SwiftのSIMDとその利用方法

iOSDC2024 Day 1 15:40~ Track D

Transcript

1. 
   LOUL 4XJGUͷ4*.%ͱͦͷར༻ํ๏ J04%$
   JPTED
   E
2. wLOULʢΧωλΧʣ
   wגࣜձࣾϝϧΧϦ ࣗݾ঺հ
3. ͸͡Ίʹ w4*.%໋ྩ
   w$16ʹଘࡏ͢Δߴ଎ԋࢉػೳ
   wը૾ɾԻ੠ॲཧ
4. w0QFO$7
   OVNQZ
   w"DDFMFSBUF
   'SBNFXPSL
   wW*NBHF
   W%41
   wʮ$16্Ͱߴ଎ͳॲཧʯ͸େମ4*.% ͸͡Ίʹ
5. w4XJGUʹ΋ଘࡏ
   w3FBMJUZ,JUͰ࢖͏
   wߴ଎ʹԋࢉ͕Ͱ͖Δݻఆ௕഑ྻ ͸͡Ίʹ let a = SIMD4<Float>(1.0, 2.0, 3.0, 4.0) let b = SIMD4<Float>(5.0, 6.0, 7.0, 8.0) let add = a * b // 6.0, 8.0, 10.0, 12.0
6. ໨࣍ w4*.%ͱ͸
   w4XJGUͷ4*.%
   wϢʔεέʔε
   wܕͱؔ܎
   w໰୊఺
7. 4*.%ͱ͸
8. 4*.%ʢγϜσΟʔʣͱ͸ w4JOHMF
   *OTUSVDUJPO
   .VMUJQMF
   %BUB
   wҰͭͷ໋ྩͰෳ਺ͷσʔλΛॲཧ͢Δ
   w֓೦ɾQSPUPDPM
9. 4*.%ʢγϜσΟʔʣͱ͸ w4JOHMF
   *OTUSVDUJPO
   .VMUJQMF
   %BUB
   wҰͭͷ໋ྩͰෳ਺ͷσʔλΛॲཧ͢Δ
   w֓೦ɾQSPUPDPM
   w$16ͷ4*.%໋ྩ
   (16
10. 4*.%ʢγϜσΟʔʣͱ͸ w4JOHMF
    *OTUSVDUJPO
    .VMUJQMF
    %BUB
    wҰͭͷ໋ྩͰෳ਺ͷσʔλΛॲཧ͢Δ
    w֓೦ɾQSPUPDPM
    w$16ͷ4*.%໋ྩ
    (16
11. $16 εΧϥԋࢉ໋ྩ wεΧϥԋࢉ໋ྩʢ௨ৗͷԋࢉ໋ྩʣ
    w࠷େCJUͷσʔλؒͰԋࢉ
    wCJUͷϨδελ
12. $16 εΧϥԋࢉͷΠϝʔδ
13. $16 εΧϥԋࢉͷΠϝʔδ        
14. $16 εΧϥԋࢉͷΠϝʔδ        
15. $16 εΧϥԋࢉͷΠϝʔδ         
16. $16 εΧϥԋࢉͷΠϝʔδ         
17.  $16 εΧϥԋࢉͷΠϝʔδ       
18.   $16 εΧϥԋࢉͷΠϝʔδ     
19.    $16 εΧϥԋࢉͷΠϝʔδ   
20.    $16 εΧϥԋࢉͷΠϝʔδ 
21. $16 4*.%ԋࢉ w4*.%ԋࢉ໋ྩ
    ʢϕΫτϧԋࢉ໋ྩʣ
    wCJUҎ্ͷσʔλؒͰԋࢉ
    wCJUҎ্ͷϨδελͱઐ༻ճ࿏
22. $16 4*.%ԋࢉ w4*.%ԋࢉ໋ྩ
    ʢϕΫτϧԋࢉ໋ྩʣ
    wCJUҎ্ͷσʔλؒͰԋࢉ
    wCJUҎ্ͷϨδελͱઐ༻ճ࿏
    wϨδελʹෳ਺ͷ஋Λ٧Ίͯԋࢉ
23. $16 4*.%ԋࢉͷΠϝʔδ
    Y@
    "79
24. $16         4*.%ԋࢉͷΠϝʔδ
    Y@
    "79
25. $16         4*.%ԋࢉͷΠϝʔδ
    Y@
    "79
26. $16         4*.%ԋࢉͷΠϝʔδ
    Y@
    "79    
27.    $16  4*.%ԋࢉͷΠϝʔδ
    Y@
    "79
28. $16 4*.%ԋࢉͷΠϝʔδ
    Y@
    "79
29. $16 4*.%ԋࢉͷ࠷େϏοτ௕ w44&
    Y@ 
    
    CJU
    w"79
    Y@ 
    
    CJU
    w"79
    Y@ 
    
    CJU
    w/&0/
    "3. 
    
