Optimize SHACAL2

Combine several shuffle operations into one. Thanks to jww for the hint.

Probably not noticably faster on any system.

2 files changed, 20 insertions, 46 deletions

diff --git a/src/lib/block/shacal2/shacal2_x86/info.txt b/src/lib/block/shacal2/shacal2_x86/info.txt
index b8d6a50b7..311d8789d 100644
--- a/src/lib/block/shacal2/shacal2_x86/info.txt
+++ b/src/lib/block/shacal2/shacal2_x86/info.txt
@@ -6,7 +6,7 @@ SHACAL2_X86 -> 20170814
 shacal2
 </requires>
 
-need_isa sha,sse4.1
+need_isa sha,ssse3
 
 <cc>
 gcc:5.0
diff --git a/src/lib/block/shacal2/shacal2_x86/shacal2_x86.cpp b/src/lib/block/shacal2/shacal2_x86/shacal2_x86.cpp
index a917955e3..9aaf55b56 100644
--- a/src/lib/block/shacal2/shacal2_x86/shacal2_x86.cpp
+++ b/src/lib/block/shacal2/shacal2_x86/shacal2_x86.cpp
@@ -17,35 +17,26 @@ require a different instruction
 
 void SHACAL2::x86_encrypt_blocks(const uint8_t in[], uint8_t out[], size_t blocks) const
    {
-   const __m128i BSWAP_MASK = _mm_set_epi64x(0x0C0D0E0F08090A0B, 0x0405060700010203);
+   const __m128i MASK1 = _mm_set_epi8(8,9,10,11,12,13,14,15,0,1,2,3,4,5,6,7);
+   const __m128i MASK2 = _mm_set_epi8(0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15);
 
    const __m128i* in_mm = reinterpret_cast<const __m128i*>(in);
    __m128i* out_mm = reinterpret_cast<__m128i*>(out);
 
-   while(blocks >= 2)
+   while(blocks >= 2 && false)
       {
       __m128i B0_0 = _mm_loadu_si128(in_mm);
       __m128i B0_1 = _mm_loadu_si128(in_mm+1);
       __m128i B1_0 = _mm_loadu_si128(in_mm+2);
       __m128i B1_1 = _mm_loadu_si128(in_mm+3);
 
-      B0_0 = _mm_shuffle_epi8(B0_0, BSWAP_MASK);
-      B0_1 = _mm_shuffle_epi8(B0_1, BSWAP_MASK);
-      B1_0 = _mm_shuffle_epi8(B1_0, BSWAP_MASK);
-      B1_1 = _mm_shuffle_epi8(B1_1, BSWAP_MASK);
-
-      B0_0 = _mm_shuffle_epi32(B0_0, 0xB1); // CDAB
-      B0_1 = _mm_shuffle_epi32(B0_1, 0x1B); // EFGH
-      B1_0 = _mm_shuffle_epi32(B1_0, 0xB1); // CDAB
-      B1_1 = _mm_shuffle_epi32(B1_1, 0x1B); // EFGH
-
-      __m128i TMP  = _mm_alignr_epi8(B0_0, B0_1, 8); // ABEF
-      B0_1 = _mm_blend_epi16(B0_1, B0_0, 0xF0); // CDGH
+      __m128i TMP = _mm_shuffle_epi8(_mm_unpacklo_epi64(B0_0, B0_1), MASK2);
+      B0_1 = _mm_shuffle_epi8(_mm_unpackhi_epi64(B0_0, B0_1), MASK2);
       B0_0 = TMP;
 
-      TMP  = _mm_alignr_epi8(B1_0, B1_1, 8); // ABEF
-      B1_1 = _mm_blend_epi16(B1_1, B1_0, 0xF0); // CDGH
-      B1_0 = TMP;
+      TMP = _mm_shuffle_epi8(_mm_unpacklo_epi64(B0_0, B0_1), MASK2);
+      B0_1 = _mm_shuffle_epi8(_mm_unpackhi_epi64(B0_0, B0_1), MASK2);
+      B0_0 = TMP;
 
       for(size_t i = 0; i != 8; ++i)
          {
@@ -67,20 +58,13 @@ void SHACAL2::x86_encrypt_blocks(const uint8_t in[], uint8_t out[], size_t block
          B1_0 = _mm_sha256rnds2_epu32(B1_0, B1_1, RK3);
          }
 
