v23.12

algo/keccak/keccak-4way.c

@@ -78,7 +78,7 @@ int scanhash_keccak_4way( struct work *work, uint32_t max_nonce,
    uint32_t *ptarget = work->target;
    uint32_t n = pdata[19];
    const uint32_t first_nonce = pdata[19];
-   __m256i  *noncev = (__m256i*)vdata + 9;   // aligned
+   __m256i *noncev = (__m256i*)vdata + 9;   // aligned
    const uint32_t Htarg = ptarget[7];
    const int thr_id = mythr->id;
    const bool bench = opt_benchmark;
@@ -108,4 +108,53 @@ int scanhash_keccak_4way( struct work *work, uint32_t max_nonce,
    return 0;
 }
 
+#elif defined(KECCAK_2WAY)
+
+void keccakhash_2x64(void *state, const void *input)
+{
+    keccak256_2x64_context ctx;
+    keccak256_2x64_init( &ctx );
+    keccak256_2x64_update( &ctx, input, 80 );
+    keccak256_2x64_close( &ctx, state );
+}
+
+int scanhash_keccak_2x64( struct work *work, uint32_t max_nonce,
+                          uint64_t *hashes_done, struct thr_info *mythr )
+{
+   uint32_t vdata[24*2] __attribute__ ((aligned (64)));
+   uint32_t hash[16*2] __attribute__ ((aligned (32)));
+   uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+   uint32_t *hash7 = &(hash[13]);   // 3*4+1
+   uint32_t *pdata = work->data;
+   uint32_t *ptarget = work->target;
+   uint32_t n = pdata[19];
+   const uint32_t first_nonce = pdata[19];
+   v128_t *noncev = (v128_t*)vdata + 9;
+   const uint32_t Htarg = ptarget[7];
+   const int thr_id = mythr->id;
+   const bool bench = opt_benchmark;
+
+   v128_bswap32_intrlv80_2x64( vdata, pdata );
+   *noncev = v128_intrlv_blend_32( v128_set32( n+1, 0, n, 0 ), *noncev );
+   do {
+      keccakhash_2x64( hash, vdata );
+
+      for ( int lane = 0; lane < 2; lane++ )
+      if unlikely( hash7[ lane<<1 ] <= Htarg && !bench )
+      {
+          extr_lane_2x64( lane_hash, hash, lane, 256 );
+          if ( valid_hash( lane_hash, ptarget ))
+          {
+              pdata[19] = bswap_32( n + lane );
+              submit_solution( work, lane_hash, mythr );
+          }
+      }
+      *noncev = v128_add32( *noncev, v128_64( 0x0000000200000000 ) );
+      n += 2;
+   } while ( (n < max_nonce-2) && !work_restart[thr_id].restart);
+   pdata[19] = n;
+   *hashes_done = n - first_nonce + 1;
+   return 0;
+}
+
 #endif

