Skip to content

lwc / candidates

Go to a project
Commit 727508dc by Wentao Zhang Committed by Sebastian Renner
Options

knot

parent 20add87b
Expand all
Inline Side-by-side
Showing with 1624 additions and 2186 deletions
knot/Implementations/crypto_aead/knot128v2/armcortexm_2/auxFormat.c
#include"auxFormat.h"
void packU96FormatToThreePacket(u32 *out, u8 *in) {
u32 t0 = U32BIG(((u32*)in)[0]), t1 = U32BIG(((u32*)in)[1]), t2 = U32BIG(((u32*)in)[2]); \
puckU32ToThree_3(t0); \
puckU32ToThree_3(t1); \
puckU32ToThree_3(t2); \
//puck begin//
void unpackU96FormatToThreePacket(u8 * out, u32 * in) {
u32 temp0[3] = { 0 };
u32 temp1[3] = { 0 };
u32 temp2[3] = { 0 };
u32 t1_32, t2_64, t2_65;
u32 t[3] = { 0 };
temp0[0] = in[0] & 0xffe00000;
temp1[0] = (in[0] & 0x001ffc00) << 11;
temp2[0] = (in[0] & 0x000003ff) << 22;
temp0[1] = in[1] & 0xffe00000;
temp1[1] = (in[1] & 0x001ff800) << 11;
t2_64 = ((in[1] & 0x00000400) << 21);
temp2[1] = (in[1] & 0x000003ff) << 22;
temp0[2] = in[2] & 0xffc00000;
t1_32 = ((in[2] & 0x00200000) << 10);
temp1[2] = (in[2] & 0x001ff800) << 11;
t2_65 = ((in[2] & 0x00000400) << 20);
temp2[2] = (in[2] & 0x000003ff) << 22;
unpuckU32ToThree(temp0[0]);
unpuckU32ToThree(temp0[1]);
unpuckU32ToThree(temp0[2]);
t[2] = temp0[0] | temp0[1] >> 1 | temp0[2] >> 2;
unpuckU32ToThree(temp1[0]);
unpuckU32ToThree(temp1[1]);
unpuckU32ToThree(temp1[2]);
t[1] = t1_32 | ((temp1[0] | temp1[1] >> 1 | temp1[2] >> 2) >> 1);
unpuckU32ToThree(temp2[0]);
unpuckU32ToThree(temp2[1]);
unpuckU32ToThree(temp2[2]);
t[0] = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2);
memcpy(out, t, 12 * sizeof(unsigned char));
out[0] = ((((t2 >> 21) & 0x400) | (((t2 >> 18) & 0x300)) | ((t2 & 0xff00) >> 8)) << 21) |
((((t1 >> 20) & 0x400) | ((t1 >> 16) & 0x300) | (t1 & 0xff)) << 10) |
(((t0 >> 20) & 0x300) | ((t0 & 0xff0000) >> 16));
out[1] = ((((t2 >> 20) & 0x400) | ((t2 >> 16) & 0x300) | (t2 & 0xff)) << 21) |
((((t1 >> 20) & 0x300) | ((t1 & 0xff0000) >> 16)) << 11) |
(((t0 >> 21) & 0x400) | ((t0 >> 18) & 0x300) | ((t0 & 0xff00) >> 8));
out[2] = ((((t2 >> 20) & 0x300) | ((t2 & 0xff0000) >> 16)) << 22) |
((((t1 >> 21) & 0x400) | ((t1 >> 18) & 0x300)|((t1 & 0xff00) >> 8)) << 11) |
(((t0 >> 20) & 0x400) | ((t0 >> 16) & 0x300) | (t0 & 0xff));
}
void packU96FormatToThreePacket(u32 * out, u8 * in) {
u32 t0 = U32BIG(((u32*)in)[2]);
u32 t1 = U32BIG(((u32*)in)[1]);
u32 t2 = U32BIG(((u32*)in)[0]);
u32 temp0[3] = { 0 };
u32 temp1[3] = { 0 };
u32 temp2[3] = { 0 };
u8 t1_32 = (in[7] & 0x80) >> 7, t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6;
