DISTRHO
/
JUCE
mirror of https://github.com/DISTRHO/JUCE
1
0
Fork 0
Code Releases 1 Activity
The JUCE cross-platform C++ framework, with DISTRHO/KXStudio specific changes
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
9257 Commits
9 Branches
11GB
Tree: 7ed282f314
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '7ed282f314'
${ noResults }
JUCE/modules/juce_dsp/native/juce_avx_SIMDNativeOps.h

650 lines
48KB
Raw Blame History 

  1. /*

  2.   ==============================================================================

  3. 

  4.    This file is part of the JUCE library.

  5.    Copyright (c) 2017 - ROLI Ltd.

  6. 

  7.    JUCE is an open source library subject to commercial or open-source

  8.    licensing.

  9. 

  10.    By using JUCE, you agree to the terms of both the JUCE 5 End-User License

  11.    Agreement and JUCE 5 Privacy Policy (both updated and effective as of the

  12.    27th April 2017).

  13. 

  14.    End User License Agreement: www.juce.com/juce-5-licence

  15.    Privacy Policy: www.juce.com/juce-5-privacy-policy

  16. 

  17.    Or: You may also use this code under the terms of the GPL v3 (see

  18.    www.gnu.org/licenses).

  19. 

  20.    JUCE IS PROVIDED "AS IS" WITHOUT ANY WARRANTY, AND ALL WARRANTIES, WHETHER

  21.    EXPRESSED OR IMPLIED, INCLUDING MERCHANTABILITY AND FITNESS FOR PURPOSE, ARE

  22.    DISCLAIMED.

  23. 

  24.   ==============================================================================

  25. */

  26. 

  27. namespace juce

  28. {

  29. namespace dsp

  30. {

  31. 

  32. #ifndef DOXYGEN

  33. 

  34. #if JUCE_GCC && (__GNUC__ >= 6)

  35.  #pragma GCC diagnostic push

  36.  #pragma GCC diagnostic ignored "-Wignored-attributes"

  37. #endif

  38. 

  39. #ifdef _MSC_VER

  40.  #define DECLARE_AVX_SIMD_CONST(type, name) \

  41.     static __declspec(align(32)) const type name[32 / sizeof (type)]

  42. 

  43.  #define DEFINE_AVX_SIMD_CONST(type, class_type, name) \

  44.     __declspec(align(32)) const type SIMDNativeOps<class_type>:: name[32 / sizeof (type)]

  45. 

  46. #else

  47.  #define DECLARE_AVX_SIMD_CONST(type, name) \

  48.     static const type name[32 / sizeof (type)] __attribute__((aligned(32)))

  49. 

  50.  #define DEFINE_AVX_SIMD_CONST(type, class_type, name) \

  51.     const type SIMDNativeOps<class_type>:: name[32 / sizeof (type)] __attribute__((aligned(32)))

  52. 

  53. #endif

  54. 

  55. template <typename type>

  56. struct SIMDNativeOps;

  57. 

  58. //==============================================================================

  59. /** Single-precision floating point AVX intrinsics.

  60. 

  61.     @tags{DSP}

  62. */

  63. template <>

  64. struct SIMDNativeOps<float>

  65. {

  66.     using vSIMDType = __m256;

  67. 

  68.     //==============================================================================

  69.     DECLARE_AVX_SIMD_CONST (int32_t, kAllBitsSet);

  70.     DECLARE_AVX_SIMD_CONST (int32_t, kEvenHighBit);

  71.     DECLARE_AVX_SIMD_CONST (float, kOne);

  72. 

  73.     //==============================================================================

  74.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE vconst (const float* a) noexcept                     { return load (a); }

  75.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE vconst (const int32_t* a) noexcept                   { return _mm256_castsi256_ps (_mm256_load_si256 ((const __m256i*) a)); }

  76.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE expand (float s) noexcept                            { return _mm256_broadcast_ss (&s); }

  77.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE load (const float* a) noexcept                       { return _mm256_load_ps (a); }

  78.     static forcedinline void   JUCE_VECTOR_CALLTYPE store (__m256 value, float* dest) noexcept           { _mm256_store_ps (dest, value); }

  79.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE add (__m256 a, __m256 b) noexcept                    { return _mm256_add_ps (a, b); }

  80.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE sub (__m256 a, __m256 b) noexcept                    { return _mm256_sub_ps (a, b); }

  81.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE mul (__m256 a, __m256 b) noexcept                    { return _mm256_mul_ps (a, b); }

