diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/Conv.hpp b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/Conv.hpp
index 450106b1dc1bd7e7ba32f1129213019c107fd91a..fc1b97e504aa4ebd9683b5269cebd48718c2d248 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/Conv.hpp
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/Conv.hpp
@@ -21,15 +21,14 @@
#ifndef __N2D2_EXPORT_CPP_CONV_HPP__
#define __N2D2_EXPORT_CPP_CONV_HPP__
-#include "typedefs.h"
-#include "assert.h"
-#include "utils.hpp"
+#include "kernels/typedefs.hpp"
+#include "kernels/utils.hpp"
#include "kernels/Macs.hpp"
namespace N2D2_Export {
-template<int NB_CHANNELS,
+template<int NB_CHANNELS,
int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
@@ -69,7 +68,7 @@ void convcellPropagate(const Input_T* __restrict inputs,
: max(PADDING_Y - (oy * STRIDE_Y), 0);
const int syMax = (PADDING_Y == 0
&& OUTPUTS_HEIGHT == OUTPUTS_HEIGHT_NOPAD) ? KERNEL_HEIGHT
- : clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
+ : clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
0, KERNEL_HEIGHT);
const int iy = (oy * STRIDE_Y) - PADDING_Y;
@@ -78,7 +77,7 @@ void convcellPropagate(const Input_T* __restrict inputs,
: max(PADDING_X - (ox * STRIDE_X), 0);
const int sxMax = (PADDING_X == 0
&& OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD) ? KERNEL_WIDTH
- : clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
+ : clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
0, KERNEL_WIDTH);
const int ix = (ox * STRIDE_X) - PADDING_X;
@@ -132,8 +131,8 @@ void convcellPropagate(const Input_T* __restrict inputs,
|| sxMax - sxMin == KERNEL_WIDTH)))
{
macsOnRange<KERNEL_WIDTH * NB_CHANNELS>(
- inputs + iOffset,
- weights + wOffset,
+ inputs + iOffset,
+ weights + wOffset,
weightedSum);
}
else {
@@ -158,8 +157,8 @@ void convcellPropagate(const Input_T* __restrict inputs,
macsOnRange<NB_CHANNELS>(
// same input line so no wrapping can occur
- inputs + iOffsetInRange,
- weights + wOffset + sx * NB_CHANNELS,
+ inputs + iOffsetInRange,
+ weights + wOffset + sx * NB_CHANNELS,
weightedSum);
}
}
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/FullyConnected/Fc.hpp b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/FullyConnected/Fc.hpp
index 9dce7efb5d2ac3b9ee287f82fbb51e46e00d5365..528a6f10ea70a8a9b2643ccbe6eec23301ff9c7b 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/FullyConnected/Fc.hpp
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/FullyConnected/Fc.hpp
@@ -21,15 +21,14 @@
#ifndef __N2D2_EXPORT_CPP_FC_HPP__
#define __N2D2_EXPORT_CPP_FC_HPP__
-#include "typedefs.h"
-#include "assert.h"
-#include "utils.hpp"
+#include "kernels/typedefs.hpp"
+#include "kernels/utils.hpp"
#include "kernels/Macs.hpp"
namespace N2D2_Export {
-template<int NB_CHANNELS,
+template<int NB_CHANNELS,
int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
@@ -84,8 +83,8 @@ void fccellPropagate(const Input_T* __restrict inputs,
if (!wrapInRange && INPUT_MEM_STRIDE == NB_CHANNELS) {
macsOnRange<NB_CHANNELS * CHANNELS_WIDTH>(
- inputs + iOffset,
- weights + wOffset,
+ inputs + iOffset,
+ weights + wOffset,
weightedSum);
}
else {
@@ -102,7 +101,7 @@ void fccellPropagate(const Input_T* __restrict inputs,
macsOnRange<NB_CHANNELS>(
inputs + iOffsetInRange,
- weights + wOffset + ix * NB_CHANNELS,
+ weights + wOffset + ix * NB_CHANNELS,
weightedSum);
}
}
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Pooling/Pooling.hpp b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Pooling/Pooling.hpp
index 15126719090027775dcd3c9204d9233232e4bedf..30ee9e316e755ebb44b2775627a3a2cdf66e13a5 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Pooling/Pooling.hpp
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Pooling/Pooling.hpp
@@ -23,13 +23,13 @@
#include <cassert>
#include <cstdio>
-#include "typedefs.h"
-#include "utils.hpp"
+#include "kernels/typedefs.hpp"
+#include "kernels/utils.hpp"
namespace N2D2_Export {
-template<int NB_CHANNELS,
+template<int NB_CHANNELS,
int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
@@ -74,7 +74,7 @@ void poolcellPropagate(const Input_T* __restrict inputs,
: max(PADDING_Y - (oy * STRIDE_Y), 0);
const int syMax = (PADDING_Y == 0
&& OUTPUTS_HEIGHT == OUTPUTS_HEIGHT_NOPAD) ? POOL_HEIGHT
- : clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
+ : clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
0, POOL_HEIGHT);
const int iy = (oy * STRIDE_Y) - PADDING_Y;
@@ -86,7 +86,7 @@ void poolcellPropagate(const Input_T* __restrict inputs,
const int sxMax = (PADDING_X == 0
&& OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD)
? POOL_WIDTH
- : clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
+ : clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
0, POOL_WIDTH);
const int ix = (ox * STRIDE_X) - PADDING_X;
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Utils/Macs.hpp b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Utils/Macs.hpp
index a0a1f85fe87c7a27ad685cdce9b49314d41e13bd..58a930536d770d4b4e43c316767efd5c8c102194 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Utils/Macs.hpp
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Utils/Macs.hpp
@@ -1,3262 +1,211 @@
-// /*
-// (C) Copyright 2017 CEA LIST. All Rights Reserved.
-// Contributor(s): N2D2 Team
+/*
+ (C) Copyright 2017 CEA LIST. All Rights Reserved.
+ Contributor(s): N2D2 Team
-// This software is governed by the CeCILL-C license under French law and
-// abiding by the rules of distribution of free software. You can use,
-// modify and/ or redistribute the software under the terms of the CeCILL-C
-// license as circulated by CEA, CNRS and INRIA at the following URL
-// "http://www.cecill.info".
+ This software is governed by the CeCILL-C license under French law and
+ abiding by the rules of distribution of free software. You can use,
+ modify and/ or redistribute the software under the terms of the CeCILL-C
+ license as circulated by CEA, CNRS and INRIA at the following URL
+ "http://www.cecill.info".
-// As a counterpart to the access to the source code and rights to copy,
-// modify and redistribute granted by the license, users are provided only
-// with a limited warranty and the software's author, the holder of the
-// economic rights, and the successive licensors have only limited
-// liability.
+ As a counterpart to the access to the source code and rights to copy,
+ modify and redistribute granted by the license, users are provided only
+ with a limited warranty and the software's author, the holder of the
+ economic rights, and the successive licensors have only limited
+ liability.
-// The fact that you are presently reading this means that you have had
-// knowledge of the CeCILL-C license and that you accept its terms.
-// */
+ The fact that you are presently reading this means that you have had
+ knowledge of the CeCILL-C license and that you accept its terms.
+*/
-// #ifndef __N2D2_EXPORT_CPP_MACS_HPP__
-// #define __N2D2_EXPORT_CPP_MACS_HPP__
+#ifndef __N2D2_EXPORT_CPP_MACS_HPP__
+#define __N2D2_EXPORT_CPP_MACS_HPP__
-// #include <cstdint>
-// #include <limits>
-// #include <type_traits>
-// #include <cmsis_compiler.h>
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+#include <cmsis_compiler.h>
-// #include "swar_arm_acle.h"
+namespace N2D2_Export {
-// namespace N2D2_Export {
-
-// template<typename Input_T>
-// inline static
-// uint32_t XTB16(uint32_t val)
-// {
-// return std::is_unsigned<Input_T>::value ? __UXTB16(val) : __SXTB16(val);
-// }
-
-// template<int INPUTS_INC = 1,
-// int WEIGHTS_INC = 1,
-// typename Input_T,
-// typename Weight_T,
-// typename Sum_T>
-// inline static
-// Sum_T dualMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// weightedSum += inputs[0] * weights[0]
-// + inputs[INPUTS_INC] * weights[WEIGHTS_INC];
-
-// 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::is_floating_point<Input_T>::value>::type* = nullptr>
-// inline 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]
-// + inputs[1*INPUTS_INC] * weights[1*WEIGHTS_INC]
-// + inputs[2*INPUTS_INC] * weights[2*WEIGHTS_INC]
-// + inputs[3*INPUTS_INC] * weights[3*WEIGHTS_INC];
-
-// 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::is_floating_point<Input_T>::value>::type* = nullptr>
-// inline static
-// Sum_T quadMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// if(INPUTS_INC != 1 || WEIGHTS_INC != 1) {
-// weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC]
-// + inputs[1*INPUTS_INC] * weights[1*WEIGHTS_INC]
-// + inputs[2*INPUTS_INC] * weights[2*WEIGHTS_INC]
-// + inputs[3*INPUTS_INC] * weights[3*WEIGHTS_INC];
-
-// return weightedSum;
-// }
-
-// // Inputs loading & preparation
-// uint32_t in;
-// memcpy((void*) &in, inputs, sizeof(in));
-
-// uint32_t in1 = XTB16<Input_T>(in);
-// uint32_t in2 = XTB16<Input_T>(in >> 8);
-
-// // Weights loading & preparation
-// uint32_t wt;
-// memcpy((void*) &wt, weights, sizeof(wt));
-
-// uint32_t wt1 = XTB16<Weight_T>(wt);
-// uint32_t wt2 = XTB16<Weight_T>(wt >> 8);
-
-// // Computation
-// if(std::is_same<Sum_T, int32_t>::value) {
-// weightedSum = __SMLAD(in1, wt1, weightedSum);
-// weightedSum = __SMLAD(in2, wt2, weightedSum);
-// }
-// else {
-// weightedSum = __SMLALD(in1, wt1, weightedSum);
-// weightedSum = __SMLALD(in2, wt2, weightedSum);
-
-// }
-
-// return weightedSum;
-// }
-
-
-
-
-// // ----------------------------------------------------------------------------
-// // -------------- MAC computing functions for kernel 4W-4A --------------------
-// // ----------------------------------------------------------------------------
-
-// /**
-// * @brief Unsigned mono mac operation (4W/4A version)
-// * @details Performs one mac operation for signed 4-bits weights
-// * and unsigned 4-bits inputs.
-// *
-// * @tparam Input_T Input type (should be udata<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to input vector
-// * @param[in] weights Pointer to kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the dual mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T monoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// weightedSum += __UBFX(inputs[0], 4, 4) * __SBFX(weights[0], 4, 4);
-// return weightedSum;
-// }
-
-// /**
-// * @brief Signed mono mac operation (4W/4A version)
-// * @details Performs one mac operation for signed 4-bits weights
-// * and signed 4-bits inputs.