algo/keccak/keccak-gate.c

@@ -17,6 +17,9 @@ bool register_keccak_algo( algo_gate_t* gate )
 #elif defined (KECCAK_4WAY)
   gate->scanhash  = (void*)&scanhash_keccak_4way;
   gate->hash      = (void*)&keccakhash_4way;
+#elif defined (KECCAK_2WAY)
+  gate->scanhash  = (void*)&scanhash_keccak_2x64;
+  gate->hash      = (void*)&keccakhash_2x64;
 #else
   gate->scanhash  = (void*)&scanhash_keccak;
   gate->hash      = (void*)&keccakhash;
@@ -37,6 +40,9 @@ bool register_keccakc_algo( algo_gate_t* gate )
 #elif defined (KECCAK_4WAY)
   gate->scanhash  = (void*)&scanhash_keccak_4way;
   gate->hash      = (void*)&keccakhash_4way;
+#elif defined (KECCAK_2WAY)
+  gate->scanhash  = (void*)&scanhash_keccak_2x64;
+  gate->hash      = (void*)&keccakhash_2x64;
 #else
   gate->scanhash  = (void*)&scanhash_keccak;
   gate->hash      = (void*)&keccakhash;
@@ -75,15 +81,17 @@ void sha3d_gen_merkle_root( char* merkle_root, struct stratum_ctx* sctx )
 bool register_sha3d_algo( algo_gate_t* gate )
 {
   hard_coded_eb = 6;
-//  opt_extranonce = false;
-  gate->optimizations = AVX2_OPT | AVX512_OPT;
+  gate->optimizations = SSE2_OPT | AVX2_OPT | AVX512_OPT | NEON_OPT;
   gate->gen_merkle_root = (void*)&sha3d_gen_merkle_root;
-#if defined (KECCAK_8WAY)
+#if defined (SHA3D_8WAY)
   gate->scanhash  = (void*)&scanhash_sha3d_8way;
   gate->hash      = (void*)&sha3d_hash_8way;
-#elif defined (KECCAK_4WAY)
+#elif defined (SHA3D_4WAY)
   gate->scanhash  = (void*)&scanhash_sha3d_4way;
   gate->hash      = (void*)&sha3d_hash_4way;
+#elif defined (SHA3D_2WAY)
+  gate->scanhash  = (void*)&scanhash_sha3d_2x64;
+  gate->hash      = (void*)&sha3d_hash_2x64;
 #else
   gate->scanhash  = (void*)&scanhash_sha3d;
   gate->hash      = (void*)&sha3d_hash;

algo/keccak/keccak-gate.h

@@ -8,6 +8,16 @@
   #define KECCAK_8WAY 1
 #elif defined(__AVX2__)
   #define KECCAK_4WAY 1
+#elif defined(__SSE2__) || defined(__ARM_NEON)
+  #define KECCAK_2WAY 1
+#endif
+
+#if defined(__AVX512F__) && defined(__AVX512VL__) && defined(__AVX512DQ__) && defined(__AVX512BW__)
+  #define SHA3D_8WAY 1
+#elif defined(__AVX2__)
+  #define SHA3D_4WAY 1
+#elif defined(__SSE2__) || defined(__ARM_NEON)
+  #define SHA3D_2WAY 1
 #endif
 
 extern int hard_coded_eb;
@@ -16,27 +26,47 @@ extern int hard_coded_eb;
 
 void keccakhash_8way( void *state, const void *input );
 int scanhash_keccak_8way( struct work *work, uint32_t max_nonce,
-                         uint64_t *hashes_done, struct thr_info *mythr );
-
-void sha3d_hash_8way( void *state, const void *input );
-int scanhash_sha3d_8way( struct work *work, uint32_t max_nonce,
-                         uint64_t *hashes_done, struct thr_info *mythr );
+                          uint64_t *hashes_done, struct thr_info *mythr );
 
 #elif defined(KECCAK_4WAY)
 
 void keccakhash_4way( void *state, const void *input );
 int scanhash_keccak_4way( struct work *work, uint32_t max_nonce,
-                         uint64_t *hashes_done, struct thr_info *mythr );
+                          uint64_t *hashes_done, struct thr_info *mythr );
 
-void sha3d_hash_4way( void *state, const void *input );
-int scanhash_sha3d_4way( struct work *work, uint32_t max_nonce,
-                         uint64_t *hashes_done, struct thr_info *mythr );
+#elif defined(KECCAK_2WAY)
+
+void keccakhash_2x64( void *state, const void *input );
+int scanhash_keccak_2x64( struct work *work, uint32_t max_nonce,
+                          uint64_t *hashes_done, struct thr_info *mythr );
 