  82.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE bit_and (__m256 a, __m256 b) noexcept                { return _mm256_and_ps (a, b); }

  83.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE bit_or  (__m256 a, __m256 b) noexcept                { return _mm256_or_ps  (a, b); }

  84.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE bit_xor (__m256 a, __m256 b) noexcept                { return _mm256_xor_ps (a, b); }

  85.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE bit_notand (__m256 a, __m256 b) noexcept             { return _mm256_andnot_ps (a, b); }

  86.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE bit_not (__m256 a) noexcept                          { return bit_notand (a, vconst (kAllBitsSet)); }

  87.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE min (__m256 a, __m256 b) noexcept                    { return _mm256_min_ps (a, b); }

  88.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE max (__m256 a, __m256 b) noexcept                    { return _mm256_max_ps (a, b); }

  89.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE equal (__m256 a, __m256 b) noexcept                  { return _mm256_cmp_ps (a, b, _CMP_EQ_OQ); }

  90.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE notEqual (__m256 a, __m256 b) noexcept               { return _mm256_cmp_ps (a, b, _CMP_NEQ_OQ); }

  91.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE greaterThan (__m256 a, __m256 b) noexcept            { return _mm256_cmp_ps (a, b, _CMP_GT_OQ); }

  92.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256 a, __m256 b) noexcept     { return _mm256_cmp_ps (a, b, _CMP_GE_OQ); }

  93.     static forcedinline bool   JUCE_VECTOR_CALLTYPE allEqual (__m256 a, __m256 b) noexcept               { return (_mm256_movemask_ps (equal (a, b)) == 0xff); }

  94.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE multiplyAdd (__m256 a, __m256 b, __m256 c) noexcept  { return _mm256_fmadd_ps (b, c, a); }

  95.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE dupeven (__m256 a) noexcept                          { return _mm256_shuffle_ps (a, a, _MM_SHUFFLE (2, 2, 0, 0)); }

  96.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE dupodd (__m256 a) noexcept                           { return _mm256_shuffle_ps (a, a, _MM_SHUFFLE (3, 3, 1, 1)); }

  97.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE swapevenodd (__m256 a) noexcept                      { return _mm256_shuffle_ps (a, a, _MM_SHUFFLE (2, 3, 0, 1)); }

  98.     static forcedinline float  JUCE_VECTOR_CALLTYPE get (__m256 v, size_t i) noexcept                    { return SIMDFallbackOps<float, __m256>::get (v, i); }

  99.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE set (__m256 v, size_t i, float s) noexcept           { return SIMDFallbackOps<float, __m256>::set (v, i, s); }

  100.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE oddevensum (__m256 a) noexcept

  101.     {

  102.         a = _mm256_add_ps (_mm256_shuffle_ps (a, a, _MM_SHUFFLE (1, 0, 3, 2)), a);

  103.         return add (_mm256_permute2f128_ps (a, a, 1), a);

  104.     }

  105. 

  106.     //==============================================================================

  107.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE cmplxmul (__m256 a, __m256 b) noexcept

  108.     {

  109.         __m256 rr_ir = mul (a, dupeven (b));

  110.         __m256 ii_ri = mul (swapevenodd (a), dupodd (b));

  111.         return add (rr_ir, bit_xor (ii_ri, vconst (kEvenHighBit)));

  112.     }

  113. 

  114.     static forcedinline float JUCE_VECTOR_CALLTYPE sum (__m256 a) noexcept

  115.     {

  116.        __m256 retval = _mm256_dp_ps (a, vconst (kOne), 0xff);

  117.        __m256 tmp = _mm256_permute2f128_ps (retval, retval, 1);

  118.        retval = _mm256_add_ps (retval, tmp);

  119. 

  120.       #if JUCE_GCC

  121.        return retval[0];

  122.       #else

  123.        return _mm256_cvtss_f32 (retval);

  124.       #endif

  125.     }

  126. };

  127. 

  128. //==============================================================================

  129. /** Double-precision floating point AVX intrinsics.

  130. 

  131.     @tags{DSP}

  132. */

  133. template <>

  134. struct SIMDNativeOps<double>

  135. {

  136.     using vSIMDType = __m256d;

  137. 

  138.     //==============================================================================

  139.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  140.     DECLARE_AVX_SIMD_CONST (int64_t, kEvenHighBit);

  141.     DECLARE_AVX_SIMD_CONST (double, kOne);

  142. 

