OLD | NEW |
1 #ifndef Sk4x_DEFINED | 1 #ifndef Sk4x_DEFINED |
2 #define Sk4x_DEFINED | 2 #define Sk4x_DEFINED |
3 | 3 |
4 #include "SkTypes.h" | 4 #include "SkTypes.h" |
5 | 5 |
6 #define SK4X_PREAMBLE 1 | 6 #define SK4X_PREAMBLE 1 |
7 #include "Sk4x_portable.h" | 7 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 8 #include "Sk4x_sse.h" |
| 9 #else |
| 10 #include "Sk4x_portable.h" |
| 11 #endif |
8 #undef SK4X_PREAMBLE | 12 #undef SK4X_PREAMBLE |
9 | 13 |
10 template <typename T> class Sk4x; | 14 template <typename T> class Sk4x; |
11 typedef Sk4x<float> Sk4f; | 15 typedef Sk4x<float> Sk4f; |
12 typedef Sk4x<int32_t> Sk4i; | 16 typedef Sk4x<int32_t> Sk4i; |
13 | 17 |
| 18 // Some Sk4x methods are implemented only for Sk4f or Sk4i. |
| 19 // They might be unavailable, really slow, or just a bad idea. |
| 20 // Talk to mtklein if you find yourself unable to link and |
| 21 // really need one of those methods. |
| 22 |
14 template <typename T> class Sk4x { | 23 template <typename T> class Sk4x { |
15 public: | 24 public: |
16 Sk4x(); // Uninitialized; use Sk4x(0,0,0,0) for zero. | 25 Sk4x(); // Uninitialized; use Sk4x(0,0,0,0) for zero. |
17 Sk4x(T, T, T, T); | 26 Sk4x(T, T, T, T); |
18 | 27 |
19 Sk4x(const Sk4x&); | 28 Sk4x(const Sk4x&); |
20 Sk4x& operator=(const Sk4x&); | 29 Sk4x& operator=(const Sk4x&); |
21 | 30 |
22 static Sk4x Load (const T[4]); | 31 static Sk4x Load (const T[4]); |
23 static Sk4x LoadAligned(const T[4]); | 32 static Sk4x LoadAligned(const T[4]); |
24 | 33 |
25 void store (T[4]) const; | 34 void store (T[4]) const; |
26 void storeAligned(T[4]) const; | 35 void storeAligned(T[4]) const; |
27 | 36 |
28 template <typename Dst> Dst reinterpret() const; | 37 template <typename Dst> Dst reinterpret() const; |
29 template <typename Dst> Dst cast() const; | 38 template <typename Dst> Dst cast() const; |
30 | 39 |
31 bool allTrue() const; | 40 bool allTrue() const; |
32 bool anyTrue() const; | 41 bool anyTrue() const; |
33 | 42 |
34 Sk4x bitNot() const; | 43 Sk4x bitNot() const; |
35 Sk4x bitAnd(const Sk4x&) const; | 44 Sk4x bitAnd(const Sk4x&) const; |
36 Sk4x bitOr(const Sk4x&) const; | 45 Sk4x bitOr(const Sk4x&) const; |
| 46 // TODO: Sk4x bitAndNot(const Sk4x&) const; is efficient in SSE. |
37 Sk4x add(const Sk4x&) const; | 47 Sk4x add(const Sk4x&) const; |
38 Sk4x subtract(const Sk4x&) const; | 48 Sk4x subtract(const Sk4x&) const; |
39 Sk4x multiply(const Sk4x&) const; | 49 Sk4x multiply(const Sk4x&) const; |
40 Sk4x divide(const Sk4x&) const; | 50 Sk4x divide(const Sk4x&) const; |
41 | 51 |
42 Sk4i equal(const Sk4x&) const; | 52 Sk4i equal(const Sk4x&) const; |
43 Sk4i notEqual(const Sk4x&) const; | 53 Sk4i notEqual(const Sk4x&) const; |
44 Sk4i lessThan(const Sk4x&) const; | 54 Sk4i lessThan(const Sk4x&) const; |
45 Sk4i greaterThan(const Sk4x&) const; | 55 Sk4i greaterThan(const Sk4x&) const; |
46 Sk4i lessThanEqual(const Sk4x&) const; | 56 Sk4i lessThanEqual(const Sk4x&) const; |
47 Sk4i greaterThanEqual(const Sk4x&) const; | 57 Sk4i greaterThanEqual(const Sk4x&) const; |
48 | 58 |
49 static Sk4x Min(const Sk4x& a, const Sk4x& b); | 59 static Sk4x Min(const Sk4x& a, const Sk4x& b); |
50 static Sk4x Max(const Sk4x& a, const Sk4x& b); | 60 static Sk4x Max(const Sk4x& a, const Sk4x& b); |
51 | 61 |
52 // Swizzles follow OpenCL xyzw convention. | 62 // Swizzles follow OpenCL xyzw convention. |
53 Sk4x zwxy() const; | 63 Sk4x zwxy() const; |
54 | 64 |
55 // When there's a second argument, it's abcd. | 65 // When there's a second argument, it's abcd. |
56 static Sk4x XYAB(const Sk4x& xyzw, const Sk4x& abcd); | 66 static Sk4x XYAB(const Sk4x& xyzw, const Sk4x& abcd); |
57 static Sk4x ZWCD(const Sk4x& xyzw, const Sk4x& abcd); | 67 static Sk4x ZWCD(const Sk4x& xyzw, const Sk4x& abcd); |
58 | 68 |
| 69 // TODO: these are particularly efficient in SSE. Useful? Also efficient i
n NEON? |
| 70 // static Sk4x XAYB(const Sk4x& xyzw, const Sk4x& abcd); |
| 71 // static Sk4x ZCWD(const Sk4x& xyzw, const Sk4x& abcd); |
| 72 |
59 private: | 73 private: |
60 // It's handy to have Sk4f and Sk4i be mutual friends. | 74 // It's handy to have Sk4f and Sk4i be mutual friends. |
61 template <typename S> friend class Sk4x; | 75 template <typename S> friend class Sk4x; |
62 | 76 |
63 #define SK4X_PRIVATE 1 | 77 #define SK4X_PRIVATE 1 |
64 #include "Sk4x_portable.h" | 78 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 79 #include "Sk4x_sse.h" |
| 80 #else |
| 81 #include "Sk4x_portable.h" |
| 82 #endif |
65 #undef SK4X_PRIVATE | 83 #undef SK4X_PRIVATE |
66 }; | 84 }; |
67 | 85 |
68 #include "Sk4x_portable.h" | 86 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 87 #include "Sk4x_sse.h" |
| 88 #else |
| 89 #include "Sk4x_portable.h" |
| 90 #endif |
69 | 91 |
70 #endif//Sk4x_DEFINED | 92 #endif//Sk4x_DEFINED |
OLD | NEW |