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.
10119 Commits
9 Branches
1.8GB
Tree: 84c05cbecc
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '84c05cbecc'
${ noResults }
JUCE/modules/juce_dsp/native/juce_avx_SIMDNativeOps.h

660 lines
50KB
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 truncate (__m256 a) noexcept                         { return _mm256_cvtepi32_ps (_mm256_cvttps_epi32 (a)); }

  101. 

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

  103.     {

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

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

  106.     }

  107. 

  108.     //==============================================================================

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

  110.     {

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

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

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

  114.     }

  115. 

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

  117.     {

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

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

  120.        retval = _mm256_add_ps (retval, tmp);

  121. 

  122.       #if JUCE_GCC

  123.        return retval[0];

  124.       #else

  125.        return _mm256_cvtss_f32 (retval);

  126.       #endif

  127.     }

  128. };

  129. 

  130. //==============================================================================

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

  132. 

  133.     @tags{DSP}

  134. */

  135. template <>

  136. struct SIMDNativeOps<double>

  137. {

  138.     using vSIMDType = __m256d;

  139. 

  140.     //==============================================================================

  141.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  142.     DECLARE_AVX_SIMD_CONST (int64_t, kEvenHighBit);

  143.     DECLARE_AVX_SIMD_CONST (double, kOne);

  144. 

  145.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  174. 

  175.     //==============================================================================

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

  177.     {

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

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

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

  181.     }

  182. 

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

  184.     {

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

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

  187.         retval = _mm256_add_pd (retval, tmp);

  188. 

  189.        #if JUCE_GCC

  190.         return retval[0];

  191.        #else

  192.         return _mm256_cvtsd_f64 (retval);

  193.        #endif

  194.     }

  195. };

  196. 

  197. //==============================================================================

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

  199. 

  200.     @tags{DSP}

  201. */

  202. template <>

  203. struct SIMDNativeOps<int8_t>

  204. {

  205.     using vSIMDType = __m256i;

  206. 

  207.     //==============================================================================

  208.     DECLARE_AVX_SIMD_CONST (int8_t, kAllBitsSet);

  209. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  229.     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); }

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

  231. 

  232.     //==============================================================================

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

  234.     {

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

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

  237. 

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

  239.         {

  240.             lo = _mm256_hadd_epi16 (lo, lo);

  241.             hi = _mm256_hadd_epi16 (hi, hi);

  242.         }

  243. 

  244.        #if JUCE_GCC

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

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

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

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

  249.        #else

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

  251. 

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

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

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

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

  256.        #endif

  257.     }

  258. 

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

  260.     {

  261.         // unpack and multiply

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

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

  264. 

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

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

  267.     }

  268. };

  269. 

  270. //==============================================================================

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

  272. 

  273.     @tags{DSP}

  274. */

  275. template <>

  276. struct SIMDNativeOps<uint8_t>

  277. {

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

  279.     using vSIMDType = __m256i;

  280. 

  281.     //==============================================================================

  282.     DECLARE_AVX_SIMD_CONST (uint8_t, kHighBit);

  283.     DECLARE_AVX_SIMD_CONST (uint8_t, kAllBitsSet);

  284. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  305.     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); }

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

  307. 

  308.     //==============================================================================

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

  310.     {

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

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

  313. 

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

  315.         {

  316.             lo = _mm256_hadd_epi16 (lo, lo);

  317.             hi = _mm256_hadd_epi16 (hi, hi);

  318.         }

  319. 

  320.        #if JUCE_GCC

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

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

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

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

  325.        #else

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

  327. 

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

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

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

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

  332.        #endif

  333.     }

  334. 

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

  336.     {

  337.         // unpack and multiply

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

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

  340. 

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

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

  343.     }

  344. };

  345. 

  346. //==============================================================================

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

  348. 

  349.     @tags{DSP}

  350. */

  351. template <>

  352. struct SIMDNativeOps<int16_t>

  353. {

  354.     //==============================================================================

  355.     using vSIMDType = __m256i;

  356. 

  357.     //==============================================================================

  358.     DECLARE_AVX_SIMD_CONST (int16_t, kAllBitsSet);

  359. 

