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.
7639 Commits
11 Branches
2.0GB
Tree: 0b67a40bb9
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '0b67a40bb9'
${ noResults }
JUCE/modules/juce_dsp/native/juce_neon_SIMDNativeOps.h

387 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. #ifndef DOXYGEN

  28. 

  29. #ifdef _MSC_VER

  30.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  32. 

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

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

  35. 

  36. #else

  37.  #define DECLARE_NEON_SIMD_CONST(type, name) \

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

  39. 

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

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

  42. 

  43. #endif

  44. 

  45. template <typename type>

  46. struct SIMDNativeOps;

  47. 

  48. //==============================================================================

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

  50. template <>

  51. struct SIMDNativeOps<float>

  52. {

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

  54.     typedef float32x4_t vSIMDType;

  55.     typedef uint32x4_t vMaskType;

  56.     typedef SIMDFallbackOps<float, vSIMDType> fb;

  57. 

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

  59.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  60.     DECLARE_NEON_SIMD_CONST (int32_t, kEvenHighBit);

  61.     DECLARE_NEON_SIMD_CONST (float, kOne);

  62. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  85. 

  86.     //==============================================================================

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

  88.     {

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

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

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

  92.     }

  93. };

  94. 

  95. //==============================================================================

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

  97.     so we need to emulate this.

  98. */

  99. template <>

  100. struct SIMDNativeOps<double>

  101. {

  102.     //==============================================================================

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

  104.     typedef SIMDFallbackOps<double, vSIMDType> fb;

  105. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  125. };

  126. 

  127. //==============================================================================

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

  129. template <>

  130. struct SIMDNativeOps<int8_t>

  131. {

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

  133.     typedef int8x16_t vSIMDType;

  134.     typedef SIMDFallbackOps<int8_t, vSIMDType> fb;

  135. 

  136.     //==============================================================================

  137.     DECLARE_NEON_SIMD_CONST (int8_t, kAllBitsSet);

  138. 

  139.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  157. };

  158. 

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

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

  161. template <>

  162. struct SIMDNativeOps<uint8_t>

  163. {

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

  165.     typedef uint8x16_t vSIMDType;

  166.     typedef SIMDFallbackOps<uint8_t, vSIMDType> fb;

  167. 

  168.     //==============================================================================

  169.     DECLARE_NEON_SIMD_CONST (uint8_t, kAllBitsSet);

  170. 

  171.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  189. };

  190. 

  191. //==============================================================================

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

  193. template <>

  194. struct SIMDNativeOps<int16_t>

  195. {

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

  197.     typedef int16x8_t vSIMDType;

  198.     typedef SIMDFallbackOps<int16_t, vSIMDType> fb;

  199. 

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

  201.     DECLARE_NEON_SIMD_CONST (int16_t, kAllBitsSet);

  202. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  221. };

  222. 

  223. 

  224. //==============================================================================

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

  226. template <>

  227. struct SIMDNativeOps<uint16_t>

  228. {

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

  230.     typedef uint16x8_t vSIMDType;

  231.     typedef SIMDFallbackOps<uint16_t, vSIMDType> fb;

  232. 

  233.     //==============================================================================

  234.     DECLARE_NEON_SIMD_CONST (uint16_t, kAllBitsSet);

  235. 

  236.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  254. };

  255. 

  256. //==============================================================================

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

  258. template <>

  259. struct SIMDNativeOps<int32_t>

  260. {

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

  262.     typedef int32x4_t vSIMDType;

  263.     typedef SIMDFallbackOps<int32_t, vSIMDType> fb;

  264. 

  265.     //==============================================================================

  266.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  267. 

  268.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  286. };

  287. 

  288. 

  289. //==============================================================================

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

  291. template <>

  292. struct SIMDNativeOps<uint32_t>

  293. {

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

  295.     typedef uint32x4_t vSIMDType;

  296.     typedef SIMDFallbackOps<uint32_t, vSIMDType> fb;

  297. 

  298.     //==============================================================================

  299.     DECLARE_NEON_SIMD_CONST (uint32_t, kAllBitsSet);

  300. 

  301.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  319. };

  320. 

  321. //==============================================================================

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

  323. template <>

  324. struct SIMDNativeOps<int64_t>

  325. {

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

  327.     typedef int64x2_t vSIMDType;

  328.     typedef SIMDFallbackOps<int64_t, vSIMDType> fb;

  329. 

  330.     //==============================================================================

  331.     DECLARE_NEON_SIMD_CONST (int64_t, kAllBitsSet);

  332. 

  333.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  351. };

  352. 

  353. 

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

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

  356. template <>

  357. struct SIMDNativeOps<uint64_t>

  358. {

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

  360.     typedef uint64x2_t vSIMDType;

  361.     typedef SIMDFallbackOps<uint64_t, vSIMDType> fb;

  362. 

  363.     //==============================================================================

  364.     DECLARE_NEON_SIMD_CONST (uint64_t, kAllBitsSet);

  365. 

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

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

  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 vSIMDType multiplyAdd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept  { return fb::multiplyAdd (a, b, c); }

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

  384. };

  385. 

  386. #endif