t1 = t1 << 1;
t2 = t2 << 2;
temp0[0] = t0; temp0[1] = t0 << 1; temp0[2] = t0 << 2;
puckU32ToThree(temp0[0]);
puckU32ToThree(temp0[1]);
puckU32ToThree(temp0[2]);
temp1[0] = t1; temp1[1] = t1 << 1; temp1[2] = t1 << 2;
puckU32ToThree(temp1[0]);
puckU32ToThree(temp1[1]);
puckU32ToThree(temp1[2]);
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2;
puckU32ToThree(temp2[0]);
puckU32ToThree(temp2[1]);
puckU32ToThree(temp2[2]);
out[0] = (temp0[0]) | (temp1[0] >> 11) | (temp2[0] >> 22);
out[1] = (temp0[1]) | (temp1[1] >> 11) | (((u32)t2_64) << 10) | (temp2[1] >> 22);
out[2] = (temp0[2]) | (((u32)t1_32) << 21) | (temp1[2] >> 11) | (((u32)t2_65) << 10) | (temp2[2] >> 22);
}
void packU32FormatToThreePacket(u32 * out, u8 * in) {
u32 t2 = U32BIG(((u32*)in)[0]);
u32 temp2[3] = { 0 };
u8 t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6;
t2 = t2 << 2;
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2;
puckU32ToThree(temp2[0]);
puckU32ToThree(temp2[1]);
puckU32ToThree(temp2[2]);
out[0] = (temp2[0] >> 22);
out[1] = (((u32)t2_64) << 10) | (temp2[1] >> 22);
out[2] =(((u32)t2_65) << 10) | (temp2[2] >> 22);
}
void unpackU32FormatToThreePacket(u8 * out, u32 * in) {
u32 temp2[3] = { 0 };
u32 t2_64, t2_65;
u32 t2;
temp2[0] = (in[0] & 0x000003ff) << 22;
t2_64 = ((in[1] & 0x00000400) << 21);
temp2[1] = (in[1] & 0x000003ff) << 22;
t2_65 = ((in[2] & 0x00000400) << 20);
temp2[2] = (in[2] & 0x000003ff) << 22;
unpuckU32ToThree(temp2[0]);
unpuckU32ToThree(temp2[1]);
unpuckU32ToThree(temp2[2]);
t2 = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2);
*(u32*)(out) = U32BIG(t2);
void unpackU96FormatToThreePacket(u8 *out, u32 *in) {
u32 t[3] = { 0 } ;\
u32 t0 = in[0], t1 = in[1], t2 = in[2]; \
t[0] = ((t1 & 0x400) << 21) | ((t2 & 0x400) << 20) | \
((t0 & 0x300) << 20) | ((t1 & 0x300) << 18) | ((t2 & 0x300) << 16) | \
((t0 & 0xff) << 16) | ((t1 & 0xff) << 8) | (t2 & 0xff); \
t[1] = ((t2 & 0x200000) << 10) | ((t0 & 0x100000) << 10) | ((t1 & 0x180000) << 9) | ((t2 & 0x180000) << 7) | \
((t0 & 0xc0000) << 6) | ((t1 & 0x7f800) << 5) | ((t2 & 0x7f800) >> 3) | \
((t0 & 0x3fc00) >> 10); \
t[2] = ((t0 & 0x80000000)) | ((t1 & 0x80000000) >> 1) | ((t2 & 0xc0000000) >> 2) | \
((t0 & 0x60000000) >> 3) | ((t1 & 0x60000000) >> 5) | ((t2 & 0x3fc00000) >> 6) | \
((t0 & 0x1fe00000) >> 13) | ((t1 & 0x1fe00000) >> 21); \
unpuckU32ToThree_3(t[0]); \
unpuckU32ToThree_3(t[1]); \
unpuckU32ToThree_3(t[2]); \
memcpy(out, t, 12 * sizeof(unsigned char)); \
}
void P384(unsigned int *s, unsigned char *round, unsigned char lunNum) {
u32 rci,t1,t2;
......
