falkTX
/
Carla
mirror of https://github.com/falkTX/Carla
1
0
Fork 0
Code Releases 42 Activity
Audio plugin host https://kx.studio/carla
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.
6842 Commits
16 Branches
14GB
Tree: 62e17bf0e1
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '62e17bf0e1'
${ noResults }
Carla/source/modules/juce_dsp/native/juce_neon_SIMDNativeOps.h

502 lines
38KB
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

  27. {

  28. namespace dsp

  29. {

  30. 

  31. #ifndef DOXYGEN

  32. 

  33. JUCE_BEGIN_IGNORE_WARNINGS_GCC_LIKE ("-Wignored-attributes")

  34. 

  35. #ifdef _MSC_VER

  36.  #define DECLARE_NEON_SIMD_CONST(type, name) \

  37.     static __declspec(align(16)) const type name [16 / sizeof (type)]

  38. 

  39.  #define DEFINE_NEON_SIMD_CONST(type, class_type, name) \

  40.     __declspec(align(16)) const type SIMDNativeOps<class_type>:: name [16 / sizeof (type)]

  41. 

  42. #else

  43.  #define DECLARE_NEON_SIMD_CONST(type, name) \

  44.     static const type name [16 / sizeof (type)] __attribute__((aligned(16)))

  45. 

  46.  #define DEFINE_NEON_SIMD_CONST(type, class_type, name) \

  47.     const type SIMDNativeOps<class_type>:: name [16 / sizeof (type)] __attribute__((aligned(16)))

  48. 

  49. #endif

  50. 

  51. template <typename type>

  52. struct SIMDNativeOps;

  53. 

  54. //==============================================================================

  55. /** Unsigned 32-bit integer NEON intrinsics.

  56. 

  57.     @tags{DSP}

  58. */

  59. template <>

  60. struct SIMDNativeOps<uint32_t>

  61. {

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

  63.     using vSIMDType = uint32x4_t;

  64.     using fb = SIMDFallbackOps<uint32_t, vSIMDType>;

  65. 

  66.     //==============================================================================

  67.     DECLARE_NEON_SIMD_CONST (uint32_t, kAllBitsSet);

  68. 

  69.     //==============================================================================

  70.     static forcedinline vSIMDType expand (uint32_t s) noexcept                                  { return vdupq_n_u32 (s); }

  71.     static forcedinline vSIMDType load (const uint32_t* a) noexcept                             { return vld1q_u32 (a); }

  72.     static forcedinline void store (vSIMDType value, uint32_t* a) noexcept                      { vst1q_u32 (a, value); }

  73.     static forcedinline uint32_t get (vSIMDType v, size_t i) noexcept                           { return v[i]; }

  74.     static forcedinline vSIMDType set (vSIMDType v, size_t i, uint32_t s) noexcept              { v[i] = s; return v; }

  75.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                       { return vaddq_u32 (a, b); }

  76.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                       { return vsubq_u32 (a, b); }

  77.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                       { return vmulq_u32 (a, b); }

  78.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                   { return vandq_u32 (a, b); }

  79.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                   { return vorrq_u32  (a, b); }

  80.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                   { return veorq_u32 (a, b); }

  81.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept                { return vbicq_u32 (b, a); }

  82.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                                { return bit_notand (a, vld1q_u32 ((uint32_t*) kAllBitsSet)); }

  83.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                       { return vminq_u32 (a, b); }

  84.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                       { return vmaxq_u32 (a, b); }

  85.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                     { return (vSIMDType) vceqq_u32 (a, b); }

  86.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                  { return (sum (notEqual (a, b)) == 0); }

  87.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                  { return bit_not (equal (a, b)); }

  88.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept               { return (vSIMDType) vcgtq_u32 (a, b); }

  89.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept        { return (vSIMDType) vcgeq_u32 (a, b); }

  90.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept  { return vmlaq_u32 (a, b, c); }

  91.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                               { return a; }

  92. 

  93.     static forcedinline uint32_t sum (vSIMDType a) noexcept

  94.     {

  95.         auto rr = vadd_u32 (vget_high_u32 (a), vget_low_u32 (a));

  96.         return vget_lane_u32 (vpadd_u32 (rr, rr), 0);

  97.     }

  98. };

  99. 