-// *
-// * @tparam Input_T Input type (should be data<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to input vector
-// * @param[in] weights Pointer to kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the dual mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(!std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T monoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// weightedSum += __SBFX(inputs[0], 4, 4) * __SBFX(weights[0], 4, 4);
-// return weightedSum;
-// }
-
-// /**
-// * @brief Unsigned dual mac operation (4W/4A version)
-// * @details Performs two mac operations for signed 4-bits weights
-// * and unsigned 4-bits inputs. Extracts the two 4-bits weights
-// * from a stored 8-bits weight and associates them into
-// * a 32-bits value. Then extracts the two 4-bits inputs
-// * from a stored 8-bits input and associates them into
-// * a 32-bits value. Finally performs a dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be udata<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the dual mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T dualMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint8_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// uint32_t wght = __BFI(w1, w0, 16, 16);
-
-// uint8_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// int32_t a0 = __UBFX(in, 0, 4);
-// int32_t a1 = __UBFX(in, 4, 4);
-// uint32_t act = __BFI(a1, a0, 16, 16);
-
-// weightedSum = __SMLAD(act, wght, weightedSum);
-
-// return weightedSum;
-// }
-
-// /**
-// * @brief Signed dual mac operation (4W/4A version)
-// * @details Performs two mac operations for signed 4-bits weights
-// * and signed 4-bits inputs. Extracts the two 4-bits weights
-// * from a stored 8-bits weight and associates them into
-// * a 32-bits value. Then extracts the two 4-bits inputs
-// * from a stored 8-bits input and associates them into
-// * a 32-bits value. Finally performs a dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be data<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the dual mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(!std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T dualMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint8_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// uint32_t wght = __BFI(w1, w0, 16, 16);
-
-// uint8_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// int32_t a0 = __SBFX(in, 0, 4);
-// int32_t a1 = __SBFX(in, 4, 4);
-// uint32_t act = __BFI(a1, a0, 16, 16);
-
-// weightedSum = __SMLAD(act, wght, weightedSum);
-
-// return weightedSum;
-// }
-
-// /**
-// * @brief Unsigned quad mac operation (4W/4A version)
-// * @details Performs four mac operations for signed 4-bits weights
-// * and unsigned 4-bits inputs. Extracts the four 4-bits weights
-// * from two stored 8-bits weights and associates them into
-// * two 32-bits values. Then extracts the four 4-bits inputs
-// * from two stored 8-bits inputs and associates them into
-// * two 32-bits values. Finally performs a double dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be udata<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the quad mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T quadMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint16_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// int32_t w2 = __SBFX(wt, 8, 4);
-// int32_t w3 = __SBFX(wt, 12, 4);
-
-// uint32_t evenW1 = __BFI(w2, w0, 16, 16);
-// uint32_t oddW1 = __BFI(w3, w1, 16, 16);
-
-// uint16_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// int32_t a0 = __UBFX(in, 0, 4);
-// int32_t a1 = __UBFX(in, 4, 4);
-// int32_t a2 = __UBFX(in, 8, 4);
-// int32_t a3 = __UBFX(in, 12, 4);
-
-// uint32_t evenA1 = __BFI(a2, a0, 16, 16);
-// uint32_t oddA1 = __BFI(a3, a1, 16, 16);
-
-// weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
-// weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
-
-// return weightedSum;
-// }
-
-// /**
-// * @brief Signed quad mac operation (4W/4A version)
-// * @details Performs four mac operations for signed 4-bits weights
-// * and signed 4-bits inputs. Extracts the four 4-bits weights
-// * from two stored 8-bits weights and associates them into
-// * two 32-bits values. Then extracts the four 4-bits inputs
-// * from two stored 8-bits inputs and associates them into
-// * two 32-bits values. Finally performs a double dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be data<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the quad mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(!std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T quadMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint16_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// int32_t w2 = __SBFX(wt, 8, 4);
-// int32_t w3 = __SBFX(wt, 12, 4);
-
-// uint32_t evenW1 = __PKHBT(w2, w0, 16);
-// uint32_t oddW1 = __PKHBT(w3, w1, 16);
-
-// uint16_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// int32_t a0 = __SBFX(in, 0, 4);
-// int32_t a1 = __SBFX(in, 4, 4);
-// int32_t a2 = __SBFX(in, 8, 4);
-// int32_t a3 = __SBFX(in, 12, 4);
-
-// uint32_t evenA1 = __PKHBT(a2, a0, 16);
-// uint32_t oddA1 = __PKHBT(a3, a1, 16);
-
-// weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
-// weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
-
-// return weightedSum;
-// }
-
-// /**
-// * @brief Unsigned octo mac operation (4W/4A version)
-// * @details Performs eight mac operations for signed 4-bits weights
-// * and unsigned 4-bits inputs. Extracts the eight 4-bits weights
-// * from four stored 8-bits weights and associates them into
-// * four 32-bits values. Then extracts the eight 4-bits inputs
-// * from four stored 8-bits inputs and associates them into
-// * four 32-bits values. Finally performs a quadruple dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be udata<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the octo mac operation
-// */
-// // template<typename Input_T, typename Weight_T,
-// // typename std::enable_if<(std::is_unsigned<Input_T>::value
-// // && std::numeric_limits<Weight_T>::digits == 4
-// // && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// // __attribute__((always_inline)) static inline
-// // Sum_T octoMac(const Input_T* __restrict inputs,
-// // const Weight_T* __restrict weights,
-// // Sum_T weightedSum)
-// // {
-// // uint32_t wt;
-// // std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// // int32_t w0 = __SBFX(wt, 0, 4);
-// // int32_t w1 = __SBFX(wt, 4, 4);
-// // int32_t w2 = __SBFX(wt, 8, 4);
-// // int32_t w3 = __SBFX(wt, 12, 4);
-// // int32_t w4 = __SBFX(wt, 16, 4);
-// // int32_t w5 = __SBFX(wt, 20, 4);
-// // int32_t w6 = __SBFX(wt, 24, 4);
-// // int32_t w7 = __SBFX(wt, 28, 4);
-
-// // // uint32_t weight0 = __BFI(w4, w0, 16, 16);
-// // // uint32_t weight1 = __BFI(w5, w1, 16, 16);
-// // // uint32_t weight2 = __BFI(w6, w2, 16, 16);
-// // // uint32_t weight3 = __BFI(w7, w3, 16, 16);
-
-// // uint32_t weight0 = __PKHBT(w0, w4, 16);
-// // uint32_t weight1 = __PKHBT(w1, w5, 16);
-// // uint32_t weight2 = __PKHBT(w2, w6, 16);
-// // uint32_t weight3 = __PKHBT(w3, w7, 16);
-
-// // uint32_t in;
-// // std::memcpy((void*) &in, inputs, sizeof(in));
-
-// // uint32_t act0 = in & 0xF000F;
-// // uint32_t act1 = (in >> 4) & 0xF000F;
-// // uint32_t act2 = (in >> 8) & 0xF000F;
-// // uint32_t act3 = (in >> 12) & 0xF000F;
-
-// // weightedSum = __SMLAD(act0, weight0, weightedSum);
-// // weightedSum = __SMLAD(act1, weight1, weightedSum);
-// // weightedSum = __SMLAD(act2, weight2, weightedSum);
-// // weightedSum = __SMLAD(act3, weight3, weightedSum);
-
-// // return weightedSum;
-// // }
-
-// // template<typename Input_T, typename Weight_T,
-// // typename std::enable_if<(std::is_unsigned<Input_T>::value
-// // && std::numeric_limits<Weight_T>::digits == 4
-// // && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// // __attribute__((always_inline)) static inline
-// // Sum_T octoMac(const Input_T* __restrict inputs,
-// // const Weight_T* __restrict weights,
-// // Sum_T weightedSum)
-// // {
-// // union n2d2_dataword wt;
-// // std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// // union n2d2_udataword in;
-// // std::memcpy((void*) &in, inputs, sizeof(in));
-
-// // for (int i = 0; i < 4; ++i) {
-// // weightedSum += (data<32>)(in.half_bytes[i].fields.op0) * wt.half_bytes[i].fields.op0;
-// // weightedSum += (data<32>)(in.half_bytes[i].fields.op1) * wt.half_bytes[i].fields.op1;
-// // }
-
-// // // weightedSum += (data<32>)(in.half_bytes[0].fields.op0) * wt.half_bytes[0].fields.op0;
-// // // weightedSum += (data<32>)(in.half_bytes[0].fields.op1) * wt.half_bytes[0].fields.op1;
-// // // weightedSum += (data<32>)(in.half_bytes[1].fields.op0) * wt.half_bytes[1].fields.op0;
-// // // weightedSum += (data<32>)(in.half_bytes[1].fields.op1) * wt.half_bytes[1].fields.op1;
-// // // weightedSum += (data<32>)(in.half_bytes[2].fields.op0) * wt.half_bytes[2].fields.op0;
-// // // weightedSum += (data<32>)(in.half_bytes[2].fields.op1) * wt.half_bytes[2].fields.op1;
-// // // weightedSum += (data<32>)(in.half_bytes[3].fields.op0) * wt.half_bytes[3].fields.op0;
-// // // weightedSum += (data<32>)(in.half_bytes[3].fields.op1) * wt.half_bytes[3].fields.op1;
-
-// // return weightedSum;
-// // }
-
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T octoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint32_t wt;
-// memcpy((void*) &wt, weights, sizeof(wt));
-
-// // Works with weights * 4096 (weights << 12)
-// const uint32_t WeightMask = 0xF000F000;
-// uint32_t weight0 = WeightMask & (wt << 12);
-// uint32_t weight1 = WeightMask & (wt << 8);
-// uint32_t weight2 = WeightMask & (wt << 4);
-// uint32_t weight3 = WeightMask & (wt);
-
-// uint32_t in;
-// memcpy((void*) &in, inputs, sizeof(in));
-
-// const uint32_t ActMask = 0x000F000F; // to explicit instructions
-// uint32_t act0 = in & ActMask;
-// // Expect second operand shift
-// uint32_t act1 = ActMask & (in >> 4);
-// uint32_t act2 = ActMask & (in >> 8);
-// uint32_t act3 = ActMask & (in >> 12);
-
-// Sum_T sum = 0;
-// sum = __SMLAD(act0, weight0, sum);
-// sum = __SMLAD(act1, weight1, sum);
-// sum = __SMLAD(act2, weight2, sum);
-// sum = __SMLAD(act3, weight3, sum);
-
-// return weightedSum + (sum >> 12);
-// }
-
-// /**
-// * @brief Signed octo mac operation (4W/4A version)
-// * @details Performs eight mac operations for signed 4-bits weights
-// * and signed 4-bits inputs. Extracts the eight 4-bits weights
-// * from four stored 8-bits weights and associates them into
-// * four 32-bits values. Then extracts the eight 4-bits inputs
-// * from four stored 8-bits inputs and associates them into
-// * four 32-bits values. Finally performs a quadruple dual mac operation
-// * with the __SMLAD instruction
-// *
-// * @tparam Input_T Input type (should be data<4>)
-// * @tparam Weight_T Weight type (should be data<4>)
-// *
-// * @param[in] inputs Pointer to compressed input vector
-// * @param[in] weights Pointer to compressed kernel weights
-// * @param[in,out] weightedSum Accumulating sum from the
-// * previous mac operations
-// * @returns Updated weightedSum with
-// * the result of the octo mac operation
-// */
-// template<typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(!std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T octoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint32_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// int32_t w2 = __SBFX(wt, 8, 4);
-// int32_t w3 = __SBFX(wt, 12, 4);
-// int32_t w4 = __SBFX(wt, 16, 4);
-// int32_t w5 = __SBFX(wt, 20, 4);
-// int32_t w6 = __SBFX(wt, 24, 4);
-// int32_t w7 = __SBFX(wt, 28, 4);
-
-// uint32_t evenW1 = __PKHBT(w2, w0, 16);
-// uint32_t oddW1 = __PKHBT(w3, w1, 16);
-// uint32_t evenW2 = __PKHBT(w6, w4, 16);
-// uint32_t oddW2 = __PKHBT(w7, w5, 16);
-
-// uint32_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// int32_t a0 = __SBFX(in, 0, 4);
-// int32_t a1 = __SBFX(in, 4, 4);
-// int32_t a2 = __SBFX(in, 8, 4);
-// int32_t a3 = __SBFX(in, 12, 4);
-// int32_t a4 = __SBFX(in, 16, 4);
-// int32_t a5 = __SBFX(in, 20, 4);
-// int32_t a6 = __SBFX(in, 24, 4);
-// int32_t a7 = __SBFX(in, 28, 4);
-
-// uint32_t evenA1 = __PKHBT(a2, a0, 16);
-// uint32_t oddA1 = __PKHBT(a3, a1, 16);
-// uint32_t evenA2 = __PKHBT(a6, a4, 16);
-// uint32_t oddA2 = __PKHBT(a7, a5, 16);
-
-// weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
-// weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
-// weightedSum = __SMLAD(evenA2, evenW2, weightedSum);
-// weightedSum = __SMLAD(oddA2, oddW2, weightedSum);
-
-// return weightedSum;
-// }
-
-
-// // template<typename Input_T, typename Weight_T, typename Sum_T,
-// // typename std::enable_if<(std::is_unsigned<Input_T>::value
-// // && std::numeric_limits<Weight_T>::digits == 4
-// // && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// // void macsOnParallel(const Input_T* __restrict inputs,
-// // const Weight_T* __restrict weights,
-// // Sum_T* weightedSums,
-// // const int nb_data)
-// // {
-// // uint32_t wt = 0;
-// // std::memcpy((void*) &wt, weights, ceil((double)nb_data/2));
-
-// // uint32_t in = 0;
-// // std::memcpy((void*) &in, inputs, ceil((double)nb_data/2));
-
-// // for (int i = 0; i < nb_data; ++i) {
-// // weightedSums[i] += __SBFX(wt, 4*i, 4) * __UBFX(in, 4*i, 4);
-// // }
-// // }
-
-// // template<typename Input_T, typename Weight_T, typename Sum_T,
-// // typename std::enable_if<(!std::is_unsigned<Input_T>::value
-// // && std::numeric_limits<Weight_T>::digits == 4
-// // && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// // void macsOnParallel(const Input_T* __restrict inputs,
-// // const Weight_T* __restrict weights,
-// // Sum_T* weightedSums,
-// // const int nb_data)
-// // {
-// // uint32_t wt = 0;
-// // std::memcpy((void*) &wt, weights, ceil((double)nb_data/2));
-
-// // uint32_t in = 0;
-// // std::memcpy((void*) &in, inputs, ceil((double)nb_data/2));
-
-// // for (int i = 0; i < nb_data; ++i) {
-// // weightedSums[i] += __SBFX(wt, 4*i, 4) * __SBFX(in, 4*i, 4);
-// // }
-// // }
-
-
-
-
-// // **************************************************************************
-// // * Multiply-accumulate the values in inputs and weights for NB_ITERATIONS *
-// // **************************************************************************
-
-// template<int NB_ITERATIONS,
-// int INPUTS_INC = 1,
-// int WEIGHTS_INC = 1,
-// class Input_T,
-// class Weight_T,
-// class Sum_T,
-// typename std::enable_if<(NB_ITERATIONS == 0)>::type* = nullptr>
-// inline static
-// void macsOnRange(const Input_T* __restrict /*inputs*/,
-// const Weight_T* __restrict /*weights*/,
-// Sum_T& __restrict /*weightedSum*/)
-// {
-// // Nothing to do
-// }
-
-// template<int NB_ITERATIONS,
-// int INPUTS_INC = 1,
-// int WEIGHTS_INC = 1,
-// class Input_T,
-// class Weight_T,
-// class Sum_T,
-// typename std::enable_if<(NB_ITERATIONS == 1)>::type* = nullptr>
-// inline static
-// void macsOnRange(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T& __restrict weightedSum)
-// {
-// weightedSum += (*weights) * (*inputs);
-// }
-
-// template<int NB_ITERATIONS,
-// int INPUTS_INC = 1,
-// int WEIGHTS_INC = 1,
-// class Input_T,
-// class Weight_T,
-// class Sum_T,
-// typename std::enable_if<(NB_ITERATIONS >= 2 && NB_ITERATIONS < 4)>::type* = nullptr>
-// inline 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 + 2*WEIGHTS_INC,
-// weightedSum);
-// }
-
-// /**
-// * @brief MACs Processing
-// * @details Performs NB_ITERATIONS MACs operations, storing results into the
-// * weightedSum variable.
