fumiama/blake2b-simd
1
0
Fork 0
mirror of https://github.com/fumiama/blake2b-simd.git synced 2026-06-05 02:00:26 +08:00
Code Activity

G1 macro for Round 2

Browse Source
This commit is contained in:
frankw
2016-06-26 19:45:07 +02:00
parent e1903c8762
commit 55b55e023b
2 changed files with 63 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
block.go
View File

@@ -152,34 +152,34 @@ func blocks(d *digest, p []uint8) {
v5 = v5<<(64-63) | v5>>63 v5 = v5<<(64-63) | v5>>63
//// Round 2. //// Round 2.
//v0 += m[14] v0 += m[14]
//v0 += v4 v0 += v4
//v12 ^= v0 v12 ^= v0
//v12 = v12<<(64-32) | v12>>32 v12 = v12<<(64-32) | v12>>32
//v8 += v12 v8 += v12
//v4 ^= v8 v4 ^= v8
//v4 = v4<<(64-24) | v4>>24 v4 = v4<<(64-24) | v4>>24
//v1 += m[4] v1 += m[4]
//v1 += v5 v1 += v5
//v13 ^= v1 v13 ^= v1
//v13 = v13<<(64-32) | v13>>32 v13 = v13<<(64-32) | v13>>32
//v9 += v13 v9 += v13
//v5 ^= v9 v5 ^= v9
//v5 = v5<<(64-24) | v5>>24 v5 = v5<<(64-24) | v5>>24
//v2 += m[9] v2 += m[9]
//v2 += v6 v2 += v6
//v14 ^= v2 v14 ^= v2
//v14 = v14<<(64-32) | v14>>32 v14 = v14<<(64-32) | v14>>32
//v10 += v14 v10 += v14
//v6 ^= v10 v6 ^= v10
//v6 = v6<<(64-24) | v6>>24 v6 = v6<<(64-24) | v6>>24
//v3 += m[13] v3 += m[13]
//v3 += v7 v3 += v7
//v15 ^= v3 v15 ^= v3
//v15 = v15<<(64-32) | v15>>32 v15 = v15<<(64-32) | v15>>32
//v11 += v15 v11 += v15
//v7 ^= v11 v7 ^= v11
//v7 = v7<<(64-24) | v7>>24 v7 = v7<<(64-24) | v7>>24
//v2 += m[15] //v2 += m[15]
//v2 += v6 //v2 += v6
//v14 ^= v2 //v14 ^= v2

