diff --git a/include/aidge/backend/cpu/data/TensorImpl.hpp b/include/aidge/backend/cpu/data/TensorImpl.hpp
index 15e8ca34c05dd61e094b589956c187f89abcf450..0b77f2e44c9ab22f222f9a6b7b073bf12d42e8f0 100644
--- a/include/aidge/backend/cpu/data/TensorImpl.hpp
+++ b/include/aidge/backend/cpu/data/TensorImpl.hpp
@@ -55,6 +55,10 @@ class TensorImpl_cpu : public TensorImpl {
}
void copyCast(const void *src, NbElts_t length, const DataType srcDt) override {
+ if (length == 0) {
+ return;
+ }
+
if (srcDt == DataType::Float64) {
std::copy(static_cast<const double*>(src), static_cast<const double*>(src) + length,
static_cast<T *>(rawPtr()));
@@ -151,8 +155,6 @@ class TensorImpl_cpu : public TensorImpl {
private:
void lazyInit() {
- AIDGE_INTERNAL_ASSERT(mTensor.dataType() == NativeType<T>::type);
-
if (mData.size() < mTensor.size()) {
// Need more data, a re-allocation will occur
AIDGE_ASSERT(mData.empty() || mDataOwner != nullptr, "trying to enlarge non-owned data");
diff --git a/src/operator/AddImpl.cpp b/src/operator/AddImpl.cpp
index fc54eca65a2638c53498c47ea8a21682f31e9eea..a6446641867d69551a8ba410016bd3e03c35e735 100644
--- a/src/operator/AddImpl.cpp
+++ b/src/operator/AddImpl.cpp
@@ -33,16 +33,36 @@ void Aidge::AddImpl_cpu::forward() {
assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->dataType() == datatypeFirstInput);
}
- auto kernelFunc = Registrar<AddImplForward_cpu>::create({
+ // Find the correct kernel type
+ const auto outputDataType = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType();
+ const Registrar<AddImplForward_cpu>::registrar_key registrarKey = {
datatypeFirstInput,
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType()});
+ outputDataType};
+ Registrar<AddImplForward_cpu>::registrar_type kernelFunc;
+ if (Registrar<AddImplForward_cpu>::exists(registrarKey)) {
+ // One exists with the right inputs/output types
+ kernelFunc = Registrar<AddImplForward_cpu>::create(registrarKey);
+ }
+ else {
+ // Otherwise, fallback to the kernel with all types matching output type
+ kernelFunc = Registrar<AddImplForward_cpu>::create({
+ outputDataType, outputDataType});
+ }
+
+ // Convert input data (no overhead if not needed!)
+ // TODO: right now, if needed, memory will be allocated/deallocated at each
+ // call to forward(). We might put the following shared_ptr as members of
+ // this class to avoid that.
std::vector<const void*> opInputs;
+ std::vector<std::shared_ptr<Tensor>> inputsFallback(mOp.nbInputs());
for (IOIndex_t i = 0; i < mOp.nbInputs(); ++i) {
- opInputs.push_back(std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->getImpl()->rawPtr());
+ const auto& input = std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->refCast(inputsFallback[i], *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
+ opInputs.push_back(input.getImpl()->rawPtr());
}
+ // Call kernel
kernelFunc(std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->size(),
opInputs,
std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->getImpl()->rawPtr());
-}
\ No newline at end of file
+}
diff --git a/src/operator/FCImpl.cpp b/src/operator/FCImpl.cpp
index 1e5450d330ee89bdceb30aca846800d7764ca911..e97fe4ccc2f12b34d9b799b0435547d329182dae 100644
--- a/src/operator/FCImpl.cpp
+++ b/src/operator/FCImpl.cpp
@@ -28,29 +28,37 @@ void Aidge::FCImpl_cpu::forward()
assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(2)) && "missing input #2");
// Find the correct kernel type
- auto kernelFunc = Registrar<FCImplForward_cpu>::create(
- {std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType()});
+ const auto outputDataType = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType();
+ const Registrar<FCImplForward_cpu>::registrar_key registrarKey = {
+ std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType(),
+ std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->dataType(),
+ std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->dataType(),
+ outputDataType};
+
+ Registrar<FCImplForward_cpu>::registrar_type kernelFunc;
+ if (Registrar<FCImplForward_cpu>::exists(registrarKey)) {
+ // One exists with the right inputs/output types
+ kernelFunc = Registrar<FCImplForward_cpu>::create(registrarKey);
+ }
+ else {
+ // Otherwise, fallback to the kernel with all types matching output type
+ kernelFunc = Registrar<FCImplForward_cpu>::create({
+ outputDataType, outputDataType, outputDataType, outputDataType});
+ }
+
+ // Convert input data (no overhead if not needed!)