-// *
-// * @tparam NB_ITERATIONS Number of MACs to perform
-// * @tparam INPUTS_INC Input Stride
-// * @tparam WEIGHTS_INC Weights Stride
-// * @tparam Input_T Input Type
-// *
-// * @param inputs Pointer to inputs vector
-// * @param weights Pointer to weights vector
-// * @param weightedSum Pointer to weightedSum
-// */
-// template<int NB_ITERATIONS,
-// int INPUTS_INC = 1,
-// int WEIGHTS_INC = 1,
-// class Input_T,
-// class Weight_T,
-// class Sum_T,
-// typename std::enable_if<(NB_ITERATIONS >= 4)>::type* = nullptr>
-// inline 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 + 4*WEIGHTS_INC,
-// weightedSum);
-// }
-
-
-// template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(NB_ITERATIONS >= 2 && NB_ITERATIONS < 4 && std::numeric_limits<Weight_T>::digits > 1)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// void macsOnRange(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T& weightedSum)
-// {
-// constexpr unsigned int idxI
-// = (std::numeric_limits<Input_T>::digits > 4) ? 2 : 1;
-// constexpr unsigned int idxW
-// = (std::numeric_limits<Weight_T>::digits > 4) ? 2 : 1;
-
-// weightedSum = dualMac(inputs, weights, weightedSum);
-// macsOnRange<NB_ITERATIONS - 2>(inputs + idxI, weights + idxW, weightedSum);
-// }
-
-// template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<NB_ITERATIONS >= 4
-// && (std::numeric_limits<Weight_T>::digits > 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// void macsOnRange(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T& weightedSum)
-// {
-// constexpr unsigned int idxI
-// = (std::numeric_limits<Input_T>::digits > 4)
-// ? 4 : (std::numeric_limits<Input_T>::digits == 4) ? 2 : 1;
-
-// constexpr unsigned int idxW = 4;
-
-// weightedSum = quadMac(inputs, weights, weightedSum);
-// macsOnRange<NB_ITERATIONS - 4>(inputs + idxI, weights + idxW, weightedSum);
-// }
-
-// template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<(NB_ITERATIONS >= 4 && NB_ITERATIONS < 8)
-// && (std::numeric_limits<Weight_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// void macsOnRange(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T& weightedSum)
-// {
-// constexpr unsigned int idxI
-// = (std::numeric_limits<Input_T>::digits > 4)
-// ? 4 : (std::numeric_limits<Input_T>::digits == 4) ? 2 : 1;
-
-// constexpr unsigned int idxW = 2;
-
-// weightedSum = quadMac(inputs, weights, weightedSum);
-// macsOnRange<NB_ITERATIONS - 4>(inputs + idxI, weights + idxW, weightedSum);
-// }
-
-// template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
-// typename std::enable_if<NB_ITERATIONS >= 8
-// && (std::numeric_limits<Weight_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// void macsOnRange(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T& weightedSum)
-// {
-// constexpr unsigned int idxI
-// = (std::numeric_limits<Input_T>::digits > 4)
-// ? 8 : (std::numeric_limits<Input_T>::digits == 4)
-// ? 4 : (std::numeric_limits<Input_T>::digits == 2)
-// ? 2 : 1;
-
-// constexpr unsigned int idxW = 4;
-
-// weightedSum = octoMac(inputs, weights, weightedSum);
-// macsOnRange<NB_ITERATIONS - 8>(inputs + idxI, weights + idxW, weightedSum);
-// }
-
-
-// } // N2D2_Export
-
-// #endif // __N2D2_EXPORT_CPP_MACS_HPP__
-
-
-
-
-/**
- ******************************************************************************
- * @file mac_functions.hpp
- * @brief Mac operation functions for ARM Cortex m7 and m4
- * This file provides different functions to perform
- * signed and unsigned mac operations. Those functions can calculate
- * up to eight mac operations at once.
- * The file also provides two general mac operation which can be
- * used in other files, especially in Network.hpp
- *
- ******************************************************************************
- * @attention
- *
- * (C) Copyright 2021 CEA LIST. All Rights Reserved.
- * Contributor(s): Vincent TEMPLIER (vincent.templier@cea.fr)
- * Philippe DORE (philippe.dore@cea.fr)
- * David BRIAND (david.briand@cea.fr)
- *
- * This file is not part of the open source version of N2D2 and is NOT under
- * the CeCILL-C license. This code is the property of the CEA. It can not be
- * copied or disseminated without its authorization.
- *
- ******************************************************************************
- */
-
-#ifndef __N2D2_MAC_FUNCTIONS_HPP__
-#define __N2D2_MAC_FUNCTIONS_HPP__
-
-#include <cstring>
-#include "swar_arm_acle.h"
-#include "kernels/typedefs.hpp"
-
-
-// ----------------------------------------------------------------------------
-// --------------- MAC computing functions for all kernels --------------------
-// ----------------------------------------------------------------------------
-
-
-// ----------------------------------------------------------------------------
-// -------------- MAC computing functions for kernel 8W-8A --------------------
-// ----------------------------------------------------------------------------
-
-/**
- * @brief Mono mac operation (8W/8A version)
- * @details Performs one mac operation for signed 8-bits weights
- * and 8-bits inputs (signed or not).
- *
- * @tparam Input_T Input type (udata<8> or data<8>)
- * @tparam Weight_T Weight type (should be data<8>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<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 == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T monoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += (Sum_T)inputs[0] * weights[0];
- return weightedSum;
-}
-
-/**
- * @brief Dual mac operation (8W/8A version)
- * @details Performs two mac operations for signed 8-bits weights
- * and 8-bits inputs (signed or not).
- *
- * @tparam Input_T Input type (udata<8> or data<8>)
- * @tparam Weight_T Weight type (should be data<8>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<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 == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T dualMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += (Sum_T)inputs[0] * weights[0] + (Sum_T)inputs[1] * weights[1];
- return weightedSum;
-}
-
-/**
- * @brief Unsigned quad mac operation (8W/8A version)
- * @details Performs four mac operations for signed 8-bits weights
- * and unsigned 8-bits inputs. Sign extends four 8-bits weights
- * and associates them into two 32-bits values. Then zero extends
- * four 8-bits inputs and associates them into two 32-bits values.
- * Finally performs a double dual mac operation
- * with the __SMLAD instruction.
- *
- * @tparam Input_T Input type (should be udata<8>)
- * @tparam Weight_T Weight type (should be data<8>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 8
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t in1 = __UXTB16(in);
- uint32_t in2 = __UXTB16_RORn(in, 8);
-
- uint32_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- uint32_t wt1 = __SXTB16(wt);
- uint32_t wt2 = __SXTB16_RORn(wt, 8);
-
- weightedSum = __SMLAD(in1, wt1, weightedSum);
- weightedSum = __SMLAD(in2, wt2, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Signed quad mac operation (8W/8A version)
- * @details Performs four mac operations for signed 8-bits weights
- * and signed 8-bits inputs. Sign extends four 8-bits weights
- * and associates them into two 32-bits values. Then sign extends
- * four 8-bits inputs and associates them into two 32-bits values.
- * Finally performs a double dual mac operation
- * with the __SMLAD instruction.
- *
- * @tparam Input_T Input type (should be data<8>)
- * @tparam Weight_T Weight type (should be data<8>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 8
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t in1 = __SXTB16(in);
- uint32_t in2 = __SXTB16_RORn(in, 8);
-
- uint32_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- uint32_t wt1 = __SXTB16(wt);
- uint32_t wt2 = __SXTB16_RORn(wt, 8);
-
- weightedSum = __SMLAD(in1, wt1, weightedSum);
- weightedSum = __SMLAD(in2, wt2, weightedSum);
-
- return weightedSum;
-}
-
-template<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 == 8)>::type* = nullptr>
-void macsOnParallel(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T* weightedSums,
- const int nb_data)
-{
- union n2d2_dataword wt = {0};
- std::memcpy((void*) &wt, weights, nb_data);
-
- typename std::conditional<(!std::is_unsigned<Input_T>::value),
- union n2d2_dataword, union n2d2_udataword>::type in = {0};
- std::memcpy((void*) &in, inputs, nb_data);
-
- for (int i = 0; i < nb_data; ++i) {
- weightedSums[i] += (Sum_T)wt.bytes[i] * in.bytes[i];
- }
-}
-
-
-
-// ----------------------------------------------------------------------------
-// -------------- MAC computing functions for kernel 4W-8A --------------------
-// ----------------------------------------------------------------------------
-
-/**
- * @brief Mono mac operation (4W/8A version)
- * @details Performs one mac operation for signed 4-bits weights
- * and 8-bits inputs (signed or not).
