v3.22.2

algo/luffa/luffa-hash-2way.c

@@ -554,20 +554,10 @@ int luffa_4way_update_close( luffa_4way_context *state,
     a = _mm256_xor_si256( a, c0 ); \
     b = _mm256_xor_si256( b, c1 );
 
-/*
-#define MULT2( a0, a1, mask ) \
-do { \
-  __m256i b = _mm256_xor_si256( a0, \
-                   _mm256_shuffle_epi32( _mm256_and_si256(a1,mask), 16 ) ); \
-  a0 = _mm256_or_si256( _mm256_srli_si256(b,4), _mm256_slli_si256(a1,12) ); \
-  a1 = _mm256_or_si256( _mm256_srli_si256(a1,4), _mm256_slli_si256(b,12) );  \
-} while(0)
-*/
-
-#define MULT2( a0, a1, mask ) \
+#define MULT2( a0, a1 ) \
 { \
-  __m256i b = _mm256_xor_si256( a0, \
-                 _mm256_shuffle_epi32( _mm256_and_si256( a1, mask ), 16 ) ); \
+  __m256i b = _mm256_xor_si256( a0, _mm256_shuffle_epi32( \
+                         _mm256_blend_epi32( a1, m256_zero, 0xee ), 16 ) ); \
   a0 = _mm256_alignr_epi8( a1,  b, 4 ); \
   a1 = _mm256_alignr_epi8(  b, a1, 4 ); \
 }
@@ -682,7 +672,6 @@ void rnd512_2way( luffa_2way_context *state, __m256i *msg )
     __m256i *chainv = state->chainv;
     __m256i msg0, msg1;
     __m256i x0, x1, x2, x3, x4, x5, x6, x7;
-    const __m256i MASK = m256_const1_i128( 0xffffffff );
 
     t0 = chainv[0];
     t1 = chainv[1];
@@ -696,7 +685,7 @@ void rnd512_2way( luffa_2way_context *state, __m256i *msg )
     t0 = _mm256_xor_si256( t0, chainv[8] );
     t1 = _mm256_xor_si256( t1, chainv[9] );
 
-    MULT2( t0, t1, MASK );
+    MULT2( t0, t1 );
 
     msg0 = _mm256_shuffle_epi32( msg[0], 27 );
     msg1 = _mm256_shuffle_epi32( msg[1], 27 );
@@ -715,66 +704,66 @@ void rnd512_2way( luffa_2way_context *state, __m256i *msg )
     t0 = chainv[0];
     t1 = chainv[1];
 
-    MULT2( chainv[0], chainv[1], MASK );
+    MULT2( chainv[0], chainv[1] );
     chainv[0] = _mm256_xor_si256( chainv[0], chainv[2] );
     chainv[1] = _mm256_xor_si256( chainv[1], chainv[3] );
 
-    MULT2( chainv[2], chainv[3], MASK );
+    MULT2( chainv[2], chainv[3] );
     chainv[2] = _mm256_xor_si256(chainv[2], chainv[4]);
     chainv[3] = _mm256_xor_si256(chainv[3], chainv[5]);
 
-    MULT2( chainv[4], chainv[5], MASK );
+    MULT2( chainv[4], chainv[5] );
     chainv[4] = _mm256_xor_si256(chainv[4], chainv[6]);
     chainv[5] = _mm256_xor_si256(chainv[5], chainv[7]);
 
-    MULT2( chainv[6], chainv[7], MASK );
+    MULT2( chainv[6], chainv[7] );
     chainv[6] = _mm256_xor_si256(chainv[6], chainv[8]);
     chainv[7] = _mm256_xor_si256(chainv[7], chainv[9]);
 
-    MULT2( chainv[8], chainv[9], MASK );
+    MULT2( chainv[8], chainv[9] );
     chainv[8] = _mm256_xor_si256( chainv[8], t0 );
     chainv[9] = _mm256_xor_si256( chainv[9], t1 );
 
     t0 = chainv[8];
     t1 = chainv[9];
 
-    MULT2( chainv[8], chainv[9], MASK );
+    MULT2( chainv[8], chainv[9] );
     chainv[8] = _mm256_xor_si256( chainv[8], chainv[6] );
     chainv[9] = _mm256_xor_si256( chainv[9], chainv[7] );
 
