From 8ebc6ad5013f260ac8f4844b259021b4d6f3c5d9 Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Fri, 11 Jul 2025 22:43:48 +0200
Subject: [PATCH] Added utils functions
---
aidge_export_cpp/static/activation_utils.hpp | 5 +-
aidge_export_cpp/static/macs.hpp | 1142 +++++++++++++++++-
aidge_export_cpp/static/rescaling_utils.hpp | 189 ++-
aidge_export_cpp/static/typedefs.hpp | 510 +++++++-
4 files changed, 1798 insertions(+), 48 deletions(-)
diff --git a/aidge_export_cpp/static/activation_utils.hpp b/aidge_export_cpp/static/activation_utils.hpp
index ddac7fb..28cd5f6 100644
--- a/aidge_export_cpp/static/activation_utils.hpp
+++ b/aidge_export_cpp/static/activation_utils.hpp
@@ -1,4 +1,5 @@
-#pragma once
+#ifndef __AIDGE_EXPORT_CPP_ACTIVATION_UTILS_HPP__
+#define __AIDGE_EXPORT_CPP_ACTIVATION_UTILS_HPP__
#include <type_traits>
#include "network/typedefs.hpp"
@@ -50,3 +51,5 @@ Output_T activation_forward_value (Sum_T weightedSum,
return saturate<Output_T>(rescaling(weightedSum, output), 8 * sizeof(Output_T));
}
+
+#endif
diff --git a/aidge_export_cpp/static/macs.hpp b/aidge_export_cpp/static/macs.hpp
index 6efba37..af7fb9b 100644
--- a/aidge_export_cpp/static/macs.hpp
+++ b/aidge_export_cpp/static/macs.hpp
@@ -1,19 +1,1145 @@
#ifndef __AIDGE_EXPORT_CPP_MACS_HPP__
#define __AIDGE_EXPORT_CPP_MACS_HPP__
+/***************************************************************************************
+*****************************************w4-a8******************************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += (*inputs) * weights[0*WEIGHTS_INC].fields.op1;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op1
+ + inputs[1*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op0;
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && NB_ITERATIONS >= 2 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = dualMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-2, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 2*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************w4-a4******************************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC].fields.op1
+ * weights[0*WEIGHTS_INC].fields.op1;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum
+ += inputs[0*INPUTS_INC].fields.op1
+ * weights[0*WEIGHTS_INC].fields.op1
+ + inputs[0*INPUTS_INC].fields.op0
+ * weights[0*WEIGHTS_INC].fields.op0;
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 4
+ && NB_ITERATIONS >=2 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = dualMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-2, INPUTS_INC, WEIGHTS_INC>(
+ inputs + INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+//
+/***************************************************************************************
+*********************************final_FC:w8-a4******************************************
+***************************************************************************************/
+
+ template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC].fields.op1 * weights[0*WEIGHTS_INC];
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC].fields.op1 * weights[0*WEIGHTS_INC]
+ + inputs[0*INPUTS_INC].fields.op0 * weights[1*WEIGHTS_INC];
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS >= 2 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = dualMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-2, INPUTS_INC, WEIGHTS_INC>(
+ inputs + INPUTS_INC, weights + 2*WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************w2-a8******************************************
+***************************************************************************************/
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op3
+ + inputs[1*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op2;
+
+ return weightedSum;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static Sum_T tripleMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op3
+ + inputs[1*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op2
+ + inputs[2*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op1;
+
+ return weightedSum;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static Sum_T quadMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op3
+ + inputs[1*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op2
+ + inputs[2*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op1
+ + inputs[3*INPUTS_INC] * weights[0*WEIGHTS_INC].fields.op0;
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += (*inputs) * weights[0*WEIGHTS_INC].fields.op3;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && NB_ITERATIONS == 2 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = dualMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-2, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 2*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && NB_ITERATIONS == 3 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = tripleMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-3, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 3*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 2
+ && NB_ITERATIONS >= 4 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = quadMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-4, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 4*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************w1-a8******************************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-(*inputs)) : (Sum_T)(*inputs));
