Skip to content

Commit ef283e3

Browse files
awxkeeawxkee
committed
Upgrade hwy, fix chroma
1 parent 50f7f86 commit ef283e3

File tree

16 files changed

+799
-257
lines changed

16 files changed

+799
-257
lines changed

‎avif-coder/src/main/cpp/JniDecoder.cpp‎

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -204,7 +204,7 @@ jobject decodeImplementationNative(JNIEnv *env, jobject thiz,
204204
GammaCurve gammaCurve = sRGB;
205205
GamutTransferFunction function = SKIP;
206206

207-
const float gamma = 1.0f;
207+
float gamma = 1.0f;
208208

209209
if (nclx->transfer_characteristics ==
210210
heif_transfer_characteristic_ITU_R_BT_2100_0_HLG) {
@@ -219,10 +219,12 @@ jobject decodeImplementationNative(JNIEnv *env, jobject thiz,
219219
function = SKIP;
220220
} else if (nclx->transfer_characteristics ==
221221
heif_transfer_characteristic_ITU_R_BT_470_6_System_M) {
222-
function = Gamma2p2;
222+
function = EOTF_GAMMA;
223+
gamma = 2.2f;
223224
} else if (nclx->transfer_characteristics ==
224225
heif_transfer_characteristic_ITU_R_BT_470_6_System_B_G) {
225-
function = Gamma2p8;
226+
function = EOTF_GAMMA;
227+
gamma = 2.8f;
226228
} else if (nclx->transfer_characteristics ==
227229
heif_transfer_characteristic_ITU_R_BT_601_6) {
228230
function = EOTF_BT601;

‎avif-coder/src/main/cpp/colorspace/HDRTransferAdapter.cpp‎

Lines changed: 34 additions & 45 deletions
Original file line numberDiff line numberDiff line change
@@ -55,6 +55,12 @@ namespace coder::HWY_NAMESPACE {
5555

5656
static const float sdrReferencePoint = 203.0f;
5757

58+
float half_to_float(const uint16_t x) {
59+
half f;
60+
f.data_ = x;
61+
return f;
62+
}
63+
5864
HWY_FAST_MATH_INLINE void
5965
TransferROWU16HFloats(uint16_t *data,
6066
const GamutTransferFunction function,
@@ -63,9 +69,9 @@ namespace coder::HWY_NAMESPACE {
6369
Eigen::Matrix3f *conversion,
6470
const float gamma,
6571
const bool useChromaticAdaptation) {
66-
auto r = (float) half(data[0]);
67-
auto g = (float) half(data[1]);
68-
auto b = (float) half(data[2]);
72+
auto r = (float) half_to_float(data[0]);
73+
auto g = (float) half_to_float(data[1]);
74+
auto b = (float) half_to_float(data[2]);
6975

7076
const auto adopt = getBradfordAdaptation();
7177

@@ -85,12 +91,8 @@ namespace coder::HWY_NAMESPACE {
8591
g = SMPTE428Eotf(g);
8692
b = SMPTE428Eotf(b);
8793
break;
88-
case Gamma2p2:
89-
case Gamma2p8: {
90-
float gammaValue = 2.2f;
91-
if (function == Gamma2p8) {
92-
gammaValue = 2.8f;
93-
}
94+
case EOTF_GAMMA: {
95+
float gammaValue = gamma;
9496
r = gammaOtf(r, gammaValue);
9597
g = gammaOtf(g, gammaValue);
9698
b = gammaOtf(b, gammaValue);
@@ -213,15 +215,10 @@ namespace coder::HWY_NAMESPACE {
213215
g = SMPTE428Eotf(g);
214216
b = SMPTE428Eotf(b);
215217
break;
216-
case Gamma2p2:
217-
case Gamma2p8: {
218-
float gammaValue = 2.2f;
219-
if (function == Gamma2p8) {
220-
gammaValue = 2.8f;
221-
}
222-
r = gammaOtf(r, gammaValue);
223-
g = gammaOtf(g, gammaValue);
224-
b = gammaOtf(b, gammaValue);
218+
case EOTF_GAMMA: {
219+
r = gammaOtf(r, gamma);
220+
g = gammaOtf(g, gamma);
221+
b = gammaOtf(b, gamma);
225222
}
226223
break;
227224
case EOTF_BT601: {
@@ -284,40 +281,40 @@ namespace coder::HWY_NAMESPACE {
284281
}
285282

