@@ -217,7 +217,7 @@ void cpSMS4_CFB_dec_gfni512(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBlkS
217
217
int itr ;
218
218
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
219
219
/* initial xors */
220
- TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [0 ]);
220
+ TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [0 ]);
221
221
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [1 ] );
222
222
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [2 ] );
223
223
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [3 ] );
@@ -242,7 +242,7 @@ void cpSMS4_CFB_dec_gfni512(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBlkS
242
242
TMP [12 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [12 ], TMP [19 ]), L512 (TMP [19 ]));
243
243
244
244
/* initial xors */
245
- TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [1 ]);
245
+ TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [1 ]);
246
246
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [2 ] );
247
247
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [3 ] );
248
248
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [0 ] );
@@ -267,7 +267,7 @@ void cpSMS4_CFB_dec_gfni512(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBlkS
267
267
TMP [13 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [13 ], TMP [19 ]), L512 (TMP [19 ]));
268
268
269
269
/* initial xors */
270
- TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [2 ]);
270
+ TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [2 ]);
271
271
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [3 ] );
272
272
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [0 ] );
273
273
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [1 ] );
@@ -292,7 +292,7 @@ void cpSMS4_CFB_dec_gfni512(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBlkS
292
292
TMP [14 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [14 ], TMP [19 ]), L512 (TMP [19 ]));
293
293
294
294
/* initial xors */
295
- TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [3 ]);
295
+ TMP [19 ] = TMP [18 ] = TMP [17 ] = TMP [16 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [3 ]);
296
296
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [0 ] );
297
297
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [1 ] );
298
298
TMP [16 ] = _mm512_xor_si512 (TMP [16 ], TMP [2 ] );
@@ -512,7 +512,7 @@ void cpSMS4_CFB_dec_gfni512x48(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
512
512
int itr ;
513
513
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
514
514
/* initial xors */
515
- TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [0 ]);
515
+ TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [0 ]);
516
516
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [1 ] );
517
517
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [2 ] );
518
518
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [3 ] );
@@ -532,7 +532,7 @@ void cpSMS4_CFB_dec_gfni512x48(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
532
532
TMP [8 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [8 ], TMP [14 ]), L512 (TMP [14 ]));
533
533
534
534
/* initial xors */
535
- TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [1 ]);
535
+ TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [1 ]);
536
536
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [2 ] );
537
537
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [3 ] );
538
538
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [0 ] );
@@ -552,7 +552,7 @@ void cpSMS4_CFB_dec_gfni512x48(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
552
552
TMP [9 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [9 ], TMP [14 ]), L512 (TMP [14 ]));
553
553
554
554
/* initial xors */
555
- TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [2 ]);
555
+ TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [2 ]);
556
556
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [3 ] );
557
557
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [0 ] );
558
558
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [1 ] );
@@ -572,7 +572,7 @@ void cpSMS4_CFB_dec_gfni512x48(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
572
572
TMP [10 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [10 ], TMP [14 ]), L512 (TMP [14 ]));
573
573
574
574
/* initial xors */
575
- TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [3 ]);
575
+ TMP [14 ] = TMP [13 ] = TMP [12 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [3 ]);
576
576
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [0 ] );
577
577
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [1 ] );
578
578
TMP [12 ] = _mm512_xor_si512 (TMP [12 ], TMP [2 ] );
@@ -746,7 +746,7 @@ void cpSMS4_CFB_dec_gfni512x32(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
746
746
int itr ;
747
747
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
748
748
/* initial xors */
749
- TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [0 ]);
749
+ TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [0 ]);
750
750
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [1 ] );
751
751
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [2 ] );
752
752
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [3 ] );
@@ -761,7 +761,7 @@ void cpSMS4_CFB_dec_gfni512x32(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