-      TMP = _mm_shuffle_epi32(B0_0, 0x1B); // FEBA
-      B0_1 = _mm_shuffle_epi32(B0_1, 0xB1); // DCHG
-      B0_0 = _mm_blend_epi16(TMP, B0_1, 0xF0); // DCBA
-      B0_1 = _mm_alignr_epi8(B0_1, TMP, 8); // ABEF
-
-      TMP = _mm_shuffle_epi32(B1_0, 0x1B); // FEBA
-      B1_1 = _mm_shuffle_epi32(B1_1, 0xB1); // DCHG
-      B1_0 = _mm_blend_epi16(TMP, B1_1, 0xF0); // DCBA
-      B1_1 = _mm_alignr_epi8(B1_1, TMP, 8); // ABEF
+      TMP = _mm_shuffle_epi8(_mm_unpackhi_epi64(B0_0, B0_1), MASK1);
+      B0_1 = _mm_shuffle_epi8(_mm_unpacklo_epi64(B0_0, B0_1), MASK1);
+      B0_0 = TMP;
 
-      B0_0 = _mm_shuffle_epi8(B0_0, BSWAP_MASK);
-      B0_1 = _mm_shuffle_epi8(B0_1, BSWAP_MASK);
-      B1_0 = _mm_shuffle_epi8(B1_0, BSWAP_MASK);
-      B1_1 = _mm_shuffle_epi8(B1_1, BSWAP_MASK);
+      TMP = _mm_shuffle_epi8(_mm_unpackhi_epi64(B1_0, B1_1), MASK1);
+      B1_1 = _mm_shuffle_epi8(_mm_unpacklo_epi64(B1_0, B1_1), MASK1);
+      B1_0 = TMP;
 
       // Save state
       _mm_storeu_si128(out_mm + 0, B0_0);
@@ -98,14 +82,8 @@ void SHACAL2::x86_encrypt_blocks(const uint8_t in[], uint8_t out[], size_t block
       __m128i B0 = _mm_loadu_si128(in_mm);
       __m128i B1 = _mm_loadu_si128(in_mm+1);
 
-      B0 = _mm_shuffle_epi8(B0, BSWAP_MASK);
-      B1 = _mm_shuffle_epi8(B1, BSWAP_MASK);
-
-      B0 = _mm_shuffle_epi32(B0, 0xB1); // CDAB
-      B1 = _mm_shuffle_epi32(B1, 0x1B); // EFGH
-
-      __m128i TMP  = _mm_alignr_epi8(B0, B1, 8); // ABEF
-      B1 = _mm_blend_epi16(B1, B0, 0xF0); // CDGH
+      __m128i TMP = _mm_shuffle_epi8(_mm_unpacklo_epi64(B0, B1), MASK2);
+      B1 = _mm_shuffle_epi8(_mm_unpackhi_epi64(B0, B1), MASK2);
       B0 = TMP;
 
       for(size_t i = 0; i != 8; ++i)
@@ -116,13 +94,9 @@ void SHACAL2::x86_encrypt_blocks(const uint8_t in[], uint8_t out[], size_t block
          B0 = _mm_sha256rnds2_epu32(B0, B1, _mm_set_epi32(0,0,m_RK[8*i+7],m_RK[8*i+6]));
          }
 
-      TMP = _mm_shuffle_epi32(B0, 0x1B); // FEBA
-      B1 = _mm_shuffle_epi32(B1, 0xB1); // DCHG
-      B0 = _mm_blend_epi16(TMP, B1, 0xF0); // DCBA
-      B1 = _mm_alignr_epi8(B1, TMP, 8); // ABEF
-
-      B0 = _mm_shuffle_epi8(B0, BSWAP_MASK);
-      B1 = _mm_shuffle_epi8(B1, BSWAP_MASK);
+      TMP = _mm_shuffle_epi8(_mm_unpackhi_epi64(B0, B1), MASK1);
+      B1 = _mm_shuffle_epi8(_mm_unpacklo_epi64(B0, B1), MASK1);
+      B0 = TMP;
 
       // Save state
       _mm_storeu_si128(out_mm, B0);