threefry_avx2_4x32.hpp File Reference

#include <mckl/random/internal/common.hpp>
#include <mckl/random/internal/threefry_common.hpp>
#include <mckl/random/internal/threefry_constants.hpp>
#include <mckl/random/internal/threefry_generic_4x32.hpp>

Go to the source code of this file.

Classes
class	mckl::internal::Threefry4x32GeneratorAVX2Impl< T >

Namespaces
	mckl
	mckl::internal

Macros
#define	MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_4(N)
#define	MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_8(N)
#define	MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX(N)
#define	MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N)

Macro Definition Documentation

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_4

#define MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_4

(

N

)

Value:

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 1);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 2);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 3);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 4);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX(N * 8 + 4);

Definition at line 88 of file threefry_avx2_4x32.hpp.

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_8

#define MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_CYCLE_8

(

N

)

Value:

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 1);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 2);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 3);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 4);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX(N * 8 + 4);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 5);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 6);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 7);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX(N * 8 + 8);                  \
    MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX(N * 8 + 8);

Definition at line 95 of file threefry_avx2_4x32.hpp.

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX

#define MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_KBOX

(

N

)

Value:

ymmt0 = _mm256_set1_epi32(                                                \
        static_cast<int>(ThreefryKBox<T, K, N>::template key<0>(par)));       \
    ymmt1 = _mm256_set1_epi32(                                                \
        static_cast<int>(ThreefryKBox<T, K, N>::template key<1>(par)));       \
    ymmt2 = _mm256_set1_epi32(                                                \
        static_cast<int>(ThreefryKBox<T, K, N>::template key<2>(par)));       \
    ymmt3 = _mm256_set1_epi32(                                                \
        static_cast<int>(ThreefryKBox<T, K, N>::template key<3>(par)));       \
    ymms0 = _mm256_add_epi32(ymms0, ymmt0);                                   \
    ymms1 = _mm256_add_epi32(ymms1, ymmt1);                                   \
    ymms2 = _mm256_add_epi32(ymms2, ymmt2);                                   \
    ymms3 = _mm256_add_epi32(ymms3, ymmt3);                                   \
    ymms4 = _mm256_add_epi32(ymms4, ymmt0);                                   \
    ymms5 = _mm256_add_epi32(ymms5, ymmt1);                                   \
    ymms6 = _mm256_add_epi32(ymms6, ymmt2);                                   \
    ymms7 = _mm256_add_epi32(ymms7, ymmt3);

Definition at line 42 of file threefry_avx2_4x32.hpp.

MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX

#define MCKL_RANDOM_INTERNAL_THREEFRY_AVX2_4X32_RBOX

(

N

)

Value:

{                                                                         \
        constexpr int L0 = Constants::rotate::value[0][(N - 1) % 8];          \
        constexpr int L1 = Constants::rotate::value[1][(N - 1) % 8];          \
        constexpr int R0 = 32 - L0;                                           \
        constexpr int R1 = 32 - L1;                                           \
        ymms0 = _mm256_add_epi32(ymms0, ymms1);                               \
        ymms2 = _mm256_add_epi32(ymms2, ymms3);                               \
        ymms4 = _mm256_add_epi32(ymms4, ymms5);                               \
        ymms6 = _mm256_add_epi32(ymms6, ymms7);                               \
        ymmt1 = _mm256_slli_epi32(ymms1, L0);                                 \
        ymmt3 = _mm256_slli_epi32(ymms3, L1);                                 \
        ymmt5 = _mm256_slli_epi32(ymms5, L0);                                 \
        ymmt7 = _mm256_slli_epi32(ymms7, L1);                                 \
        ymms1 = _mm256_srli_epi32(ymms1, R0);                                 \
        ymms3 = _mm256_srli_epi32(ymms3, R1);                                 \
        ymms5 = _mm256_srli_epi32(ymms5, R0);                                 \
        ymms7 = _mm256_srli_epi32(ymms7, R1);                                 \
        ymmt1 = _mm256_or_si256(ymms1, ymmt1);                                \
        ymmt3 = _mm256_or_si256(ymms3, ymmt3);                                \
        ymmt5 = _mm256_or_si256(ymms5, ymmt5);                                \
        ymmt7 = _mm256_or_si256(ymms7, ymmt7);                                \
        ymms1 = _mm256_xor_si256(ymms2, ymmt3);                               \
        ymms3 = _mm256_xor_si256(ymms0, ymmt1);                               \
        ymms5 = _mm256_xor_si256(ymms6, ymmt7);                               \
        ymms7 = _mm256_xor_si256(ymms4, ymmt5);                               \
    }

Definition at line 60 of file threefry_avx2_4x32.hpp.