Add experimental CV-CUDA resize

Signed-off-by: Rafal Banas <[email protected]>

cmake/Dependencies.common.cmake

@@ -264,7 +264,7 @@ if (BUILD_CVCUDA)
   set(DALI_BUILD_PYTHON ${BUILD_PYTHON})
   set(BUILD_PYTHON OFF)
   # for now we use only median blur from CV-CUDA
-  set(CV_CUDA_SRC_PATERN medianblur median_blur morphology warp)
+  set(CV_CUDA_SRC_PATERN medianblur median_blur morphology warp HQResize)
   check_and_add_cmake_submodule(${PROJECT_SOURCE_DIR}/third_party/cvcuda)
   set(BUILD_PYTHON ${DALI_BUILD_PYTHON})
 endif()

dali/operators/image/resize/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2017-2024, NVIDIA CORPORATION. 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.
@@ -12,7 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+if (BUILD_CVCUDA)
+  add_subdirectory(experimental)
+endif()
+
 # Get all the source files and dump test files
 collect_headers(DALI_INST_HDRS PARENT_SCOPE)
 collect_sources(DALI_OPERATOR_SRCS PARENT_SCOPE)
 collect_test_sources(DALI_OPERATOR_TEST_SRCS PARENT_SCOPE)
+

dali/operators/image/resize/experimental/CMakeLists.txt

@@ -0,0 +1,18 @@
+# Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. 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.
+
+
+collect_headers(DALI_INST_HDRS PARENT_SCOPE)
+collect_sources(DALI_OPERATOR_SRCS PARENT_SCOPE)
+collect_test_sources(DALI_OPERATOR_TEST_SRCS PARENT_SCOPE)

dali/operators/image/resize/experimental/resize.cc

@@ -0,0 +1,79 @@
+// Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. 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.
+
+#define DALI_RESIZE_BASE_CC
+
+#include "dali/operators/image/resize/experimental/resize.h"
+#include <cassert>
+#include "dali/pipeline/data/views.h"
+
+namespace dali {
+
+DALI_SCHEMA(experimental__Resize)
+  .DocStr(R"code(Resize images.)code")
+  .NumInput(1)
+  .NumOutput(1)
+  .AdditionalOutputsFn([](const OpSpec& spec) {
+    return static_cast<int>(spec.GetArgument<bool>("save_attrs"));
+  })
+  .InputLayout(0, {"HWC",  "FHWC",  "CHW",  "FCHW",  "CFHW" ,
+                   "DHWC", "FDHWC", "CDHW", "FCDHW", "CFDHW"  })
+  .AddOptionalArg("save_attrs",
+      R"code(Save reshape attributes for testing.)code", false)
+  .AddOptionalArg<DALIImageType>("image_type", "Image type", nullptr)
+  .DeprecateArg("image_type")  // deprecated since 0.25dev
+  .SupportVolumetric()
+  .AllowSequences()
+  .AddParent("ResizeAttr")
+  .AddParent("ResamplingFilterAttr");
+
+CvCudaResize::CvCudaResize(const OpSpec &spec)
+    : StatelessOperator<GPUBackend>(spec)
+    , ResizeBase<GPUBackend>(spec) {
+  save_attrs_ = this->spec_.HasArgument("save_attrs");
+  InitializeBackend();
+}
+
+void CvCudaResize::InitializeBackend() {
+  InitializeGPU(spec_.GetArgument<int>("minibatch_size"),
+                spec_.GetArgument<int64_t>("temp_buffer_hint"));
+}
+
+void CvCudaResize::RunImpl(Workspace &ws) {
+  const auto &input = ws.Input<GPUBackend>(0);
+  auto &output = ws.Output<GPUBackend>(0);
+
+  RunResize(ws, output, input);
+  output.SetLayout(input.GetLayout());
+
+  if (save_attrs_) {
+    auto &attr_out = ws.Output<GPUBackend>(1);
+    const auto &attr_shape = attr_out.shape();
+    assert(attr_shape.num_samples() == input.shape().num_samples() &&
+           attr_shape.sample_dim() == 1 &&
+           is_uniform(attr_shape) &&
+           attr_shape[0][0] == NumSpatialDims());
+
+    if (!attr_staging_.has_data())
+      attr_staging_.set_pinned(true);
+    attr_staging_.Resize(attr_out.shape(), DALI_INT32);
+    auto attr_view = view<int, 1>(attr_staging_);
+    SaveAttrs(attr_view, input.shape());
+    attr_out.Copy(attr_staging_, ws.stream());
+  }
+}
+
+DALI_REGISTER_OPERATOR(experimental__Resize, CvCudaResize, GPU);
+
+}  // namespace dali

