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.
8384 Commits
11 Branches
588MB
Tree: 8d8c69beb8
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '8d8c69beb8'
${ noResults }
JUCE/modules/juce_dsp/native/juce_neon_SIMDNativeOps.h

438 lines
32KB
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. /** Unsigned 32-bit integer NEON intrinsics. */

  55. template <>

  56. struct SIMDNativeOps<uint32_t>

  57. {

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

  59.     typedef uint32x4_t vSIMDType;

  60.     typedef SIMDFallbackOps<uint32_t, vSIMDType> fb;

  61. 

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

  63.     DECLARE_NEON_SIMD_CONST (uint32_t, kAllBitsSet);

  64. 

  65.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  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_u32 (a, b); }

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

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

  85.     static forcedinline uint32_t sum (vSIMDType a) noexcept

  86.     {

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

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

  89.     }

  90. };

  91. 

  92. //==============================================================================

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

  94. template <>

  95. struct SIMDNativeOps<int32_t>

  96. {

  97.     //==============================================================================

  98.     typedef int32x4_t vSIMDType;

  99.     typedef SIMDFallbackOps<int32_t, vSIMDType> fb;

  100. 

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

  102.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  103. 

  104.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  124.     static forcedinline int32_t sum (vSIMDType a) noexcept

  125.     {

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

  127.         rr = vpadd_s32 (rr, rr);

  128.         return vget_lane_s32 (rr, 0);

  129.     }

  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 load (const int8_t* a) noexcept                { return vld1q_s8 (a); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  165. };

  166. 

  167. //==============================================================================

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

  169. template <>

  170. struct SIMDNativeOps<uint8_t>

  171. {

  172.     //==============================================================================

  173.     typedef uint8x16_t vSIMDType;

  174.     typedef SIMDFallbackOps<uint8_t, vSIMDType> fb;

  175. 

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

  177.     DECLARE_NEON_SIMD_CONST (uint8_t, kAllBitsSet);

  178. 

  179.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  200. };

  201. 

  202. //==============================================================================

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

  204. template <>

  205. struct SIMDNativeOps<int16_t>

  206. {

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

  208.     typedef int16x8_t vSIMDType;

  209.     typedef SIMDFallbackOps<int16_t, vSIMDType> fb;

  210. 

  211.     //==============================================================================

  212.     DECLARE_NEON_SIMD_CONST (int16_t, kAllBitsSet);

  213. 

  214.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  235. };

  236. 

  237. 

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

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

  240. template <>

  241. struct SIMDNativeOps<uint16_t>

  242. {

  243.     //==============================================================================

  244.     typedef uint16x8_t vSIMDType;

  245.     typedef SIMDFallbackOps<uint16_t, vSIMDType> fb;

  246. 

  247.     //==============================================================================

  248.     DECLARE_NEON_SIMD_CONST (uint16_t, kAllBitsSet);

  249. 

  250.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  271. };

  272. 

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

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

  275. template <>

  276. struct SIMDNativeOps<int64_t>

  277. {

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

  279.     typedef int64x2_t vSIMDType;

  280.     typedef SIMDFallbackOps<int64_t, vSIMDType> fb;

  281. 

  282.     //==============================================================================

  283.     DECLARE_NEON_SIMD_CONST (int64_t, kAllBitsSet);

  284. 

  285.     //==============================================================================

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  306. };

  307. 

  308. 

  309. //==============================================================================

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

  311. template <>

  312. struct SIMDNativeOps<uint64_t>

  313. {

  314.     //==============================================================================

  315.     typedef uint64x2_t vSIMDType;

  316.     typedef SIMDFallbackOps<uint64_t, vSIMDType> fb;

  317. 

  318.     //==============================================================================

  319.     DECLARE_NEON_SIMD_CONST (uint64_t, kAllBitsSet);

  320. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  342. };

  343. 

  344.     //==============================================================================

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

  346. template <>

  347. struct SIMDNativeOps<float>

  348. {

  349.     //==============================================================================

  350.     typedef float32x4_t vSIMDType;

  351.     typedef uint32x4_t vMaskType;

  352.     typedef SIMDFallbackOps<float, vSIMDType> fb;

  353. 

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

  355.     DECLARE_NEON_SIMD_CONST (int32_t, kAllBitsSet);

  356.     DECLARE_NEON_SIMD_CONST (int32_t, kEvenHighBit);

  357.     DECLARE_NEON_SIMD_CONST (float, kOne);

  358. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  383. 

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

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

  386.     {

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

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

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

  390.     }

  391. 

  392.     static forcedinline float sum (vSIMDType a) noexcept

  393.     {

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

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

  396.     }

  397. };

  398. 

  399. //==============================================================================

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

  401.     so we need to emulate this.

  402. */

  403. template <>

  404. struct SIMDNativeOps<double>

  405. {

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

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

  408.     typedef SIMDFallbackOps<double, vSIMDType> fb;

  409. 

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

  411.     static forcedinline vSIMDType load (const double* a) noexcept                { return fb::load (a); }

  412.     static forcedinline void store (vSIMDType value, double* a) noexcept         { fb::store (value, a); }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  432. };

  433. 

  434. #endif

  435. 

  436. } // namespace dsp

  437. } // namespace juce