popov/cpuminer-opt-gpu
1
0
Fork 0
mirror of https://github.com/JayDDee/cpuminer-opt.git synced 2025-09-17 23:44:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

v3.9.5.4

Browse Source
This commit is contained in:
Jay D Dee
2019-07-15 17:00:26 -04:00
parent e625ed5420
commit e2d5762ef2
63 changed files with 1973 additions and 2980 deletions
Download Patch File Download Diff File Expand all files Collapse all files

911
simd-utils/intrlv-avx.h
View File

@@ -1,911 +0,0 @@
#if !defined(INTRLV_AVX_H__)
#define INTRLV_AVX_H__ 1
// philosophical discussion
//
// transitions:
//
// int32 <-> int64
// uint64_t = (uint64_t)int32_lo | ( (uint64_t)int32_hi << 32 )
// Efficient transition and post processing, 32 bit granularity is lost.
//
// int32 <-> m64
// More complex, 32 bit granularity maintained, limited number of mmx regs.
// int32 <-> int64 <-> m64 might be more efficient.
//
// int32 <-> m128
// Expensive, current implementation.
//
// int32 <-> m256
// Very expensive multi stage, current implementation.
//
// int64/m64 <-> m128
// Efficient, agnostic to native element size. Common.
//
// m128 <-> m256
// Expensive for a single instruction, unavoidable. Common.
//
// Multi stage options
//
// int32 <-> int64 -> m128
// More efficient than insert32, granularity maintained. Common.
//
// int64 <-> m128 -> m256
// Unavoidable, reasonably efficient. Common
//
// int32 <-> int64 -> m128 -> m256
// Seems inevitable, most efficient despite number of stages. Common.
//
// Implementation plan.
//
// 1. Complete m128 <-> m256
// 2. Implement int64 <-> m128
// 3. Combine int64 <-> m128 <-> m256
// 4. Implement int32 <-> int64 <-> m128
// 5. Combine int32 <-> int64 <-> m128 <-> m256
//
#if defined(__AVX__)
// Convenient short cuts for local use only
// Extract 64 bits from the low 128 bits of 256 bit vector.
#define extr64_cast128_256( a, n ) \
_mm_extract_epi64( _mm256_castsi256_si128( a ), n )
// Extract 32 bits from the low 128 bits of 256 bit vector.
#define extr32_cast128_256( a, n ) \
_mm_extract_epi32( _mm256_castsi256_si128( a ), n )
///////////////////////////////////////////////////////////
//
// AVX 256 Bit Vectors
//
// 256 bit interleaving can be done with AVX.
#define mm256_put_64( s0, s1, s2, s3) \
_mm256_set_epi64x( *((const uint64_t*)(s3)), *((const uint64_t*)(s2)), \
*((const uint64_t*)(s1)), *((const uint64_t*)(s0)) )
#define mm256_put_32( s00, s01, s02, s03, s04, s05, s06, s07 ) \
_mm256_set_epi32( *((const uint32_t*)(s07)), *((const uint32_t*)(s06)), \
*((const uint32_t*)(s05)), *((const uint32_t*)(s04)), \
*((const uint32_t*)(s03)), *((const uint32_t*)(s02)), \
*((const uint32_t*)(s01)), *((const uint32_t*)(s00)) )
#define mm256_get_64( s, i0, i1, i2, i3 ) \
_mm256_set_epi64x( ((const uint64_t*)(s))[i3], ((const uint64_t*)(s))[i2], \
((const uint64_t*)(s))[i1], ((const uint64_t*)(s))[i0] )
#define mm256_get_32( s, i0, i1, i2, i3, i4, i5, i6, i7 ) \
_mm256_set_epi32( ((const uint32_t*)(s))[i7], ((const uint32_t*)(s))[i6], \
((const uint32_t*)(s))[i5], ((const uint32_t*)(s))[i4], \
((const uint32_t*)(s))[i3], ((const uint32_t*)(s))[i2], \
((const uint32_t*)(s))[i1], ((const uint32_t*)(s))[i0] )
/*
// Blend 2 vectors alternating hi & lo: { hi[n], lo[n-1], ... hi[1], lo[0] }
#define mm256_intrlv_blend_128( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x0f )
#define mm256_intrlv_blend_64( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x33 )
#define mm256_intrlv_blend_32( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x55 )
*/
// Interleave 8x32_256
#define mm256_intrlv_8x32_256( d, s0, s1, s2, s3, s4, s5, s6, s7 ) \
{ \
__m128i s0hi = mm128_extr_hi128_256( s0 ); \
__m128i s1hi = mm128_extr_hi128_256( s1 ); \
__m128i s2hi = mm128_extr_hi128_256( s2 ); \
__m128i s3hi = mm128_extr_hi128_256( s3 ); \
__m128i s4hi = mm128_extr_hi128_256( s4 ); \
__m128i s5hi = mm128_extr_hi128_256( s5 ); \
__m128i s6hi = mm128_extr_hi128_256( s6 ); \
__m128i s7hi = mm128_extr_hi128_256( s7 ); \
casti_m256i( d,0 ) = _mm256_set_epi32( \
extr32_cast128_256(s7,0), extr32_cast128_256(s6,0), \
extr32_cast128_256(s5,0), extr32_cast128_256(s4,0), \
extr32_cast128_256(s3,0), extr32_cast128_256(s2,0), \
extr32_cast128_256(s1,0), extr32_cast128_256(s0,0) ); \
casti_m256i( d,1 ) = _mm256_set_epi32( \
extr32_cast128_256(s7,1), extr32_cast128_256(s6,1), \
extr32_cast128_256(s5,1), extr32_cast128_256(s4,1), \
extr32_cast128_256(s3,1), extr32_cast128_256(s2,1), \
extr32_cast128_256(s1,1), extr32_cast128_256(s0,1) ); \
casti_m256i( d,2 ) = _mm256_set_epi32( \
extr32_cast128_256(s7,2), extr32_cast128_256(s6,2), \
extr32_cast128_256(s5,2), extr32_cast128_256(s4,2), \
extr32_cast128_256(s3,2), extr32_cast128_256(s2,2), \
extr32_cast128_256(s1,2), extr32_cast128_256(s0,2) ); \
casti_m256i( d,3 ) = _mm256_set_epi32( \
extr32_cast128_256(s7,3), extr32_cast128_256(s6,3), \
extr32_cast128_256(s5,3), extr32_cast128_256(s4,3), \
extr32_cast128_256(s3,3), extr32_cast128_256(s2,3), \
extr32_cast128_256(s1,3), extr32_cast128_256(s0,3) ); \
casti_m256i( d,4 ) = _mm256_set_epi32( \
mm128_extr_32(s7hi,0), mm128_extr_32(s6hi,0), \
mm128_extr_32(s5hi,0), mm128_extr_32(s4hi,0), \
mm128_extr_32(s3hi,0), mm128_extr_32(s2hi,0), \
mm128_extr_32(s1hi,0), mm128_extr_32(s0hi,0) ); \
casti_m256i( d,5 ) = _mm256_set_epi32( \
mm128_extr_32(s7hi,1), mm128_extr_32(s6hi,1), \
mm128_extr_32(s5hi,1), mm128_extr_32(s4hi,1), \
mm128_extr_32(s3hi,1), mm128_extr_32(s2hi,1), \
mm128_extr_32(s1hi,1), mm128_extr_32(s0hi,1) ); \
casti_m256i( d,6 ) = _mm256_set_epi32( \
mm128_extr_32(s7hi,2), mm128_extr_32(s6hi,2), \
mm128_extr_32(s5hi,2), mm128_extr_32(s4hi,2), \
mm128_extr_32(s3hi,2), mm128_extr_32(s2hi,2), \
mm128_extr_32(s1hi,2), mm128_extr_32(s0hi,2) ); \
casti_m256i( d,7 ) = _mm256_set_epi32( \
mm128_extr_32(s7hi,3), mm128_extr_32(s6hi,3), \
mm128_extr_32(s5hi,3), mm128_extr_32(s4hi,3), \
mm128_extr_32(s3hi,3), mm128_extr_32(s2hi,3), \
mm128_extr_32(s1hi,3), mm128_extr_32(s0hi,3) ); \
} while(0)
#define mm256_intrlv_8x32_128( d, s0, s1, s2, s3, s4, s5, s6, s7 ) \
{ \
casti_m256i( d,0 ) = _mm256_set_epi32( \
mm128_extr_32(s7,0), mm128_extr_32(s6,0), \
mm128_extr_32(s5,0), mm128_extr_32(s4,0), \
mm128_extr_32(s3,0), mm128_extr_32(s2,0), \
mm128_extr_32(s1,0), mm128_extr_32(s0,0) ); \
casti_m256i( d,1 ) = _mm256_set_epi32( \
mm128_extr_32(s7,1), mm128_extr_32(s6,1), \
mm128_extr_32(s5,1), mm128_extr_32(s4,1), \
mm128_extr_32(s3,1), mm128_extr_32(s2,1), \
mm128_extr_32(s1,1), mm128_extr_32(s0,1) ); \
casti_m256i( d,2 ) = _mm256_set_epi32( \
mm128_extr_32(s7,2), mm128_extr_32(s6,2), \
mm128_extr_32(s5,2), mm128_extr_32(s4,2), \
mm128_extr_32(s3,2), mm128_extr_32(s2,2), \
mm128_extr_32(s1,2), mm128_extr_32(s0,2) ); \
casti_m256i( d,3 ) = _mm256_set_epi32( \
mm128_extr_32(s7,3), mm128_extr_32(s6,3), \
mm128_extr_32(s5,3), mm128_extr_32(s4,3), \
mm128_extr_32(s3,3), mm128_extr_32(s2,3), \
mm128_extr_32(s1,3), mm128_extr_32(s0,3) ); \
} while(0)
/*
#define mm256_bswap_intrlv_8x32_256( d, src ) \
do { \
__m256i s0 = mm256_bswap_32( src ); \
__m128i s1 = _mm256_extracti128_si256( s0, 1 ); \
casti_m256i( d, 0 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 0 ) ); \
casti_m256i( d, 1 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 1 ) ); \
casti_m256i( d, 2 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 2 ) ); \
casti_m256i( d, 3 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 3 ) ); \
casti_m256i( d, 4 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 0 ) ); \
casti_m256i( d, 5 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 1 ) ); \
casti_m256i( d, 6 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 2 ) ); \
casti_m256i( d, 7 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 3 ) ); \
} while(0)
#define mm256_bswap_intrlv_8x32_128( d, src ) \
do { \
__m128i ss = mm128_bswap_32( src ); \
casti_m256i( d, 0 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 0 ) ); \
casti_m256i( d, 1 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 1 ) ); \
casti_m256i( d, 2 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 2 ) ); \
casti_m256i( d, 3 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 3 ) ); \
} while(0)
*/
#define mm256_dintrlv_8x32_256( d0, d1, d2, d3, d4, d5, d6, d7, s ) \
do { \
__m256i s0 = casti_m256i(s,0); \
__m256i s1 = casti_m256i(s,1); \
__m256i s2 = casti_m256i(s,2); \
__m256i s3 = casti_m256i(s,3); \
__m256i s4 = casti_m256i(s,4); \
__m256i s5 = casti_m256i(s,5); \
__m256i s6 = casti_m256i(s,6); \
__m256i s7 = casti_m256i(s,7); \
__m128i s0hi = _mm256_extracti128_si256( s0, 1 ); \
__m128i s1hi = _mm256_extracti128_si256( s1, 1 ); \
__m128i s2hi = _mm256_extracti128_si256( s2, 1 ); \
__m128i s3hi = _mm256_extracti128_si256( s3, 1 ); \
__m128i s4hi = _mm256_extracti128_si256( s4, 1 ); \
__m128i s5hi = _mm256_extracti128_si256( s5, 1 ); \
__m128i s6hi = _mm256_extracti128_si256( s6, 1 ); \
__m128i s7hi = _mm256_extracti128_si256( s7, 1 ); \
d0 = _mm256_set_epi32( \
extr32_cast128_256( s7, 0 ), extr32_cast128_256( s6, 0 ), \
extr32_cast128_256( s5, 0 ), extr32_cast128_256( s4, 0 ), \
extr32_cast128_256( s3, 0 ), extr32_cast128_256( s2, 0 ), \
extr32_cast128_256( s1, 0 ), extr32_cast128_256( s0, 0 ) );\
d1 = _mm256_set_epi32( \
extr32_cast128_256( s7, 1 ), extr32_cast128_256( s6, 1 ), \
extr32_cast128_256( s5, 1 ), extr32_cast128_256( s4, 1 ), \
extr32_cast128_256( s3, 1 ), extr32_cast128_256( s2, 1 ), \
extr32_cast128_256( s1, 1 ), extr32_cast128_256( s0, 1 ) );\
d2 = _mm256_set_epi32( \
extr32_cast128_256( s7, 2 ), extr32_cast128_256( s6, 2 ), \
extr32_cast128_256( s5, 2 ), extr32_cast128_256( s4, 2 ), \
extr32_cast128_256( s3, 2 ), extr32_cast128_256( s2, 2 ), \
extr32_cast128_256( s1, 2 ), extr32_cast128_256( s0, 2 ) );\
d3 = _mm256_set_epi32( \
extr32_cast128_256( s7, 3 ), extr32_cast128_256( s6, 3 ), \
extr32_cast128_256( s5, 3 ), extr32_cast128_256( s4, 3 ), \
extr32_cast128_256( s3, 3 ), extr32_cast128_256( s2, 3 ), \
extr32_cast128_256( s1, 3 ), extr32_cast128_256( s0, 3 ) );\
d4 = _mm256_set_epi32( \
_mm_extract_epi32( s7hi, 0 ), _mm_extract_epi32( s6hi, 0 ), \
_mm_extract_epi32( s5hi, 0 ), _mm_extract_epi32( s4hi, 0 ), \
_mm_extract_epi32( s3hi, 0 ), _mm_extract_epi32( s2hi, 0 ), \
_mm_extract_epi32( s1hi, 0 ), _mm_extract_epi32( s0hi, 0 ) ); \
d5 = _mm256_set_epi32( \
_mm_extract_epi32( s7hi, 1 ), _mm_extract_epi32( s6hi, 1 ), \
_mm_extract_epi32( s5hi, 1 ), _mm_extract_epi32( s4hi, 1 ), \
_mm_extract_epi32( s3hi, 1 ), _mm_extract_epi32( s2hi, 1 ), \
_mm_extract_epi32( s1hi, 1 ), _mm_extract_epi32( s0hi, 1 ) ); \
d6 = _mm256_set_epi32( \
_mm_extract_epi32( s7hi, 2 ), _mm_extract_epi32( s6hi, 2 ), \
_mm_extract_epi32( s5hi, 2 ), _mm_extract_epi32( s4hi, 2 ), \
_mm_extract_epi32( s3hi, 2 ), _mm_extract_epi32( s2hi, 2 ), \
_mm_extract_epi32( s1hi, 2 ), _mm_extract_epi32( s0hi, 2 ) ); \
d7 = _mm256_set_epi32( \
_mm_extract_epi32( s7hi, 3 ), _mm_extract_epi32( s6hi, 3 ), \
_mm_extract_epi32( s5hi, 3 ), _mm_extract_epi32( s4hi, 3 ), \
_mm_extract_epi32( s3hi, 3 ), _mm_extract_epi32( s2hi, 3 ), \
_mm_extract_epi32( s1hi, 3 ), _mm_extract_epi32( s0hi, 3 ) ); \
} while(0)
#define mm128_dintrlv_8x32_128( d0, d1, d2, d3, d4, d5, d6, d7, s ) \
do { \
__m128i s0 = casti_m128i(s,0); \
__m128i s1 = casti_m128i(s,1); \
__m128i s2 = casti_m128i(s,2); \
__m128i s3 = casti_m128i(s,3); \
d0 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d1 = _mm_set_epi32( \
_mm_extract_epi32( s3, 1 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 1 ), _mm_extract_epi32( s0, 0 ) ); \
d2 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d3 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d4 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d5 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d6 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
d7 = _mm_set_epi32( \
_mm_extract_epi32( s3, 0 ), _mm_extract_epi32( s2, 0 ), \
_mm_extract_epi32( s1, 0 ), _mm_extract_epi32( s0, 0 ) ); \
} while(0)
#define mm256_intrlv_4x64_256( d, s0, s1, s2, s3 ) \
do { \
__m128i s0hi = _mm256_extracti128_si256( s0, 1 ); \
__m128i s1hi = _mm256_extracti128_si256( s1, 1 ); \
__m128i s2hi = _mm256_extracti128_si256( s2, 1 ); \
__m128i s3hi = _mm256_extracti128_si256( s3, 1 ); \
casti_m256i( d,0 ) = _mm256_set_epi64x( \
extr64_cast128_256( s3, 0 ), extr64_cast128_256( s2, 0 ), \
extr64_cast128_256( s1, 0 ), extr64_cast128_256( s0, 0 ) ); \
casti_m256i( d,1 ) = _mm256_set_epi64x( \
extr64_cast128_256( s3, 1 ), extr64_cast128_256( s2, 1 ), \
extr64_cast128_256( s1, 1 ), extr64_cast128_256( s0, 1 ) ); \
casti_m256i( d,2 ) = _mm256_set_epi64x( \
_mm_extract_epi64( s3hi,0 ), _mm_extract_epi64( s2hi,0 ), \
_mm_extract_epi64( s1hi,0 ), _mm_extract_epi64( s0hi,0 ) ); \
casti_m256i( d,3 ) = _mm256_set_epi64x( \
_mm_extract_epi64( s3hi,1 ), _mm_extract_epi64( s2hi,1 ), \
_mm_extract_epi64( s1hi,1 ), _mm_extract_epi64( s0hi,1 ) ); \
} while(0)
#define mm256_intrlv_4x64_128( d, s0, s1, s2, s3 ) \
do { \
casti_m256i( d,0 ) = _mm256_set_epi64x( \
_mm_extract_epi64( s3, 0 ), _mm_extract_epi64( s2, 0 ), \
_mm_extract_epi64( s1, 0 ), _mm_extract_epi64( s0, 0 ) ); \
casti_m256i( d,1 ) = _mm256_set_epi64x( \
_mm_extract_epi64( s3, 1 ), _mm_extract_epi64( s2, 1 ), \
_mm_extract_epi64( s1, 1 ), _mm_extract_epi64( s0, 1 ) ); \
} while(0)
/*
#define mm256_bswap_intrlv_4x64_256( d, src ) \
do { \
__m256i s0 = mm256_bswap_32( src ); \
__m128i s1 = _mm256_extracti128_si256( s0, 1 ); \
casti_m256i( d,0 ) = _mm256_set1_epi64x( _mm_extract_epi64( \
_mm256_castsi256_si128( s0 ), 0 ) ); \
casti_m256i( d,1 ) = _mm256_set1_epi64x( _mm_extract_epi64( \
_mm256_castsi256_si128( s0 ), 1 ) ); \
casti_m256i( d,2 ) = _mm256_set1_epi64x( _mm_extract_epi64( s1, 0 ) ); \
casti_m256i( d,3 ) = _mm256_set1_epi64x( _mm_extract_epi64( s1, 1 ) ); \
} while(0)
#define mm256_bswap_intrlv_4x64_128( d, src ) \
do { \
__m128i ss = mm128_bswap_32( src ); \
casti_m256i( d,0 ) = _mm256_set1_epi64x( _mm_extract_epi64( ss, 0 ) ); \
casti_m256i( d,1 ) = _mm256_set1_epi64x( _mm_extract_epi64( ss, 1 ) ); \
} while(0)
*/
// 4 lanes of 256 bits using 64 bit interleaving (standard final hash size)
static inline void mm256_dintrlv_4x64_256( void *d0, void *d1, void *d2,
void *d3, const int n, const void *src )
{
__m256i s0 = *( (__m256i*) src ); // s[0][1:0]
__m256i s1 = *( (__m256i*)(src+32) ); // s[1][1:0]
__m256i s2 = *( (__m256i*)(src+64) ); // s[2][1:0]
__m256i s3 = *( (__m256i*)(src+96) ); // s[3][2:0]
__m128i s0hi = _mm256_extracti128_si256( s0, 1 ); // s[0][3:2]
__m128i s1hi = _mm256_extracti128_si256( s1, 1 ); // s[1][3:2]
__m128i s2hi = _mm256_extracti128_si256( s2, 1 ); // s[2][3:2]
__m128i s3hi = _mm256_extracti128_si256( s3, 1 ); // s[3][3:2]
casti_m256i( d0,n ) = _mm256_set_epi64x(
extr64_cast128_256( s3, 0 ), extr64_cast128_256( s2, 0 ),
extr64_cast128_256( s1, 0 ), extr64_cast128_256( s0, 0 ) );
casti_m256i( d1,n ) = _mm256_set_epi64x(
extr64_cast128_256( s3, 1 ), extr64_cast128_256( s2, 1 ),
extr64_cast128_256( s1, 1 ), extr64_cast128_256( s0, 1 ) );
casti_m256i( d2,n ) = _mm256_set_epi64x(
_mm_extract_epi64( s3hi, 0 ), _mm_extract_epi64( s2hi, 0 ),
_mm_extract_epi64( s1hi, 0 ), _mm_extract_epi64( s0hi, 0 ) );
casti_m256i( d3,n ) = _mm256_set_epi64x(
_mm_extract_epi64( s3hi, 1 ), _mm_extract_epi64( s2hi, 1 ),
_mm_extract_epi64( s1hi, 1 ), _mm_extract_epi64( s0hi, 1 ) );
}
// quarter avx2 block, 16 bytes * 4 lanes
// 4 lanes of 128 bits using 64 bit interleaving
// Used for last 16 bytes of 80 byte input, only used for testing.
static inline void mm128_dintrlv_4x64_128( void *d0, void *d1, void *d2,
void *d3, const int n, const void *src )
{
__m256i s0 = *( (__m256i*) src );
__m256i s1 = *( (__m256i*)(src+32) );
__m128i s0hi = _mm256_extracti128_si256( s0, 1 );
__m128i s1hi = _mm256_extracti128_si256( s1, 1 );
casti_m128i( d0,n ) = _mm_set_epi64x( extr64_cast128_256( s1 , 0 ),
extr64_cast128_256( s0 , 0 ) );
casti_m128i( d1,n ) = _mm_set_epi64x( extr64_cast128_256( s1 , 1 ),
extr64_cast128_256( s0 , 1 ) );
casti_m128i( d2,n ) = _mm_set_epi64x( _mm_extract_epi64( s1hi, 0 ),
_mm_extract_epi64( s0hi, 0 ) );
casti_m128i( d3,n ) = _mm_set_epi64x( _mm_extract_epi64( s1hi, 1 ),
_mm_extract_epi64( s0hi, 1 ) );
}
/*
static inline void mm256_dintrlv_2x128x256( void *d0, void *d1,
const int n, const void *s )
{
casti_m256i( d0,n ) = mm256_get_64( s, 0, 1, 4, 5 );
casti_m256i( d1,n ) = mm256_get_64( s, 2, 3, 6, 7 );
}
*/
//
#define mm256_intrlv_4x32_256( d, s0, s1, s2, s3 ) \
do { \
casti_m256i( d,0 ) = _mm256_set_epi32( \
mm128_extr_32( s3, 1 ), mm128_extr_32( s2, 1 ), \
mm128_extr_32( s1, 1 ), mm128_extr_32( s0, 1 ), \
mm128_extr_32( s3, 0 ), mm128_extr_32( s2, 0 ), \
mm128_extr_32( s1, 0 ), mm128_extr_32( s0, 0 ) ); \
casti_m256i( d,1 ) = _mm256_set_epi32( \
mm128_extr_32( s3, 3 ), mm128_extr_32( s2, 3 ), \
mm128_extr_32( s1, 3 ), mm128_extr_32( s0, 3 ), \
mm128_extr_32( s3, 2 ), mm128_extr_32( s2, 2 ), \
mm128_extr_32( s1, 2 ), mm128_extr_32( s0, 2 ) ); \
casti_m256i( d,2 ) = _mm256_set_epi32( \
mm128_extr_32( s3, 5 ), mm128_extr_32( s2, 5 ), \
mm128_extr_32( s1, 5 ), mm128_extr_32( s0, 5 ), \
mm128_extr_32( s3, 4 ), mm128_extr_32( s2, 4 ), \
mm128_extr_32( s1, 4 ), mm128_extr_32( s0, 4 ) ); \
casti_m256i( d,3 ) = _mm256_set_epi32( \
mm128_extr_32( s3, 7 ), mm128_extr_32( s2, 7 ), \
mm128_extr_32( s1, 7 ), mm128_extr_32( s0, 7 ), \
mm128_extr_32( s3, 6 ), mm128_extr_32( s2, 6 ), \
mm128_extr_32( s1, 6 ), mm128_extr_32( s0, 6 ) ); \
} while(0)
// 256 bit versions of commmon 128 bit functions.
static inline void mm256_intrlv_4x32( void *d, const void *s0,
const void *s1, const void *s2, const void *s3, int bit_len )
{
mm256_intrlv_4x32_256( d ,casti_m256i(s0,0), casti_m256i(s1,0),
casti_m256i(s2,0), casti_m256i(s3,0) );
if ( bit_len <= 256 ) return;
mm256_intrlv_4x32_256( d+128 ,casti_m256i(s0,1), casti_m256i(s1,1),
