LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE
void stream_vector_copy( const double* i_a,
double* io_c,
const int i_length) {
int l_n = 0;
int l_trip_prolog = 0;
int l_trip_stream = 0;
/* init the trip counts */
stream_init( i_length, (size_t)io_c, &l_trip_prolog, &l_trip_stream );
/* run the prologue */
for ( ; l_n < l_trip_prolog; l_n++ ) {
io_c[l_n] = i_a[l_n];
}
/* run the bulk, hopefully using streaming stores */
#if defined(__SSE3__) && defined(__AVX__) && !defined(__AVX512F__)
{
/* we need manual unrolling as the compiler otherwise generates
too many dependencies */
for ( ; l_n < l_trip_stream; l_n+=8 ) {
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n]), _mm256_loadu_pd(&(i_a[l_n])) );
_mm256_store_pd( &(io_c[l_n+4]), _mm256_loadu_pd(&(i_a[l_n+4])) );
#else
_mm256_stream_pd( &(io_c[l_n]), _mm256_loadu_pd(&(i_a[l_n])) );
_mm256_stream_pd( &(io_c[l_n+4]), _mm256_loadu_pd(&(i_a[l_n+4])) );
#endif
}
}
#elif defined(__SSE3__) && defined(__AVX__) && defined(__AVX512F__)
{
for ( ; l_n < l_trip_stream; l_n+=8 ) {
#ifdef DISABLE_NONTEMPORAL_STORES
_mm512_store_pd( &(io_c[l_n]), _mm512_loadu_pd(&(i_a[l_n])) );
#else
_mm512_stream_pd( &(io_c[l_n]), _mm512_loadu_pd(&(i_a[l_n])) );
#endif
}
}
#else
for ( ; l_n < l_trip_stream; l_n++ ) {
io_c[l_n] = i_a[l_n];
}
#endif
/* run the epilogue */
for ( ; l_n < i_length; l_n++ ) {
io_c[l_n] = i_a[l_n];
}
}
int main(int, char**argv)
{
/* AVX512 Foundation */
__m512i i;
__m512d d;
__m512 f;
__mmask16 m = ~1;
i = _mm512_maskz_loadu_epi32(0, argv);
d = _mm512_loadu_pd((double *)argv + 64);
f = _mm512_loadu_ps((float *)argv + 128);
#ifdef __AVX512ER__
/* AVX512 Exponential and Reciprocal */
f = _mm512_exp2a23_round_ps(f, 8);
#endif
#ifdef __AVX512CD__
/* AVX512 Conflict Detection */
i = _mm512_maskz_conflict_epi32(m, i);
#endif
#ifdef __AVX512PF__
/* AVX512 Prefetch */
_mm512_mask_prefetch_i64scatter_pd(argv, 0xf, i, 2, 2);
#endif
#ifdef __AVX512DQ__
/* AVX512 Doubleword and Quadword support */
m = _mm512_movepi32_mask(i);
#endif
#ifdef __AVX512BW__
/* AVX512 Byte and Word support */
i = _mm512_mask_loadu_epi8(i, m, argv - 8);
#endif
#ifdef __AVX512VL__
/* AVX512 Vector Length */
__m256i i2 = _mm256_maskz_loadu_epi32(0, argv);
_mm256_mask_storeu_epi32(argv + 1, m, i2);
#endif
#ifdef __AVX512IFMA__
/* AVX512 Integer Fused Multiply-Add */
i = _mm512_madd52lo_epu64(i, i, i);
#endif
#ifdef __AVX512VBMI__
/* AVX512 Vector Byte Manipulation Instructions */
i = _mm512_permutexvar_epi8(i, i);
#endif
_mm512_mask_storeu_epi64(argv, m, i);
_mm512_mask_storeu_ps(argv + 64, m, f);
_mm512_mask_storeu_pd(argv + 128, m, d);
return 0;
}
KFR_INTRINSIC f64avx512 read(cunaligned_t, csize_t<8>, const f64* ptr) { return _mm512_loadu_pd(ptr); }
LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE
void stream_update_helmholtz_no_h2( const double* i_g1,
const double* i_g2,
const double* i_g3,
const double* i_tm1,
const double* i_tm2,
const double* i_tm3,
double* io_c,
const double i_h1,
const int i_length) {
int l_n = 0;
int l_trip_prolog = 0;
int l_trip_stream = 0;
/* init the trip counts */
