digest-kangarootwelve 0.2.0 → 0.3.1

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-times2-SnP.h

@@ -0,0 +1,47 @@
+/*
+Implementation by Gilles Van Assche, hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+
+---
+
+Please refer to PlSnP-documentation.h for more details.
+*/
+
+#ifndef _KeccakP_1600_times2_SnP_h_
+#define _KeccakP_1600_times2_SnP_h_
+
+#include "SIMD128-config.h"
+
+#define KeccakP1600times2_implementation        "128-bit SIMD implementation (" KeccakP1600times2_implementation_config ")"
+#define KeccakP1600times2_statesSizeInBytes     400
+#define KeccakP1600times2_statesAlignment       16
+#define KeccakF1600times2_FastLoop_supported
+
+#include <stddef.h>
+
+#define KeccakP1600times2_StaticInitialize()
+void KeccakP1600times2_InitializeAll(void *states);
+#define KeccakP1600times2_AddByte(states, instanceIndex, byte, offset) \
+    ((unsigned char*)(states))[(instanceIndex)*8 + ((offset)/8)*2*8 + (offset)%8] ^= (byte)
+void KeccakP1600times2_AddBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times2_AddLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times2_OverwriteBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times2_OverwriteLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times2_OverwriteWithZeroes(void *states, unsigned int instanceIndex, unsigned int byteCount);
+void KeccakP1600times2_PermuteAll_4rounds(void *states);
+void KeccakP1600times2_PermuteAll_6rounds(void *states);
+void KeccakP1600times2_PermuteAll_12rounds(void *states);
+void KeccakP1600times2_PermuteAll_24rounds(void *states);
+void KeccakP1600times2_ExtractBytes(const void *states, unsigned int instanceIndex, unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times2_ExtractLanesAll(const void *states, unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times2_ExtractAndAddBytes(const void *states, unsigned int instanceIndex,  const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+void KeccakP1600times2_ExtractAndAddLanesAll(const void *states, const unsigned char *input, unsigned char *output, unsigned int laneCount, unsigned int laneOffset);
+size_t KeccakF1600times2_FastLoop_Absorb(void *states, unsigned int laneCount, unsigned int laneOffsetParallel, unsigned int laneOffsetSerial, const unsigned char *data, size_t dataByteLen);
+
+#endif

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-times4-SnP.h

@@ -0,0 +1,45 @@
+/*
+Implementation by the Keccak Team, namely, Guido Bertoni, Joan Daemen,
+Michaël Peeters, Gilles Van Assche and Ronny Van Keer,
+hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+
+---
+
+Please refer to PlSnP-documentation.h for more details.
+*/
+
+#ifndef _KeccakP_1600_times4_SnP_h_
+#define _KeccakP_1600_times4_SnP_h_
+
+#include "KeccakP-1600-times2-SnP.h"
+
+#define KeccakP1600times4_implementation        "fallback on times-2 implementation (" KeccakP1600times2_implementation ")"
+#define KeccakP1600times4_statesSizeInBytes     (((KeccakP1600times2_statesSizeInBytes+(KeccakP1600times2_statesAlignment-1))/KeccakP1600times2_statesAlignment)*KeccakP1600times2_statesAlignment*2)
+#define KeccakP1600times4_statesAlignment       KeccakP1600times2_statesAlignment
+#define KeccakP1600times4_isFallback
+
+void KeccakP1600times4_StaticInitialize( void );
+void KeccakP1600times4_InitializeAll(void *states);
+void KeccakP1600times4_AddByte(void *states, unsigned int instanceIndex, unsigned char data, unsigned int offset);
+void KeccakP1600times4_AddBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times4_AddLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times4_OverwriteBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times4_OverwriteLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times4_OverwriteWithZeroes(void *states, unsigned int instanceIndex, unsigned int byteCount);
+void KeccakP1600times4_PermuteAll_4rounds(void *states);
+void KeccakP1600times4_PermuteAll_6rounds(void *states);
+void KeccakP1600times4_PermuteAll_12rounds(void *states);
+void KeccakP1600times4_PermuteAll_24rounds(void *states);
+void KeccakP1600times4_ExtractBytes(const void *states, unsigned int instanceIndex, unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times4_ExtractLanesAll(const void *states, unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times4_ExtractAndAddBytes(const void *states, unsigned int instanceIndex,  const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+void KeccakP1600times4_ExtractAndAddLanesAll(const void *states, const unsigned char *input, unsigned char *output, unsigned int laneCount, unsigned int laneOffset);
+
+#endif