casti_m256i(s2,1), casti_m256i(s3,1) );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm128_intrlv_4x32_128( d+256, casti_m128i(s0,4), casti_m128i(s1,4),
casti_m128i(s2,4), casti_m128i(s3,4) );
return;
}
mm256_intrlv_4x32_256( d+256 ,casti_m256i(s0,2), casti_m256i(s1,2),
casti_m256i(s2,2), casti_m256i(s3,2) );
mm256_intrlv_4x32_256( d+384 ,casti_m256i(s0,3), casti_m256i(s1,3),
casti_m256i(s2,3), casti_m256i(s3,3) );
}
static inline void mm256_dintrlv_4x32_256( void *d0, void *d1, void *d2,
void *d3, const void *src )
{
__m256i s0 = *(__m256i*) src;
__m256i s1 = *(__m256i*)(src+32);
__m256i s2 = *(__m256i*)(src+64);
__m256i s3 = *(__m256i*)(src+96);
*(__m256i*)d0 = _mm256_set_epi32(
_mm256_extract_epi32( s3,4 ), _mm256_extract_epi32( s3,0 ),
_mm256_extract_epi32( s2,4 ), _mm256_extract_epi32( s2,0 ),
_mm256_extract_epi32( s1,4 ), _mm256_extract_epi32( s1,0 ),
_mm256_extract_epi32( s0,4 ), _mm256_extract_epi32( s0,0 ) );
*(__m256i*)d1 = _mm256_set_epi32(
_mm256_extract_epi32( s3,5 ), _mm256_extract_epi32( s3,1 ),
_mm256_extract_epi32( s2,5 ), _mm256_extract_epi32( s2,1 ),
_mm256_extract_epi32( s1,5 ), _mm256_extract_epi32( s1,1 ),
_mm256_extract_epi32( s0,5 ), _mm256_extract_epi32( s0,1 ) );
*(__m256i*)d2 = _mm256_set_epi32(
_mm256_extract_epi32( s3,6 ), _mm256_extract_epi32( s3,2 ),
_mm256_extract_epi32( s2,6 ), _mm256_extract_epi32( s2,2 ),
_mm256_extract_epi32( s1,6 ), _mm256_extract_epi32( s1,2 ),
_mm256_extract_epi32( s0,6 ), _mm256_extract_epi32( s0,2 ) );
*(__m256i*)d3 = _mm256_set_epi32(
_mm256_extract_epi32( s3,7 ), _mm256_extract_epi32( s3,3 ),
_mm256_extract_epi32( s2,7 ), _mm256_extract_epi32( s2,3 ),
_mm256_extract_epi32( s1,7 ), _mm256_extract_epi32( s1,3 ),
_mm256_extract_epi32( s0,7 ), _mm256_extract_epi32( s0,3 ) );
}
static inline void mm256_dintrlv_4x32( void *d0, void *d1, void *d2,
void *d3, const void *s, int bit_len )
{
mm256_dintrlv_4x32_256( d0 , d1 , d2 , d3 , s );
if ( bit_len <= 256 ) return;
mm256_dintrlv_4x32_256( d0+ 32, d1+ 32, d2+ 32, d3+ 32, s+128 );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm128_dintrlv_4x32_128( d0+ 64, d1+ 64, d2+ 64, d3+ 64, s+256 );
return;
}
mm256_dintrlv_4x32_256( d0+ 64, d1+ 64, d2+ 64, d3+ 64, s+256 );
mm256_dintrlv_4x32_256( d0+ 96, d1+ 96, d2+ 96, d3+ 96, s+384 );
}
static inline void mm256_extr_lane_4x32( void *d, const void *s,
const int lane, const int bit_len )
{
casti_m256i( d, 0 ) = mm256_get_32( s, lane , lane+ 4, lane+ 8, lane+12,
lane+16, lane+20, lane+24, lane+28 );
if ( bit_len <= 256 ) return;
casti_m256i( d, 1 ) = mm256_get_32( s, lane+32, lane+36, lane+40, lane+44,
lane+48, lane+52, lane+56, lane+60 );
}
// Interleave 8 source buffers containing 32 bit data into the destination
// vector
static inline void mm256_intrlv_8x32( void *d, const void *s0,
const void *s1, const void *s2, const void *s3, const void *s4,
const void *s5, const void *s6, const void *s7, int bit_len )
{
mm256_intrlv_8x32_256( d , casti_m256i( s0,0 ), casti_m256i( s1,0 ),
casti_m256i( s2,0 ), casti_m256i( s3,0 ), casti_m256i( s4,0 ),
casti_m256i( s5,0 ), casti_m256i( s6,0 ), casti_m256i( s7,0 ) );
if ( bit_len <= 256 ) return;
mm256_intrlv_8x32_256( d+256, casti_m256i( s0,1 ), casti_m256i( s1,1 ),
casti_m256i( s2,1 ), casti_m256i( s3,1 ), casti_m256i( s4,1 ),
casti_m256i( s5,1 ), casti_m256i( s6,1 ), casti_m256i( s7,1 ) );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm256_intrlv_8x32_128( d+512, casti_m128i( s0,4 ), casti_m128i( s1,4 ),
casti_m128i( s2,4 ), casti_m128i( s3,4 ), casti_m128i( s4,4 ),
casti_m128i( s5,4 ), casti_m128i( s6,4 ), casti_m128i( s7,4 ) );
return;
}
mm256_intrlv_8x32_256( d+512, casti_m256i( s0,2 ), casti_m256i( s1,2 ),
casti_m256i( s2,2 ), casti_m256i( s3,2 ), casti_m256i( s4,2 ),
casti_m256i( s5,2 ), casti_m256i( s6,2 ), casti_m256i( s7,2 ) );
mm256_intrlv_8x32_256( d+768, casti_m256i( s0,3 ), casti_m256i( s1,3 ),
casti_m256i( s2,3 ), casti_m256i( s3,3 ), casti_m256i( s4,3 ),
casti_m256i( s5,3 ), casti_m256i( s6,3 ), casti_m256i( s7,3 ) );
// bit_len == 1024
}
// A couple of mining specifi functions.
/*
// Interleave 80 bytes of 32 bit data for 8 lanes.
static inline void mm256_bswap_intrlv80_8x32( void *d, const void *s )
{
mm256_bswap_intrlv_8x32_256( d , casti_m256i( s, 0 ) );
mm256_bswap_intrlv_8x32_256( d+256, casti_m256i( s, 1 ) );
mm256_bswap_intrlv_8x32_128( d+512, casti_m128i( s, 4 ) );
}
*/
// Deinterleave 8 buffers of 32 bit data from the source buffer.
// Sub-function can be called directly for 32 byte final hash.
static inline void mm256_dintrlv_8x32( void *d0, void *d1, void *d2,
void *d3, void *d4, void *d5, void *d6, void *d7,
const void *s, int bit_len )
{
mm256_dintrlv_8x32_256( casti_m256i(d0,0), casti_m256i(d1,0),
casti_m256i(d2,0), casti_m256i(d3,0), casti_m256i(d4,0),
casti_m256i(d5,0), casti_m256i(d6,0), casti_m256i(d7,0), s );
if ( bit_len <= 256 ) return;
mm256_dintrlv_8x32_256( casti_m256i(d0,1), casti_m256i(d1,1),
casti_m256i(d2,1), casti_m256i(d3,1), casti_m256i(d4,1),
casti_m256i(d5,1), casti_m256i(d6,1), casti_m256i(d7,1), s+256 );
if ( bit_len <= 512 ) return;
// short block, final 16 bytes of input data
if ( bit_len <= 640 )
{
mm128_dintrlv_8x32_128( casti_m128i(d0,2), casti_m128i(d1,2),
casti_m128i(d2,2), casti_m128i(d3,2), casti_m128i(d4,2),
casti_m128i(d5,2), casti_m128i(d6,2), casti_m128i(d7,2), s+512 );
return;
}
// bitlen == 1024
mm256_dintrlv_8x32_256( casti_m256i(d0,2), casti_m256i(d1,2),
casti_m256i(d2,2), casti_m256i(d3,2), casti_m256i(d4,2),
casti_m256i(d5,2), casti_m256i(d6,2), casti_m256i(d7,2), s+512 );
mm256_dintrlv_8x32_256( casti_m256i(d0,3), casti_m256i(d1,3),
casti_m256i(d2,3), casti_m256i(d3,3), casti_m256i(d4,3),
casti_m256i(d5,3), casti_m256i(d6,3), casti_m256i(d7,3), s+768 );
}
static inline void mm256_extr_lane_8x32( void *d, const void *s,
const int lane, const int bit_len )
{
casti_m256i( d,0 ) = mm256_get_32(s, lane , lane+ 8, lane+ 16, lane+ 24,
lane+32, lane+ 40, lane+ 48, lane+ 56 );
if ( bit_len <= 256 ) return;
casti_m256i( d,1 ) = mm256_get_32(s, lane+64, lane+ 72, lane+ 80, lane+ 88,
lane+96, lane+104, lane+112, lane+120 );
// bit_len == 512
}
// Interleave 4 source buffers containing 64 bit data into the destination
// buffer. Only bit_len 256, 512, 640 & 1024 are supported.
static inline void mm256_intrlv_4x64( void *d, const void *s0,
const void *s1, const void *s2, const void *s3, int bit_len )
{
mm256_intrlv_4x64_256( d , casti_m256i(s0,0), casti_m256i(s1,0),
casti_m256i(s2,0), casti_m256i(s3,0) );
if ( bit_len <= 256 ) return;
mm256_intrlv_4x64_256( d+128, casti_m256i(s0,1), casti_m256i(s1,1),
casti_m256i(s2,1), casti_m256i(s3,1) );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm256_intrlv_4x64_128( d+256, casti_m128i(s0,4), casti_m128i(s1,4),
casti_m128i(s2,4), casti_m128i(s3,4) );
return;
}
// bit_len == 1024
mm256_intrlv_4x64_256( d+256, casti_m256i(s0,2), casti_m256i(s1,2),
casti_m256i(s2,2), casti_m256i(s3,2) );
mm256_intrlv_4x64_256( d+384, casti_m256i(s0,3), casti_m256i(s1,3),
casti_m256i(s2,3), casti_m256i(s3,3) );
}
/*
// Interleave 80 bytes of 32 bit data for 8 lanes.
static inline void mm256_bswap_intrlv80_4x64( void *d, const void *s )
{
mm256_bswap_intrlv_4x64_256( d , casti_m256i( s, 0 ) );
mm256_bswap_intrlv_4x64_256( d+128, casti_m256i( s, 1 ) );
mm256_bswap_intrlv_4x64_128( d+256, casti_m128i( s, 4 ) );
}
// Blend 32 byte lanes of hash from 2 sources according to control mask.
// macro due to 256 bit value arg.
#define mm256_blend_hash_4x64( dst, a, b, mask ) \
do { \
dst[0] = _mm256_blendv_epi8( a[0], b[0], mask ); \
dst[1] = _mm256_blendv_epi8( a[1], b[1], mask ); \
dst[2] = _mm256_blendv_epi8( a[2], b[2], mask ); \
dst[3] = _mm256_blendv_epi8( a[3], b[3], mask ); \
dst[4] = _mm256_blendv_epi8( a[4], b[4], mask ); \
dst[5] = _mm256_blendv_epi8( a[5], b[5], mask ); \
dst[6] = _mm256_blendv_epi8( a[6], b[6], mask ); \
dst[7] = _mm256_blendv_epi8( a[7], b[7], mask ); \
} while(0)
*/
// Deinterleave 4 buffers of 64 bit data from the source buffer.
// bit_len must be 256, 512, 640 or 1024 bits.
// Requires overrun padding for 640 bit len.
static inline void mm256_dintrlv_4x64( void *d0, void *d1, void *d2,
void *d3, const void *s, int bit_len )
{
mm256_dintrlv_4x64_256( d0, d1, d2, d3, 0, s );
if ( bit_len <= 256 ) return;
mm256_dintrlv_4x64_256( d0, d1, d2, d3, 1, s+128 );
if ( bit_len <= 512 ) return;
// short block, final 16 bytes of input data
if ( bit_len <= 640 )
{
mm128_dintrlv_4x64_128( d0, d1, d2, d3, 4, s+256 );
return;
}
// bit_len == 1024
mm256_dintrlv_4x64_256( d0, d1, d2, d3, 2, s+256 );
mm256_dintrlv_4x64_256( d0, d1, d2, d3, 3, s+384 );
}
// extract and deinterleave specified lane.
#define mm256_extr_lane_4x64_256 \
casti_m256i( d, 0 ) = mm256_get_64( s, lane, lane+4, lane+8, lane+12 )
static inline void mm256_extr_lane_4x64( void *d, const void *s,
const int lane, const int bit_len )
{
casti_m256i( d, 0 ) = mm256_get_64( s, lane, lane+4, lane+8, lane+12 );
if ( bit_len <= 256 ) return;
casti_m256i( d, 1 ) = mm256_get_64( s, lane+16, lane+20, lane+24, lane+28 );
return;
}
// Convert from 4x32 SSE2 interleaving to 4x64 AVX2.
// Can't do it in place
static inline void mm256_rintrlv_4x32_4x64( void *dst, void *src,
int bit_len )
{
__m256i* d = (__m256i*)dst;
uint32_t *s = (uint32_t*)src;
d[0] = _mm256_set_epi32( s[ 7],s[ 3],s[ 6],s[ 2],s[ 5],s[ 1],s[ 4],s[ 0] );
d[1] = _mm256_set_epi32( s[15],s[11],s[14],s[10],s[13],s[ 9],s[12],s[ 8] );
d[2] = _mm256_set_epi32( s[23],s[19],s[22],s[18],s[21],s[17],s[20],s[16] );
d[3] = _mm256_set_epi32( s[31],s[27],s[30],s[26],s[29],s[25],s[28],s[24] );
if ( bit_len <= 256 ) return;
d[4] = _mm256_set_epi32( s[39],s[35],s[38],s[34],s[37],s[33],s[36],s[32] );
d[5] = _mm256_set_epi32( s[47],s[43],s[46],s[42],s[45],s[41],s[44],s[40] );
d[6] = _mm256_set_epi32( s[55],s[51],s[54],s[50],s[53],s[49],s[52],s[48] );
d[7] = _mm256_set_epi32( s[63],s[59],s[62],s[58],s[61],s[57],s[60],s[56] );
if ( bit_len <= 512 ) return;
d[8] = _mm256_set_epi32( s[71],s[67],s[70],s[66],s[69],s[65],s[68],s[64] );
d[9] = _mm256_set_epi32( s[79],s[75],s[78],s[74],s[77],s[73],s[76],s[72] );
if ( bit_len <= 640 ) return;
d[10] = _mm256_set_epi32(s[87],s[83],s[86],s[82],s[85],s[81],s[84],s[80]);
d[11] = _mm256_set_epi32(s[95],s[91],s[94],s[90],s[93],s[89],s[92],s[88]);
d[12] = _mm256_set_epi32(s[103],s[99],s[102],s[98],s[101],s[97],s[100],s[96]);
d[13] = _mm256_set_epi32(s[111],s[107],s[110],s[106],s[109],s[105],s[108],s[104]);
d[14] = _mm256_set_epi32(s[119],s[115],s[118],s[114],s[117],s[113],s[116],s[112]);
d[15] = _mm256_set_epi32(s[127],s[123],s[126],s[122],s[125],s[121],s[124],s[120]);
// bit_len == 1024
}
// Convert 4x64 byte (256 bit) vectors to 4x32 (128 bit) vectors for AVX
// bit_len must be multiple of 64
static inline void mm256_rintrlv_4x64_4x32( void *dst, void *src,
int bit_len )
{
__m256i *d = (__m256i*)dst;
uint32_t *s = (uint32_t*)src;
d[0] = _mm256_set_epi32( s[ 7],s[ 5],s[ 3],s[ 1],s[ 6],s[ 4],s[ 2],s[ 0] );
d[1] = _mm256_set_epi32( s[15],s[13],s[11],s[ 9],s[14],s[12],s[10],s[ 8] );
d[2] = _mm256_set_epi32( s[23],s[21],s[19],s[17],s[22],s[20],s[18],s[16] );
d[3] = _mm256_set_epi32( s[31],s[29],s[27],s[25],s[30],s[28],s[26],s[24] );
if ( bit_len <= 256 ) return;
d[4] = _mm256_set_epi32( s[39],s[37],s[35],s[33],s[38],s[36],s[34],s[32] );
d[5] = _mm256_set_epi32( s[47],s[45],s[43],s[41],s[46],s[44],s[42],s[40] );
d[6] = _mm256_set_epi32( s[55],s[53],s[51],s[49],s[54],s[52],s[50],s[48] );
d[7] = _mm256_set_epi32( s[63],s[61],s[59],s[57],s[62],s[60],s[58],s[56] );
if ( bit_len <= 512 ) return;
d[8] = _mm256_set_epi32( s[71],s[69],s[67],s[65],s[70],s[68],s[66],s[64] );
d[9] = _mm256_set_epi32( s[79],s[77],s[75],s[73],s[78],s[76],s[74],s[72] );
if ( bit_len <= 640 ) return;
d[10] = _mm256_set_epi32( s[87],s[85],s[83],s[81],s[86],s[84],s[82],s[80] );
d[11] = _mm256_set_epi32( s[95],s[93],s[91],s[89],s[94],s[92],s[90],s[88] );
d[12] = _mm256_set_epi32( s[103],s[101],s[99],s[97],s[102],s[100],s[98],s[96] );
d[13] = _mm256_set_epi32( s[111],s[109],s[107],s[105],s[110],s[108],s[106],s[104] );
d[14] = _mm256_set_epi32( s[119],s[117],s[115],s[113],s[118],s[116],s[114],s[112] );
d[15] = _mm256_set_epi32( s[127],s[125],s[123],s[121],s[126],s[124],s[122],s[120] );
// bit_len == 1024
}
static inline void mm256_rintrlv_4x64_2x128( void *dst0, void *dst1,
const void *src, int bit_len )
{
__m256i* d0 = (__m256i*)dst0;
__m256i* d1 = (__m256i*)dst1;
uint64_t *s = (uint64_t*)src;
d0[0] = _mm256_set_epi64x( s[ 5], s[ 1], s[ 4], s[ 0] );
d1[0] = _mm256_set_epi64x( s[ 7], s[ 3], s[ 6], s[ 2] );
d0[1] = _mm256_set_epi64x( s[13], s[ 9], s[12], s[ 8] );
d1[1] = _mm256_set_epi64x( s[15], s[11], s[14], s[10] );
if ( bit_len <= 256 ) return;
d0[2] = _mm256_set_epi64x( s[21], s[17], s[20], s[16] );
d1[2] = _mm256_set_epi64x( s[23], s[19], s[22], s[18] );
d0[3] = _mm256_set_epi64x( s[29], s[25], s[28], s[24] );
d1[3] = _mm256_set_epi64x( s[31], s[27], s[30], s[26] );
if ( bit_len <= 512 ) return;
d0[4] = _mm256_set_epi64x( s[37], s[33], s[36], s[32] );
d1[4] = _mm256_set_epi64x( s[39], s[35], s[38], s[34] );
d0[5] = _mm256_set_epi64x( s[45], s[41], s[44], s[40] );
d1[5] = _mm256_set_epi64x( s[47], s[43], s[46], s[42] );
d0[6] = _mm256_set_epi64x( s[53], s[49], s[52], s[48] );
d1[6] = _mm256_set_epi64x( s[55], s[51], s[54], s[50] );
d0[7] = _mm256_set_epi64x( s[61], s[57], s[60], s[56] );
d1[7] = _mm256_set_epi64x( s[63], s[59], s[62], s[58] );
}
static inline void mm256_rintrlv_2x128_4x64( void *dst, const void *src0,
const void *src1, int bit_len )
{
__m256i* d = (__m256i*)dst;
uint64_t *s0 = (uint64_t*)src0;
uint64_t *s1 = (uint64_t*)src1;
d[ 0] = _mm256_set_epi64x( s1[2], s1[0], s0[2], s0[0] );
d[ 1] = _mm256_set_epi64x( s1[3], s1[1], s0[3], s0[1] );
d[ 2] = _mm256_set_epi64x( s1[6], s1[4], s0[6], s0[4] );
d[ 3] = _mm256_set_epi64x( s1[7], s1[5], s0[7], s0[5] );
if ( bit_len <= 256 ) return;
d[ 4] = _mm256_set_epi64x( s1[10], s1[ 8], s0[10], s0[ 8] );
d[ 5] = _mm256_set_epi64x( s1[11], s1[ 9], s0[11], s0[ 9] );
d[ 6] = _mm256_set_epi64x( s1[14], s1[12], s0[14], s0[12] );
d[ 7] = _mm256_set_epi64x( s1[15], s1[13], s0[15], s0[13] );
if ( bit_len <= 512 ) return;
d[ 8] = _mm256_set_epi64x( s1[18], s1[16], s0[18], s0[16] );
d[ 9] = _mm256_set_epi64x( s1[19], s1[17], s0[19], s0[17] );
d[10] = _mm256_set_epi64x( s1[22], s1[20], s0[22], s0[20] );
d[11] = _mm256_set_epi64x( s1[23], s1[21], s0[23], s0[21] );
d[12] = _mm256_set_epi64x( s1[26], s1[24], s0[26], s0[24] );
d[13] = _mm256_set_epi64x( s1[27], s1[25], s0[27], s0[25] );
d[14] = _mm256_set_epi64x( s1[30], s1[28], s0[30], s0[28] );
d[15] = _mm256_set_epi64x( s1[31], s1[29], s0[31], s0[29] );
}
static inline void mm256_intrlv_2x128( const void *d, const void *s0,
void *s1, const int bit_len )
{
__m128i s1hi = _mm256_extracti128_si256( casti_m256i( s1,0 ), 1 );
__m128i s0hi = _mm256_extracti128_si256( casti_m256i( s0,0 ), 1 );
casti_m256i( d,0 ) = mm256_concat_128(
_mm256_castsi256_si128( casti_m256i( s1,0 ) ),
_mm256_castsi256_si128( casti_m256i( s0,0 ) ) );
casti_m256i( d,1 ) = mm256_concat_128( s1hi, s0hi );
if ( bit_len <= 256 ) return;
s0hi = _mm256_extracti128_si256( casti_m256i( s0,1 ), 1 );
s1hi = _mm256_extracti128_si256( casti_m256i( s1,1 ), 1 );
casti_m256i( d,2 ) = mm256_concat_128(
_mm256_castsi256_si128( casti_m256i( s1,1 ) ),
_mm256_castsi256_si128( casti_m256i( s0,1 ) ) );
casti_m256i( d,3 ) = mm256_concat_128( s1hi, s0hi );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
casti_m256i( d,4 ) = mm256_concat_128(
_mm256_castsi256_si128( casti_m256i( s1,2 ) ),
_mm256_castsi256_si128( casti_m256i( s0,2 ) ) );
return;
}
s0hi = _mm256_extracti128_si256( casti_m256i( s0,2 ), 1 );
s1hi = _mm256_extracti128_si256( casti_m256i( s1,2 ), 1 );
casti_m256i( d,4 ) = mm256_concat_128(
_mm256_castsi256_si128( casti_m256i( s1,2 ) ),
_mm256_castsi256_si128( casti_m256i( s0,2 ) ) );
casti_m256i( d,5 ) = mm256_concat_128( s1hi, s0hi );
s0hi = _mm256_extracti128_si256( casti_m256i( s0,3 ), 1 );
s1hi = _mm256_extracti128_si256( casti_m256i( s1,3 ), 1 );
casti_m256i( d,6 ) = mm256_concat_128(
_mm256_castsi256_si128( casti_m256i( s1,3 ) ),
_mm256_castsi256_si128( casti_m256i( s0,3 ) ) );
casti_m256i( d,7 ) = mm256_concat_128( s1hi, s0hi );
}
// 512 is the bit len used by most, eliminate the conditionals
static inline void mm256_dintrlv_2x128_512( void *dst0, void *dst1,
const void *s )
{
__m256i *d0 = (__m256i*)dst0;
__m256i *d1 = (__m256i*)dst1;
__m256i s0 = casti_m256i( s, 0 );
__m256i s1 = casti_m256i( s, 1 );
d0[0] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[0] = _mm256_permute2x128_si256( s0, s1, 0x31 );
s0 = casti_m256i( s, 2 );
s1 = casti_m256i( s, 3 );
d0[1] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[1] = _mm256_permute2x128_si256( s0, s1, 0x31 );
}
// Phase out usage for all 512 bit data lengths
static inline void mm256_dintrlv_2x128( void *dst0, void *dst1, const void *s,
int bit_len )
{
__m256i *d0 = (__m256i*)dst0;
__m256i *d1 = (__m256i*)dst1;
__m256i s0 = casti_m256i( s, 0 );
__m256i s1 = casti_m256i( s, 1 );
d0[0] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[0] = _mm256_permute2x128_si256( s0, s1, 0x31 );
if ( bit_len <= 256 ) return;
s0 = casti_m256i( s, 2 );
s1 = casti_m256i( s, 3 );
d0[1] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[1] = _mm256_permute2x128_si256( s0, s1, 0x31 );
if ( bit_len <= 512 ) return;
s0 = casti_m256i( s, 4 );
s1 = casti_m256i( s, 5 );
d0[2] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[2] = _mm256_permute2x128_si256( s0, s1, 0x31 );
s0 = casti_m256i( s, 6 );
s1 = casti_m256i( s, 7 );
d0[3] = _mm256_permute2x128_si256( s0, s1, 0x20 );
d1[3] = _mm256_permute2x128_si256( s0, s1, 0x31 );
}
#undef extr64_cast128_256
#undef extr32_cast128_256
#endif // AVX
#endif // INTRLV_AVX_H__