knot/Implementations/crypto_aead/knot128v2/armcortexm_2/auxFormat.h
......@@ -14,26 +14,40 @@ typedef unsigned long long u64;
#define PR0_ROUNDS 76
#define PR_ROUNDS 28
#define PRF_ROUNDS 32
/*
#define PR0_ROUNDS 76
#define PR_ROUNDS 40
#define PRF_ROUNDS 44
#define PR0_ROUNDS 76
#define PR_ROUNDS 28
#define PRF_ROUNDS 32
* */
#define ARR_SIZE(a) (sizeof((a))/sizeof((a[0])))
#define LOTR32(x,n) (((x)<<(n))|((x)>>(32-(n))))
void packU96FormatToThreePacket(u32 * out, u8 * in);
void unpackU96FormatToThreePacket(u8 * out, u32 * in);
void P384(unsigned int *s, unsigned char *round, unsigned char lunNum);
//////////////////puck begin
//&:5 <<:4 |:4
#define puckU32ToThree(x){\
x &= 0x92492492;\
x = (x | (x << 2)) & 0xc30c30c3;\
x = (x | (x << 4)) & 0xf00f00f0;\
x = (x | (x << 8)) & 0xff0000ff;\
x = (x | (x << 16)) & 0xfff00000;\
#define puckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
}
#define unpuckU32ToThree(x){\
x &= 0xfff00000;\
x = (x | (x >> 16)) & 0xff0000ff;\
x = (x | (x >> 8)) & 0xf00f00f0;\
x = (x | (x >> 4)) & 0xc30c30c3;\
x = (x | (x >> 2)) & 0x92492492;\
#define unpuckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
}
unsigned char constant7Format[80];
......
knot/Implementations/crypto_aead/knot128v2/armcortexm_2/encrypt.c
......@@ -96,7 +96,7 @@ void ProcessCiphertext(u32 *s, unsigned char *m, const unsigned char *c, unsigne
{
u32 dataFormat[12] = { 0 };
u32 dataFormat_1[12] = { 0 };
u8 tempU8[24] = { 0 },i;
u8 tempU8[24] = { 0 },tempData[24] = { 0 };
if (clen) {
while (clen >= aead_RATE) {
packU96FormatToThreePacket(dataFormat, c);
......@@ -123,12 +123,18 @@ void ProcessCiphertext(u32 *s, unsigned char *m, const unsigned char *c, unsigne
}
unpackU96FormatToThreePacket(tempU8, s);
unpackU96FormatToThreePacket(tempU8 + 12, s + 3);
for (i = 0; i < clen; ++i, ++m, ++c)
{
*m = tempU8[i] ^ *c;
tempU8[i] = *c;
}
tempU8[i] ^= 0x01;
memset(tempData, 0, sizeof(tempData));
memcpy(tempData, c, clen * sizeof(unsigned char));
tempData[clen] = 0x01;
U32BIG(((u32*)tempU8)[0]) ^= U32BIG(((u32* )tempData)[0]);
U32BIG(((u32*)tempU8)[1]) ^= U32BIG(((u32* )tempData)[1]);
U32BIG(((u32*)tempU8)[2]) ^= U32BIG(((u32* )tempData)[2]);
U32BIG(((u32*)tempU8)[3]) ^= U32BIG(((u32* )tempData)[3]);
U32BIG(((u32*)tempU8)[4]) ^= U32BIG(((u32* )tempData)[4]);
U32BIG(((u32*)tempU8)[5]) ^= U32BIG(((u32* )tempData)[5]);
memcpy(m, tempU8, clen * sizeof(unsigned char));
memcpy(tempU8, tempData, clen * sizeof(unsigned char));
c += clen;
packU96FormatToThreePacket(s, tempU8);
packU96FormatToThreePacket(s + 3, tempU8 + 12);
}
......
