feat : [ADD] convtranspose forward 1D & 2D

include/aidge/backend/cpu.hpp

@@ -27,6 +27,7 @@
 #include "aidge/backend/cpu/operator/ClipImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvDepthWiseImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvImpl.hpp"
+#include "aidge/backend/cpu/operator/ConvTransposeImpl.hpp"
 #include "aidge/backend/cpu/operator/ConstantOfShapeImpl.hpp"
 #include "aidge/backend/cpu/operator/DivImpl.hpp"
 #include "aidge/backend/cpu/operator/EqualImpl.hpp"

include/aidge/backend/cpu/operator/ConvTransposeImpl.hpp

+
+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#ifndef AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_H_
+#define AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_H_
+
+#include <array>
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/ConvTranspose.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+
+using std::array;
+
+// Operator implementation entry point for the backend
+using ConvTranspose1D_Op = ConvTranspose_Op<1>;
+using ConvTransposeImpl1D_cpu =
+    OperatorImpl_cpu<ConvTranspose1D_Op,
+                     void(const array<DimSize_t,1> &,
+                          const array<DimSize_t,1> &,
+                          const array<DimSize_t,1> &,
+                          const array<DimSize_t, 3> &,
+                          const array<DimSize_t, 3> &,
+                          const void *,
+                          const void *,
+                          const void *,
+                          void *)>;
+
+using ConvTranspose2D_Op = ConvTranspose_Op<2>;
+using ConvTransposeImpl2D_cpu =
+ OperatorImpl_cpu<ConvTranspose2D_Op,
+                                        void(const array<DimSize_t, 2> &,
+                                             const array<DimSize_t, 2> &,
+                                             const array<DimSize_t, 2> &,
+                                             const array<DimSize_t, 4> &,
+                                             const array<DimSize_t, 4> &,
+                                             const void *,
+                                             const void *,
+                                             const void *,
+                                             void *)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(ConvTranspose1D_Op, "cpu", ConvTransposeImpl1D_cpu::create);
+REGISTRAR(ConvTranspose2D_Op, "cpu", ConvTransposeImpl2D_cpu::create);
+} // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_H_ */