- *
- * @tparam Input_T Input type (udata<8> or data<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T monoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += (Sum_T)inputs[0] * __SBFX(weights[0], 4, 4);
- return weightedSum;
-}
-
-/**
- * @brief Unsigned dual mac operation (4W/8A version)
- * @details Performs two mac operations for signed 4-bits weights
- * and unsigned 8-bits inputs. Extracts the two 4-bits weights
- * from a stored 8-bits weight and associates them into
- * a 32-bits value. Then zero extends two 8-bits inputs and
- * associates them into a 32-bits value. Finally performs a
- * dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T dualMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint8_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- uint32_t wght = __BFI(w0, w1, 16, 16);
-
- uint16_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t act = ((in << 8) | in);
- act = __UXTB16(act);
-
- weightedSum = __SMLAD(act, wght, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Signed dual mac operation (4W/8A version)
- * @details Performs two mac operations for signed 4-bits weights
- * and signed 8-bits inputs. Extracts the two 4-bits weights
- * from a stored 8-bits weight and associates them into
- * a 32-bits value. Then sign extends two 8-bits inputs and
- * associates them into a 32-bits value. Finally performs a
- * dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T dualMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint8_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- uint32_t wght = __BFI(w0, w1, 16, 16);
-
- uint16_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t act = ((in << 8) | in);
- act = __SXTB16(act);
-
- weightedSum = __SMLAD(act, wght, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Unsigned quad mac operation (4W/8A version)
- * @details Performs four mac operations for signed 4-bits weights
- * and unsigned 8-bits inputs. Extracts the four 4-bits weights
- * from two stored 8-bits weights and associates them into
- * two 32-bits values. Then zero extends four 8-bits inputs and
- * associates them into two 32-bits values. Finally performs a
- * double dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint16_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
-
- uint32_t evenW1 = __PKHBT(w0, w2, 16);
- uint32_t oddW1 = __PKHBT(w1, w3, 16);
-
- uint32_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t evenA1 = __UXTB16(in);
- uint32_t oddA1 = __UXTB16_RORn(in, 8);
-
- weightedSum = __SMLAD(evenA1, oddW1, weightedSum);
- weightedSum = __SMLAD(oddA1, evenW1, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Signed quad mac operation (4W/8A version)
- * @details Performs four mac operations for signed 4-bits weights
- * and signed 8-bits inputs. Extracts the four 4-bits weights
- * from two stored 8-bits weights and associates them into
- * two 32-bits values. Then sign extends four 8-bits inputs and
- * associates them into two 32-bits values. Finally performs a
- * double dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint16_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
-
- uint32_t evenW1 = __BFI(w2, w0, 16, 16);
- uint32_t oddW1 = __BFI(w3, w1, 16, 16);
-
- uint32_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- uint32_t evenA1 = __SXTB16(in);
- uint32_t oddA1 = __SXTB16_RORn(in, 8);
-
- weightedSum = __SMLAD(evenA1, oddW1, weightedSum);
- weightedSum = __SMLAD(oddA1, evenW1, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Unsigned octo mac operation (4W/8A version)
- * @details Performs eight mac operations for signed 4-bits weights
- * and unsigned 8-bits inputs. Extracts the eight 4-bits weights
- * from four stored 8-bits weights and associates them into
- * four 32-bits values. Then zero extends eights 8-bits inputs and
- * associates them into four 32-bits values. Finally performs a
- * quadruple dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the octo mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- // uint32_t wt;
- // std::memcpy((void*) &wt, weights, sizeof(wt));
-
- // int32_t w0 = __SBFX(wt, 0, 4);
- // int32_t w1 = __SBFX(wt, 4, 4);
- // int32_t w2 = __SBFX(wt, 8, 4);
- // int32_t w3 = __SBFX(wt, 12, 4);
- // int32_t w4 = __SBFX(wt, 16, 4);
- // int32_t w5 = __SBFX(wt, 20, 4);
- // int32_t w6 = __SBFX(wt, 24, 4);
- // int32_t w7 = __SBFX(wt, 28, 4);
-
- // // uint32_t evenW1 = __BFI(w2, w0, 16, 16);
- // // uint32_t oddW1 = __BFI(w3, w1, 16, 16);
- // // uint32_t evenW2 = __BFI(w6, w4, 16, 16);
- // // uint32_t oddW2 = __BFI(w7, w5, 16, 16);
-
- // uint32_t evenW1 = __PKHBT(w0, w2, 16);
- // uint32_t oddW1 = __PKHBT(w1, w3, 16);
- // uint32_t evenW2 = __PKHBT(w4, w6, 16);
- // uint32_t oddW2 = __PKHBT(w5, w7, 16);
-
- // uint32_t in1, in2;
- // std::memcpy((void*) &in1, inputs, sizeof(in1));
- // std::memcpy((void*) &in2, (inputs + 4), sizeof(in2));
-
- // uint32_t evenA1 = __UXTB16(in1);
- // uint32_t oddA1 = __UXTB16_RORn(in1, 8);
- // uint32_t evenA2 = __UXTB16(in2);
- // uint32_t oddA2 = __UXTB16_RORn(in2, 8);
-
- // weightedSum = __SMLAD(evenA1, oddW1, weightedSum);
- // weightedSum = __SMLAD(oddA1, evenW1, weightedSum);
- // weightedSum = __SMLAD(evenA2, oddW2, weightedSum);
- // weightedSum = __SMLAD(oddA2, evenW2, weightedSum);
-
- // 2nd implementation
- // union n2d2_dataword wt;
- // std::memcpy((void*) &wt, weights, sizeof(wt));
-
- // union n2d2_udataword in1, in2;
- // std::memcpy((void*) &in1, inputs, sizeof(in1));
- // std::memcpy((void*) &in2, inputs + 4, sizeof(in2));
-
- // weightedSum += (data<32>)(in1.bytes[0]) * wt.half_bytes[0].fields.op1;
- // weightedSum += (data<32>)(in1.bytes[1]) * wt.half_bytes[0].fields.op0;
- // weightedSum += (data<32>)(in1.bytes[2]) * wt.half_bytes[1].fields.op1;
- // weightedSum += (data<32>)(in1.bytes[3]) * wt.half_bytes[1].fields.op0;
- // weightedSum += (data<32>)(in2.bytes[0]) * wt.half_bytes[2].fields.op1;
- // weightedSum += (data<32>)(in2.bytes[1]) * wt.half_bytes[2].fields.op0;
- // weightedSum += (data<32>)(in2.bytes[2]) * wt.half_bytes[3].fields.op1;
- // weightedSum += (data<32>)(in2.bytes[3]) * wt.half_bytes[3].fields.op0;
-
- uint32_t wt;
- memcpy((void*) &wt, weights, sizeof(wt));
-
- // Works with weights * 4096 (weights << 12)
- const uint32_t WeightMask = 0xF000F000;
- uint32_t weight0 = WeightMask & (wt << 12);
- uint32_t weight1 = WeightMask & (wt << 8);
- uint32_t weight2 = WeightMask & (wt << 4);
- uint32_t weight3 = WeightMask & (wt);
-
- uint32_t in1, in2;
- std::memcpy((void*) &in1, inputs, sizeof(in1));
- std::memcpy((void*) &in2, (inputs + 4), sizeof(in2));
-
- uint32_t in_a = __PKHBT(in1, in2, 16);
- uint32_t in_b = __PKHTB(in2, in1, 16);
-
- uint32_t evenA1 = __UXTB16(in_a);
- uint32_t oddA1 = __UXTB16_RORn(in_a, 8);
- uint32_t evenA2 = __UXTB16(in_b);
- uint32_t oddA2 = __UXTB16_RORn(in_b, 8);
-
- Sum_T sum = 0;
- sum = __SMLAD(oddA1, weight0, sum);
- sum = __SMLAD(evenA1, weight1, sum);
- sum = __SMLAD(oddA2, weight2, sum);
- sum = __SMLAD(evenA2, weight3, sum);
- weightedSum += sum >> 12;
-
- return weightedSum;
-}
-
-/**
- * @brief Signed octo mac operation (4W/8A version)
- * @details Performs eight mac operations for signed 4-bits weights
- * and signed 8-bits inputs. Extracts the eight 4-bits weights
- * from four stored 8-bits weights and associates them into
- * four 32-bits values. Then sign extends eights 8-bits inputs and
- * associates them into four 32-bits values. Finally performs a
- * quadruple dual mac operation with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<8>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the octo mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 8)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
- int32_t w4 = __SBFX(wt, 16, 4);
- int32_t w5 = __SBFX(wt, 20, 4);
- int32_t w6 = __SBFX(wt, 24, 4);
- int32_t w7 = __SBFX(wt, 28, 4);
-
- uint32_t evenW1 = __BFI(w2, w0, 16, 16);
- uint32_t oddW1 = __BFI(w3, w1, 16, 16);
- uint32_t evenW2 = __BFI(w6, w4, 16, 16);
- uint32_t oddW2 = __BFI(w7, w5, 16, 16);
-
- uint32_t in1, in2;
- std::memcpy((void*) &in1, inputs, sizeof(in1));
- std::memcpy((void*) &in2, (inputs + 4), sizeof(in2));
-
- uint32_t evenA1 = __SXTB16(in1);
- uint32_t oddA1 = __SXTB16_RORn(in1, 8);
- uint32_t evenA2 = __SXTB16(in2);
- uint32_t oddA2 = __SXTB16_RORn(in2, 8);
-
- weightedSum = __SMLAD(evenA1, oddW1, weightedSum);
- weightedSum = __SMLAD(oddA1, evenW1, weightedSum);
- weightedSum = __SMLAD(evenA2, oddW2, weightedSum);
- weightedSum = __SMLAD(oddA2, evenW2, weightedSum);
-
- return weightedSum;
-}
-
-
-template<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>
-void macsOnParallel(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T* weightedSums,
- const int nb_data)
-{
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, ceil((double)nb_data/2));
-
- for (int i = 0; i < nb_data; ++i) {
- weightedSums[i] += __SBFX(wt, 4*i, 4) * inputs[i];
- }
-}
-
-
-// ----------------------------------------------------------------------------
-// -------------- MAC computing functions for kernel 4W-4A --------------------
-// ----------------------------------------------------------------------------
-
-/**
- * @brief Unsigned mono mac operation (4W/4A version)
- * @details Performs one mac operation for signed 4-bits weights
- * and unsigned 4-bits inputs.
- *
- * @tparam Input_T Input type (should be udata<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T monoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += __UBFX(inputs[0], 4, 4) * __SBFX(weights[0], 4, 4);
- return weightedSum;
-}
-
-/**
- * @brief Signed mono mac operation (4W/4A version)
- * @details Performs one mac operation for signed 4-bits weights
- * and signed 4-bits inputs.
