Update slow-hash.c

1. Added huge pages support and optimized scratchpad twiddling. (credits to dga).
2. Added aes-ni key expansion support.
3. Minor speedup to scratchpad initialization/finalization.

1 files changed, 288 insertions, 110 deletions

diff --git a/src/crypto/slow-hash.c b/src/crypto/slow-hash.c
index 35438dcab..cda66af52 100644
--- a/src/crypto/slow-hash.c
+++ b/src/crypto/slow-hash.c
@@ -13,8 +13,9 @@
 
 #include <emmintrin.h>
 
-#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
+#if defined(_MSC_VER)
 #include <intrin.h>
+#include <Windows.h>
 #define STATIC
 #define INLINE __inline
 #if !defined(RDATA_ALIGN16)
@@ -22,6 +23,7 @@
 #endif
 #else
 #include <wmmintrin.h>
+#include <sys/mman.h>
 #define STATIC static
 #define INLINE inline
 #if !defined(RDATA_ALIGN16)
@@ -29,15 +31,58 @@
 #endif
 #endif
 
+#if defined(__INTEL_COMPILER)
+#define ASM __asm__
+#elif !defined(_MSC_VER)
+#define ASM __asm__
+#else
+#define ASM __asm
+#endif
+
 #define MEMORY         (1 << 21) // 2MB scratchpad
 #define ITER           (1 << 20)
 #define AES_BLOCK_SIZE  16
 #define AES_KEY_SIZE    32
 #define INIT_SIZE_BLK   8
 #define INIT_SIZE_BYTE (INIT_SIZE_BLK * AES_BLOCK_SIZE)
+#define TOTALBLOCKS (MEMORY / AES_BLOCK_SIZE)
 
 #define U64(x) ((uint64_t *) (x))
 #define R128(x) ((__m128i *) (x))
+#define SWAP(a, b) (((a) -= (b)), ((b) += (a)), ((a) = (b) - (a)))
+
+#define state_index(x) (((*((uint64_t *)x) >> 4) & (TOTALBLOCKS - 1)) << 4)
+#if defined(_MSC_VER)
+#define __mul() lo = _umul128(c[0], b[0], &hi);
+#else
+#define __mul() ASM("mulq %3\n\t" : "=d"(hi), "=a"(lo) : "%a" (c[0]), "rm" (b[0]) : "cc");
+#endif
+
+#define pre_aes() \
+    j = state_index(a); \
+	_c = _mm_load_si128(R128(&hp_state[j])); \
+	_a = _mm_load_si128(R128(a)); \
+ 
+// dga's optimized scratchpad twiddling
+#define post_aes() \
+	_mm_store_si128(R128(c), _c); \
+	_b = _mm_xor_si128(_b, _c); \
+	_mm_store_si128(R128(&hp_state[j]), _b); \
+	j = state_index(c); \
+	p = U64(&hp_state[j]); \
+	b[0] = p[0]; b[1] = p[1]; \
+	__mul(); \
+	a[0] += hi; a[1] += lo; \
+	p = U64(&hp_state[j]); \
+	p[0] = a[0];  p[1] = a[1]; \
+	a[0] ^= b[0]; a[1] ^= b[1]; \
+	_b = _c; \
+ 
+#if defined(_MSC_VER)
+#define THREADV __declspec(thread)
+#else
+#define THREADV __thread
+#endif
 
 extern int aesb_single_round(const uint8_t *in, uint8_t*out, const uint8_t *expandedKey);
 extern int aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *expandedKey);
@@ -54,59 +99,26 @@ union cn_slow_hash_state
 };
 #pragma pack(pop)
 
-#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
+THREADV uint8_t *hp_state = NULL;
+THREADV int hp_allocated = 0;
+
+#if defined(_MSC_VER)
 #define cpuid(info,x)    __cpuidex(info,x,0)
 #else
 void cpuid(int CPUInfo[4], int InfoType)
 {
-    __asm__ __volatile__
+    ASM __volatile__
     (
-        "cpuid":
+    "cpuid":
         "=a" (CPUInfo[0]),
         "=b" (CPUInfo[1]),
         "=c" (CPUInfo[2]),
         "=d" (CPUInfo[3]) :
-        "a" (InfoType), "c" (0)
-    );
+            "a" (InfoType), "c" (0)
+        );
 }
 #endif
 
