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.
14070 Commits
9 Branches
1.8GB
Tree: e6509fcb51
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'e6509fcb51'
${ noResults }
JUCE/modules/juce_dsp/native/juce_SIMDNativeOps_avx.h

659 lines
50KB
Raw Blame History 

  1. /*

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

  3. 

  4.    This file is part of the JUCE library.

  5.    Copyright (c) 2022 - Raw Material Software Limited

  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 7 End-User License

  11.    Agreement and JUCE Privacy Policy.

  12. 

  13.    End User License Agreement: www.juce.com/juce-7-licence

  14.    Privacy Policy: www.juce.com/juce-privacy-policy

  15. 

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

  17.    www.gnu.org/licenses).

  18. 

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

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

  21.    DISCLAIMED.

  22. 

  23.   ==============================================================================

  24. */

  25. 

  26. namespace juce::dsp

  27. {

  28. 

  29. #ifndef DOXYGEN

  30. 

  31. JUCE_BEGIN_IGNORE_WARNINGS_GCC_LIKE ("-Wignored-attributes")

  32. 

  33. #ifdef _MSC_VER

  34.  #define DECLARE_AVX_SIMD_CONST(type, name) \

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

  36. 

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

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

  39. 

  40. #else

  41.  #define DECLARE_AVX_SIMD_CONST(type, name) \

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

  43. 

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

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

  46. 

  47. #endif

  48. 

  49. template <typename type>

  50. struct SIMDNativeOps;

  51. 

  52. //==============================================================================

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

  54. 

  55.     @tags{DSP}

  56. */

  57. template <>

  58. struct SIMDNativeOps<float>

  59. {

  60.     using vSIMDType = __m256;

  61. 

  62.     //==============================================================================

  63.     DECLARE_AVX_SIMD_CONST (int32_t, kAllBitsSet);

  64.     DECLARE_AVX_SIMD_CONST (int32_t, kEvenHighBit);

  65.     DECLARE_AVX_SIMD_CONST (float, kOne);

  66. 

  67.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  93.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE truncate (__m256 a) noexcept                         { return _mm256_cvtepi32_ps (_mm256_cvttps_epi32 (a)); }

  94. 

  95.     static forcedinline __m256 JUCE_VECTOR_CALLTYPE multiplyAdd (__m256 a, __m256 b, __m256 c) noexcept

  96.     {

  97.        #if __FMA__

  98.         return _mm256_fmadd_ps (b, c, a);

  99.        #else

  100.         return add (a, mul (b, c));

  101.        #endif

  102.     }

  103. 

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

  105.     {

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

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

  108.     }

  109. 

  110.     //==============================================================================

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

  112.     {

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

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

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

  116.     }

  117. 

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

  119.     {

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

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

  122.        retval = _mm256_add_ps (retval, tmp);

  123. 

  124.       #if JUCE_GCC

  125.        return retval[0];

  126.       #else

  127.        return _mm256_cvtss_f32 (retval);

  128.       #endif

  129.     }

  130. };

  131. 

  132. //==============================================================================

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

  134. 

  135.     @tags{DSP}

  136. */

  137. template <>

  138. struct SIMDNativeOps<double>

  139. {

  140.     using vSIMDType = __m256d;

  141. 

  142.     //==============================================================================

  143.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  144.     DECLARE_AVX_SIMD_CONST (int64_t, kEvenHighBit);

  145.     DECLARE_AVX_SIMD_CONST (double, kOne);

  146. 

  147.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  175.     static forcedinline __m256d JUCE_VECTOR_CALLTYPE truncate (__m256d a) noexcept                          { return _mm256_cvtepi32_pd (_mm256_cvttpd_epi32 (a)); }

  176. 

  177.     //==============================================================================

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

  179.     {

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

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

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

  183.     }

  184. 

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

  186.     {

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

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

  189.         retval = _mm256_add_pd (retval, tmp);

  190. 

  191.        #if JUCE_GCC

  192.         return retval[0];

  193.        #else

  194.         return _mm256_cvtsd_f64 (retval);

  195.        #endif

  196.     }

  197. };

  198. 