- *
- * @tparam Input_T Input type (should be data<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to input vector
- * @param[in] weights Pointer to kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T monoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += __SBFX(inputs[0], 4, 4) * __SBFX(weights[0], 4, 4);
- return weightedSum;
-}
-
-/**
- * @brief Unsigned dual mac operation (4W/4A version)
- * @details Performs two mac operations for signed 4-bits weights
- * and unsigned 4-bits inputs. Extracts the two 4-bits weights
- * from a stored 8-bits weight and associates them into
- * a 32-bits value. Then extracts the two 4-bits inputs
- * from a stored 8-bits input and associates them into
- * a 32-bits value. Finally performs a dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T dualMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint8_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- uint32_t wght = __BFI(w1, w0, 16, 16);
-
- uint8_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- int32_t a0 = __UBFX(in, 0, 4);
- int32_t a1 = __UBFX(in, 4, 4);
- uint32_t act = __BFI(a1, a0, 16, 16);
-
- weightedSum = __SMLAD(act, wght, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Signed dual mac operation (4W/4A version)
- * @details Performs two mac operations for signed 4-bits weights
- * and signed 4-bits inputs. Extracts the two 4-bits weights
- * from a stored 8-bits weight and associates them into
- * a 32-bits value. Then extracts the two 4-bits inputs
- * from a stored 8-bits input and associates them into
- * a 32-bits value. Finally performs a dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the dual mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T dualMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint8_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- uint32_t wght = __BFI(w1, w0, 16, 16);
-
- uint8_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- int32_t a0 = __SBFX(in, 0, 4);
- int32_t a1 = __SBFX(in, 4, 4);
- uint32_t act = __BFI(a1, a0, 16, 16);
-
- weightedSum = __SMLAD(act, wght, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Unsigned quad mac operation (4W/4A version)
- * @details Performs four mac operations for signed 4-bits weights
- * and unsigned 4-bits inputs. Extracts the four 4-bits weights
- * from two stored 8-bits weights and associates them into
- * two 32-bits values. Then extracts the four 4-bits inputs
- * from two stored 8-bits inputs and associates them into
- * two 32-bits values. Finally performs a double dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint16_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
-
- uint32_t evenW1 = __BFI(w2, w0, 16, 16);
- uint32_t oddW1 = __BFI(w3, w1, 16, 16);
-
- uint16_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- int32_t a0 = __UBFX(in, 0, 4);
- int32_t a1 = __UBFX(in, 4, 4);
- int32_t a2 = __UBFX(in, 8, 4);
- int32_t a3 = __UBFX(in, 12, 4);
-
- uint32_t evenA1 = __BFI(a2, a0, 16, 16);
- uint32_t oddA1 = __BFI(a3, a1, 16, 16);
-
- weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
- weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Signed quad mac operation (4W/4A version)
- * @details Performs four mac operations for signed 4-bits weights
- * and signed 4-bits inputs. Extracts the four 4-bits weights
- * from two stored 8-bits weights and associates them into
- * two 32-bits values. Then extracts the four 4-bits inputs
- * from two stored 8-bits inputs and associates them into
- * two 32-bits values. Finally performs a double dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the quad mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint16_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
-
- uint32_t evenW1 = __PKHBT(w2, w0, 16);
- uint32_t oddW1 = __PKHBT(w3, w1, 16);
-
- uint16_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- int32_t a0 = __SBFX(in, 0, 4);
- int32_t a1 = __SBFX(in, 4, 4);
- int32_t a2 = __SBFX(in, 8, 4);
- int32_t a3 = __SBFX(in, 12, 4);
-
- uint32_t evenA1 = __PKHBT(a2, a0, 16);
- uint32_t oddA1 = __PKHBT(a3, a1, 16);
-
- weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
- weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
-
- return weightedSum;
-}
-
-/**
- * @brief Unsigned octo mac operation (4W/4A version)
- * @details Performs eight mac operations for signed 4-bits weights
- * and unsigned 4-bits inputs. Extracts the eight 4-bits weights
- * from four stored 8-bits weights and associates them into
- * four 32-bits values. Then extracts the eight 4-bits inputs
- * from four stored 8-bits inputs and associates them into
- * four 32-bits values. Finally performs a quadruple dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be udata<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the octo mac operation
- */
-// template<typename Input_T, typename Weight_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T octoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// uint32_t wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// int32_t w0 = __SBFX(wt, 0, 4);
-// int32_t w1 = __SBFX(wt, 4, 4);
-// int32_t w2 = __SBFX(wt, 8, 4);
-// int32_t w3 = __SBFX(wt, 12, 4);
-// int32_t w4 = __SBFX(wt, 16, 4);
-// int32_t w5 = __SBFX(wt, 20, 4);
-// int32_t w6 = __SBFX(wt, 24, 4);
-// int32_t w7 = __SBFX(wt, 28, 4);
-
-// // uint32_t weight0 = __BFI(w4, w0, 16, 16);
-// // uint32_t weight1 = __BFI(w5, w1, 16, 16);
-// // uint32_t weight2 = __BFI(w6, w2, 16, 16);
-// // uint32_t weight3 = __BFI(w7, w3, 16, 16);
-
-// uint32_t weight0 = __PKHBT(w0, w4, 16);
-// uint32_t weight1 = __PKHBT(w1, w5, 16);
-// uint32_t weight2 = __PKHBT(w2, w6, 16);
-// uint32_t weight3 = __PKHBT(w3, w7, 16);
-
-// uint32_t in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// uint32_t act0 = in & 0xF000F;
-// uint32_t act1 = (in >> 4) & 0xF000F;
-// uint32_t act2 = (in >> 8) & 0xF000F;
-// uint32_t act3 = (in >> 12) & 0xF000F;
-
-// weightedSum = __SMLAD(act0, weight0, weightedSum);
-// weightedSum = __SMLAD(act1, weight1, weightedSum);
-// weightedSum = __SMLAD(act2, weight2, weightedSum);
-// weightedSum = __SMLAD(act3, weight3, weightedSum);
-
-// return weightedSum;
-// }
-
-// template<typename Input_T, typename Weight_T,
-// typename std::enable_if<(std::is_unsigned<Input_T>::value
-// && std::numeric_limits<Weight_T>::digits == 4
-// && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-// __attribute__((always_inline)) static inline
-// Sum_T octoMac(const Input_T* __restrict inputs,
-// const Weight_T* __restrict weights,
-// Sum_T weightedSum)
-// {
-// union n2d2_dataword wt;
-// std::memcpy((void*) &wt, weights, sizeof(wt));
-
-// union n2d2_udataword in;
-// std::memcpy((void*) &in, inputs, sizeof(in));
-
-// for (int i = 0; i < 4; ++i) {
-// weightedSum += (data<32>)(in.half_bytes[i].fields.op0) * wt.half_bytes[i].fields.op0;
-// weightedSum += (data<32>)(in.half_bytes[i].fields.op1) * wt.half_bytes[i].fields.op1;
-// }
-
-// // weightedSum += (data<32>)(in.half_bytes[0].fields.op0) * wt.half_bytes[0].fields.op0;
-// // weightedSum += (data<32>)(in.half_bytes[0].fields.op1) * wt.half_bytes[0].fields.op1;
-// // weightedSum += (data<32>)(in.half_bytes[1].fields.op0) * wt.half_bytes[1].fields.op0;
-// // weightedSum += (data<32>)(in.half_bytes[1].fields.op1) * wt.half_bytes[1].fields.op1;
-// // weightedSum += (data<32>)(in.half_bytes[2].fields.op0) * wt.half_bytes[2].fields.op0;
-// // weightedSum += (data<32>)(in.half_bytes[2].fields.op1) * wt.half_bytes[2].fields.op1;
-// // weightedSum += (data<32>)(in.half_bytes[3].fields.op0) * wt.half_bytes[3].fields.op0;
-// // weightedSum += (data<32>)(in.half_bytes[3].fields.op1) * wt.half_bytes[3].fields.op1;
-
-// return weightedSum;
-// }
-
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t wt;
- memcpy((void*) &wt, weights, sizeof(wt));
-
- // Works with weights * 4096 (weights << 12)
- const uint32_t WeightMask = 0xF000F000;
- uint32_t weight0 = WeightMask & (wt << 12);
- uint32_t weight1 = WeightMask & (wt << 8);
- uint32_t weight2 = WeightMask & (wt << 4);
- uint32_t weight3 = WeightMask & (wt);
-
- uint32_t in;
- memcpy((void*) &in, inputs, sizeof(in));
-
- const uint32_t ActMask = 0x000F000F; // to explicit instructions
- uint32_t act0 = in & ActMask;
- // Expect second operand shift
- uint32_t act1 = ActMask & (in >> 4);
- uint32_t act2 = ActMask & (in >> 8);
- uint32_t act3 = ActMask & (in >> 12);
-
- Sum_T sum = 0;
- sum = __SMLAD(act0, weight0, sum);
- sum = __SMLAD(act1, weight1, sum);
- sum = __SMLAD(act2, weight2, sum);
- sum = __SMLAD(act3, weight3, sum);
-
- return weightedSum + (sum >> 12);
-}
-
-/**
- * @brief Signed octo mac operation (4W/4A version)
- * @details Performs eight mac operations for signed 4-bits weights
- * and signed 4-bits inputs. Extracts the eight 4-bits weights
- * from four stored 8-bits weights and associates them into
- * four 32-bits values. Then extracts the eight 4-bits inputs
- * from four stored 8-bits inputs and associates them into
- * four 32-bits values. Finally performs a quadruple dual mac operation
- * with the __SMLAD instruction
- *
- * @tparam Input_T Input type (should be data<4>)
- * @tparam Weight_T Weight type (should be data<4>)
- *
- * @param[in] inputs Pointer to compressed input vector
- * @param[in] weights Pointer to compressed kernel weights
- * @param[in,out] weightedSum Accumulating sum from the
- * previous mac operations
- * @returns Updated weightedSum with
- * the result of the octo mac operation
- */
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t wt;
- std::memcpy((void*) &wt, weights, sizeof(wt));
-
- int32_t w0 = __SBFX(wt, 0, 4);
- int32_t w1 = __SBFX(wt, 4, 4);
- int32_t w2 = __SBFX(wt, 8, 4);
- int32_t w3 = __SBFX(wt, 12, 4);
- int32_t w4 = __SBFX(wt, 16, 4);
- int32_t w5 = __SBFX(wt, 20, 4);
- int32_t w6 = __SBFX(wt, 24, 4);
- int32_t w7 = __SBFX(wt, 28, 4);
-
- uint32_t evenW1 = __PKHBT(w2, w0, 16);
- uint32_t oddW1 = __PKHBT(w3, w1, 16);
- uint32_t evenW2 = __PKHBT(w6, w4, 16);
- uint32_t oddW2 = __PKHBT(w7, w5, 16);
-
- uint32_t in;
- std::memcpy((void*) &in, inputs, sizeof(in));
-
- int32_t a0 = __SBFX(in, 0, 4);
- int32_t a1 = __SBFX(in, 4, 4);
- int32_t a2 = __SBFX(in, 8, 4);
- int32_t a3 = __SBFX(in, 12, 4);
- int32_t a4 = __SBFX(in, 16, 4);
- int32_t a5 = __SBFX(in, 20, 4);
- int32_t a6 = __SBFX(in, 24, 4);
- int32_t a7 = __SBFX(in, 28, 4);
-
- uint32_t evenA1 = __PKHBT(a2, a0, 16);
- uint32_t oddA1 = __PKHBT(a3, a1, 16);
- uint32_t evenA2 = __PKHBT(a6, a4, 16);
- uint32_t oddA2 = __PKHBT(a7, a5, 16);
-
- weightedSum = __SMLAD(evenA1, evenW1, weightedSum);
- weightedSum = __SMLAD(oddA1, oddW1, weightedSum);
- weightedSum = __SMLAD(evenA2, evenW2, weightedSum);
- weightedSum = __SMLAD(oddA2, oddW2, weightedSum);
-
- return weightedSum;
-}
-
-
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-void macsOnParallel(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T* weightedSums,
- const int nb_data)
-{
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, ceil((double)nb_data/2));
-
- uint32_t in = 0;
- std::memcpy((void*) &in, inputs, ceil((double)nb_data/2));
-
- for (int i = 0; i < nb_data; ++i) {
- weightedSums[i] += __SBFX(wt, 4*i, 4) * __UBFX(in, 4*i, 4);
- }
-}
-
-template<typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(!std::is_unsigned<Input_T>::value
- && std::numeric_limits<Weight_T>::digits == 4
- && std::numeric_limits<Input_T>::digits == 4)>::type* = nullptr>
-void macsOnParallel(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T* weightedSums,
- const int nb_data)
-{
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, ceil((double)nb_data/2));
-
- uint32_t in = 0;
- std::memcpy((void*) &in, inputs, ceil((double)nb_data/2));
-
- for (int i = 0; i < nb_data; ++i) {
- weightedSums[i] += __SBFX(wt, 4*i, 4) * __SBFX(in, 4*i, 4);
- }
-}
-
-
-// ----------------------------------------------------------------------------
-// ------------------ Notes about performing MAC operations -------------------
-// --------------------------- with 1-bit weights -----------------------------
-// ----------------------------------------------------------------------------
-
-/**
- * @note How to perform MAC operations with 1-bit weight
- *
- * Working with an 1-bit weight means working only with two possible values
- * for each weight. Thus, it has been defined a convention that will be used
- * in the following functions in this file.
- * Convention: when the value of a weight is 0, it means 1
- * when the value of a weight is 1, it means -1
- *
- * Example: let's take a simple dual MAC operation
- * weightedSum = w0 * a0 + w1 * a1;
- *
- * if w0 = 0x00 and w1 = 0x01 then weightedSum should be:
- * weightedSum = a0 - a1;
- *
- * To easily perform MAC operations and use as often as possible
- * SIMD instructions to parallelize and speed up MAC calculations, most of
- * the following functions use the same scheme:
- *
- * - Perform a parallel subtraction of 0 and the weights
- * Some SIMD instructions as __USUB16 and __USUB8 can perform
- * parallel subtractions and activate a Greater or Equal flag (GE) if
- * the results of each subtraction is positive.
