diff --git a/include/aidge/backend/cpu/operator/ScalingImpl_forward_kernels.hpp b/include/aidge/backend/cpu/operator/ScalingImpl_forward_kernels.hpp
index 5258c4c3e7376c3883b119503ee9e6765de844d5..df8e1a7e7b02a4ad032d6f09fae3ae2cd8a42eff 100644
--- a/include/aidge/backend/cpu/operator/ScalingImpl_forward_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/ScalingImpl_forward_kernels.hpp
@@ -9,13 +9,12 @@
*
********************************************************************************/
-#ifndef AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H
-#define AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H
+#ifndef AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H_
+#define AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H_
#include <cmath>
-
+#include <cstddef>
#include "aidge/utils/Registrar.hpp"
-
#include "aidge/backend/cpu/operator/ScalingImpl.hpp"
//TODO : improve propagate, n2d2 :
@@ -61,12 +60,13 @@ const O& clamp(const O& x, const O& min, const O& max)
}
template<class O>
-O saturate(O value, std::size_t quantizedNbBits, bool isOutputUnsigned) {
+O saturate(const O value, const std::size_t quantizedNbBits, const bool isOutputUnsigned) {
+ // TODO: no assertions in kernel
assert(quantizedNbBits > 0);
- const O min = isOutputUnsigned?0:
+ const O min = isOutputUnsigned ? 0 :
-(1ll << (quantizedNbBits - 1ll));
- const O max = isOutputUnsigned?(1ll << quantizedNbBits) - 1ll:
+ const O max = isOutputUnsigned ? (1ll << quantizedNbBits) - 1ll :
(1ll << (quantizedNbBits - 1ll)) - 1ll;
return clamp(value, min, max);
@@ -81,8 +81,8 @@ void ScalingImpl_cpu_forward_kernel(const Scaling_Op::Attrs& attrs,
const I* input = static_cast<const I*>(input_);
O* output = static_cast<O*>(output_);
const I& scalingFactor = static_cast<const I&>(std::get<0>(attrs));
- std::size_t quantizedNbBits = static_cast<std::size_t>(std::get<1>(attrs));
- bool isOutputUnsigned = static_cast<bool>(std::get<2>(attrs));
+ const std::size_t quantizedNbBits = static_cast<std::size_t>(std::get<1>(attrs));
+ const bool isOutputUnsigned = static_cast<bool>(std::get<2>(attrs));
for (std::size_t i = 0; i < inputLenght; ++i) {
output[i] = input[i] * scalingFactor;
@@ -103,4 +103,4 @@ static Registrar<ScalingImplForward_cpu> registrarScalingImplForward_cpu_Float64
} // namespace
} // namespace Aidge
-#endif /* AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H */
+#endif /* AIDGE_CPU_OPERATOR_SCALINGIMPL_FORWARD_KERNEL_H_ */
\ No newline at end of file
diff --git a/include/aidge/backend/cpu/operator/SliceImpl.hpp b/include/aidge/backend/cpu/operator/SliceImpl.hpp
index d6f7caad4febf72764892f956886c8fb6875d0ab..1cba5906064c51a4f0da2f1f3682b0828a080d43 100644
--- a/include/aidge/backend/cpu/operator/SliceImpl.hpp
+++ b/include/aidge/backend/cpu/operator/SliceImpl.hpp
@@ -26,15 +26,17 @@ namespace Aidge {
// compute kernel registry for forward and backward
class SliceImplForward_cpu
- : public Registrable<SliceImplForward_cpu,
- std::tuple<DataType, DataType>,