    CJU
30. 4XJGUͷ4*.%
31. 4XJGUͷ4*.% w4XJGU
    ʙ
    wඪ४ϥΠϒϥϦ
32. 4XJGUͷ4*.%
    Ϣʔεέʔε
33. 4XJGUͷ4*.% wύϑΥʔϚϯε͕ඞཁʹͳΔॲཧ
    wը૾ɾԻ੠ॲཧ Ϣʔεέʔε
    4*.%ϓϩάϥϛϯά
34. 4XJGUͷ4*.% Ϣʔεέʔε
    4*.%ϓϩάϥϛϯάʢը૾ͷϘοΫεϑΟϧλʣ let input: [[UInt8]] = … var output: [[UInt8]] = … for j in 0..<height { for i in 0..<width { var sum = 0 for ry in -R...R { for rx in -R...R { sum += Int(input[j + ry][i + rx]) } } output[j][i] = UInt8(sum / (2 * R + 1) * (2 * R + 1)) } }
35. 4XJGUͷ4*.% Ϣʔεέʔε
    4*.%ϓϩάϥϛϯάʢը૾ͷϘοΫεϑΟϧλʣ let input: [[UInt8]] = … var output: [[UInt8]] = … for j in 0..<height { for i in 0..<width { var sum = 0 for ry in -R...R { for rx in -R...R { sum += Int(input[j + ry][i + rx]) } } output[j][i] = UInt8(sum / (2 * R + 1) * (2 * R + 1)) } }
36. 4XJGUͷ4*.% Ϣʔεέʔε
    4*.%ϓϩάϥϛϯάʢը૾ͷϘοΫεϑΟϧλʣ for y in 0..<height { do { var x = 0 while x < width { var sum = SIMD8<UInt16>.zero for k in 0..<L { let startIndex = x + k sum &+= SIMD8<UInt16>(yresult[startIndex..<startIndex+8]) } sum /= weightSIMD for k in 0..<8 { output[y][x+k] = sum[k] }
37. 4XJGUͷ4*.% Ϣʔεέʔε
    4*.%ϓϩάϥϛϯάʢը૾ͷϘοΫεϑΟϧλʣ IUUQTHJUIVCDPNLOULZNUTXJGUJNBHFQSPDFTTJOHTBNQMF IUUQTTQFBLFSEFDLDPNLOULZNUNFFUIJHIQFSGPSNBODFJNBHF fi MUFSJOHJOTXJGU
38. 4XJGUͷ4*.% w࠲ඪදݱʢओʹ"3
    73
    $(ͳͲʣ
    w%΍%ͷ࠲ඪදݱ͕Ͱ͖Δσʔλߏ଄
    wࠓ·Ͱ4XJGUඪ४Ͱଘࡏ͠ͳ͔ͬͨ Ϣʔεέʔε
    ࠲ඪදݱ let point = SIMD2<Float>(x: 1.0, y: 2.0) let position = SIMD3<Float>(x: 1.0, y: 2.0, z: 3.0)
39. 4XJGUͷ4*.% w3FBMJUZ,JU)BT5SBOTGPSN
    