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 2 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC]);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 3 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC]);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 4 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC]) : (Sum_T)inputs[3*INPUTS_INC]);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 5 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC]) : (Sum_T)inputs[3*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[4*INPUTS_INC]) : (Sum_T)inputs[4*INPUTS_INC]);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 6 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC]) : (Sum_T)inputs[3*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[4*INPUTS_INC]) : (Sum_T)inputs[4*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[5*INPUTS_INC]) : (Sum_T)inputs[5*INPUTS_INC]);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 7 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC]) : (Sum_T)inputs[3*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[4*INPUTS_INC]) : (Sum_T)inputs[4*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[5*INPUTS_INC]) : (Sum_T)inputs[5*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[6*INPUTS_INC]) : (Sum_T)inputs[6*INPUTS_INC]);
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static Sum_T octoMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC]) : (Sum_T)inputs[0*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC]) : (Sum_T)inputs[1*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC]) : (Sum_T)inputs[2*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC]) : (Sum_T)inputs[3*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[4*INPUTS_INC]) : (Sum_T)inputs[4*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[5*INPUTS_INC]) : (Sum_T)inputs[5*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[6*INPUTS_INC]) : (Sum_T)inputs[6*INPUTS_INC])
+ + ((weights[0*WEIGHTS_INC].fields.op0 == 0)
+ ? (Sum_T)(-inputs[7*INPUTS_INC]) : (Sum_T)inputs[7*INPUTS_INC]);
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS >= 8 && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = octoMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-8, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 8*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************w1-a4******************************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 2 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 3 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 4 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 5 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op1) : (Sum_T)inputs[2*INPUTS_INC].fields.op1);
+}
+
template<int NB_ITERATIONS,
- int INPUTS_INC = 1,
- int WEIGHTS_INC = 1,
- class Input_T,
- class Weight_T,
- class Sum_T>
-static void macsOnRange(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& __restrict weightedSum)
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 6 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op1) : (Sum_T)inputs[2*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op0) : (Sum_T)inputs[2*INPUTS_INC].fields.op0);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 7 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op1) : (Sum_T)inputs[2*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op0) : (Sum_T)inputs[2*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC].fields.op1) : (Sum_T)inputs[3*INPUTS_INC].fields.op1);
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static Sum_T octoMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op1) : (Sum_T)inputs[2*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[2*INPUTS_INC].fields.op0) : (Sum_T)inputs[2*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC].fields.op1) : (Sum_T)inputs[3*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op0 == 0)
+ ? (Sum_T)(-inputs[3*INPUTS_INC].fields.op0) : (Sum_T)inputs[3*INPUTS_INC].fields.op0);
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS >= 8 && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = octoMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-8, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 4*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************w1-a2******************************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 2 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 3 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 4 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 5 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op3) : (Sum_T)inputs[1*INPUTS_INC].fields.op3);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 6 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op3) : (Sum_T)inputs[1*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op2) : (Sum_T)inputs[1*INPUTS_INC].fields.op2);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS == 7 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op3) : (Sum_T)inputs[1*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op2) : (Sum_T)inputs[1*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1);
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static Sum_T octoMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += ((weights[0*WEIGHTS_INC].fields.op7 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op3) : (Sum_T)inputs[0*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op6 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op2) : (Sum_T)inputs[0*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op5 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op1) : (Sum_T)inputs[0*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op4 == 0)
+ ? (Sum_T)(-inputs[0*INPUTS_INC].fields.op0) : (Sum_T)inputs[0*INPUTS_INC].fields.op0)
+ + ((weights[0*WEIGHTS_INC].fields.op3 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op3) : (Sum_T)inputs[1*INPUTS_INC].fields.op3)