286283
if (gammaCorrection == Rec2020) {
287-
data[0] = (uint8_t) clamp((float) bt2020GammaCorrection(r * maxColors), 0.0f,
284+
data[0] = (uint8_t) clamp((float) bt2020GammaCorrection(r) * maxColors, 0.0f,
288285
maxColors);
289-
data[1] = (uint8_t) clamp((float) bt2020GammaCorrection(g * maxColors), 0.0f,
286+
data[1] = (uint8_t) clamp((float) bt2020GammaCorrection(g) * maxColors, 0.0f,
290287
maxColors);
291-
data[2] = (uint8_t) clamp((float) bt2020GammaCorrection(b * maxColors), 0.0f,
288+
data[2] = (uint8_t) clamp((float) bt2020GammaCorrection(b) * maxColors, 0.0f,
292289
maxColors);
293290
} else if (gammaCorrection == DCIP3) {
294-
data[0] = (uint8_t) clamp((float) dciP3PQGammaCorrection(r * maxColors), 0.0f,
291+
data[0] = (uint8_t) clamp((float) dciP3PQGammaCorrection(r) * maxColors, 0.0f,
295292
maxColors);
296-
data[1] = (uint8_t) clamp((float) dciP3PQGammaCorrection(g * maxColors), 0.0f,
293+
data[1] = (uint8_t) clamp((float) dciP3PQGammaCorrection(g) * maxColors, 0.0f,
297294
maxColors);
298-
data[2] = (uint8_t) clamp((float) dciP3PQGammaCorrection(b * maxColors), 0.0f,
295+
data[2] = (uint8_t) clamp((float) dciP3PQGammaCorrection(b) * maxColors, 0.0f,
299296
maxColors);
300297
} else if (gammaCorrection == GAMMA) {
301298
const float gammaEval = 1.f / gamma;
302-
data[0] = (uint8_t) clamp((float) gammaOtf(r * maxColors, gammaEval), 0.0f,
299+
data[0] = (uint8_t) clamp((float) gammaOtf(r, gammaEval) * maxColors, 0.0f,
303300
maxColors);
304-
data[1] = (uint8_t) clamp((float) gammaOtf(g * maxColors, gammaEval), 0.0f,
301+
data[1] = (uint8_t) clamp((float) gammaOtf(g, gammaEval) * maxColors, 0.0f,
305302
maxColors);
306-
data[2] = (uint8_t) clamp((float) gammaOtf(b * maxColors, gammaEval), 0.0f,
303+
data[2] = (uint8_t) clamp((float) gammaOtf(b, gammaEval) * maxColors, 0.0f,
307304
maxColors);
308305
} else if (gammaCorrection == Rec709) {
309-
data[0] = (uint8_t) clamp((float) LinearITUR709ToITUR709(r * maxColors), 0.0f,
306+
data[0] = (uint8_t) clamp((float) LinearITUR709ToITUR709(r) * maxColors, 0.0f,
310307
maxColors);
311-
data[1] = (uint8_t) clamp((float) LinearITUR709ToITUR709(g * maxColors), 0.0f,
308+
data[1] = (uint8_t) clamp((float) LinearITUR709ToITUR709(g) * maxColors, 0.0f,
312309
maxColors);
313-
data[2] = (uint8_t) clamp((float) LinearITUR709ToITUR709(b * maxColors), 0.0f,
310+
data[2] = (uint8_t) clamp((float) LinearITUR709ToITUR709(b) * maxColors, 0.0f,
314311
maxColors);
315312
} else if (gammaCorrection == sRGB) {
316-
data[0] = (uint8_t) clamp((float) LinearSRGBTosRGB(r * maxColors), 0.0f,
313+
data[0] = (uint8_t) clamp((float) LinearSRGBTosRGB(r) * maxColors, 0.0f,
317314
maxColors);
318-
data[1] = (uint8_t) clamp((float) LinearSRGBTosRGB(g * maxColors), 0.0f,
315+
data[1] = (uint8_t) clamp((float) LinearSRGBTosRGB(g) * maxColors, 0.0f,
319316
maxColors);
320-
data[2] = (uint8_t) clamp((float) LinearSRGBTosRGB(b * maxColors), 0.0f,
317+
data[2] = (uint8_t) clamp((float) LinearSRGBTosRGB(b) * maxColors, 0.0f,
321318
maxColors);
322319
} else {
323320
data[0] = (uint8_t) clamp((float) r * maxColors, 0.0f, maxColors);
@@ -410,12 +407,8 @@ namespace coder::HWY_NAMESPACE {
410407
pqHighG = SMPTE428Eotf(df32, gHigh32);
411408
pqHighB = SMPTE428Eotf(df32, bHigh32);
412409
break;
413-
case Gamma2p2:
414-
case Gamma2p8: {
415-
float gammaValue = 2.2f;
416-
if (function == Gamma2p8) {
417-
gammaValue = 2.8f;
418-
}
410+
case EOTF_GAMMA: {
411+
const float gammaValue = gamma;
419412
pqLowR = gammaOtf(df32, rLow32, gammaValue);
420413
pqLowG = gammaOtf(df32, gLow32, gammaValue);
421414
pqLowB = gammaOtf(df32, bLow32, gammaValue);
@@ -610,12 +603,8 @@ namespace coder::HWY_NAMESPACE {
610603
pqG = SMPTE428Eotf(df32, G);
611604
pqB = SMPTE428Eotf(df32, B);
612605
break;
613-
case Gamma2p2:
614-
case Gamma2p8: {
615-
float gammaValue = 2.2f;
616-
if (function == Gamma2p8) {
617-
gammaValue = 2.8f;
618-
}
606+
case EOTF_GAMMA: {
607+
const float gammaValue = gamma;
619608
pqR = gammaOtf(df32, R, gammaValue);
620609
pqG = gammaOtf(df32, G, gammaValue);
621610
pqB = gammaOtf(df32, B, gammaValue);