stream_init( i_length, (size_t)io_c, &l_trip_prolog, &l_trip_stream );
/* run the prologue */
for ( ; l_n < l_trip_prolog; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
/* run the bulk, hopefully using streaming stores */
#if defined(__SSE3__) && defined(__AVX__) && !defined(__AVX512F__)
{
const __m256d vec_h1 = _mm256_broadcast_sd(&i_h1);
/* we need manual unrolling as the compiler otherwise generates
too many dependencies */
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m256d vec_g1_1, vec_g2_1, vec_g3_1, vec_tm1_1, vec_tm2_1, vec_tm3_1;
__m256d vec_g1_2, vec_g2_2, vec_g3_2, vec_tm1_2, vec_tm2_2, vec_tm3_2;
vec_g1_1 = _mm256_loadu_pd(&(i_g1[l_n]));
vec_tm1_1 = _mm256_loadu_pd(&(i_tm1[l_n]));
vec_g1_2 = _mm256_loadu_pd(&(i_g1[l_n+4]));
vec_tm1_2 = _mm256_loadu_pd(&(i_tm1[l_n+4]));
vec_g1_1 = _mm256_mul_pd(vec_g1_1, vec_tm1_1);
vec_g2_1 = _mm256_loadu_pd(&(i_g2[l_n]));
vec_g1_2 = _mm256_mul_pd(vec_g1_2, vec_tm1_2);
vec_g2_2 = _mm256_loadu_pd(&(i_g2[l_n+4]));
vec_tm2_1 = _mm256_loadu_pd(&(i_tm2[l_n]));
vec_g2_1 = _mm256_mul_pd(vec_g2_1, vec_tm2_1);
vec_tm2_2 = _mm256_loadu_pd(&(i_tm2[l_n+4]));
vec_g2_2 = _mm256_mul_pd(vec_g2_2, vec_tm2_2);
vec_g3_1 = _mm256_loadu_pd(&(i_g3[l_n]));
vec_tm3_1 = _mm256_loadu_pd(&(i_tm3[l_n]));
vec_g3_2 = _mm256_loadu_pd(&(i_g3[l_n+4]));
vec_tm3_2 = _mm256_loadu_pd(&(i_tm3[l_n+4]));
vec_g3_1 = _mm256_mul_pd(vec_g3_1, vec_tm3_1);
vec_g3_2 = _mm256_mul_pd(vec_g3_2, vec_tm3_2);
vec_g1_1 = _mm256_add_pd(vec_g1_1, vec_g2_1);
vec_g1_2 = _mm256_add_pd(vec_g1_2, vec_g2_2);
vec_g1_1 = _mm256_add_pd(vec_g1_1, vec_g3_1);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n]), _mm256_mul_pd(vec_g1_1, vec_h1) );
#else
_mm256_stream_pd( &(io_c[l_n]), _mm256_mul_pd(vec_g1_1, vec_h1) );
#endif
vec_g1_2 = _mm256_add_pd(vec_g1_2, vec_g3_2);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n+4]), _mm256_mul_pd(vec_g1_2, vec_h1) );
#else
_mm256_stream_pd( &(io_c[l_n+4]), _mm256_mul_pd(vec_g1_2, vec_h1) );
#endif
}
}
#elif defined(__SSE3__) && defined(__AVX__) && defined(__AVX512F__)
{
const __m512d vec_h1 = _mm512_broadcastsd_pd(_mm_load_sd(&i_h1));
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m512d vec_g1, vec_g2, vec_g3, vec_tm1, vec_tm2, vec_tm3;
vec_g1 = _mm512_loadu_pd(&(i_g1[l_n]));
vec_tm1 = _mm512_loadu_pd(&(i_tm1[l_n]));
vec_g1 = _mm512_mul_pd(vec_g1, vec_tm1);
vec_g2 = _mm512_loadu_pd(&(i_g2[l_n]));
vec_tm2 = _mm512_loadu_pd(&(i_tm2[l_n]));
vec_g2 = _mm512_mul_pd(vec_g2, vec_tm2);
vec_g3 = _mm512_loadu_pd(&(i_g3[l_n]));
vec_tm3 = _mm512_loadu_pd(&(i_tm3[l_n]));
vec_g3 = _mm512_mul_pd(vec_g3, vec_tm3);
vec_g1 = _mm512_add_pd(vec_g1, vec_g2);
vec_g1 = _mm512_add_pd(vec_g1, vec_g3);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm512_store_pd( &(io_c[l_n]), _mm512_mul_pd(vec_g1, vec_h1) );
#else
_mm512_stream_pd( &(io_c[l_n]), _mm512_mul_pd(vec_g1, vec_h1) );
#endif
}
}
#else
for ( ; l_n < l_trip_stream; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
#endif
/* run the epilogue */
for ( ; l_n < i_length; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
}
inline
void load(const double *data)
{
val = _mm512_loadu_pd(data);
}
