param.pool.stride,
+        NULL,
+        self->nn_param.pool.pad_type,
+        self->nn_param.pool.pad
+    );
 
     if( VSI_NN_DIM_AUTO == outputs[0]->attr.dim_num )
     {

src/tim/vx/ops/README.md

@@ -25,7 +25,7 @@ Reorg|REORG|Mapped|[darknet.reorg](https://github.com/pjreddie/darknet/blob/mast
 ||VARIABLE|Unmapped|[tf.variable](https://tensorflow.google.cn/api_docs/python/tf/Variable)
 L2Normalization|L2_NORMALIZE|Mapped|[tf.math.l2_normalize](https://tensorflow.google.cn/api_docs/python/tf/math/l2_normalize)
 FullyConnected|FCL2|Mapped|[tf.keras.layers.Dense](https://tensorflow.google.cn/api_docs/python/tf/keras/layers/Dense)
-||POOLWITHARGMAX|Unmapped|[tf.nn.max_pool_with_argmax](https://tensorflow.google.cn/api_docs/python/tf/nn/max_pool_with_argmax)
+|MaxpoolWithArgmax|POOLWITHARGMAX|Mapped|[tf.nn.max_pool_with_argmax](https://tensorflow.google.cn/api_docs/python/tf/nn/max_pool_with_argmax)
 ArgMax|ARGMAX|Mapped|[tf.math.argmax](https://tensorflow.google.cn/api_docs/python/tf/math/argmax)
 Maximum|MAXIMUM|Mapped|[tf.math.maximum](https://tensorflow.google.cn/api_docs/python/tf/math/maximum)
 ||CROP|Unmapped

src/tim/vx/ops/maxpoolwithargmax.cc

@@ -0,0 +1,57 @@
+/****************************************************************************
+*
+*    Copyright (c) 2021 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+#include "tim/vx/ops/maxpoolwithargmax.h"
+
+#include "operation_private.h"
+#include "type_utils.h"
+#include "vsi_nn_pub.h"
+
+namespace tim {
+namespace vx {
+namespace ops {
+
+MaxpoolWithArgmax::MaxpoolWithArgmax(Graph* graph, PadType padding,
+                                   const std::array<uint32_t, 2>& ksize,
+                                   const std::array<uint32_t, 2>& stride,
+                                   RoundType round_type,
+                                   DataLayout layout)
+    : Operation(graph, VSI_NN_OP_POOLWITHARGMAX, 1, 2, layout),
+      padding_(padding),
+      ksize_(ksize),
+      stride_(stride),
+      round_type_(round_type) {
+  this->impl()->node()->nn_param.pool.type = TranslatePoolType(PoolType::MAX);
+  this->impl()->node()->nn_param.pool.round_type =
+      TranslateRoundType(round_type_);
+  this->impl()->node()->nn_param.pool.ksize[0] = ksize_[0];
+  this->impl()->node()->nn_param.pool.ksize[1] = ksize_[1];
+  this->impl()->node()->nn_param.pool.stride[0] = stride_[0];
+  this->impl()->node()->nn_param.pool.stride[1] = stride_[1];
+  this->impl()->node()->nn_param.pool.pad_type = TranslatePadType(padding_);
+  this->SetRoundingPolicy(OverflowPolicy::SATURATE, RoundingPolicy::RTNE, round_type_);
+}
+
+}  // namespace ops
+}  // namespace vx
+}  // namespace tim

