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.

intrin_avx.h 2.3 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970
  1. #define V_SIMD 256
  2. #define V_SIMD_F64 1
  3. /***************************
  4. * Data Type
  5. ***************************/
  6. typedef __m256 v_f32;
  7. typedef __m256d v_f64;
  8. #define v_nlanes_f32 8
  9. #define v_nlanes_f64 4
  10. /***************************
  11. * Arithmetic
  12. ***************************/
  13. #define v_add_f32 _mm256_add_ps
  14. #define v_add_f64 _mm256_add_pd
  15. #define v_sub_f32 _mm256_sub_ps
  16. #define v_sub_f64 _mm256_sub_pd
  17. #define v_mul_f32 _mm256_mul_ps
  18. #define v_mul_f64 _mm256_mul_pd
  19. #ifdef HAVE_FMA3
  20. // multiply and add, a*b + c
  21. #define v_muladd_f32 _mm256_fmadd_ps
  22. #define v_muladd_f64 _mm256_fmadd_pd
  23. // multiply and subtract, a*b - c
  24. #define v_mulsub_f32 _mm256_fmsub_ps
  25. #define v_mulsub_f64 _mm256_fmsub_pd
  26. #else
  27. // multiply and add, a*b + c
  28. BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
  29. { return v_add_f32(v_mul_f32(a, b), c); }
  30. BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
  31. { return v_add_f64(v_mul_f64(a, b), c); }
  32. // multiply and subtract, a*b - c
  33. BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
  34. { return v_sub_f32(v_mul_f32(a, b), c); }
  35. BLAS_FINLINE v_f64 v_mulsub_f64(v_f64 a, v_f64 b, v_f64 c)
  36. { return v_sub_f64(v_mul_f64(a, b), c); }
  37. #endif // !HAVE_FMA3
  38. // Horizontal add: Calculates the sum of all vector elements.
  39. BLAS_FINLINE float v_sum_f32(__m256 a)
  40. {
  41. __m256 sum_halves = _mm256_hadd_ps(a, a);
  42. sum_halves = _mm256_hadd_ps(sum_halves, sum_halves);
  43. __m128 lo = _mm256_castps256_ps128(sum_halves);
  44. __m128 hi = _mm256_extractf128_ps(sum_halves, 1);
  45. __m128 sum = _mm_add_ps(lo, hi);
  46. return _mm_cvtss_f32(sum);
  47. }
  48. BLAS_FINLINE double v_sum_f64(__m256d a)
  49. {
  50. __m256d sum_halves = _mm256_hadd_pd(a, a);
  51. __m128d lo = _mm256_castpd256_pd128(sum_halves);
  52. __m128d hi = _mm256_extractf128_pd(sum_halves, 1);
  53. __m128d sum = _mm_add_pd(lo, hi);
  54. return _mm_cvtsd_f64(sum);
  55. }
  56. /***************************
  57. * memory
  58. ***************************/
  59. // unaligned load
  60. #define v_loadu_f32 _mm256_loadu_ps
  61. #define v_loadu_f64 _mm256_loadu_pd
  62. #define v_storeu_f32 _mm256_storeu_ps
  63. #define v_storeu_f64 _mm256_storeu_pd
  64. #define v_setall_f32(VAL) _mm256_set1_ps(VAL)
  65. #define v_setall_f64(VAL) _mm256_set1_pd(VAL)
  66. #define v_zero_f32 _mm256_setzero_ps
  67. #define v_zero_f64 _mm256_setzero_pd