104
simd-utils/intrlv-avx2.h
View File

@@ -1,104 +0,0 @@
#if !defined(INTRLV_AVX2_H__)
#define INTRLV_AVX2_H__ 1
#if defined(__AVX2__)
///////////////////////////////////////////////////////////
//
// AVX2 256 Bit Vectors
//
// A few functions that need AVX2 for 256 bit.
// Blend 2 vectors alternating hi & lo: { hi[n], lo[n-1], ... hi[1], lo[0] }
#define mm256_intrlv_blend_128( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x0f )
#define mm256_intrlv_blend_64( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x33 )
#define mm256_intrlv_blend_32( hi, lo ) \
_mm256_blend_epi32( hi, lo, 0x55 )
#define mm256_bswap_intrlv_8x32_256( d, src ) \
do { \
__m256i s0 = mm256_bswap_32( src ); \
__m128i s1 = _mm256_extracti128_si256( s0, 1 ); \
casti_m256i( d, 0 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 0 ) ); \
casti_m256i( d, 1 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 1 ) ); \
casti_m256i( d, 2 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 2 ) ); \
casti_m256i( d, 3 ) = _mm256_set1_epi32( _mm_extract_epi32( \
_mm256_castsi256_si128( s0 ), 3 ) ); \
casti_m256i( d, 4 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 0 ) ); \
casti_m256i( d, 5 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 1 ) ); \
casti_m256i( d, 6 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 2 ) ); \
casti_m256i( d, 7 ) = _mm256_set1_epi32( _mm_extract_epi32( s1, 3 ) ); \
} while(0)
#define mm256_bswap_intrlv_8x32_128( d, src ) \
do { \
__m128i ss = mm128_bswap_32( src ); \
casti_m256i( d, 0 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 0 ) ); \
casti_m256i( d, 1 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 1 ) ); \
casti_m256i( d, 2 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 2 ) ); \
casti_m256i( d, 3 ) = _mm256_set1_epi32( _mm_extract_epi32( ss, 3 ) ); \
} while(0)
#define mm256_bswap_intrlv_4x64_256( d, src ) \
do { \
__m256i s0 = mm256_bswap_32( src ); \
__m128i s1 = _mm256_extracti128_si256( s0, 1 ); \
casti_m256i( d,0 ) = _mm256_set1_epi64x( _mm_extract_epi64( \
_mm256_castsi256_si128( s0 ), 0 ) ); \
casti_m256i( d,1 ) = _mm256_set1_epi64x( _mm_extract_epi64( \
_mm256_castsi256_si128( s0 ), 1 ) ); \
casti_m256i( d,2 ) = _mm256_set1_epi64x( _mm_extract_epi64( s1, 0 ) ); \
casti_m256i( d,3 ) = _mm256_set1_epi64x( _mm_extract_epi64( s1, 1 ) ); \
} while(0)
#define mm256_bswap_intrlv_4x64_128( d, src ) \
do { \
__m128i ss = mm128_bswap_32( src ); \
casti_m256i( d,0 ) = _mm256_set1_epi64x( _mm_extract_epi64( ss, 0 ) ); \
casti_m256i( d,1 ) = _mm256_set1_epi64x( _mm_extract_epi64( ss, 1 ) ); \
} while(0)
// A couple of mining specifi functions.
// Interleave 80 bytes of 32 bit data for 8 lanes.
static inline void mm256_bswap_intrlv80_8x32( void *d, const void *s )
{
mm256_bswap_intrlv_8x32_256( d , casti_m256i( s, 0 ) );
mm256_bswap_intrlv_8x32_256( d+256, casti_m256i( s, 1 ) );
mm256_bswap_intrlv_8x32_128( d+512, casti_m128i( s, 4 ) );
}
// Interleave 80 bytes of 32 bit data for 8 lanes.
static inline void mm256_bswap_intrlv80_4x64( void *d, const void *s )
{
mm256_bswap_intrlv_4x64_256( d , casti_m256i( s, 0 ) );
mm256_bswap_intrlv_4x64_256( d+128, casti_m256i( s, 1 ) );
mm256_bswap_intrlv_4x64_128( d+256, casti_m128i( s, 4 ) );
}
// Blend 32 byte lanes of hash from 2 sources according to control mask.
// macro due to 256 bit value arg.
#define mm256_blend_hash_4x64( dst, a, b, mask ) \
do { \
dst[0] = _mm256_blendv_epi8( a[0], b[0], mask ); \
dst[1] = _mm256_blendv_epi8( a[1], b[1], mask ); \
dst[2] = _mm256_blendv_epi8( a[2], b[2], mask ); \
dst[3] = _mm256_blendv_epi8( a[3], b[3], mask ); \
dst[4] = _mm256_blendv_epi8( a[4], b[4], mask ); \
dst[5] = _mm256_blendv_epi8( a[5], b[5], mask ); \
dst[6] = _mm256_blendv_epi8( a[6], b[6], mask ); \
dst[7] = _mm256_blendv_epi8( a[7], b[7], mask ); \
} while(0)
#endif // AVX2
#endif // INTRLV_AVX2_H__