  143.     //==============================================================================

  144.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE vconst (const double* a) noexcept                      { return load (a); }

  145.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE vconst (const int64_t* a) noexcept                     { return _mm256_castsi256_pd (_mm256_load_si256 ((const __m256i*) a)); }

  146.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE expand (double s) noexcept                             { return _mm256_broadcast_sd (&s); }

  147.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE load (const double* a) noexcept                        { return _mm256_load_pd (a); }

  148.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256d value, double* dest) noexcept              { _mm256_store_pd (dest, value); }

  149.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE add (__m256d a, __m256d b) noexcept                    { return _mm256_add_pd (a, b); }

  150.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE sub (__m256d a, __m256d b) noexcept                    { return _mm256_sub_pd (a, b); }

  151.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE mul (__m256d a, __m256d b) noexcept                    { return _mm256_mul_pd (a, b); }

  152.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE bit_and (__m256d a, __m256d b) noexcept                { return _mm256_and_pd (a, b); }

  153.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE bit_or  (__m256d a, __m256d b) noexcept                { return _mm256_or_pd  (a, b); }

  154.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE bit_xor (__m256d a, __m256d b) noexcept                { return _mm256_xor_pd (a, b); }

  155.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE bit_notand (__m256d a, __m256d b) noexcept             { return _mm256_andnot_pd (a, b); }

  156.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE bit_not (__m256d a) noexcept                           { return bit_notand (a, vconst (kAllBitsSet)); }

  157.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE min (__m256d a, __m256d b) noexcept                    { return _mm256_min_pd (a, b); }

  158.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE max (__m256d a, __m256d b) noexcept                    { return _mm256_max_pd (a, b); }

  159.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE equal (__m256d a, __m256d b) noexcept                  { return _mm256_cmp_pd (a, b, _CMP_EQ_OQ); }

  160.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE notEqual (__m256d a, __m256d b) noexcept               { return _mm256_cmp_pd (a, b, _CMP_NEQ_OQ); }

  161.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE greaterThan (__m256d a, __m256d b) noexcept            { return _mm256_cmp_pd (a, b, _CMP_GT_OQ); }

  162.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256d a, __m256d b) noexcept     { return _mm256_cmp_pd (a, b, _CMP_GE_OQ); }

  163.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256d a, __m256d b) noexcept               { return (_mm256_movemask_pd (equal (a, b)) == 0xf); }

  164.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE multiplyAdd (__m256d a, __m256d b, __m256d c) noexcept { return _mm256_add_pd (a, _mm256_mul_pd (b, c)); }

  165.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE dupeven (__m256d a) noexcept                           { return _mm256_shuffle_pd (a, a, 0); }

  166.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE dupodd (__m256d a) noexcept                            { return _mm256_shuffle_pd (a, a, (1 << 0) | (1 << 1) | (1 << 2) | (1 << 3)); }

  167.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE swapevenodd (__m256d a) noexcept                       { return _mm256_shuffle_pd (a, a, (1 << 0) | (0 << 1) | (1 << 2) | (0 << 3)); }

  168.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE oddevensum (__m256d a) noexcept                        { return _mm256_add_pd (_mm256_permute2f128_pd (a, a, 1), a); }

  169.     static forcedinline double  JUCE_VECTOR_CALLTYPE get (__m256d v, size_t i) noexcept                     { return SIMDFallbackOps<double, __m256d>::get (v, i); }

  170.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE set (__m256d v, size_t i, double s) noexcept           { return SIMDFallbackOps<double, __m256d>::set (v, i, s); }

  171. 

  172. 

  173.     //==============================================================================

  174.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE cmplxmul (__m256d a, __m256d b) noexcept

  175.     {

  176.         __m256d rr_ir = mul (a, dupeven (b));

  177.         __m256d ii_ri = mul (swapevenodd (a), dupodd (b));

  178.         return add (rr_ir, bit_xor (ii_ri, vconst (kEvenHighBit)));

  179.     }

  180. 

  181.     static forcedinline double JUCE_VECTOR_CALLTYPE sum (__m256d a) noexcept

  182.     {

  183.         __m256d retval = _mm256_hadd_pd (a, a);

  184.         __m256d tmp = _mm256_permute2f128_pd (retval, retval, 1);

  185.         retval = _mm256_add_pd (retval, tmp);

  186. 

  187.        #if JUCE_GCC

  188.         return retval[0];

  189.        #else

  190.         return _mm256_cvtsd_f64 (retval);

  191.        #endif

  192.     }

  193. };

  194. 

