diff --git a/include/aidge/backend/cpu/operator/TanhImpl.hpp b/include/aidge/backend/cpu/operator/TanhImpl.hpp
index 3e88a3d00b5829fc24d8dc77ce53cb358551c7e4..4169b1a533a8b2382644246ea295a683e6f83f1d 100644
--- a/include/aidge/backend/cpu/operator/TanhImpl.hpp
+++ b/include/aidge/backend/cpu/operator/TanhImpl.hpp
@@ -33,7 +33,7 @@ class TanhImplBackward_cpu
class TanhImpl_cpu : public OperatorImpl {
public:
- TanhImpl_cpu(const Tanh_Op& op) : OperatorImpl(op) {}
+ TanhImpl_cpu(const Tanh_Op& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TanhImpl_cpu> create(const Tanh_Op& op) {
return std::make_unique<TanhImpl_cpu>(op);
diff --git a/include/aidge/backend/cpu/operator/TransposeImpl.hpp b/include/aidge/backend/cpu/operator/TransposeImpl.hpp
index 712e672752648f5ff8a3c073f6c81bbe7cc85d9d..3c6913dd71d6642d8b76198a272d64bfaba833e8 100644
--- a/include/aidge/backend/cpu/operator/TransposeImpl.hpp
+++ b/include/aidge/backend/cpu/operator/TransposeImpl.hpp
@@ -57,7 +57,7 @@ class TransposeImpl6DBackward_cpu
class TransposeImpl2D_cpu : public OperatorImpl {
public:
- TransposeImpl2D_cpu(const Transpose_Op<2>& op) : OperatorImpl(op) {}
+ TransposeImpl2D_cpu(const Transpose_Op<2>& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TransposeImpl2D_cpu> create(const Transpose_Op<2>& op) {
return std::make_unique<TransposeImpl2D_cpu>(op);
@@ -68,7 +68,7 @@ public:
};
class TransposeImpl3D_cpu : public OperatorImpl {
public:
- TransposeImpl3D_cpu(const Transpose_Op<3>& op) : OperatorImpl(op) {}
+ TransposeImpl3D_cpu(const Transpose_Op<3>& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TransposeImpl3D_cpu> create(const Transpose_Op<3>& op) {
return std::make_unique<TransposeImpl3D_cpu>(op);
@@ -79,7 +79,7 @@ public:
};
class TransposeImpl4D_cpu : public OperatorImpl {
public:
- TransposeImpl4D_cpu(const Transpose_Op<4>& op) : OperatorImpl(op) {}
+ TransposeImpl4D_cpu(const Transpose_Op<4>& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TransposeImpl4D_cpu> create(const Transpose_Op<4>& op) {
return std::make_unique<TransposeImpl4D_cpu>(op);
@@ -90,7 +90,7 @@ public:
};
class TransposeImpl5D_cpu : public OperatorImpl {
public:
- TransposeImpl5D_cpu(const Transpose_Op<5>& op) : OperatorImpl(op) {}
+ TransposeImpl5D_cpu(const Transpose_Op<5>& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TransposeImpl5D_cpu> create(const Transpose_Op<5>& op) {
return std::make_unique<TransposeImpl5D_cpu>(op);
@@ -101,7 +101,7 @@ public:
};
class TransposeImpl6D_cpu : public OperatorImpl {
public:
- TransposeImpl6D_cpu(const Transpose_Op<6>& op) : OperatorImpl(op) {}
+ TransposeImpl6D_cpu(const Transpose_Op<6>& op) : OperatorImpl(op, "cpu") {}
static std::unique_ptr<TransposeImpl6D_cpu> create(const Transpose_Op<6>& op) {
return std::make_unique<TransposeImpl6D_cpu>(op);
diff --git a/src/operator/AddImpl.cpp b/src/operator/AddImpl.cpp
index 7355ebcb3e8fb68bf74dbd1ce831bf471d285cb7..abd40bd6af06c52945815fd6245e661710fa1127 100644
--- a/src/operator/AddImpl.cpp
+++ b/src/operator/AddImpl.cpp
@@ -9,17 +9,18 @@
*
********************************************************************************/
+#include "aidge/backend/cpu/operator/AddImpl.hpp"
+
#include <cassert>
#include <numeric> // std::accumulate
#include <vector>
-#include "aidge/utils/Types.h"
#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include "aidge/backend/cpu/operator/AddImpl_forward_kernels.hpp"
#include "aidge/data/Data.hpp"
#include "aidge/data/Tensor.hpp"
-
-#include "aidge/backend/cpu/operator/AddImpl.hpp"
-#include "aidge/backend/cpu/operator/AddImpl_forward_kernels.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/utils/ErrorHandling.hpp"
Aidge::NbElts_t Aidge::AddImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex_t /*inputIdx*/) const {
// this implementation can be in-place