  100. //==============================================================================

  101. /** Signed 32-bit integer NEON intrinsics.

  102. 

  103.     @tags{DSP}

  104. */

  105. template <>

  106. struct SIMDNativeOps<int32_t>

  107. {

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

  109.     using vSIMDType = int32x4_t;

  110.     using fb = SIMDFallbackOps<int32_t, vSIMDType>;

  111. 

  112.     //==============================================================================

  113.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  114. 

  115.     //==============================================================================

  116.     static forcedinline vSIMDType expand (int32_t s) noexcept                                   { return vdupq_n_s32 (s); }

  117.     static forcedinline vSIMDType load (const int32_t* a) noexcept                              { return vld1q_s32 (a); }

  118.     static forcedinline void store (vSIMDType value, int32_t* a) noexcept                       { vst1q_s32 (a, value); }

  119.     static forcedinline int32_t get (vSIMDType v, size_t i) noexcept                            { return v[i]; }

  120.     static forcedinline vSIMDType set (vSIMDType v, size_t i, int32_t s) noexcept               { v[i] = s; return v; }

  121.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                       { return vaddq_s32 (a, b); }

  122.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                       { return vsubq_s32 (a, b); }

  123.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                       { return vmulq_s32 (a, b); }

  124.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                   { return vandq_s32 (a, b); }

  125.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                   { return vorrq_s32 (a, b); }

  126.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                   { return veorq_s32 (a, b); }

  127.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept                { return vbicq_s32 (b, a); }

  128.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                                { return bit_notand (a, vld1q_s32 ((int32_t*) kAllBitsSet)); }

  129.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                       { return vminq_s32 (a, b); }

  130.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                       { return vmaxq_s32 (a, b); }

  131.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                     { return (vSIMDType) vceqq_s32 (a, b); }

  132.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                  { return (sum (notEqual (a, b)) == 0); }

  133.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                  { return bit_not (equal (a, b)); }

  134.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept               { return (vSIMDType) vcgtq_s32 (a, b); }

  135.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept        { return (vSIMDType) vcgeq_s32 (a, b); }

  136.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept  { return vmlaq_s32 (a, b, c); }

  137.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                               { return a; }

  138. 

  139.     static forcedinline int32_t sum (vSIMDType a) noexcept

  140.     {

  141.         auto rr = vadd_s32 (vget_high_s32 (a), vget_low_s32 (a));

  142.         rr = vpadd_s32 (rr, rr);

  143.         return vget_lane_s32 (rr, 0);

  144.     }

  145. };

  146. 

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

  148. /** Signed 8-bit integer NEON intrinsics.

  149. 

  150.     @tags{DSP}

  151. */

  152. template <>

  153. struct SIMDNativeOps<int8_t>

  154. {

  155.     //==============================================================================

  156.     using vSIMDType = int8x16_t;

  157.     using fb = SIMDFallbackOps<int8_t, vSIMDType>;

  158. 

  159.     //==============================================================================

  160.     DECLARE_NEON_SIMD_CONST (int8_t, kAllBitsSet);

  161. 

  162.     //==============================================================================

  163.     static forcedinline vSIMDType expand (int8_t s) noexcept                                   { return vdupq_n_s8 (s); }

  164.     static forcedinline vSIMDType load (const int8_t* a) noexcept                              { return vld1q_s8 (a); }

  165.     static forcedinline void store (vSIMDType value, int8_t* a) noexcept                       { vst1q_s8 (a, value); }

  166.     static forcedinline int8_t get (vSIMDType v, size_t i) noexcept                            { return v[i]; }

  167.     static forcedinline vSIMDType set (vSIMDType v, size_t i, int8_t s) noexcept               { v[i] = s; return v; }

  168.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_s8 (a, b); }

  169.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_s8 (a, b); }

  170.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return vmulq_s8 (a, b); }

  171.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return vandq_s8 (a, b); }

  172.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return vorrq_s8 (a, b); }

  173.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return veorq_s8 (a, b); }

  174.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return vbicq_s8 (b, a); }