knot/Implementations/crypto_aead/knot128v2/armcortexm_3/auxFormat.h
......@@ -12,99 +12,6 @@ typedef unsigned long long u64;
#define ARR_SIZE(a) (sizeof((a))/sizeof((a[0])))
#define LOTR32(x,n) (((x)<<(n))|((x)>>(32-(n))))
#define puckU32ToThree(x){\
x &= 0x92492492;\
x = (x | (x << 2)) & 0xc30c30c3;\
x = (x | (x << 4)) & 0xf00f00f0;\
x = (x | (x << 8)) & 0xff0000ff;\
x = (x | (x << 16)) & 0xfff00000;\
}
#define unpuckU32ToThree(x){\
x &= 0xfff00000;\
x = (x | (x >> 16)) & 0xff0000ff;\
x = (x | (x >> 8)) & 0xf00f00f0;\
x = (x | (x >> 4)) & 0xc30c30c3;\
x = (x | (x >> 2)) & 0x92492492;\
}
#define packU32FormatToThreePacket( out, in) {\
t2 = U32BIG(((u32*)in)[0]); \
t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6; \
t2 = t2 << 2; \
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2; \
puckU32ToThree(temp2[0]); \
puckU32ToThree(temp2[1]); \
puckU32ToThree(temp2[2]); \
out[0] = (temp2[0] >> 22); \
out[1] = (((u32)t2_64) << 10) | (temp2[1] >> 22); \
out[2] =(((u32)t2_65) << 10) | (temp2[2] >> 22); \
}
#define packU96FormatToThreePacket(out, in) {\
t9 = U32BIG(((u32*)in)[2]); \
t1 = U32BIG(((u32*)in)[1]); \
t2 = U32BIG(((u32*)in)[0]); \
t1_32 = (in[7] & 0x80) >> 7, t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6; \
t1 = t1 << 1; \
t2 = t2 << 2; \
temp0[0] = t9; temp0[1] = t9 << 1; temp0[2] = t9 << 2; \
puckU32ToThree(temp0[0]); \
puckU32ToThree(temp0[1]); \
puckU32ToThree(temp0[2]); \
temp1[0] = t1; temp1[1] = t1 << 1; temp1[2] = t1 << 2; \
puckU32ToThree(temp1[0]); \
puckU32ToThree(temp1[1]); \
puckU32ToThree(temp1[2]); \
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2; \
puckU32ToThree(temp2[0]); \
puckU32ToThree(temp2[1]); \
puckU32ToThree(temp2[2]); \
out[0] = (temp0[0]) | (temp1[0] >> 11) | (temp2[0] >> 22); \
out[1] = (temp0[1]) | (temp1[1] >> 11) | (((u32)t2_64) << 10) | (temp2[1] >> 22); \
out[2] = (temp0[2]) | (((u32)t1_32) << 21) | (temp1[2] >> 11) | (((u32)t2_65) << 10) | (temp2[2] >> 22); \
}
#define unpackU32FormatToThreePacket(out, in) {\
temp2[0] = (in[0] & 0x000003ff) << 22; \
t2_64 = ((in[1] & 0x00000400) << 21); \
temp2[1] = (in[1] & 0x000003ff) << 22; \
t2_65 = ((in[2] & 0x00000400) << 20); \
temp2[2] = (in[2] & 0x000003ff) << 22; \
unpuckU32ToThree(temp2[0]); \
unpuckU32ToThree(temp2[1]); \
unpuckU32ToThree(temp2[2]); \
t2 = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2); \
*(u32*)(out) = U32BIG(t2); \
}
#define unpackU96FormatToThreePacket( out, in) {\
temp0[0] = in[0] & 0xffe00000; \
temp1[0] = (in[0] & 0x001ffc00) << 11; \
temp2[0] = (in[0] & 0x000003ff) << 22; \
temp0[1] = in[1] & 0xffe00000; \
temp1[1] = (in[1] & 0x001ff800) << 11; \
