From e342d7879a8df53b18477119814cf01060bd1d39 Mon Sep 17 00:00:00 2001
From: NAUD Maxence <maxence.naud@cea.fr>
Date: Wed, 26 Mar 2025 17:10:26 +0100
Subject: [PATCH] fix: error related to 'fma' call with different input types.
Also simplify GlobalAvgPooling logic
---
.../cpu/operator/GlobalAveragePoolingImpl.hpp | 3 +-
.../GlobalAveragePoolingImpl_kernels.hpp | 95 ++++++++-----------
.../backend/cpu/operator/ReduceMeanImpl.hpp | 1 -
.../cpu/operator/ReduceMeanImpl_kernels.hpp | 20 ++--
src/operator/GlobalAveragePoolingImpl.cpp | 6 +-
5 files changed, 52 insertions(+), 73 deletions(-)
diff --git a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp
index 4e04b1a5..a71174c0 100644
--- a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp
+++ b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp
@@ -18,12 +18,11 @@
#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
#include "aidge/operator/GlobalAveragePooling.hpp"
#include "aidge/utils/Registrar.hpp"
-#include "aidge/utils/Types.h"
namespace Aidge {
// Operator implementation entry point for the backend
using GlobalAveragePoolingImpl_cpu = OperatorImpl_cpu<GlobalAveragePooling_Op,
- void(const std::vector<DimSize_t> &, const void *, void *)>;
+ void(const std::shared_ptr<Tensor>&, void *)>;
// Implementation entry point registration to Operator
REGISTRAR(GlobalAveragePooling_Op, "cpu", Aidge::GlobalAveragePoolingImpl_cpu::create);
diff --git a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
index 7a47ccf3..40dd3a69 100644
--- a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
@@ -12,92 +12,71 @@
#ifndef AIDGE_CPU_OPERATOR_GLOBALAVERAGEPOOLINGIMPL_KERNELS_H_
#define AIDGE_CPU_OPERATOR_GLOBALAVERAGEPOOLINGIMPL_KERNELS_H_
-#include <cstddef>
-#include <functional> // std::multiplies
-#include <numeric> // std::accumulate
+#include <cstddef> // std::size_t
#include <vector>
#include "aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp"
-#include "aidge/data/Data.hpp"
-#include "aidge/utils/ErrorHandling.hpp"
+#include "aidge/data/Tensor.hpp"
#include "aidge/utils/Registrar.hpp"
#include "aidge/utils/Types.h"
-
namespace Aidge {
template <typename T>
-typename std::enable_if<std::is_floating_point<T>::value, T>::type
-stableMean(const T* vec, size_t size) {
- T mean = 0;
- for (size_t i = 0; i < size; ++i) {
- mean = std::fma<T>(vec[i] - mean, 1.0f / (i + 1), mean);
- }
- return mean;
+typename std::enable_if_t<std::is_floating_point<T>::value, T>
+static stableMean(const T* vec, std::size_t size) {
+ T mean{0};
+ for (std::size_t i = 0; i < size; ++i) {
+ mean = std::fma(vec[i] - mean, static_cast<T>(1) / static_cast<T>(i + 1), mean);
+ }
+ return mean;
}
// Specialization for integers: perform the mean computation in float
template <typename T>
-typename std::enable_if<!std::is_floating_point<T>::value, double>::type
-stableMean(const T* vec, size_t size) {
- double mean = 0;
- for (size_t i = 0; i < size; ++i) {
- mean = std::fma<double>(vec[i] - mean, 1.0f / (i + 1), mean);
- }
- return mean;
-}
-
-template <typename T>
-typename std::enable_if<std::is_floating_point<T>::value, T>::type
-castFromFloat(T value) {
- return value;
-}
-
-template <typename T>
-typename std::enable_if<!std::is_floating_point<T>::value, T>::type
-castFromFloat(double value) {
- return static_cast<T>(std::nearbyint(value));
+typename std::enable_if_t<!std::is_floating_point<T>::value, double>
+static stableMean(const T* vec, std::size_t size) {
+ double mean{0};
+ for (std::size_t i = 0; i < size; ++i) {
+ mean = std::fma<double>(static_cast<double>(vec[i]) - mean, 1.0 / static_cast<double>(i + 1), mean);
+ }
+ return mean;
}
-template <class I, class O>
-void GlobalAveragePoolingImpl_cpu_forward_kernel(
- const std::vector<DimSize_t> &dims, const void *input_, void *output_) {
- // error checking
- AIDGE_ASSERT(dims.size() >= 3,"GlobalAveragePool needs at least a 3 dimensions "
- "input, number of input dim : {}",
- dims.size());
+template <DataType DT_I, DataType DT_O = DT_I>
+void GlobalAveragePoolingImpl_cpu_forward_kernel(const std::shared_ptr<Tensor>& inputTensor, void *output_) {
- // computation
- const I *input = static_cast<const I *>(input_);
- O *output = static_cast<O *>(output_);
+ // computation
+ using I = cpptype_t<DT_I>;
+ using O = cpptype_t<DT_O>;
+ const I *input = static_cast<const I *>(inputTensor->getImpl()->rawPtr());
+ O *output = static_cast<O *>(output_);
+ const auto& dims = inputTensor->dims();
- DimSize_t nb_elems = std::accumulate(dims.begin(), dims.end(), std::size_t(1),
- std::multiplies<std::size_t>());
+ const DimSize_t strides_channels = inputTensor->strides()[1];
- const DimSize_t in_batch_nb_elems{nb_elems / dims[0]};
- const DimSize_t in_channel_nb_elems{in_batch_nb_elems / dims[1]};
- const DimSize_t out_batch_nb_elems{dims[1]};
- // parse channel by channel and fill each output with the average of the
- // values in the channel
- for (DimSize_t batch = 0; batch < dims[0]; ++batch) {
- for (DimSize_t channel = 0; channel < dims[1]; ++channel) {