  175.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_s8 ((int8_t*) kAllBitsSet)); }

  176.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return vminq_s8 (a, b); }

  177.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return vmaxq_s8 (a, b); }

  178.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return (vSIMDType) vceqq_s8 (a, b); }

  179.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return bit_not (equal (a, b)); }

  180.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return (vSIMDType) vcgtq_s8 (a, b); }

  181.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return (vSIMDType) vcgeq_s8 (a, b); }

  182.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<int32_t>::sum ((SIMDNativeOps<int32_t>::vSIMDType) notEqual (a, b)) == 0); }

  183.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return vmlaq_s8 (a, b, c); }

  184.     static forcedinline int8_t sum (vSIMDType a) noexcept                                      { return fb::sum (a); }

  185.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return a; }

  186. };

  187. 

  188. //==============================================================================

  189. /** Unsigned 8-bit integer NEON intrinsics.

  190. 

  191.     @tags{DSP}

  192. */

  193. template <>

  194. struct SIMDNativeOps<uint8_t>

  195. {

  196.     //==============================================================================

  197.     using vSIMDType = uint8x16_t;

  198.     using fb = SIMDFallbackOps<uint8_t, vSIMDType>;

  199. 

  200.     //==============================================================================

  201.     DECLARE_NEON_SIMD_CONST (uint8_t, kAllBitsSet);

  202. 

  203.     //==============================================================================

  204.     static forcedinline vSIMDType expand (uint8_t s) noexcept                                  { return vdupq_n_u8 (s); }

  205.     static forcedinline vSIMDType load (const uint8_t* a) noexcept                             { return vld1q_u8 (a); }

  206.     static forcedinline void store (vSIMDType value, uint8_t* a) noexcept                      { vst1q_u8 (a, value); }

  207.     static forcedinline uint8_t get (vSIMDType v, size_t i) noexcept                           { return v[i]; }

  208.     static forcedinline vSIMDType set (vSIMDType v, size_t i, uint8_t s) noexcept              { v[i] = s; return v; }

  209.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_u8 (a, b); }

  210.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_u8 (a, b); }

  211.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return vmulq_u8 (a, b); }

  212.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return vandq_u8 (a, b); }

  213.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return vorrq_u8 (a, b); }

  214.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return veorq_u8 (a, b); }

  215.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return vbicq_u8 (b, a); }

  216.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_u8 ((uint8_t*) kAllBitsSet)); }

  217.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return vminq_u8 (a, b); }

  218.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return vmaxq_u8 (a, b); }

  219.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return (vSIMDType) vceqq_u8 (a, b); }

  220.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return bit_not (equal (a, b)); }

  221.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return (vSIMDType) vcgtq_u8 (a, b); }

  222.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return (vSIMDType) vcgeq_u8 (a, b); }

  223.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<uint32_t>::sum ((SIMDNativeOps<uint32_t>::vSIMDType) notEqual (a, b)) == 0); }

  224.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return vmlaq_u8 (a, b, c); }

  225.     static forcedinline uint8_t sum (vSIMDType a) noexcept                                     { return fb::sum (a); }

  226.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return a; }

  227. };

  228. 

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

  230. /** Signed 16-bit integer NEON intrinsics.

  231. 

  232.     @tags{DSP}

  233. */

  234. template <>

  235. struct SIMDNativeOps<int16_t>

  236. {

  237.     //==============================================================================

  238.     using vSIMDType = int16x8_t;

  239.     using fb = SIMDFallbackOps<int16_t, vSIMDType>;

  240. 

  241.     //==============================================================================

  242.     DECLARE_NEON_SIMD_CONST (int16_t, kAllBitsSet);

  243. 

