Line data    Source code

       1             : #include "fd_chacha20rng.h"
       2             : #include "../../util/simd/fd_avx512.h"
       3             : #include <assert.h>
       4             : 
       5   599366640 : #define wwu_rol16(a) wwb_exch_adj_pair( (a) )
       6   599366640 : #define wwu_rol12(a) wwu_rol( (a), 12 )
       7   599366640 : #define wwu_rol7(a)  wwu_rol( (a),  7 )
       8             : 
       9             : static inline __attribute__((always_inline)) wwu_t
      10   599366640 : wwu_rol8( wwu_t x ) {
      11   599366640 :   wwb_t const mask =
      12   599366640 :     wwb_bcast_hex( 3,0,1,2, 7,4,5,6, 11,8,9,10, 15,12,13,14 );
      13   599366640 :   return _mm512_shuffle_epi8( x, mask );
      14   599366640 : }
      15             : 
      16             : void
      17     7492083 : fd_chacha20rng_refill_avx512( fd_chacha20rng_t * rng ) {
      18             : 
      19             :   /* This function should only be called if the buffer is empty. */
      20     7492083 :   if( FD_UNLIKELY( rng->buf_off != rng->buf_fill ) ) {
      21           0 :     FD_LOG_CRIT(( "refill out of sync: buf_off=%lu buf_fill=%lu", rng->buf_off, rng->buf_fill ));
      22           0 :   }
      23             : 
      24     7492083 :   wwu_t iv0  = wwu_bcast( 0x61707865U );
      25     7492083 :   wwu_t iv1  = wwu_bcast( 0x3320646eU );
      26     7492083 :   wwu_t iv2  = wwu_bcast( 0x79622d32U );
      27     7492083 :   wwu_t iv3  = wwu_bcast( 0x6b206574U );
      28     7492083 :   wwu_t zero = wwu_zero();
      29             : 
      30             :   /* Unpack key equivalent to:
      31             : 
      32             :        c4 = wwu_bcast( (uint const *)(rng->key)[0] );
      33             :        c5 = wwu_bcast( (uint const *)(rng->key)[1] );
      34             :        ...
      35             :        cB = wwu_bcast( (uint const *)(rng->key)[7] ); */
      36             : 
      37     7492083 :   wwu_t key_lo = _mm512_broadcast_i32x4( _mm_load_epi32( rng->key    ) );  /* [0,1,2,3,0,1,2,3] */
      38     7492083 :   wwu_t key_hi = _mm512_broadcast_i32x4( _mm_load_epi32( rng->key+16 ) );  /* [4,5,6,7,4,5,6,7] */
      39     7492083 :   wwu_t k0 = _mm512_shuffle_epi32( key_lo, 0x00 );
      40     7492083 :   wwu_t k1 = _mm512_shuffle_epi32( key_lo, 0x55 );
      41     7492083 :   wwu_t k2 = _mm512_shuffle_epi32( key_lo, 0xaa );
      42     7492083 :   wwu_t k3 = _mm512_shuffle_epi32( key_lo, 0xff );
      43     7492083 :   wwu_t k4 = _mm512_shuffle_epi32( key_hi, 0x00 );
      44     7492083 :   wwu_t k5 = _mm512_shuffle_epi32( key_hi, 0x55 );
      45     7492083 :   wwu_t k6 = _mm512_shuffle_epi32( key_hi, 0xaa );
      46     7492083 :   wwu_t k7 = _mm512_shuffle_epi32( key_hi, 0xff );
      47             : 
      48             :   /* Derive block index */
      49             : 
      50     7492083 :   ulong idx = rng->buf_fill / FD_CHACHA20_BLOCK_SZ;  /* really a right shift */
      51     7492083 :   wwu_t idxs = wwu_add( wwu_bcast( idx ), wwu( 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 ) );
      52             : 
      53             :   /* Run through the round function */
      54             : 
      55     7492083 :   wwu_t c0 = iv0;   wwu_t c1 = iv1;   wwu_t c2 = iv2;   wwu_t c3 = iv3;
      56     7492083 :   wwu_t c4 = k0;    wwu_t c5 = k1;    wwu_t c6 = k2;    wwu_t c7 = k3;
      57     7492083 :   wwu_t c8 = k4;    wwu_t c9 = k5;    wwu_t cA = k6;    wwu_t cB = k7;
