Reduced OpenMP thresholds

include/aidge/backend/cpu/operator/AvgPoolingImpl_kernels.hpp

@@ -77,7 +77,7 @@ void AvgPoolingImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideD
     using signedsize = std::make_signed<std::size_t>::type;
 
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] > 32)
+    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] >= 16)
 #endif
     for (int batch = 0; batch < static_cast<int>(dims[0]); ++batch) {
         for (int ch = 0; ch < static_cast<int>(dims[1]); ++ch) {

include/aidge/backend/cpu/operator/BatchNormImpl_kernels.hpp

@@ -54,7 +54,7 @@ void BatchNormImpl2D_cpu_forward_kernel(float epsilon, float momentum, const std
 
     if ((freeze == true) || (momentum == 0.0f)) {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (nbBatch * nbChannels > 32)
+        #pragma omp parallel for collapse(2) if (nbBatch * nbChannels >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(nbBatch); ++batch) {
             for (int ch = 0; ch < static_cast<int>(nbChannels); ++ch) {

include/aidge/backend/cpu/operator/ConvDepthWiseImpl_kernels.hpp

@@ -66,7 +66,7 @@ void ConvDepthWiseImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& stri
     // does not take Dilation attribute into account
     using signedsize = std::make_signed<std::size_t>::type;
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] > 32)
+    #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] >= 16)
 #endif
     for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
         for (int ch = 0; ch < static_cast<int>(inputDims[1]); ++ch) {
@@ -156,7 +156,7 @@ void ConvDepthWiseImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& stri
 
     if (dilated_kernel_x ==3 && dilated_kernel_y == 3) {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int ch = 0; ch < static_cast<int>(inputDims[1]); ++ch) {
@@ -199,7 +199,7 @@ void ConvDepthWiseImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& stri
         }
     } else if (dilated_kernel_x == 1 && dilated_kernel_y == 1) {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int ch = 0; ch < static_cast<int>(inputDims[1]); ++ch) {
@@ -224,7 +224,7 @@ void ConvDepthWiseImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& stri
         }
     } else {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * inputDims[1] >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int ch = 0; ch < static_cast<int>(inputDims[1]); ++ch) {

include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp

@@ -60,7 +60,7 @@ void ConvImpl1D_cpu_forward_kernel(const array<DimSize_t, 1> &strideDim,
 
     using signedsize = std::make_signed<std::size_t>::type;
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels > 32)
+    #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels >= 16)
 #endif
     for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
         for (int outCh = 0; outCh < static_cast<int>(outChannels); ++outCh) {
@@ -482,7 +482,7 @@ void ConvImpl2D_cpu_forward_kernel(const array<DimSize_t, 2> &strideDims,
 
     if (dilated_kernel_x == 3 && dilated_kernel_y == 3) {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int outCh = 0; outCh < static_cast<int>(outChannels); ++outCh) {
@@ -571,7 +571,7 @@ void ConvImpl2D_cpu_forward_kernel(const array<DimSize_t, 2> &strideDims,
         }
     } else if (dilated_kernel_x == 1 && dilated_kernel_y == 1) {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int outCh = 0; outCh < static_cast<int>(outChannels); ++outCh) {
@@ -607,7 +607,7 @@ void ConvImpl2D_cpu_forward_kernel(const array<DimSize_t, 2> &strideDims,
         }
     } else {
 #ifdef _OPENMP
-        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels > 32)
+        #pragma omp parallel for collapse(2) if (inputDims[0] * outChannels >= 16)
 #endif
         for (int batch = 0; batch < static_cast<int>(inputDims[0]); ++batch) {
             for (int outCh = 0; outCh < static_cast<int>(outChannels); ++outCh) {

include/aidge/backend/cpu/operator/FCImpl_kernels.hpp

@@ -97,7 +97,7 @@ void FCImpl_cpu_forward_kernel(const DimSize_t batchSize,
     O* output = static_cast<O*>(output_);
 
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (batchSize * outputFeatureSize > 32)
+    #pragma omp parallel for collapse(2) if (batchSize * outputFeatureSize >= 16)
 #endif
     for (int batch = 0; batch < static_cast<int>(batchSize); ++batch) {
         for (int out = 0; out < static_cast<int>(outputFeatureSize); ++out) {

include/aidge/backend/cpu/operator/GlobalAveragePoolingImpl_kernels.hpp

@@ -75,7 +75,7 @@ void GlobalAveragePoolingImpl_cpu_forward_kernel(const std::shared_ptr<Tensor>&
     // parse channel by channel and fill each output with the average of the
     // values in the channel
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] > 32)
+    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] >= 16)
 #endif
     for (int batch = 0; batch < static_cast<int>(dims[0]); ++batch) {
         for (int channel = 0; channel < static_cast<int>(dims[1]); ++channel) {

include/aidge/backend/cpu/operator/MatMulImpl_kernels.hpp

@@ -27,7 +27,7 @@ void MatMulImpl_cpu_forward_kernel(const std::size_t n, const std::size_t k, con
     std::memset(output, O(0), n * m * sizeof(O));
 
 #ifdef _OPENMP
-    #pragma omp parallel for if (n > 32)
+    #pragma omp parallel for if (n >= 16)
 #endif
     for (int i = 0; i < static_cast<int>(n); ++i) {
         for (std::size_t l = 0; l < k; ++l) {

include/aidge/backend/cpu/operator/MaxPoolingImpl_kernels.hpp

@@ -67,7 +67,7 @@ void MaxPoolingImpl2D_cpu_forward_kernel(
   using signedsize = std::make_signed<std::size_t>::type;
 
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] > 32)
+    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] >= 16)
 #endif
   for (int batch = 0; batch < static_cast<int>(dims[0]); ++batch){
     for (int channel = 0; channel < static_cast<int>(dims[1]); ++channel){

include/aidge/backend/cpu/operator/SoftmaxImpl_kernels.hpp

@@ -38,7 +38,7 @@ void SoftmaxImpl_cpu_forward_kernel(std::size_t axisIdx, const std::vector<DimSi
     }
 
 #ifdef _OPENMP
-    #pragma omp parallel for collapse(2) if (preAxisElems * postAxisElems > 32)
+    #pragma omp parallel for collapse(2) if (preAxisElems * postAxisElems >= 16)
 #endif
     for (int i = 0; i < static_cast<int>(preAxisElems); ++i) {
         for (int j = 0; j < static_cast<int>(postAxisElems); ++j) {