  244.     //==============================================================================

  245.     static forcedinline vSIMDType expand (int16_t s) noexcept                                  { return vdupq_n_s16 (s); }

  246.     static forcedinline vSIMDType load (const int16_t* a) noexcept                             { return vld1q_s16 (a); }

  247.     static forcedinline void store (vSIMDType value, int16_t* a) noexcept                      { vst1q_s16 (a, value); }

  248.     static forcedinline int16_t get (vSIMDType v, size_t i) noexcept                           { return v[i]; }

  249.     static forcedinline vSIMDType set (vSIMDType v, size_t i, int16_t s) noexcept              { v[i] = s; return v; }

  250.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_s16 (a, b); }

  251.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_s16 (a, b); }

  252.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return vmulq_s16 (a, b); }

  253.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return vandq_s16 (a, b); }

  254.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return vorrq_s16 (a, b); }

  255.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return veorq_s16 (a, b); }

  256.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return vbicq_s16 (b, a); }

  257.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_s16 ((int16_t*) kAllBitsSet)); }

  258.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return vminq_s16 (a, b); }

  259.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return vmaxq_s16 (a, b); }

  260.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return (vSIMDType) vceqq_s16 (a, b); }

  261.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return bit_not (equal (a, b)); }

  262.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return (vSIMDType) vcgtq_s16 (a, b); }

  263.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return (vSIMDType) vcgeq_s16 (a, b); }

  264.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<int32_t>::sum ((SIMDNativeOps<int32_t>::vSIMDType) notEqual (a, b)) == 0); }

  265.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return vmlaq_s16 (a, b, c); }

  266.     static forcedinline int16_t sum (vSIMDType a) noexcept                                     { return fb::sum (a); }

  267.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return a; }

  268. };

  269. 

  270. 

  271. //==============================================================================

  272. /** Unsigned 16-bit integer NEON intrinsics.

  273. 

  274.     @tags{DSP}

  275. */

  276. template <>

  277. struct SIMDNativeOps<uint16_t>

  278. {

  279.     //==============================================================================

  280.     using vSIMDType = uint16x8_t;

  281.     using fb = SIMDFallbackOps<uint16_t, vSIMDType>;

  282. 

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

  284.     DECLARE_NEON_SIMD_CONST (uint16_t, kAllBitsSet);

  285. 

  286.     //==============================================================================

  287.     static forcedinline vSIMDType expand (uint16_t s) noexcept                                 { return vdupq_n_u16 (s); }

  288.     static forcedinline vSIMDType load (const uint16_t* a) noexcept                            { return vld1q_u16 (a); }

  289.     static forcedinline void store (vSIMDType value, uint16_t* a) noexcept                     { vst1q_u16 (a, value); }

  290.     static forcedinline uint16_t get (vSIMDType v, size_t i) noexcept                          { return v[i]; }

  291.     static forcedinline vSIMDType set (vSIMDType v, size_t i, uint16_t s) noexcept             { v[i] = s; return v; }

  292.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_u16 (a, b); }

  293.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_u16 (a, b); }

  294.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return vmulq_u16 (a, b); }

  295.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return vandq_u16 (a, b); }

  296.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return vorrq_u16 (a, b); }

  297.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return veorq_u16 (a, b); }

  298.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return vbicq_u16 (b, a); }

  299.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_u16 ((uint16_t*) kAllBitsSet)); }

  300.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return vminq_u16 (a, b); }

  301.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return vmaxq_u16 (a, b); }

  302.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return (vSIMDType) vceqq_u16 (a, b); }

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

  304.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return (vSIMDType) vcgtq_u16 (a, b); }

  305.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return (vSIMDType) vcgeq_u16 (a, b); }

  306.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<uint32_t>::sum ((SIMDNativeOps<uint32_t>::vSIMDType) notEqual (a, b)) == 0); }

  307.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return vmlaq_u16 (a, b, c); }

  308.     static forcedinline uint16_t sum (vSIMDType a) noexcept                                    { return fb::sum (a); }

  309.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return a; }

  310. };

  311. 

  312. //==============================================================================

  313. /** Signed 64-bit integer NEON intrinsics.

  314. 

  315.     @tags{DSP}

  316. */

  317. template <>

  318. struct SIMDNativeOps<int64_t>

  319. {

  320.     //==============================================================================

  321.     using vSIMDType = int64x2_t;

  322.     using fb = SIMDFallbackOps<int64_t, vSIMDType>;

  323. 

  324.     //==============================================================================

  325.     DECLARE_NEON_SIMD_CONST (int64_t, kAllBitsSet);

  326. 