679
simd-utils/intrlv-avx512.h
View File

@@ -1,679 +0,0 @@
#if !defined(INTRLV_AVX512_H__)
#define INTRLV_AVX512_H__ 1
#if defined(__AVX512VL__) && defined(__AVX512DQ__) && defined(__AVX512BW__)
// SSE2 functions used in AVX512 interleaving
// AVX512 block is 64 * 64 bytes
// quarter avx512 block, 16 bytes * 16 lanes
static inline void mm128_dintrlv_16x32x128( void *d00, void *d01,
void *d02, void *d03, void *d04, void *d05, void *d06, void *d07,
void *d08, void *d09, void *d10, void *d11, void *d12, void *d13,
void *d14, void *d15, const int n, const void *s )
{
cast_m128i( d00 ) = mm128_get_32( s, 0, 16, 32, 48 );
cast_m128i( d01 ) = mm128_get_32( s, 1, 17, 33, 49 );
cast_m128i( d02 ) = mm128_get_32( s, 2, 18, 34, 50 );
cast_m128i( d03 ) = mm128_get_32( s, 3, 19, 35, 51 );
cast_m128i( d04 ) = mm128_get_32( s, 4, 20, 36, 52 );
cast_m128i( d05 ) = mm128_get_32( s, 5, 21, 37, 53 );
cast_m128i( d06 ) = mm128_get_32( s, 6, 22, 38, 54 );
cast_m128i( d07 ) = mm128_get_32( s, 7, 23, 39, 55 );
cast_m128i( d08 ) = mm128_get_32( s, 8, 24, 40, 56 );
cast_m128i( d09 ) = mm128_get_32( s, 9, 25, 41, 57 );
cast_m128i( d10 ) = mm128_get_32( s, 10, 26, 42, 58 );
cast_m128i( d11 ) = mm128_get_32( s, 11, 27, 43, 59 );
cast_m128i( d12 ) = mm128_get_32( s, 12, 28, 44, 60 );
cast_m128i( d13 ) = mm128_get_32( s, 13, 29, 45, 61 );
cast_m128i( d14 ) = mm128_get_32( s, 14, 30, 46, 62 );
cast_m128i( d15 ) = mm128_get_32( s, 15, 31, 47, 63 );
}
// quarter avx512 block, 32 bytes * 8 lanes
// 8 lanes of 128 bits using 64 bit interleaving
// Used for last 16 bytes of 80 byte input, only used for testing.
static inline void mm128_dintrlv_8x64x128( void *d0, void *d1, void *d2,
void *d3, void *d4, void *d5, void *d6, void *d7,
const int n, const void *s )
{
casti_m128i( d0,n ) = mm128_get_64( s, 0, 8 );
casti_m128i( d1,n ) = mm128_get_64( s, 1, 9 );
casti_m128i( d2,n ) = mm128_get_64( s, 2, 10 );
casti_m128i( d3,n ) = mm128_get_64( s, 3, 11 );
casti_m128i( d4,n ) = mm128_get_64( s, 4, 12 );
casti_m128i( d5,n ) = mm128_get_64( s, 5, 13 );
casti_m128i( d6,n ) = mm128_get_64( s, 6, 14 );
casti_m128i( d7,n ) = mm128_get_64( s, 7, 15 );
}
static inline void mm128_dintrlv_4x128x128( void *d0, void *d1, void *d2,
void *d3, const int n, const void *s )
{
casti_m128i( d0,n ) = mm128_get_64( s, 0, 1 );
casti_m128i( d1,n ) = mm128_get_64( s, 2, 3 );
casti_m128i( d2,n ) = mm128_get_64( s, 4, 5 );
casti_m128i( d3,n ) = mm128_get_64( s, 5, 7 );
}
// AVX2 functions Used in AVX512 interleaving
static inline void mm256_dintrlv_16x32x256( void *d00, void *d01,
void *d02, void *d03, void *d04, void *d05,
void *d06, void *d07, void *d08, void *d09,
void *d10, void *d11, void *d12, void *d13,
void *d14, void *d15, const int n, const void *s )
{
casti_m256i( d00,n ) = mm256_get_32( s, 0, 16, 32, 48, 64, 80, 96,112 );
casti_m256i( d01,n ) = mm256_get_32( s, 1, 17, 33, 49, 65, 81, 97,113 );
casti_m256i( d02,n ) = mm256_get_32( s, 2, 18, 34, 50, 66, 82, 98,114 );
casti_m256i( d03,n ) = mm256_get_32( s, 3, 19, 35, 51, 67, 83, 99,115 );
casti_m256i( d04,n ) = mm256_get_32( s, 4, 20, 36, 52, 68, 84,100,116 );
casti_m256i( d05,n ) = mm256_get_32( s, 5, 21, 37, 53, 69, 85,101,117 );
casti_m256i( d06,n ) = mm256_get_32( s, 6, 22, 38, 54, 70, 86,102,118 );
casti_m256i( d07,n ) = mm256_get_32( s, 7, 23, 39, 55, 71, 87,103,119 );
casti_m256i( d08,n ) = mm256_get_32( s, 8, 24, 40, 56, 72, 88,104,120 );
casti_m256i( d09,n ) = mm256_get_32( s, 9, 25, 41, 57, 73, 89,105,121 );
casti_m256i( d10,n ) = mm256_get_32( s, 10, 26, 42, 58, 74, 90,106,122 );
casti_m256i( d11,n ) = mm256_get_32( s, 11, 27, 43, 59, 75, 91,107,123 );
casti_m256i( d12,n ) = mm256_get_32( s, 12, 28, 44, 60, 76, 92,108,124 );
casti_m256i( d13,n ) = mm256_get_32( s, 13, 29, 45, 61, 77, 93,109,125 );
casti_m256i( d14,n ) = mm256_get_32( s, 14, 30, 46, 62, 78, 94,110,126 );
casti_m256i( d15,n ) = mm256_get_32( s, 15, 31, 47, 63, 79, 95,111,127 );
}
// 8 lanes of 256 bits using 64 bit interleaving (standard final hash size)
static inline void mm256_dintrlv_8x64x256( void *d0, void *d1, void *d2,
void *d3, void *d4, void *d5, void *d6, void *d7,
const int n, const void *s )
{
casti_m256i( d0,n ) = mm256_get_64( s, 0, 8, 16, 24 );
casti_m256i( d1,n ) = mm256_get_64( s, 1, 9, 17, 25 );
casti_m256i( d2,n ) = mm256_get_64( s, 2, 10, 18, 26 );
casti_m256i( d3,n ) = mm256_get_64( s, 3, 11, 19, 27 );
casti_m256i( d4,n ) = mm256_get_64( s, 4, 12, 20, 28 );
casti_m256i( d5,n ) = mm256_get_64( s, 5, 13, 21, 29 );
casti_m256i( d6,n ) = mm256_get_64( s, 6, 14, 22, 30 );
casti_m256i( d7,n ) = mm256_get_64( s, 7, 15, 23, 31 );
}
static inline void mm256_dintrlv_4x128x256( void *d0, void *d1, void *d2,
void *d3, const int n, const void *s )
{
casti_m256i( d0,n ) = mm256_get_64( s, 0, 1, 8, 9 );
casti_m256i( d1,n ) = mm256_get_64( s, 2, 3, 10, 11 );
casti_m256i( d2,n ) = mm256_get_64( s, 4, 5, 12, 13 );
casti_m256i( d3,n ) = mm256_get_64( s, 6, 7, 14, 15 );
}
// AVX 512 helper functions.
//
// Macro functions returning vector.
// Abstracted typecasting, avoid temp pointers.
// Source arguments may be any 64 or 32 byte aligned pointer as appropriate.
#define mm512_put_64( s0, s1, s2, s3, s4, s5, s6, s7 ) \
_mm512_set_epi64( *((const uint64_t*)(s7)), *((const uint64_t*)(s6)), \
*((const uint64_t*)(s5)), *((const uint64_t*)(s4)), \
*((const uint64_t*)(s3)), *((const uint64_t*)(s2)), \
*((const uint64_t*)(s1)), *((const uint64_t*)(s0)) )
#define mm512_put_32( s00, s01, s02, s03, s04, s05, s06, s07, \
s08, s09, s10, s11, s12, s13, s14, s15 ) \
_mm512_set_epi32( *((const uint32_t*)(s15)), *((const uint32_t*)(s14)), \
*((const uint32_t*)(s13)), *((const uint32_t*)(s12)), \
*((const uint32_t*)(s11)), *((const uint32_t*)(s10)), \
*((const uint32_t*)(s09)), *((const uint32_t*)(s08)), \
*((const uint32_t*)(s07)), *((const uint32_t*)(s06)), \
*((const uint32_t*)(s05)), *((const uint32_t*)(s04)), \
*((const uint32_t*)(s03)), *((const uint32_t*)(s02)), \
*((const uint32_t*)(s01)), *((const uint32_t*)(s00)) )
#define mm512_get_64( s, i0, i1, i2, i3, i4, i5, i6, i7 ) \
_mm512_set_epi64( ((const uint64_t*)(s))[i7], ((const uint64_t*)(s))[i6], \
((const uint64_t*)(s))[i5], ((const uint64_t*)(s))[i4], \
((const uint64_t*)(s))[i3], ((const uint64_t*)(s))[i2], \
((const uint64_t*)(s))[i1], ((const uint64_t*)(s))[i0] )
#define mm512_get_32( s, i00, i01, i02, i03, i04, i05, i06, i07, \
i08, i09, i10, i11, i12, i13, i14, i15 ) \
_mm512_set_epi32( ((const uint32_t*)(s))[i15], ((const uint32_t*)(s))[i14], \
((const uint32_t*)(s))[i13], ((const uint32_t*)(s))[i12], \
((const uint32_t*)(s))[i11], ((const uint32_t*)(s))[i10], \
((const uint32_t*)(s))[i09], ((const uint32_t*)(s))[i08], \
((const uint32_t*)(s))[i07], ((const uint32_t*)(s))[i06], \
((const uint32_t*)(s))[i05], ((const uint32_t*)(s))[i04], \
((const uint32_t*)(s))[i03], ((const uint32_t*)(s))[i02], \
((const uint32_t*)(s))[i01], ((const uint32_t*)(s))[i00] )
// AVX512 has no blend, can be done with permute2xvar but at what cost?
// Can also be done with shifting and mask-or'ing for 3 instructins with
// 1 dependency. Finally it can be done with 1 _mm512_set but with 8 64 bit
// array index calculations and 8 pointer reads.
// Blend 2 vectors alternating hi & lo: { hi[n], lo[n-1], ... hi[1]. lo[0] }
#define mm512_interleave_blend_128( hi, lo ) \
_mm256_permute2xvar_epi64( hi, lo, _mm512_set_epi64( \
0x7, 0x6, 0x5, 0x4, 0xb, 0xa, 0x9, 0x8 )
#define mm512_interleave_blend_64( hi, lo ) \
_mm256_permute2xvar_epi64( hi, lo, _mm512_set_epi64( \
0x7, 0x6, 0xd, 0xc, 0x3, 0x2, 0x9, 0x8 )
#define mm512_interleave_blend_32( hi, lo ) \
_mm256_permute2xvar_epi32( hi, lo, _mm512_set_epi32( \
0x0f, 0x1e, 0x0d, 0x1c, 0x0b, 0x1a, 0x09, 0x18, \
0x07, 0x16, 0x05, 0x14, 0x03, 0x12, 0x01, 0x10 )
//
static inline void mm512_intrlv_16x32x512( void *d, const void *s00,
const void *s01, const void *s02, const void *s03, const void *s04,
const void *s05, const void *s06, const void *s07, const void *s08,
const void *s09, const void *s10, const void *s11, const void *s12,
const void *s13, const void *s14, const void *s15 )
{
casti_m512i( d, 0 ) = mm512_put_32(
s00, s01, s02, s03, s04, s05, s06, s07,
s08, s09, s10, s11, s12, s13, s14, s15 );
casti_m512i( d, 1 ) = mm512_put_32(
s00+ 4, s01+ 4, s02+ 4, s03+ 4, s04+ 4, s05+ 4, s06+ 4, s07+ 4,
s08+ 4, s09+ 4, s10+ 4, s11+ 4, s12+ 4, s13+ 4, s14+ 4, s15+ 4 );
casti_m512i( d, 2 ) = mm512_put_32(
s00+ 8, s01+ 8, s02+ 8, s03+ 8, s04+ 8, s05+ 8, s06+ 8, s07+ 8,
s08+ 8, s09+ 8, s10+ 8, s11+ 8, s12+ 8, s13+ 8, s14+ 8, s15+ 8 );
casti_m512i( d, 3 ) = mm512_put_32(
s00+12, s01+12, s02+12, s03+12, s04+12, s05+12, s06+12, s07+12,
s08+12, s09+12, s10+12, s11+12, s12+12, s13+12, s14+12, s15+12 );
casti_m512i( d, 4 ) = mm512_put_32(
s00+16, s01+16, s02+16, s03+16, s04+16, s05+16, s06+16, s07+16,
s08+16, s09+16, s10+16, s11+16, s12+16, s13+16, s14+16, s15+16 );
casti_m512i( d, 5 ) = mm512_put_32(
s00+20, s01+20, s02+20, s03+20, s04+20, s05+20, s06+20, s07+20,
s08+20, s09+20, s10+20, s11+20, s12+20, s13+20, s14+20, s15+20 );
casti_m512i( d, 6 ) = mm512_put_32(
s00+24, s01+24, s02+24, s03+24, s04+24, s05+24, s06+24, s07+24,
s08+24, s09+24, s10+24, s11+24, s12+24, s13+24, s14+24, s15+24 );
casti_m512i( d, 7 ) = mm512_put_32(
s00+28, s01+28, s02+28, s03+28, s04+28, s05+28, s06+28, s07+28,
s08+28, s09+28, s10+28, s11+28, s12+28, s13+28, s14+28, s15+28 );
casti_m512i( d, 8 ) = mm512_put_32(
s00+32, s01+28, s02+28, s03+28, s04+32, s05+28, s06+28, s07+28,
s08+32, s09+28, s10+28, s11+28, s12+32, s13+28, s14+28, s15+28 );
casti_m512i( d, 9 ) = mm512_put_32(
s00+36, s01+28, s02+28, s03+28, s04+36, s05+28, s06+28, s07+28,
s08+36, s09+28, s10+28, s11+28, s12+36, s13+28, s14+28, s15+28 );
casti_m512i( d,10 ) = mm512_put_32(
s00+40, s01+28, s02+28, s03+28, s04+40, s05+28, s06+28, s07+28,
s08+40, s09+28, s10+28, s11+28, s12+40, s13+28, s14+28, s15+28 );
casti_m512i( d,11 ) = mm512_put_32(
s00+44, s01+28, s02+28, s03+28, s04+44, s05+28, s06+28, s07+28,
s08+44, s09+28, s10+28, s11+28, s12+44, s13+28, s14+28, s15+28 );
casti_m512i( d,12 ) = mm512_put_32(
s00+48, s01+28, s02+28, s03+28, s04+48, s05+28, s06+28, s07+28,
s08+48, s09+28, s10+28, s11+28, s12+48, s13+28, s14+28, s15+28 );
casti_m512i( d,13 ) = mm512_put_32(
s00+52, s01+28, s02+28, s03+28, s04+52, s05+28, s06+28, s07+28,
s08+52, s09+28, s10+28, s11+28, s12+52, s13+28, s14+28, s15+28 );
casti_m512i( d,14 ) = mm512_put_32(
s00+56, s01+28, s02+28, s03+28, s04+56, s05+28, s06+28, s07+28,
s08+56, s09+28, s10+28, s11+28, s12+56, s13+28, s14+28, s15+28 );
casti_m512i( d,15 ) = mm512_put_32(
s00+60, s01+28, s02+28, s03+28, s04+60, s05+28, s06+28, s07+28,
s08+60, s09+28, s10+28, s11+28, s12+60, s13+28, s14+28, s15+28 );
}
static inline void mm512_intrlv_16x32x256( void *d, const void *s00,
const void *s01, const void *s02, const void *s03, const void *s04,
const void *s05, const void *s06, const void *s07, const void *s08,
const void *s09, const void *s10, const void *s11, const void *s12,
const void *s13, const void *s14, const void *s15 )
{
casti_m512i( d, 0 ) = mm512_put_32(
s00, s01, s02, s03, s04, s05, s06, s07,
s08, s09, s10, s11, s12, s13, s14, s15 );
casti_m512i( d, 1 ) = mm512_put_32(
s00+ 4, s01+ 4, s02+ 4, s03+ 4, s04+ 4, s05+ 4, s06+ 4, s07+ 4,
s08+ 4, s09+ 4, s10+ 4, s11+ 4, s12+ 4, s13+ 4, s14+ 4, s15+ 4 );
casti_m512i( d, 2 ) = mm512_put_32(
s00+ 8, s01+ 8, s02+ 8, s03+ 8, s04+ 8, s05+ 8, s06+ 8, s07+ 8,
s08+ 8, s09+ 8, s10+ 8, s11+ 8, s12+ 8, s13+ 8, s14+ 8, s15+ 8 );
casti_m512i( d, 3 ) = mm512_put_32(
s00+12, s01+12, s02+12, s03+12, s04+12, s05+12, s06+12, s07+12,
s08+12, s09+12, s10+12, s11+12, s12+12, s13+12, s14+12, s15+12 );
casti_m512i( d, 4 ) = mm512_put_32(
s00+16, s01+16, s02+16, s03+16, s04+16, s05+16, s06+16, s07+16,
s08+16, s09+16, s10+16, s11+16, s12+16, s13+16, s14+16, s15+16 );
casti_m512i( d, 5 ) = mm512_put_32(
s00+20, s01+20, s02+20, s03+20, s04+20, s05+20, s06+20, s07+20,
s08+20, s09+20, s10+20, s11+20, s12+20, s13+20, s14+20, s15+20 );
casti_m512i( d, 6 ) = mm512_put_32(
s00+24, s01+24, s02+24, s03+24, s04+24, s05+24, s06+24, s07+24,
s08+24, s09+24, s10+24, s11+24, s12+24, s13+24, s14+24, s15+24 );
casti_m512i( d, 7 ) = mm512_put_32(
s00+28, s01+28, s02+28, s03+28, s04+28, s05+28, s06+28, s07+28,
s08+28, s09+28, s10+28, s11+28, s12+28, s13+28, s14+28, s15+28 );
}
// Last 16 bytes of input
static inline void mm512_intrlv_16x32x128( void *d, const void *s00,
const void *s01, const void *s02, const void *s03, const void *s04,
const void *s05, const void *s06, const void *s07, const void *s08,
const void *s09, const void *s10, const void *s11, const void *s12,
const void *s13, const void *s14, const void *s15 )
{
casti_m512i( d, 0 ) = mm512_put_32(
s00, s01, s02, s03, s04, s05, s06, s07,
s08, s09, s10, s11, s12, s13, s14, s15 );
casti_m512i( d, 1 ) = mm512_put_32(
s00+ 4, s01+ 4, s02+ 4, s03+ 4, s04+ 4, s05+ 4, s06+ 4, s07+ 4,
s08+ 4, s09+ 4, s10+ 4, s11+ 4, s12+ 4, s13+ 4, s14+ 4, s15+ 4 );
casti_m512i( d, 2 ) = mm512_put_32(
s00+ 8, s01+ 8, s02+ 8, s03+ 8, s04+ 8, s05+ 8, s06+ 8, s07+ 8,
s08+ 8, s09+ 8, s10+ 8, s11+ 8, s12+ 8, s13+ 8, s14+ 8, s15+ 8 );
casti_m512i( d, 3 ) = mm512_put_32(
s00+12, s01+12, s02+12, s03+12, s04+12, s05+12, s06+12, s07+12,
s08+12, s09+12, s10+12, s11+12, s12+12, s13+12, s14+12, s15+12 );
}
// can be called directly for 64 byte hash.
static inline void mm512_dintrlv_16x32x512( void *d00, void *d01,
void *d02, void *d03, void *d04, void *d05, void *d06,
void *d07, void *d08, void *d09, void *d10, void *d11,
void *d12, void *d13, void *d14, void *d15, const int n,
const void *s )
{
casti_m512i(d00,n) = mm512_get_32( s, 0, 16, 32, 48, 64, 80, 96,112,