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-times4-on2.c

@@ -0,0 +1,38 @@
+/*
+Implementation by the Keccak Team, namely, Guido Bertoni, Joan Daemen,
+Michaël Peeters, Gilles Van Assche and Ronny Van Keer,
+hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+
+---
+
+This file implements Keccak-p[1600]×4 in a PlSnP-compatible way.
+Please refer to PlSnP-documentation.h for more details.
+
+This implementation comes with KeccakP-1600-times4-SnP.h in the same folder.
+Please refer to LowLevel.build for the exact list of other files it must be combined with.
+*/
+
+#include "KeccakP-1600-times2-SnP.h"
+
+#define prefix                          KeccakP1600times4
+#define PlSnP_baseParallelism           2
+#define PlSnP_targetParallelism         4
+#define SnP_laneLengthInBytes           8
+#define SnP                             KeccakP1600times2
+#define SnP_PermuteAll                  KeccakP1600times2_PermuteAll_24rounds
+#define SnP_PermuteAll_12rounds         KeccakP1600times2_PermuteAll_12rounds
+#define SnP_PermuteAll_6rounds          KeccakP1600times2_PermuteAll_6rounds
+#define SnP_PermuteAll_4rounds          KeccakP1600times2_PermuteAll_4rounds
+#define PlSnP_PermuteAll                KeccakP1600times4_PermuteAll_24rounds
+#define PlSnP_PermuteAll_12rounds       KeccakP1600times4_PermuteAll_12rounds
+#define PlSnP_PermuteAll_6rounds        KeccakP1600times4_PermuteAll_6rounds
+#define PlSnP_PermuteAll_4rounds        KeccakP1600times4_PermuteAll_4rounds
+
+#include "PlSnP-Fallback.inc"

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-times8-SnP.h

@@ -0,0 +1,45 @@
+/*
+Implementation by the Keccak Team, namely, Guido Bertoni, Joan Daemen,
+Michaël Peeters, Gilles Van Assche and Ronny Van Keer,
+hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+
+---
+
+Please refer to PlSnP-documentation.h for more details.
+*/
+
+#ifndef _KeccakP_1600_times8_SnP_h_
+#define _KeccakP_1600_times8_SnP_h_
+
+#include "KeccakP-1600-times2-SnP.h"
+
+#define KeccakP1600times8_implementation        "fallback on times-2 implementation (" KeccakP1600times2_implementation ")"
+#define KeccakP1600times8_statesSizeInBytes     (((KeccakP1600times2_statesSizeInBytes+(KeccakP1600times2_statesAlignment-1))/KeccakP1600times2_statesAlignment)*KeccakP1600times2_statesAlignment*4)
+#define KeccakP1600times8_statesAlignment       KeccakP1600times2_statesAlignment
+#define KeccakP1600times8_isFallback
+
+void KeccakP1600times8_StaticInitialize( void );
+void KeccakP1600times8_InitializeAll(void *states);
+void KeccakP1600times8_AddByte(void *states, unsigned int instanceIndex, unsigned char data, unsigned int offset);
+void KeccakP1600times8_AddBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times8_AddLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times8_OverwriteBytes(void *states, unsigned int instanceIndex, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times8_OverwriteLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times8_OverwriteWithZeroes(void *states, unsigned int instanceIndex, unsigned int byteCount);
+void KeccakP1600times8_PermuteAll_4rounds(void *states);
+void KeccakP1600times8_PermuteAll_6rounds(void *states);
+void KeccakP1600times8_PermuteAll_12rounds(void *states);
+void KeccakP1600times8_PermuteAll_24rounds(void *states);
+void KeccakP1600times8_ExtractBytes(const void *states, unsigned int instanceIndex, unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600times8_ExtractLanesAll(const void *states, unsigned char *data, unsigned int laneCount, unsigned int laneOffset);
+void KeccakP1600times8_ExtractAndAddBytes(const void *states, unsigned int instanceIndex,  const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+void KeccakP1600times8_ExtractAndAddLanesAll(const void *states, const unsigned char *input, unsigned char *output, unsigned int laneCount, unsigned int laneOffset);
+
+#endif