  327.     //==============================================================================

  328.     static forcedinline vSIMDType expand (int64_t s) noexcept                                  { return vdupq_n_s64 (s); }

  329.     static forcedinline vSIMDType load (const int64_t* a) noexcept                             { return vld1q_s64 (a); }

  330.     static forcedinline void store (vSIMDType value, int64_t* a) noexcept                      { vst1q_s64 (a, value); }

  331.     static forcedinline int64_t get (vSIMDType v, size_t i) noexcept                           { return v[i]; }

  332.     static forcedinline vSIMDType set (vSIMDType v, size_t i, int64_t s) noexcept              { v[i] = s; return v; }

  333.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_s64 (a, b); }

  334.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_s64 (a, b); }

  335.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return fb::mul (a, b); }

  336.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return vandq_s64 (a, b); }

  337.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return vorrq_s64 (a, b); }

  338.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return veorq_s64 (a, b); }

  339.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return vbicq_s64 (b, a); }

  340.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_s64 ((int64_t*) kAllBitsSet)); }

  341.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return fb::min (a, b); }

  342.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return fb::max (a, b); }

  343.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return fb::equal (a, b); }

  344.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return fb::notEqual (a, b); }

  345.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return fb::greaterThan (a, b); }

  346.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return fb::greaterThanOrEqual (a, b); }

  347.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<int32_t>::sum ((SIMDNativeOps<int32_t>::vSIMDType) notEqual (a, b)) == 0); }

  348.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return fb::multiplyAdd (a, b, c); }

  349.     static forcedinline int64_t sum (vSIMDType a) noexcept                                     { return fb::sum (a); }

  350.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return a; }

  351. };

  352. 

  353. 

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

  355. /** Unsigned 64-bit integer NEON intrinsics.

  356. 

  357.     @tags{DSP}

  358. */

  359. template <>

  360. struct SIMDNativeOps<uint64_t>

  361. {

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

  363.     using vSIMDType = uint64x2_t;

  364.     using fb = SIMDFallbackOps<uint64_t, vSIMDType>;

  365. 

  366.     //==============================================================================

  367.     DECLARE_NEON_SIMD_CONST (uint64_t, kAllBitsSet);

  368. 

  369.     //==============================================================================

  370.     static forcedinline vSIMDType expand (uint64_t s) noexcept                                  { return vdupq_n_u64 (s); }

  371.     static forcedinline vSIMDType load (const uint64_t* a) noexcept                             { return vld1q_u64 (a); }

  372.     static forcedinline void store (vSIMDType value, uint64_t* a) noexcept                      { vst1q_u64 (a, value); }

  373.     static forcedinline uint64_t get (vSIMDType v, size_t i) noexcept                           { return v[i]; }

  374.     static forcedinline vSIMDType set (vSIMDType v, size_t i, uint64_t s) noexcept              { v[i] = s; return v; }

  375.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                       { return vaddq_u64 (a, b); }

  376.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                       { return vsubq_u64 (a, b); }

  377.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                       { return fb::mul (a, b); }

  378.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                   { return vandq_u64 (a, b); }

  379.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                   { return vorrq_u64 (a, b); }

  380.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                   { return veorq_u64 (a, b); }

  381.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept                { return vbicq_u64 (b, a); }

  382.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                                { return bit_notand (a, vld1q_u64 ((uint64_t*) kAllBitsSet)); }

  383.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                       { return fb::min (a, b); }

  384.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                       { return fb::max (a, b); }

  385.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                     { return fb::equal (a, b); }

  386.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                  { return fb::notEqual (a, b); }

  387.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept               { return fb::greaterThan (a, b); }

  388.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept        { return fb::greaterThanOrEqual (a, b); }

  389.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                  { return (SIMDNativeOps<uint32_t>::sum ((SIMDNativeOps<uint32_t>::vSIMDType) notEqual (a, b)) == 0); }

  390.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept  { return fb::multiplyAdd (a, b, c); }

  391.     static forcedinline uint64_t sum (vSIMDType a) noexcept                                     { return fb::sum (a); }

  392.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                               { return a; }

  393. };

  394. 