+ + ((weights[0*WEIGHTS_INC].fields.op2 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op2) : (Sum_T)inputs[1*INPUTS_INC].fields.op2)
+ + ((weights[0*WEIGHTS_INC].fields.op1 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op1) : (Sum_T)inputs[1*INPUTS_INC].fields.op1)
+ + ((weights[0*WEIGHTS_INC].fields.op0 == 0)
+ ? (Sum_T)(-inputs[1*INPUTS_INC].fields.op0) : (Sum_T)inputs[1*INPUTS_INC].fields.op0);
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 1
+ && NB_ITERATIONS >= 8 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = octoMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-8, INPUTS_INC, WEIGHTS_INC>(
+ inputs + 2*INPUTS_INC, weights + WEIGHTS_INC, weightedSum);
+}
+
+//
+/***************************************************************************************
+*********************************final_FC:w8-a2******************************************
+***************************************************************************************/
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += weights[0*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op3
+ + weights[1*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op2;
+
+ return weightedSum;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static Sum_T tripleMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += weights[0*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op3
+ + weights[1*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op2
+ + weights[2*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op1;
+
+ return weightedSum;
+}
+
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T, typename Sum_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static Sum_T quadMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
+{
+ weightedSum += weights[0*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op3
+ + weights[1*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op2
+ + weights[2*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op1
+ + weights[3*WEIGHTS_INC] * inputs[0*INPUTS_INC].fields.op0;
+
+ return weightedSum;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS == 1 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum += (*weights) * inputs[0*INPUTS_INC].fields.op3;
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS == 2 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = dualMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-2, INPUTS_INC, WEIGHTS_INC>(
+ inputs + INPUTS_INC, weights + 2*WEIGHTS_INC, weightedSum);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS == 3 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = tripleMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-3, INPUTS_INC, WEIGHTS_INC>(
+ inputs + INPUTS_INC, weights + 3*WEIGHTS_INC, weightedSum);
+}
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits == 8
+ && NB_ITERATIONS >= 4 && std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+ weightedSum = quadMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS-4, INPUTS_INC, WEIGHTS_INC>(
+ inputs + INPUTS_INC, weights + 4*WEIGHTS_INC, weightedSum);
+}
+
+/***************************************************************************************
+*****************************************0-iterations***********************************
+***************************************************************************************/
+
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<NB_ITERATIONS == 0>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
+{
+}
+
+/***************************************************************************************
+**************************************>=8-bit******************************************
+***************************************************************************************/
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Weight_T, typename Sum_T,
+ class Input_T,
+ typename std::enable_if<(std::numeric_limits<Weight_T>::digits >= 8
+ && std::numeric_limits<Input_T>::digits >= 8)>::type* = nullptr>
+static void macsOnRange(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T& __restrict weightedSum)
{
for (int iter = 0; iter < NB_ITERATIONS; ++iter) {
weightedSum += inputs[iter*INPUTS_INC] * weights[iter*WEIGHTS_INC];
}
}
+/***************************************************************************************
+**************************************PACK_ACTIVATIONS**********************************
+***************************************************************************************/
+template<typename Output_T>
+static void compact_data_during_loop (const uint8_t value,
+ Output_T* __restrict outputs,
+ int* outputOffset,
+ PackSupport* infoPack)
+{
+ constexpr unsigned int mask = (1L << std::numeric_limits<Output_T>::digits) - 1;
+ constexpr unsigned int nbSlot = ceil((double)8/std::numeric_limits<Output_T>::digits);
+
+ infoPack->accumulator |= value & mask;
+ infoPack->cptAccumulator += 1;
+
+ if (infoPack->cptAccumulator == nbSlot) {
+ outputs[*(outputOffset)] = infoPack->accumulator;
+ infoPack->cptAccumulator = 0;
+ infoPack->accumulator = 0;
+ }
+ else {
+ infoPack->accumulator <<= std::numeric_limits<Output_T>::digits;
+ }
+}
+
+template<typename Output_T>
+static void compact_data_end_loop (Output_T* __restrict outputs,
+ int* outputOffset,
+ PackSupport* infoPack)
+{
+ // if data still accumulated but not stored
+ if (infoPack->cptAccumulator != 0) {
+ constexpr unsigned int nbSlot = ceil((double)8/std::numeric_limits<Output_T>::digits);
+
+ // Add extra zero to shift data to the left
+ infoPack->cptAccumulator += 1;
+ while (infoPack->cptAccumulator < nbSlot) {
