Merge branch 'scheduling' into 'dev'

Improved scheduling

See merge request !45

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

@@ -40,7 +40,7 @@ public:
         return std::make_unique<AddImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t /*inputIdx*/) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t /*inputIdx*/) const override final;
     void forward() override;
 };
 

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

@@ -44,7 +44,7 @@ public:
         return std::make_unique<AvgPoolingImpl2D_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -59,7 +59,7 @@ public:
         return std::make_unique<BatchNormImpl2D_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -48,18 +48,6 @@ public:
     }
 
 public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbRequiredProtected(const IOIndex_t /*inputIdx*/) const override final;
-
-    NbElts_t getRequiredMemory(const IOIndex_t outputIdx, const std::vector<DimSize_t>& /*inputsSize*/) const override final;
-
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbProducedData(const IOIndex_t outputIdx) const override final;
-
-    void updateConsummerProducer() override final;
-
     void forward() override;
 
     void backward() override;

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

@@ -46,7 +46,7 @@ public:
         return std::make_unique<ConvDepthWiseImpl2D_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -47,7 +47,7 @@ class ConvImpl2D_cpu : public OperatorImpl {
     }
 
    public:
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -40,7 +40,7 @@ public:
         return std::make_unique<DivImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 
     void forward() override final;
 };

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

@@ -38,7 +38,7 @@ public:
         return std::make_unique<ErfImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -38,7 +38,6 @@ public:
         return std::make_unique<GatherImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -49,12 +49,12 @@ 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 sum = 0;
+      I mean = 0;
       for (size_t i = 0; i < in_channel_nb_elems; ++i) {
-        sum += filter_start[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] =
-          sum / static_cast<I>(in_channel_nb_elems);
+      output[batch * out_batch_nb_elems + channel] = mean;
     }
   }
 }

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

@@ -39,7 +39,7 @@ public:
         return std::make_unique<LeakyReLUImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 
     void forward() override final;
 

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

@@ -44,7 +44,7 @@ public:
         return std::make_unique<MaxPoolingImpl2D_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -29,8 +29,8 @@ public:
         return std::make_unique<MemorizeImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-    NbElts_t getRequiredMemory(const Aidge::IOIndex_t outputIdx,
+    Elts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
+    Elts_t getRequiredMemory(const Aidge::IOIndex_t outputIdx,
                                const std::vector<Aidge::DimSize_t> &/*inputsSize*/) const override final;
     void updateConsummerProducer() override final;
     void forward() override;

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

@@ -39,7 +39,7 @@ public:
         return std::make_unique<MulImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -46,7 +46,7 @@ public:
         return std::make_unique<PadImpl2D_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -39,7 +39,7 @@ public:
         return std::make_unique<PopImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

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

@@ -39,7 +39,7 @@ public:
         return std::make_unique<PowImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
     void backward() override;
 };

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

@@ -41,7 +41,7 @@ public:
         return std::make_unique<ReLUImpl_cpu>(op);
     }
 
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 
     void forward() override final;
 

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

@@ -44,7 +44,6 @@ class ReduceMeanImpl_cpu : public OperatorImpl {
     }
 
    public:
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 
@@ -87,7 +86,6 @@ class ReduceMeanImpl_cpu : public OperatorImpl {
 //     }
 
 //    public:
-//     NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 //     void forward() override;
 // };
 
@@ -100,7 +98,6 @@ class ReduceMeanImpl_cpu : public OperatorImpl {
 //     }
 
 //    public:
-//     NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 //     void forward() override;
 // };
 
@@ -113,7 +110,6 @@ class ReduceMeanImpl_cpu : public OperatorImpl {
 //     }
 
 //    public:
-//     NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
 //     void forward() override;
 // };
 namespace {

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

@@ -47,22 +47,23 @@ void ReduceMeanImpl_cpu_forward_kernel(const typename ReduceMean_Op::Attrs& attr
             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;
-                output[idx_o] = input[idx_i];
-                for (std::size_t i = 1; i < dim_i; ++i) {
-                    output[idx_o] += input[idx_i + i*stride_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] /= dim_i;
+                output[idx_o]  = mean;
             }
         }
     } else {
         std::size_t outputElements = totalElements;
 
-        std::size_t *stride_post = new std::size_t[nb_dims];
+        auto stride_post = std::unique_ptr<std::size_t[]>(new std::size_t[nb_dims]);
         stride_post[nb_dims - 1] = 1;
         for (std::size_t i = nb_dims-2; i != static_cast<std::size_t>(-1); --i) {
             stride_post[i] = stride_post[i+1]*inputDims[i+1];
         }
-        std::size_t *stride_pre = new std::size_t[nb_dims];
+        auto stride_pre = std::unique_ptr<std::size_t[]>(new std::size_t[nb_dims]);
         stride_pre[0] = 1;
         for (std::size_t i = 1; i < nb_dims; ++i) {
             stride_pre[i] = stride_pre[i-1]*inputDims[i-1];
@@ -80,13 +81,15 @@ void ReduceMeanImpl_cpu_forward_kernel(const typename ReduceMean_Op::Attrs& attr
                 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;
-                    outputAccumulation[idx_o] = inputAccumulation[idx_i];
-                    for (std::size_t i = 1; i < dim_i; ++i) {
-                        outputAccumulation[idx_o] += inputAccumulation[idx_i + i*stride_post[a]];
+                    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;
                 }
             }
-            std::for_each(stride_pre+a+1, stride_pre+nb_dims, [dim_i] (std::size_t& val) { val /= dim_i; });
+            std::for_each(stride_pre.get()+a+1, stride_pre.get()+nb_dims, [dim_i] (std::size_t& val) { val /= dim_i; });
             if (inputAccumulation != input) {
                 delete[] inputAccumulation;
             }
@@ -94,14 +97,10 @@ void ReduceMeanImpl_cpu_forward_kernel(const typename ReduceMean_Op::Attrs& attr
         }
 
         // Copy elements from inputAccumulation to output while dividing by divisor
-        I divisor = totalElements / outputElements;
-        std::transform(inputAccumulation, inputAccumulation + outputElements, output,
-                    [divisor](I element) { return element / divisor; });
+        std::copy(inputAccumulation, inputAccumulation + outputElements, output);
         if (outputAccumulation) {
             delete[] outputAccumulation;
         }
-        delete[] stride_post;
-        delete[] stride_pre;
     }
 }