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.
7852 Commits
9 Branches
8.6GB
Tree: 71d10e750a
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '71d10e750a'
${ noResults }
JUCE/modules/juce_dsp/native/juce_neon_SIMDNativeOps.h

395 lines
28KB
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. #ifdef _MSC_VER

  35.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  37. 

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

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

  40. 

  41. #else

  42.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  44. 

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

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

  47. 

  48. #endif

  49. 

  50. template <typename type>

  51. struct SIMDNativeOps;

  52. 

  53. //==============================================================================

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

  55. template <>

  56. struct SIMDNativeOps<float>

  57. {

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

  59.     typedef float32x4_t vSIMDType;

  60.     typedef uint32x4_t vMaskType;

  61.     typedef SIMDFallbackOps<float, vSIMDType> fb;

  62. 

  63.     //==============================================================================

  64.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  65.     DECLARE_NEON_SIMD_CONST (int32_t, kEvenHighBit);

  66.     DECLARE_NEON_SIMD_CONST (float, kOne);

  67. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  88.     static forcedinline float sum (vSIMDType a) noexcept { return fb::sum (a); }

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

  90. 

  91.     //==============================================================================

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

  93.     {

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

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

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

  97.     }

  98. };

  99. 

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

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

  102.     so we need to emulate this.

  103. */

  104. template <>

  105. struct SIMDNativeOps<double>

  106. {

  107.     //==============================================================================

  108.     typedef struct { double values [2]; } vSIMDType;

  109.     typedef SIMDFallbackOps<double, vSIMDType> fb;

  110. 

  111.     static forcedinline vSIMDType expand (double s) noexcept                     { return fb::expand (s); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  130. };

  131. 

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

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

  134. template <>

  135. struct SIMDNativeOps<int8_t>

  136. {

  137.     //==============================================================================

  138.     typedef int8x16_t vSIMDType;

  139.     typedef SIMDFallbackOps<int8_t, vSIMDType> fb;

  140. 

  141.     //==============================================================================

  142.     DECLARE_NEON_SIMD_CONST (int8_t, kAllBitsSet);

  143. 

  144.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  162. };

  163. 

  164. //==============================================================================

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

  166. template <>

  167. struct SIMDNativeOps<uint8_t>

  168. {

  169.     //==============================================================================

  170.     typedef uint8x16_t vSIMDType;

  171.     typedef SIMDFallbackOps<uint8_t, vSIMDType> fb;

  172. 

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

  174.     DECLARE_NEON_SIMD_CONST (uint8_t, kAllBitsSet);

  175. 

  176.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  194. };

  195. 

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

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

  198. template <>

  199. struct SIMDNativeOps<int16_t>

  200. {

  201.     //==============================================================================

  202.     typedef int16x8_t vSIMDType;

  203.     typedef SIMDFallbackOps<int16_t, vSIMDType> fb;

  204. 

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

  206.     DECLARE_NEON_SIMD_CONST (int16_t, kAllBitsSet);

  207. 

  208.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  226. };

  227. 

  228. 

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

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

  231. template <>

  232. struct SIMDNativeOps<uint16_t>

  233. {

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

  235.     typedef uint16x8_t vSIMDType;

  236.     typedef SIMDFallbackOps<uint16_t, vSIMDType> fb;

  237. 

  238.     //==============================================================================

  239.     DECLARE_NEON_SIMD_CONST (uint16_t, kAllBitsSet);

  240. 

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

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

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

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

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

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

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

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

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

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

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

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

  253.     static forcedinline vSIMDType equal (vSIMDType a, vSIMDType b) noexcept                  { return (vSIMDType) vceqq_u16 (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_u16 (a, b); }

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

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

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

  259. };

  260. 

  261. //==============================================================================

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

  263. template <>

  264. struct SIMDNativeOps<int32_t>

  265. {

  266.     //==============================================================================

  267.     typedef int32x4_t vSIMDType;

  268.     typedef SIMDFallbackOps<int32_t, vSIMDType> fb;

  269. 

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

  271.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  272. 

  273.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  290.     static forcedinline int32_t sum (vSIMDType a) noexcept { return fb::sum (a); }

  291. };

  292. 

  293. 

  294. //==============================================================================

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

  296. template <>

  297. struct SIMDNativeOps<uint32_t>

  298. {

  299.     //==============================================================================

  300.     typedef uint32x4_t vSIMDType;

  301.     typedef SIMDFallbackOps<uint32_t, vSIMDType> fb;

  302. 

  303.     //==============================================================================

  304.     DECLARE_NEON_SIMD_CONST (uint32_t, kAllBitsSet);

  305. 

  306.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  323.     static forcedinline uint32_t sum (vSIMDType a) noexcept { return fb::sum (a); }

  324. };

  325. 

  326. //==============================================================================

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

  328. template <>

  329. struct SIMDNativeOps<int64_t>

  330. {

  331.     //==============================================================================

  332.     typedef int64x2_t vSIMDType;

  333.     typedef SIMDFallbackOps<int64_t, vSIMDType> fb;

  334. 

  335.     //==============================================================================

  336.     DECLARE_NEON_SIMD_CONST (int64_t, kAllBitsSet);

  337. 

  338.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  356. };

  357. 

  358. 

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

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

  361. template <>

  362. struct SIMDNativeOps<uint64_t>

  363. {

  364.     //==============================================================================

  365.     typedef uint64x2_t vSIMDType;

  366.     typedef SIMDFallbackOps<uint64_t, vSIMDType> fb;

  367. 

  368.     //==============================================================================

  369.     DECLARE_NEON_SIMD_CONST (uint64_t, kAllBitsSet);

  370. 

  371.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  389. };

  390. 

  391. #endif

  392. 

  393. } // namespace dsp

  394. } // namespace juce