  199. //==============================================================================

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

  201. 

  202.     @tags{DSP}

  203. */

  204. template <>

  205. struct SIMDNativeOps<int8_t>

  206. {

  207.     using vSIMDType = __m256i;

  208. 

  209.     //==============================================================================

  210.     DECLARE_AVX_SIMD_CONST (int8_t, kAllBitsSet);

  211. 

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

  213.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int8_t* p) noexcept                        { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  214.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, int8_t* dest) noexcept              { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  231.     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); }

  232.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                          { return a; }

  233. 

  234.     //==============================================================================

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

  236.     {

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

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

  239. 

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

  241.         {

  242.             lo = _mm256_hadd_epi16 (lo, lo);

  243.             hi = _mm256_hadd_epi16 (hi, hi);

  244.         }

  245. 

  246.        #if JUCE_GCC

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

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

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

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

  251.        #else

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

  253. 

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

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

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

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

  258.        #endif

  259.     }

  260. 

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

  262.     {

  263.         // unpack and multiply

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

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

  266. 

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

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

  269.     }

  270. };

  271. 

  272. //==============================================================================

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

  274. 

  275.     @tags{DSP}

  276. */

  277. template <>

  278. struct SIMDNativeOps<uint8_t>

  279. {

  280.     //==============================================================================

  281.     using vSIMDType = __m256i;

  282. 

  283.     //==============================================================================

  284.     DECLARE_AVX_SIMD_CONST (uint8_t, kHighBit);

  285.     DECLARE_AVX_SIMD_CONST (uint8_t, kAllBitsSet);

  286. 

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

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

  289.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const uint8_t* p) noexcept                        { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  290.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, uint8_t* dest) noexcept              { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  307.     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); }

  308.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                           { return a; }

  309. 

  310.     //==============================================================================

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

  312.     {

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

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

  315. 

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

  317.         {

  318.             lo = _mm256_hadd_epi16 (lo, lo);

  319.             hi = _mm256_hadd_epi16 (hi, hi);

  320.         }

  321. 

  322.        #if JUCE_GCC

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

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

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

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

  327.        #else

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

  329. 

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

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

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

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

  334.        #endif

  335.     }

  336. 

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

  338.     {

  339.         // unpack and multiply

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

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

  342. 

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

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

  345.     }

  346. };

  347. 

  348. //==============================================================================

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

  350. 

  351.     @tags{DSP}

  352. */

  353. template <>

  354. struct SIMDNativeOps<int16_t>

  355. {

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

  357.     using vSIMDType = __m256i;

  358. 

  359.     //==============================================================================

  360.     DECLARE_AVX_SIMD_CONST (int16_t, kAllBitsSet);

  361. 

  362.     //==============================================================================

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

  364.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int16_t* p) noexcept                        { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  365.     static forcedinline void JUCE_VECTOR_CALLTYPE store (__m256i value, int16_t* dest) noexcept              { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  383.     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); }

  384.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                           { return a; }

  385. 

  386.     //==============================================================================

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

  388.     {

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

  390.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  391.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  392. 

  393.        #if JUCE_GCC

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

  395.        #else

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

  397. 

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

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

  400.        #endif

  401.     }

  402. };

  403. 

  404. //==============================================================================

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

  406. 

  407.     @tags{DSP}

  408. */

  409. template <>

  410. struct SIMDNativeOps<uint16_t>

  411. {

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

  413.     using vSIMDType = __m256i;

  414. 

  415.     //==============================================================================

  416.     DECLARE_AVX_SIMD_CONST (uint16_t, kHighBit);

  417.     DECLARE_AVX_SIMD_CONST (uint16_t, kAllBitsSet);

  418. 

  419.     //==============================================================================

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

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

  422.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint16_t* p) noexcept                       { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  423.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint16_t* dest) noexcept          { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  441.     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); }

  442.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                            { return a; }

  443. 

  444.     //==============================================================================

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

  446.     {

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

  448.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  449.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  450. 