 #else
 
 void keccakhash( void *state, const void *input );
 int scanhash_keccak( struct work *work, uint32_t max_nonce,
-                    uint64_t *hashes_done, struct thr_info *mythr );
+                     uint64_t *hashes_done, struct thr_info *mythr );
+
+#endif
+
+#if defined(SHA3D_8WAY)
+
+void sha3d_hash_8way( void *state, const void *input );
+int scanhash_sha3d_8way( struct work *work, uint32_t max_nonce,
+                         uint64_t *hashes_done, struct thr_info *mythr );
+
+#elif defined(SHA3D_4WAY)
+
+void sha3d_hash_4way( void *state, const void *input );
+int scanhash_sha3d_4way( struct work *work, uint32_t max_nonce,
+                         uint64_t *hashes_done, struct thr_info *mythr );
+
+#elif defined(SHA3D_2WAY)
+
+void sha3d_hash_2x64( void *state, const void *input );
+int scanhash_sha3d_2x64( struct work *work, uint32_t max_nonce,
+                         uint64_t *hashes_done, struct thr_info *mythr );
+
+#else
 
 void sha3d_hash( void *state, const void *input );
 int scanhash_sha3d( struct work *work, uint32_t max_nonce,

algo/keccak/sha3d-4way.c

@@ -4,7 +4,7 @@
 #include <stdint.h>
 #include "keccak-hash-4way.h"
 
-#if defined(KECCAK_8WAY)
+#if defined(SHA3D_8WAY)
 
 void sha3d_hash_8way(void *state, const void *input)
 {
@@ -64,7 +64,7 @@ int scanhash_sha3d_8way( struct work *work, uint32_t max_nonce,
    return 0;
 }
 
-#elif defined(KECCAK_4WAY)
+#elif defined(SHA3D_4WAY)
 
 void sha3d_hash_4way(void *state, const void *input)
 {
@@ -122,4 +122,60 @@ int scanhash_sha3d_4way( struct work *work, uint32_t max_nonce,
    return 0;
 }
 
+#elif defined(SHA3D_2WAY)
+
+void sha3d_hash_2x64(void *state, const void *input)
+{
+    uint32_t buffer[16*4] __attribute__ ((aligned (64)));
+    keccak256_2x64_context ctx;
+
+    keccak256_2x64_init( &ctx );
+    keccak256_2x64_update( &ctx, input, 80 );
+    keccak256_2x64_close( &ctx, buffer );
+
+    keccak256_2x64_init( &ctx );
+    keccak256_2x64_update( &ctx, buffer, 32 );
+    keccak256_2x64_close( &ctx, state );
+}
+
+int scanhash_sha3d_2x64( struct work *work, uint32_t max_nonce,
+                          uint64_t *hashes_done, struct thr_info *mythr )
+{
+   uint32_t vdata[24*2] __attribute__ ((aligned (64)));
+   uint32_t hash[16*2] __attribute__ ((aligned (32)));
+   uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+   uint32_t *hash7 = &(hash[13]);   // 3*4+1
+   uint32_t *pdata = work->data;
+   uint32_t *ptarget = work->target;
+   uint32_t n = pdata[19];
+   const uint32_t first_nonce = pdata[19];
+   const uint32_t last_nonce = max_nonce - 2;
+   v128_t *noncev = (v128_t*)vdata + 9;
+   const uint32_t Htarg = ptarget[7];
+   const int thr_id = mythr->id;
+   const bool bench = opt_benchmark;
+
+   v128_bswap32_intrlv80_2x64( vdata, pdata );
+   *noncev = v128_intrlv_blend_32( v128_set32( n+1, 0, n, 0 ), *noncev );
+   do {
+      sha3d_hash_2x64( hash, vdata );
+
+      for ( int lane = 0; lane < 2; lane++ )
+      if ( unlikely( hash7[ lane<<1 ] <= Htarg && !bench ) )
+      {
+          extr_lane_2x64( lane_hash, hash, lane, 256 );
+          if ( valid_hash( lane_hash, ptarget ) )
+          {
+              pdata[19] = bswap_32( n + lane );
+              submit_solution( work, lane_hash, mythr );
+          }
+      }
+      *noncev = v128_add32( *noncev, v128_64( 0x0000000200000000 ) );
+      n += 2;
+   } while ( likely( (n < last_nonce) && !work_restart[thr_id].restart ) );
+   pdata[19] = n;
+   *hashes_done = n - first_nonce;
+   return 0;
+}
+
 #endif