+ infoPack->accumulator <<= std::numeric_limits<Output_T>::digits;
+ infoPack->cptAccumulator += 1;
+ }
+ outputs[*(outputOffset)] = infoPack->accumulator;
+ infoPack->cptAccumulator = 0;
+ infoPack->accumulator = 0;
+ }
+}
+
+
+/*************************************************************************************
+***************************************MAX_POOL_4bits*********************************
+***************************************************************************************/
+template<class Input_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void fillMaxVal(Input_T& __restrict maxVal)
+{
+ maxVal.fields.op0 = std::numeric_limits<Input_T>::lowest();
+ maxVal.fields.op1 = std::numeric_limits<Input_T>::lowest();
+}
+
+template<class Input_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void findMaxVal(const Input_T& __restrict inputs, Input_T& __restrict maxVal)
+{
+ if(inputs.fields.op1 > maxVal.fields.op1){
+ maxVal.fields.op1 = inputs.fields.op1;
+ }
+ if(inputs.fields.op0 > maxVal.fields.op0){
+ maxVal.fields.op0 = inputs.fields.op0;
+ }
+}
+
+template<class Input_T, class Output_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
+static void packMaxVal(Input_T& __restrict maxVal,
+ Output_T* __restrict outputs,
+ int* outputOffset,
+ PackSupport* infoPack)
+{
+ compact_data_during_loop(maxVal.fields.op1, outputs, outputOffset, infoPack);
+ compact_data_during_loop(maxVal.fields.op0, outputs, outputOffset, infoPack);
+}
+
+/*************************************************************************************
+***************************************MAX_POOL_2bits*********************************
+***************************************************************************************/
+template<class Input_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void fillMaxVal(Input_T& __restrict maxVal)
+{
+ maxVal.fields.op3 = std::numeric_limits<Input_T>::lowest();
+ maxVal.fields.op2 = std::numeric_limits<Input_T>::lowest();
+ maxVal.fields.op0 = std::numeric_limits<Input_T>::lowest();
+ maxVal.fields.op1 = std::numeric_limits<Input_T>::lowest();
+}
+
+template<class Input_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void findMaxVal(const Input_T& __restrict inputs, Input_T& __restrict maxVal)
+{
+ if(inputs.fields.op3 > maxVal.fields.op3){
+ maxVal.fields.op3 = inputs.fields.op3;
+ }
+ if(inputs.fields.op2 > maxVal.fields.op2){
+ maxVal.fields.op2 = inputs.fields.op2;
+ }
+ if(inputs.fields.op1 > maxVal.fields.op1){
+ maxVal.fields.op1 = inputs.fields.op1;
+ }
+ if(inputs.fields.op0 > maxVal.fields.op0){
+ maxVal.fields.op0 = inputs.fields.op0;
+ }
+}
+
+template<class Input_T, class Output_T,
+ typename std::enable_if<(std::numeric_limits<Input_T>::digits == 2)>::type* = nullptr>
+static void packMaxVal(Input_T& __restrict maxVal,
+ Output_T* __restrict outputs,
+ int* outputOffset,
+ PackSupport* infoPack)
+{
+ compact_data_during_loop(maxVal.fields.op3, outputs, outputOffset, infoPack);
+ compact_data_during_loop(maxVal.fields.op2, outputs, outputOffset, infoPack);
+ compact_data_during_loop(maxVal.fields.op1, outputs, outputOffset, infoPack);
+ compact_data_during_loop(maxVal.fields.op0, outputs, outputOffset, infoPack);
+}
+
#endif
diff --git a/aidge_export_cpp/static/rescaling_utils.hpp b/aidge_export_cpp/static/rescaling_utils.hpp
index 01377bb..bfd229f 100644
--- a/aidge_export_cpp/static/rescaling_utils.hpp
+++ b/aidge_export_cpp/static/rescaling_utils.hpp
@@ -1,4 +1,5 @@
-#pragma once
+#ifndef __AIDGE_EXPORT_CPP_RESCALING_UTILS_HPP__
+#define __AIDGE_EXPORT_CPP_RESCALING_UTILS_HPP__
// ---------------------------------------------------
// ----------------- Saturate Utils ------------------
@@ -15,64 +16,176 @@ constexpr int64_t smlal(int32_t lhs, int32_t rhs,
}
// ---------------------------------------------------
-// --------------- Scaling by Shifting ---------------
+// ------------------- No Scaling --------------------
// ---------------------------------------------------
-template<int SHIFT>
-struct SingleShiftScaling {
+struct NoScaling {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ return weightedSum;
+ }
+};
+struct FloatingPointScaling {
template<typename Sum_T>
- Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const
- {
- return (SHIFT != 0) ? ((weightedSum >> (SHIFT - 1)) + 1) >> 1 // Rounding
- : weightedSum;
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ return round(weightedSum*mScaling);
}
- // // Shift attribute
- // static const int mShift = SHIFT;
- // static const Scaling_T mScalingType = SingleShift;
+ double mScaling;
+};
- // // FP Attribute
- // static const int32_t mScaling = 0;
- // static const int64_t mFractionalBits = 0;
+template<size_t SIZE>
+struct FloatingPointScalingPerChannel {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ return round(weightedSum * mScaling[output]);
+ }
+ double mScaling[SIZE];
};
-// ---------------------------------------------------
-// --------------- Fixed Point Scaling ---------------
-// ---------------------------------------------------
+template<size_t SIZE>
+struct FloatingPointClippingAndScaling {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ Sum_T clipValue = weightedSum;
+ clipValue = (clipValue < Sum_T(0)) ?
+ Sum_T(0) : (clipValue > Sum_T(mClipping)) ?
+ Sum_T(mClipping) : clipValue;
+ return round(clipValue * mScaling);
+ }
+
+ double mScaling;
+ int32_t mClipping;
+};
-template<int64_t SHIFT, int32_t COEF>
+template<size_t SIZE>
+struct FloatingPointClippingAndScalingPerChannel {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ Sum_T clipValue = weightedSum;
+ clipValue = (clipValue < Sum_T(0)) ?
+ Sum_T(0) : (clipValue > Sum_T(mClipping[output])) ?
