OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: 668f48f4fc
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '668f48f4fc'
${ noResults }
OpenBLAS/kernel/simd/intrin_avx512.h

60 lines
2.1 kB
Raw Blame History 

  1. #define V_SIMD 512

  2. #define V_SIMD_F64 1

  3. /***************************

  4.  * Data Type

  5.  ***************************/

  6. typedef __m512  v_f32;

  7. typedef __m512d  v_f64;

  8. #define v_nlanes_f32 16

  9. #define v_nlanes_f64 8

  10. /***************************

  11.  * Arithmetic

  12.  ***************************/

  13. #define v_add_f32 _mm512_add_ps

  14. #define v_add_f64 _mm512_add_pd

  15. #define v_sub_f32 _mm512_sub_ps

  16. #define v_sub_f64 _mm512_sub_pd

  17. #define v_mul_f32 _mm512_mul_ps

  18. #define v_mul_f64 _mm512_mul_pd

  19. // multiply and add, a*b + c

  20. #define v_muladd_f32 _mm512_fmadd_ps

  21. #define v_muladd_f64 _mm512_fmadd_pd

  22. // multiply and subtract, a*b - c

  23. #define v_mulsub_f32 _mm512_fmsub_ps

  24. #define v_mulsub_f64 _mm512_fmsub_pd

  25. BLAS_FINLINE float v_sum_f32(v_f32 a)

  26. {

  27.     __m512 h64 = _mm512_shuffle_f32x4(a, a, _MM_SHUFFLE(3, 2, 3, 2));

  28.     __m512 sum32 = _mm512_add_ps(a, h64);

  29.     __m512 h32 = _mm512_shuffle_f32x4(sum32, sum32, _MM_SHUFFLE(1, 0, 3, 2));

  30.     __m512 sum16 = _mm512_add_ps(sum32, h32);

  31.     __m512 h16 = _mm512_permute_ps(sum16, _MM_SHUFFLE(1, 0, 3, 2));

  32.     __m512 sum8 = _mm512_add_ps(sum16, h16);

  33.     __m512 h4 = _mm512_permute_ps(sum8, _MM_SHUFFLE(2, 3, 0, 1));

  34.     __m512 sum4 = _mm512_add_ps(sum8, h4);

  35.     return _mm_cvtss_f32(_mm512_castps512_ps128(sum4));

  36. }

  37. 

  38. BLAS_FINLINE double v_sum_f64(v_f64 a)

  39. {

  40.     __m512d h64   = _mm512_shuffle_f64x2(a, a, _MM_SHUFFLE(3, 2, 3, 2));

  41.     __m512d sum32 = _mm512_add_pd(a, h64);

  42.     __m512d h32   = _mm512_permutex_pd(sum32, _MM_SHUFFLE(1, 0, 3, 2));

  43.     __m512d sum16 = _mm512_add_pd(sum32, h32);

  44.     __m512d h16   = _mm512_permute_pd(sum16, _MM_SHUFFLE(2, 3, 0, 1));

  45.     __m512d sum8  = _mm512_add_pd(sum16, h16);

  46.     return _mm_cvtsd_f64(_mm512_castpd512_pd128(sum8));

  47. }

  48. /***************************

  49.  * memory

  50.  ***************************/

  51. // unaligned load

  52. #define v_loadu_f32(PTR) _mm512_loadu_ps((const __m512*)(PTR))

  53. #define v_loadu_f64(PTR) _mm512_loadu_pd((const __m512*)(PTR))

  54. #define v_storeu_f32 _mm512_storeu_ps

  55. #define v_storeu_f64 _mm512_storeu_pd

  56. #define v_setall_f32(VAL) _mm512_set1_ps(VAL)

  57. #define v_setall_f64(VAL) _mm512_set1_pd(VAL)

  58. #define v_zero_f32 _mm512_setzero_ps

  59. #define v_zero_f64 _mm512_setzero_pd