From 2bb9f35b309bc9b910a78fdb81e5342399be4573 Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Thu, 6 Feb 2025 10:04:25 +0100
Subject: [PATCH] Fix bug eclipse/aidge/aidge_backend_cpu#40
---
.../GlobalAveragePoolingImpl_kernels.hpp | 29 +++++++++++----
.../cpu/operator/ReduceMeanImpl_kernels.hpp | 36 ++++++++++++-------
2 files changed, 47 insertions(+), 18 deletions(-)
diff --git a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
index ed838a94..e396020f 100644
--- a/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp
@@ -25,6 +25,28 @@
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;
+}
+
+// Specialization for integers: perform the mean computation in float
+template <typename T>
+typename std::enable_if<!std::is_floating_point<T>::value, T>::type
+stableMean(const T* vec, size_t size) {
+ float mean = 0;
+ for (size_t i = 0; i < size; ++i) {
+ mean = std::fma<float>(vec[i] - mean, 1.0f / (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_) {
@@ -49,12 +71,7 @@ void GlobalAveragePoolingImpl_cpu_forward_kernel(
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));
- I mean = 0;
- for (size_t i = 0; i < in_channel_nb_elems; ++i) {
- // Single pass numerically stable mean, using the fmaf
- mean = fmaf(filter_start[i] - mean, 1.0f/(i+1), mean);
- }
- output[batch * out_batch_nb_elems + channel] = mean;
+ output[batch * out_batch_nb_elems + channel] = stableMean<I>(filter_start, in_channel_nb_elems);
}
}
}
diff --git a/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp b/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
index 5a143164..0ab23655 100644
--- a/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/ReduceMeanImpl_kernels.hpp
@@ -25,6 +25,28 @@
#include "aidge/utils/Registrar.hpp"
namespace Aidge {
+
+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) {
+ T mean = 0;
+ for (size_t i = 0; i < len; ++i) {
+ mean = std::fma<T>(vec[i * stride] - mean, 1.0f / (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, T>::type
+stableMean(const T* vec, size_t len, size_t stride) {
+ float mean = 0;
+ for (size_t i = 0; i < len; ++i) {
+ mean = std::fma<float>(vec[i * stride] - mean, 1.0f / (i + 1), mean);
+ }
+ return mean;
+}
+
template <class I, class O>
void ReduceMeanImpl_cpu_forward_kernel(const std::vector<std::int32_t>& axes,
DimSize_t /*keepDims*/,
@@ -50,12 +72,7 @@ void ReduceMeanImpl_cpu_forward_kernel(const std::vector<std::int32_t>& axes,
for (std::size_t post = 0; post < stride_post; ++post) {
const std::size_t idx_i = pre * dim_i * stride_post + post;
const std::size_t idx_o = pre * stride_post + post;
- O mean = 0;
- for (std::size_t i = 0; i < dim_i; ++i) {
- // Single pass numerically stable mean, using the fmaf
- mean = fmaf(input[idx_i + i*stride_post] - mean, 1.0f/(i+1), mean);
- }
- output[idx_o] = mean;
+ output[idx_o] = stableMean<I>(input + idx_i, dim_i, stride_post);
}
}
} else {
@@ -84,12 +101,7 @@ void ReduceMeanImpl_cpu_forward_kernel(const std::vector<std::int32_t>& axes,
for (std::size_t post = 0; post < stride_post[a]; ++post) {
const std::size_t idx_i = pre * dim_i * stride_post[a] + post;
const std::size_t idx_o = pre * stride_post[a] + post;
- I mean = 0;
- for (std::size_t i = 0; i < dim_i; ++i) {
- // Single pass numerically stable mean, using the fmaf
- mean = fmaf(inputAccumulation[idx_i + i*stride_post[a]] - mean, 1.0f/(i+1), mean);
- }
- outputAccumulation[idx_o] = mean;
+ outputAccumulation[idx_o] = stableMean<I>(inputAccumulation + idx_i, dim_i, stride_post[a]);
}
}
std::for_each(stride_pre.get()+a+1, stride_pre.get()+nb_dims, [dim_i] (std::size_t& val) { val /= dim_i; });
--
GitLab