t2_64 = ((in[1] & 0x00000400) << 21); \
temp2[1] = (in[1] & 0x000003ff) << 22; \
temp0[2] = in[2] & 0xffc00000; \
t1_32 = ((in[2] & 0x00200000) << 10); \
temp1[2] = (in[2] & 0x001ff800) << 11; \
t2_65 = ((in[2] & 0x00000400) << 20); \
temp2[2] = (in[2] & 0x000003ff) << 22; \
unpuckU32ToThree(temp0[0]); \
unpuckU32ToThree(temp0[1]); \
unpuckU32ToThree(temp0[2]); \
t9 = temp0[0] | temp0[1] >> 1 | temp0[2] >> 2; \
unpuckU32ToThree(temp1[0]); \
unpuckU32ToThree(temp1[1]); \
unpuckU32ToThree(temp1[2]); \
t1 = t1_32 | ((temp1[0] | temp1[1] >> 1 | temp1[2] >> 2) >> 1); \
unpuckU32ToThree(temp2[0]); \
unpuckU32ToThree(temp2[1]); \
unpuckU32ToThree(temp2[2]); \
t2 = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2); \
*(u32*)(out) = U32BIG(t2); \
*(u32*)(out + 4) = U32BIG(t1); \
*(u32*)(out + 8) = U32BIG(t9); \
}
#define ARR_SIZE(a) (sizeof((a))/sizeof((a[0])))
#define sbox(a, b, c, d, f, g, h) \
{ \
t1 = ~a; t2 = b & t1;t3 = c ^ t2; h = d ^ t3; t5 = b | c; t6 = d ^ t1; g = t5 ^ t6; t8 = b ^ d; t9 = t3 & t6; a = t8 ^ t9; t11 = g & t8; f = t3 ^ t11; \
......@@ -120,11 +27,86 @@ t3= LOTR32(t1, 18); \
t4 = LOTR32(t2, 18);\
t5 = LOTR32(t0, 19); \
}
/*
s0 s1 s2
s3 s4 s5
s6 s7 s8
s9 s10 s11
*/
#define ROUND384(lunNum) {\
s[0] ^= (constant7Format[lunNum] >> 6) & 0x3;\
s[1] ^= (constant7Format[lunNum] >> 3) & 0x7;\
s[2] ^= constant7Format[lunNum] & 0x7;\
sbox(s[0], s[3], s[6], s[9] , s_temp[3], s_temp[6], s_temp[9]);\
sbox(s[1], s[4], s[7], s[10], s[3] , s_temp[7], s_temp[10]);\
sbox(s[2], s[5], s[8], s[11], s[4] , s_temp[8], s_temp[11]);\
s[5] = LOTR32(s_temp[3], 1); \
U96_BIT_LOTR32_8(s_temp[6], s_temp [7], s_temp[ 8], s[6], s[7], s[8]);\
U96_BIT_LOTR32_55(s_temp[9], s_temp[10], s_temp[11], s[9], s[10], s[11]);\
}
//Processing_Data:
#define Processing_Data(data) \
do { \
packU96FormatToThreePacket(dataFormat, data); \
s[0] ^= dataFormat[0]; \
s[1] ^= dataFormat[1]; \
s[2] ^= dataFormat[2]; \
packU96FormatToThreePacket((dataFormat + 3), (data + 12)); \
s[3] ^= dataFormat[3]; \
s[4] ^= dataFormat[4]; \
s[5] ^= dataFormat[5]; \
} while (0)
#define puckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
}
#define unpuckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
}
#define packU96FormatToThreePacket( out, in) { \
u32 t0 = U32BIG(((u32*)in)[0]), t1 = U32BIG(((u32*)in)[1]), t2 = U32BIG(((u32*)in)[2]); \
puckU32ToThree_3(t0); \
puckU32ToThree_3(t1); \
puckU32ToThree_3(t2); \
out[0] = ((((t2 >> 21) & 0x400) | (((t2 >> 18) & 0x300)) | ((t2 & 0xff00) >> 8)) << 21) | \
((((t1 >> 20) & 0x400) | ((t1 >> 16) & 0x300) | (t1 & 0xff)) << 10) | \
(((t0 >> 20) & 0x300) | ((t0 & 0xff0000) >> 16)); \