  195. //==============================================================================

  196. /** Signed 8-bit integer AVX intrinsics

  197. 

  198.     @tags{DSP}

  199. */

  200. template <>

  201. struct SIMDNativeOps<int8_t>

  202. {

  203.     using vSIMDType = __m256i;

  204. 

  205.     //==============================================================================

  206.     DECLARE_AVX_SIMD_CONST (int8_t, kAllBitsSet);

  207. 

  208.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE expand (int8_t s) noexcept                             { return _mm256_set1_epi8 (s); }

  209.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int8_t* p) noexcept                        { return _mm256_load_si256 ((const __m256i*) p); }

  210.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, int8_t* dest) noexcept              { _mm256_store_si256 ((__m256i*) dest, value); }

  211.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                    { return _mm256_add_epi8 (a, b); }

  212.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                    { return _mm256_sub_epi8 (a, b); }

  213.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                { return _mm256_and_si256 (a, b); }

  214.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                { return _mm256_or_si256  (a, b); }

  215.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                { return _mm256_xor_si256 (a, b); }

  216.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept             { return _mm256_andnot_si256 (a, b); }

  217.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                           { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  218.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                    { return _mm256_min_epi8 (a, b); }

  219.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                    { return _mm256_max_epi8 (a, b); }

  220.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                  { return _mm256_cmpeq_epi8 (a, b); }

  221.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept            { return _mm256_cmpgt_epi8 (a, b); }

  222.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept     { return bit_or (greaterThan (a, b), equal (a,b)); }

  223.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept               { return _mm256_movemask_epi8 (equal (a, b)) == -1; }

  224.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept { return add (a, mul (b, c)); }

  225.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept               { return bit_not (equal (a, b)); }

  226.     static forcedinline int8_t  JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                     { return SIMDFallbackOps<int8_t, __m256i>::get (v, i); }

  227.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, int8_t s) noexcept           { return SIMDFallbackOps<int8_t, __m256i>::set (v, i, s); }

  228. 

  229.     //==============================================================================

  230.     static forcedinline int8_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  231.     {

  232.         __m256i lo = _mm256_unpacklo_epi8 (a, _mm256_setzero_si256());

  233.         __m256i hi = _mm256_unpackhi_epi8 (a, _mm256_setzero_si256());

  234. 

  235.         for (int i = 0; i < 3; ++i)

  236.         {

  237.             lo = _mm256_hadd_epi16 (lo, lo);

  238.             hi = _mm256_hadd_epi16 (hi, hi);

  239.         }

  240. 

  241.        #if JUCE_GCC

  242.         return (int8_t) ((lo[0] & 0xff) +

  243.                          (hi[0] & 0xff) +

  244.                          (lo[2] & 0xff) +

  245.                          (hi[2] & 0xff));

  246.        #else

  247.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  248. 

  249.         return (int8_t) ((_mm256_cvtsi256_si32 (lo) & 0xff) +

  250.                          (_mm256_cvtsi256_si32 (hi) & 0xff) +

  251.                          (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (lo, mask)) & 0xff) +

  252.                          (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (hi, mask)) & 0xff));

  253.        #endif

  254.     }

  255. 

  256.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b)

  257.     {

  258.         // unpack and multiply

  259.         __m256i even = _mm256_mullo_epi16 (a, b);

  260.         __m256i odd  = _mm256_mullo_epi16 (_mm256_srli_epi16 (a, 8), _mm256_srli_epi16 (b, 8));

  261. 

  262.         return _mm256_or_si256 (_mm256_slli_epi16 (odd, 8),

  263.                              _mm256_srli_epi16 (_mm256_slli_epi16 (even, 8), 8));

  264.     }

  265. };

  266. 

  267. //==============================================================================

  268. /** Unsigned 8-bit integer AVX intrinsics.

  269. 

  270.     @tags{DSP}

  271. */

  272. template <>

  273. struct SIMDNativeOps<uint8_t>

  274. {

  275.     //==============================================================================

  276.     using vSIMDType = __m256i;

  277. 

  278.     //==============================================================================

  279.     DECLARE_AVX_SIMD_CONST (uint8_t, kHighBit);

  280.     DECLARE_AVX_SIMD_CONST (uint8_t, kAllBitsSet);

  281. 

