OLD | NEW |
1 // It is important _not_ to put header guards here. | 1 // It is important _not_ to put header guards here. |
2 // This file will be intentionally included three times. | 2 // This file will be intentionally included three times. |
3 | 3 |
4 #include "SkTypes.h" // Keep this before any #ifdef for skbug.com/3362 | 4 #include "SkTypes.h" // Keep this before any #ifdef for skbug.com/3362 |
5 | 5 |
6 #if defined(SK4X_PREAMBLE) | 6 #if defined(SK4X_PREAMBLE) |
7 #include <arm_neon.h> | 7 #include <arm_neon.h> |
8 | 8 |
9 // Template metaprogramming to map scalar types to vector types. | 9 // Template metaprogramming to map scalar types to vector types. |
10 template <typename T> struct SkScalarToSIMD; | 10 template <typename T> struct SkScalarToSIMD; |
(...skipping 57 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
68 M(Sk4i) cast<Sk4i>() const { return vcvtq_s32_f32(fVec); } | 68 M(Sk4i) cast<Sk4i>() const { return vcvtq_s32_f32(fVec); } |
69 | 69 |
70 // We're going to skip allTrue(), anyTrue(), and bit-manipulators | 70 // We're going to skip allTrue(), anyTrue(), and bit-manipulators |
71 // for Sk4f. Code that calls them probably does so accidentally. | 71 // for Sk4f. Code that calls them probably does so accidentally. |
72 // Ask msarett or mtklein to fill these in if you really need them. | 72 // Ask msarett or mtklein to fill these in if you really need them. |
73 M(Sk4f) add (const Sk4f& o) const { return vaddq_f32(fVec, o.fVec); } | 73 M(Sk4f) add (const Sk4f& o) const { return vaddq_f32(fVec, o.fVec); } |
74 M(Sk4f) subtract(const Sk4f& o) const { return vsubq_f32(fVec, o.fVec); } | 74 M(Sk4f) subtract(const Sk4f& o) const { return vsubq_f32(fVec, o.fVec); } |
75 M(Sk4f) multiply(const Sk4f& o) const { return vmulq_f32(fVec, o.fVec); } | 75 M(Sk4f) multiply(const Sk4f& o) const { return vmulq_f32(fVec, o.fVec); } |
76 | 76 |
77 M(Sk4f) divide (const Sk4f& o) const { | 77 M(Sk4f) divide (const Sk4f& o) const { |
78 float32x4_t est0 = vrecpeq_f32(o.fVec); | 78 #if defined(SK_CPU_ARM64) |
79 float32x4_t est1 = vmulq_f32(vrecpsq_f32(est0, o.fVec), est0); | 79 return vdivq_f32(fVec, o.fVec); |
80 float32x4_t est2 = vmulq_f32(vrecpsq_f32(est1, o.fVec), est1); | 80 #else |
| 81 float32x4_t est0 = vrecpeq_f32(o.fVec), |
| 82 est1 = vmulq_f32(vrecpsq_f32(est0, o.fVec), est0), |
| 83 est2 = vmulq_f32(vrecpsq_f32(est1, o.fVec), est1); |
81 return vmulq_f32(est2, fVec); | 84 return vmulq_f32(est2, fVec); |
| 85 #endif |
82 } | 86 } |
83 | 87 |
84 M(Sk4f) rsqrt() const { | 88 M(Sk4f) rsqrt() const { |
85 float32x4_t est0 = vrsqrteq_f32(fVec), | 89 float32x4_t est0 = vrsqrteq_f32(fVec), |
86 est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0
); | 90 est1 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0
); |
87 return est1; | 91 return est1; |
88 } | 92 } |