@@ -27,15 +28,18 @@ Aidge::NbElts_t Aidge::AddImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex
}
void Aidge::AddImpl_cpu::forward() {
- assert(mOp.getRawInput(0) && "missing input in Add operator");
- DataType datatypeFirstInput = std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType();
- for (IOIndex_t i = 1; i < mOp.nbInputs(); ++i) {
- assert(mOp.getRawInput(i) && "missing input in Add operator");
- assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->dataType() == datatypeFirstInput);
+ const auto& opTensor = static_cast<const OperatorTensor&>(mOp);
+ AIDGE_ASSERT(opTensor.getInput(0)->hasImpl(), "cannot run Add forward because the 0-th input has no implementation.");
+ assert(opTensor.getInput(0) && "missing input in Add operator");
+ DataType datatypeFirstInput = opTensor.getInput(0)->dataType();
+ for (IOIndex_t i = 1; i < opTensor.nbInputs(); ++i) {
+ AIDGE_ASSERT(opTensor.getInput(i)->hasImpl(), "cannot run Add forward because the {}-th input has no implementation.", i);
+ assert(opTensor.getInput(i) && "missing input in Add operator");
+ assert(opTensor.getInput(i)->dataType() == datatypeFirstInput);
}
// Find the correct kernel type
- const auto outputDataType = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType();
+ const auto outputDataType = opTensor.getOutput(0)->dataType();
const Registrar<AddImplForward_cpu>::registrar_key registrarKey = {
datatypeFirstInput,
outputDataType};
@@ -55,26 +59,26 @@ void Aidge::AddImpl_cpu::forward() {
// 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::size_t nbDims = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->nbDims();
+ const std::size_t nbDims = opTensor.getOutput(0)->nbDims();
std::vector<std::vector<std::size_t>> inputsDims;
std::vector<const void*> opInputs;
- std::vector<std::shared_ptr<Tensor>> inputsFallback(mOp.nbInputs());
- for (IOIndex_t i = 0; i < mOp.nbInputs(); ++i) {
+ std::vector<std::shared_ptr<Tensor>> inputsFallback(opTensor.nbInputs());
+ for (IOIndex_t i = 0; i < opTensor.nbInputs(); ++i) {
std::vector<std::size_t> inputDims(nbDims, 1);
- auto dims = std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->dims();
+ auto dims = opTensor.getInput(i)->dims();
for(std::size_t j=dims.size()-1; j+1>0; --j)
{
std::size_t idx = nbDims - (dims.size()-j);
inputDims[idx] = dims[j];
}
inputsDims.push_back(inputDims);
- const auto& input = std::static_pointer_cast<Tensor>(mOp.getRawInput(i))->refCastFrom(inputsFallback[i], *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
+ const auto& input = opTensor.getInput(i)->refCastFrom(inputsFallback[i], *opTensor.getOutput(0));
opInputs.push_back(input.getImpl()->rawPtr());
}
kernelFunc(opInputs,
inputsDims,
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->size(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims(),
- getCPUPtr(mOp.getRawOutput(0)));
+ opTensor.getOutput(0)->size(),
+ opTensor.getOutput(0)->dims(),
+ getCPUPtr(opTensor.getRawOutput(0)));
}
diff --git a/src/operator/ConvImpl.cpp b/src/operator/ConvImpl.cpp
index b849142dd3abe0131fb0c6c448530a7669ce27dc..34ea7b37ec9929908192bde6f31d84ae581640a2 100644
--- a/src/operator/ConvImpl.cpp
+++ b/src/operator/ConvImpl.cpp
@@ -28,17 +28,19 @@ Aidge::NbElts_t Aidge::ConvImpl2D_cpu::getNbRequiredProtected(IOIndex_t /*inputI
}
void Aidge::ConvImpl2D_cpu::forward() {
+ const auto& opTensor = static_cast<const OperatorTensor&>(mOp);
+
// FIXME: uncomment the following code once memory handling will work
assert(mOp.getRawInput(0) && "missing input #0");
assert(mOp.getRawInput(1) && "missing input #1");
assert(mOp.getRawInput(2) && "missing input #2");
// Find the correct kernel type
- const auto outputDataType = std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType();
+ const auto outputDataType = opTensor.getOutput(0)->dataType();
const Registrar<ConvImpl2DForward_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(),