  282.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE ssign (__m256i a) noexcept                              { return _mm256_xor_si256 (a, load (kHighBit)); }

  283.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE expand (uint8_t s) noexcept                             { return _mm256_set1_epi8 ((int8_t) s); }

  284.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const uint8_t* p) noexcept                        { return _mm256_load_si256 ((const __m256i*) p); }

  285.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, uint8_t* dest) noexcept              { _mm256_store_si256 ((__m256i*) dest, value); }

  286.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi8 (a, b); }

  287.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi8 (a, b); }

  288.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  289.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  290.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  291.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  292.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  293.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { return _mm256_min_epu8 (a, b); }

  294.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { return _mm256_max_epu8 (a, b); }

  295.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi8 (a, b); }

  296.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi8 (ssign (a), ssign (b)); }

  297.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  298.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  299.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  300.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  301.     static forcedinline uint8_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<uint8_t, __m256i>::get (v, i); }

  302.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, uint8_t s) noexcept           { return SIMDFallbackOps<uint8_t, __m256i>::set (v, i, s); }

  303. 

  304.     //==============================================================================

  305.     static forcedinline uint8_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  306.     {

  307.         __m256i lo = _mm256_unpacklo_epi8 (a, _mm256_setzero_si256());

  308.         __m256i hi = _mm256_unpackhi_epi8 (a, _mm256_setzero_si256());

  309. 

  310.         for (int i = 0; i < 3; ++i)

  311.         {

  312.             lo = _mm256_hadd_epi16 (lo, lo);

  313.             hi = _mm256_hadd_epi16 (hi, hi);

  314.         }

  315. 

  316.        #if JUCE_GCC

  317.         return (uint8_t) ((static_cast<uint32_t> (lo[0]) & 0xffu) +

  318.                           (static_cast<uint32_t> (hi[0]) & 0xffu) +

  319.                           (static_cast<uint32_t> (lo[2]) & 0xffu) +

  320.                           (static_cast<uint32_t> (hi[2]) & 0xffu));

  321.        #else

  322.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  323. 

  324.         return (uint8_t) ((static_cast<uint32_t> (_mm256_cvtsi256_si32 (lo)) & 0xffu) +

  325.                           (static_cast<uint32_t> (_mm256_cvtsi256_si32 (hi)) & 0xffu) +

  326.                           (static_cast<uint32_t> (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (lo, mask))) & 0xffu) +

  327.                           (static_cast<uint32_t> (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (hi, mask))) & 0xffu));

  328.        #endif

  329.     }

  330. 

  331.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b)

  332.     {

  333.         // unpack and multiply

  334.         __m256i even = _mm256_mullo_epi16 (a, b);

  335.         __m256i odd  = _mm256_mullo_epi16 (_mm256_srli_epi16 (a, 8), _mm256_srli_epi16 (b, 8));

  336. 

  337.         return _mm256_or_si256 (_mm256_slli_epi16 (odd, 8),

  338.                              _mm256_srli_epi16 (_mm256_slli_epi16 (even, 8), 8));

  339.     }

  340. };

  341. 

  342. //==============================================================================

  343. /** Signed 16-bit integer AVX intrinsics.

  344. 

  345.     @tags{DSP}

  346. */

  347. template <>

  348. struct SIMDNativeOps<int16_t>

  349. {

  350.     //==============================================================================

  351.     using vSIMDType = __m256i;

  352. 

  353.     //==============================================================================

  354.     DECLARE_AVX_SIMD_CONST (int16_t, kAllBitsSet);

  355. 

  356.     //==============================================================================

  357.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE expand (int16_t s) noexcept                             { return _mm256_set1_epi16 (s); }

  358.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int16_t* p) noexcept                        { return _mm256_load_si256 ((const __m256i*) p); }

  359.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, int16_t* dest) noexcept              { _mm256_store_si256 ((__m256i*) dest, value); }

  360.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi16 (a, b); }

  361.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi16 (a, b); }

  362.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return _mm256_mullo_epi16 (a, b); }

  363.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  364.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  365.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  366.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  367.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  368.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { return _mm256_min_epi16 (a, b); }

  369.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { return _mm256_max_epi16 (a, b); }

  370.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi16 (a, b); }

  371.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi16 (a, b); }

  372.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  373.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  374.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  375.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  376.     static forcedinline int16_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<int16_t, __m256i>::get (v, i); }

  377.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, int16_t s) noexcept           { return SIMDFallbackOps<int16_t, __m256i>::set (v, i, s); }

  378. 

  379.     //==============================================================================

  380.     static forcedinline int16_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  381.     {

  382.         __m256i tmp = _mm256_hadd_epi16 (a, a);

  383.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  384.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  385. 