-    MULT2( chainv[6], chainv[7], MASK );
+    MULT2( chainv[6], chainv[7] );
     chainv[6] = _mm256_xor_si256( chainv[6], chainv[4] );
     chainv[7] = _mm256_xor_si256( chainv[7], chainv[5] );
 
-    MULT2( chainv[4], chainv[5], MASK );
+    MULT2( chainv[4], chainv[5] );
     chainv[4] = _mm256_xor_si256( chainv[4], chainv[2] );
     chainv[5] = _mm256_xor_si256( chainv[5], chainv[3] );
 
-    MULT2( chainv[2], chainv[3], MASK );
+    MULT2( chainv[2], chainv[3] );
     chainv[2] = _mm256_xor_si256( chainv[2], chainv[0] );
     chainv[3] = _mm256_xor_si256( chainv[3], chainv[1] );
 
-    MULT2( chainv[0], chainv[1], MASK );
+    MULT2( chainv[0], chainv[1] );
     chainv[0] = _mm256_xor_si256( _mm256_xor_si256( chainv[0], t0 ), msg0 );
     chainv[1] = _mm256_xor_si256( _mm256_xor_si256( chainv[1], t1 ), msg1 );
 
-    MULT2( msg0, msg1, MASK );
+    MULT2( msg0, msg1 );
     chainv[2] = _mm256_xor_si256( chainv[2], msg0 );
     chainv[3] = _mm256_xor_si256( chainv[3], msg1 );
 
-    MULT2( msg0, msg1, MASK );
+    MULT2( msg0, msg1 );
     chainv[4] = _mm256_xor_si256( chainv[4], msg0 );
     chainv[5] = _mm256_xor_si256( chainv[5], msg1 );
 
-    MULT2( msg0, msg1, MASK );
+    MULT2( msg0, msg1 );
     chainv[6] = _mm256_xor_si256( chainv[6], msg0 );
     chainv[7] = _mm256_xor_si256( chainv[7], msg1 );
 
-    MULT2( msg0, msg1, MASK );
+    MULT2( msg0, msg1 );
     chainv[8] = _mm256_xor_si256( chainv[8], msg0 );
     chainv[9] = _mm256_xor_si256( chainv[9], msg1 );
 
-    MULT2( msg0, msg1, MASK );
+    MULT2( msg0, msg1 );
 
     chainv[3] = mm256_rol_32( chainv[3], 1 );
     chainv[5] = mm256_rol_32( chainv[5], 2 );

algo/lyra2/lyra2rev2-4way.c

@@ -75,7 +75,7 @@ void lyra2rev2_16way_hash( void *state, const void *input )
    keccak256_8way_close( &ctx.keccak, vhash );
 
    dintrlv_8x64( hash8,  hash9,  hash10,  hash11,
-                 hash12, hash13, hash14, hash5, vhash, 256 );
+                 hash12, hash13, hash14, hash15, vhash, 256 );
 
    cubehash_full( &ctx.cube, (byte*) hash0,  256, (const byte*) hash0,  32 );
    cubehash_full( &ctx.cube, (byte*) hash1,  256, (const byte*) hash1,  32 );