  451.        #if JUCE_GCC

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

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

  454.        #else

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

  456. 

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

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

  459.        #endif

  460.     }

  461. };

  462. 

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

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

  465. 

  466.     @tags{DSP}

  467. */

  468. template <>

  469. struct SIMDNativeOps<int32_t>

  470. {

  471.     //==============================================================================

  472.     using vSIMDType = __m256i;

  473. 

  474.     //==============================================================================

  475.     DECLARE_AVX_SIMD_CONST (int32_t, kAllBitsSet);

  476. 

  477.     //==============================================================================

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

  479.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int32_t* p) noexcept                        { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  480.     static forcedinline void    JUCE_VECTOR_CALLTYPE store (__m256i value, int32_t* dest) noexcept           { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  498.     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); }

  499.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                           { return a; }

  500. 

  501.     //==============================================================================

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

  503.     {

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

  505.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  506. 

  507.        #if JUCE_GCC

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

  509.        #else

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

  511. 

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

  513.        #endif

  514.     }

  515. };

  516. 

  517. //==============================================================================

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

  519. 

  520.     @tags{DSP}

  521. */

  522. template <>

  523. struct SIMDNativeOps<uint32_t>

  524. {

  525.     //==============================================================================

  526.     using vSIMDType = __m256i;

  527. 

  528.     //==============================================================================

  529.     DECLARE_AVX_SIMD_CONST (uint32_t, kAllBitsSet);

  530.     DECLARE_AVX_SIMD_CONST (uint32_t, kHighBit);

  531. 

  532.     //==============================================================================

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

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

  535.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint32_t* p) noexcept                       { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  536.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint32_t* dest) noexcept          { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  554.     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); }

  555.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                            { return a; }

  556. 

  557.     //==============================================================================

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

  559.     {

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

  561.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  562. 

  563.        #if JUCE_GCC

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

  565.        #else

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

  567. 

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

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

  570.        #endif

  571.     }

  572. };

  573. 

  574. //==============================================================================

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

  576. 

  577.     @tags{DSP}

  578. */

  579. template <>

  580. struct SIMDNativeOps<int64_t>

  581. {

  582.     //==============================================================================

  583.     using vSIMDType = __m256i;

  584. 

  585.     //==============================================================================

  586.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  587. 

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

  589.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE load (const int64_t* p) noexcept                        { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  590.     static forcedinline void    JUCE_VECTOR_CALLTYPE store (__m256i value, int64_t* dest) noexcept           { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

  598.     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)); }

  599.     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)); }

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

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

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

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

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

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

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

  607.     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); }

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

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

  610.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                           { return a; }

  611. };

  612. 

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

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

  615. 

  616.     @tags{DSP}

  617. */

  618. template <>

  619. struct SIMDNativeOps<uint64_t>

  620. {

  621.     //==============================================================================

  622.     using vSIMDType = __m256i;

  623. 

  624.     //==============================================================================

  625.     DECLARE_AVX_SIMD_CONST (uint64_t, kAllBitsSet);

  626.     DECLARE_AVX_SIMD_CONST (uint64_t, kHighBit);

  627. 

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

  629.     static forcedinline __m256i  JUCE_VECTOR_CALLTYPE load (const uint64_t* p) noexcept                       { return _mm256_load_si256 (reinterpret_cast<const __m256i*> (p)); }

  630.     static forcedinline void     JUCE_VECTOR_CALLTYPE store (__m256i value, uint64_t* dest) noexcept          { _mm256_store_si256 (reinterpret_cast<__m256i*> (dest), value); }

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

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

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

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

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

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

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

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

  639.     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)); }

  640.     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)); }

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

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

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

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

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

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

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

  648.     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); }

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

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

  651.     static forcedinline __m256i JUCE_VECTOR_CALLTYPE truncate (__m256i a) noexcept                            { return a; }

  652. };

  653. 

  654. #endif

  655. 

  656. JUCE_END_IGNORE_WARNINGS_GCC_LIKE

  657. 

  658. } // namespace juce::dsp