+ // TODO: right now, if needed, memory will be allocated/deallocated at each
+ // call to forward(). We might put the following shared_ptr as members of
+ // this class to avoid that.
+ std::shared_ptr<Tensor> input0Fallback, input1Fallback, input2Fallback;
+ const auto& input0 = std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->refCast(input0Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
+ const auto& input1 = std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->refCast(input1Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
+ const auto& input2 = std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->refCast(input2Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
// Call kernel
- // if (std::static_pointer_cast<Tensor>(mOp.getRawInput(0)->nbDims() == 4) {
- // kernelFunc(
- // mOp.getStaticAttributes(),
- // std::static_pointer_cast<Tensor>(std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->template dims<4>(),
- // std::static_pointer_cast<Tensor>(mOp.getRawInput(0)->getImpl()->rawPtr(),
- // mOp.mInputs[1]->getImpl()->rawPtr(),
- // mOp.mInputs[2]->getImpl()->rawPtr(),
- // mOp.getOutput(0)->getImpl()->rawPtr());
- // }
- // else
- kernelFunc(
- dynamic_cast<const FC_Op&>(mOp).getStaticAttributes(),
+ kernelFunc(dynamic_cast<const FC_Op&>(mOp).getStaticAttributes(),
std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims()[0],
std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->sizeM1(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->getImpl()->rawPtr(),
+ input0.getImpl()->rawPtr(), input1.getImpl()->rawPtr(), input2.getImpl()->rawPtr(),
std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->getImpl()->rawPtr());
}
diff --git a/unit_tests/recipies/Test_ExplicitConvert.cpp b/unit_tests/recipies/Test_ExplicitConvert.cpp
index 80548aac027c9c719049240b18afefd4ca2eb678..7c5a970ef5a96ec327cd504bc4f51c38de0499a9 100644
--- a/unit_tests/recipies/Test_ExplicitConvert.cpp
+++ b/unit_tests/recipies/Test_ExplicitConvert.cpp
@@ -42,5 +42,5 @@ TEST_CASE("[ExplicitConvert] conv") {
explicitConvert(g1);
g1->save("ExplicitConvert_after");
- REQUIRE(g1->getNodes().size() == 5);
+ REQUIRE(g1->getNodes().size() == 13);
}
diff --git a/unit_tests/scheduler/Test_Convert.cpp b/unit_tests/scheduler/Test_Convert.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..df3db1c2d6467dc328878929e172c480a109ac60
--- /dev/null
+++ b/unit_tests/scheduler/Test_Convert.cpp
@@ -0,0 +1,239 @@
+/********************************************************************************
+ * 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 <memory>
+#include <string>
+
+#include "aidge/data/Tensor.hpp"
+#include "aidge/utils/TensorUtils.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/graph/GraphView.hpp"
+#include "aidge/graph/OpArgs.hpp"
+#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/recipies/Recipies.hpp"
+
+#include "aidge/backend/cpu.hpp"
+
+using namespace Aidge;
+
+TEST_CASE("[cpu/convert] Convert(forward)") {
+ std::shared_ptr<Tensor> inputTensor =
+ std::make_shared<Tensor>(Array4D<int, 2, 1, 5, 5>{{{{{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}}}}});
+
+ std::shared_ptr<Tensor> weight1 = std::make_shared<Tensor>(
+ Array4D<int, 3, 1, 3, 3>{{{{{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}}}}});
+
+ std::shared_ptr<Tensor> bias1 = std::make_shared<Tensor>(Array1D<int, 3>{{1, 2, 3}});
+
+ SECTION("Test implicit") {
+ std::shared_ptr<GraphView> g =
+ Sequential({
+ Conv(1, 3, {3, 3}, "conv1"),
+ Conv(3, 4, {1, 1}, "conv2"),