algo/sha/sha256dt.c

@@ -0,0 +1,268 @@
+#include "algo-gate-api.h"
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdio.h>
+#include "sha-hash-4way.h"
+
+#if defined(__AVX512F__) && defined(__AVX512VL__) && defined(__AVX512DQ__) && defined(__AVX512BW__)
+  #define SHA256DT_16WAY 1
+#elif defined(__AVX2__)
+  #define SHA256DT_8WAY 1
+#else
+  #define SHA256DT_4WAY 1
+#endif
+
+#if defined(SHA256DT_16WAY)
+
+int scanhash_sha256dt_16way( struct work *work, const uint32_t max_nonce,
+                           uint64_t *hashes_done, struct thr_info *mythr )
+{
+   __m512i  vdata[32]    __attribute__ ((aligned (128)));
+   __m512i  block[16]    __attribute__ ((aligned (64)));
+   __m512i  hash32[8]    __attribute__ ((aligned (64)));
+   __m512i  initstate[8] __attribute__ ((aligned (64)));
+   __m512i  midstate1[8] __attribute__ ((aligned (64)));
+   __m512i  midstate2[8] __attribute__ ((aligned (64)));
+   __m512i  mexp_pre[16] __attribute__ ((aligned (64)));
+   uint32_t lane_hash[8] __attribute__ ((aligned (64)));
+   uint32_t *hash32_d7 =  (uint32_t*)&( hash32[7] );
+   uint32_t *pdata = work->data;
+   const uint32_t *ptarget = work->target;
+   const uint32_t targ32_d7 = ptarget[7];
+   const uint32_t first_nonce = pdata[19];
+   const uint32_t last_nonce = max_nonce - 16;
+   uint32_t n = first_nonce;
+   __m512i *noncev = vdata + 19; 
+   const int thr_id = mythr->id;
+   const bool bench = opt_benchmark;
+   const __m512i last_byte = m512_const1_32( 0x80000000 );
+   const __m512i sixteen = m512_const1_32( 16 );
+
+   for ( int i = 0; i < 19; i++ )
+      vdata[i] = mm512_bcast_i32( pdata[i] );
+
+   *noncev = _mm512_set_epi32( n+15, n+14, n+13, n+12, n+11, n+10, n+9, n+8,
+                               n+ 7, n+ 6, n+ 5, n+ 4, n+ 3, n+ 2, n+1, n );
+
+   vdata[16+4] = last_byte;
+   memset_zero_512( vdata+16 + 5, 10 );
+   vdata[16+15] = mm512_bcast_i32( 0x480 ); 
+   
+   block[ 8] = last_byte;
+   memset_zero_512( block + 9, 6 );
+   block[15] = mm512_bcast_i32( 0x300 ); 
+   
+   initstate[0] = mm512_bcast_i64( 0xdfa9bf2cdfa9bf2c );
+   initstate[1] = mm512_bcast_i64( 0xb72074d4b72074d4 );
+   initstate[2] = mm512_bcast_i64( 0x6bb011226bb01122 );
+   initstate[3] = mm512_bcast_i64( 0xd338e869d338e869 );
+   initstate[4] = mm512_bcast_i64( 0xaa3ff126aa3ff126 );
+   initstate[5] = mm512_bcast_i64( 0x475bbf30475bbf30 );
+   initstate[6] = mm512_bcast_i64( 0x8fd52e5b8fd52e5b );
+   initstate[7] = mm512_bcast_i64( 0x9f75c9ad9f75c9ad );
+
+   sha256_16way_transform_le( midstate1, vdata, initstate );
+   
+   // Do 3 rounds on the first 12 bytes of the next block
+   sha256_16way_prehash_3rounds( midstate2, mexp_pre, vdata+16, midstate1 );
+
+   do
+   {
+      sha256_16way_final_rounds( block, vdata+16, midstate1, midstate2,
+                                 mexp_pre );
+      sha256_16way_transform_le( hash32, block, initstate );
+      mm512_block_bswap_32( hash32, hash32 );    
+
+      for ( int lane = 0; lane < 16; lane++ )
+      if ( hash32_d7[ lane ] <= targ32_d7 )
+      {
+         extr_lane_16x32( lane_hash, hash32, lane, 256 );
+         if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
+         {
+            pdata[19] = n + lane;