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-times8-on2.c

@@ -0,0 +1,38 @@
+/*
+Implementation by the Keccak Team, namely, Guido Bertoni, Joan Daemen,
+Michaël Peeters, Gilles Van Assche and Ronny Van Keer,
+hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+
+---
+
+This file implements Keccak-p[1600]×8 in a PlSnP-compatible way.
+Please refer to PlSnP-documentation.h for more details.
+
+This implementation comes with KeccakP-1600-times8-SnP.h in the same folder.
+Please refer to LowLevel.build for the exact list of other files it must be combined with.
+*/
+
+#include "KeccakP-1600-times2-SnP.h"
+
+#define prefix                          KeccakP1600times8
+#define PlSnP_baseParallelism           2
+#define PlSnP_targetParallelism         8
+#define SnP_laneLengthInBytes           8
+#define SnP                             KeccakP1600times2
+#define SnP_PermuteAll                  KeccakP1600times2_PermuteAll_24rounds
+#define SnP_PermuteAll_12rounds         KeccakP1600times2_PermuteAll_12rounds
+#define SnP_PermuteAll_6rounds          KeccakP1600times2_PermuteAll_6rounds
+#define SnP_PermuteAll_4rounds          KeccakP1600times2_PermuteAll_4rounds
+#define PlSnP_PermuteAll                KeccakP1600times8_PermuteAll_24rounds
+#define PlSnP_PermuteAll_12rounds       KeccakP1600times8_PermuteAll_12rounds
+#define PlSnP_PermuteAll_6rounds        KeccakP1600times8_PermuteAll_6rounds
+#define PlSnP_PermuteAll_4rounds        KeccakP1600times8_PermuteAll_4rounds
+
+#include "PlSnP-Fallback.inc"

data/ext/digest/kangarootwelve/keccak/sandybridge/KeccakP-1600-unrolling.macros

@@ -0,0 +1,302 @@
+/*
+Implementation by the Keccak Team, namely, Guido Bertoni, Joan Daemen,
+Michaël Peeters, Gilles Van Assche and Ronny Van Keer,
+hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to our website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+*/
+
+#if (defined(FullUnrolling))
+#define rounds24 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta( 0, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 1, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 2, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 3, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 4, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 5, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 6, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 7, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 8, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 9, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(10, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(11, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(12, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(13, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(14, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(15, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(16, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(17, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds12 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(12, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(13, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(14, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(15, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(16, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(17, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds6 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds4 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#elif (Unrolling == 12)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i+=12) { \
+        thetaRhoPiChiIotaPrepareTheta(i   , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 2, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 3, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 4, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 5, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 6, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 7, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 8, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+ 9, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+10, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+11, E, A) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(12, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(13, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(14, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(15, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(16, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(17, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds6 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds4 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#elif (Unrolling == 6)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i+=6) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+4, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+5, E, A) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=12; i<24; i+=6) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+4, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+5, E, A) \
+    } \
+
+#define rounds6 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#define rounds4 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#elif (Unrolling == 4)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i+=4) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=12; i<24; i+=4) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
+    } \
+
+#define rounds6 \
+    prepareTheta \
+    for(i=18; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#define rounds4 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(21, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(22, A, E) \
+    thetaRhoPiChiIota(23, E, A) \
+
+#elif (Unrolling == 3)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i+=3) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=12; i<24; i+=3) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds6 \
+    prepareTheta \
+    for(i=18; i<24; i+=3) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+        thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds4 \
+    prepareTheta \
+    for(i=20; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#elif (Unrolling == 2)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=12; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#define rounds6 \
+    prepareTheta \
+    for(i=18; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#define rounds4 \
+    prepareTheta \
+    for(i=20; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#elif (Unrolling == 1)
+#define rounds24 \
+    prepareTheta \
+    for(i=0; i<24; i++) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=12; i<24; i++) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds6 \
+    prepareTheta \
+    for(i=18; i<24; i++) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#define rounds4 \
+    prepareTheta \
+    for(i=20; i<24; i++) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        copyStateVariables(A, E) \
+    } \
+
+#else
+#error "Unrolling is not correctly specified!"
+#endif
+
+#define roundsN(__nrounds) \
+    prepareTheta \
+    i = 24 - (__nrounds); \
+    if ((i&1) != 0) { \
+        thetaRhoPiChiIotaPrepareTheta(i, A, E) \
+        copyStateVariables(A, E) \
+        ++i; \
+    } \
+    for( /* empty */; i<24; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    }