  360.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  381.     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); }

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

  383. 

  384.     //==============================================================================

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

  386.     {

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

  388.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  389.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  390. 

  391.        #if JUCE_GCC

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

  393.        #else

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

  395. 

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

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

  398.        #endif

  399.     }

  400. };

  401. 

  402. //==============================================================================

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

  404. 

  405.     @tags{DSP}

  406. */

  407. template <>

  408. struct SIMDNativeOps<uint16_t>

  409. {

  410.     //==============================================================================

  411.     using vSIMDType = __m256i;

  412. 

  413.     //==============================================================================

  414.     DECLARE_AVX_SIMD_CONST (uint16_t, kHighBit);

  415.     DECLARE_AVX_SIMD_CONST (uint16_t, kAllBitsSet);

  416. 

  417.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  439.     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); }

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

  441. 

  442.     //==============================================================================

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

  444.     {

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

  446.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  447.         tmp = _mm256_hadd_epi16 (tmp, tmp);

  448. 

  449.        #if JUCE_GCC

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

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

  452.        #else

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

  454. 

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

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

  457.        #endif

  458.     }

  459. };

  460. 

  461. //==============================================================================

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

  463. 

  464.     @tags{DSP}

  465. */

  466. template <>

  467. struct SIMDNativeOps<int32_t>

  468. {

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

  470.     using vSIMDType = __m256i;

  471. 

  472.     //==============================================================================

  473.     DECLARE_AVX_SIMD_CONST (int32_t, kAllBitsSet);

  474. 

  475.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  496.     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); }

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

  498. 

  499.     //==============================================================================

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

  501.     {

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

  503.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  504. 

  505.        #if JUCE_GCC

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

  507.        #else

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

  509. 

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

  511.        #endif

  512.     }

  513. };

  514. 

  515. //==============================================================================

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

  517. 

  518.     @tags{DSP}

  519. */

  520. template <>

  521. struct SIMDNativeOps<uint32_t>

  522. {

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

  524.     using vSIMDType = __m256i;

  525. 

  526.     //==============================================================================

  527.     DECLARE_AVX_SIMD_CONST (uint32_t, kAllBitsSet);

  528.     DECLARE_AVX_SIMD_CONST (uint32_t, kHighBit);

  529. 

  530.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  552.     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); }

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

  554. 

  555.     //==============================================================================

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

  557.     {

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

  559.         tmp = _mm256_hadd_epi32 (tmp, tmp);

  560. 

  561.        #if JUCE_GCC

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

  563.        #else

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

  565. 

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

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

  568.        #endif

  569.     }

  570. };

  571. 

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

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

  574. 

  575.     @tags{DSP}

  576. */

  577. template <>

  578. struct SIMDNativeOps<int64_t>

  579. {

  580.     //==============================================================================

  581.     using vSIMDType = __m256i;

  582. 

  583.     //==============================================================================

  584.     DECLARE_AVX_SIMD_CONST (int64_t, kAllBitsSet);

  585. 

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

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

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

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

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

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

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

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

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

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

  596.     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)); }

  597.     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)); }

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

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

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

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

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

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

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

  605.     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); }

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

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

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

  609. };

  610. 

  611. //==============================================================================

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

  613. 

  614.     @tags{DSP}

  615. */

  616. template <>

  617. struct SIMDNativeOps<uint64_t>

  618. {

  619.     //==============================================================================

  620.     using vSIMDType = __m256i;

  621. 

  622.     //==============================================================================

  623.     DECLARE_AVX_SIMD_CONST (uint64_t, kAllBitsSet);

  624.     DECLARE_AVX_SIMD_CONST (uint64_t, kHighBit);

  625. 

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

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

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

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

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

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

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

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

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

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

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

  637.     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)); }

  638.     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)); }

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

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

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

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

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

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

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

  646.     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); }

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

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

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

  650. };

  651. 

  652. #endif

  653. 

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

  655.  #pragma GCC diagnostic pop

  656. #endif

  657. 

  658. } // namespace dsp

  659. } // namespace juce