+            submit_solution( work, lane_hash, mythr );
+         }
+      }
+      *noncev = _mm512_add_epi32( *noncev, sixteen );
+      n += 16;
+   } while ( (n < last_nonce) && !work_restart[thr_id].restart );
+   pdata[19] = n;
+   *hashes_done = n - first_nonce;
+   return 0;
+}
+
+
+#endif
+
+#if defined(SHA256DT_8WAY)
+
+int scanhash_sha256dt_8way( struct work *work, const uint32_t max_nonce,
+                           uint64_t *hashes_done, struct thr_info *mythr )
+{
+   __m256i  vdata[32]    __attribute__ ((aligned (64)));
+   __m256i  block[16]    __attribute__ ((aligned (32)));
+   __m256i  hash32[8]    __attribute__ ((aligned (32)));
+   __m256i  initstate[8] __attribute__ ((aligned (32)));
+   __m256i  midstate1[8] __attribute__ ((aligned (32)));
+   __m256i  midstate2[8] __attribute__ ((aligned (32)));
+   __m256i  mexp_pre[16] __attribute__ ((aligned (32)));
+   uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+   uint32_t *hash32_d7 =  (uint32_t*)&( hash32[7] );
+   uint32_t *pdata = work->data;
+   const uint32_t *ptarget = work->target;
+   const uint32_t targ32_d7 = ptarget[7];
+   const uint32_t first_nonce = pdata[19];
+   const uint32_t last_nonce = max_nonce - 8;
+   uint32_t n = first_nonce;
+   __m256i *noncev = vdata + 19;
+   const int thr_id = mythr->id;
+   const bool bench = opt_benchmark;
+   const __m256i last_byte = m256_const1_32( 0x80000000 );
+   const __m256i eight = m256_const1_32( 8 );
+
+   for ( int i = 0; i < 19; i++ )
+      vdata[i] = mm256_bcast_i32( pdata[i] );
+
+   *noncev = _mm256_set_epi32( n+ 7, n+ 6, n+ 5, n+ 4, n+ 3, n+ 2, n+1, n );
+
+   vdata[16+4] = last_byte;
+   memset_zero_256( vdata+16 + 5, 10 );
+   vdata[16+15] = mm256_bcast_i32( 0x480 );
+
+   block[ 8] = last_byte;
+   memset_zero_256( block + 9, 6 );
+   block[15] = mm256_bcast_i32( 0x300 ); 
+   
+   // initialize state
+   initstate[0] = mm256_bcast_i64( 0xdfa9bf2cdfa9bf2c );
+   initstate[1] = mm256_bcast_i64( 0xb72074d4b72074d4 );
+   initstate[2] = mm256_bcast_i64( 0x6bb011226bb01122 );
+   initstate[3] = mm256_bcast_i64( 0xd338e869d338e869 );
+   initstate[4] = mm256_bcast_i64( 0xaa3ff126aa3ff126 );
+   initstate[5] = mm256_bcast_i64( 0x475bbf30475bbf30 );
+   initstate[6] = mm256_bcast_i64( 0x8fd52e5b8fd52e5b );
+   initstate[7] = mm256_bcast_i64( 0x9f75c9ad9f75c9ad );
+
+   sha256_8way_transform_le( midstate1, vdata, initstate );
+
+   // Do 3 rounds on the first 12 bytes of the next block
+   sha256_8way_prehash_3rounds( midstate2, mexp_pre, vdata + 16, midstate1 );
+   
+   do
+   {
+      sha256_8way_final_rounds( block, vdata+16, midstate1, midstate2,
+                                mexp_pre );
+      sha256_8way_transform_le( hash32, block, initstate );
+      mm256_block_bswap_32( hash32, hash32 );
+
+      for ( int lane = 0; lane < 8; lane++ )
+      if ( hash32_d7[ lane ] <= targ32_d7 )
+      {
+         extr_lane_8x32( lane_hash, hash32, lane, 256 );
+         if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
+         {
+            pdata[19] = n + lane;
+            submit_solution( work, lane_hash, mythr );
+         }
+      }
+      *noncev = _mm256_add_epi32( *noncev, eight );
+      n += 8;
+   } while ( (n < last_nonce) && !work_restart[thr_id].restart );