- * Thus, if the result of 0 - w0 >= 0 ==> GE[0] = 1
- * 0 - w0 < 0 ==> GE[0] = 0
- * (the results of the subtractions are not saved because only the
- * GE flags trigger is required)
- *
- * - Use of the __SEL instruction to read the GE flags
- * The __SEL can select an input from two values according to the
- * the GE flag provided by the previous subtraction. In the case of
- * the 1W/8A project, the two possible values selected by __SEL are
- * (+input) or (-input). Thus, __SEL is often used like "__SEL(in, -in)"
- * The results of __SEL are saved as MAC results
- *
- * - Addition of the accumuling sums with the results of the MAC operations
- * Use of __SADD16 or __SADD8 for signed additions
- *
- */
-
-// ----------------------------------------------------------------------------
-// ----------------- MAC computing functions for kernel -----------------------
-// ------------------------------- 1W / 8A ------------------------------------
-// ------------------------------- 1W / 7A ------------------------------------
-// ------------------------------- 1W / 6A ------------------------------------
-// ------------------------------- 1W / 5A ------------------------------------
-// ----------------------------------------------------------------------------
-
-template<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
- && std::numeric_limits<Input_T>::digits > 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T monoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- return weightedSum;
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 2)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 3)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[2])) : (Sum_T)(inputs[2]);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[2])) : (Sum_T)(inputs[2]);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[3])) : (Sum_T)(inputs[3]);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[2])) : (Sum_T)(inputs[2]);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[3])) : (Sum_T)(inputs[3]);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[4])) : (Sum_T)(inputs[4]);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 6)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[2])) : (Sum_T)(inputs[2]);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[3])) : (Sum_T)(inputs[3]);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[4])) : (Sum_T)(inputs[4]);
- weightedSum += (weights[0].fields.op2) ? (Sum_T)(-(inputs[5])) : (Sum_T)(inputs[5]);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0])) : (Sum_T)(inputs[0]);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[1])) : (Sum_T)(inputs[1]);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[2])) : (Sum_T)(inputs[2]);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[3])) : (Sum_T)(inputs[3]);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[4])) : (Sum_T)(inputs[4]);
- weightedSum += (weights[0].fields.op2) ? (Sum_T)(-(inputs[5])) : (Sum_T)(inputs[5]);
- weightedSum += (weights[0].fields.op1) ? (Sum_T)(-(inputs[6])) : (Sum_T)(inputs[6]);
-}
-
-
-// ----------------------------------------------------------------------------
-// ----------------- MAC computing functions for kernel -----------------------
-// ------------------------------- 1W / 8A ------------------------------------
-// ----------------------------------------------------------------------------
-
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T octoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 1);
- wt |= wt << 16;
-
- memcpy((void*) &in, inputs, sizeof(in));
- uint32_t evenA1 = __UXTB16(in);
- uint32_t oddA1 = __UXTB16_RORn(in, 8);
- uint32_t neg_evenA1 = __SSUB16(0, evenA1);
- uint32_t neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x40001);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x80002);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x400010);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x800020);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- return weightedSum;
-}
-
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T quadquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 2);
- wt |= wt << 16;
-
- memcpy((void*) &in, inputs, sizeof(in));
- uint32_t evenA1 = __UXTB16(in);
- uint32_t oddA1 = __UXTB16_RORn(in, 8);
- uint32_t neg_evenA1 = __SSUB16(0, evenA1);
- uint32_t neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x40001);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x80002);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x400010);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x800020);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x4000100);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x8000200);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x40001000);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x80002000);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- return weightedSum;
-}
-
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T octoquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t wt;
- memcpy((void*) &wt, weights, 4);
- uint32_t wt1 = __PKHBT(wt, wt, 16);
- uint32_t wt2 = __PKHTB(wt, wt, 16);
-
- memcpy((void*) &in, inputs, sizeof(in));
- uint32_t evenA1 = __UXTB16(in);
- uint32_t oddA1 = __UXTB16_RORn(in, 8);
- uint32_t neg_evenA1 = __SSUB16(0, evenA1);
- uint32_t neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x40001);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x80002);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x400010);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x800020);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x4000100);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x8000200);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt & 0x40001000);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt & 0x80002000);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 16, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt2 & 0x40001);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt2 & 0x80002);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 20, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt2 & 0x400010);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt2 & 0x800020);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 24, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt2 & 0x4000100);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt2 & 0x8000200);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
-
- memcpy((void*) &in, inputs + 28, sizeof(in));
- evenA1 = __UXTB16(in);
- oddA1 = __UXTB16_RORn(in, 8);
- neg_evenA1 = __SSUB16(0, evenA1);
- neg_oddA1 = __SSUB16(0, oddA1);
-
- __USUB16(0, wt2 & 0x40001000);
- mac_result = __SEL(evenA1, neg_evenA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- __USUB16(0, wt2 & 0x80002000);
- mac_result = __SEL(oddA1, neg_oddA1);
- weightedSum = __SADD16(mac_result, weightedSum);
-
- return weightedSum;
-}
-
-// ----------------------------------------------------------------------------
-// ----------------- MAC computing functions for kernel -----------------------
-// ------------------------------- 1W / 7A ------------------------------------
-// ----------------------------------------------------------------------------
-
-template<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 == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 1);
- wt |= wt << 8;
- wt |= wt << 16;
-
- memcpy((void*) &in, inputs, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0xC0C0C0C0) ^ 0xC0C0C0C0;
-
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08040201);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0xC0C0C0C0) ^ 0xC0C0C0C0;
-
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x80402010);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- return weightedSum;
-}
-
-template<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 == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 4);
-
- memcpy((void*) &in, inputs, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x01010101);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- return weightedSum;
-}
-
-template<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 == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 4);
-
- memcpy((void*) &in, inputs, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x01010101);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 16, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x10101010);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 20, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x20202020);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 24, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x40404040);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- memcpy((void*) &in, inputs + 28, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x80808080);
- mac_result = __SEL(in, neg_in);
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- return weightedSum;
-}
-
-// ----------------------------------------------------------------------------
-// ----------------- MAC computing functions for kernel -----------------------
-// ------------------------------- 1W / 5A ------------------------------------
-// ----------------------------------------------------------------------------
-
-template<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 == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 1);
- wt |= wt << 8;
- wt |= wt << 16;
-
- memcpy((void*) &in, inputs, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0x70707070) ^ 0x70707070;
-
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08040201);
- sum = __SEL(in, neg_in);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0x70707070) ^ 0x70707070;
-
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x80402010);
- mac_result = __SEL(in, neg_in);
-
- sum = __QADD8(sum, mac_result);
-
- return weightedSum;
-}
-
-template<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 == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T quadquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 2);
-
- memcpy((void*) &in, inputs, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x01010101);
- sum = __SEL(in, neg_in);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- weightedSum = __SXTAB16(weightedSum, sum);
- weightedSum = __SXTAB16_RORn(weightedSum, sum, 8);
-
- return weightedSum;
-}
-
-template<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 == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-Sum_T octoquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 4);
-
- memcpy((void*) &in, inputs, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x01010101);
- sum = __SEL(in, neg_in);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 16, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x10101010);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 20, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x20202020);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 24, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x40404040);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 28, sizeof(in));
- neg_in = __SSUB8(0, in);
- __USUB8(0, wt & 0x80808080);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- weightedSum = __SXTAB16(weightedSum, sum);
- weightedSum = __SXTAB16_RORn(weightedSum, sum, 8);
-
- return weightedSum;
-}
-
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS >= 8 && NB_ITERATIONS < 16)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum = octoMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-8>(inputs + 8, weights + 1, weightedSum);
-}
-
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS >= 16 && NB_ITERATIONS < 32)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
+template<typename Input_T>
+inline static
+uint32_t XTB16(uint32_t val)
{
- weightedSum = quadquadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-16>(inputs + 16, weights + 2, weightedSum);
+ return std::is_unsigned<Input_T>::value ? __UXTB16(val) : __SXTB16(val);
}
-template<int NB_ITERATIONS,
- 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
- && std::numeric_limits<Input_T>::digits > 4
- && NB_ITERATIONS >= 32)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T,
+ typename Sum_T>
+inline static
+Sum_T dualMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
{
- weightedSum = octoquadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-32>(inputs + 32, weights + 4, weightedSum);
-}
-
+ weightedSum += inputs[0] * weights[0]
+ + inputs[INPUTS_INC] * weights[WEIGHTS_INC];
-// ----------------------------------------------------------------------------
-// ----------------- MAC computing functions for kernel -----------------------
-// ------------------------------- 1W / 4A ------------------------------------
-// ----------------------------------------------------------------------------
-
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T monoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
return weightedSum;
}
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 2)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
-}
-
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 3)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[1].fields.op1)) : (Sum_T)(inputs[1].fields.op1);
-}
-
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[1].fields.op1)) : (Sum_T)(inputs[1].fields.op1);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[1].fields.op0)) : (Sum_T)(inputs[1].fields.op0);
-}
-
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[1].fields.op1)) : (Sum_T)(inputs[1].fields.op1);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[1].fields.op0)) : (Sum_T)(inputs[1].fields.op0);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[2].fields.op1)) : (Sum_T)(inputs[2].fields.op1);
-}
-
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 6)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[1].fields.op1)) : (Sum_T)(inputs[1].fields.op1);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[1].fields.op0)) : (Sum_T)(inputs[1].fields.op0);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[2].fields.op1)) : (Sum_T)(inputs[2].fields.op1);