128,144,160,176,192,208,224,240 );
casti_m512i(d01,n) = mm512_get_32( s, 1, 17, 33, 49, 65, 81, 97,113,
129,145,161,177,193,209,225,241 );
casti_m512i(d02,n) = mm512_get_32( s, 2, 18, 34, 50, 66, 82, 98,114,
130,146,162,178,194,210,226,242 );
casti_m512i(d03,n) = mm512_get_32( s, 3, 19, 35, 51, 67, 83, 99,115,
131,147,163,179,195,211,227,243 );
casti_m512i(d04,n) = mm512_get_32( s, 4, 20, 36, 52, 68, 84,100,116,
132,148,164,180,196,212,228,244 );
casti_m512i(d05,n) = mm512_get_32( s, 5, 21, 37, 53, 69, 85,101,117,
133,149,165,181,197,213,229,245 );
casti_m512i(d06,n) = mm512_get_32( s, 6, 22, 38, 54, 70, 86,102,118,
134,150,166,182,198,214,230,246 );
casti_m512i(d07,n) = mm512_get_32( s, 7, 23, 39, 55, 71, 87,103,119,
135,151,167,183,199,215,231,247 );
casti_m512i(d08,n) = mm512_get_32( s, 8, 24, 40, 56, 72, 88,104,120,
136,152,168,184,200,216,232,248 );
casti_m512i(d09,n) = mm512_get_32( s, 9, 25, 41, 57, 73, 89,105,121,
137,153,169,185,201,217,233,249 );
casti_m512i(d10,n) = mm512_get_32( s, 10, 26, 42, 58, 74, 90,106,122,
138,154,170,186,202,218,234,250 );
casti_m512i(d11,n) = mm512_get_32( s, 11, 27, 43, 59, 75, 91,107,123,
139,155,171,187,203,219,235,251 );
casti_m512i(d12,n) = mm512_get_32( s, 12, 28, 44, 60, 76, 92,108,124,
140,156,172,188,204,220,236,252 );
casti_m512i(d13,n) = mm512_get_32( s, 13, 29, 45, 61, 77, 93,109,125,
141,157,173,189,205,221,237,253 );
casti_m512i(d14,n) = mm512_get_32( s, 14, 30, 46, 62, 78, 94,110,126,
142,158,174,190,206,222,238,254 );
casti_m512i(d15,n) = mm512_get_32( s, 15, 31, 47, 63, 79, 95,111,127,
143,159,175,191,207,223,239,255 );
}
static inline void mm512_intrlv_8x64x512( void *d, const void *s0,
const void *s1, const void *s2, const void *s3,
const void *s4, const void *s5, const void *s6,
const void *s7 )
{
casti_m512i( d,0 ) = mm512_put_64( s0, s1, s2, s3,
s4, s5, s6, s7 );
casti_m512i( d,1 ) = mm512_put_64( s0+ 8, s1+ 8, s2+ 8, s3+ 8,
s4+ 8, s5+ 8, s6+ 8, s7+ 8 );
casti_m512i( d,2 ) = mm512_put_64( s0+16, s1+16, s2+16, s3+16,
s4+16, s5+16, s6+16, s7+16 );
casti_m512i( d,3 ) = mm512_put_64( s0+24, s1+24, s2+24, s3+24,
s4+24, s5+24, s6+24, s7+24 );
casti_m512i( d,4 ) = mm512_put_64( s0+32, s1+32, s2+32, s3+32,
s4+32, s5+32, s6+32, s7+32 );
casti_m512i( d,5 ) = mm512_put_64( s0+40, s1+40, s2+40, s3+40,
s4+40, s5+40, s6+40, s7+40 );
casti_m512i( d,6 ) = mm512_put_64( s0+48, s1+48, s2+48, s3+48,
s4+48, s5+48, s6+48, s7+48 );
casti_m512i( d,7 ) = mm512_put_64( s0+56, s1+56, s2+56, s3+56,
s4+56, s5+56, s6+56, s7+56 );
}
static inline void mm512_intrlv_8x64x256( void *d, const void *s0,
const void *s1, const void *s2, const void *s3,
const void *s4, const void *s5, const void *s6,
const void *s7 )
{
casti_m512i( d,0 ) = mm512_put_64( s0, s1, s2, s3,
s4, s5, s6, s7 );
casti_m512i( d,1 ) = mm512_put_64( s0+ 8, s1+ 8, s2+ 8, s3+ 8,
s4+ 8, s5+ 8, s6+ 8, s7+ 8 );
casti_m512i( d,2 ) = mm512_put_64( s0+16, s1+16, s2+16, s3+16,
s4+16, s5+16, s6+16, s7+16 );
casti_m512i( d,3 ) = mm512_put_64( s0+24, s1+24, s2+24, s3+24,
s4+24, s5+24, s6+24, s7+24 );
}
// 8 lanes of 512 bits using 64 bit interleaving (typical intermediate hash)
static inline void mm512_dintrlv_8x64x512( void *d0, void *d1, void *d2,
void *d3, void *d4, void *d5, void *d6, void *d7,
const int n, const void *s )
{
casti_m512i( d0,n ) = mm512_get_64( s, 0, 8, 16, 24, 32, 40, 48, 56 );
casti_m512i( d1,n ) = mm512_get_64( s, 1, 9, 17, 25, 33, 41, 49, 57 );
casti_m512i( d2,n ) = mm512_get_64( s, 2, 10, 18, 26, 34, 42, 50, 58 );
casti_m512i( d3,n ) = mm512_get_64( s, 3, 11, 19, 27, 35, 43, 51, 59 );
casti_m512i( d4,n ) = mm512_get_64( s, 4, 12, 20, 28, 36, 44, 52, 60 );
casti_m512i( d5,n ) = mm512_get_64( s, 5, 13, 21, 29, 37, 45, 53, 61 );
casti_m512i( d6,n ) = mm512_get_64( s, 6, 14, 22, 30, 38, 46, 54, 62 );
casti_m512i( d7,n ) = mm512_get_64( s, 7, 15, 23, 31, 39, 47, 55, 63 );
}
static inline void mm512_dintrlv_4x128x512( void *d0, void *d1, void *d2,
void *d3, const int n, const void *s )
{
casti_m512i( d0,n ) = mm512_get_64( s, 0, 1, 8, 9, 16, 17, 24, 25 );
casti_m512i( d1,n ) = mm512_get_64( s, 2, 3, 10, 11, 18, 19, 16, 27 );
casti_m512i( d2,n ) = mm512_get_64( s, 4, 5, 12, 13, 20, 21, 28, 29 );
casti_m512i( d3,n ) = mm512_get_64( s, 6, 7, 14, 15, 22, 23, 30, 31 );
}
// AVX-512 user facing functions.
static inline void mm512_intrlv_16x32( void *d, const void *s00,
const void *s01, const void *s02, const void *s03, const void *s04,
const void *s05, const void *s06, const void *s07, const void *s08,
const void *s09, const void *s10, const void *s11, const void *s12,
const void *s13, const void *s14, const void *s15, int bit_len )
{
if ( bit_len <= 256 )
{
mm512_intrlv_16x32x256( d, s00, s01, s02, s03, s04, s05, s06, s07,
s08, s09, s10, s11, s12, s13, s14, s15 );
return;
}
mm512_intrlv_16x32x512( d, s00, s01, s02, s03, s04, s05, s06, s07,
s08, s09, s10, s11, s12, s13, s14, s15 );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm512_intrlv_16x32x128( d+1024, s00+64, s01+64, s02+64, s03+64,
s04+64, s05+64, s06+64, s07+64, s08+64, s09+64,
s10+64, s11+64, s12+64, s13+64, s14+64, s15+64 );
return;
}
mm512_intrlv_16x32x512( d+1024, s00+64, s01+64, s02+64, s03+64,
s04+64, s05+64, s06+64, s07+64, s08+64, s09+64,
s10+64, s11+64, s12+64, s13+64, s14+64, s15+64 );
// bit_len == 1024
}
// sub-functions can be called directly for 32 & 64 byte hash.
static inline void mm512_dintrlv_16x32( void *d00, void *d01, void *d02,
void *d03, void *d04, void *d05, void *d06, void *d07, void *d08,
void *d09, void *d10, void *d11, void *d12, void *d13, void *d14,
void *d15, const void *src, const int bit_len )
{
if ( bit_len <= 256 )
{
mm256_dintrlv_16x32x256( d00, d01, d02, d03, d04, d05, d06, d07,
d08, d09, d10, d11, d12, d13, d14, d15,
0,src );
return;
}
mm512_dintrlv_16x32x512( d00, d01, d02, d03, d04, d05, d06, d07,
d08, d09, d10, d11, d12, d13, d14, d15,
0, src );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
// short block, final 16 bytes of input data.
mm128_dintrlv_16x32x128( d00, d01, d02, d03, d04, d05, d06, d07,
d08, d09, d10, d11, d12, d13, d14, d15,
1, src+1024 );
return;
}
// bit_len == 1024
mm512_dintrlv_16x32x512( d00, d01, d02, d03, d04, d05, d06, d07,
d08, d09, d10, d11, d12, d13, d14, d15,
1, src+1024 );
}
static inline void mm512_extr_lane_16x32( void *dst, const void *src,
const int lane, const int bit_len )
{
if ( bit_len <= 256 )
{
cast_m256i( dst ) = mm256_get_32( src, lane, lane+16, lane+32, lane+48,
lane+64, lane+80, lane+96, lane+112 );
return;
}
cast_m512i( dst ) = mm512_get_32( src, lane, lane+ 16, lane+ 32, lane+ 48,
lane+ 64, lane+ 80, lane+ 96, lane+112, lane+128, lane+144,
lane+160, lane+176, lane+192, lane+208, lane+224, lane+248 );
}
//
static inline void mm512_intrlv_8x64( void *d, const void *s0,
const void *s1, const void *s2, const void *s3,
const void *s4, const void *s5, const void *s6,
const void *s7, int bit_len )
{
if ( bit_len <= 256 )
{
mm512_intrlv_8x64x256( d, s0, s1, s2, s3, s4, s5, s6, s7 );
return;
}
mm512_intrlv_8x64x512( d, s0, s1, s2, s3, s4, s5, s6, s7 );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
casti_m512i( d, 8 ) = mm512_put_64( s7+64, s6+64, s5+64, s4+64,
s3+64, s2+64, s1+64, s0+64 );
casti_m512i( d, 9 ) = mm512_put_64( s7+72, s6+72, s5+72, s4+72,
s3+72, s2+72, s1+72, s0+72 );
return;
}
// bitlen == 1024
mm512_intrlv_8x64x512( d+512, s0+64, s1+64, s2+64, s3+64,
s4+64, s5+64, s6+64, s7+64 );
}
static inline void mm512_dintrlv_8x64( void *d0, void *d1, void *d2,
void *d3, void *d4, void *d5, void *d6, void *d7,
const void *s, const int bit_len )
{
if ( bit_len <= 256 )
{
mm256_dintrlv_8x64x256( d0, d1, d2, d3, d4, d5, d6, d7, 0, s );
return;
}
mm512_dintrlv_8x64x512( d0, d1, d2, d3, d4, d5, d6, d7, 0, s );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
// short block, final 16 bytes of input data.
mm128_dintrlv_8x64x128( d0, d1, d2, d3, d4, d5, d6, d7, 1, s+512 );
return;
}
// bit_len == 1024
mm512_dintrlv_8x64x512( d0, d1, d2, d3, d4, d5, d6, d7, 1, s+512 );
}
// Extract one lane from 64 bit interleaved data
static inline void mm512_extr_lane_8x64( void *d, const void *s,
const int lane, const int bit_len )
{
if ( bit_len <= 256 )
{
cast_m256i( d ) = mm256_get_64( s, lane, lane+8, lane+16, lane+24 );
return;
}
// else bit_len == 512
cast_m512i( d ) = mm512_get_64( s, lane , lane+ 8, lane+16, lane+24,
lane+32, lane+40, lane+48, lane+56 );
}
//
static inline void mm512_intrlv_4x128( void *d, const void *s0,
const void *s1, const void *s2, const void *s3, const int bit_len )
{
casti_m512i( d, 0 ) = mm512_put_64( s0, s0+8, s1, s1+8,
s2, s2+8, s3, s3+8 );
casti_m512i( d, 1 ) = mm512_put_64( s0+16, s0+24, s1+16, s1+24,
s2+16, s2+24, s3+16, s3+24 );
if ( bit_len <= 256 ) return;
casti_m512i( d, 2 ) = mm512_put_64( s0+32, s0+40, s1+32, s1+40,
s2+32, s2+40, s3+32, s3+40 );
casti_m512i( d, 3 ) = mm512_put_64( s0+48, s0+56, s1+48, s1+56,
s2+48, s2+56, s3+48, s3+56 );
if ( bit_len <= 512 ) return;
casti_m512i( d, 4 ) = mm512_put_64( s0+64, s0+72, s1+64, s1+72,
s2+64, s2+72, s3+64, s3+72 );
if ( bit_len <= 640 ) return;
casti_m512i( d, 5 ) = mm512_put_64( s0+ 80, s0+ 88, s1+ 80, s1+ 88,
s2+ 80, s2+ 88, s3+ 80, s3+ 88 );
casti_m512i( d, 6 ) = mm512_put_64( s0+ 96, s0+104, s1+ 96, s1+104,
s2+ 96, s2+104, s3+ 96, s3+104 );
casti_m512i( d, 7 ) = mm512_put_64( s0+112, s0+120, s1+112, s1+120,
s2+112, s2+120, s3+112, s3+120 );
// bit_len == 1024
}
static inline void mm512_dintrlv_4x128( void *d0, void *d1, void *d2,
void *d3, const void *s, const int bit_len )
{
if ( bit_len <= 256 )
{
mm256_dintrlv_4x128x256( d0, d1, d2, d3, 0, s );
return;
}
mm512_dintrlv_4x128x512( d0, d1, d2, d3, 0, s );
if ( bit_len <= 512 ) return;
if ( bit_len <= 640 )
{
mm128_dintrlv_4x128x128( d0, d1, d2, d3, 1, s+256 );
return;
}
// bit_len == 1024
mm512_dintrlv_4x128x512( d0, d1, d2, d3, 1, s+256 );
}
// input one 8x64 buffer and return 2*4*128
static inline void mm512_rintrlv_8x64_4x128( void *dst0, void *dst1,
const void *src, int bit_len )
{
__m512i* d0 = (__m512i*)dst0;
__m512i* d1 = (__m512i*)dst1;
uint64_t *s = (uint64_t*)src;
d0[0] = _mm512_set_epi64( s[ 11], s[ 3], s[ 10], s[ 2],
s[ 9], s[ 1], s[ 8], s[ 0] );
d0[1] = _mm512_set_epi64( s[ 27], s[ 19], s[ 26], s[ 18],
s[ 25], s[ 17], s[ 24], s[ 16] );
d0[2] = _mm512_set_epi64( s[ 15], s[ 7], s[ 14], s[ 6],
s[ 13], s[ 5], s[ 12], s[ 4] );
d0[3] = _mm512_set_epi64( s[ 31], s[ 23], s[ 30], s[ 22],
s[ 29], s[ 21], s[ 28], s[ 20] );
d1[0] = _mm512_set_epi64( s[ 43], s[ 35], s[ 42], s[ 34],
s[ 41], s[ 33], s[ 40], s[ 32] );
d1[1] = _mm512_set_epi64( s[ 59], s[ 51], s[ 58], s[ 50],
s[ 57], s[ 49], s[ 56], s[ 48] );
d1[2] = _mm512_set_epi64( s[ 47], s[ 39], s[ 46], s[ 38],
s[ 45], s[ 37], s[ 44], s[ 36] );
d1[3] = _mm512_set_epi64( s[ 63], s[ 55], s[ 62], s[ 54],
s[ 61], s[ 53], s[ 60], s[ 52] );
if ( bit_len <= 512 ) return;
d0[4] = _mm512_set_epi64( s[ 75], s[ 67], s[ 74], s[ 66],
s[ 73], s[ 65], s[ 72], s[ 64] );
d0[5] = _mm512_set_epi64( s[ 91], s[ 83], s[ 90], s[ 82],
s[ 89], s[ 81], s[ 88], s[ 80] );
d0[6] = _mm512_set_epi64( s[ 79], s[ 71], s[ 78], s[ 70],
s[ 77], s[ 69], s[ 76], s[ 68] );
d0[7] = _mm512_set_epi64( s[ 95], s[ 87], s[ 94], s[ 86],
s[ 93], s[ 85], s[ 92], s[ 84] );
d1[4] = _mm512_set_epi64( s[107], s[ 99], s[106], s[ 98],
s[105], s[ 97], s[104], s[ 96] );
d1[5] = _mm512_set_epi64( s[123], s[115], s[122], s[114],
s[121], s[113], s[120], s[112] );
d1[6] = _mm512_set_epi64( s[111], s[103], s[110], s[102],
s[109], s[101], s[108], s[100] );
d1[7] = _mm512_set_epi64( s[127], s[119], s[126], s[118],
s[125], s[117], s[124], s[116] );
}
// input 2 4x128 return 8x64
static inline void mm512_rintrlv_4x128_8x64( void *dst, const void *src0,
const void *src1, int bit_len )
{
__m512i* d = (__m512i*)dst;
uint64_t *s0 = (uint64_t*)src0;
uint64_t *s1 = (uint64_t*)src1;
d[0] = _mm512_set_epi64( s1[ 6], s1[ 4], s1[ 2], s1[ 0],
s0[ 6], s0[ 4], s0[ 2], s0[ 0] );
d[1] = _mm512_set_epi64( s1[ 7], s1[ 5], s1[ 3], s1[ 1],
s0[ 7], s0[ 5], s0[ 3], s0[ 1] );
d[2] = _mm512_set_epi64( s1[14], s1[12], s1[10], s1[ 8],
s0[14], s0[12], s0[10], s0[ 8] );
d[3] = _mm512_set_epi64( s1[15], s1[13], s1[11], s1[ 9],
s0[15], s0[13], s0[11], s0[ 9] );
d[4] = _mm512_set_epi64( s1[22], s1[20], s1[18], s1[16],
s0[22], s0[20], s0[18], s0[16] );
d[5] = _mm512_set_epi64( s1[23], s1[21], s1[19], s1[17],
s0[24], s0[21], s0[19], s0[17] );
d[6] = _mm512_set_epi64( s1[22], s1[28], s1[26], s1[24],
s0[22], s0[28], s0[26], s0[24] );
d[7] = _mm512_set_epi64( s1[31], s1[29], s1[27], s1[25],
s0[31], s0[29], s0[27], s0[25] );
if ( bit_len <= 512 ) return;
d[0] = _mm512_set_epi64( s1[38], s1[36], s1[34], s1[32],
s0[38], s0[36], s0[34], s0[32] );
d[1] = _mm512_set_epi64( s1[39], s1[37], s1[35], s1[33],
s0[39], s0[37], s0[35], s0[33] );
d[2] = _mm512_set_epi64( s1[46], s1[44], s1[42], s1[40],
s0[46], s0[44], s0[42], s0[40] );
d[3] = _mm512_set_epi64( s1[47], s1[45], s1[43], s1[41],
s0[47], s0[45], s0[43], s0[41] );
d[4] = _mm512_set_epi64( s1[54], s1[52], s1[50], s1[48],
s0[54], s0[52], s0[50], s0[48] );
d[5] = _mm512_set_epi64( s1[55], s1[53], s1[51], s1[49],
s0[55], s0[53], s0[51], s0[49] );
d[6] = _mm512_set_epi64( s1[62], s1[60], s1[58], s1[56],
s0[62], s0[60], s0[58], s0[56] );
d[7] = _mm512_set_epi64( s1[63], s1[61], s1[59], s1[57],
s0[63], s0[61], s0[59], s0[57] );
}
static inline void mm512_extr_lane_4x128( void *d, const void *s,
const int lane, const int bit_len )
{
int l = lane<<1;
if ( bit_len <= 256 )
{
cast_m256i( d ) = mm256_get_64( s, l, l+1, l+8, l+9 );
return;
}
// else bit_len == 512
cast_m512i( d ) = mm512_get_64( s, l , l+ 1, l+ 8, l+ 9,
l+16, l+17, l+24, l+25 );
}
#endif // AVX512
#endif // INTRLV_AVX512_H__