-STATIC INLINE void mul(const uint8_t *a, const uint8_t *b, uint8_t *res)
-{
-    uint64_t a0, b0;
-    uint64_t hi, lo;
-
-    a0 = U64(a)[0];
-    b0 = U64(b)[0];
-    lo = mul128(a0, b0, &hi);
-    U64(res)[0] = hi;
-    U64(res)[1] = lo;
-}
-
-STATIC INLINE void sum_half_blocks(uint8_t *a, const uint8_t *b)
-{
-    uint64_t a0, a1, b0, b1;
-    a0 = U64(a)[0];
-    a1 = U64(a)[1];
-    b0 = U64(b)[0];
-    b1 = U64(b)[1];
-    a0 += b0;
-    a1 += b1;
-    U64(a)[0] = a0;
-    U64(a)[1] = a1;
-}
-
-STATIC INLINE void swap_blocks(uint8_t *a, uint8_t *b)
-{
-    uint64_t t[2];
-    U64(t)[0] = U64(a)[0];
-    U64(t)[1] = U64(a)[1];
-    U64(a)[0] = U64(b)[0];
-    U64(a)[1] = U64(b)[1];
-    U64(b)[0] = U64(t)[0];
-    U64(b)[1] = U64(t)[1];
-}
-
 STATIC INLINE void xor_blocks(uint8_t *a, const uint8_t *b)
 {
     U64(a)[0] ^= U64(b)[0];
@@ -125,74 +137,248 @@ STATIC INLINE int check_aes_hw(void)
     return supported = cpuid_results[2] & (1 << 25);
 }
 
-STATIC INLINE void aesni_pseudo_round(const uint8_t *in, uint8_t *out,
-                                      const uint8_t *expandedKey)
+STATIC INLINE void aes_256_assist1(__m128i* t1, __m128i * t2)
+{
+    __m128i t4;
+    *t2 = _mm_shuffle_epi32(*t2, 0xff);
+    t4 = _mm_slli_si128(*t1, 0x04);
+    *t1 = _mm_xor_si128(*t1, t4);
+    t4 = _mm_slli_si128(t4, 0x04);
+    *t1 = _mm_xor_si128(*t1, t4);
+    t4 = _mm_slli_si128(t4, 0x04);
+    *t1 = _mm_xor_si128(*t1, t4);
+    *t1 = _mm_xor_si128(*t1, *t2);
+}
+
+STATIC INLINE void aes_256_assist2(__m128i* t1, __m128i * t3)
+{
+    __m128i t2, t4;
+    t4 = _mm_aeskeygenassist_si128(*t1, 0x00);
+    t2 = _mm_shuffle_epi32(t4, 0xaa);
+    t4 = _mm_slli_si128(*t3, 0x04);
+    *t3 = _mm_xor_si128(*t3, t4);
+    t4 = _mm_slli_si128(t4, 0x04);
+    *t3 = _mm_xor_si128(*t3, t4);
+    t4 = _mm_slli_si128(t4, 0x04);
+    *t3 = _mm_xor_si128(*t3, t4);
+    *t3 = _mm_xor_si128(*t3, t2);
+}
+
+STATIC INLINE void aes_expand_key(const uint8_t *key, uint8_t *expandedKey)
+{
+    __m128i *ek = R128(expandedKey);
+    __m128i t1, t2, t3;
+
+    t1 = _mm_loadu_si128(R128(key));
+    t3 = _mm_loadu_si128(R128(key + 16));
+
+    ek[0] = t1;
+    ek[1] = t3;
+
+    t2 = _mm_aeskeygenassist_si128(t3, 0x01);
+    aes_256_assist1(&t1, &t2);
+    ek[2] = t1;
+    aes_256_assist2(&t1, &t3);
+    ek[3] = t3;
+
+    t2 = _mm_aeskeygenassist_si128(t3, 0x02);
+    aes_256_assist1(&t1, &t2);
+    ek[4] = t1;
+    aes_256_assist2(&t1, &t3);
+    ek[5] = t3;
+
+    t2 = _mm_aeskeygenassist_si128(t3, 0x04);
+    aes_256_assist1(&t1, &t2);
+    ek[6] = t1;
+    aes_256_assist2(&t1, &t3);
+    ek[7] = t3;
+
+    t2 = _mm_aeskeygenassist_si128(t3, 0x08);
+    aes_256_assist1(&t1, &t2);
+    ek[8] = t1;
+    aes_256_assist2(&t1, &t3);
+    ek[9] = t3;
+
+    t2 = _mm_aeskeygenassist_si128(t3, 0x10);
+    aes_256_assist1(&t1, &t2);
+    ek[10] = t1;
+}
+
+STATIC INLINE void aes_pseudo_round(const uint8_t *in, uint8_t *out,
+                                    const uint8_t *expandedKey, int nblocks)
+{
+    __m128i *k = R128(expandedKey);
+    __m128i d;
+    int i;
+
+    for(i = 0; i < nblocks; i++)
+    {
+        d = _mm_loadu_si128(R128(in + i * AES_BLOCK_SIZE));
+        d = _mm_aesenc_si128(d, *R128(&k[0]));
+        d = _mm_aesenc_si128(d, *R128(&k[1]));
+        d = _mm_aesenc_si128(d, *R128(&k[2]));
+        d = _mm_aesenc_si128(d, *R128(&k[3]));
+        d = _mm_aesenc_si128(d, *R128(&k[4]));
+        d = _mm_aesenc_si128(d, *R128(&k[5]));
+        d = _mm_aesenc_si128(d, *R128(&k[6]));
+        d = _mm_aesenc_si128(d, *R128(&k[7]));
+        d = _mm_aesenc_si128(d, *R128(&k[8]));
+        d = _mm_aesenc_si128(d, *R128(&k[9]));
+        _mm_storeu_si128((R128(out + i * AES_BLOCK_SIZE)), d);
+    }
+}
+
+STATIC INLINE void aes_pseudo_round_xor(const uint8_t *in, uint8_t *out,
+                                        const uint8_t *expandedKey, const uint8_t *xor, int nblocks)
 {
     __m128i *k = R128(expandedKey);
+    __m128i *x = R128(xor);
     __m128i d;
+    int i;
+
+    for(i = 0; i < nblocks; i++)
+    {
+        d = _mm_loadu_si128(R128(in + i * AES_BLOCK_SIZE));
+        d = _mm_xor_si128(d, *R128(x++));
+        d = _mm_aesenc_si128(d, *R128(&k[0]));
+        d = _mm_aesenc_si128(d, *R128(&k[1]));
+        d = _mm_aesenc_si128(d, *R128(&k[2]));
+        d = _mm_aesenc_si128(d, *R128(&k[3]));
+        d = _mm_aesenc_si128(d, *R128(&k[4]));
+        d = _mm_aesenc_si128(d, *R128(&k[5]));
+        d = _mm_aesenc_si128(d, *R128(&k[6]));
+        d = _mm_aesenc_si128(d, *R128(&k[7]));
+        d = _mm_aesenc_si128(d, *R128(&k[8]));
+        d = _mm_aesenc_si128(d, *R128(&k[9]));
+        _mm_storeu_si128((R128(out + i * AES_BLOCK_SIZE)), d);
+    }
+}
+
+#if defined(_MSC_VER)
+BOOL SetLockPagesPrivilege(HANDLE hProcess, BOOL bEnable)
+{
+    struct
+    {
+        DWORD count;
+        LUID_AND_ATTRIBUTES privilege[1];
+    } info;
+
+    HANDLE token;
+    if(!OpenProcessToken(hProcess, TOKEN_ADJUST_PRIVILEGES, &token))
+        return FALSE;
+
+    info.count = 1;
+    info.privilege[0].Attributes = bEnable ? SE_PRIVILEGE_ENABLED : 0;
 