- void(const std::vector<DimSize_t>&, const void*, DimSize_t, const void*, const void*, const void*, void*)> {
-};
+ : public Registrable<SliceImplForward_cpu, std::tuple<DataType>,
+ void(const typename Slice_Op::Attrs&,
+ const std::vector<std::size_t>,
+ const void*,
+ void*)> {};
class SliceImplBackward_cpu
- : public Registrable<SliceImplBackward_cpu,
- std::tuple<DataType, DataType>,
- void(const std::vector<DimSize_t>&, const void*, DimSize_t, const void*, const void*, const void*, void*)> {
-};
+ : public Registrable<SliceImplBackward_cpu, std::tuple<DataType>,
+ void(const typename Slice_Op::Attrs&,
+ const std::vector<std::size_t>,
+ const void*,
+ void*)> {};
class SliceImpl_cpu : public OperatorImpl {
public:
@@ -61,4 +63,4 @@ static Registrar<Slice_Op> registrarSliceImpl_cpu("cpu", Aidge::SliceImpl_cpu::c
}
} // namespace Aidge
-#endif /* AIDGE_CPU_OPERATOR_SLICEIMPL_H_ */
\ No newline at end of file
+#endif /* AIDGE_CPU_OPERATOR_SLICEIMPL_H_ */
diff --git a/include/aidge/backend/cpu/operator/SliceImpl_forward_kernels.hpp b/include/aidge/backend/cpu/operator/SliceImpl_forward_kernels.hpp
index 9381039bb4d8fcdc5aa34c9d34d96ec50a73d651..80c036cded168aea017f3ae8e2c004474b3977f3 100644
--- a/include/aidge/backend/cpu/operator/SliceImpl_forward_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/SliceImpl_forward_kernels.hpp
@@ -13,94 +13,82 @@
#define AIDGE_CPU_OPERATOR_SLICEIMPL_FORWARD_KERNEL_H_
#include "aidge/utils/Registrar.hpp"
-
+#include "aidge/operator/Slice.hpp"
#include "aidge/backend/cpu/operator/SliceImpl.hpp"
+#include <vector>
+#include <cstddef>
-namespace Aidge {
-template <class I, class O>
-void SliceImpl_cpu_forward_kernel(const std::vector<DimSize_t>& inputDims,
- const void* input_,
- DimSize_t nbSlices,
- const void* axes_,
- const void* starts_,
- const void* ends_,
- void* output_) {
- const I* input = static_cast<const I*>(input_);
- const int* axes = static_cast<const int*>(axes_);
- const int* starts = static_cast<const int*>(starts_);
- const int* ends = static_cast<const int*>(ends_);
- O* output = static_cast<O*>(output_);
-
- // Calculate the total number of elements in the input array
- size_t totalElements = 1;
- for (size_t dimSize : inputDims) {
- totalElements *= dimSize;
- }
-
- // Create a temporary arrays to store intermediate input/output for each slice op
- std::vector<I> tempInArray(input, input + totalElements);
- std::vector<I> tempOutArray(input, input + totalElements);
- std::vector<size_t> currentDims = inputDims;
+#include "aidge/data/Data.hpp"
- size_t copiedElems = 0;
- // Loop over each slice operation
- for(size_t i=0; i< nbSlices; ++i)
+namespace Aidge {
+template <class I>
+void SliceImpl_cpu_forward_kernel(const typename Slice_Op::Attrs& attrs,
+ const std::vector<std::size_t> inputDims,
+ const void* input_,
+ void* output_) {
+ std::vector<std::size_t> slicedDims = inputDims;
+
+ std::size_t beginning = 0;
+ DimSize_t nbAxes = std::get<2>(attrs).size();
+ for(std::size_t i=0; i<nbAxes;++i)
{