    w4DFOF,JU4$//PEF Ϣʔεέʔε
    ࠲ඪදݱ import RealityKit let modelEntity = ModelEntity() modelEntity.position = SIMD3<Float>(0, 1, 2)
40. 4XJGUͷ4*.% wཁૉຖͷԋࢉΛαϙʔτ Ϣʔεέʔε
    ࠲ඪදݱ let a = SIMD2<Float>(x: 1.0, y: 2.0) let b = SIMD2<Float>(x: 1.0, y: 3.0) a + b // (2.0, 5.0) ཁૉຖͷ࿨ a .== b // (true, false) ཁૉຖͷҰக pointwiseMax(a, b) // (1.0, 3.0) ཁૉຖͷmax
41. 4XJGUͷ4*.%
    ܕͱؔ܎
42. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
43. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
44. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.%
    wQSPUPDPM
    4*.%
    4*.%4UPSBHF ܕͱؔ܎
45. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
46. 4XJGUͷ4*.% wԋࢉ͕Ͱ͖Δ let a = SIMD4<Float>(1.0, 2.0, 3.0, 4.0) let b = SIMD4<Float>(1.0, 2.0, 3.0, 4.0) a + b // SIMD4<Float>(2.0, 4.0, 6.0, 8.0) a - b // SIMD4<Float>(0.0, 0.0, 0.0, 0.0) a * b // SIMD4<Float>(1.0, 4.0, 9.0, 16.0) a / b // SIMD4<Float>(1.0, 1.0, 1.0, 1.0) ࢖༻ྫ
    4*.%O
47. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.% ܕͱؔ܎
    TUSVDU
    4*.%O
48. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wO
    ཁૉ਺
           let a = SIMD2<Float>(1.0, 2.0) let b = SIMD3<Float>(1.0, 2.0, 3.0) let c = SIMD4<Float>(1.0, 2.0, 3.0, 4.0) ܕͱؔ܎
    TUSVDU
    4*.%O
49. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wO
    ཁૉ਺
           public struct SIMD2<Scalar: SIMDScalar>: SIMD public struct SIMD3<Scalar: SIMDScalar>: SIMD public struct SIMD4<Scalar: SIMDScalar>: SIMD ... ܕͱؔ܎
    TUSVDU
    4*.%O
50. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wO
    ཁૉ਺
          