-    d = _mm_loadu_si128(R128(in));
-    d = _mm_aesenc_si128(d, *R128(&k[0]));
-    d = _mm_aesenc_si128(d, *R128(&k[1]));
-    d = _mm_aesenc_si128(d, *R128(&k[2]));
-    d = _mm_aesenc_si128(d, *R128(&k[3]));
-    d = _mm_aesenc_si128(d, *R128(&k[4]));
-    d = _mm_aesenc_si128(d, *R128(&k[5]));
-    d = _mm_aesenc_si128(d, *R128(&k[6]));
-    d = _mm_aesenc_si128(d, *R128(&k[7]));
-    d = _mm_aesenc_si128(d, *R128(&k[8]));
-    d = _mm_aesenc_si128(d, *R128(&k[9]));
-    _mm_storeu_si128((R128(out)), d);
+    if(!LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &(info.privilege[0].Luid)))
+        return FALSE;
+
+    if(!AdjustTokenPrivileges(token, FALSE, (PTOKEN_PRIVILEGES) &info, 0, NULL, NULL))
+        return FALSE;
+
+    if (GetLastError() != ERROR_SUCCESS)
+        return FALSE;
+
+    CloseHandle(token);
+
+    return TRUE;
+
+}
+#endif
+
+void slow_hash_allocate_state(void)
+{
+    int state = 0;
+    if(hp_state != NULL)
+        return;
+
+#if defined(_MSC_VER)
+    SetLockPagesPrivilege(GetCurrentProcess(), TRUE);
+    hp_state = (uint8_t *) VirtualAlloc(hp_state, MEMORY, MEM_LARGE_PAGES |
+                                        MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
+#else
+    hp_state = mmap(0, MEMORY, PROT_READ | PROT_WRITE,
+                    MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB, 0, 0);
+    if(hp_state == MAP_FAILED)
+        hp_state = NULL;
+#endif
+    hp_allocated = 1;
+    if(hp_state == NULL)
+    {
+        hp_allocated = 0;
+        hp_state = (uint8_t *) malloc(MEMORY);
+    }
+}
+
+void slow_hash_free_state(void)
+{
+    if(hp_state == NULL)
+        return;
+
+    if(!hp_allocated)
+        free(hp_state);
+    else
+    {
+#if defined(_MSC_VER)
+        VirtualFree(hp_state, MEMORY, MEM_RELEASE);
+#else
+        munmap(hp_state, MEMORY);
+#endif
+    }
+
+    hp_state = NULL;
+    hp_allocated = 0;
 }
 