+ opTensor.getInput(0)->dataType(),
+ opTensor.getInput(1)->dataType(),
+ opTensor.getInput(2)->dataType(),
outputDataType};
Registrar<ConvImpl2DForward_cpu>::registrar_type kernelFunc;
@@ -57,12 +59,12 @@ void Aidge::ConvImpl2D_cpu::forward() {
// 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))->refCastFrom(input0Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
- const auto& input1 = std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->refCastFrom(input1Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
- const auto& input2 = std::static_pointer_cast<Tensor>(mOp.getRawInput(2))->refCastFrom(input2Fallback, *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0)));
+ const auto& input0 = opTensor.getInput(0)->refCastFrom(input0Fallback, *opTensor.getOutput(0));
+ const auto& input1 = opTensor.getInput(1)->refCastFrom(input1Fallback, *opTensor.getOutput(0));
+ const auto& input2 = opTensor.getInput(2)->refCastFrom(input2Fallback, *opTensor.getOutput(0));
// Call kernel
- kernelFunc(dynamic_cast<const Conv_Op<2>&>(mOp).getStaticAttributes(), std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->template dims<4>(),
+ kernelFunc(dynamic_cast<const Conv_Op<2>&>(mOp).getStaticAttributes(), opTensor.getInput(0)->template dims<4>(),
input0.getImpl()->rawPtr(), input1.getImpl()->rawPtr(), input2.getImpl()->rawPtr(),
getCPUPtr(mOp.getRawOutput(0)));
}
diff --git a/src/operator/DivImpl.cpp b/src/operator/DivImpl.cpp
index 729aff2452b46f00eb6d3e0b558c0b3d58ea2f0e..8e2118e9e78fd364189769ead2eb01f1c55b3c58 100644
--- a/src/operator/DivImpl.cpp
+++ b/src/operator/DivImpl.cpp
@@ -57,17 +57,18 @@ void Aidge::DivImpl_cpu::forward() {
// 3. Compute the highest number of contiguous data -> 7
// 4. Compute stride and offset step for the broadcast mechnism
// 5. Call a simple kernel
+ const auto& opTensor = static_cast<const Div_Op&>(mOp);
// Find the correct kernel type
auto kernelFunc = Registrar<DivImplForward_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.getRawOutput(0))->dataType()});
+ opTensor.getInput(0)->dataType(),
+ opTensor.getInput(1)->dataType(),
+ opTensor.getOutput(0)->dataType()});
// Compute compatible input dimensions
- std::vector<std::size_t> dims0 = static_cast<const Div_Op&>(mOp).getInput(0)->dims();
- std::vector<std::size_t> dims1 = static_cast<const Div_Op&>(mOp).getInput(1)->dims();
- const std::vector<std::size_t>& outDims = static_cast<const Div_Op&>(mOp).getOutput(0)->dims();
+ std::vector<std::size_t> dims0 = opTensor.getInput(0)->dims();
+ std::vector<std::size_t> dims1 = opTensor.getInput(1)->dims();
+ const std::vector<std::size_t>& outDims = opTensor.getOutput(0)->dims();
// if (dims0 == dims1) {
// const std::size_t input0_contiguous_size = std::accumulate(dims0.cbegin(), dims0.cend(), std::size_t(1), std::multiplies<std::size_t>());
@@ -108,24 +109,24 @@ void Aidge::DivImpl_cpu::forward() {
const std::size_t output_contiguous_size = std::accumulate(outDims.cbegin()+contiguousIdx, outDims.cend(), std::size_t(1), std::multiplies<std::size_t>());
// initialize strides to iterate through data because of broadcasting
- std::size_t *stride_post0;
- std::size_t *stride_post1;
+ std::int32_t *stride_post0;
+ std::int32_t *stride_post1;
std::int32_t *stride_step0;
std::int32_t *stride_step1;
if (contiguousIdx > 0) {
- stride_post0 = new std::size_t[contiguousIdx];
+ stride_post0 = new std::int32_t[contiguousIdx];
stride_post0[contiguousIdx - 1] = 1;
- stride_post1 = new std::size_t[contiguousIdx];
+ stride_post1 = new std::int32_t[contiguousIdx];
stride_post1[contiguousIdx - 1] = 1;
for (std::size_t i = contiguousIdx - 2; i != static_cast<std::size_t>(-1); --i) {
- stride_post0[i] = stride_post0[i+1]*dims0[i+1];
- stride_post1[i] = stride_post1[i+1]*dims1[i+1];
+ stride_post0[i] = stride_post0[i+1]*static_cast<std::int32_t>(dims0[i+1]);