    w4DBMBS
    ཁૉͷܕ let a = SIMD2<Float>(1.0, 2.0) let b = SIMD2<Double>(1.0, 2.0) let c = SIMD4<Int32>(1, 2, 3, 4) let d = SIMD4<UInt32>(1, 2, 3, 4) ܕͱؔ܎
    TUSVDU
    4*.%O
51. 4XJGUͷ4*.% // ϥϕϧ͋Γ (SIMD2,3,4ͷΈ) SIMD3<Float>(x: 1.0, y: 2.0, z: 3.0) // ϥϕϧͳ͠ SIMD3<Float>(1.0, 2.0, 3.0) // arrayLiteral let a: SIMD3<Int> = [1, 2, 3] // Sequence SIMD3<Int>([1,2,3]) let array = [1,2,3,4,5,6] SIMD3<Int>(array[1..<4]) SIMD3<Int>(array.suffix(3)) ࢖༻ྫ
    4*.%Oͷੜ੒
52. 4XJGUͷ4*.% // ֦ு let xy = SIMD2<Int>(10, 20) let xyz = SIMD3<Int>(xy, 30) // SIMD3<Int>(10, 20, 30) ࢖༻ྫ
    4*.%Oͷੜ੒
53. 4XJGUͷ4*.% // ֦ு let xy = SIMD2<Int>(10, 20) let xyz = SIMD3<Int>(xy, 30) // SIMD3<Int>(10, 20, 30) // γϟοϑϧ let yxz = xyz[.init(1, 0, 2)] // SIMD3<Int>(20, 10, 30) ࢖༻ྫ
    4*.%Oͷੜ੒
54. 4XJGUͷ4*.% // ֦ு let xy = SIMD2<Int>(10, 20) let xyz = SIMD3<Int>(xy, 30) // SIMD3<Int>(10, 20, 30) // γϟοϑϧ let yxz = xyz[.init(1, 0, 2)] // SIMD3<Int>(20, 10, 30) // ෦෼நग़ let xy = xyz[.init(0, 1)] // SIMD2<Int>(10, 20) ࢖༻ྫ
    4*.%Oͷੜ੒
55. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
56. 4XJGUͷ4*.% wQSPUPDPM
    4*.%4DBMBS
    w4*.%Oͷཁૉͱͯ͠࢖͑Δܕ ܕͱؔ܎
    QSPUPDPM
    4*.%4DBMBS
57. 4XJGUͷ4*.% wQSPUPDPM
    4*.%4DBMBS
    w4*.%Oͷཁૉͱͯ͠࢖͑Δܕ ܕͱؔ܎
    QSPUPDPM
    4*.%4DBMBS Double Float Float16 Int Int8 Int16 Int32 Int64 UInt UInt8 UInt16 UInt32 UInt64
58. 4XJGUͷ4*.% wQSPUPDPM
    4*.%4DBMBS
    w'MPBUJOH1PJOU
    w'JYFE8JEUI*OUFHFS ܕͱؔ܎
    QSPUPDPM
    4*.%4DBMBS Int Int8 Int16 Int32 Int64 UInt UInt8 UInt16 UInt32 UInt64 Double Float Float16
59. 4XJGUͷ4*.% w'JYFE8JEUI*OUFHFS޲͚ૢ࡞ let a = SIMD3<UInt8>(x: 1, y: 2, z: 3) let b = SIMD3<UInt8>(x: 1, y: 1, z: 1) ~a // SIMD3<UInt8>(254, 253, 252) a & b // SIMD3<UInt8>(1, 0, 1) ܕͱؔ܎
    QSPUPDPM
    4*.%4DBMBS
60. 4XJGUͷ4*.% w'JYFE8JEUI*OUFHFS޲͚ૢ࡞
    w'MPBUJOH1PJOU޲͚ૢ࡞ let a = SIMD3<Float>(x: 2.0, y: 3.0, z: 4.0) a.squareRoot() // SIMD3<Float>(1.4142135, 1.7320508, 2.0) let a = SIMD3<Float>(x: 1.0, y: 1.4, z: 1.8) a.rounded(.up) // SIMD3<Float>(1.0, 2.0, 2.0) let a = SIMD3<UInt8>(x: 1, y: 2, z: 3) let b = SIMD3<UInt8>(x: 1, y: 1, z: 1) ~a // SIMD3<UInt8>(254, 253, 252) a & b // SIMD3<UInt8>(1, 0, 1) ܕͱؔ܎
    QSPUPDPM
    4*.%4DBMBS
61. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
62. 4XJGUͷ4*.% wQSPUPDPM
    4*.%
    w4*.%OΛந৅Խ
    wWBS
    TDBMBS$PVOU
    *OU
    wGVOD
    TVCTDSJQU JOEFY
    *OU ܕͱؔ܎
    QSPUPDPM
    4*.%
63. wશͯͷ4*.%OͰ࢖͑Δؔ਺ 4XJGUͷ4*.% func abs<Vector: SIMD>(_ simd: Vector) -> Vector where Vector.Scalar: FloatingPoint { simd.replacing(with: -simd, where: simd .< 0) } let simd2 = SIMD2<Float>(1.0, -2.0) abs(simd2) // SIMD2<Float>(1.0, 2.0) let simd3 = SIMD3<Float>(1.0, -2.0, 3.0) abs(simd3) // SIMD2<Float>(1.0, 2.0, 3.0) ࢖༻ྫ
    QSPUPDPM
    4*.%