 void cn_slow_hash(const void *data, size_t length, char *hash)
 {
-    uint8_t long_state[MEMORY];
-    uint8_t text[INIT_SIZE_BYTE];
-    uint8_t a[AES_BLOCK_SIZE];
-    uint8_t b[AES_BLOCK_SIZE];
-    uint8_t d[AES_BLOCK_SIZE];
-    uint8_t aes_key[AES_KEY_SIZE];
-    RDATA_ALIGN16 uint8_t expandedKey[256];
+    RDATA_ALIGN16 uint8_t expandedKey[240];
 
+    uint8_t text[INIT_SIZE_BYTE];
+    RDATA_ALIGN16 uint64_t a[2];
+    RDATA_ALIGN16 uint64_t b[2];
+    RDATA_ALIGN16 uint64_t c[2];
+    RDATA_ALIGN16 uint8_t aes_key[AES_KEY_SIZE];
     union cn_slow_hash_state state;
+    __m128i _a, _b, _c;
+    uint64_t hi, lo;
 
     size_t i, j;
-    uint8_t *p = NULL;
+    uint64_t *p = NULL;
     oaes_ctx *aes_ctx;
-
     int useAes = check_aes_hw();
+
     static void (*const extra_hashes[4])(const void *, size_t, char *) =
     {
         hash_extra_blake, hash_extra_groestl, hash_extra_jh, hash_extra_skein
     };
 
+    // this isn't supposed to happen, but guard against it for now.
+    if(hp_state == NULL)
+        slow_hash_allocate_state();
+
     hash_process(&state.hs, data, length);
     memcpy(text, state.init, INIT_SIZE_BYTE);
 
-    aes_ctx = (oaes_ctx *) oaes_alloc();
-    oaes_key_import_data(aes_ctx, state.hs.b, AES_KEY_SIZE);
-
-    // use aligned data
-    memcpy(expandedKey, aes_ctx->key->exp_data, aes_ctx->key->exp_data_len);
-
     if(useAes)
     {
+        aes_expand_key(state.hs.b, expandedKey);
         for(i = 0; i < MEMORY / INIT_SIZE_BYTE; i++)
         {
-            for(j = 0; j < INIT_SIZE_BLK; j++)
-                aesni_pseudo_round(&text[AES_BLOCK_SIZE * j], &text[AES_BLOCK_SIZE * j], expandedKey);
-            memcpy(&long_state[i * INIT_SIZE_BYTE], text, INIT_SIZE_BYTE);
+            aes_pseudo_round(text, text, expandedKey, INIT_SIZE_BLK);
+            memcpy(&hp_state[i * INIT_SIZE_BYTE], text, INIT_SIZE_BYTE);
         }
     }
     else
     {
+        aes_ctx = (oaes_ctx *) oaes_alloc();
+        oaes_key_import_data(aes_ctx, state.hs.b, AES_KEY_SIZE);
         for(i = 0; i < MEMORY / INIT_SIZE_BYTE; i++)
         {
             for(j = 0; j < INIT_SIZE_BLK; j++)
-                aesb_pseudo_round(&text[AES_BLOCK_SIZE * j], &text[AES_BLOCK_SIZE * j], expandedKey);
+                aesb_pseudo_round(&text[AES_BLOCK_SIZE * j], &text[AES_BLOCK_SIZE * j], aes_ctx->key->exp_data);
 