src/tim/vx/ops/maxpoolwithargmax_test.cc

@@ -0,0 +1,123 @@
+/****************************************************************************
+*
+*    Copyright (c) 2021 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+#include "tim/vx/context.h"
+#include "tim/vx/graph.h"
+#include "tim/vx/ops/maxpoolwithargmax.h"
+
+#include "gtest/gtest.h"
+
+TEST(MaxpoolWithArgmax, shape_3_3_1_fp32_kernel_2_stride_2) {
+    auto ctx = tim::vx::Context::Create();
+    auto graph = ctx->CreateGraph();
+
+    tim::vx::ShapeType in_shape({3, 3, 1});
+    tim::vx::ShapeType out_shape({2, 2, 1});
+    tim::vx::TensorSpec input_spec(tim::vx::DataType::FLOAT32,
+                            in_shape, tim::vx::TensorAttribute::INPUT);
+    tim::vx::TensorSpec output_spec_indices(tim::vx::DataType::UINT8,
+                            out_shape, tim::vx::TensorAttribute::OUTPUT);
+    tim::vx::TensorSpec output_spec_values(tim::vx::DataType::FLOAT32,
+                            out_shape, tim::vx::TensorAttribute::OUTPUT);
+
+    auto input_tensor = graph->CreateTensor(input_spec);
+    auto output_tensor_indices = graph->CreateTensor(output_spec_indices);
+    auto output_tensor_values = graph->CreateTensor(output_spec_values);
+
+    std::vector<float> in_data = {
+        1, 2, 3,
+        4, 5, 6,
+        7, 8, 9 };
+    std::vector<float> values_golden = {
+        5, 6,
+        8, 9 };
+    std::vector<uint8_t> indices_golden = {
+        3, 2,
+        1, 0 };
+
+    EXPECT_TRUE(input_tensor->CopyDataToTensor(in_data.data(), in_data.size()*4));
+    std::array<uint32_t, 2> ksize = {2, 2};
+    std::array<uint32_t, 2> stride = {2, 2};
+    auto op = graph->CreateOperation<tim::vx::ops::MaxpoolWithArgmax>(
+        tim::vx::PadType::VALID, ksize, stride);
+    (*op).BindInputs({input_tensor}).BindOutputs({output_tensor_values, output_tensor_indices});
+
+    EXPECT_TRUE(graph->Compile());
+    EXPECT_TRUE(graph->Run());
+    std::vector<float> output_values(4);
+    std::vector<uint8_t> output_indices(4);
+
+    EXPECT_TRUE(output_tensor_values->CopyDataFromTensor(output_values.data()));
+    EXPECT_TRUE(output_tensor_indices->CopyDataFromTensor(output_indices.data()));
+    EXPECT_EQ(values_golden, output_values);
+    EXPECT_EQ(indices_golden, output_indices);
+}
+
+TEST(MaxpoolWithArgmax, shape_4_4_1_uint8_kernel_2_stride_2) {
+    auto ctx = tim::vx::Context::Create();
+    auto graph = ctx->CreateGraph();
+
+    tim::vx::ShapeType in_shape({4, 4, 1});
+    tim::vx::ShapeType out_shape({2, 2, 1});
+    tim::vx::Quantization io_quant(tim::vx::QuantType::ASYMMETRIC, 1, 0);
+    tim::vx::TensorSpec input_spec(tim::vx::DataType::UINT8,
+                            in_shape, tim::vx::TensorAttribute::INPUT, io_quant);
+    tim::vx::TensorSpec output_spec_indices(tim::vx::DataType::UINT8,
+                            out_shape, tim::vx::TensorAttribute::OUTPUT);
+    tim::vx::TensorSpec output_spec_values(tim::vx::DataType::UINT8,
+                            out_shape, tim::vx::TensorAttribute::OUTPUT, io_quant);
+
+    auto input_tensor = graph->CreateTensor(input_spec);
+    auto output_tensor_indices = graph->CreateTensor(output_spec_indices);
+    auto output_tensor_values = graph->CreateTensor(output_spec_values);
+
+    std::vector<uint8_t> in_data = {
+        1, 2, 3, 3,
+        4, 5, 6, 6,
+        7, 8, 9, 9,
+        10, 11, 12, 12 };
+    std::vector<uint8_t> values_golden = {
+        5, 6,
+        11, 12};
+    std::vector<uint8_t> indices_golden = {
+        3, 2,
+        3, 2};
+
+    EXPECT_TRUE(input_tensor->CopyDataToTensor(in_data.data(), in_data.size()));
+    std::array<uint32_t, 2> ksize = {2, 2};
+    std::array<uint32_t, 2> stride = {2, 2};
+    auto op = graph->CreateOperation<tim::vx::ops::MaxpoolWithArgmax>(
+        tim::vx::PadType::VALID, ksize, stride);
+    (*op).BindInputs({input_tensor}).BindOutputs({output_tensor_values, output_tensor_indices});
+
+    EXPECT_TRUE(graph->Compile());
+    EXPECT_TRUE(graph->Run());
+    std::vector<uint8_t> output_values(4);
+    std::vector<uint8_t> output_indices(4);
+
+    EXPECT_TRUE(output_tensor_values->CopyDataFromTensor(output_values.data()));
+    EXPECT_TRUE(output_tensor_indices->CopyDataFromTensor(output_indices.data()));
+    EXPECT_EQ(values_golden, output_values);
+    EXPECT_EQ(indices_golden, output_indices);
+}
+