- const I *filter_start = std::next(
- input, (batch * in_batch_nb_elems) + (channel * in_channel_nb_elems));
- output[batch * out_batch_nb_elems + channel] = castFromFloat<O>(stableMean<I>(filter_start, in_channel_nb_elems));
+ // parse channel by channel and fill each output with the average of the
+ // values in the channel
+ std::size_t input_idx = 0;
+ std::size_t output_idx = 0;
+ for (DimSize_t batch = 0; batch < dims[0]; ++batch) {
+ for (DimSize_t channel = 0; channel < dims[1]; ++channel) {
+ output[output_idx++] = static_cast<O>(stableMean<I>(input + input_idx, strides_channels));
+ input_idx += strides_channels;
+ }
}
- }
}
// Kernels registration to implementation entry point
REGISTRAR(GlobalAveragePoolingImpl_cpu,
{DataType::Float32},
- {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<float, float>, nullptr});
+ {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<DataType::Float32>, nullptr});
REGISTRAR(GlobalAveragePoolingImpl_cpu,
{DataType::Float64},
- {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<double, double>, nullptr});
+ {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<DataType::Float64>, nullptr});
REGISTRAR(GlobalAveragePoolingImpl_cpu,
{DataType::Int32},
- {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<int32_t, int32_t>, nullptr});
+ {ProdConso::defaultModel, Aidge::GlobalAveragePoolingImpl_cpu_forward_kernel<DataType::Int32>, nullptr});
} // namespace Aidge
#endif /* AIDGE_CPU_OPERATOR_GLOBALAVERAGEPOOLINGIMPL_KERNELS_H_ */
diff --git a/include/aidge/backend/cpu/operator/ReduceMeanImpl.hpp b/include/aidge/backend/cpu/operator/ReduceMeanImpl.hpp
index 1c50805d..d6c60c35 100644
--- a/include/aidge/backend/cpu/operator/ReduceMeanImpl.hpp
+++ b/include/aidge/backend/cpu/operator/ReduceMeanImpl.hpp
@@ -12,7 +12,6 @@
#ifndef AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_H_
#define AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_H_
-#include <array>
#include <memory>
#include <tuple>
#include <vector>
diff --git a/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp b/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
index a1562322..73aa283d 100644
--- a/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
@@ -25,39 +25,39 @@
#include "aidge/utils/Registrar.hpp"
namespace Aidge {
-
+
template <typename T>
-using Acc_T = typename std::conditional<std::is_floating_point<T>::value, T, double>::type;
+using Acc_T = typename std::conditional_t<std::is_floating_point<T>::value, T, double>;
template <typename T>
typename std::enable_if<std::is_floating_point<T>::value, T>::type
-stableMean(const T* vec, size_t len, size_t stride) {
+stableMean(const T* vec, std::size_t len, std::size_t stride) {
T mean = 0;
- for (size_t i = 0; i < len; ++i) {
- mean = std::fma<T>(vec[i * stride] - mean, 1.0f / (i + 1), mean);
+ for (std::size_t i = 0; i < len; ++i) {
+ mean = std::fma(vec[i * stride] - mean, static_cast<T>(1) / static_cast<T>(i + 1), mean);
}
return mean;
}
// Specialization for integers: perform the mean computation in float
template <typename T>
-typename std::enable_if<!std::is_floating_point<T>::value, double>::type
-stableMean(const T* vec, size_t len, size_t stride) {
+typename std::enable_if_t<!std::is_floating_point<T>::value, double>
+stableMean(const T* vec, std::size_t len, std::size_t stride) {
double mean = 0;
for (size_t i = 0; i < len; ++i) {
- mean = std::fma<double>(vec[i * stride] - mean, 1.0f / (i + 1), mean);
+ mean = std::fma<double>(static_cast<double>(vec[i * stride]) - mean, 1.0 / static_cast<double>(i + 1), mean);
}
return mean;
}
template <typename T>
-typename std::enable_if<std::is_floating_point<T>::value, T>::type
+typename std::enable_if_t<std::is_floating_point<T>::value, T>
castFromFloat(T value) {
return value;
}
template <typename T>
-typename std::enable_if<!std::is_floating_point<T>::value, T>::type
+typename std::enable_if_t<!std::is_floating_point<T>::value, T>
castFromFloat(double value) {
return static_cast<T>(std::nearbyint(value));
}
diff --git a/src/operator/GlobalAveragePoolingImpl.cpp b/src/operator/GlobalAveragePoolingImpl.cpp
index c53f92e1..1b6d9a06 100644
--- a/src/operator/GlobalAveragePoolingImpl.cpp
+++ b/src/operator/GlobalAveragePoolingImpl.cpp
@@ -30,13 +30,15 @@ void Aidge::GlobalAveragePoolingImpl_cpu::forward()
const GlobalAveragePooling_Op& op_ = static_cast<const GlobalAveragePooling_Op&>(mOp);
// Check if input is provided
AIDGE_ASSERT(op_.getInput(0), "missing input 0");
+ // error checking
+ AIDGE_ASSERT(op_.getInput(0)->nbDims() >= 3,"GlobalAveragePool needs at least a 3 dimensions "
+ "input. Got input dims {}", op_.getInput(0)->dims());
// Find the correct kernel type
const auto impl = Registrar<GlobalAveragePoolingImpl_cpu>::create(getBestMatch(getRequiredSpec()));
// Call kernel
- impl.forward(op_.getInput(0)->dims(),
- op_.getInput(0)->getImpl()->rawPtr(),
+ impl.forward(op_.getInput(0),
op_.getOutput(0)->getImpl()->rawPtr());
}
--
GitLab