+ stride_post1[i] = stride_post1[i+1]*static_cast<std::int32_t>(dims1[i+1]);
}
stride_step0 = new std::int32_t[contiguousIdx];
stride_step1 = new std::int32_t[contiguousIdx];
for (std::size_t i = 0; i != contiguousIdx; ++i) {
- stride_step0[i] = (dims0[i] == 1) ? 1 - static_cast<std::int32_t>(stride_post0[i]) : 1;
- stride_step1[i] = (dims1[i] == 1) ? 1 - static_cast<std::int32_t>(stride_post1[i]) : 1;
+ stride_step0[i] = (dims0[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (dims1[i] == 1) ? 1 - stride_post1[i] : 1;
}
}
diff --git a/src/operator/ErfImpl.cpp b/src/operator/ErfImpl.cpp
index 06ec65008aee41215192cd05e126ac4f82388c1b..55752e4f5b9f798a6901e108ddcba2f61fdf9774 100644
--- a/src/operator/ErfImpl.cpp
+++ b/src/operator/ErfImpl.cpp
@@ -9,32 +9,34 @@
*
********************************************************************************/
-#include <cassert>
-#include <chrono> // std::chrono::milliseconds
-#include <numeric> // std::accumulate
-#include <thread> // std::this_thread::sleep_for
+#include "aidge/backend/cpu/operator/ErfImpl.hpp"
+
+#include <memory>
#include <vector>
+#include "aidge/backend/cpu/operator/ErfImpl_forward_kernels.hpp"
+#include "aidge/data/Tensor.hpp"
#include "aidge/operator/Erf.hpp"
#include "aidge/utils/Types.h"
-#include "aidge/backend/cpu/operator/ErfImpl.hpp"
-#include "aidge/backend/cpu/operator/ErfImpl_forward_kernels.hpp"
-
Aidge::NbElts_t Aidge::ErfImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex_t /*inputIdx*/) const {
// this implementation can be in-place
return 0;
}
void Aidge::ErfImpl_cpu::forward() {
+ const Erf_Op& op = static_cast<const Erf_Op&>(mOp);
// Find the correct kernel type
auto kernelFunc = Registrar<ErfImplForward_cpu>::create({
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType()});
+ op.getInput(0)->dataType(),
+ op.getOutput(0)->dataType()
+ });
// Call kernel
- kernelFunc(std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->size(),
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->getImpl()->rawPtr(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->getImpl()->rawPtr());
+ kernelFunc(
+ op.getInput(0)->size(),
+ op.getInput(0)->getImpl()->rawPtr(),
+ op.getOutput(0)->getImpl()->rawPtr()
+ );
}
diff --git a/src/operator/GatherImpl.cpp b/src/operator/GatherImpl.cpp
index ce98627d95e0d05541db1ccaf4896abe756431b0..d80b53e7e864faf3fca289f94aba4f511bcba161 100644
--- a/src/operator/GatherImpl.cpp
+++ b/src/operator/GatherImpl.cpp
@@ -9,32 +9,34 @@
*
********************************************************************************/
-#include <cassert>
-#include <chrono> // std::chrono::milliseconds
-#include <numeric> // std::accumulate
-#include <thread> // std::this_thread::sleep_for
+#include "aidge/backend/cpu/operator/GatherImpl.hpp"
+
+#include <memory>
#include <vector>
+#include "aidge/backend/cpu/operator/GatherImpl_forward_kernels.hpp"
+#include "aidge/data/Data.hpp"
+#include "aidge/data/Tensor.hpp"
#include "aidge/operator/Gather.hpp"
#include "aidge/utils/Types.h"
-#include "aidge/backend/cpu/operator/GatherImpl.hpp"
-#include "aidge/backend/cpu/operator/GatherImpl_forward_kernels.hpp"
-
Aidge::NbElts_t Aidge::GatherImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex_t /*inputIdx*/) const {
// this implementation can be in-place
return 0;
}
void Aidge::GatherImpl_cpu::forward() {
+ const Gather_Op& op = static_cast<const Gather_Op&>(mOp);
auto kernelFunc = Registrar<GatherImplForward_cpu>::create({
- std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dataType(),
- std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dataType()});
+ op.getInput(0)->dataType(),
+ op.getOutput(0)->dataType()
+ });
// Call kernel
kernelFunc(dynamic_cast<const Gather_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());
+ op.getInput(0)->dims(),
+ op.getInput(0)->getImpl()->rawPtr(),
+ op.getOutput(0)->getImpl()->rawPtr()
+ );
}