+ Conv(4, 3, {1, 1}, "conv3"),
+ FC(27, 5, false, "fc")});
+
+ g->getNode("conv1")->getOperator()->setInput(0, inputTensor);
+ g->getNode("conv1")->getOperator()->setInput(1, weight1);
+ g->getNode("conv1")->getOperator()->setInput(2, bias1);
+
+ std::shared_ptr<Tensor> weight2 =
+ std::make_shared<Tensor>(Array4D<int, 4, 3, 1, 1>{{{{{1}}, {{2}}, {{3}}},
+ {{{4}}, {{5}}, {{6}}},
+ {{{7}}, {{8}}, {{9}}},
+ {{{10}}, {{11}}, {{12}}}}});
+ std::shared_ptr<Tensor> bias2 = std::make_shared<Tensor>(Array1D<int, 4>{{1, 2, 3, 4}});
+ g->getNode("conv2")->getOperator()->setInput(1, weight2);
+ g->getNode("conv2")->getOperator()->setInput(2, bias2);
+ // *(g->getNode("conv2")->getOperator()->input(1, weight2);
+
+ std::shared_ptr<Tensor> weight3 = std::make_shared<Tensor>(
+ Array4D<int, 3, 4, 1, 1>{{{{{1}}, {{2}}, {{3}}, {{4}}},
+ {{{5}}, {{6}}, {{7}}, {{8}}},
+ {{{9}}, {{10}}, {{11}}, {{12}}}}});
+ std::shared_ptr<Tensor> bias3 = std::make_shared<Tensor>(Array1D<int, 3>{{1, 2, 3}});
+ g->getNode("conv3")->getOperator()->setInput(1, weight3);
+ g->getNode("conv3")->getOperator()->setInput(2, bias3);
+
+ std::shared_ptr<Tensor> weightfc = std::make_shared<Tensor>(
+ Array2D<int, 5, 27>{{{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+ 15, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12},
+ {13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+ 12, 13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8, 9},
+ {10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8,
+ 9, 10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5, 6},
+ {7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5,
+ 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2, 3},
+ {4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2,
+ 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}}});
+ std::shared_ptr<Tensor> biasfc = std::make_shared<Tensor>(Array1D<int, 5>{{1, 2, 3, 4, 5}});
+ g->getNode("fc")->getOperator()->setInput(1, weightfc);
+ g->getNode("fc")->getOperator()->setInput(2, biasfc);
+
+ // input->addChild(g);
+ g->setDataType(Aidge::DataType::Int32);
+ g->getNode("conv1")->getOperator()->setDataType(DataType::Float32);
+ g->getNode("conv3")->getOperator()->setDataType(DataType::Float64);
+
+ g->setBackend("cpu");
+ g->forwardDims();
+ SequentialScheduler scheduler(g);
+ REQUIRE_NOTHROW(scheduler.forward());
+ scheduler.saveSchedulingDiagram("schedulingSequential");
+
+ std::shared_ptr<Tensor> expectedOutput1 = std::make_shared<Tensor>(Array4D<int, 2, 3, 3, 3>{
+ {{{{367, 412, 457}, {592, 637, 682}, {817, 862, 907}},
+ {{854, 980, 1106}, {1484, 1610, 1736}, {2114, 2240, 2366}},
+ {{1341, 1548, 1755}, {2376, 2583, 2790}, {3411, 3618, 3825}}},
+ {{{1492, 1537, 1582}, {1717, 1762, 1807}, {1942, 1987, 2032}},
+ {{4004, 4130, 4256}, {4634, 4760, 4886}, {5264, 5390, 5516}},
+ {{6516, 6723, 6930}, {7551, 7758, 7965}, {8586, 8793, 9000}}}}});
+
+ std::shared_ptr<Tensor> expectedOutput2 = std::make_shared<Tensor>(Array4D<int, 2, 4, 3, 3>{
+ {{{{6099, 7017, 7935}, {10689, 11607, 12525}, {15279, 16197, 17115}},
+ {{13786, 15838, 17890}, {24046, 26098, 28150}, {34306, 36358, 38410}},
+ {{21473, 24659, 27845}, {37403, 40589, 43775}, {53333, 56519, 59705}},
+ {{29160, 33480, 37800}, {50760, 55080, 59400}, {72360, 76680, 81000}}},
+ {{{29049, 29967, 30885}, {33639, 34557, 35475}, {38229, 39147, 40065}},
+ {{65086, 67138, 69190}, {75346, 77398, 79450}, {85606, 87658, 89710}},
+ {{101123, 104309, 107495}, {117053, 120239, 123425}, {132983, 136169, 139355}},