+ Sum_T(mClipping[output]) : clipValue;
+ return round(clipValue * mScaling[output]);
+ }
+
+ double mScaling[SIZE];
+ int32_t mClipping[SIZE];
+};
+
+template<int32_t SCALING, int64_t FRACTIONAL_BITS>
struct FixedPointScaling {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ // Different rounding if weightesSum < 0
+ // if(weightedSum < 0) {
+ // HALF--;
+ // }
+
+ return smlal(weightedSum, SCALING, HALF_LO, HALF_HI) >> FRACTIONAL_BITS;
+ }
+ static const uint32_t HALF_LO = (FRACTIONAL_BITS > 0)
+ ? (1ull << (FRACTIONAL_BITS - 1)) & 0xFFFFFFFF : 0;
+ static const uint32_t HALF_HI = (FRACTIONAL_BITS > 0)
+ ? (1ull << (FRACTIONAL_BITS - 1)) >> 32u : 0;
+};
+
+template<size_t SIZE, int64_t FRACTIONAL_BITS>
+struct FixedPointScalingScalingPerChannel {
template<typename Sum_T>
- Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const
- {
- return smlal(weightedSum, COEF, HALF_LO, HALF_HI) >> SHIFT;
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ // Different rounding if weightesSum < 0
+ // if(weightedSum < 0) {
+ // HALF--;
+ // }
+
+ return smlal(weightedSum, mScaling[output], HALF_LO, HALF_HI) >> FRACTIONAL_BITS;
}
- // Attributes
- static constexpr uint32_t HALF_LO = (SHIFT > 0)
- ? (1ull << (SHIFT - 1)) & 0xFFFFFFFF : 0;
- static constexpr uint32_t HALF_HI = (SHIFT > 0)
- ? (1ull << (SHIFT - 1)) >> 32u : 0;
-
- // static const int32_t mScaling = SCALING;
- // static const int64_t mFractionalBits = FRACTIONAL_BITS;
- // static const Scaling_T mScalingType = FixedPoint;
- // static const int mShift = 0;
+ static const uint32_t HALF_LO = (FRACTIONAL_BITS > 0)
+ ? (1ull << (FRACTIONAL_BITS - 1)) & 0xFFFFFFFF : 0;
+ static const uint32_t HALF_HI = (FRACTIONAL_BITS > 0)
+ ? (1ull << (FRACTIONAL_BITS - 1)) >> 32u : 0;
+
+ int32_t mScaling[SIZE];
};
-// ---------------------------------------------------
-// ------------------- No Scaling --------------------
-// ---------------------------------------------------
+template<size_t SIZE, int64_t FRACTIONAL_BITS>
+struct FixedPointClippingAndScalingPerChannel {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ // Different rounding if weightesSum < 0
+ // if(weightedSum < 0) {
+ // HALF--;
+ // }
+ Sum_T clipValue = weightedSum;
+ clipValue = (clipValue < Sum_T(0)) ?
+ Sum_T(0) : (clipValue > Sum_T(mClipping[output])) ?
+ Sum_T(mClipping[output]) : clipValue;
+
+ return smlal(clipValue, mScaling[output], HALF_LO, HALF_HI) >> FRACTIONAL_BITS;
+ }
-struct NoScaling {
+ static const uint32_t HALF_LO = (1ull << (FRACTIONAL_BITS - 1)) & 0xFFFFFFFF;
+ static const uint32_t HALF_HI = (1ull << (FRACTIONAL_BITS - 1)) >> 32u;
+ int32_t mScaling[SIZE];
+ int32_t mClipping[SIZE];
+};
+
+template<size_t SHIFT>
+struct SingleShiftScaling {
template<typename Sum_T>
- Sum_T operator()(Sum_T weightedSum, unsigned int /*output*/) const
- {
- return weightedSum;
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ return (SHIFT != 0) ? ((weightedSum >> (SHIFT - 1)) + 1) >> 1 // Rounding
+ : weightedSum;
+ }
+};
+
+template<size_t SIZE>
+struct SingleShiftScalingPerChannel {
+ template<typename Sum_T>
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ return (mScaling[output] != 0) ? ((weightedSum >> (mScaling[output] - 1)) + 1) >> 1 // Rounding
+ : weightedSum;
}
+ unsigned char mScaling[SIZE];
};
+
+template<size_t SHIFT1, size_t SHIFT2, typename Sum_T>
+struct DoubleShiftScaling {
+ Sum_T operator()(Sum_T weightedSum, size_t /*output*/) const {
+ // Different rounding if weightesSum < 0
+ // if(weightedSum < 0) {
+ // HALF--;
+ // }
+
+ return (weightedSum + (weightedSum << SHIFT1) + HALF) >> SHIFT2;
+ }
+
+ static const Sum_T HALF = ((Sum_T) 1) << (SHIFT2 - 1);
+};
+
+template<size_t SIZE, bool UNSIGNED_WEIGHTED_SUM, typename Sum_T>
+struct DoubleShiftScalingPerChannel {
+ Sum_T operator()(Sum_T weightedSum, size_t output) const {
+ const Sum_T SHIFT1 = mScaling[output][0];
+ const Sum_T SHIFT2 = mScaling[output][1];
+ const Sum_T HALF = mScaling[output][2];