35
compress_amd64.s
View File

@@ -216,6 +216,41 @@ TEXT ·compressSSE(SB), 7, $0
BYTE $0xc4; BYTE $0x41; BYTE $0x09; BYTE $0x6c; BYTE $0xfe // VPUNPCKLQDQ XMM15, XMM14, XMM14 /* _mm_unpacklo_epi64(t1, t1) */ BYTE $0xc4; BYTE $0x41; BYTE $0x09; BYTE $0x6c; BYTE $0xfe // VPUNPCKLQDQ XMM15, XMM14, XMM14 /* _mm_unpacklo_epi64(t1, t1) */
BYTE $0xc4; BYTE $0xc1; BYTE $0x41; BYTE $0x6d; BYTE $0xff // VPUNPCKHQDQ XMM7, XMM7, XMM15 /* row4h = _mm_unpackhi_epi64(row4h, ) */ BYTE $0xc4; BYTE $0xc1; BYTE $0x41; BYTE $0x6d; BYTE $0xff // VPUNPCKHQDQ XMM7, XMM7, XMM15 /* row4h = _mm_unpackhi_epi64(row4h, ) */
///////////////////////////////////////////////////////////////////////////
// R O U N D 2
///////////////////////////////////////////////////////////////////////////
// LOAD_MSG_ ##r ##_1(b0, b1);
// LOAD_MSG_ ##r ##_2(b0, b1);
// (X12 used as additional temp register)
MOVQ message+0(FP), DX // DX: &p (message)
MOVOU 112(DX), X12 // X12 = m[14]+m[15]
MOVOU 32(DX), X13 // X13 = m[4]+ m[5]
MOVOU 64(DX), X14 // X14 = m[8]+ m[9]
MOVOU 96(DX), X15 // X15 = m[12]+m[13]
BYTE $0xc4; BYTE $0x41; BYTE $0x19; BYTE $0x6c; BYTE $0xc5 // VPUNPCKLQDQ XMM8, XMM12, XMM13 /* m[14], m[4] */
BYTE $0xc4; BYTE $0x41; BYTE $0x09; BYTE $0x6d; BYTE $0xcf // VPUNPCKHQDQ XMM9, XMM14, XMM15 /* m[9], m[13] */
// Load shuffle value
MOVQ shffle+120(FP), SI // SI: &shuffle
MOVOU 0(SI), X12 // X12 = 03040506 07000102 0b0c0d0e 0f08090a
// G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1);
BYTE $0xc4; BYTE $0xc1; BYTE $0x79; BYTE $0xd4; BYTE $0xc0 // VPADDQ XMM0,XMM0,XMM8 /* v0 += m[0], v1 += m[2] */
BYTE $0xc4; BYTE $0xc1; BYTE $0x71; BYTE $0xd4; BYTE $0xc9 // VPADDQ XMM1,XMM1,XMM9 /* v2 += m[4], v3 += m[6] */
BYTE $0xc5; BYTE $0xf9; BYTE $0xd4; BYTE $0xc2 // VPADDQ XMM0,XMM0,XMM2 /* v0 += v4, v1 += v5 */
BYTE $0xc5; BYTE $0xf1; BYTE $0xd4; BYTE $0xcb // VPADDQ XMM1,XMM1,XMM3 /* v2 += v6, v3 += v7 */
BYTE $0xc5; BYTE $0xc9; BYTE $0xef; BYTE $0xf0 // VPXOR XMM6,XMM6,XMM0 /* v12 ^= v0, v13 ^= v1 */
BYTE $0xc5; BYTE $0xc1; BYTE $0xef; BYTE $0xf9 // VPXOR XMM7,XMM7,XMM1 /* v14 ^= v2, v15 ^= v3 */
BYTE $0xc5; BYTE $0xf9; BYTE $0x70; BYTE $0xf6; BYTE $0xb1 // VPSHUFD XMM6,XMM6,0xb1 /* v12 = v12<<(64-32) | v12>>32, v13 = v13<<(64-32) | v13>>32 */
BYTE $0xc5; BYTE $0xf9; BYTE $0x70; BYTE $0xff; BYTE $0xb1 // VPSHUFD XMM7,XMM7,0xb1 /* v14 = v14<<(64-32) | v14>>32, v15 = v15<<(64-32) | v15>>32 */
BYTE $0xc5; BYTE $0xd9; BYTE $0xd4; BYTE $0xe6 // VPADDQ XMM4,XMM4,XMM6 /* v8 += v12, v9 += v13 */
BYTE $0xc5; BYTE $0xd1; BYTE $0xd4; BYTE $0xef // VPADDQ XMM5,XMM5,XMM7 /* v10 += v14, v11 += v15 */
BYTE $0xc5; BYTE $0xe9; BYTE $0xef; BYTE $0xd4 // VPXOR XMM2,XMM2,XMM4 /* v4 ^= v8, v5 ^= v9 */
BYTE $0xc5; BYTE $0xe1; BYTE $0xef; BYTE $0xdd // VPXOR XMM3,XMM3,XMM5 /* v6 ^= v10, v7 ^= v11 */
BYTE $0xc4; BYTE $0xc2; BYTE $0x69; BYTE $0x00; BYTE $0xd4 // VPSHUFB XMM2,XMM2,XMM12 /* v4 = v4<<(64-24) | v4>>24, v5 = v5<<(64-24) | v5>>24 */
BYTE $0xc4; BYTE $0xc2; BYTE $0x61; BYTE $0x00; BYTE $0xdc // VPSHUFB XMM3,XMM3,XMM12 /* v6 = v6<<(64-24) | v6>>24, v7 = v7<<(64-24) | v7>>24 */
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////
// R O U N D 1 1 // R O U N D 1 1
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////