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.
11016 Commits
9 Branches
3.6GB
Tree: bf51d2c076
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bf51d2c076'
${ noResults }
JUCE/modules/juce_dsp/native/juce_neon_SIMDNativeOps.h

495 lines
38KB
Raw Blame History 

  1. /*

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

  3. 

  4.    This file is part of the JUCE 6 technical preview.

  5.    Copyright (c) 2020 - Raw Material Software Limited

  6. 

  7.    You may use this code under the terms of the GPL v3

  8.    (see www.gnu.org/licenses).

  9. 

  10.    For this technical preview, this file is not subject to commercial licensing.

  11. 

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

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

  14.    DISCLAIMED.

  15. 

  16.   ==============================================================================

  17. */

  18. 

  19. namespace juce

  20. {

  21. namespace dsp

  22. {

  23. 

  24. #ifndef DOXYGEN

  25. 

  26. JUCE_BEGIN_IGNORE_WARNINGS_GCC_LIKE ("-Wignored-attributes")

  27. 

  28. #ifdef _MSC_VER

  29.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  31. 

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

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

  34. 

  35. #else

  36.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  38. 

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

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

  41. 

  42. #endif

  43. 

  44. template <typename type>

  45. struct SIMDNativeOps;

  46. 

  47. //==============================================================================

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

  49. 

  50.     @tags{DSP}

  51. */

  52. template <>

  53. struct SIMDNativeOps<uint32_t>

  54. {

  55.     //==============================================================================

  56.     using vSIMDType = uint32x4_t;

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

  58. 

  59.     //==============================================================================

  60.     DECLARE_NEON_SIMD_CONST (uint32_t, kAllBitsSet);

  61. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  85. 

  86.     static forcedinline uint32_t sum (vSIMDType a) noexcept

  87.     {

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

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

  90.     }

  91. };

  92. 

  93. //==============================================================================

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

  95. 

  96.     @tags{DSP}

  97. */

  98. template <>

  99. struct SIMDNativeOps<int32_t>

  100. {

  101.     //==============================================================================

  102.     using vSIMDType = int32x4_t;

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

  104. 

  105.     //==============================================================================

  106.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  107. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  131. 

  132.     static forcedinline int32_t sum (vSIMDType a) noexcept

  133.     {

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

  135.         rr = vpadd_s32 (rr, rr);

  136.         return vget_lane_s32 (rr, 0);

  137.     }

  138. };

  139. 

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

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

  142. 

  143.     @tags{DSP}

  144. */

  145. template <>

  146. struct SIMDNativeOps<int8_t>

  147. {

  148.     //==============================================================================

  149.     using vSIMDType = int8x16_t;

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

  151. 

  152.     //==============================================================================

  153.     DECLARE_NEON_SIMD_CONST (int8_t, kAllBitsSet);

  154. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  179. };

  180. 

  181. //==============================================================================

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

  183. 

  184.     @tags{DSP}

  185. */

  186. template <>

  187. struct SIMDNativeOps<uint8_t>

  188. {

  189.     //==============================================================================

  190.     using vSIMDType = uint8x16_t;

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

  192. 

  193.     //==============================================================================

  194.     DECLARE_NEON_SIMD_CONST (uint8_t, kAllBitsSet);

  195. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  220. };

  221. 

  222. //==============================================================================

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

  224. 

  225.     @tags{DSP}

  226. */

  227. template <>

  228. struct SIMDNativeOps<int16_t>

  229. {

  230.     //==============================================================================

  231.     using vSIMDType = int16x8_t;

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

  233. 

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

  235.     DECLARE_NEON_SIMD_CONST (int16_t, kAllBitsSet);

  236. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  261. };

  262. 

  263. 

  264. //==============================================================================

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

  266. 

  267.     @tags{DSP}

  268. */

  269. template <>

  270. struct SIMDNativeOps<uint16_t>

  271. {

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

  273.     using vSIMDType = uint16x8_t;

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

  275. 

  276.     //==============================================================================

  277.     DECLARE_NEON_SIMD_CONST (uint16_t, kAllBitsSet);

  278. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  303. };

  304. 

  305. //==============================================================================

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

  307. 

  308.     @tags{DSP}

  309. */

  310. template <>

  311. struct SIMDNativeOps<int64_t>

  312. {

  313.     //==============================================================================

  314.     using vSIMDType = int64x2_t;

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

  316. 

  317.     //==============================================================================

  318.     DECLARE_NEON_SIMD_CONST (int64_t, kAllBitsSet);

  319. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  344. };

  345. 

  346. 

  347. //==============================================================================

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

  349. 

  350.     @tags{DSP}

  351. */

  352. template <>

  353. struct SIMDNativeOps<uint64_t>

  354. {

  355.     //==============================================================================

  356.     using vSIMDType = uint64x2_t;

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

  358. 

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

  360.     DECLARE_NEON_SIMD_CONST (uint64_t, kAllBitsSet);

  361. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  386. };

  387. 

  388.     //==============================================================================

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

  390. 

  391.     @tags{DSP}

  392. */

  393. template <>

  394. struct SIMDNativeOps<float>

  395. {

  396.     //==============================================================================

  397.     using vSIMDType = float32x4_t;

  398.     using vMaskType = uint32x4_t;

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

  400. 

  401.     //==============================================================================

  402.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  403.     DECLARE_NEON_SIMD_CONST (int32_t, kEvenHighBit);

  404.     DECLARE_NEON_SIMD_CONST (float, kOne);

  405. 

  406.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  433. 

  434.     //==============================================================================

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

  436.     {

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

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

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

  440.     }

  441. 

  442.     static forcedinline float sum (vSIMDType a) noexcept

  443.     {

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

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

  446.     }

  447. };

  448. 

  449. //==============================================================================

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

  451.     so we need to emulate this.

  452. 

  453.     @tags{DSP}

  454. */

  455. template <>

  456. struct SIMDNativeOps<double>

  457. {

  458.     //==============================================================================

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

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

  461. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  487. };

  488. 

  489. #endif

  490. 

  491. JUCE_END_IGNORE_WARNINGS_GCC_LIKE

  492. 

  493. } // namespace dsp

  494. } // namespace juce