out[1] = ((((t2 >> 20) & 0x400) | ((t2 >> 16) & 0x300) | (t2 & 0xff)) << 21) | \
((((t1 >> 20) & 0x300) | ((t1 & 0xff0000) >> 16)) << 11) | \
(((t0 >> 21) & 0x400) | ((t0 >> 18) & 0x300) | ((t0 & 0xff00) >> 8)); \
out[2] = ((((t2 >> 20) & 0x300) | ((t2 & 0xff0000) >> 16)) << 22) | \
((((t1 >> 21) & 0x400) | ((t1 >> 18) & 0x300)|((t1 & 0xff00) >> 8)) << 11) | \
(((t0 >> 20) & 0x400) | ((t0 >> 16) & 0x300) | (t0 & 0xff)); \
}
#define unpackU96FormatToThreePacket( out, in) {\
u32 t[3] = { 0 };\
u32 t0 = in[0], t1 = in[1], t2 = in[2]; \
t[0] = ((t1 & 0x400) << 21) | ((t2 & 0x400) << 20) | \
((t0 & 0x300) << 20) | ((t1 & 0x300) << 18) | ((t2 & 0x300) << 16) | \
((t0 & 0xff) << 16) | ((t1 & 0xff) << 8) | (t2 & 0xff); \
t[1] = ((t2 & 0x200000) << 10) | ((t0 & 0x100000) << 10) | ((t1 & 0x180000) << 9) | ((t2 & 0x180000) << 7) | \
((t0 & 0xc0000) << 6) | ((t1 & 0x7f800) << 5) | ((t2 & 0x7f800) >> 3) | \
((t0 & 0x3fc00) >> 10); \
t[2] = ((t0 & 0x80000000)) | ((t1 & 0x80000000) >> 1) | ((t2 & 0xc0000000) >> 2) | \
((t0 & 0x60000000) >> 3) | ((t1 & 0x60000000) >> 5) | ((t2 & 0x3fc00000) >> 6) | \
((t0 & 0x1fe00000) >> 13) | ((t1 & 0x1fe00000) >> 21); \
unpuckU32ToThree_3(t[0]); \
unpuckU32ToThree_3(t[1]); \
unpuckU32ToThree_3(t[2]); \
memcpy(out, t, 12 * sizeof(unsigned char)); \
}
knot/Implementations/crypto_aead/knot128v2/armcortexm_4/auxFormat.c
#include"auxFormat.h"
void packU96FormatToThreePacket(u32 *out, u8 *in) {
u32 t0 = U32BIG(((u32*)in)[0]), t1 = U32BIG(((u32*)in)[1]), t2 = U32BIG(((u32*)in)[2]); \
puckU32ToThree_3(t0); \
puckU32ToThree_3(t1); \
puckU32ToThree_3(t2); \
out[0] = ((((t2 >> 21) & 0x400) | (((t2 >> 18) & 0x300)) | ((t2 & 0xff00) >> 8)) << 21) | \
((((t1 >> 20) & 0x400) | ((t1 >> 16) & 0x300) | (t1 & 0xff)) << 10) | \
(((t0 >> 20) & 0x300) | ((t0 & 0xff0000) >> 16)); \
out[1] = ((((t2 >> 20) & 0x400) | ((t2 >> 16) & 0x300) | (t2 & 0xff)) << 21) | \
((((t1 >> 20) & 0x300) | ((t1 & 0xff0000) >> 16)) << 11) | \
(((t0 >> 21) & 0x400) | ((t0 >> 18) & 0x300) | ((t0 & 0xff00) >> 8)); \
out[2] = ((((t2 >> 20) & 0x300) | ((t2 & 0xff0000) >> 16)) << 22) | \
((((t1 >> 21) & 0x400) | ((t1 >> 18) & 0x300)|((t1 & 0xff00) >> 8)) << 11) | \
(((t0 >> 20) & 0x400) | ((t0 >> 16) & 0x300) | (t0 & 0xff)); \
//puck begin//
void unpackU96FormatToThreePacket(u8 * out, u32 * in) {
u32 temp0[3] = { 0 };
u32 temp1[3] = { 0 };
u32 temp2[3] = { 0 };
u32 t1_32, t2_64, t2_65;
u32 t[3] = { 0 };
temp0[0] = in[0] & 0xffe00000;
temp1[0] = (in[0] & 0x001ffc00) << 11;
temp2[0] = (in[0] & 0x000003ff) << 22;
temp0[1] = in[1] & 0xffe00000;
temp1[1] = (in[1] & 0x001ff800) << 11;
t2_64 = ((in[1] & 0x00000400) << 21);
temp2[1] = (in[1] & 0x000003ff) << 22;