761
761
TMP [4 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [4 ], TMP [9 ]), L512 (TMP [9 ]));
762
762
763
763
/* initial xors */
764
- TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [1 ]);
764
+ TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [1 ]);
765
765
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [2 ] );
766
766
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [3 ] );
767
767
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [0 ] );
@@ -776,7 +776,7 @@ void cpSMS4_CFB_dec_gfni512x32(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
776
776
TMP [5 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [5 ], TMP [9 ]), L512 (TMP [9 ]));
777
777
778
778
/* initial xors */
779
- TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [2 ]);
779
+ TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [2 ]);
780
780
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [3 ] );
781
781
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [0 ] );
782
782
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [1 ] );
@@ -791,7 +791,7 @@ void cpSMS4_CFB_dec_gfni512x32(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
791
791
TMP [6 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [6 ], TMP [9 ]), L512 (TMP [9 ]));
792
792
793
793
/* initial xors */
794
- TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [3 ]);
794
+ TMP [9 ] = TMP [8 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [3 ]);
795
795
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [0 ] );
796
796
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [1 ] );
797
797
TMP [8 ] = _mm512_xor_si512 (TMP [8 ], TMP [2 ] );
@@ -922,7 +922,7 @@ void cpSMS4_CFB_dec_gfni512x16(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
922
922
int itr ;
923
923
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
924
924
/* initial xors */
925
- TMP [4 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [0 ]);
925
+ TMP [4 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [0 ]);
926
926
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [1 ]);
927
927
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [2 ]);
928
928
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [3 ]);
@@ -932,7 +932,7 @@ void cpSMS4_CFB_dec_gfni512x16(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
932
932
TMP [0 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [0 ], TMP [4 ]), L512 (TMP [4 ]));
933
933
934
934
/* initial xors */
935
- TMP [4 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [1 ]);
935
+ TMP [4 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [1 ]);
936
936
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [2 ]);
937
937
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [3 ]);
938
938
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [0 ]);
@@ -942,7 +942,7 @@ void cpSMS4_CFB_dec_gfni512x16(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
942
942
TMP [1 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [1 ], TMP [4 ]), L512 (TMP [4 ]));
943
943
944
944
/* initial xors */
945
- TMP [4 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [2 ]);
945
+ TMP [4 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [2 ]);
946
946
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [3 ]);
947
947
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [0 ]);
948
948
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [1 ]);
@@ -952,7 +952,7 @@ void cpSMS4_CFB_dec_gfni512x16(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
952
952
TMP [2 ] = _mm512_xor_si512 (_mm512_xor_si512 (TMP [2 ], TMP [4 ]), L512 (TMP [4 ]));
953
953
954
954
/* initial xors */
955
- TMP [4 ] = _mm512_set1_epi32 ((Ipp16u )pRKey [3 ]);
955
+ TMP [4 ] = _mm512_set1_epi32 ((Ipp32s )pRKey [3 ]);
956
956
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [0 ]);
957
957
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [1 ]);
958
958
TMP [4 ] = _mm512_xor_si512 (TMP [4 ], TMP [2 ]);
@@ -1105,7 +1105,7 @@ void cpSMS4_CFB_dec_gfni128x12(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
1105
1105
int itr ;
1106
1106
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
1107
1107
/* initial xors */
1108
- TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [0 ]), 0 );
1108
+ TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [0 ]), 0 );
1109
1109
TMP [14 ] = TMP [13 ];
1110
1110
TMP [12 ] = TMP [13 ];
1111
1111
TMP [12 ] = _mm_xor_si128 (TMP [12 ], TMP [1 ] );
@@ -1127,7 +1127,7 @@ void cpSMS4_CFB_dec_gfni128x12(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
1127
1127
TMP [8 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [8 ], TMP [14 ]), L128 (TMP [14 ]));
1128
1128
1129
1129
/* initial xors */
1130
- TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [1 ]), 0 );
1130
+ TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [1 ]), 0 );
1131
1131
TMP [14 ] = TMP [13 ];
1132
1132
TMP [12 ] = TMP [13 ];
1133
1133
TMP [12 ] = _mm_xor_si128 (TMP [12 ], TMP [2 ] );
@@ -1149,7 +1149,7 @@ void cpSMS4_CFB_dec_gfni128x12(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
1149
1149
TMP [9 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [9 ], TMP [14 ]), L128 (TMP [14 ]));
1150
1150
1151
1151
/* initial xors */
1152
- TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [2 ]), 0 );
1152
+ TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [2 ]), 0 );
1153
1153
TMP [14 ] = TMP [13 ];
1154
1154
TMP [12 ] = TMP [13 ];
1155
1155
TMP [12 ] = _mm_xor_si128 (TMP [12 ], TMP [3 ] );
@@ -1171,7 +1171,7 @@ void cpSMS4_CFB_dec_gfni128x12(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbB
1171
1171
TMP [10 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [10 ], TMP [14 ]), L128 (TMP [14 ]));
1172
1172
1173
1173
/* initial xors */
1174
- TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [3 ]), 0 );
1174
+ TMP [13 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [3 ]), 0 );
1175
1175
TMP [14 ] = TMP [13 ];
1176
1176
TMP [12 ] = TMP [13 ];
1177
1177
TMP [12 ] = _mm_xor_si128 (TMP [12 ], TMP [0 ] );
@@ -1358,7 +1358,7 @@ void cpSMS4_CFB_dec_gfni128x8(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1358
1358
int itr ;
1359
1359
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
1360
1360
/* initial xors */
1361
- TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [0 ]), 0 );
1361
+ TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [0 ]), 0 );
1362
1362
TMP [8 ] = TMP [9 ];
1363
1363
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [1 ] );
1364
1364
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [2 ] );
@@ -1374,7 +1374,7 @@ void cpSMS4_CFB_dec_gfni128x8(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1374
1374
TMP [4 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [4 ], TMP [9 ]), L128 (TMP [9 ]));
1375
1375
1376
1376
/* initial xors */
1377
- TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [1 ]), 0 );
1377
+ TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [1 ]), 0 );
1378
1378
TMP [8 ] = TMP [9 ];
1379
1379
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [2 ] );
1380
1380
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [3 ] );
@@ -1390,7 +1390,7 @@ void cpSMS4_CFB_dec_gfni128x8(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1390
1390
TMP [5 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [5 ], TMP [9 ]), L128 (TMP [9 ]));
1391
1391
1392
1392
/* initial xors */
1393
- TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [2 ]), 0 );
1393
+ TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [2 ]), 0 );
1394
1394
TMP [8 ] = TMP [9 ];
1395
1395
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [3 ] );
1396
1396
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [0 ] );
@@ -1406,7 +1406,7 @@ void cpSMS4_CFB_dec_gfni128x8(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1406
1406
TMP [6 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [6 ], TMP [9 ]), L128 (TMP [9 ]));
1407
1407
1408
1408
/* initial xors */
1409
- TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [3 ]), 0 );
1409
+ TMP [9 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [3 ]), 0 );
1410
1410
TMP [8 ] = TMP [9 ];
1411
1411
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [0 ] );
1412
1412
TMP [8 ] = _mm_xor_si128 (TMP [8 ], TMP [1 ] );
@@ -1541,7 +1541,7 @@ void cpSMS4_CFB_dec_gfni128x4(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1541
1541
int itr ;
1542
1542
for (itr = 0 ; itr < 8 ; itr ++ , pRKey += 4 ) {
1543
1543
/* initial xors */
1544
- TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [0 ]), 0 );
1544
+ TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [0 ]), 0 );
1545
1545
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [1 ] );
1546
1546
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [2 ] );
1547
1547
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [3 ] );
@@ -1551,7 +1551,7 @@ void cpSMS4_CFB_dec_gfni128x4(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1551
1551
TMP [0 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [0 ], TMP [4 ]), L128 (TMP [4 ]));
1552
1552
1553
1553
/* initial xors */
1554
- TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [1 ]), 0 );
1554
+ TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [1 ]), 0 );
1555
1555
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [2 ] );
1556
1556
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [3 ] );
1557
1557
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [0 ] );
@@ -1561,7 +1561,7 @@ void cpSMS4_CFB_dec_gfni128x4(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1561
1561
TMP [1 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [1 ], TMP [4 ]), L128 (TMP [4 ]));
1562
1562
1563
1563
/* initial xors */
1564
- TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [2 ]), 0 );
1564
+ TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [2 ]), 0 );
1565
1565
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [3 ] );
1566
1566
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [0 ] );
1567
1567
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [1 ] );
@@ -1571,7 +1571,7 @@ void cpSMS4_CFB_dec_gfni128x4(Ipp8u* pDst, const Ipp8u* pSrc, int len, int cfbBl
1571
1571
TMP [2 ] = _mm_xor_si128 (_mm_xor_si128 (TMP [2 ], TMP [4 ]), L128 (TMP [4 ]));
1572
1572
1573
1573
/* initial xors */
1574
- TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp16u )pRKey [3 ]), 0 );
1574
+ TMP [4 ] = _mm_shuffle_epi32 (_mm_cvtsi32_si128 ((Ipp32s )pRKey [3 ]), 0 );
1575
1575
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [0 ] );
1576
1576
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [1 ] );
1577
1577
TMP [4 ] = _mm_xor_si128 (TMP [4 ], TMP [2 ] );
0 commit comments