dali/operators/image/resize/experimental/resize.h

@@ -0,0 +1,135 @@
+// Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. 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 DALI_OPERATORS_IMAGE_RESIZE_EXPERIMENTAL_RESIZE_H_
+#define DALI_OPERATORS_IMAGE_RESIZE_EXPERIMENTAL_RESIZE_H_
+
+#include <random>
+#include <utility>
+#include <vector>
+#include <memory>
+
+#include "dali/core/common.h"
+#include "dali/core/error_handling.h"
+#include "dali/kernels/context.h"
+#include "dali/kernels/imgproc/resample/params.h"
+#include "dali/kernels/scratch.h"
+#include "dali/operators/image/resize/experimental/resize_op_impl_cvcuda.h"
+#include "dali/operators/image/resize/resize_attr.h"
+#include "dali/operators/image/resize/resize_base.h"
+#include "dali/pipeline/operator/checkpointing/stateless_operator.h"
+#include "dali/pipeline/operator/common.h"
+
+namespace dali {
+
+class CvCudaResize : public StatelessOperator<GPUBackend>, protected ResizeBase<GPUBackend> {
+ public:
+  explicit CvCudaResize(const OpSpec &spec);
+
+ protected:
+  void SetupResize(TensorListShape<> &out_shape, DALIDataType out_type,
+                   const TensorListShape<> &in_shape, DALIDataType in_type,
+                   span<const kernels::ResamplingParams> params, int spatial_ndim,
+                   int first_spatial_dim) {
+    VALUE_SWITCH(spatial_ndim, static_spatial_ndim, (2, 3),
+    (
+      using ImplType = ResizeOpImplCvCuda<static_spatial_ndim>;
+      SetImpl<ImplType>([&]{ return std::make_unique<ImplType>(GetMinibatchSize()); });
+      impl_->Setup(out_shape, in_shape, first_spatial_dim, params);
+    ), // NOLINT
+    (DALI_FAIL(make_string("Unsupported number of resized dimensions: ", spatial_ndim))));
+  }
+
+
+  int NumSpatialDims() const {
+    return resize_attr_.spatial_ndim_;
+  }
+  int FirstSpatialDim() const {
+    return resize_attr_.first_spatial_dim_;
+  }
+
+  bool CanInferOutputs() const override {
+    return true;
+  }
+
+  bool SetupImpl(std::vector<OutputDesc> &output_desc, const Workspace &ws) override;
+
+  void RunImpl(Workspace &ws) override;
+
+  void SaveAttrs(const TensorListView<StorageCPU, int, 1> &shape_data,
+                 const TensorListShape<> &orig_shape) const {
+    int N = orig_shape.num_samples();
+    int D = NumSpatialDims();
+    assert(shape_data.sample_dim() == 1);
+    for (int i = 0; i < N; i++) {
+      auto sample_shape = orig_shape.tensor_shape_span(i);
+      assert(static_cast<int>(shape_data.shape[i][0]) == D);
+      int *out_shape = shape_data.data[i];
+      for (int d = 0; d < D; d++) {
+        out_shape[d] = sample_shape[FirstSpatialDim() + d];
+      }
+    }
+  }
+
+  void PrepareParams(const ArgumentWorkspace &ws, const TensorListShape<> &input_shape,
+                     const TensorLayout &layout) {
+    resize_attr_.PrepareResizeParams(spec_, ws, input_shape, layout);
+    assert(NumSpatialDims() >= 1 && NumSpatialDims() <= 3);
+    assert(FirstSpatialDim() >= 0);
+    int N = input_shape.num_samples();
+    resample_params_.resize(N * NumSpatialDims());
+    resampling_attr_.PrepareFilterParams(spec_, ws, N);
+    resampling_attr_.GetResamplingParams(make_span(resample_params_),
+                                         make_cspan(resize_attr_.params_));
+  }
+
+  void InitializeBackend();
+
+  USE_OPERATOR_MEMBERS();
+  std::vector<kernels::ResamplingParams> resample_params_;
+  TensorList<CPUBackend> attr_staging_;
+  using Operator<GPUBackend>::RunImpl;
+  bool save_attrs_ = false;
+
+  ResizeAttr resize_attr_;
+  ResamplingFilterAttr resampling_attr_;
+};
+
+bool CvCudaResize::SetupImpl(std::vector<OutputDesc> &output_desc, const Workspace &ws) {
+  output_desc.resize(save_attrs_ ? 2 : 1);
+  auto &input = ws.Input<GPUBackend>(0);
+
+  const auto &in_shape = input.shape();
+  auto in_type = input.type();
+  auto in_layout = input.GetLayout();
+  int N = in_shape.num_samples();
+
+  PrepareParams(ws, in_shape, in_layout);
+
+  auto out_type = resampling_attr_.GetOutputType(in_type);
+
+  output_desc[0].type = out_type;
+  this->SetupResize(output_desc[0].shape, out_type, in_shape, in_type,
+                    make_cspan(this->resample_params_), NumSpatialDims(), FirstSpatialDim());
+
+  if (save_attrs_) {
+    output_desc[1].shape = uniform_list_shape(N, TensorShape<1>(NumSpatialDims()));
+    output_desc[1].type = DALI_INT32;
+  }
+  return true;
+}
+
+}  // namespace dali
+
+#endif  // DALI_OPERATORS_IMAGE_RESIZE_EXPERIMENTAL_RESIZE_H_