- copiedElems = 0;
- I* tempOutArrayPtr = tempOutArray.data();
- // Extract parameters for the current slice, make sure indexes are positive
- size_t axisIdx = axes[i]>=0?axes[i]:(axes[i]+currentDims.size());
- size_t startIdx = starts[i]>=0?starts[i]:(starts[i]+currentDims[axisIdx]);
- size_t endIdx = ends[i]>=0?ends[i]:(ends[i]+currentDims[axisIdx]);
-
-
- // Compute the size of the slice over each element on the axis
- size_t strideOnCurrDim = 1;
- for(size_t j=(axisIdx+1); j<currentDims.size(); ++j)
- {
- strideOnCurrDim *= currentDims[j];
- }
- size_t sliceSize = (endIdx - startIdx + 1) * strideOnCurrDim;
+ // For each slice operation get the params and cast them to size_t
+ int axis_ = std::get<2>(attrs)[i];
+ int start_ = std::get<0>(attrs)[i];
+ int end_ = std::get<1>(attrs)[i];
+ std::size_t axis = axis_>=0?axis_:axis_+inputDims.size();
+ std::size_t start = start_>=0?start_:start_+inputDims[axis];
+ std::size_t end = end_>=0?end_:end_+inputDims[axis];
+ std::size_t stride=1;
+ for(std::size_t j = inputDims.size()-1; j>axis; --j)
+ stride*=inputDims[j];
+ beginning += start * stride;
+ std::size_t sliceLength = end - start + 1;
+ slicedDims[axis] = sliceLength;
+ }
- // For each slice operation, we will slice all elements on the axis (subSlice)
- // the number of sublices is the product of dimension previous to the slice dimension
- size_t nbSubSlices = 1;
- for(size_t j=0; j<axisIdx; ++j)
- {
- nbSubSlices*=currentDims[j];
- }
- // Operate the slice over each element of the dim we want to slice
- for(size_t s=0; s<nbSubSlices; ++s)
- {
- // Compute the pointer postion on input
- std::size_t copyStartPos = s * strideOnCurrDim * currentDims[axisIdx] + startIdx * strideOnCurrDim;
- const I* copyPtr = std::next(tempInArray.data(), copyStartPos);
- // Copy slice to output array and update pointer
- std::copy_n(copyPtr, sliceSize , tempOutArrayPtr);
- tempOutArrayPtr += sliceSize ;
- copiedElems+= sliceSize ;
- }
+ const I* input = static_cast<const I*>(input_) + beginning;
+ I* output = static_cast<I*>(output_);
+ const std::size_t nbDims = slicedDims.size();
- // Update the input for the next slice operation
- tempInArray.assign(tempOutArray.begin(), tempOutArray.begin() + copiedElems);
- currentDims[axisIdx] = endIdx - startIdx + 1;
+ // for inputDims = {4,5,5,3} & slicedDims = {3,2,2,1}, substractDims = {1,5,5,3}
+ std::vector<std::size_t> substractedDims = std::vector<std::size_t>(nbDims);
+ for (std::size_t i = 0; i < nbDims; ++i) {
+ substractedDims[i] = inputDims[i] - slicedDims[i];
}
- std::copy_n(tempInArray.data(), copiedElems, output);
+ // for slicedDims = {3,2,2,1}, prodSlicedDims = {12,4,2,1}
+ std::vector<std::size_t> prodSlicedDims = std::vector<std::size_t>(nbDims);
+ std::vector<std::size_t> prodInputDims = std::vector<std::size_t>(nbDims+1);
+ prodSlicedDims[nbDims - 1] = slicedDims[nbDims - 1];
+ prodInputDims[nbDims - 1] = inputDims[nbDims - 1];
+ prodInputDims[nbDims] = 1;
+ for (std::size_t i = 2; i <= nbDims; ++i) {
