Improved performance for cryptonight v7.

CHANGELOG.md

@@ -1,9 +1,11 @@
 # v2.6.0-beta2
+- Improved performance for `cryptonight v7` especially in double hash mode.
 - [#499](https://github.com/xmrig/xmrig/issues/499) IPv6 disabled for internal HTTP API by default, was cause issues on some systems.
 - Added short aliases for algorithm names: `cn`, `cn-lite` and `cn-heavy`.
 - Fixed regressions (v2.6.0-beta1 affected)
   - [#494](https://github.com/xmrig/xmrig/issues/494) Command line option `--donate-level` was broken.
-  - [#502](https://github.com/xmrig/xmrig/issues/502) Build without libmicrohttpd was broken.
+  - [#502](https://github.com/xmrig/xmrig/issues/502) Build without libmicrohttpd was broken.
+  - Fixed nonce calculation for `--av 4` (software AES, double hash) was cause reduction of effective hashrate and rejected shares on nicehash.
 
 # v2.6.0-beta1
  - [#476](https://github.com/xmrig/xmrig/issues/476) **Added Cryptonight-Heavy support for Sumokoin ASIC resistance fork.**

src/crypto/CryptoNight_x86.h

@@ -386,6 +386,22 @@ static inline void cn_implode_scratchpad(const __m128i *input, __m128i *output)
 }
 
 
+static inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
+{
+    mem_out[0] = EXTRACT64(tmp);
+
+    tmp = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(tmp), _mm_castsi128_ps(tmp)));
+    uint64_t vh = EXTRACT64(tmp);
+
+    uint8_t x = vh >> 24;
+    static const uint16_t table = 0x7531;
+    const uint8_t index = (((x >> 3) & 6) | (x & 1)) << 1;
+    vh ^= ((table >> index) & 0x3) << 28;
+
+    mem_out[1] = vh;
+}
+
+
 template<xmrig::Algo ALGO, bool SOFT_AES, int VARIANT>
 inline void cryptonight_single_hash(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, cryptonight_ctx *__restrict__ ctx)
 {
@@ -400,7 +416,7 @@ inline void cryptonight_single_hash(const uint8_t *__restrict__ input, size_t si
 
     keccak(input, (int) size, ctx->state0, 200);
 
-    VARIANT1_INIT(0);
+    VARIANT1_INIT(0)
 
     cn_explode_scratchpad<ALGO, MEM, SOFT_AES>((__m128i*) ctx->state0, (__m128i*) ctx->memory);
 
@@ -423,8 +439,13 @@ inline void cryptonight_single_hash(const uint8_t *__restrict__ input, size_t si
             cx = _mm_load_si128((__m128i *) &l0[idx0 & MASK]);
             cx = _mm_aesenc_si128(cx, _mm_set_epi64x(ah0, al0));
         }
-        _mm_store_si128((__m128i *) &l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
+
-        VARIANT1_1(&l0[idx0 & MASK]);
+        if (VARIANT > 0) {
+            cryptonight_monero_tweak((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
+        } else {
+            _mm_store_si128((__m128i *)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
+        }
+
         idx0 = EXTRACT64(cx);
         bx0 = cx;
 
@@ -513,10 +534,13 @@ inline void cryptonight_double_hash(const uint8_t *__restrict__ input, size_t si
             cx1 = _mm_aesenc_si128(cx1, _mm_set_epi64x(ah1, al1));
         }
 
-        _mm_store_si128((__m128i *) &l0[idx0 & MASK], _mm_xor_si128(bx0, cx0));
+        if (VARIANT > 0) {
-        _mm_store_si128((__m128i *) &l1[idx1 & MASK], _mm_xor_si128(bx1, cx1));
+            cryptonight_monero_tweak((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx0));
-        VARIANT1_1(&l0[idx0 & MASK]);
+            cryptonight_monero_tweak((uint64_t*)&l1[idx1 & MASK], _mm_xor_si128(bx1, cx1));
-        VARIANT1_1(&l1[idx1 & MASK]);
+        } else {
+            _mm_store_si128((__m128i *) &l0[idx0 & MASK], _mm_xor_si128(bx0, cx0));
+            _mm_store_si128((__m128i *) &l1[idx1 & MASK], _mm_xor_si128(bx1, cx1));
+        }
 
         idx0 = EXTRACT64(cx0);
         idx1 = EXTRACT64(cx1);