126
simd-utils/intrlv-mmx.h
View File

@@ -1,126 +0,0 @@
#if !defined(INTRLV_MMX_H__)
#define INTRLV_MMX_H__ 1
#if defined(__MMX__)
//////////////////////////////////////////////////////
//
// MMX 64 bit vectors
#define mm64_put_32( s0, s1 ) \
_mm_set_pi32( *((const uint32_t*)(s1)), *((const uint32_t*)(s0)) )
#define mm64_get_32( s, i0, i1 ) \
_mm_set_pi32( ((const uint32_t*)(s))[i1], ((const uint32_t*)(s))[i0] )
// 1 MMX block, 8 bytes * 2 lanes
static inline void mm64_intrlv_2x32( void *d, const void *s0,
const void *s1, int len )
{
casti_m64( d, 0 ) = mm64_put_32( s0 , s1 );
casti_m64( d, 1 ) = mm64_put_32( s0+ 4, s1+ 4 );
casti_m64( d, 2 ) = mm64_put_32( s0+ 8, s1+ 8 );
casti_m64( d, 3 ) = mm64_put_32( s0+ 12, s1+ 12 );
casti_m64( d, 4 ) = mm64_put_32( s0+ 16, s1+ 16 );
casti_m64( d, 5 ) = mm64_put_32( s0+ 20, s1+ 20 );
casti_m64( d, 6 ) = mm64_put_32( s0+ 24, s1+ 24 );
casti_m64( d, 7 ) = mm64_put_32( s0+ 28, s1+ 28 );
if ( len <= 256 ) return;
casti_m64( d, 8 ) = mm64_put_32( s0+ 32, s1+ 32 );
casti_m64( d, 9 ) = mm64_put_32( s0+ 36, s1+ 36 );
casti_m64( d,10 ) = mm64_put_32( s0+ 40, s1+ 40 );
casti_m64( d,11 ) = mm64_put_32( s0+ 44, s1+ 44 );
casti_m64( d,12 ) = mm64_put_32( s0+ 48, s1+ 48 );
casti_m64( d,13 ) = mm64_put_32( s0+ 52, s1+ 52 );
casti_m64( d,14 ) = mm64_put_32( s0+ 56, s1+ 56 );
casti_m64( d,15 ) = mm64_put_32( s0+ 60, s1+ 60 );
if ( len <= 512 ) return;
casti_m64( d,16 ) = mm64_put_32( s0+ 64, s1+ 64 );
casti_m64( d,17 ) = mm64_put_32( s0+ 68, s1+ 68 );
casti_m64( d,18 ) = mm64_put_32( s0+ 72, s1+ 72 );
casti_m64( d,19 ) = mm64_put_32( s0+ 76, s1+ 76 );
if ( len <= 640 ) return;
casti_m64( d,20 ) = mm64_put_32( s0+ 80, s1+ 80 );
casti_m64( d,21 ) = mm64_put_32( s0+ 84, s1+ 84 );
casti_m64( d,22 ) = mm64_put_32( s0+ 88, s1+ 88 );
casti_m64( d,23 ) = mm64_put_32( s0+ 92, s1+ 92 );
casti_m64( d,24 ) = mm64_put_32( s0+ 96, s1+ 96 );
casti_m64( d,25 ) = mm64_put_32( s0+100, s1+100 );
casti_m64( d,26 ) = mm64_put_32( s0+104, s1+104 );
casti_m64( d,27 ) = mm64_put_32( s0+108, s1+108 );
casti_m64( d,28 ) = mm64_put_32( s0+112, s1+112 );
casti_m64( d,29 ) = mm64_put_32( s0+116, s1+116 );
casti_m64( d,30 ) = mm64_put_32( s0+120, s1+120 );
casti_m64( d,31 ) = mm64_put_32( s0+124, s1+124 );
}
static inline void mm64_dintrlv_2x32( void *d00, void *d01, const int n,
const void *s, int len )
{
casti_m64( d00,0 ) = mm64_get_32( s, 0, 2 );
casti_m64( d01,0 ) = mm64_get_32( s, 1, 3 );
casti_m64( d00,1 ) = mm64_get_32( s, 4, 6 );
casti_m64( d01,1 ) = mm64_get_32( s, 5, 7 );
casti_m64( d00,2 ) = mm64_get_32( s, 8, 10 );
casti_m64( d01,2 ) = mm64_get_32( s, 9, 11 );
casti_m64( d00,3 ) = mm64_get_32( s, 12, 14 );
casti_m64( d01,3 ) = mm64_get_32( s, 13, 15 );
if ( len <= 256 ) return;
casti_m64( d00,4 ) = mm64_get_32( s, 16, 18 );
casti_m64( d01,4 ) = mm64_get_32( s, 17, 19 );
casti_m64( d00,5 ) = mm64_get_32( s, 20, 22 );
casti_m64( d01,5 ) = mm64_get_32( s, 21, 23 );
casti_m64( d00,6 ) = mm64_get_32( s, 24, 26 );
casti_m64( d01,6 ) = mm64_get_32( s, 25, 27 );
casti_m64( d00,7 ) = mm64_get_32( s, 28, 30 );
casti_m64( d01,7 ) = mm64_get_32( s, 29, 31 );
if ( len <= 512 ) return;
casti_m64( d00,8 ) = mm64_get_32( s, 32, 34 );
casti_m64( d01,8 ) = mm64_get_32( s, 33, 35 );
casti_m64( d00,9 ) = mm64_get_32( s, 36, 38 );
casti_m64( d01,9 ) = mm64_get_32( s, 37, 39 );
if ( len <= 640 ) return;
casti_m64( d00,10 ) = mm64_get_32( s, 40, 42 );
casti_m64( d01,10 ) = mm64_get_32( s, 41, 43 );
casti_m64( d00,11 ) = mm64_get_32( s, 44, 46 );
casti_m64( d01,11 ) = mm64_get_32( s, 45, 47 );
casti_m64( d00,12 ) = mm64_get_32( s, 48, 50 );
casti_m64( d01,12 ) = mm64_get_32( s, 49, 51 );
casti_m64( d00,13 ) = mm64_get_32( s, 52, 54 );
casti_m64( d01,13 ) = mm64_get_32( s, 53, 55 );
casti_m64( d00,14 ) = mm64_get_32( s, 56, 58 );
casti_m64( d01,14 ) = mm64_get_32( s, 57, 59 );
casti_m64( d00,15 ) = mm64_get_32( s, 60, 62 );
casti_m64( d01,15 ) = mm64_get_32( s, 61, 63 );
}
static inline void mm64_extr_lane_2x32( void *d, const void *s,
const int lane, const int bit_len )
{
casti_m64( d, 0 ) = mm64_get_32( s, lane , lane+ 4 );
casti_m64( d, 1 ) = mm64_get_32( s, lane+ 8, lane+12 );
casti_m64( d, 2 ) = mm64_get_32( s, lane+16, lane+20 );
casti_m64( d, 3 ) = mm64_get_32( s, lane+24, lane+28 );
if ( bit_len <= 256 ) return;
casti_m64( d, 4 ) = mm64_get_32( s, lane+32, lane+36 );
casti_m64( d, 5 ) = mm64_get_32( s, lane+40, lane+44 );
casti_m64( d, 6 ) = mm64_get_32( s, lane+48, lane+52 );
casti_m64( d, 7 ) = mm64_get_32( s, lane+56, lane+60 );
// bit_len == 512
}
#endif // MMX
#endif // INTRLV_MMX_H__