+ prodSlicedDims[nbDims - i] = prodSlicedDims[nbDims - i + 1]*slicedDims[nbDims - i];
+ prodInputDims[nbDims - i] = prodInputDims[nbDims - i + 1]*inputDims[nbDims - i];
+ }
+
+ std::size_t j = 0;
+ std::size_t i = 0;
+ for (; j < prodSlicedDims[0];) {
+ output[j] = input[i++];
+ ++j;
+ for (std::size_t idx = nbDims - 1; idx > 0; --idx) {
+ i += j % prodSlicedDims[idx] == 0 ? substractedDims[idx]*prodInputDims[idx+1] : 0;
+ }
+ }
}
namespace {
+
static Registrar<SliceImplForward_cpu> registrarSliceImplForward_cpu_Float32(
- {DataType::Float32, DataType::Float32},
- Aidge::SliceImpl_cpu_forward_kernel<float, float>);
+ {DataType::Float32}, Aidge::SliceImpl_cpu_forward_kernel<float>);
static Registrar<SliceImplForward_cpu> registrarSliceImplForward_cpu_Int32(
- {DataType::Int32, DataType::Int32},
- Aidge::SliceImpl_cpu_forward_kernel<int, int>);
+ {DataType::Int32}, Aidge::SliceImpl_cpu_forward_kernel<int>);
static Registrar<SliceImplForward_cpu> registrarSliceImplForward_cpu_Float64(
- {DataType::Float64, DataType::Float64},
- Aidge::SliceImpl_cpu_forward_kernel<double, double>);
+ {DataType::Float64}, Aidge::SliceImpl_cpu_forward_kernel<double>);
} // namespace
} // namespace Aidge
-#endif /* AIDGE_CPU_OPERATOR_SLICEIMPL_FORWARD_KERNEL_H_ */
\ No newline at end of file
+#endif /* AIDGE_CPU_OPERATOR_SLICEIMPL_FORWARD_KERNEL_H_ */
diff --git a/src/operator/SliceImpl.cpp b/src/operator/SliceImpl.cpp
index 6a528afdbc1399994139f5a1b8336d04ec582159..32d31f046465425a269d6f8e3fc52eaad31c663a 100644
--- a/src/operator/SliceImpl.cpp
+++ b/src/operator/SliceImpl.cpp
@@ -18,22 +18,29 @@
#include "aidge/backend/cpu/operator/SliceImpl.hpp"
#include "aidge/backend/cpu/operator/SliceImpl_forward_kernels.hpp"
#include "aidge/utils/Types.h"
-
+#include <vector>
+#include <cassert>
+#include <tuple>
Aidge::NbElts_t Aidge::SliceImpl_cpu::getNbRequiredData(const Aidge::IOIndex_t /*inputIdx*/) const {
assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(0)) && "requires valid input");
// Requires the whole tensors
- return std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims()[0];
+ const auto& inputDims = std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims();
+
+ return std::accumulate(inputDims.begin(), inputDims.end(), static_cast<NbElts_t>(1),
+ std::multiplies<NbElts_t>());
}
Aidge::NbElts_t Aidge::SliceImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex_t /*inputIdx*/) const { return 0; }
Aidge::NbElts_t Aidge::SliceImpl_cpu::getRequiredMemory(const Aidge::IOIndex_t outputIdx,
- const std::vector<Aidge::DimSize_t>& inputsSize) const {
+ const std::vector<Aidge::DimSize_t>& inputsSize) const {
(void)outputIdx;
(void)inputsSize;
- return std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims()[0];
+ const auto& outputDims = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims();
+ return std::accumulate(outputDims.begin(), outputDims.end(), static_cast<NbElts_t>(1),
+ std::multiplies<NbElts_t>());
}
Aidge::NbElts_t Aidge::SliceImpl_cpu::getNbConsumedData(const Aidge::IOIndex_t /*inputIdx*/) const {