+
+ // Different rounding if weightesSum < 0
+ // if(weightedSum < 0) {
+ // HALF--;
+ // }
+
+ return (weightedSum + (weightedSum << SHIFT1) + HALF) >> SHIFT2;
+ }
+
+ Sum_T mScaling[SIZE][3];
+};
+
+#endif
diff --git a/aidge_export_cpp/static/typedefs.hpp b/aidge_export_cpp/static/typedefs.hpp
index 8ecd095..7a8dd76 100644
--- a/aidge_export_cpp/static/typedefs.hpp
+++ b/aidge_export_cpp/static/typedefs.hpp
@@ -1,7 +1,10 @@
#ifndef __AIDGE_EXPORT_CPP_NETWORK_TYPEDEFS__
#define __AIDGE_EXPORT_CPP_NETWORK_TYPEDEFS__
-#include <stdint.h>
+#include <cmath>
+#include <type_traits>
+#include <limits>
+#include <cstdint>
typedef enum {
Tanh,
@@ -30,4 +33,509 @@ typedef enum {
} CoeffMode_T;
+// ----------------------------------------------------------------------------
+// -------------------- Generic custom bit-width types ------------------------
+// ----------------------------------------------------------------------------
+
+template <int BITWIDTH>
+struct data {};
+
+template <int BITWIDTH>
+struct udata{};
+
+namespace std {
+ // Specialization of STL, allows to use std::is_unsigned<> for example.
+ template <int BITWIDTH>
+ struct is_integral<data<BITWIDTH>>
+ : std::is_integral<decltype(data<BITWIDTH>::value)>::type {};
+ template <int BITWIDTH>
+ struct is_integral<udata<BITWIDTH>>
+ : std::is_integral<decltype(udata<BITWIDTH>::value)>::type {};
+ template <int BITWIDTH>
+ struct is_floating_point<data<BITWIDTH>>
+ : std::is_floating_point<decltype(data<BITWIDTH>::value)>::type {};
+ template <int BITWIDTH>
+ struct is_unsigned<data<BITWIDTH>>
+ : std::is_unsigned<decltype(data<BITWIDTH>::value)>::type {};
+ template <int BITWIDTH>
+ struct is_unsigned<udata<BITWIDTH>>
+ : std::is_unsigned<decltype(udata<BITWIDTH>::value)>::type {};
+
+ template <int BITWIDTH>
+ class numeric_limits<data<BITWIDTH>> {
+ public:
+ static constexpr int is_integer = (BITWIDTH > 0);
+ static constexpr int is_signed = true;
+ static constexpr int digits = std::abs(BITWIDTH);
+ static constexpr decltype(data<BITWIDTH>::value) min() noexcept
+ { return (BITWIDTH > 0) ? -(1 << (BITWIDTH - 1)) :
+ std::numeric_limits<decltype(data<BITWIDTH>::value)>::min(); };
+ static constexpr decltype(data<BITWIDTH>::value) lowest() noexcept
+ { return (BITWIDTH > 0) ? -(1 << (BITWIDTH - 1)) :
+ std::numeric_limits<decltype(data<BITWIDTH>::value)>::lowest(); };
+ static constexpr decltype(data<BITWIDTH>::value) max() noexcept
+ { return (BITWIDTH > 0) ? ((1 << (BITWIDTH - 1)) - 1) :
+ std::numeric_limits<decltype(data<BITWIDTH>::value)>::max(); };
+ };
+ template <int BITWIDTH>
+ class numeric_limits<udata<BITWIDTH>> {
+ public:
+ static constexpr int is_integer = true;
+ static constexpr int is_signed = false;
+ static constexpr int digits = BITWIDTH;
+ static constexpr decltype(data<BITWIDTH>::value) min() noexcept
+ { return 0; };
+ static constexpr decltype(data<BITWIDTH>::value) lowest() noexcept
+ { return 0; };
+ static constexpr decltype(data<BITWIDTH>::value) max() noexcept
+ { return ((1 << BITWIDTH) - 1); };
+ };
+}
+
+
+// ----------------------------------------------------------------------------
+// -------------- Custom bit-width types operator overloading -----------------
+// ----------------------------------------------------------------------------
+
+// data
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH>& operator+=(data<BITWIDTH>& d, T rhs)
+// {return d.value += decltype(data<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH> operator+(data<BITWIDTH> d, T rhs)
+// {return d += rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH>& operator-=(data<BITWIDTH>& d, T rhs)
+// {return d.value -= decltype(data<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH> operator-(data<BITWIDTH> d, T rhs)
+// {return d -= rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH>& operator*=(data<BITWIDTH>& d, T rhs)
+// {return d.value *= decltype(data<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH> operator*(data<BITWIDTH> d, T rhs)