+   pdata[19] = n;
+   *hashes_done = n - first_nonce;
+   return 0;
+}
+
+#endif
+
+
+#if defined(SHA256DT_4WAY)
+
+int scanhash_sha256dt_4way( struct work *work, const uint32_t max_nonce,
+                           uint64_t *hashes_done, struct thr_info *mythr )
+{
+   __m128i  vdata[32]    __attribute__ ((aligned (64)));
+   __m128i  block[16]    __attribute__ ((aligned (32)));
+   __m128i  hash32[8]    __attribute__ ((aligned (32)));
+   __m128i  initstate[8] __attribute__ ((aligned (32)));
+   __m128i  midstate[8]  __attribute__ ((aligned (32)));
+   uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+   uint32_t *hash32_d7 =  (uint32_t*)&( hash32[7] );
+   uint32_t *pdata = work->data;
+   const uint32_t *ptarget = work->target;
+   const uint32_t targ32_d7 = ptarget[7];
+   const uint32_t first_nonce = pdata[19];
+   const uint32_t last_nonce = max_nonce - 4;
+   uint32_t n = first_nonce;
+   __m128i *noncev = vdata + 19;
+   const int thr_id = mythr->id;
+   const bool bench = opt_benchmark;
+   const __m128i last_byte = m128_const1_32( 0x80000000 );
+   const __m128i four = m128_const1_32( 4 );
+
+   for ( int i = 0; i < 19; i++ )
+       vdata[i] = mm128_bcast_i32( pdata[i] );
+
+   *noncev = _mm_set_epi32( n+ 3, n+ 2, n+1, n );
+
+   vdata[16+4] = last_byte;
+   memset_zero_128( vdata+16 + 5, 10 );
+   vdata[16+15] = mm128_bcast_i32( 0x480 );
+
+   block[ 8] = last_byte;
+   memset_zero_128( block + 9, 6 );
+   block[15] = mm128_bcast_i32( 0x300 );
+   
+   // initialize state
+   initstate[0] = mm128_bcast_i64( 0xdfa9bf2cdfa9bf2c );
+   initstate[1] = mm128_bcast_i64( 0xb72074d4b72074d4 );
+   initstate[2] = mm128_bcast_i64( 0x6bb011226bb01122 );
+   initstate[3] = mm128_bcast_i64( 0xd338e869d338e869 );
+   initstate[4] = mm128_bcast_i64( 0xaa3ff126aa3ff126 );
+   initstate[5] = mm128_bcast_i64( 0x475bbf30475bbf30 );
+   initstate[6] = mm128_bcast_i64( 0x8fd52e5b8fd52e5b );
+   initstate[7] = mm128_bcast_i64( 0x9f75c9ad9f75c9ad );
+
+   // hash first 64 bytes of data
+   sha256_4way_transform_le( midstate, vdata, initstate );
+
+   do
+   {
+      sha256_4way_transform_le( block,  vdata+16, midstate  );
+      sha256_4way_transform_le( hash32, block,    initstate );
+      mm128_block_bswap_32( hash32, hash32 );
+
+      for ( int lane = 0; lane < 4; lane++ )
+      if ( unlikely( hash32_d7[ lane ] <= targ32_d7 ) )
+      {
+         extr_lane_4x32( lane_hash, hash32, lane, 256 );
+         if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
+         {
+            pdata[19] = n + lane;
+            submit_solution( work, lane_hash, mythr );
+         }
+       }
+       *noncev = _mm_add_epi32( *noncev, four );
+       n += 4;
+   } while ( (n < last_nonce) && !work_restart[thr_id].restart );
+   pdata[19] = n;
+   *hashes_done = n - first_nonce;
+   return 0;
+}
+
+#endif
+
+bool register_sha256dt_algo( algo_gate_t* gate )
+{
+    gate->optimizations = SSE2_OPT | AVX2_OPT | AVX512_OPT;
+#if defined(SHA256DT_16WAY)
+    gate->scanhash   = (void*)&scanhash_sha256dt_16way;
+#elif defined(SHA256DT_8WAY)
+    gate->scanhash   = (void*)&scanhash_sha256dt_8way;
+#else
+    gate->scanhash   = (void*)&scanhash_sha256dt_4way;
+#endif
+    return true;
+}
+