@@ -52,27 +59,24 @@ void Aidge::SliceImpl_cpu::updateConsummerProducer() {
}
void Aidge::SliceImpl_cpu::forward() {
- for (IOIndex_t i = 0; i < 4; ++i)
- assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(i)) && ("missing input"));
-
-
- assert((std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->nbDims() == 1) && "input #1 must either be a tensor of rank 1");
- assert((std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->nbDims() == 1) && "input #2 must either be a tensor of rank 1");
- assert((std::static_pointer_cast<Tensor>(mOp.getRawInput(3))->nbDims() == 1) && "input #3 must either be a tensor of rank 1");
+ // FIXME: uncomment the following code once memory handling will work
+ assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(0)) && "missing input #0");
// Find the correct kernel type
- auto kernelFunc = Registrar<SliceImplForward_cpu>::create({
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType()});
+ auto kernelFunc = Registrar<SliceImplForward_cpu>::create(
+ {std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType()});
// Call kernel
- kernelFunc(std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->dims()[0],
- std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(3))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->getImpl()->rawPtr());
+ kernelFunc(dynamic_cast<const Slice_Op&>(mOp).getStaticAttributes(),
+ std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims(),
+ std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->getImpl()->rawPtr(),
+ std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->getImpl()->rawPtr()
+ );
+
+ // each input is consumed by the minimum amount for a forward pass
+ mNbConsumedData[0] += getNbRequiredData(0);
+
+ mNbProducedData[0] += getRequiredMemory(0, {});
}
void Aidge::SliceImpl_cpu::backward() { printf("Not implemented yet.\n"); }
diff --git a/unit_tests/operator/Test_SliceImpl.cpp b/unit_tests/operator/Test_SliceImpl.cpp
index e9392d3a01dabf427813b653916a79f07093c41d..8d35c885871a83c88f2852fbd946a1fb7f74dfee 100644
--- a/unit_tests/operator/Test_SliceImpl.cpp
+++ b/unit_tests/operator/Test_SliceImpl.cpp
@@ -29,21 +29,15 @@ TEST_CASE("[cpu/operator] Slice(forward)") {
{5, 6, 7, 8}
}
});
- std::shared_ptr<Tensor> axes = std::make_shared<Tensor>(Array1D<int,2>{{0, 1}});
- std::shared_ptr<Tensor> starts = std::make_shared<Tensor>(Array1D<int,2>{{1, 1}});
- std::shared_ptr<Tensor> ends = std::make_shared<Tensor>(Array1D<int,2>{{1, 3}});
std::shared_ptr<Tensor> expectedOutput = std::make_shared<Tensor>(Array2D<int,1,3> {
{
{6, 7, 8}
}
});
- std::shared_ptr<Node> mySlice = Slice();
+ std::shared_ptr<Node> mySlice = Slice({1, 1}, {1, 3}, {0, 1});
auto op = std::static_pointer_cast<OperatorTensor>(mySlice -> getOperator());
op->associateInput(0, input);
- op->associateInput(1, axes);
- op->associateInput(2, starts);
- op->associateInput(3, ends);
op->setDataType(DataType::Int32);
op->setBackend("cpu");
op->computeOutputDims();