+// {return d *= rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH>& operator/=(data<BITWIDTH>& d, T rhs)
+// {return d.value /= decltype(data<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr data<BITWIDTH> operator/(data<BITWIDTH> d, T rhs)
+// {return d /= rhs;}
+
+
+// udata
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH>& operator+=(udata<BITWIDTH>& d, T rhs)
+// {return d.value += decltype(udata<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH> operator+(udata<BITWIDTH> d, T rhs)
+// {return d += rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH>& operator-=(udata<BITWIDTH>& d, T rhs)
+// {return d.value -= decltype(udata<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH> operator-(udata<BITWIDTH> d, T rhs)
+// {return d -= rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH>& operator*=(udata<BITWIDTH>& d, T rhs)
+// {return d.value *= decltype(udata<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH> operator*(udata<BITWIDTH> d, T rhs)
+// {return d *= rhs;}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH>& operator/=(udata<BITWIDTH>& d, T rhs)
+// {return d.value /= decltype(udata<BITWIDTH>::value)(rhs);}
+
+// template<int BITWIDTH, typename T>
+// constexpr udata<BITWIDTH> operator/(udata<BITWIDTH> d, T rhs)
+// {return d /= rhs;}
+
+
+// ----------------------------------------------------------------------------
+// ---------------- Custom bit-width types specializations --------------------
+// ----------------------------------------------------------------------------
+
+// Data structure for double
+template <>
+struct data<-64>
+{
+ data<-64>() = default;
+ constexpr data<-64>(double v): value(v) {};
+ constexpr operator double() const { return value; }
+ union {
+ double value;
+ };
+};
+
+// Data structure for float
+template <>
+struct data<-32>
+{
+ data<-32>() = default;
+ constexpr data<-32>(float v): value(v) {};
+ constexpr operator float() const { return value; }
+ union {
+ float value;
+ };
+};
+
+// Data structure for half float
+template <>
+struct data<-16>
+{
+ data<-16>() = default;
+ constexpr data<-16>(float v): value(v) {};
+ constexpr operator float() const { return value; }
+ union {
+ float value;
+ };
+};
+
+// Data structure for int32
+template <>
+struct data<32>
+{
+ data<32>() = default;
+ constexpr data<32>(int32_t v): value(v) {};
+ constexpr operator int32_t() const { return value; }
+ union {
+ int32_t value;
+ };
+};
+
+// Data structure for uint32
+template <>
+struct udata<32>
+{
+ udata<32>() = default;
+ constexpr udata<32>(uint32_t v): value(v) {};
+ constexpr operator uint32_t() const { return value; }
+ union {
+ uint32_t value;
+ };
+};
+
+// Data structure for int16
+template <>
+struct data<16>
+{
+ data<16>() = default;
+ constexpr data<16>(int16_t v): value(v) {};
+ constexpr operator int16_t() const { return value; }
+ union {
+ int16_t value;
+ };
+};
+
+// Data structure for uint16
+template <>
+struct udata<16>
+{
+ udata<16>() = default;
+ constexpr udata<16>(uint16_t v): value(v) {};
+ constexpr operator uint16_t() const { return value; }
+ union {
+ uint16_t value;
+ };
+};
+
+// Data structure for int8
+template <>
+struct data<8>
+{
+ data<8>() = default;
+ constexpr data<8>(int8_t v): value(v) {};
+ constexpr operator int8_t() const { return value; }
+ union {
+ int8_t value;
+ };
+};
+
+// Data structure for uint8
+template <>
+struct udata<8>
+{
+ udata<8>() = default;
+ constexpr udata<8>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ };
+};
+
+// Data structure for int7
+template <>
+struct data<7>
+{
+ data<7>() = default;
+ constexpr data<7>(int8_t v): value(v) {};
+ constexpr operator int8_t() const { return value; }
+ union {
+ int8_t value;
+ };
+};
+
+// Data structure for uint7
+template <>
+struct udata<7>
+{
+ udata<7>() = default;
+ constexpr udata<7>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ };
+};
+