temp0[2] = in[2] & 0xffc00000;
t1_32 = ((in[2] & 0x00200000) << 10);
temp1[2] = (in[2] & 0x001ff800) << 11;
t2_65 = ((in[2] & 0x00000400) << 20);
temp2[2] = (in[2] & 0x000003ff) << 22;
unpuckU32ToThree(temp0[0]);
unpuckU32ToThree(temp0[1]);
unpuckU32ToThree(temp0[2]);
t[2] = temp0[0] | temp0[1] >> 1 | temp0[2] >> 2;
unpuckU32ToThree(temp1[0]);
unpuckU32ToThree(temp1[1]);
unpuckU32ToThree(temp1[2]);
t[1] = t1_32 | ((temp1[0] | temp1[1] >> 1 | temp1[2] >> 2) >> 1);
unpuckU32ToThree(temp2[0]);
unpuckU32ToThree(temp2[1]);
unpuckU32ToThree(temp2[2]);
t[0] = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2);
memcpy(out, t, 12 * sizeof(unsigned char));
}
void packU96FormatToThreePacket(u32 * out, u8 * in) {
u32 t0 = U32BIG(((u32*)in)[2]);
u32 t1 = U32BIG(((u32*)in)[1]);
u32 t2 = U32BIG(((u32*)in)[0]);
u32 temp0[3] = { 0 };
u32 temp1[3] = { 0 };
u32 temp2[3] = { 0 };
u8 t1_32 = (in[7] & 0x80) >> 7, t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6;
t1 = t1 << 1;
t2 = t2 << 2;
temp0[0] = t0; temp0[1] = t0 << 1; temp0[2] = t0 << 2;
puckU32ToThree(temp0[0]);
puckU32ToThree(temp0[1]);
puckU32ToThree(temp0[2]);
temp1[0] = t1; temp1[1] = t1 << 1; temp1[2] = t1 << 2;
puckU32ToThree(temp1[0]);
puckU32ToThree(temp1[1]);
puckU32ToThree(temp1[2]);
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2;
puckU32ToThree(temp2[0]);
puckU32ToThree(temp2[1]);
puckU32ToThree(temp2[2]);
out[0] = (temp0[0]) | (temp1[0] >> 11) | (temp2[0] >> 22);
out[1] = (temp0[1]) | (temp1[1] >> 11) | (((u32)t2_64) << 10) | (temp2[1] >> 22);
out[2] = (temp0[2]) | (((u32)t1_32) << 21) | (temp1[2] >> 11) | (((u32)t2_65) << 10) | (temp2[2] >> 22);
}
void packU32FormatToThreePacket(u32 * out, u8 * in) {
u32 t2 = U32BIG(((u32*)in)[0]);
u32 temp2[3] = { 0 };
u8 t2_64 = (in[3] & 0x80) >> 7, t2_65 = (in[3] & 0x40) >> 6;
t2 = t2 << 2;
temp2[0] = t2; temp2[1] = t2 << 1; temp2[2] = t2 << 2;
puckU32ToThree(temp2[0]);
puckU32ToThree(temp2[1]);
puckU32ToThree(temp2[2]);
out[0] = (temp2[0] >> 22);
out[1] = (((u32)t2_64) << 10) | (temp2[1] >> 22);
out[2] =(((u32)t2_65) << 10) | (temp2[2] >> 22);
}
void unpackU32FormatToThreePacket(u8 * out, u32 * in) {
u32 temp2[3] = { 0 };
u32 t2_64, t2_65;
u32 t2;
temp2[0] = (in[0] & 0x000003ff) << 22;
t2_64 = ((in[1] & 0x00000400) << 21);
temp2[1] = (in[1] & 0x000003ff) << 22;
t2_65 = ((in[2] & 0x00000400) << 20);
temp2[2] = (in[2] & 0x000003ff) << 22;
unpuckU32ToThree(temp2[0]);
unpuckU32ToThree(temp2[1]);
unpuckU32ToThree(temp2[2]);
t2 = t2_65 | t2_64 | ((temp2[0] | temp2[1] >> 1 | temp2[2] >> 2) >> 2);
*(u32*)(out) = U32BIG(t2);
void unpackU96FormatToThreePacket(u8 *out, u32 *in) {
u32 t[3] = { 0 } ;\