- weightedSum += (weights[0].fields.op2) ? (Sum_T)(-(inputs[2].fields.op0)) : (Sum_T)(inputs[2].fields.op0);
-}
-
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum += (weights[0].fields.op7) ? (Sum_T)(-(inputs[0].fields.op1)) : (Sum_T)(inputs[0].fields.op1);
- weightedSum += (weights[0].fields.op6) ? (Sum_T)(-(inputs[0].fields.op0)) : (Sum_T)(inputs[0].fields.op0);
- weightedSum += (weights[0].fields.op5) ? (Sum_T)(-(inputs[1].fields.op1)) : (Sum_T)(inputs[1].fields.op1);
- weightedSum += (weights[0].fields.op4) ? (Sum_T)(-(inputs[1].fields.op0)) : (Sum_T)(inputs[1].fields.op0);
- weightedSum += (weights[0].fields.op3) ? (Sum_T)(-(inputs[2].fields.op1)) : (Sum_T)(inputs[2].fields.op1);
- weightedSum += (weights[0].fields.op2) ? (Sum_T)(-(inputs[2].fields.op0)) : (Sum_T)(inputs[2].fields.op0);
- weightedSum += (weights[0].fields.op1) ? (Sum_T)(-(inputs[3].fields.op1)) : (Sum_T)(inputs[3].fields.op1);
-}
-
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T octoMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T,
+ typename Sum_T,
+ typename std::enable_if<std::is_floating_point<Input_T>::value>::type* = nullptr>
+inline static
+Sum_T quadMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 1);
- wt |= wt << 8;
- wt |= wt << 16;
-
- memcpy((void*) &in, inputs, sizeof(in));
-
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x40100401);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- neg_in = __SSUB8(0, (in >> 4) & 0xF0F0F0F0);
- __USUB8(0, wt & 0x80200802);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC]
+ + inputs[1*INPUTS_INC] * weights[1*WEIGHTS_INC]
+ + inputs[2*INPUTS_INC] * weights[2*WEIGHTS_INC]
+ + inputs[3*INPUTS_INC] * weights[3*WEIGHTS_INC];
return weightedSum;
}
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T quadquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
+template<int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ typename Input_T,
+ typename Weight_T,
+ typename Sum_T,
+ typename std::enable_if<!std::is_floating_point<Input_T>::value>::type* = nullptr>
+inline static
+Sum_T quadMac(const Input_T* __restrict inputs,
+ const Weight_T* __restrict weights,
+ Sum_T weightedSum)
{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
- uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 4);
-
- memcpy((void*) &in, inputs, sizeof(in));
-
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x01010101);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
+ if(INPUTS_INC != 1 || WEIGHTS_INC != 1) {
+ weightedSum += inputs[0*INPUTS_INC] * weights[0*WEIGHTS_INC]
+ + inputs[1*INPUTS_INC] * weights[1*WEIGHTS_INC]
+ + inputs[2*INPUTS_INC] * weights[2*WEIGHTS_INC]
+ + inputs[3*INPUTS_INC] * weights[3*WEIGHTS_INC];
- memcpy((void*) &in, inputs + 4, sizeof(in));
-
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
-
- weightedSum = __SXTAB16(weightedSum, sum);
- weightedSum = __SXTAB16_RORn(weightedSum, sum, 8);
-
- return weightedSum;
-}
+ return weightedSum;
+ }
-template<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>
-__attribute__((always_inline)) static inline
-Sum_T octoquadMac (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T weightedSum)
-{
- uint32_t sum = 0;
- uint32_t mac_result = 0;
+ // Inputs loading & preparation
uint32_t in;
- uint32_t neg_in;
- uint32_t wt = 0;
- std::memcpy((void*) &wt, weights, 4);
-
memcpy((void*) &in, inputs, sizeof(in));
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x01010101);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x02020202);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 4, sizeof(in));
-
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x04040404);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x08080808);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 8, sizeof(in));
-
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x10101010);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
+ uint32_t in1 = XTB16<Input_T>(in);
+ uint32_t in2 = XTB16<Input_T>(in >> 8);
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x20202020);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
-
- memcpy((void*) &in, inputs + 12, sizeof(in));
+ // Weights loading & preparation
+ uint32_t wt;
+ memcpy((void*) &wt, weights, sizeof(wt));
- neg_in = __SSUB8(0, in & 0x0F0F0F0F);
- __USUB8(0, wt & 0x40404040);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
+ uint32_t wt1 = XTB16<Weight_T>(wt);
+ uint32_t wt2 = XTB16<Weight_T>(wt >> 8);
- neg_in = __SSUB8(0, (in >> 4) & 0x0F0F0F0F);
- __USUB8(0, wt & 0x80808080);
- mac_result = __SEL(in, neg_in);
- sum = __QADD8(sum, mac_result);
+ // Computation
+ if(std::is_same<Sum_T, int32_t>::value) {
+ weightedSum = __SMLAD(in1, wt1, weightedSum);
+ weightedSum = __SMLAD(in2, wt2, weightedSum);
+ }
+ else {
+ weightedSum = __SMLALD(in1, wt1, weightedSum);
+ weightedSum = __SMLALD(in2, wt2, weightedSum);
- weightedSum = __SXTAB16(weightedSum, sum);
- weightedSum = __SXTAB16_RORn(weightedSum, sum, 8);
+ }
return weightedSum;
}
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS >= 8 && NB_ITERATIONS < 16)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum = octoMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-8>(inputs + 4, weights + 1, weightedSum);
-}
-template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS >= 16 && NB_ITERATIONS < 32)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum = quadquadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-16>(inputs + 8, weights + 2, weightedSum);
-}
+// **************************************************************************
+// * Multiply-accumulate the values in inputs and weights for NB_ITERATIONS *
+// **************************************************************************
template<int NB_ITERATIONS,
- 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
- && NB_ITERATIONS >= 32)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange (const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum = octoquadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS-32>(inputs + 16, weights + 4, weightedSum);
-}
-
-
-// ----------------------------------------------------------------------------
-// -------------- MAC computing functions for kernel 1W-7A --------------------
-// ----------------------------------------------------------------------------
-
-template<typename Input_T,
- typename std::enable_if<(std::numeric_limits<Input_T>::digits == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-uint32_t quadMacInter(const Input_T* __restrict inputs,
- const uint32_t weight,
- uint32_t weightedSum)
-{
- uint32_t in;
- memcpy((void*) &in, inputs, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0xC0C0C0C0) ^ 0xC0C0C0C0;
-
- uint32_t neg_in = __SSUB8(0, in);
-
- __USUB8(0, weight);
- uint32_t mac_result = __SEL(in, neg_in);
-
- uint32_t evenA1 = __SXTB16(mac_result);
- uint32_t oddA1 = __SXTB16_RORn(mac_result, 8);
-
- weightedSum = __SADD16(evenA1, weightedSum);
- weightedSum = __SADD16(oddA1, weightedSum);
-
- return weightedSum;
-}
-
-template<typename Input_T,
- typename std::enable_if<(std::numeric_limits<Input_T>::digits == 7)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-uint32_t quadMacInterV2(const Input_T* __restrict inputs,
- const uint32_t weight,
- uint32_t weightedSum)
-{
- uint32_t in;
- memcpy((void*) &in, inputs, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0xC0C0C0C0) ^ 0xC0C0C0C0;
-
- uint32_t neg_in = __SSUB8(0, in);
-
- __USUB8(0, weight);
- uint32_t mac_result = __SEL(in, neg_in);
-
- weightedSum = __SXTAB16(weightedSum, mac_result);
- weightedSum = __SXTAB16_RORn(weightedSum, mac_result, 8);
-
- return weightedSum;
-}
-
-
-// ----------------------------------------------------------------------------
-// -------------- MAC computing functions for kernel 1W-5A --------------------
-// ----------------------------------------------------------------------------
-
-template<typename Input_T,
- typename std::enable_if<(std::numeric_limits<Input_T>::digits == 5)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-uint32_t quadMacInter(const Input_T* __restrict inputs,
- const uint32_t weight,
- uint32_t weightedSum)
-{
- uint32_t in;
- memcpy((void*) &in, inputs, sizeof(in));
-
- // Sign extend
- if (!std::is_unsigned<Input_T>::value)
- in = (in + 0x70707070) ^ 0x70707070;
-
- uint32_t neg_in = __SSUB8(0, in);
-
- __USUB8(0, weight);
- uint32_t mac_result = __SEL(in, neg_in);
-
- weightedSum = __QADD8(weightedSum, mac_result);
-
- return weightedSum;
-}
-
-
-// ----------------------------------------------------------------------------
-// ------------------- MAC computing general functions ------------------------
-// ----------------------------------------------------------------------------
-
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ class Input_T,
+ class Weight_T,
+ class Sum_T,
typename std::enable_if<(NB_ITERATIONS == 0)>::type* = nullptr>
-__attribute__((always_inline)) static inline
+inline static
void macsOnRange(const Input_T* __restrict /*inputs*/,
const Weight_T* __restrict /*weights*/,
- Sum_T& /*weightedSum*/)
+ Sum_T& __restrict /*weightedSum*/)
{
- // Nothing should happen
+ // Nothing to do
}
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ class Input_T,
+ class Weight_T,
+ class Sum_T,
typename std::enable_if<(NB_ITERATIONS == 1)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- weightedSum = monoMac(inputs, weights, weightedSum);
-}
-
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(NB_ITERATIONS >= 2 && NB_ITERATIONS < 4 && std::numeric_limits<Weight_T>::digits > 1)>::type* = nullptr>
-__attribute__((always_inline)) static inline
+inline static
void macsOnRange(const Input_T* __restrict inputs,
const Weight_T* __restrict weights,
- Sum_T& weightedSum)
+ Sum_T& __restrict weightedSum)
{
- constexpr unsigned int idxI
- = (std::numeric_limits<Input_T>::digits > 4) ? 2 : 1;
- constexpr unsigned int idxW
- = (std::numeric_limits<Weight_T>::digits > 4) ? 2 : 1;
-
- weightedSum = dualMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS - 2>(inputs + idxI, weights + idxW, weightedSum);
+ weightedSum += (*weights) * (*inputs);
}
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<NB_ITERATIONS >= 4
- && (std::numeric_limits<Weight_T>::digits > 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ class Input_T,
+ class Weight_T,
+ class Sum_T,
+ typename std::enable_if<(NB_ITERATIONS >= 2 && NB_ITERATIONS < 4)>::type* = nullptr>
+inline static
void macsOnRange(const Input_T* __restrict inputs,
const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- constexpr unsigned int idxI
- = (std::numeric_limits<Input_T>::digits > 4)
- ? 4 : (std::numeric_limits<Input_T>::digits == 4) ? 2 : 1;
-
- constexpr unsigned int idxW = 4;
-
- weightedSum = quadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS - 4>(inputs + idxI, weights + idxW, weightedSum);
-}
-
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<(NB_ITERATIONS >= 4 && NB_ITERATIONS < 8)
- && (std::numeric_limits<Weight_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
+ 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 + 2*WEIGHTS_INC,
+ weightedSum);
+}
+
+/**
+ * @brief MACs Processing
+ * @details Performs NB_ITERATIONS MACs operations, storing results into the
+ * weightedSum variable.
+ *
+ * @tparam NB_ITERATIONS Number of MACs to perform
+ * @tparam INPUTS_INC Input Stride
+ * @tparam WEIGHTS_INC Weights Stride
+ * @tparam Input_T Input Type
+ *
+ * @param inputs Pointer to inputs vector
+ * @param weights Pointer to weights vector
+ * @param weightedSum Pointer to weightedSum
+*/
+template<int NB_ITERATIONS,
+ int INPUTS_INC = 1,
+ int WEIGHTS_INC = 1,
+ class Input_T,
+ class Weight_T,
+ class Sum_T,
+ typename std::enable_if<(NB_ITERATIONS >= 4)>::type* = nullptr>
+inline static
void macsOnRange(const Input_T* __restrict inputs,
const Weight_T* __restrict weights,
- Sum_T& weightedSum)
+ Sum_T& __restrict weightedSum)
{
- constexpr unsigned int idxI
- = (std::numeric_limits<Input_T>::digits > 4)
- ? 4 : (std::numeric_limits<Input_T>::digits == 4) ? 2 : 1;
-
- constexpr unsigned int idxW = 2;
-
- weightedSum = quadMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS - 4>(inputs + idxI, weights + idxW, weightedSum);
+ weightedSum = quadMac<INPUTS_INC, WEIGHTS_INC>(inputs, weights, weightedSum);
+ macsOnRange<NB_ITERATIONS - 4, INPUTS_INC, WEIGHTS_INC>(inputs + 4*INPUTS_INC,
+ weights + 4*WEIGHTS_INC,
+ weightedSum);
}
-template<int NB_ITERATIONS, typename Input_T, typename Weight_T, typename Sum_T,
- typename std::enable_if<NB_ITERATIONS >= 8
- && (std::numeric_limits<Weight_T>::digits == 4)>::type* = nullptr>
-__attribute__((always_inline)) static inline
-void macsOnRange(const Input_T* __restrict inputs,
- const Weight_T* __restrict weights,
- Sum_T& weightedSum)
-{
- constexpr unsigned int idxI
- = (std::numeric_limits<Input_T>::digits > 4)
- ? 8 : (std::numeric_limits<Input_T>::digits == 4)
- ? 4 : (std::numeric_limits<Input_T>::digits == 2)
- ? 2 : 1;
-
- constexpr unsigned int idxW = 4;
-
- weightedSum = octoMac(inputs, weights, weightedSum);
- macsOnRange<NB_ITERATIONS - 8>(inputs + idxI, weights + idxW, weightedSum);
-}
+} // N2D2_Export
-#endif // __N2D2_MAC_FUNCTIONS_HPP__
+#endif // __N2D2_EXPORT_CPP_MACS_HPP__
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/pool_config.jinja b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/pool_config.jinja
index 363e6f9bcd3f83460a3a2533ce6e243c887442fd..1900612b65ab2c2c581886acd3894659233c64b6 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/pool_config.jinja
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/pool_config.jinja
@@ -2,7 +2,7 @@
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
-{# #include "typedefs.h" #}
+{# #include "kernels/typedefs.h" #}
{# For layer configuration -#}
{% include "./_def_io.jinja" %}
diff --git a/aidge_export_arm_cortexm/boards/stm32/H7/Inc/assert.h b/aidge_export_arm_cortexm/boards/stm32/H7/Inc/assert.h
index bb0ec19d8544bf8c8830cb85da00f7e065422cff..fab32bda495f9a19850ebd1e38830e9f0bdd858e 100644
--- a/aidge_export_arm_cortexm/boards/stm32/H7/Inc/assert.h
+++ b/aidge_export_arm_cortexm/boards/stm32/H7/Inc/assert.h
@@ -1,33 +1,33 @@
-/*
- (C) Copyright 2019 CEA LIST. All Rights Reserved.