77
simd-utils/intrlv-selector.h
View File

@@ -1,77 +0,0 @@
#if !defined(INTRLV_SELECTOR_H__)
#define INTRLV_SELECTOR_H__
//////////////////////////////////////////////////////////////
//
// Generic interface for interleaving data for parallel processing.
//
// Best tech is chosen atomatically.
/*
#if defined(__AVX512F__)
#define intrlv_4x128 mm512_intrlv_4x128
#define intrlv_4x128 mm512_intrlv_4x128
#define intrlv_8x64 mm512_intrlv_8x64
#define dintrlv_8x64 mm512_dintrlv_8x64
#define extr_lane_8x64 mm512_extr_lane_8x64
#define intrlv_16x32 mm512_intrlv_16x32
#define dintrlv_16x32 mm512_dintrlv_16x32
#define extr_lane_16x32 mm512_extr_lane_16x32
#define intrlv_2x128 mm512_intrlv_2x128
#define dintrlv_2x128 mm512_dintrlv_2x128
#define intrlv_4x64 mm512_intrlv_4x64
#define dintrlv_4x64 mm512_dintrlv_4x64
#define extr_lane_4x64 mm512_extr_lane_4x64
#define intrlv_8x32 mm512_intrlv_8x32
#define dintrlv_8x32 mm512_dintrlv_8x32
#define extr_lane_8x32 mm512_extr_lane_8x32
#elif defined(__AVX__)
*/
#if defined(__AVX__)
#define intrlv_2x128 mm256_intrlv_2x128
#define dintrlv_2x128 mm256_dintrlv_2x128
#define intrlv_4x64 mm256_intrlv_4x64
#define dintrlv_4x64 mm256_dintrlv_4x64
#define extr_lane_4x64 mm256_extr_lane_4x64
#define intrlv_8x32 mm256_intrlv_8x32
#define dintrlv_8x32 mm256_dintrlv_8x32
#define extr_lane_8x32 mm256_extr_lane_8x32
#define intrlv_4x32 mm256_intrlv_4x32
#define dintrlv_4x32 mm256_dintrlv_4x32
#define extr_lane_4x32 mm256_extr_lane_4x32
#else
#define intrlv_2x128 mm128_intrlv_2x128
#define dintrlv_2x128 mm128_dintrlv_2x128
#define intrlv_4x64 mm128_intrlv_4x64
#define dintrlv_4x64 mm128_dintrlv_4x64
#define extr_lane_4x64 mm128_extr_lane_4x64
#define intrlv_8x32 mm128_intrlv_8x32
#define dintrlv_8x32 mm128_dintrlv_8x32
#define extr_lane_8x32 mm128_extr_lane_8x32
#define intrlv_2x64 mm128_intrlv_2x64
#define dintrlv_2x64 mm128_dintrlv_2x64
#define extr_lane_2x64 mm128_extr_lane_2x64
#define intrlv_4x32 mm128_intrlv_4x32
#define dintrlv_4x32 mm128_dintrlv_4x32
#define extr_lane_4x32 mm128_extr_lane_4x32
#endif
#endif // INTRLV_SELECTOR_H__

192
simd-utils/intrlv-sse2.h
View File

@@ -1,192 +0,0 @@
#if !defined(INTRLV_SSE2_H__)
#define INTRLV_SSE2_H__ 1
// Don't call __mm_extract_epi32 directly, it needs SSE4.1.
// Use mm128_extr_32 wrapper instead, it has both SSE4.1 & SSE2 covered.
#if defined(__SSE2__)
///////////////////////////////////////////////////////////////
//
// SSE2 128 bit vectors
// Macros to abstract typecasting
// Interleave lanes
#define mm128_put_64( s0, s1) \
_mm_set_epi64x( *((const uint64_t*)(s1)), *((const uint64_t*)(s0)) )
#define mm128_put_32( s0, s1, s2, s3 ) \
_mm_set_epi32( *((const uint32_t*)(s3)), *((const uint32_t*)(s2)), \
*((const uint32_t*)(s1)), *((const uint32_t*)(s0)) )
// Deinterleave lanes
#define mm128_get_64( s, i0, i1 ) \
_mm_set_epi64x( ((const uint64_t*)(s))[i1], ((const uint64_t*)(s))[i0] )
#define mm128_get_32( s, i0, i1, i2, i3 ) \
_mm_set_epi32( ((const uint32_t*)(s))[i3], ((const uint32_t*)(s))[i2], \
((const uint32_t*)(s))[i1], ((const uint32_t*)(s))[i0] )
// blend 2 vectors while interleaving: { hi[n], lo[n-1], ... hi[1], lo[0] }
#define mm128_intrlv_blend_64( hi, lo ) \
_mm256_blend_epi16( hi, lo, 0x0f )
#define mm128_intrlv_blend_32( hi, lo ) \
_mm6_blend_epi16( hi, lo, 0x33 )
// 1 sse2 block, 16 x 16 bytes
#define mm128_intrlv_4x32_128( d, s0, s1, s2, s3 )\
do { \
casti_m128i( d,0 ) = _mm_set_epi32( \
mm128_extr_32( s3, 0 ), mm128_extr_32( s2, 0 ), \
mm128_extr_32( s1, 0 ), mm128_extr_32( s0, 0 ) ); \
casti_m128i( d,1 ) = _mm_set_epi32( \
mm128_extr_32( s3, 1 ), mm128_extr_32( s2, 1 ), \
mm128_extr_32( s1, 1 ), mm128_extr_32( s0, 1 ) ); \
casti_m128i( d,2 ) = _mm_set_epi32( \
mm128_extr_32( s3, 2 ), mm128_extr_32( s2, 2 ), \
mm128_extr_32( s1, 2 ), mm128_extr_32( s0, 2 ) ); \
casti_m128i( d,3 ) = _mm_set_epi32( \
mm128_extr_32( s3, 3 ), mm128_extr_32( s2, 3 ), \
mm128_extr_32( s1, 3 ), mm128_extr_32( s0, 3 ) ); \
} while(0)
static inline void mm128_dintrlv_4x32_128( void *d0, void *d1, void *d2,
void *d3, const void *src )
{
__m128i s0 = *(__m128i*) src;
__m128i s1 = *(__m128i*)(src+16);
__m128i s2 = *(__m128i*)(src+32);
__m128i s3 = *(__m128i*)(src+48);
*(__m128i*)d0 = _mm_set_epi32(
mm128_extr_32( s3,0 ), mm128_extr_32( s2,0 ),
mm128_extr_32( s1,0 ), mm128_extr_32( s0,0 ) );
*(__m128i*)d1 = _mm_set_epi32(
mm128_extr_32( s3,1 ), mm128_extr_32( s2,1 ),
mm128_extr_32( s1,1 ), mm128_extr_32( s0,1 ) );
*(__m128i*)d2 = _mm_set_epi32(
mm128_extr_32( s3,2 ), mm128_extr_32( s2,2 ),
mm128_extr_32( s1,2 ), mm128_extr_32( s0,2 ) );
*(__m128i*)d3 = _mm_set_epi32(
mm128_extr_32( s3,3 ), mm128_extr_32( s2,3 ),
mm128_extr_32( s1,3 ), mm128_extr_32( s0,3 ) );
}
static inline void mm128_intrlv_2x64x128( void *d, const void *s0,
const void *s1 )
{
casti_m128i( d,0 ) = mm128_put_64( s0, s1 );
casti_m128i( d,1 ) = mm128_put_64( s0+ 8, s1+ 8 );
casti_m128i( d,2 ) = mm128_put_64( s0+16, s1+16 );
casti_m128i( d,3 ) = mm128_put_64( s0+24, s1+24 );
}
#define mm128_bswap_intrlv_4x32_128( d, src ) \
do { \
__m128i ss = mm128_bswap_32( src );\
casti_m128i( d,0 ) = _mm_set1_epi32( mm128_extr_32( ss, 0 ) ); \
casti_m128i( d,1 ) = _mm_set1_epi32( mm128_extr_32( ss, 1 ) ); \
casti_m128i( d,2 ) = _mm_set1_epi32( mm128_extr_32( ss, 2 ) ); \
casti_m128i( d,3 ) = _mm_set1_epi32( mm128_extr_32( ss, 3 ) ); \
} while(0)
//
// User functions.
// interleave 4 arrays of 32 bit elements for 128 bit processing
// bit_len must be 256, 512 or 640 bits.
static inline void mm128_intrlv_4x32( void *d, const void *s0,
const void *s1, const void *s2, const void *s3, int bit_len )
{
mm128_intrlv_4x32_128( d , casti_m128i(s0,0), casti_m128i(s1,0),
casti_m128i(s2,0), casti_m128i(s3,0) );
mm128_intrlv_4x32_128( d+ 64, casti_m128i(s0,1), casti_m128i(s1,1),
casti_m128i(s2,1), casti_m128i(s3,1) );
if ( bit_len <= 256 ) return;
mm128_intrlv_4x32_128( d+128, casti_m128i(s0,2), casti_m128i(s1,2),
casti_m128i(s2,2), casti_m128i(s3,2) );
mm128_intrlv_4x32_128( d+192, casti_m128i(s0,3), casti_m128i(s1,3),
casti_m128i(s2,3), casti_m128i(s3,3) );
if ( bit_len <= 512 ) return;
mm128_intrlv_4x32_128( d+256, casti_m128i(s0,4), casti_m128i(s1,4),
casti_m128i(s2,4), casti_m128i(s3,4) );
if ( bit_len <= 640 ) return;
mm128_intrlv_4x32_128( d+320, casti_m128i(s0,5), casti_m128i(s1,5),
casti_m128i(s2,5), casti_m128i(s3,5) );
mm128_intrlv_4x32_128( d+384, casti_m128i(s0,6), casti_m128i(s1,6),
casti_m128i(s2,6), casti_m128i(s3,6) );
mm128_intrlv_4x32_128( d+448, casti_m128i(s0,7), casti_m128i(s1,7),
casti_m128i(s2,7), casti_m128i(s3,7) );
// bit_len == 1024
}
// Still used by decred due to odd data size: 180 bytes
// bit_len must be multiple of 32
static inline void mm128_intrlv_4x32x( void *dst, void *src0, void *src1,
void *src2, void *src3, int bit_len )
{
uint32_t *d = (uint32_t*)dst;
uint32_t *s0 = (uint32_t*)src0;
uint32_t *s1 = (uint32_t*)src1;
uint32_t *s2 = (uint32_t*)src2;
uint32_t *s3 = (uint32_t*)src3;
for ( int i = 0; i < bit_len >> 5; i++, d += 4 )
{
*d = *(s0+i);
*(d+1) = *(s1+i);
*(d+2) = *(s2+i);
*(d+3) = *(s3+i);
}
}
static inline void mm128_dintrlv_4x32( void *d0, void *d1, void *d2,
void *d3, const void *s, int bit_len )
{
mm128_dintrlv_4x32_128( d0 , d1 , d2 , d3 , s );
mm128_dintrlv_4x32_128( d0+ 16, d1+ 16, d2+ 16, d3+ 16, s+ 64 );
if ( bit_len <= 256 ) return;
mm128_dintrlv_4x32_128( d0+ 32, d1+ 32, d2+ 32, d3+ 32, s+128 );
mm128_dintrlv_4x32_128( d0+ 48, d1+ 48, d2+ 48, d3+ 48, s+192 );
if ( bit_len <= 512 ) return;
mm128_dintrlv_4x32_128( d0+ 64, d1+ 64, d2+ 64, d3+ 64, s+256 );
if ( bit_len <= 640 ) return;
mm128_dintrlv_4x32_128( d0+ 80, d1+ 80, d2+ 80, d3+ 80, s+320 );
mm128_dintrlv_4x32_128( d0+ 96, d1+ 96, d2+ 96, d3+ 96, s+384 );
mm128_dintrlv_4x32_128( d0+112, d1+112, d2+112, d3+112, s+448 );
// bit_len == 1024
}
// extract and deinterleave specified lane.
static inline void mm128_extr_lane_4x32( void *d, const void *s,
const int lane, const int bit_len )
{
casti_m128i( d, 0 ) =
mm128_get_32( s, lane , lane+ 4, lane+ 8, lane+12 );
casti_m128i( d, 1 ) =
mm128_get_32( s, lane+16, lane+20, lane+24, lane+28 );
if ( bit_len <= 256 ) return;
casti_m128i( d, 2 ) =
mm128_get_32( s, lane+32, lane+36, lane+40, lane+44 );
casti_m128i( d, 3 ) =
mm128_get_32( s, lane+48, lane+52, lane+56, lane+60 );
// bit_len == 512
}
// Interleave 80 bytes of 32 bit data for 4 lanes.
static inline void mm128_bswap_intrlv80_4x32( void *d, const void *s )
{
mm128_bswap_intrlv_4x32_128( d , casti_m128i( s, 0 ) );
mm128_bswap_intrlv_4x32_128( d+ 64, casti_m128i( s, 1 ) );
mm128_bswap_intrlv_4x32_128( d+128, casti_m128i( s, 2 ) );
mm128_bswap_intrlv_4x32_128( d+192, casti_m128i( s, 3 ) );
mm128_bswap_intrlv_4x32_128( d+256, casti_m128i( s, 4 ) );
}
#endif // SSE2
#endif // INTRLV_SSE2_H__

1333
simd-utils/intrlv.h Normal file
View File

File diff suppressed because it is too large Load Diff

28
simd-utils/simd-128.h
View File

@@ -44,7 +44,17 @@
// repeatedly. It may be better for the application to reimplement the
// utility to better suit its usage.
//
// More tips:
//
// Conversions from integer to vector should be avoided whenever possible.
// Extract, insert and set and set1 instructions should be avoided.
// In addition to the issues with constants set is also very inefficient with
// variables.
// Converting integer data to perform a couple of vector operations
// then converting back to integer should be avoided. Converting data in
// registers should also be avoided. Conversion should be limited to buffers
// in memory where the data is loaded directly to vector registers, bypassing
// the integer to vector conversion.
//
// Pseudo constants.
//
@@ -71,7 +81,7 @@ static inline __m128i m128_one_64_fn()
asm( "pxor %0, %0\n\t"
"pcmpeqd %%xmm1, %%xmm1\n\t"
"psubq %%xmm1, %0\n\t"
:"=x"(a)
: "=x"(a)
:
: "xmm1" );
return a;
@@ -84,7 +94,7 @@ static inline __m128i m128_one_32_fn()
asm( "pxor %0, %0\n\t"
"pcmpeqd %%xmm1, %%xmm1\n\t"
"psubd %%xmm1, %0\n\t"
:"=x"(a)
: "=x"(a)
:
: "xmm1" );
return a;
@@ -97,7 +107,7 @@ static inline __m128i m128_one_16_fn()
asm( "pxor %0, %0\n\t"
"pcmpeqd %%xmm1, %%xmm1\n\t"
"psubw %%xmm1, %0\n\t"
:"=x"(a)
: "=x"(a)
:
: "xmm1" );
return a;
@@ -110,7 +120,7 @@ static inline __m128i m128_one_8_fn()
asm( "pxor %0, %0\n\t"
"pcmpeqd %%xmm1, %%xmm1\n\t"
"psubb %%xmm1, %0\n\t"
:"=x"(a)
: "=x"(a)
:
: "xmm1" );
return a;
@@ -121,7 +131,7 @@ static inline __m128i m128_neg1_fn()
{
__m128i a;
asm( "pcmpeqd %0, %0\n\t"
:"=x"(a) );
: "=x"(a) );
return a;
}
#define m128_neg1 m128_neg1_fn()
@@ -133,7 +143,7 @@ static inline __m128i m128_one_128_fn()
__m128i a;
asm( "pinsrq $0, $1, %0\n\t"
"pinsrq $1, $0, %0\n\t"
:"=x"(a) );
: "=x"(a) );
return a;
}
#define m128_one_128 m128_one_128_fn()
@@ -145,8 +155,8 @@ static inline __m128i m128_const_64( uint64_t hi, uint64_t lo )
__m128i a;
asm( "pinsrq $0, %2, %0\n\t"
"pinsrq $1, %1, %0\n\t"
:"=x"(a)
:"r"(hi),"r"(lo) );
: "=x"(a)
: "r"(hi), "r"(lo) );
return a;
}