u32 t0 = in[0], t1 = in[1], t2 = in[2]; \
t[0] = ((t1 & 0x400) << 21) | ((t2 & 0x400) << 20) | \
((t0 & 0x300) << 20) | ((t1 & 0x300) << 18) | ((t2 & 0x300) << 16) | \
((t0 & 0xff) << 16) | ((t1 & 0xff) << 8) | (t2 & 0xff); \
t[1] = ((t2 & 0x200000) << 10) | ((t0 & 0x100000) << 10) | ((t1 & 0x180000) << 9) | ((t2 & 0x180000) << 7) | \
((t0 & 0xc0000) << 6) | ((t1 & 0x7f800) << 5) | ((t2 & 0x7f800) >> 3) | \
((t0 & 0x3fc00) >> 10); \
t[2] = ((t0 & 0x80000000)) | ((t1 & 0x80000000) >> 1) | ((t2 & 0xc0000000) >> 2) | \
((t0 & 0x60000000) >> 3) | ((t1 & 0x60000000) >> 5) | ((t2 & 0x3fc00000) >> 6) | \
((t0 & 0x1fe00000) >> 13) | ((t1 & 0x1fe00000) >> 21); \
unpuckU32ToThree_3(t[0]); \
unpuckU32ToThree_3(t[1]); \
unpuckU32ToThree_3(t[2]); \
memcpy(out, t, 12 * sizeof(unsigned char)); \
}
void P384(unsigned int *s, unsigned char *round, unsigned char lunNum) {
u32 s_temp[12] = { 0 };
......
knot/Implementations/crypto_aead/knot128v2/armcortexm_4/auxFormat.h
......@@ -14,7 +14,14 @@ typedef unsigned long long u64;
#define PR0_ROUNDS 76
#define PR_ROUNDS 28
#define PRF_ROUNDS 32
/*
#define PR0_ROUNDS 76
#define PR_ROUNDS 40
#define PRF_ROUNDS 44
* */
#define ARR_SIZE(a) (sizeof((a))/sizeof((a[0])))
#define LOTR32(x,n) (((x)<<(n))|((x)>>(32-(n))))
......@@ -35,24 +42,26 @@ t4 = LOTR32(t2, 18);\
t5 = LOTR32(t0, 19); \
}
//////////////////puck begin
//&:5 <<:4 |:4
#define puckU32ToThree(x){\
x &= 0x92492492;\
x = (x | (x << 2)) & 0xc30c30c3;\
x = (x | (x << 4)) & 0xf00f00f0;\
x = (x | (x << 8)) & 0xff0000ff;\
x = (x | (x << 16)) & 0xfff00000;\
}
#define unpuckU32ToThree(x){\
x &= 0xfff00000;\
x = (x | (x >> 16)) & 0xff0000ff;\
x = (x | (x >> 8)) & 0xf00f00f0;\
x = (x | (x >> 4)) & 0xc30c30c3;\
x = (x | (x >> 2)) & 0x92492492;\
}
void packU96FormatToThreePacket(u32 * out, u8 * in);
void unpackU96FormatToThreePacket(u8 * out, u32 * in);
unsigned char constant7Format[80];
#define puckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
}
#define unpuckU32ToThree_3(lo){\
u32 r0;\
r0 = (lo ^ (lo << 12)) & 0x000f0000, lo ^= r0 ^ (r0 >> 12);\
r0 = (lo ^ (lo << 4)) & 0x000f0f00, lo ^= r0 ^ (r0 >> 4);\
r0 = (lo ^ (lo << 6)) & 0x00300300, lo ^= r0 ^ (r0 >> 6);\
r0 = (lo ^ (lo << 2)) & 0x00330330, lo ^= r0 ^ (r0 >> 2);\
r0 = (lo ^ (lo << 3)) & 0x10410410, lo ^= r0 ^ (r0 >> 3);\
r0 = (lo ^ (lo << 1)) & 0x14514514, lo ^= r0 ^ (r0 >> 1);\
}
knot/Implementations/crypto_aead/knot128v2/armcortexm_4/encrypt.c
......@@ -96,7 +96,7 @@ void ProcessCiphertext(u32 *s, unsigned char *m, const unsigned char *c, unsigne
{
u32 dataFormat[12] = { 0 };
u32 dataFormat_1[12] = { 0 };
u8 tempU8[24] = { 0 },i;
u8 tempU8[24] = { 0 },tempData[24] = { 0 };
if (clen) {