+// Data structure for int6
+template <>
+struct data<6>
+{
+ data<6>() = default;
+ constexpr data<6>(int8_t v): value(v) {};
+ constexpr operator int8_t() const { return value; }
+ union {
+ int8_t value;
+ };
+};
+
+// Data structure for uint6
+template <>
+struct udata<6>
+{
+ udata<6>() = default;
+ constexpr udata<6>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ };
+};
+
+// Data structure for int5
+template <>
+struct data<5>
+{
+ data<5>() = default;
+ constexpr data<5>(int8_t v): value(v) {};
+ constexpr operator int8_t() const { return value; }
+ union {
+ int8_t value;
+ };
+};
+
+// Data structure for uint5
+template <>
+struct udata<5>
+{
+ udata<5>() = default;
+ constexpr udata<5>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ };
+};
+
+// Data structure for 2 * int4
+template <>
+struct data<4>
+{
+ data<4>() = default;
+ constexpr data<4>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ int8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ int8_t op0 : 4;
+ int8_t op1 : 4;
+ } fields;
+ };
+};
+
+// Data structure for 2 * uint4
+template <>
+struct udata<4>
+{
+ udata<4>() = default;
+ constexpr udata<4>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ uint8_t op0 : 4;
+ uint8_t op1 : 4;
+ } fields;
+ };
+};
+
+// Data structure for 2 * int3
+template <>
+struct data<3>
+{
+ data<3>() = default;
+ constexpr data<3>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ int8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ int8_t op0 : 4;
+ int8_t op1 : 4;
+ } fields;
+ };
+};
+
+// Data structure for 2 * uint3
+template <>
+struct udata<3>
+{
+ udata<3>() = default;
+ constexpr udata<3>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ uint8_t op0 : 4;
+ uint8_t op1 : 4;
+ } fields;
+ };
+};
+
+// Data structure for 4 * int2
+template <>
+struct data<2>
+{
+ data<2>() = default;
+ constexpr data<2>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ int8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ int8_t op0 : 2;
+ int8_t op1 : 2;
+ int8_t op2 : 2;
+ int8_t op3 : 2;
+ } fields;
+ };
+};
+
+// Data structure for 4 * uint2
+template <>
+struct udata<2>
+{
+ udata<2>() = default;
+ constexpr udata<2>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ uint8_t op0 : 2;
+ uint8_t op1 : 2;
+ uint8_t op2 : 2;
+ uint8_t op3 : 2;
+ } fields;
+ };
+};
+
+// Data structure for 8 * int1
+template <>
+struct data<1>
+{
+ data<1>() = default;
+ constexpr data<1>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ int8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ int8_t op0 : 1;
+ int8_t op1 : 1;
+ int8_t op2 : 1;
+ int8_t op3 : 1;
+ int8_t op4 : 1;
+ int8_t op5 : 1;
+ int8_t op6 : 1;
+ int8_t op7 : 1;
+ } fields;
+ };
+};
+
+// Data structure for 8 * uint1
+template <>
+struct udata<1>
+{
+ udata<1>() = default;
+ constexpr udata<1>(uint8_t v): value(v) {};
+ constexpr operator uint8_t() const { return value; }
+ union {
+ uint8_t value;
+ uint8_t uvalue;
+ struct
+ {
+ uint8_t op0 : 1;
+ uint8_t op1 : 1;
+ uint8_t op2 : 1;
+ uint8_t op3 : 1;
+ uint8_t op4 : 1;
+ uint8_t op5 : 1;
+ uint8_t op6 : 1;
+ uint8_t op7 : 1;
+ } fields;
+ };
+};
+
+
+// ----------------------------------------------------------------------------
+// ------------------------- Structures and Unions ----------------------------
+// ----------------------------------------------------------------------------
+
+/* Object for compressing the outputs after mac operations */
+typedef struct PackSupport {
+ uint8_t accumulator;
+ unsigned int cptAccumulator;
+} PackSupport;
+
+/* Union to access the data<32>/data<8>/data<4>/data<1> types */
+union dataword
+{
+ data<32> word;
+ data<8> bytes[4];
+ data<4> half_bytes[4];
+ data<1> bitfields[4];
+};
+
+/* Union to access the udata<32>/udata<8>/udata<4>/udata<1> types */
+union udataword
+{
+ udata<32> word;
+ udata<8> bytes[4];
+ udata<4> half_bytes[4];
+ udata<1> bitfields[4];
+};
+
#endif // __AIDGE_EXPORT_CPP_NETWORK_TYPEDEFS__
--
GitLab