  386.        #if JUCE_GCC

  387.         return (int16_t) ((tmp[0] & 0xffff) + (tmp[2] & 0xffff));

  388.        #else

  389.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  390. 

  391.         return (int16_t) ((_mm256_cvtsi256_si32 (tmp) & 0xffff) +

  392.                           (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (tmp, mask)) & 0xffff));

  393.        #endif

  394.     }

  395. };

  396. 

  397. //==============================================================================

  398. /** Unsigned 16-bit integer AVX intrinsics.

  399. 

  400.     @tags{DSP}

  401. */

  402. template <>

  403. struct SIMDNativeOps<uint16_t>

  404. {

  405.     //==============================================================================

  406.     using vSIMDType = __m256i;

  407. 

  408.     //==============================================================================

  409.     DECLARE_AVX_SIMD_CONST (uint16_t, kHighBit);

  410.     DECLARE_AVX_SIMD_CONST (uint16_t, kAllBitsSet);

  411. 

  412.     //==============================================================================

  413.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE ssign (__m256i a) noexcept                              { return _mm256_xor_si256 (a, load (kHighBit)); }

  414.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE expand (uint16_t s) noexcept                            { return _mm256_set1_epi16 ((int16_t) s); }

  415.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint16_t* p) noexcept                       { return _mm256_load_si256 ((const __m256i*) p); }

  416.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint16_t* dest) noexcept          { _mm256_store_si256 ((__m256i*) dest, value); }

  417.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi16 (a, b); }

  418.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi16 (a, b); }

  419.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return _mm256_mullo_epi16 (a, b); }

  420.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  421.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  422.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  423.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  424.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  425.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { return _mm256_min_epu16 (a, b); }

  426.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { return _mm256_max_epu16 (a, b); }

  427.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi16 (a, b); }

  428.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi16 (ssign (a), ssign (b)); }

  429.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  430.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  431.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  432.     static forcedinline bool     JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  433.     static forcedinline uint16_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<uint16_t, __m256i>::get (v, i); }

  434.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, uint16_t s) noexcept          { return SIMDFallbackOps<uint16_t, __m256i>::set (v, i, s); }

  435. 

  436.     //==============================================================================

  437.     static forcedinline uint16_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  438.     {

  439.         __m256i tmp = _mm256_hadd_epi16 (a, a);

  440.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  441.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  442. 

  443.        #if JUCE_GCC

  444.         return (uint16_t) ((static_cast<uint32_t> (tmp[0]) & 0xffffu) +

  445.                            (static_cast<uint32_t> (tmp[2]) & 0xffffu));

  446.        #else

  447.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  448. 

  449.         return (uint16_t) ((static_cast<uint32_t> (_mm256_cvtsi256_si32 (tmp)) & 0xffffu) +

  450.                            (static_cast<uint32_t> (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (tmp, mask))) & 0xffffu));

  451.        #endif

  452.     }

  453. };

  454. 

  455. //==============================================================================

  456. /** Signed 32-bit integer AVX intrinsics.

  457. 

  458.     @tags{DSP}

  459. */

  460. template <>

  461. struct SIMDNativeOps<int32_t>

  462. {

  463.     //==============================================================================

  464.     using vSIMDType = __m256i;

  465. 

  466.     //==============================================================================

  467.     DECLARE_AVX_SIMD_CONST (int32_t, kAllBitsSet);

  468. 

  469.     //==============================================================================

  470.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE expand (int32_t s) noexcept                             { return _mm256_set1_epi32 (s); }

  471.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int32_t* p) noexcept                        { return _mm256_load_si256 ((const __m256i*) p); }

  472.     static forcedinline void    JUCE_VECTOR_CALLTYPE store (__m256i value, int32_t* dest) noexcept           { _mm256_store_si256 ((__m256i*) dest, value); }

  473.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi32 (a, b); }

  474.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi32 (a, b); }

  475.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return _mm256_mullo_epi32 (a, b); }

  476.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  477.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  478.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  479.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  480.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  481.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { return _mm256_min_epi32 (a, b); }

  482.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { return _mm256_max_epi32 (a, b); }

  483.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi32 (a, b); }

  484.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi32 (a, b); }

  485.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  486.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  487.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  488.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  489.     static forcedinline int32_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<int32_t, __m256i>::get (v, i); }

  490.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, int32_t s) noexcept           { return SIMDFallbackOps<int32_t, __m256i>::set (v, i, s); }

  491. 