- Contributor(s): Olivier BICHLER (olivier.bichler@cea.fr)
+// /*
+// (C) Copyright 2019 CEA LIST. All Rights Reserved.
+// Contributor(s): Olivier BICHLER (olivier.bichler@cea.fr)
- This file is not part of the open source version of N2D2 and is NOT under
- the CeCILL-C license. This code is the property of the CEA. It can not be
- copied or disseminated without its authorization.
-*/
-#ifndef ASSERT_H
-#define ASSERT_H
+// This file is not part of the open source version of N2D2 and is NOT under
+// the CeCILL-C license. This code is the property of the CEA. It can not be
+// copied or disseminated without its authorization.
+// */
+// #ifndef ASSERT_H
+// #define ASSERT_H
-#ifdef __cplusplus
-extern "C" {
-#endif
+// #ifdef __cplusplus
+// extern "C" {
+// #endif
-inline void assert_failure(const char* msg, const char* file, int line) {
-// printf("Assert failure: %s in %s:%d.\r\n", msg, file, line);
- while(1) {}
-}
+// inline void assert_failure(const char* msg, const char* file, int line) {
+// // printf("Assert failure: %s in %s:%d.\r\n", msg, file, line);
+// while(1) {}
+// }
-#ifdef NDEBUG
-#define assert(test) ((void)0)
-#define assertm(test, msg) ((void)0)
-#else
-#define assert(test) do { if(!(test)) { assert_failure("error", __FILE__, __LINE__); } } while(0)
-#define assertm(test, msg) do { if(!(test)) { assert_failure(msg, __FILE__, __LINE__); } } while(0)
-#endif
+// #ifdef NDEBUG
+// #define assert(test) ((void)0)
+// #define assertm(test, msg) ((void)0)
+// #else
+// #define assert(test) do { if(!(test)) { assert_failure("error", __FILE__, __LINE__); } } while(0)
+// #define assertm(test, msg) do { if(!(test)) { assert_failure(msg, __FILE__, __LINE__); } } while(0)
+// #endif
-#ifdef __cplusplus
-}
-#endif
+// #ifdef __cplusplus
+// }
+// #endif
-#endif
+// #endif
diff --git a/aidge_export_arm_cortexm/export.py b/aidge_export_arm_cortexm/export.py
index fe7884eb3fa18e6a62bf9ad9bd792d403d1191b8..ebac2d68929399bab55234684a5d35c6c246a9dd 100644
--- a/aidge_export_arm_cortexm/export.py
+++ b/aidge_export_arm_cortexm/export.py
@@ -42,12 +42,11 @@ def gen_board_files(path:str, board:str)->None:
raise ValueError(f"Board {board} is not supported, supported board are:\n\t-{joint_board_str}")
if isinstance(path, str): path = Path(path)
- # Create dnn directory is not exist
+ # Create dnn directory if not exist
dnn_folder = path / "dnn"
os.makedirs(str(dnn_folder), exist_ok=True)
# Determine which board the user wants
# to select correct config
-
# Copy all static files in the export
shutil.copytree(BOARDS_MAP[board], str(path), dirs_exist_ok=True)
diff --git a/aidge_export_arm_cortexm/operators.py b/aidge_export_arm_cortexm/operators.py
index 5f11b8fe909a6cc57b3d31d654402159021c06e9..2903edefc53199337fb4a658a6fc27ac2ed6b7cc 100644
--- a/aidge_export_arm_cortexm/operators.py
+++ b/aidge_export_arm_cortexm/operators.py
@@ -86,11 +86,11 @@ class Producer_ARMCortexM(ExportNode):
def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
super().__init__(node, mem_info, conversion_map)
-
+
weights = self.operator.get_output(0)
self.values = np.array(weights).reshape(-1)
-
+
def export(self, export_folder: Path):
header_path = f"include/parameters/{self.attributes['name']}.hpp"
@@ -261,7 +261,7 @@ class Pad_ARMCortexM(ExportNodeCpp):
@ExportLibAidgeARM.register("ReLU", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class ReLU_ARMCortexM(ExportNodeCpp):
+class ReLU_ARMCortexM_float32(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
@@ -274,7 +274,7 @@ class ReLU_ARMCortexM(ExportNodeCpp):
]
@ExportLibAidgeARM.register("Conv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class Conv_ARMCortexM(ExportNodeCpp):
+class Conv_ARMCortexM_float32(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
self.attributes["activation"] = "Linear"
@@ -287,27 +287,33 @@ class Conv_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "conv_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "Conv.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "Conv.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
+ # str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h"),
]
@ExportLibAidgeARM.register_generic("ArmPadConv2D", aidge_core.ImplSpec([
aidge_core.IOSpec(aidge_core.dtype.any), # Input[0] : Input Spec
aidge_core.IOSpec(aidge_core.dtype.dual_int4), # Input[1] : Weight Spec
aidge_core.IOSpec(aidge_core.dtype.int32) # Input[2] : Bias Spec
- ],
+ ],
[
aidge_core.IOSpec(aidge_core.dtype.any) # Output[0] : Output spec
]))
class PadConvScaling_ARMCortexM(ExportNodeCpp):
def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
super().__init__(node, mem_info, conversion_map)
-
+
self.attributes["activation"] = "Linear"
-
+
self.attributes["padding"] = [0, 0]
if self.operator.attr.has_attr("Pad2D_0"):
self.attributes["padding"] = self.operator.attr.get_attr("Pad2D_0").get_attr("begin_end_borders")
-
+
self.attributes["kernel_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("kernel_dims")
self.attributes["stride_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("stride_dims")
self.attributes["dilation_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("dilation_dims")
@@ -329,14 +335,15 @@ class PadConvScaling_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "lowbit_conv_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "aidge_supportfunctions.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
@@ -344,20 +351,20 @@ class PadConvScaling_ARMCortexM(ExportNodeCpp):
aidge_core.IOSpec(aidge_core.dtype.any), # Input[0] : Input Spec
aidge_core.IOSpec(aidge_core.dtype.dual_int4), # Input[1] : Weight Spec
aidge_core.IOSpec(aidge_core.dtype.int32) # Input[2] : Bias Spec
- ],
+ ],
[
aidge_core.IOSpec(aidge_core.dtype.any) # Output[0] : Output spec
]))
class ConvScaling_ARMCortexM(ExportNodeCpp):
def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
super().__init__(node, mem_info, conversion_map)
-
+
self.attributes["activation"] = "Linear"
-
+
self.attributes["padding"] = [0, 0]
if self.operator.attr.has_attr("Pad2D_0"):
self.attributes["padding"] = self.operator.attr.get_attr("Pad2D_0").get_attr("begin_end_borders")
-
+
self.attributes["kernel_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("kernel_dims")
self.attributes["stride_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("stride_dims")
self.attributes["dilation_dims"] = self.operator.attr.get_attr("Conv2D_0").get_attr("dilation_dims")
@@ -378,14 +385,15 @@ class ConvScaling_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "lowbit_conv_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "aidge_supportfunctions.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
@@ -393,11 +401,11 @@ class ConvScaling_ARMCortexM(ExportNodeCpp):
aidge_core.IOSpec(aidge_core.dtype.any), # Input[0] : Input Spec
aidge_core.IOSpec(aidge_core.dtype.dual_int4), # Input[1] : Weight Spec
aidge_core.IOSpec(aidge_core.dtype.int32) # Input[2] : Bias Spec
- ],
+ ],
[
aidge_core.IOSpec(aidge_core.dtype.any) # Output[0] : Output spec
]))
-class FCScaling_ARMCortexM(ExportNodeCpp):
+class FCScaling_ARMCortexM_int4(ExportNodeCpp):
def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
super().__init__(node, mem_info, conversion_map)
self.attributes["activation"] = "Linear"
@@ -424,38 +432,42 @@ class FCScaling_ARMCortexM(ExportNodeCpp):
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
- ]
-
-@ExportLibAidgeARM.register("MaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
-class LowbitPooling_ARMCortexM(ExportNodeCpp):
- def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
- super().__init__(node, mem_info, conversion_map)
-
- self.attributes["activation"] = "Linear"
- self.attributes["pool_type"] = "Max"
- # No padding with MaxPooling or AvgPooling
- # Use PaddedMaxPooling/PaddedAvgPooling to add padding attribute
- self.attributes["padding"] = [0, 0]
-
- self.attributes["kernel_dims"] = node.get_operator().attr.kernel_dims
- self.attributes["stride_dims"] = node.get_operator().attr.stride_dims
-
- self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "pool_config.jinja")
- self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "lowbit_pool_kernel.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "LowbitPooling.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
- self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "LowbitPooling.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "aidge_supportfunctions.h"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
- ]
+# FIXME This take the precedence on float32 kernel due to poor management of IOSpec
+# Need to update the IOSpec
+# @ExportLibAidgeARM.register("MaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
+# class LowbitPooling_ARMCortexM(ExportNodeCpp):
+# def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
+# super().__init__(node, mem_info, conversion_map)
+
+# self.attributes["activation"] = "Linear"
+# self.attributes["pool_type"] = "Max"
+# # No padding with MaxPooling or AvgPooling
+# # Use PaddedMaxPooling/PaddedAvgPooling to add padding attribute
+# self.attributes["padding"] = [0, 0]
+
+# self.attributes["kernel_dims"] = node.get_operator().attr.kernel_dims
+# self.attributes["stride_dims"] = node.get_operator().attr.stride_dims
+
+# self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "pool_config.jinja")
+# self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "lowbit_pool_kernel.jinja")
+# self.include_list = []
+# self.kernels_to_copy = [
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "LowbitPooling.hpp")
+# ]
+
+# self.kernels_to_copy = [
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "LowbitPooling.hpp"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "aidge_supportfunctions.h"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
+# str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
+# ]
@@ -464,11 +476,11 @@ class LowbitPooling_ARMCortexM(ExportNodeCpp):
aidge_core.IOSpec(aidge_core.dtype.any), # Input[0] : Input Spec
aidge_core.IOSpec(aidge_core.dtype.int4), # Input[1] : Weight Spec
aidge_core.IOSpec(aidge_core.dtype.any) # Input[2] : Bias Spec
- ],
+ ],
[
aidge_core.IOSpec(aidge_core.dtype.int4) # Output[0] : Output spec
]))
-class Conv_ARMCortexM(ExportNodeCpp):
+class Conv_ARMCortexM_int4(ExportNodeCpp):
def __init__(self, node, mem_info, conversion_map = datatype_converter_aidge2arm):
super().__init__(node, mem_info, conversion_map)
self.attributes["activation"] = "Linear"
@@ -481,14 +493,15 @@ class Conv_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "lowbit_conv_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "LowbitConv.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "aidge_supportfunctions.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "subkernels_functions.hpp"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
- str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
@ExportLibAidgeARM.register("ConvDepthWise2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
@@ -574,14 +587,15 @@ class Pooling_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "pool_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "Pooling.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Pooling" / "Pooling.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp")
]
self.kernel = node.get_operator().attr.kernel_dims
self.stride = node.get_operator().attr.stride_dims
@ExportLibAidgeARM.register("FC", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class FC_ARMCortexM(ExportNodeCpp):
+class FC_ARMCortexM_float32(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
self.attributes["activation"] = "Linear"
@@ -594,7 +608,13 @@ class FC_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "fc_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "FullyConnected" / "Fc.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "FullyConnected" / "Fc.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
+ # str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
@ExportLibAidgeARM.register("MaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
@@ -624,7 +644,13 @@ class FC_ARMCortexM(ExportNodeCpp):
self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "fc_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "_Aidge_Arm" / "kernels" / "FullyConnected" / "Fc.hpp")
+ str(ROOT / "_Aidge_Arm" / "kernels" / "FullyConnected" / "Fc.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "Macs.hpp"),
+ # str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "swar_arm_acle.h"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "nn_scaling_functions.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "typedefs.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "utils.hpp"),
+ str(ROOT / "_Aidge_Arm" / "kernels" / "Utils" / "assert.h")
]
@ExportLibAidgeARM.register("Add", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))