+ {{137160, 141480, 145800}, {158760, 163080, 167400}, {180360, 184680, 189000}}}}});
+
+ std::shared_ptr<Tensor> expectedOutput3 = std::make_shared<Tensor>(Array4D<int, 2, 3, 3, 3>{
+ {{{{214731, 246591, 278451}, {374031, 405891, 437751}, {533331, 565191, 597051}},
+ {{496804, 570568, 644332}, {865624, 939388, 1013152}, {1234444, 1308208, 1381972}},
+ {{778877, 894545, 1010213}, {1357217, 1472885, 1588553}, {1935557, 2051225, 2166893}}},
+ {{{1011231, 1043091, 1074951}, {1170531, 1202391, 1234251}, {1329831, 1361691, 1393551}},
+ {{2340904, 2414668, 2488432}, {2709724, 2783488, 2857252}, {3078544, 3152308, 3226072}},
+ {{3670577, 3786245, 3901913}, {4248917, 4364585, 4480253}, {4827257, 4942925, 5058593}}}}});
+
+ Tensor expectedOutput4 = Array2D<int, 2, 5>{
+ {{205050376, 198925904, 181355097, 196978090, 238868348},
+ {598467376, 561797804, 560823897, 593043790, 698672948}}};
+ std::shared_ptr<Tensor> other1 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv1")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<float, int>(*other1, *expectedOutput1, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other2 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv2")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<int>(*other2, *expectedOutput2, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other3 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv3")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<double, int>(*other3, *expectedOutput3, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other4 = std::static_pointer_cast<OperatorTensor>(g->getNode("fc")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<int>(*other4, expectedOutput4, 0.0, 1.0e-12));
+ }
+
+ SECTION("Test explicit") {
+ std::shared_ptr<GraphView> g =
+ Sequential({
+ Conv(1, 3, {3, 3}, "conv1"),
+ Conv(3, 4, {1, 1}, "conv2"),
+ Conv(4, 3, {1, 1}, "conv3"),
+ FC(27, 5, false, "fc")});
+
+ g->getNode("conv1")->getOperator()->setInput(0, inputTensor);
+ g->getNode("conv1")->getOperator()->setInput(1, weight1);
+ g->getNode("conv1")->getOperator()->setInput(2, bias1);
+
+ std::shared_ptr<Tensor> weight2 =
+ std::make_shared<Tensor>(Array4D<int, 4, 3, 1, 1>{{{{{1}}, {{2}}, {{3}}},
+ {{{4}}, {{5}}, {{6}}},
+ {{{7}}, {{8}}, {{9}}},
+ {{{10}}, {{11}}, {{12}}}}});
+ std::shared_ptr<Tensor> bias2 = std::make_shared<Tensor>(Array1D<int, 4>{{1, 2, 3, 4}});
+ g->getNode("conv2")->getOperator()->setInput(1, weight2);
+ g->getNode("conv2")->getOperator()->setInput(2, bias2);
+ // *(g->getNode("conv2")->getOperator()->input(1, weight2);
+
+ std::shared_ptr<Tensor> weight3 = std::make_shared<Tensor>(
+ Array4D<int, 3, 4, 1, 1>{{{{{1}}, {{2}}, {{3}}, {{4}}},
+ {{{5}}, {{6}}, {{7}}, {{8}}},
+ {{{9}}, {{10}}, {{11}}, {{12}}}}});
+ std::shared_ptr<Tensor> bias3 = std::make_shared<Tensor>(Array1D<int, 3>{{1, 2, 3}});
+ g->getNode("conv3")->getOperator()->setInput(1, weight3);
+ g->getNode("conv3")->getOperator()->setInput(2, bias3);
+
+ std::shared_ptr<Tensor> weightfc = std::make_shared<Tensor>(
+ Array2D<int, 5, 27>{{{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+ 15, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12},
+ {13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+ 12, 13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8, 9},
+ {10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5, 6, 7, 8,
+ 9, 10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5, 6},
+ {7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2, 3, 4, 5,
+ 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2, 3},