      58     7492083 :   wwu_t cC = idxs;  wwu_t cD = zero;  wwu_t cE = zero;  wwu_t cF = zero;
      59             : 
      60     7492083 : # define QUARTER_ROUND(a,b,c,d)                                   \
      61   599366640 :   do {                                                            \
      62   599366640 :     a = wwu_add( a, b ); d = wwu_xor( d, a ); d = wwu_rol16( d ); \
      63   599366640 :     c = wwu_add( c, d ); b = wwu_xor( b, c ); b = wwu_rol12( b ); \
      64   599366640 :     a = wwu_add( a, b ); d = wwu_xor( d, a ); d = wwu_rol8( d );  \
      65   599366640 :     c = wwu_add( c, d ); b = wwu_xor( b, c ); b = wwu_rol7( b );  \
      66   599366640 :   } while(0)
      67             : 
      68    82412913 :   for( ulong i=0UL; i<10UL; i++ ) {
      69    74920830 :     QUARTER_ROUND( c0, c4, c8, cC );
      70    74920830 :     QUARTER_ROUND( c1, c5, c9, cD );
      71    74920830 :     QUARTER_ROUND( c2, c6, cA, cE );
      72    74920830 :     QUARTER_ROUND( c3, c7, cB, cF );
      73    74920830 :     QUARTER_ROUND( c0, c5, cA, cF );
      74    74920830 :     QUARTER_ROUND( c1, c6, cB, cC );
      75    74920830 :     QUARTER_ROUND( c2, c7, c8, cD );
      76    74920830 :     QUARTER_ROUND( c3, c4, c9, cE );
      77    74920830 :   }
      78     7492083 : # undef QUARTER_ROUND
      79             : 
      80             :   /* Finalize */
      81             : 
      82     7492083 :   c0 = wwu_add( c0, iv0  );
      83     7492083 :   c1 = wwu_add( c1, iv1  );
      84     7492083 :   c2 = wwu_add( c2, iv2  );
      85     7492083 :   c3 = wwu_add( c3, iv3  );
      86     7492083 :   c4 = wwu_add( c4, k0   );
      87     7492083 :   c5 = wwu_add( c5, k1   );
      88     7492083 :   c6 = wwu_add( c6, k2   );
      89     7492083 :   c7 = wwu_add( c7, k3   );
      90     7492083 :   c8 = wwu_add( c8, k4   );
      91     7492083 :   c9 = wwu_add( c9, k5   );
      92     7492083 :   cA = wwu_add( cA, k6   );
      93     7492083 :   cB = wwu_add( cB, k7   );
      94     7492083 :   cC = wwu_add( cC, idxs );
      95             :   //cD = wwu_add( cD, zero );
      96             :   //cE = wwu_add( cE, zero );
      97             :   //cF = wwu_add( cF, zero );
      98             : 
      99             :   /* Transpose matrix to get output vector */
     100             : 
     101     7492083 :   wwu_transpose_16x16( c0, c1, c2, c3, c4, c5, c6, c7,
     102     7492083 :                        c8, c9, cA, cB, cC, cD, cE, cF,
     103     7492083 :                        c0, c1, c2, c3, c4, c5, c6, c7,
     104     7492083 :                        c8, c9, cA, cB, cC, cD, cE, cF );
     105             : 
     106             :   /* Update ring buffer */
     107             : 
     108     7492083 :   uint * out = (uint *)rng->buf;
     109     7492083 :   wwu_st( out+0x00, c0 ); wwu_st( out+0x10, c1 );
     110     7492083 :   wwu_st( out+0x20, c2 ); wwu_st( out+0x30, c3 );
     111     7492083 :   wwu_st( out+0x40, c4 ); wwu_st( out+0x50, c5 );
     112     7492083 :   wwu_st( out+0x60, c6 ); wwu_st( out+0x70, c7 );
     113     7492083 :   wwu_st( out+0x80, c8 ); wwu_st( out+0x90, c9 );
     114     7492083 :   wwu_st( out+0xa0, cA ); wwu_st( out+0xb0, cB );
     115     7492083 :   wwu_st( out+0xc0, cC ); wwu_st( out+0xd0, cD );
     116     7492083 :   wwu_st( out+0xe0, cE ); wwu_st( out+0xf0, cF );
     117             : 
     118             :   /* Update ring descriptor */
     119             : 
     120     7492083 :   rng->buf_fill += 16*FD_CHACHA20_BLOCK_SZ;
     121     7492083 : }