  395.     //==============================================================================

  396. /** Single-precision floating point NEON intrinsics.

  397. 

  398.     @tags{DSP}

  399. */

  400. template <>

  401. struct SIMDNativeOps<float>

  402. {

  403.     //==============================================================================

  404.     using vSIMDType = float32x4_t;

  405.     using vMaskType = uint32x4_t;

  406.     using fb = SIMDFallbackOps<float, vSIMDType>;

  407. 

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

  409.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  410.     DECLARE_NEON_SIMD_CONST (int32_t, kEvenHighBit);

  411.     DECLARE_NEON_SIMD_CONST (float, kOne);

  412. 

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

  414.     static forcedinline vSIMDType expand (float s) noexcept                                    { return vdupq_n_f32 (s); }

  415.     static forcedinline vSIMDType load (const float* a) noexcept                               { return vld1q_f32 (a); }

  416.     static forcedinline float get (vSIMDType v, size_t i) noexcept                             { return v[i]; }

  417.     static forcedinline vSIMDType set (vSIMDType v, size_t i, float s) noexcept                { v[i] = s; return v; }

  418.     static forcedinline void store (vSIMDType value, float* a) noexcept                        { vst1q_f32 (a, value); }

  419.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return vaddq_f32 (a, b); }

  420.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return vsubq_f32 (a, b); }

  421.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return vmulq_f32 (a, b); }

  422.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return (vSIMDType) vandq_u32 ((vMaskType) a, (vMaskType) b); }

  423.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return (vSIMDType) vorrq_u32 ((vMaskType) a, (vMaskType) b); }

  424.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return (vSIMDType) veorq_u32 ((vMaskType) a, (vMaskType) b); }

  425.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return (vSIMDType) vbicq_u32 ((vMaskType) b, (vMaskType) a); }

  426.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return bit_notand (a, vld1q_f32 ((float*) kAllBitsSet)); }

  427.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return vminq_f32 (a, b); }

  428.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return vmaxq_f32 (a, b); }

  429.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return (vSIMDType) vceqq_f32 (a, b); }

  430.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return bit_not (equal (a, b)); }

  431.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return (vSIMDType) vcgtq_f32 (a, b); }

  432.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return (vSIMDType) vcgeq_f32 (a, b); }

  433.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return (SIMDNativeOps<uint32_t>::sum ((SIMDNativeOps<uint32_t>::vSIMDType) notEqual (a, b)) == 0); }

  434.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return vmlaq_f32 (a, b, c); }

  435.     static forcedinline vSIMDType dupeven (vSIMDType a) noexcept                               { return fb::shuffle<(0 << 0) | (0 << 2) | (2 << 4) | (2 << 6)>     (a); }

  436.     static forcedinline vSIMDType dupodd  (vSIMDType a) noexcept                               { return fb::shuffle<(1 << 0) | (1 << 2) | (3 << 4) | (3 << 6)>     (a); }

  437.     static forcedinline vSIMDType swapevenodd (vSIMDType a) noexcept                           { return fb::shuffle<(1 << 0) | (0 << 2) | (3 << 4) | (2 << 6)> (a); }

  438.     static forcedinline vSIMDType oddevensum (vSIMDType a) noexcept                            { return add (fb::shuffle<(2 << 0) | (3 << 2) | (0 << 4) | (1 << 6)> (a), a); }

  439.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return vcvtq_f32_s32 (vcvtq_s32_f32 (a)); }

  440. 

  441.     //==============================================================================

  442.     static forcedinline vSIMDType cmplxmul (vSIMDType a, vSIMDType b) noexcept

  443.     {

  444.         vSIMDType rr_ir = mul (a, dupeven (b));

  445.         vSIMDType ii_ri = mul (swapevenodd (a), dupodd (b));

  446.         return add (rr_ir, bit_xor (ii_ri, vld1q_f32 ((float*) kEvenHighBit)));

  447.     }

  448. 

  449.     static forcedinline float sum (vSIMDType a) noexcept

  450.     {

  451.         auto rr = vadd_f32 (vget_high_f32 (a), vget_low_f32 (a));

  452.         return vget_lane_f32 (vpadd_f32 (rr, rr), 0);

  453.     }

  454. };

  455. 