89 | 93 |
90 M(Sk4f) sqrt() const { | 94 M(Sk4f) sqrt() const { |
| 95 #if defined(SK_CPU_ARM64) |
| 96 return vsqrtq_f32(fVec); |
| 97 #else |
91 float32x4_t est1 = this->rsqrt().fVec, | 98 float32x4_t est1 = this->rsqrt().fVec, |
92 // An extra step of Newton's method to refine the estimate of 1/sqrt(this). | 99 // An extra step of Newton's method to refine the estimate of 1/sqrt(this). |
93 est2 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1
); | 100 est2 = vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1
); |
94 return vmulq_f32(fVec, est2); | 101 return vmulq_f32(fVec, est2); |
| 102 #endif |
95 } | 103 } |
96 | 104 |
97 M(Sk4i) equal (const Sk4f& o) const { return vreinterpretq_s32_u32(vce
qq_f32(fVec, o.fVec)); } | 105 M(Sk4i) equal (const Sk4f& o) const { return vreinterpretq_s32_u32(vce
qq_f32(fVec, o.fVec)); } |
98 M(Sk4i) notEqual (const Sk4f& o) const { return vreinterpretq_s32_u32(vmv
nq_u32(vceqq_f32(fVec, o.fVec))); } | 106 M(Sk4i) notEqual (const Sk4f& o) const { return vreinterpretq_s32_u32(vmv
nq_u32(vceqq_f32(fVec, o.fVec))); } |
99 M(Sk4i) lessThan (const Sk4f& o) const { return vreinterpretq_s32_u32(vcl
tq_f32(fVec, o.fVec)); } | 107 M(Sk4i) lessThan (const Sk4f& o) const { return vreinterpretq_s32_u32(vcl
tq_f32(fVec, o.fVec)); } |
100 M(Sk4i) greaterThan (const Sk4f& o) const { return vreinterpretq_s32_u32(vcg
tq_f32(fVec, o.fVec)); } | 108 M(Sk4i) greaterThan (const Sk4f& o) const { return vreinterpretq_s32_u32(vcg
tq_f32(fVec, o.fVec)); } |
101 M(Sk4i) lessThanEqual (const Sk4f& o) const { return vreinterpretq_s32_u32(vcl
eq_f32(fVec, o.fVec)); } | 109 M(Sk4i) lessThanEqual (const Sk4f& o) const { return vreinterpretq_s32_u32(vcl
eq_f32(fVec, o.fVec)); } |
102 M(Sk4i) greaterThanEqual(const Sk4f& o) const { return vreinterpretq_s32_u32(vcg
eq_f32(fVec, o.fVec)); } | 110 M(Sk4i) greaterThanEqual(const Sk4f& o) const { return vreinterpretq_s32_u32(vcg
eq_f32(fVec, o.fVec)); } |
103 | 111 |
104 M(Sk4f) Min(const Sk4f& a, const Sk4f& b) { return vminq_f32(a.fVec, b.fVec); } | 112 M(Sk4f) Min(const Sk4f& a, const Sk4f& b) { return vminq_f32(a.fVec, b.fVec); } |
(...skipping 112 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
217 } | 225 } |
218 M(Sk4i) ZWCD(const Sk4i& xyzw, const Sk4i& abcd) { | 226 M(Sk4i) ZWCD(const Sk4i& xyzw, const Sk4i& abcd) { |
219 int32x4x2_t xayb_zcwd = vzipq_s32(xyzw.fVec, abcd.fVec); | 227 int32x4x2_t xayb_zcwd = vzipq_s32(xyzw.fVec, abcd.fVec); |
220 int32x4x2_t axby_czdw = vzipq_s32(abcd.fVec, xyzw.fVec); | 228 int32x4x2_t axby_czdw = vzipq_s32(abcd.fVec, xyzw.fVec); |
221 return vuzpq_s32(xayb_zcwd.val[1], axby_czdw.val[1]).val[0]; | 229 return vuzpq_s32(xayb_zcwd.val[1], axby_czdw.val[1]).val[0]; |
222 } | 230 } |
223 | 231 |
224 #undef M | 232 #undef M |
225 | 233 |
226 #endif | 234 #endif |
OLD | NEW |