include/aidge/backend/cpu/operator/ConvTransposeImpl_kernels.hpp

+/********************************************************************************
+ * Copyright (c) 2025 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#ifndef AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_KERNELS_H_
+
+#include <array>
+
+#include "aidge/backend/cpu/operator/ConvTransposeImpl.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include <aidge/backend/cpu/operator/ConvImpl_kernels.hpp>
+#include <aidge/data/Data.hpp>
+#include <aidge/data/half.hpp>
+#include <aidge/scheduler/ProdConso.hpp>
+#include <aidge/utils/Types.h>
+
+namespace Aidge {
+
+using std::array;
+
+////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////
+// 1D
+////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////
+
+/**
+ * @brief performs forward bias operation for convtranspose operator
+ *
+ * @tparam B Bias data type.
+ * @tparam O Output data type.
+ * @param[in] bias bias values
+ * @param[in] oDims dimensions of the output
+ * @param[in] oStrides nb of elements contained per dimension of the output
+ * @param[out] output
+ */
+template <class B, class O>
+static void convTranspose1DForwardBias(const B *biases,
+                                       const array<DimSize_t, 3> &oDims,
+                                       const array<DimSize_t, 2> &oStrides,
+                                       O *output) {
+    array<DimSize_t, 2> outOffsets{0, 0};
+    for (DimSize_t batch = 0; batch < oDims[0]; ++batch) {
+        outOffsets[0] = batch * oStrides[0];
+        for (DimSize_t outCh = 0; outCh < oDims[1]; ++outCh) {
+            outOffsets[1] = outCh * oStrides[1] + outOffsets[0];
+            // If bias = nullptr, set B(0)
+            B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
+            std::fill(output + outOffsets[1],
+                      output + (outOffsets[1] + oDims[2]),
+                      biasVal);
+        }
+    }
+}
+
+/**
+ * @brief forward kernel for convtranspose
+ * @note ConvTranspose forward is simply convolution backward kernel.
+ * Check convolution functions for more in-depth details on how the
+ subfunctions are built.
+ * @tparam I Input data type.
+ * @tparam W Weight data type.
+ * @tparam B Bias data type.
+ * @tparam O Output data type.
+ * @param[in] stride stride parameter of the convTranspose operator
+ * @param[in] dilation dilation parameter of the convTranspose operator
+ * @param[in] inputDims input dimensions
+ * @param[in] outputDims output tensor dimensions
+ * @param[in] oStrides nb of elements contained per dimension of the output
+ * @param[in] input_ values
+ * @param[in] weight_ values
+ * @param[in] biases_ values
+ * @param[out] output
+ */
+template <class I, class W, class B, class O>
+void ConvTransposeImpl1D_cpu_forward_kernel(
+    const array<DimSize_t, 1> &stride,
+    const array<DimSize_t, 1> &dilation,
+    const array<DimSize_t, 1> &kernelDim,
+    const array<DimSize_t, 3> &inputDims,
+    const array<DimSize_t, 3> &outputDims,
+    const void *input_,
+    const void *weights_,
+    const void *biases_,
+    void *output_) {
+
+    const I *input = static_cast<const I *>(input_);
+    const W *weights = static_cast<const W *>(weights_);
+    O *output = static_cast<O *>(output_);
+
+    // {batch_stride, channel_stride, dim0_stride}
+    const array<DimSize_t, 2> inputStrides{inputDims[1] * inputDims[2],
+                                           inputDims[2]};
+
+    // {batch_stride, channel_stride, dim0_stride}
+    const array<DimSize_t, 2> outputStrides{outputDims[1] * outputDims[2],
+                                            outputDims[2]};
+
+    // NOTE: kernel dims = {inChannels, outChannels, kernelDims[0]}
+    const array<DimSize_t, 2> kernelStrides{
+        outputDims[1] * kernelDim[0],
+        kernelDim[0],
+    };
+
+    if (biases_ != nullptr) {
+        const B *biases = static_cast<const B *>(biases_);
+        convTranspose1DForwardBias(biases, outputDims, outputStrides, output);
+    }
+
+    conv1DBackwardInput(stride,
+                        dilation,
+                        kernelDim,
+                        kernelStrides,
+                        weights,
+                        inputDims,
+                        inputStrides,
+                        input,
+                        outputDims,
+                        outputStrides,
+                        output);
+}
+
+REGISTRAR(ConvTransposeImpl1D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Int32, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl1D_cpu_forward_kernel<std::int32_t,
+                                                  std::int32_t,
+                                                  std::int32_t,
+                                                  std::int32_t>,
+           nullptr});
+REGISTRAR(ConvTransposeImpl1D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Float32, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl1D_cpu_forward_kernel<float, float, float, float>,
+           nullptr});
+REGISTRAR(ConvTransposeImpl1D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Float16, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl1D_cpu_forward_kernel<half_float::half,
+                                                  half_float::half,
+                                                  half_float::half,
+                                                  half_float::half>,
+           nullptr});
+REGISTRAR(
+    ConvTransposeImpl1D_cpu,