algo/sha/sha512256d-4way.c

@@ -0,0 +1,221 @@
+#include "algo-gate-api.h"
+#include "sha-hash-4way.h"
+#include <string.h>
+#include <stdint.h>
+
+#if defined(__AVX512F__) && defined(__AVX512VL__) && defined(__AVX512DQ__) && defined(__AVX512BW__)
+#define SHA512256D_8WAY 1
+#elif defined(__AVX2__)
+#define SHA512256D_4WAY 1
+#endif
+
+#if defined(SHA512256D_8WAY)
+
+static void sha512256d_8way_init( sha512_8way_context *ctx )
+{
+  ctx->count = 0;
+  ctx->initialized = true;
+  ctx->val[0] = mm512_bcast_i64( 0x22312194FC2BF72C );
+  ctx->val[1] = mm512_bcast_i64( 0x9F555FA3C84C64C2 );
+  ctx->val[2] = mm512_bcast_i64( 0x2393B86B6F53B151 );
+  ctx->val[3] = mm512_bcast_i64( 0x963877195940EABD );
+  ctx->val[4] = mm512_bcast_i64( 0x96283EE2A88EFFE3 );
+  ctx->val[5] = mm512_bcast_i64( 0xBE5E1E2553863992 );
+  ctx->val[6] = mm512_bcast_i64( 0x2B0199FC2C85B8AA );
+  ctx->val[7] = mm512_bcast_i64( 0x0EB72DDC81C52CA2 );
+}
+
+int scanhash_sha512256d_8way( struct work *work, uint32_t max_nonce,
+                          uint64_t *hashes_done, struct thr_info *mythr )
+{
+    uint64_t hash[8*8] __attribute__ ((aligned (128)));
+    uint32_t vdata[20*8] __attribute__ ((aligned (64)));
+    sha512_8way_context ctx; 
+    uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+    uint64_t *hash_q3 = &(hash[3*8]);
+    uint32_t *pdata = work->data;
+    uint32_t *ptarget = work->target;
+    const uint64_t targ_q3 = ((uint64_t*)ptarget)[3];
+    const uint32_t first_nonce = pdata[19];
+    const uint32_t last_nonce = max_nonce - 8;
+    uint32_t n = first_nonce;
+    __m512i  *noncev = (__m512i*)vdata + 9;
+    const int thr_id = mythr->id;
+    const bool bench = opt_benchmark;
+    const __m512i eight = mm512_bcast_i64( 0x0000000800000000 );
+
+    mm512_bswap32_intrlv80_8x64( vdata, pdata );
+    *noncev = mm512_intrlv_blend_32(
+                _mm512_set_epi32( n+7, 0, n+6, 0, n+5, 0, n+4, 0,
+                                  n+3, 0, n+2, 0, n+1, 0, n  , 0 ), *noncev );
+    do
+    {
+       sha512256d_8way_init( &ctx );
+       sha512_8way_update( &ctx, vdata, 80 );
+       sha512_8way_close( &ctx, hash );        
+
+       sha512256d_8way_init( &ctx );
+       sha512_8way_update( &ctx, hash, 32 );
+       sha512_8way_close( &ctx, hash );
+
+       for ( int lane = 0; lane < 8; lane++ )
+       if ( unlikely( hash_q3[ lane ] <= targ_q3 && !bench ) )
+       {
+          extr_lane_8x64( lane_hash, hash, lane, 256 );
+          if ( valid_hash( lane_hash, ptarget ) && !bench )
+          {
+             pdata[19] = bswap_32( n + lane );
+             submit_solution( work, lane_hash, mythr );
+          }
+       }
+       *noncev = _mm512_add_epi32( *noncev, eight );
+       n += 8;
+    } while ( likely( (n < last_nonce) && !work_restart[thr_id].restart ) );
+
+    pdata[19] = n;
+    *hashes_done = n - first_nonce;
+    return 0;
+}
+
+#elif defined(SHA512256D_4WAY)
+
+static void sha512256d_4way_init( sha512_4way_context *ctx )
+{
+  ctx->count = 0;
+  ctx->initialized = true;
+  ctx->val[0] = mm256_bcast_i64( 0x22312194FC2BF72C );
+  ctx->val[1] = mm256_bcast_i64( 0x9F555FA3C84C64C2 );
+  ctx->val[2] = mm256_bcast_i64( 0x2393B86B6F53B151 );
+  ctx->val[3] = mm256_bcast_i64( 0x963877195940EABD );
+  ctx->val[4] = mm256_bcast_i64( 0x96283EE2A88EFFE3 );
+  ctx->val[5] = mm256_bcast_i64( 0xBE5E1E2553863992 );
+  ctx->val[6] = mm256_bcast_i64( 0x2B0199FC2C85B8AA );
+  ctx->val[7] = mm256_bcast_i64( 0x0EB72DDC81C52CA2 );
+}
+
+int scanhash_sha512256d_4way( struct work *work, uint32_t max_nonce,
+                          uint64_t *hashes_done, struct thr_info *mythr )
+{
