1 /*
2 * Copyright 2014 Google Inc.
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8 #ifndef SkHalf_DEFINED
9 #define SkHalf_DEFINED
10
11 #include "SkNx.h"
12 #include "SkTypes.h"
13
14 // 16-bit floating point value
15 // format is 1 bit sign, 5 bits exponent, 10 bits mantissa
16 // only used for storage
17 typedef uint16_t SkHalf;
18
19 static constexpr uint16_t SK_HalfMin = 0x0400; // 2^-14 (minimum positive normal value)
20 static constexpr uint16_t SK_HalfMax = 0x7bff; // 65504
21 static constexpr uint16_t SK_HalfEpsilon = 0x1400; // 2^-10
22 static constexpr uint16_t SK_Half1 = 0x3C00; // 1
23
24 // convert between half and single precision floating point
25 float SkHalfToFloat(SkHalf h);
26 SkHalf SkFloatToHalf(float f);
27
28 // Convert between half and single precision floating point,
29 // assuming inputs and outputs are both finite, and may
30 // flush values which would be denormal half floats to zero.
31 static inline Sk4f SkHalfToFloat_finite_ftz(uint64_t);
32 static inline Sk4h SkFloatToHalf_finite_ftz(const Sk4f&);
33
34 // ~~~~~~~~~~~ impl ~~~~~~~~~~~~~~ //
35
36 // Like the serial versions in SkHalf.cpp, these are based on
37 // https://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
38
39 // GCC 4.9 lacks the intrinsics to use ARMv8 f16<->f32 instructions, so we use inline assembly.
40
SkHalfToFloat_finite_ftz(uint64_t rgba)41 static inline Sk4f SkHalfToFloat_finite_ftz(uint64_t rgba) {
42 Sk4h hs = Sk4h::Load(&rgba);
43 #if !defined(SKNX_NO_SIMD) && defined(SK_CPU_ARM64)
44 float32x4_t fs;
45 asm ("fcvtl %[fs].4s, %[hs].4h \n" // vcvt_f32_f16(...)
46 : [fs] "=w" (fs) // =w: write-only NEON register
47 : [hs] "w" (hs.fVec)); // w: read-only NEON register
48 return fs;
49 #else
50 Sk4i bits = SkNx_cast<int>(hs), // Expand to 32 bit.
51 sign = bits & 0x00008000, // Save the sign bit for later...
52 positive = bits ^ sign, // ...but strip it off for now.
53 is_norm = 0x03ff < positive; // Exponent > 0?
54
55 // For normal half floats, extend the mantissa by 13 zero bits,
56 // then adjust the exponent from 15 bias to 127 bias.
57 Sk4i norm = (positive << 13) + ((127 - 15) << 23);
58
59 Sk4i merged = (sign << 16) | (norm & is_norm);
60 return Sk4f::Load(&merged);
61 #endif
62 }
63
SkFloatToHalf_finite_ftz(const Sk4f & fs)64 static inline Sk4h SkFloatToHalf_finite_ftz(const Sk4f& fs) {
65 #if !defined(SKNX_NO_SIMD) && defined(SK_CPU_ARM64)
66 float32x4_t vec = fs.fVec;
67 asm ("fcvtn %[vec].4h, %[vec].4s \n" // vcvt_f16_f32(vec)
68 : [vec] "+w" (vec)); // +w: read-write NEON register
69 return vreinterpret_u16_f32(vget_low_f32(vec));
70 #else
71 Sk4i bits = Sk4i::Load(&fs),
72 sign = bits & 0x80000000, // Save the sign bit for later...
73 positive = bits ^ sign, // ...but strip it off for now.
74 will_be_norm = 0x387fdfff < positive; // greater than largest denorm half?
75
76 // For normal half floats, adjust the exponent from 127 bias to 15 bias,
77 // then drop the bottom 13 mantissa bits.
78 Sk4i norm = (positive - ((127 - 15) << 23)) >> 13;
79
80 Sk4i merged = (sign >> 16) | (will_be_norm & norm);
81 return SkNx_cast<uint16_t>(merged);
82 #endif
83 }
84
85 #endif
86