+    {{DataType::Any, DataFormat::NCHW}, {DataType::Float64, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel,
+     ConvTransposeImpl1D_cpu_forward_kernel<double, double, double, double>,
+     nullptr});
+
+////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////
+// 2D
+////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////
+
+/**
+ * @brief performs forward bias operation for convtranspose operator
+ *
+ * @tparam B Bias data type.
+ * @tparam O Output data type.
+ * @param[in] bias bias values
+ * @param[in] oDims dimensions of the output
+ * @param[in] oStrides nb of elements contained per dimension of the output
+ * @param[out] output
+ */
+template <class B, class O>
+static void convTranspose2DForwardBias(const B *biases,
+                                       const array<DimSize_t, 4> &oDims,
+                                       const array<DimSize_t, 3> &oStrides,
+                                       O *output) {
+    array<DimSize_t, 2> outOffsets{0, 0};
+
+    for (DimSize_t batch = 0; batch < oDims[0]; ++batch) {
+        outOffsets[0] = batch * oStrides[0];
+
+        for (DimSize_t outCh = 0; outCh < oDims[1]; ++outCh) {
+            outOffsets[1] = outCh * oStrides[1] + outOffsets[0];
+            // If bias = nullptr, set B(0)
+            B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
+            std::fill(output + outOffsets[1],
+                      (output + outOffsets[1]) + oStrides[1],
+                      biasVal);
+        }
+    }
+}
+
+/**
+ * @brief forward kernel for convtranspose
+ * @note ConvTranspose forward is simply convolution backward kernel.
+ * Check convolution functions for more in-depth details on how the
+ subfunctions are built.
+ * @tparam I Input data type.
+ * @tparam W Weight data type.
+ * @tparam B Bias data type.
+ * @tparam O Output data type.
+ * @param[in] stride stride parameter of the convTranspose operator
+ * @param[in] dilation dilation parameter of the convTranspose operator
+ * @param[in] inputDims input dimensions
+ * @param[in] outputDims output tensor dimensions
+ * @param[in] oStrides nb of elements contained per dimension of the output
+ * @param[in] input_ values
+ * @param[in] weight_ values
+ * @param[in] biases_ values
+ * @param[out] output
+ */
+template <class I, class W, class B, class O>
+void ConvTransposeImpl2D_cpu_forward_kernel(
+    const array<DimSize_t, 2> &stride,
+    const array<DimSize_t, 2> &dilation,
+    const array<DimSize_t, 2> &kernelDims,
+    const array<DimSize_t, 4> &inputDims,
+    const array<DimSize_t, 4> &outputDims,
+    const void *input_,
+    const void *weights_,
+    const void *biases_,
+    void *output_) {
+
+    auto input = static_cast<const I *>(input_);
+    auto weights = static_cast<const W *>(weights_);
+    auto output = static_cast<O *>(output_);
+
+    // {channel_stride, dim0_stride, dim1_stride}
+    const array<DimSize_t, 3> inputStrides{
+        inputDims[1] * inputDims[2] * inputDims[3],
+        inputDims[2] * inputDims[3],
+        inputDims[3]};
+
+    // {channel_stride, dim0_stride, dim1_stride}
+    const array<DimSize_t, 3> outputStrides{
+        outputDims[1] * outputDims[2] * outputDims[3],
+        outputDims[2] * outputDims[3],
+        outputDims[3]};
+
+    // NOTE: kernel dims = {inChannels, outChannels, kernelDims[0],
+    // kernelDims[1]}
+    const array<DimSize_t, 3> kernelStrides{
+        outputDims[1] * kernelDims[0] * kernelDims[1],
+        kernelDims[0] * kernelDims[1],
+        kernelDims[1],
+    };
+
+    if (biases_ != nullptr) {
+        auto biases = static_cast<const B *>(biases_);
+        convTranspose2DForwardBias(biases, outputDims, outputStrides, output);
+    }
+
+    conv2DBackwardInput(stride,
+                        dilation,
+                        kernelDims,
+                        kernelStrides,
+                        weights,
+                        inputDims,
+                        inputStrides,
+                        input,
+                        outputDims,
+                        outputStrides,
+                        output);
+}
+
+REGISTRAR(ConvTransposeImpl2D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Int32, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl2D_cpu_forward_kernel<std::int32_t,
+                                                  std::int32_t,
+                                                  std::int32_t,
+                                                  std::int32_t>,
+           nullptr});
+REGISTRAR(ConvTransposeImpl2D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Float16, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl2D_cpu_forward_kernel<half_float::half,
+                                                  half_float::half,
+                                                  half_float::half,
+                                                  half_float::half>,
+           nullptr});
+REGISTRAR(ConvTransposeImpl2D_cpu,
+          {{DataType::Any, DataFormat::NCHW},
+           {DataType::Float32, DataFormat::NCHW}},
+          {ProdConso::inPlaceModel,
+           ConvTransposeImpl2D_cpu_forward_kernel<float, float, float, float>,
+           nullptr});
+REGISTRAR(
+    ConvTransposeImpl2D_cpu,
+    {{DataType::Any, DataFormat::NCHW}, {DataType::Float64, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel,
+     ConvTransposeImpl2D_cpu_forward_kernel<double, double, double, double>,
+     nullptr});
+
+} // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_CONVTRANSPOSEIMPL_KERNELS_H_ */