263
simd-utils/simd-512.h
View File

@@ -41,79 +41,6 @@
// Experimental, not fully tested.
//
// Compile time vector constants and initializers.
//
// The following macro constants and functions should only be used
// for compile time initialization of constant and variable vector
// arrays. These constants use memory, use set instruction or pseudo
// constants at run time to avoid using memory.
// Constant initializers
#define mm512_const_64( x7, x6, x5, x4, x3, x2, x1, x0 ) \
{{ x7, x6, x5, x4, x3, x2, x1, x0 }}
#define mm512_const1_64( x ) {{ x,x,x,x,x,x,x }}
#define mm512_const_32( x15, x14, x13, x12, x11, x10, x09, x08, \
x07, x06, x05, x04, x03, x02, x01, x00 ) \
{{ x15, x14, x13, x12, x11, x10, x09, x08, }} \
x07, x06, x05, x04, x03, x02, x01, x00 }}
#define mm512_const1_32( x ) {{ x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x }}
#define mm512_const_16( x31, x30, x29, x28, x27, x26, x25, x24, \
x23, x22, x21, x20, x19, x18, x17, x16, \
x15, x14, x13, x12, x11, x10, x09, x08, \
x07, x06, x05, x04, x03, x02, x01, x00 ) \
{{ x31, x30, x29, x28, x27, x26, x25, x24, \
x23, x22, x21, x20, x19, x18, x17, x16, \
x15, x14, x13, x12, x11, x10, x09, x08, \
x07, x06, x05, x04, x03, x02, x01, x00 }}
#define mm512_const1_16( x ) {{ x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x, \
x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x }}
#define mm512_const_8( x63, x62, x61, x60, x59, x58, x57, x56, \
x55, x54, x53, x52, x51, x50, x49, x48, \
x47, x46, x45, x44, x43, x42, x41, x40, \
x39, x38, x37, x36, x35, x34, x33, x32, \
x31, x30, x29, x28, x27, x26, x25, x24, \
x23, x22, x21, x20, x19, x18, x17, x16, \
x15, x14, x13, x12, x11, x10, x09, x08, \
x07, x06, x05, x04, x03, x02, x01, x00 ) \
{{ x63, x62, x61, x60, x59, x58, x57, x56, \
x55, x54, x53, x52, x51, x50, x49, x48, \
x47, x46, x45, x44, x43, x42, x41, x40, \
x39, x38, x37, x36, x35, x34, x33, x32, \
x31, x30, x29, x28, x27, x26, x25, x24, \
x23, x22, x21, x20, x19, x18, x17, x16, \
x15, x14, x13, x12, x11, x10, x09, x08, \
x07, x06, x05, x04, x03, x02, x01, x00 }}
#define mm512_const1_8( x ) {{ x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x, \
x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x, \
x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x, \
x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x }}
// Predefined compile time constant vectors.
#define c512_zero mm512_const1_64( 0ULL )
#define c512_neg1 mm512_const1_64( 0xFFFFFFFFFFFFFFFFULL )
#define c512_one_512 mm512_const_epi64( 0ULL, 0ULL, 0ULL, 0ULL, \
0ULL, 0ULL, 0ULL, 1ULL )
#define c512_one_256 mm512_const_64( 0ULL, 0ULL, 0ULL, 1ULL, \
0ULL, 0ULL, 0ULL, 1ULL )
#define c512_one_128 mm512_const_64( 0ULL, 1ULL, 0ULL, 1ULL, \
0ULL, 1ULL, 0ULL, 1ULL )
#define c512_one_64 mm512_const1_64( 1ULL )
#define c512_one_32 mm512_const1_32( 1UL )
#define c512_one_16 mm512_const1_16( 1U )
#define c512_one_8 mm512_const1_8( 1U )
#define c512_neg1_64 mm512_const1_64( 0xFFFFFFFFFFFFFFFFULL )
#define c512_neg1_32 mm512_const1_32( 0xFFFFFFFFUL )
#define c512_neg1_16 mm512_const1_32( 0xFFFFU )
#define c512_neg1_8 mm512_const1_32( 0xFFU )
//
// Pseudo constants.
@@ -127,11 +54,77 @@
0ULL, 0ULL, 0ULL, 1ULL )
#define m512_one_256 _mm512_set4_epi64( 0ULL, 0ULL, 0ULL, 1ULL )
#define m512_one_128 _mm512_set4_epi64( 0ULL, 1ULL, 0ULL, 1ULL )
#define m512_one_64 _mm512_set1_epi64( 1ULL )
#define m512_one_32 _mm512_set1_epi32( 1UL )
#define m512_one_16 _mm512_set1_epi16( 1U )
#define m512_one_8 _mm512_set1_epi8( 1U )
#define m512_neg1 _mm512_set1_epi64( 0xFFFFFFFFFFFFFFFFULL )
//#define m512_one_64 _mm512_set1_epi64( 1ULL )
//#define m512_one_32 _mm512_set1_epi32( 1UL )
//#define m512_one_16 _mm512_set1_epi16( 1U )
//#define m512_one_8 _mm512_set1_epi8( 1U )
//#define m512_neg1 _mm512_set1_epi64( 0xFFFFFFFFFFFFFFFFULL )
#define mi512_const_64( i7, i6, i5, i4, i3, i2, i1, i0 ) \
_mm512_inserti64x4( _mm512_castsi512_si256( m256_const_64( i3.i2,i1,i0 ) ), \
m256_const_64( i7,i6,i5,i4 ), 1 )
#define m512_const1_64( i ) m256_const_64( i, i, i, i, i, i, i, i )
static inline __m512i m512_one_64_fn()
{
__m512i a;
asm( "vpxorq %0, %0, %0\n\t"
"vpcmpeqd %%zmm1, %%zmm1, %%zmm1\n\t"
"vpsubq %%zmm1, %0, %0\n\t"
:"=x"(a)
:
: "zmm1" );
return a;
}
#define m512_one_64 m512_one_64_fn()
static inline __m512i m512_one_32_fn()
{
__m512i a;
asm( "vpxord %0, %0, %0\n\t"
"vpcmpeqd %%zmm1, %%zmm1, %%zmm1\n\t"
"vpsubd %%zmm1, %0, %0\n\t"
:"=x"(a)
:
: "zmm1" );
return a;
}
#define m512_one_32 m512_one_32_fn()
static inline __m512i m512_one_16_fn()
{
__m512i a;
asm( "vpxord %0, %0, %0\n\t"
"vpcmpeqd %%zmm1, %%zmm1, %%zmm1\n\t"
"vpsubw %%zmm1, %0, %0\n\t"
:"=x"(a)
:
: "zmm1" );
return a;
}
#define m512_one_16 m512_one_16_fn()
static inline __m512i m512_one_8_fn()
{
__m512i a;
asm( "vpxord %0, %0, %0\n\t"
"vpcmpeqd %%zmm1, %%zmm1, %%zmm1\n\t"
"vpsubb %%zmm1, %0, %0\n\t"
:"=x"(a)
:
: "zmm1" );
return a;
}
#define m512_one_8 m512_one_8_fn()
static inline __m512i m512_neg1_fn()
{
__m512i a;
asm( "vpcmpeqq %0, %0, %0\n\t"
:"=x"(a) );
return a;
}
#define m512_neg1 m512_neg1_fn()
//
@@ -142,6 +135,15 @@
#define mm512_negate_32( x ) _mm512_sub_epi32( m512_zero, x )
#define mm512_negate_16( x ) _mm512_sub_epi16( m512_zero, x )
#define mm256_extr_lo256_512( a ) _mm512_castsi512_si256( a )
#define mm256_extr_hi256_512( a ) _mm512_extracti64x4_epi64( a, 1 )
#define mm128_extr_lo128_512( a ) _mm512_castsi512_si256( a )
//
// Pointer casting
@@ -225,6 +227,25 @@
*((uint32_t*)(d06)) = ((uint32_t*)(s))[14]; \
*((uint32_t*)(d07)) = ((uint32_t*)(s))[15];
// Add 4 values, fewer dependencies than sequential addition.
#define mm512_add4_64( a, b, c, d ) \
_mm512_add_epi64( _mm512_add_epi64( a, b ), _mm512_add_epi64( c, d ) )
#define mm512_add4_32( a, b, c, d ) \
_mm512_add_epi32( _mm512_add_epi32( a, b ), _mm512_add_epi32( c, d ) )
#define mm512_add4_16( a, b, c, d ) \
_mm512_add_epi16( _mm512_add_epi16( a, b ), _mm512_add_epi16( c, d ) )
#define mm512_add4_8( a, b, c, d ) \
_mm512_add_epi8( _mm512_add_epi8( a, b ), _mm512_add_epi8( c, d ) )
#define mm512_xor4( a, b, c, d ) \
_mm512_xor_si512( _mm512_xor_si256( a, b ), _mm512_xor_si256( c, d ) )
//
// Bit rotations.
@@ -263,45 +284,41 @@
#define mm512_ror_x32( v, n ) _mm512_alignr_epi32( v, v, n )
// Although documented to exist in AVX512F the _mm512_set_epi8 &
// _mm512_set_epi16 intrinsics fail to compile. Seems usefull to have
// for endian byte swapping. Workaround by using _mm512_set_epi32.
// Ugly but it works.
#define mm512_ror_1x16( v ) \
_mm512_permutexvar_epi16( v, _mm512_set_epi32( \
0x0000001F, 0x001E001D, 0x001C001B, 0x001A0019, \
0X00180017, 0X00160015, 0X00140013, 0X00120011, \
0X0010000F, 0X000E000D, 0X000C000B, 0X000A0009, \
0X00080007, 0X00060005, 0X00040003, 0X00020001 ) )
_mm512_permutexvar_epi16( v, m512_const_64( \
0x0000001F001E001D, 0x001C001B001A0019, \
0X0018001700160015, 0X0014001300120011, \
0X0010000F000E000D, 0X000C000B000A0009, \
0X0008000700060005, 0X0004000300020001 ) )
#define mm512_rol_1x16( v ) \
_mm512_permutexvar_epi16( v, _mm512_set_epi32( \
0x001E001D, 0x001C001B, 0x001A0019, 0x00180017, \
0X00160015, 0X00140013, 0X00120011, 0x0010000F, \
0X000E000D, 0X000C000B, 0X000A0009, 0X00080007, \
0X00060005, 0X00040003, 0X00020001, 0x0000001F ) )
_mm512_permutexvar_epi16( v, m512_const_64( \
0x001E001D001C001B, 0x001A001900180017, \
0X0016001500140013, 0X001200110010000F, \
0X000E000D000C000B, 0X000A000900080007, \
0X0006000500040003, 0X000200010000001F ) )
#define mm512_ror_1x8( v ) \
_mm512_permutexvar_epi8( v, _mm512_set_epi32( \
0x003F3E3D, 0x3C3B3A39, 0x38373635, 0x34333231, \
0x302F2E2D, 0x2C2B2A29, 0x28272625, 0x24232221, \
0x201F1E1D, 0x1C1B1A19. 0x18171615, 0x14131211, \
0x100F0E0D, 0x0C0B0A09, 0x08070605, 0x04030201 ) )
_mm512_permutexvar_epi8( v, m512_const_64( \
0x003F3E3D3C3B3A39, 0x3837363534333231, \
0x302F2E2D2C2B2A29, 0x2827262524232221, \
0x201F1E1D1C1B1A19. 0x1817161514131211, \
0x100F0E0D0C0B0A09, 0x0807060504030201 ) )
#define mm512_rol_1x8( v ) \
_mm512_permutexvar_epi8( v, _mm512_set_epi32( \
0x3E3D3C3B, 0x3A393837, 0x36353433, 0x3231302F. \
0x2E2D2C2B, 0x2A292827, 0x26252423, 0x2221201F, \
0x1E1D1C1B, 0x1A191817, 0x16151413, 0x1211100F, \
0x0E0D0C0B, 0x0A090807, 0x06050403, 0x0201003F ) )
_mm512_permutexvar_epi8( v, m512_const_64( \
0x3E3D3C3B3A393837, 0x363534333231302F. \
0x2E2D2C2B2A292827, 0x262524232221201F, \
0x1E1D1C1B1A191817, 0x161514131211100F, \
0x0E0D0C0B0A090807, 0x060504030201003F ) )
// Invert vector: {3,2,1,0} -> {0,1,2,3}
#define mm512_invert_128( v ) _mm512_permute4f128_epi32( a, 0x1b )
#define mm512_invert_64( v ) \
_mm512_permutex_epi64( v, _mm512_set_epi64( 0,1,2,3,4,5,6,7 ) )
_mm512_permutex_epi64( v, m512_const_64( 0,1,2,3,4,5,6,7 ) )
#define mm512_invert_32( v ) \
_mm512_permutexvar_epi32( v, _mm512_set_epi32( \
@@ -378,32 +395,32 @@
#define mm512_rol1x32_128( v ) _mm512_shuffle_epi32( v, 0x93 )
#define mm512_ror1x16_128( v ) \
_mm512_permutexvar_epi16( v, _mm512_set_epi32( \
0x0018001F, 0x001E001D, 0x001C001B, 0x001A0019, \
0x00100017, 0x00160015, 0x00140013, 0x00120011, \
0x0008000F, 0x000E000D, 0x000C000B, 0x000A0009, \
0x00000007, 0x00060005, 0x00040003, 0x00020001 ) )
_mm512_permutexvar_epi16( v, m512_const_64( \
0x0018001F001E001D, 0x001C001B001A0019, \
0x0010001700160015, 0x0014001300120011, \
0x0008000F000E000D, 0x000C000B000A0009, \
0x0000000700060005, 0x0004000300020001 ) )
#define mm512_rol1x16_128( v ) \
_mm512_permutexvar_epi16( v, _mm512_set_epi32( \
0x001E001D, 0x001C001B, 0x001A0019, 0x0018001F, \
0x00160015, 0x00140013, 0x00120011, 0x00100017, \
0x000E000D, 0x000C000B, 0x000A0009, 0x0008000F, \
0x00060005, 0x00040003, 0x00020001, 0x00000007 ) )
_mm512_permutexvar_epi16( v, m512_const_64( \
0x001E001D001C001B, 0x001A00190018001F, \
0x0016001500140013, 0x0012001100100017, \
0x000E000D000C000B, 0x000A00090008000F, \
0x0006000500040003, 0x0002000100000007 ) )
#define mm512_ror1x8_128( v ) \
_mm512_permutexvar_epi8( v, _mm512_set_epi32( \
0x303F3E3D, 0x3C3B3A39, 0x38373635, 0x34333231, \
0x202F2E2D, 0x2C2B2A29, 0x28272625, 0x24232221, \
0x101F1E1D, 0x1C1B1A19, 0x18171615, 0x14131211, \
0x000F0E0D, 0x0C0B0A09, 0x08070605, 0x04030201 ) )
_mm512_permutexvar_epi8( v, m512_const_64( \
0x303F3E3D3C3B3A39, 0x3837363534333231, \
0x202F2E2D2C2B2A29, 0x2827262524232221, \
0x101F1E1D1C1B1A19, 0x1817161514131211, \
0x000F0E0D0C0B0A09, 0x0807060504030201 ) )
#define mm512_rol1x8_128( v ) \
_mm512_permutexvar_epi8( v, _mm512_set_epi32( \
0x3E3D3C3B, 0x3A393837, 0x36353433. 0x3231303F, \
0x2E2D2C2B, 0x2A292827, 0x26252423, 0x2221202F, \
0x1E1D1C1B, 0x1A191817, 0x16151413, 0x1211101F, \
0x0E0D0C0B, 0x0A090807, 0x06050403, 0x0201000F ) )
_mm512_permutexvar_epi8( v, m512_const_64( \
0x3E3D3C3B3A393837, 0x363534333231303F, \
0x2E2D2C2B2A292827, 0x262524232221202F, \
0x1E1D1C1B1A191817, 0x161514131211101F, \
0x0E0D0C0B0A090807, 0x060504030201000F ) )
// Rotate 128 bit lanes by c bytes.
#define mm512_bror_128( v, c ) \