64. wશͯͷ4*.%OͰ࢖͑Δؔ਺ 4XJGUͷ4*.% extension SIMD where Scalar: SIMDScalar { var xy: SIMD2<Scalar> { self[.init(0, 1)] } var xyz: SIMD3<Scalar> { self[.init(0, 1, 2)] } } let simd4 = SIMD4<Float>(0.0, 1.0, 2.0, 3.0) let xy = simd4.xy // SIMD2<Float>(0.0, 1.0) let xyz = simd4.xyz // SIMD3<Float>(0.0, 1.0, 2.0) ࢖༻ྫ
    QSPUPDPM
    4*.%
65. 4XJGUͷ4*.% wTUSVDU
    4*.%O4DBMBS
    4*.%4DBMBS
    4*.%
    wTUSVDU
    4*.%.BTL4UPSBHF
    4*.%
    4*.% ܕͱؔ܎
66. 4XJGUͷ4*.% wTUSVDU
    4*.%.BTL4UPSBHF
    wཁૉຖͷൺֱ݁Ռ ܕͱؔ܎
    4*.%.BTL let a = SIMD3<UInt8>(1, 2, 3) let b = SIMD3<UInt8>(3, 2, 1) a == b // false a .== b // SIMDMask<SIMD3<Int8>(false, true, false))
67. 4XJGUͷ4*.% wTUSVDU
    4*.%.BTL ܕͱؔ܎
    4*.%.BTL let a = SIMD4<Int32>(1, -1, 2, -2) let mask = a.<0 //SIMDMask<SIMD4<Int32>>(false,true,false,true) // mask͕trueͷཁૉ͚ͩ`-a`Ͱஔ͖׵͑Δ a.replacing(with: -a, where: mask) // SIMD4<Int32>(1, 1, 2, 2) // trueͳ΋ͷ͕ҰͭͰ΋͋Δ͔ any(mask) // true // ͢΂ͯtrue͔ all(mask) // false
68. 4XJGUͷ4*.% wTUSVDU
    4*.%O
    ԋࢉ͕Ͱ͖Δݻఆ௕഑ྻ
    wTUSVDU
    4*.%.BTL
    4*.%Oͷൺֱ݁Ռ
    wQSPUPDPM
    4*.%4DBMBS
    4*.%Oʹ࢖͑Δཁૉ
    wQSPUPDPM
    4*.%
    4*.%OΛந৅Խ ܕͱؔ܎
69. extension SIMD2 where Scalar == UInt8 { public static func &+(a: Self, b: Self) -> Self { Self(Builtin.add_Vec2xInt8(a._storage._value, b._storage._value)) } public static func &-(a: Self, b: Self) -> Self { Self(Builtin.sub_Vec2xInt8(a._storage._value, b._storage._value)) } public static func &*(a: Self, b: Self) -> Self { Self(Builtin.mul_Vec2xInt8(a._storage._value, b._storage._value)) } } 4XJGUͷ4*.% Ͳ͏΍ͬͯ4*.%໋ྩΛൃߦ͍ͯ͠Δʁ
70. w#VJMUJOʢ૊ΈࠐΈʣؔ਺
    w--7.্ͷ4*.%໋ྩʹม׵
    w--7.͕ΞηϯϒϦʹ4*.%໋ྩΛग़ྗ 4XJGUͷ4*.% Ͳ͏΍ͬͯ4*.%໋ྩΛൃߦ͍ͯ͠Δʁ
71. 4XJGUͷ4*.%
    ໰୊఺
72. 4XJGUͷ4*.% w૊ΈࠐΈؔ਺Ͱ࣮૷͍ͯ͠ͳ͍ॲཧ͕ଟ͍ ໰୊఺
73. 4XJGUͷ4*.% w૊ΈࠐΈؔ਺Ͱ࣮૷͍ͯ͠ͳ͍ॲཧ͕ଟ͍
    w'MPBUJOH1PJOU͸#VJMUJOͷ࣮૷͕ͳ͍ʂ ໰୊఺ // Implementations of floating-point operations. These should eventually all // be replaced with @_semantics to lower directly to vector IR nodes. extension SIMD where Scalar: FloatingPoint { public static func +(a: Self, b: Self) -> Self { var result = Self() for i in result.indices { result[i] = a[i] + b[i] } return result } }
74. 4XJGUͷ4*.% w૊ΈࠐΈؔ਺Ͱ࣮૷͍ͯ͠ͳ͍ॲཧ͕ଟ͍
    w'MPBUJOH1PJOU͸#VJMUJOͷ࣮૷͕ͳ͍ʂ
    wʢҰԠʣࣗಈϕΫτϧԽ͸ޮ͖΍͍͢ ໰୊఺ extension SIMD where Scalar: FloatingPoint { public static func +(a: Self, b: Self) -> Self { var result = Self() for i in result.indices { result[i] = a[i] + b[i] } return result } }
75. 4XJGUͷ4*.% w.BUSJY
    RVBUFSOJPO͕ඇରԠ
    wճసΛදݱ
    w"DDFMFSBUFTJNEΛ࢖͏ඞཁ͋Γ ໰୊఺
76. ·ͱΊ w4XJGUͷ4*.%
    wߴ଎ʹԋࢉͰ͖Δݻఆ௕഑ྻ
    w༻్
    wը૾ॲཧͳͲߴ଎ͳॲཧ
    w࠲ඪදݱ
77. ·ͱΊ w4XJGUͷ4*.%
    wߴ଎ʹԋࢉͰ͖Δݻఆ௕഑ྻ
    w༻్
    wը૾ॲཧͳͲߴ଎ͳॲཧʢະ࣮૷ଟ਺ʣ
    w࠲ඪදݱ
78. ·ͱΊ w4XJGUͷ4*.%
    wߴ଎ʹԋࢉͰ͖Δݻఆ௕഑ྻ
    w༻్
    wը૾ॲཧͳͲߴ଎ͳॲཧʢະ࣮૷ଟ਺ʣ
    w࠲ඪදݱ
79. ࢀߟ
80. ࢀߟ w4&
    4*.%
    7FDUPST
    w IUUQTHJUIVCDPNBQQMFTXJGUFWPMVUJPOCMPCNBJOQSPQPTBMTTJNENE
    w4&
    4*.%
    BEEJUJPOT
    w IUUQTHJUIVCDPNTXJGUMBOHTXJGUFWPMVUJPOCMPCNBJOQSPQPTBMTTJNEBEEJUJPOTNE
    w"3.
    $16ʹ͓͚Δ4*.%Λ༻͍ͨߴ଎ܭࢉೖ໳
    w IUUQTXXXTMJEFTIBSFOFUTMJEFTIPXBSNDQVTJNE