src/operator/ConvTransposeImpl.cpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#include "aidge/backend/cpu/operator/ConvTransposeImpl.hpp"
+#include "aidge/backend/cpu/operator/ConvTransposeImpl_kernels.hpp"
+
+template <> void Aidge::ConvTransposeImpl1D_cpu::forward() {
+    const auto &op = static_cast<const ConvTranspose_Op<1> &>(mOp);
+
+    AIDGE_ASSERT(op.getInput(0), "{}: missing data input (#0).", op.type());
+    AIDGE_ASSERT(op.getInput(1), "{}: missing bias input (#1).", op.type());
+    AIDGE_ASSERT(op.getInput(2), "{}: missing weight input (#1).", op.type());
+
+    std::shared_ptr<Tensor> inputDataFallback, inputWeightFallback,
+        inputBiasFallback;
+    const auto &inputData =
+        op.getInput(0)->refCastFrom(inputDataFallback, *op.getOutput(0));
+    const auto &inputWeight =
+        op.getInput(1)->refCastFrom(inputWeightFallback, *op.getOutput(0));
+    const auto &inputBias =
+        (op.getInput(2))
+            ? op.getInput(2)->refCastFrom(inputBiasFallback, *op.getOutput(0))
+            : Tensor();
+
+    // Call kernel
+    const auto impl = Registrar<ConvTransposeImpl1D_cpu>::create(
+        getBestMatch(getRequiredSpec()));
+    impl.forward(op.strideDims(),
+                 op.dilationDims(),
+                 op.kernelDims(),
+                 op.getInput(0)->template dims<3>(),
+                 op.getOutput(0)->template dims<3>(),
+                 inputData.getImpl()->hostPtr(),
+                 inputWeight.getImpl()->hostPtr(),
+                 op.getInput(2) ? inputBias.getImpl()->hostPtr() : nullptr,
+                 op.getOutput(0)->getImpl()->rawPtr());
+}
+
+template <> void Aidge::ConvTransposeImpl1D_cpu::backward() {
+    AIDGE_THROW_OR_ABORT(
+        std::runtime_error,
+        "Backward not yet implemented for Conv_Op<1> on backend cpu");
+}
+
+template <> void Aidge::ConvTransposeImpl2D_cpu::forward() {
+    const auto &op = static_cast<const ConvTranspose_Op<2> &>(mOp);
+
+    AIDGE_ASSERT(op.getInput(0), "{}: missing data input (#0).", op.type());
+    AIDGE_ASSERT(op.getInput(1), "{}: missing bias input (#1).", op.type());
+    AIDGE_ASSERT(op.getInput(2), "{}: missing weight input (#1).", op.type());
+
+    std::shared_ptr<Tensor> inputDataFallback, inputWeightFallback,
+        inputBiasFallback;
+    const auto &inputData =
+        op.getInput(0)->refCastFrom(inputDataFallback, *op.getOutput(0));
+    const auto &inputWeight =
+        op.getInput(1)->refCastFrom(inputWeightFallback, *op.getOutput(0));
+    const auto &inputBias =
+        (op.getInput(2))
+            ? op.getInput(2)->refCastFrom(inputBiasFallback, *op.getOutput(0))
+            : Tensor();
+
+    // Call kernel
+    const auto impl = Registrar<ConvTransposeImpl2D_cpu>::create(
+        getBestMatch(getRequiredSpec()));
+
+    impl.forward(op.strideDims(),
+                 op.dilationDims(),
+                 op.kernelDims(),
+                 op.getInput(0)->template dims<4>(),
+                 op.getOutput(0)->template dims<4>(),
+                 inputData.getImpl()->hostPtr(),
+                 inputWeight.getImpl()->hostPtr(),
+                 op.getInput(2) ? inputBias.getImpl()->hostPtr() : nullptr,
+                 op.getOutput(0)->getImpl()->rawPtr());
+}
+
+template <> void Aidge::ConvTransposeImpl2D_cpu::backward() {
+    AIDGE_THROW_OR_ABORT(
+        std::runtime_error,
+        "Backward not yet implemented for Conv_Op<2> on backend cpu");
+}
+