+    uint64_t hash[8*4] __attribute__ ((aligned (64)));
+    uint32_t vdata[20*4] __attribute__ ((aligned (64)));
+    sha512_4way_context ctx;
+    uint32_t lane_hash[8] __attribute__ ((aligned (32)));
+    uint64_t *hash_q3 = &(hash[3*4]);
+    uint32_t *pdata = work->data;
+    uint32_t *ptarget = work->target;
+    const uint64_t targ_q3 = ((uint64_t*)ptarget)[3];
+    const uint32_t first_nonce = pdata[19];
+    const uint32_t last_nonce = max_nonce - 4;
+    uint32_t n = first_nonce;
+    __m256i  *noncev = (__m256i*)vdata + 9;
+    const int thr_id = mythr->id;
+    const bool bench = opt_benchmark;
+    const __m256i four = mm256_bcast_i64( 0x0000000400000000 );
+
+    mm256_bswap32_intrlv80_4x64( vdata, pdata );
+    *noncev = mm256_intrlv_blend_32(
+                _mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ), *noncev );
+    do
+    {
+       sha512256d_4way_init( &ctx );
+       sha512_4way_update( &ctx, vdata, 80 );
+       sha512_4way_close( &ctx, hash );
+
+       sha512256d_4way_init( &ctx );
+       sha512_4way_update( &ctx, hash, 32 );
+       sha512_4way_close( &ctx, hash );
+
+       for ( int lane = 0; lane < 4; lane++ )
+       if ( hash_q3[ lane ] <= targ_q3 )
+       {
+          extr_lane_4x64( lane_hash, hash, lane, 256 );
+          if ( valid_hash( lane_hash, ptarget ) && !bench )
+          {
+             pdata[19] = bswap_32( n + lane );
+             submit_solution( work, lane_hash, mythr );
+          }
+       }
+       *noncev = _mm256_add_epi32( *noncev, four );
+       n += 4;
+    } while ( (n < last_nonce) && !work_restart[thr_id].restart );
+
+    pdata[19] = n;
+    *hashes_done = n - first_nonce;
+    return 0;
+}
+
+#else
+
+#include "sph_sha2.h"
+
+static const uint64_t H512_256[8] =
+{
+   0x22312194FC2BF72C, 0x9F555FA3C84C64C2,
+   0x2393B86B6F53B151, 0x963877195940EABD,
+   0x96283EE2A88EFFE3, 0xBE5E1E2553863992,
+   0x2B0199FC2C85B8AA, 0x0EB72DDC81C52CA2,
+};
+
+static void sha512256d_init( sph_sha512_context *ctx )
+{
+   memcpy( ctx->val, H512_256, sizeof H512_256 );
+   ctx->count = 0;
+}
+
+int scanhash_sha512256d( struct work *work,   uint32_t max_nonce,
+                     uint64_t *hashes_done, struct thr_info *mythr )
+{
+   uint32_t *pdata = work->data;
+   uint32_t *ptarget = work->target;
+   uint32_t hash64[8] __attribute__ ((aligned (64)));
+   uint32_t endiandata[20] __attribute__ ((aligned (64)));
+   sph_sha512_context ctx;
+   const uint32_t Htarg = ptarget[7];
+   const uint32_t first_nonce = pdata[19];
+   uint32_t n = first_nonce;
+   int thr_id = mythr->id;
+
+   swab32_array( endiandata, pdata, 20 );
+
+   do {
+      be32enc( &endiandata[19], n );
+
+      sha512256d_init( &ctx );
+      sph_sha512( &ctx, endiandata, 80 );
+      sph_sha512_close( &ctx, hash64 );
+
+      sha512256d_init( &ctx );
+      sph_sha512( &ctx, hash64, 32 );
+      sph_sha512_close( &ctx, hash64 );
+      
+      if ( hash64[7] <= Htarg )
+      if ( fulltest( hash64, ptarget ) && !opt_benchmark )
+      {
+         pdata[19] = n;
+         submit_solution( work, hash64, mythr );
+      }
+      n++;
+
+   } while (n < max_nonce && !work_restart[thr_id].restart);
+
+   *hashes_done = n - first_nonce + 1;
+   pdata[19] = n;
+
+   return 0;
+}
+
+#endif
+
+bool register_sha512256d_algo( algo_gate_t* gate )
+{
+   gate->optimizations = AVX2_OPT | AVX512_OPT;
+#if defined(SHA512256D_8WAY)
+   gate->scanhash = (void*)&scanhash_sha512256d_8way;
+#elif defined(SHA512256D_4WAY)
+   gate->scanhash = (void*)&scanhash_sha512256d_4way;
+#else
+   gate->scanhash = (void*)&scanhash_sha512256d;
+#endif
+   return true;
+};
+