diff options
| author | Jeffrey Walton <noloader@gmail.com> | 2020-06-29 05:17:59 -0400 |
|---|---|---|
| committer | Jeffrey Walton <noloader@gmail.com> | 2020-06-29 05:17:59 -0400 |
| commit | 827f2ebcadd02b3474bdbda23a4d0c733ace8f3b (patch) | |
| tree | dff0bd3c121841e8ec373a224f551c6517a2eda2 /blake2s_simd.cpp | |
| parent | a3aefbb1dcd35260718e0f78aafd3ac7d68b6aeb (diff) | |
| download | cryptopp-git-827f2ebcadd02b3474bdbda23a4d0c733ace8f3b.tar.gz | |
Speedup BLAKE2s message loading on PowerPC
Diffstat (limited to 'blake2s_simd.cpp')
| -rw-r--r-- | blake2s_simd.cpp | 60 |
1 files changed, 56 insertions, 4 deletions
diff --git a/blake2s_simd.cpp b/blake2s_simd.cpp index a2c641b6..2f67f5c8 100644 --- a/blake2s_simd.cpp +++ b/blake2s_simd.cpp @@ -992,12 +992,64 @@ void BLAKE2_Compress32_ALTIVEC(const byte* input, BLAKE2s_State& state) BLAKE2S_G2(row1,row2,row3,row4,buf4); \
BLAKE2S_UNDIAGONALIZE(row1,row2,row3,row4);
+ /* Possibly unaligned user messages */
+ uint32x4_p m0, m4, m8, m12;
+ /* Endian conversion mask */
const uint8x16_p le_mask = {3,2,1,0, 7,6,5,4, 11,10,9,8, 15,14,13,12};
- const uint32x4_p m0 = VecLoad32LE(input + 0, le_mask);
- const uint32x4_p m4 = VecLoad32LE(input + 16, le_mask);
- const uint32x4_p m8 = VecLoad32LE(input + 32, le_mask);
- const uint32x4_p m12 = VecLoad32LE(input + 48, le_mask);
+#if defined(_ARCH_PWR9)
+ /* POWER9 provides loads for char's and short's */
+ m0 = (uint32x4_p) vec_xl( 0, CONST_V8_CAST( input ));
+ m4 = (uint32x4_p) vec_xl( 16, CONST_V8_CAST( input ));
+ m8 = (uint32x4_p) vec_xl( 32, CONST_V8_CAST( input ));
+ m12 = (uint32x4_p) vec_xl( 48, CONST_V8_CAST( input ));
+
+# if defined(CRYPTOPP_BIG_ENDIAN)
+ m0 = vec_perm(m0, m0, le_mask);
+ m4 = vec_perm(m4, m4, le_mask);
+ m8 = vec_perm(m8, m8, le_mask);
+ m12 = vec_perm(m12, m12, le_mask);
+# endif
+#else
+ /* Altivec only provides 16-byte aligned loads */
+ /* http://www.nxp.com/docs/en/reference-manual/ALTIVECPEM.pdf, Section 3.16 */
+ m0 = (uint32x4_p) vec_ld( 0, CONST_V8_CAST( input ));
+ m4 = (uint32x4_p) vec_ld( 16, CONST_V8_CAST( input ));
+ m8 = (uint32x4_p) vec_ld( 32, CONST_V8_CAST( input ));
+ m12 = (uint32x4_p) vec_ld( 48, CONST_V8_CAST( input ));
+
+ /* Alignment check for load of the message buffer */
+ const uintptr_t addr = (uintptr_t)input;
+ if (addr%16 == 0)
+ {
+ /* Already aligned. Perform a little-endian swap as required */
+# if defined(CRYPTOPP_BIG_ENDIAN)
+ m0 = vec_perm(m0, m0, le_mask);
+ m4 = vec_perm(m4, m4, le_mask);
+ m8 = vec_perm(m8, m8, le_mask);
+ m12 = vec_perm(m12, m12, le_mask);
+# endif
+ }
+ else
+ {
+ /* Not aligned. Fix vectors and perform a little-endian swap as required */
+ // http://mirror.informatimago.com/next/developer.apple.com/
+ // hardwaredrivers/ve/code_optimization.html
+ uint32x4_p ex; uint8x16_p perm;
+ ex = (uint32x4_p) vec_ld(48+15, CONST_V8_CAST( input ));
+ perm = vec_lvsl(0, CONST_V8_CAST( addr ));
+
+# if defined(CRYPTOPP_BIG_ENDIAN)
+ /* Combine the vector permute with the little-endian swap */
+ perm = vec_perm(perm, perm, le_mask);
+# endif
+
+ m0 = vec_perm(m0, m4, perm);
+ m4 = vec_perm(m4, m8, perm);
+ m8 = vec_perm(m8, m12, perm);
+ m12 = vec_perm(m12, ex, perm);
+ }
+#endif
uint32x4_p row1, row2, row3, row4;
uint32x4_p buf1, buf2, buf3, buf4;
|