-            memcpy(&long_state[i * INIT_SIZE_BYTE], text, INIT_SIZE_BYTE);
+            memcpy(&hp_state[i * INIT_SIZE_BYTE], text, INIT_SIZE_BYTE);
         }
     }
 
@@ -201,60 +387,52 @@ void cn_slow_hash(const void *data, size_t length, char *hash)
     U64(b)[0] = U64(&state.k[16])[0] ^ U64(&state.k[48])[0];
     U64(b)[1] = U64(&state.k[16])[1] ^ U64(&state.k[48])[1];
 
-    for(i = 0; i < ITER / 2; i++)
+    _b = _mm_load_si128(R128(b));
+    // this is ugly but the branching affects the loop somewhat so put it outside.
+    if(useAes)
     {
-				#define TOTALBLOCKS (MEMORY / AES_BLOCK_SIZE)
-				#define state_index(x) (((*((uint64_t *)x) >> 4) & (TOTALBLOCKS - 1)) << 4)
-
-        // Iteration 1
-        p = &long_state[state_index(a)];
-
-        if(useAes)
-            _mm_storeu_si128(R128(p), _mm_aesenc_si128(_mm_loadu_si128(R128(p)), _mm_loadu_si128(R128(a))));
-        else
-            aesb_single_round(p, p, a);
-
-        xor_blocks(b, p);
-        swap_blocks(b, p);
-        swap_blocks(a, b);
-
-        // Iteration 2
-        p = &long_state[state_index(a)];
-
-        mul(a, p, d);
-        sum_half_blocks(b, d);
-        swap_blocks(b, p);
-        xor_blocks(b, p);
-        swap_blocks(a, b);
+        for(i = 0; i < ITER / 2; i++)
+        {
+            pre_aes();
+            _c = _mm_aesenc_si128(_c, _a);
+            // post_aes(), optimized scratchpad twiddling (credits to dga)
+            post_aes();
+        }
+    }
+    else
+    {
+        for(i = 0; i < ITER / 2; i++)
+        {
+            pre_aes();
+            aesb_single_round((uint8_t *) &_c, (uint8_t *) &_c, (uint8_t *) &_a);
+            post_aes();
+        }
     }
 
     memcpy(text, state.init, INIT_SIZE_BYTE);
-    oaes_key_import_data(aes_ctx, &state.hs.b[32], AES_KEY_SIZE);
-    memcpy(expandedKey, aes_ctx->key->exp_data, aes_ctx->key->exp_data_len);
     if(useAes)
     {
+        aes_expand_key(&state.hs.b[32], expandedKey);
         for(i = 0; i < MEMORY / INIT_SIZE_BYTE; i++)
         {
-            for(j = 0; j < INIT_SIZE_BLK; j++)
-            {
-                xor_blocks(&text[j * AES_BLOCK_SIZE], &long_state[i * INIT_SIZE_BYTE + j * AES_BLOCK_SIZE]);
-                aesni_pseudo_round(&text[j * AES_BLOCK_SIZE], &text[j * AES_BLOCK_SIZE], expandedKey);
-            }
+            // add the xor to the pseudo round
+            aes_pseudo_round_xor(text, text, expandedKey, &hp_state[i * INIT_SIZE_BYTE], INIT_SIZE_BLK);
         }
     }
     else
     {
+        oaes_key_import_data(aes_ctx, &state.hs.b[32], AES_KEY_SIZE);
         for(i = 0; i < MEMORY / INIT_SIZE_BYTE; i++)
         {
             for(j = 0; j < INIT_SIZE_BLK; j++)
             {
-                xor_blocks(&text[j * AES_BLOCK_SIZE], &long_state[i * INIT_SIZE_BYTE + j * AES_BLOCK_SIZE]);
-                aesb_pseudo_round(&text[AES_BLOCK_SIZE * j], &text[AES_BLOCK_SIZE * j], expandedKey);
+                xor_blocks(&text[j * AES_BLOCK_SIZE], &hp_state[i * INIT_SIZE_BYTE + j * AES_BLOCK_SIZE]);
+                aesb_pseudo_round(&text[AES_BLOCK_SIZE * j], &text[AES_BLOCK_SIZE * j], aes_ctx->key->exp_data);
             }
         }
+        oaes_free((OAES_CTX **) &aes_ctx);
     }
 
-    oaes_free((OAES_CTX **) &aes_ctx);
     memcpy(state.init, text, INIT_SIZE_BYTE);
     hash_permutation(&state.hs);
     extra_hashes[state.hs.b[0] & 3](&state, 200, hash);