  492.     //==============================================================================

  493.     static forcedinline int32_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  494.     {

  495.         __m256i tmp = _mm256_hadd_epi32 (a, a);

  496.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  497. 

  498.        #if JUCE_GCC

  499.         return (int32_t) (tmp[0] + tmp[2]);

  500.        #else

  501.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  502. 

  503.         return _mm256_cvtsi256_si32 (tmp) + _mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (tmp, mask));

  504.        #endif

  505.     }

  506. };

  507. 

  508. //==============================================================================

  509. /** Unsigned 32-bit integer AVX intrinsics.

  510. 

  511.     @tags{DSP}

  512. */

  513. template <>

  514. struct SIMDNativeOps<uint32_t>

  515. {

  516.     //==============================================================================

  517.     using vSIMDType = __m256i;

  518. 

  519.     //==============================================================================

  520.     DECLARE_AVX_SIMD_CONST (uint32_t, kAllBitsSet);

  521.     DECLARE_AVX_SIMD_CONST (uint32_t, kHighBit);

  522. 

  523.     //==============================================================================

  524.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE ssign (__m256i a) noexcept                              { return _mm256_xor_si256 (a, load (kHighBit)); }

  525.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE expand (uint32_t s) noexcept                            { return _mm256_set1_epi32 ((int32_t) s); }

  526.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint32_t* p) noexcept                       { return _mm256_load_si256 ((const __m256i*) p); }

  527.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint32_t* dest) noexcept          { _mm256_store_si256 ((__m256i*) dest, value); }

  528.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi32 (a, b); }

  529.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi32 (a, b); }

  530.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return _mm256_mullo_epi32 (a, b); }

  531.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  532.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  533.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  534.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  535.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  536.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { return _mm256_min_epu32 (a, b); }

  537.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { return _mm256_max_epu32 (a, b); }

  538.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi32 (a, b); }

  539.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi32 (ssign (a), ssign (b)); }

  540.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  541.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  542.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  543.     static forcedinline bool     JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  544.     static forcedinline uint32_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<uint32_t, __m256i>::get (v, i); }

  545.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, uint32_t s) noexcept          { return SIMDFallbackOps<uint32_t, __m256i>::set (v, i, s); }

  546. 

  547.     //==============================================================================

  548.     static forcedinline uint32_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept

  549.     {

  550.         __m256i tmp = _mm256_hadd_epi32 (a, a);

  551.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  552. 

  553.        #if JUCE_GCC

  554.         return static_cast<uint32_t> (tmp[0]) + static_cast<uint32_t> (tmp[2]);

  555.        #else

  556.         constexpr int mask = (2 << 0) | (3 << 2) | (0 << 4) | (1 << 6);

  557. 

  558.         return static_cast<uint32_t> (_mm256_cvtsi256_si32 (tmp))

  559.             + static_cast<uint32_t> (_mm256_cvtsi256_si32 (_mm256_permute4x64_epi64 (tmp, mask)));

  560.        #endif

  561.     }

  562. };

  563. 

  564. //==============================================================================

  565. /** Signed 64-bit integer AVX intrinsics.

  566. 

  567.     @tags{DSP}

  568. */

  569. template <>

  570. struct SIMDNativeOps<int64_t>

  571. {

  572.     //==============================================================================

  573.     using vSIMDType = __m256i;

  574. 

  575.     //==============================================================================

  576.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  577. 

  578.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE expand (int64_t s) noexcept                             { return _mm256_set1_epi64x ((int64_t) s); }

  579.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int64_t* p) noexcept                        { return _mm256_load_si256 ((const __m256i*) p); }

  580.     static forcedinline void    JUCE_VECTOR_CALLTYPE store (__m256i value, int64_t* dest) noexcept           { _mm256_store_si256 ((__m256i*) dest, value); }

  581.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi64 (a, b); }

  582.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi64 (a, b); }

  583.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  584.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  585.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  586.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  587.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  588.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { __m256i lt = greaterThan (b, a); return bit_or (bit_and (lt, a), bit_andnot (lt, b)); }

  589.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { __m256i gt = greaterThan (a, b); return bit_or (bit_and (gt, a), bit_andnot (gt, b)); }

  590.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi64 (a, b); }

  591.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi64 (a, b); }

  592.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  593.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  594.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  595.     static forcedinline bool    JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  596.     static forcedinline int64_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<int64_t, __m256i>::get (v, i); }

  597.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, int64_t s) noexcept           { return SIMDFallbackOps<int64_t, __m256i>::set (v, i, s); }

  598.     static forcedinline int64_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept                                { return SIMDFallbackOps<int64_t, __m256i>::sum (a); }

  599.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return SIMDFallbackOps<int64_t, __m256i>::mul (a, b); }

  600. };

  601. 