+ {4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1, 2,
+ 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}}});
+ std::shared_ptr<Tensor> biasfc = std::make_shared<Tensor>(Array1D<int, 5>{{1, 2, 3, 4, 5}});
+ g->getNode("fc")->getOperator()->setInput(1, weightfc);
+ g->getNode("fc")->getOperator()->setInput(2, biasfc);
+
+ // input->addChild(g);
+ g->setDataType(Aidge::DataType::Int32);
+ g->getNode("conv1")->getOperator()->setDataType(DataType::Float32);
+ g->getNode("conv3")->getOperator()->setDataType(DataType::Float64);
+
+ explicitConvert(g);
+ g->setBackend("cpu");
+ g->forwardDims();
+
+ SequentialScheduler scheduler(g);
+ REQUIRE_NOTHROW(scheduler.forward());
+ scheduler.saveSchedulingDiagram("schedulingSequential");
+
+ std::shared_ptr<Tensor> expectedOutput1 = std::make_shared<Tensor>(Array4D<int, 2, 3, 3, 3>{
+ {{{{367, 412, 457}, {592, 637, 682}, {817, 862, 907}},
+ {{854, 980, 1106}, {1484, 1610, 1736}, {2114, 2240, 2366}},
+ {{1341, 1548, 1755}, {2376, 2583, 2790}, {3411, 3618, 3825}}},
+ {{{1492, 1537, 1582}, {1717, 1762, 1807}, {1942, 1987, 2032}},
+ {{4004, 4130, 4256}, {4634, 4760, 4886}, {5264, 5390, 5516}},
+ {{6516, 6723, 6930}, {7551, 7758, 7965}, {8586, 8793, 9000}}}}});
+
+ std::shared_ptr<Tensor> expectedOutput2 = std::make_shared<Tensor>(Array4D<int, 2, 4, 3, 3>{
+ {{{{6099, 7017, 7935}, {10689, 11607, 12525}, {15279, 16197, 17115}},
+ {{13786, 15838, 17890}, {24046, 26098, 28150}, {34306, 36358, 38410}},
+ {{21473, 24659, 27845}, {37403, 40589, 43775}, {53333, 56519, 59705}},
+ {{29160, 33480, 37800}, {50760, 55080, 59400}, {72360, 76680, 81000}}},
+ {{{29049, 29967, 30885}, {33639, 34557, 35475}, {38229, 39147, 40065}},
+ {{65086, 67138, 69190}, {75346, 77398, 79450}, {85606, 87658, 89710}},
+ {{101123, 104309, 107495}, {117053, 120239, 123425}, {132983, 136169, 139355}},
+ {{137160, 141480, 145800}, {158760, 163080, 167400}, {180360, 184680, 189000}}}}});
+
+ std::shared_ptr<Tensor> expectedOutput3 = std::make_shared<Tensor>(Array4D<int, 2, 3, 3, 3>{
+ {{{{214731, 246591, 278451}, {374031, 405891, 437751}, {533331, 565191, 597051}},
+ {{496804, 570568, 644332}, {865624, 939388, 1013152}, {1234444, 1308208, 1381972}},
+ {{778877, 894545, 1010213}, {1357217, 1472885, 1588553}, {1935557, 2051225, 2166893}}},
+ {{{1011231, 1043091, 1074951}, {1170531, 1202391, 1234251}, {1329831, 1361691, 1393551}},
+ {{2340904, 2414668, 2488432}, {2709724, 2783488, 2857252}, {3078544, 3152308, 3226072}},
+ {{3670577, 3786245, 3901913}, {4248917, 4364585, 4480253}, {4827257, 4942925, 5058593}}}}});
+
+ Tensor expectedOutput4 = Array2D<int, 2, 5>{
+ {{205050376, 198925904, 181355097, 196978090, 238868348},
+ {598467376, 561797804, 560823897, 593043790, 698672948}}};
+ std::shared_ptr<Tensor> other1 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv1")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<float, int>(*other1, *expectedOutput1, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other2 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv2")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<int>(*other2, *expectedOutput2, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other3 = std::static_pointer_cast<OperatorTensor>(g->getNode("conv3")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<double, int>(*other3, *expectedOutput3, 0.0, 1.0e-12));
+ std::shared_ptr<Tensor> other4 = std::static_pointer_cast<OperatorTensor>(g->getNode("fc")->getOperator())->getOutput(0);
+ REQUIRE(approxEq<int>(*other4, expectedOutput4, 0.0, 1.0e-12));
+ }
+}