@@ -69,9 +63,6 @@ TEST_CASE("[cpu/operator] Slice(forward)") {
}
}
});
- std::shared_ptr<Tensor> axes = std::make_shared<Tensor>(Array1D<int,2>{{1, 2}});
- std::shared_ptr<Tensor> starts = std::make_shared<Tensor>(Array1D<int,2>{{0, 2}});
- std::shared_ptr<Tensor> ends = std::make_shared<Tensor>(Array1D<int,2>{{2, 2}});
std::shared_ptr<Tensor> expectedOutput = std::make_shared<Tensor>(Array3D<int,2,3,1> {
{
{
@@ -87,12 +78,9 @@ TEST_CASE("[cpu/operator] Slice(forward)") {
}
});
- std::shared_ptr<Node> mySlice = Slice();
+ std::shared_ptr<Node> mySlice = Slice({0, 2}, {2, 2}, {1, 2});
auto op = std::static_pointer_cast<OperatorTensor>(mySlice -> getOperator());
op->associateInput(0, input);
- op->associateInput(1, axes);
- op->associateInput(2, starts);
- op->associateInput(3, ends);
op->setDataType(DataType::Int32);
op->setBackend("cpu");
op->computeOutputDims();
diff --git a/unit_tests/recipies/Test_HorizontalTiling.cpp b/unit_tests/recipies/Test_HorizontalTiling.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..b71a01d130a783caf5c643dfb0c3757b1c524e5e
--- /dev/null
+++ b/unit_tests/recipies/Test_HorizontalTiling.cpp
@@ -0,0 +1,208 @@
+/********************************************************************************
+ * 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 <catch2/catch_test_macros.hpp>
+#include <set>
+
+#include "aidge/graph/GraphView.hpp"
+#include "aidge/graph/OpArgs.hpp"
+#include "aidge/operator/Conv.hpp"
+#include "aidge/operator/ReLU.hpp"
+#include "aidge/recipies/Recipies.hpp"
+#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/operator/Concat.hpp"
+
+
+namespace Aidge {
+
+TEST_CASE("[core/recipies] Tiling(transformation)", "[Tiling][Recipies]") {
+
+ SECTION("Transform a pre-generated GraphView") {
+
+ SECTION("Simple Node: Conv") {
+ std::shared_ptr<Node> myReLU = ReLU("myReLU");
+ std::shared_ptr<Node> myConv = Conv(3,4,{3,3}, "myconv");
+ std::shared_ptr<Tensor> myWeights = std::make_shared<Tensor>(Array4D<int,4,3,3,3> {
+ {
+ {
+ {{ 0, 1, 2},
+ { 3, 4, 5},
+ { 6, 7, 8}},
+ {{ 9, 10, 11},
+ { 12, 13, 14},
+ { 15, 16, 17}},
+ {{ 18, 19, 20},
+ { 21, 22, 23},
+ { 24, 25, 26}}
+ },
+ {
+ {{ 27, 28, 29},
+ { 30, 31, 32},
+ { 33, 34, 35}},
+ {{ 36, 37, 38},
+ { 39, 40, 41},
+ { 42, 43, 44}},
+ {{ 45, 46, 47},
+ { 48, 49, 50},
+ { 51, 52, 53}}
+ },
+ {
+ {{ 54, 55, 56},
+ { 57, 58, 59},
+ { 60, 61, 62}},
+ {{ 63, 64, 65},
+ { 66, 67, 68},
+ { 69, 70, 71}},
+ {{ 72, 73, 74},
+ { 75, 76, 77},
+ { 78, 79, 80}}
+ },
+ {
+ {{ 81, 82, 83},
+ { 84, 85, 86},
+ { 87, 88, 89}},
+ {{ 90, 91, 92},
+ { 93, 94, 95},
+ { 96, 97, 98}},
+ {{ 99, 100, 101},
+ {102, 103, 104},
+ {105, 106, 107}}
+ }
+ }
+ });
+ std::shared_ptr<Tensor> myBias = std::make_shared<Tensor>(Array1D<int,4> {{7,0,9,0}});
+ std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(Array4D<int,2,3,5,5> { //NCHW
+ {
+ {
+ {{ 0, 1, 2, 3, 4},
+ { 5, 6, 7, 8, 9},
+ { 10, 11, 12, 13, 14},
+ { 15, 16, 17, 18, 19},
+ { 20, 21, 22, 23, 24}},
+
+ {{ 25, 26, 27, 28, 29},