  602. //==============================================================================

  603. /** Unsigned 64-bit integer AVX intrinsics.

  604. 

  605.     @tags{DSP}

  606. */

  607. template <>

  608. struct SIMDNativeOps<uint64_t>

  609. {

  610.     //==============================================================================

  611.     using vSIMDType = __m256i;

  612. 

  613.     //==============================================================================

  614.     DECLARE_AVX_SIMD_CONST (uint64_t, kAllBitsSet);

  615.     DECLARE_AVX_SIMD_CONST (uint64_t, kHighBit);

  616. 

  617.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE expand (uint64_t s) noexcept                            { return _mm256_set1_epi64x ((int64_t) s); }

  618.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint64_t* p) noexcept                       { return _mm256_load_si256 ((const __m256i*) p); }

  619.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint64_t* dest) noexcept          { _mm256_store_si256 ((__m256i*) dest, value); }

  620.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE ssign (__m256i a) noexcept                              { return _mm256_xor_si256 (a, load (kHighBit)); }

  621.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE add (__m256i a, __m256i b) noexcept                     { return _mm256_add_epi64 (a, b); }

  622.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE sub (__m256i a, __m256i b) noexcept                     { return _mm256_sub_epi64 (a, b); }

  623.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_and (__m256i a, __m256i b) noexcept                 { return _mm256_and_si256 (a, b); }

  624.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_or  (__m256i a, __m256i b) noexcept                 { return _mm256_or_si256  (a, b); }

  625.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_xor (__m256i a, __m256i b) noexcept                 { return _mm256_xor_si256 (a, b); }

  626.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_andnot (__m256i a, __m256i b) noexcept              { return _mm256_andnot_si256 (a, b); }

  627.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE bit_not (__m256i a) noexcept                            { return _mm256_andnot_si256 (a, load (kAllBitsSet)); }

  628.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE min (__m256i a, __m256i b) noexcept                     { __m256i lt = greaterThan (b, a); return bit_or (bit_and (lt, a), bit_andnot (lt, b)); }

  629.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE max (__m256i a, __m256i b) noexcept                     { __m256i gt = greaterThan (a, b); return bit_or (bit_and (gt, a), bit_andnot (gt, b)); }

  630.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE equal (__m256i a, __m256i b) noexcept                   { return _mm256_cmpeq_epi64 (a, b); }

  631.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThan (__m256i a, __m256i b) noexcept             { return _mm256_cmpgt_epi64 (ssign (a), ssign (b)); }

  632.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE greaterThanOrEqual (__m256i a, __m256i b) noexcept      { return bit_or (greaterThan (a, b), equal (a,b)); }

  633.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE multiplyAdd (__m256i a, __m256i b, __m256i c) noexcept  { return add (a, mul (b, c)); }

  634.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE notEqual (__m256i a, __m256i b) noexcept                { return bit_not (equal (a, b)); }

  635.     static forcedinline bool     JUCE_VECTOR_CALLTYPE allEqual (__m256i a, __m256i b) noexcept                { return (_mm256_movemask_epi8 (equal (a, b)) == -1); }

  636.     static forcedinline uint64_t JUCE_VECTOR_CALLTYPE get (__m256i v, size_t i) noexcept                      { return SIMDFallbackOps<uint64_t, __m256i>::get (v, i); }

  637.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE set (__m256i v, size_t i, uint64_t s) noexcept          { return SIMDFallbackOps<uint64_t, __m256i>::set (v, i, s); }

  638.     static forcedinline uint64_t JUCE_VECTOR_CALLTYPE sum (__m256i a) noexcept                                { return SIMDFallbackOps<uint64_t, __m256i>::sum (a); }

  639.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE mul (__m256i a, __m256i b) noexcept                     { return SIMDFallbackOps<uint64_t, __m256i>::mul (a, b); }

  640. };

  641. 

  642. #endif

  643. 

  644. #if JUCE_GCC && (__GNUC__ >= 6)

  645.  #pragma GCC diagnostic pop

  646. #endif

  647. 

  648. } // namespace dsp

  649. } // namespace juce