  456. //==============================================================================

  457. /** Double-precision floating point NEON intrinsics does not exist in NEON

  458.     so we need to emulate this.

  459. 

  460.     @tags{DSP}

  461. */

  462. template <>

  463. struct SIMDNativeOps<double>

  464. {

  465.     //==============================================================================

  466.     using vSIMDType = struct { double v[2]; };

  467.     using fb = SIMDFallbackOps<double, vSIMDType>;

  468. 

  469.     static forcedinline vSIMDType expand (double s) noexcept                                   { return {{s, s}}; }

  470.     static forcedinline vSIMDType load (const double* a) noexcept                              { return {{a[0], a[1]}}; }

  471.     static forcedinline void store (vSIMDType v, double* a) noexcept                           { a[0] = v.v[0]; a[1] = v.v[1]; }

  472.     static forcedinline double get (vSIMDType v, size_t i) noexcept                            { return v.v[i]; }

  473.     static forcedinline vSIMDType set (vSIMDType v, size_t i, double s) noexcept               { v.v[i] = s; return v; }

  474.     static forcedinline vSIMDType add (vSIMDType a, vSIMDType b) noexcept                      { return {{a.v[0] + b.v[0], a.v[1] + b.v[1]}}; }

  475.     static forcedinline vSIMDType sub (vSIMDType a, vSIMDType b) noexcept                      { return {{a.v[0] - b.v[0], a.v[1] - b.v[1]}}; }

  476.     static forcedinline vSIMDType mul (vSIMDType a, vSIMDType b) noexcept                      { return {{a.v[0] * b.v[0], a.v[1] * b.v[1]}}; }

  477.     static forcedinline vSIMDType bit_and (vSIMDType a, vSIMDType b) noexcept                  { return fb::bit_and (a, b); }

  478.     static forcedinline vSIMDType bit_or  (vSIMDType a, vSIMDType b) noexcept                  { return fb::bit_or  (a, b); }

  479.     static forcedinline vSIMDType bit_xor (vSIMDType a, vSIMDType b) noexcept                  { return fb::bit_xor (a, b); }

  480.     static forcedinline vSIMDType bit_notand (vSIMDType a, vSIMDType b) noexcept               { return fb::bit_notand (a, b); }

  481.     static forcedinline vSIMDType bit_not (vSIMDType a) noexcept                               { return fb::bit_not (a); }

  482.     static forcedinline vSIMDType min (vSIMDType a, vSIMDType b) noexcept                      { return fb::min (a, b); }

  483.     static forcedinline vSIMDType max (vSIMDType a, vSIMDType b) noexcept                      { return fb::max (a, b); }

  484.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                    { return fb::equal (a, b); }

  485.     static forcedinline vSIMDType notEqual (vSIMDType a, vSIMDType b) noexcept                 { return fb::notEqual (a, b); }

  486.     static forcedinline vSIMDType greaterThan (vSIMDType a, vSIMDType b) noexcept              { return fb::greaterThan (a, b); }

  487.     static forcedinline vSIMDType greaterThanOrEqual (vSIMDType a, vSIMDType b) noexcept       { return fb::greaterThanOrEqual (a, b); }

  488.     static forcedinline bool      allEqual (vSIMDType a, vSIMDType b) noexcept                 { return fb::allEqual (a, b); }

  489.     static forcedinline vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept { return fb::multiplyAdd (a, b, c); }

  490.     static forcedinline vSIMDType cmplxmul (vSIMDType a, vSIMDType b) noexcept                 { return fb::cmplxmul (a, b); }

  491.     static forcedinline double sum (vSIMDType a) noexcept                                      { return fb::sum (a); }

  492.     static forcedinline vSIMDType oddevensum (vSIMDType a) noexcept                            { return a; }

  493.     static forcedinline vSIMDType truncate (vSIMDType a) noexcept                              { return fb::truncate (a); }

  494. };

  495. 

  496. #endif

  497. 

  498. JUCE_END_IGNORE_WARNINGS_GCC_LIKE

  499. 

  500. } // namespace dsp

  501. } // namespace juce