+ { 30, 31, 32, 33, 34},
+ { 35, 36, 37, 38, 39},
+ { 40, 41, 42, 43, 44},
+ { 45, 46, 47, 48, 49}},
+
+ {{ 50, 51, 52, 53, 54},
+ { 55, 56, 57, 58, 59},
+ { 60, 61, 62, 63, 64},
+ { 65, 66, 67, 68, 69},
+ { 70, 71, 72, 73, 74}}
+ },
+ {
+ {{ 75, 76, 77, 78, 79},
+ { 80, 81, 82, 83, 84},
+ { 85, 86, 87, 88, 89},
+ { 90, 91, 92, 93, 94},
+ { 95, 96, 97, 98, 99}},
+
+ {{100, 101, 102, 103, 104},
+ {105, 106, 107, 108, 109},
+ {110, 111, 112, 113, 114},
+ {115, 116, 117, 118, 119},
+ {120, 121, 122, 123, 124}},
+
+ {{125, 126, 127, 128, 129},
+ {130, 131, 132, 133, 134},
+ {135, 136, 137, 138, 139},
+ {140, 141, 142, 143, 144},
+ {145, 146, 147, 148, 149}}
+ }
+ }
+ });
+ std::shared_ptr<Tensor> myOutput = std::make_shared<Tensor>(Array4D<int,2,4,3,3> {
+ {
+ {
+ {{ 15226, 15577, 15928},
+ { 16981, 17332, 17683},
+ { 18736, 19087, 19438}},
+
+ {{ 37818, 38898, 39978},
+ { 43218, 44298, 45378},
+ { 48618, 49698, 50778}},
+
+ {{ 60426, 62235, 64044},
+ { 69471, 71280, 73089},
+ { 78516, 80325, 82134}},
+
+ {{ 83016, 85554, 88092},
+ { 95706, 98244, 100782},
+ {108396, 110934, 113472}}
+ },
+ {
+ {{ 41551, 41902, 42253},
+ { 43306, 43657, 44008},
+ { 45061, 45412, 45763}},
+
+ {{118818, 119898, 120978},
+ {124218, 125298, 126378},
+ {129618, 130698, 131778}},
+
+ {{196101, 197910, 199719},
+ {205146, 206955, 208764},
+ {214191, 216000, 217809}},
+
+ {{273366, 275904, 278442},
+ {286056, 288594, 291132},
+ {298746, 301284, 303822}}
+ }
+ }
+ });
+ myReLU->getOperator()->associateInput(0, myInput);
+ myReLU->addChild(myConv, 0, 0);
+ myConv->getOperator()->setInput(1, myWeights);
+ myConv->getOperator()->setInput(2, myBias);
+ std::dynamic_pointer_cast<Conv_Op<2>>(myConv->getOperator())->computeOutputDims();
+
+ std::shared_ptr<GraphView> g = std::make_shared<GraphView>();
+ g->add({myReLU, myConv});
+ g->compile("cpu", DataType::Int32);
+ std::set<std::shared_ptr<Node>> tiledConv = getConvHorizontalTiling(myConv, 2, 3);
+
+ SequentialScheduler s(g);
+ s.forward();
+ REQUIRE(*(std::dynamic_pointer_cast<Conv_Op<2>>(myConv->getOperator())->getOutput(0)) == *myOutput);
+
+ GraphView::replace({myConv, myConv->getParent(1), myConv->getParent(2)}, tiledConv);
+ g->compile("cpu", DataType::Int32);
+ s.resetScheduling();
+ s.forward();
+
+ REQUIRE(*(std::dynamic_pointer_cast<OperatorTensor>((*g->outputNodes().begin())->getOperator())->getOutput(0)) == *myOutput);
+ }
+ }
+}
+}
+ // std::shared_ptr<GraphView> g = Sequential({
+ // Conv(3, 16, {3,3}, "conv1"),
+ // ReLU("relu1"),
+ // Conv(16, 32, {1,1}, "conv2"),
+ // Conv(32, 16, {1,1}, "conv3"),
+ // Conv(16, 10, {3,3}, "conv4"),
+ // ReLU("relu2")
+ // });
+
+ // for (auto& individualConv : g->match("Conv")) {
+ // auto tiledConv = horizontalTiling(individualConv);
+ // g->replace(individualConv, tiledConv);
+ // }
+ // }
+
+ // SECTION("Create the GraphView with tiled layers") {
+ // std::shared_ptr<GraphView> g;
+ // g->addChild(horizontalTiling(Conv()))
+ // }
+
+// }
+// } // namespace Aidge
\ No newline at end of file