‎avif-coder/src/main/cpp/colorspace/HDRTransferAdapter.h‎

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -39,7 +39,7 @@ enum GammaCurve {
3939
};
4040

4141
enum GamutTransferFunction {
42-
SKIP, PQ, HLG, SMPTE428, Gamma2p2, Gamma2p8, EOTF_BT601, EOTF_BT709, EOTF_SMPTE240,
42+
SKIP, PQ, HLG, SMPTE428, EOTF_GAMMA, EOTF_BT601, EOTF_BT709, EOTF_SMPTE240,
4343
EOTF_LOG100, EOTF_LOG100SRT10, EOTF_IEC_61966, EOTF_BT1361
4444
};
4545

@@ -56,7 +56,9 @@ class HDRTransferAdapter {
5656
float gamma,
5757
bool useChromaticAdaptation)
5858
: function(function),
59-
gammaCorrection(gammaCorrection), bitDepth(bitDepth), halfFloats(halfFloats),
59+
gammaCorrection(gammaCorrection),
60+
bitDepth(bitDepth),
61+
halfFloats(halfFloats),
6062
rgbaData(rgbaData),
6163
stride(stride),
6264
width(width),

‎avif-coder/src/main/cpp/hwy/base.h‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -381,8 +381,8 @@ static constexpr HWY_MAYBE_UNUSED size_t kMaxVectorSize = 16;
381381
// hwy::float16_t and hwy::bfloat16_t are forward declared here to allow
382382
// BitCastScalar to be implemented before the implementations of the
383383
// hwy::float16_t and hwy::bfloat16_t types
384-
struct alignas(2) float16_t;
385-
struct alignas(2) bfloat16_t;
384+
struct float16_t;
385+
struct bfloat16_t;
386386

387387
using float32_t = float;
388388
using float64_t = double;
@@ -1259,7 +1259,7 @@ struct BitCastScalarSrcCastHelper<hwy::float16_t> {
12591259
#endif // HWY_HAVE_SCALAR_F16_OPERATORS
12601260

12611261
HWY_API HWY_F16_CONSTEXPR float F32FromF16(float16_t f16) {
1262-
#if HWY_HAVE_SCALAR_F16_OPERATORS
1262+
#if HWY_HAVE_SCALAR_F16_OPERATORS && !HWY_IDE
12631263
return static_cast<float>(f16);
12641264
#endif
12651265
#if !HWY_HAVE_SCALAR_F16_OPERATORS || HWY_IDE
@@ -1329,7 +1329,7 @@ HWY_API HWY_F16_CONSTEXPR float F32FromF16(float16_t f16) {
13291329
// HWY_COMPILER_MSVC >= 1926)
13301330

13311331
HWY_API HWY_F16_CONSTEXPR float16_t F16FromF32(float f32) {
1332-
#if HWY_HAVE_SCALAR_F16_OPERATORS
1332+
#if HWY_HAVE_SCALAR_F16_OPERATORS && !HWY_IDE
13331333
return float16_t(static_cast<float16_t::Native>(f32));
13341334
#endif
13351335
#if !HWY_HAVE_SCALAR_F16_OPERATORS || HWY_IDE

‎avif-coder/src/main/cpp/hwy/contrib/unroller/unroller_test.cc‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -132,7 +132,7 @@ struct FindUnit : UnrollerUnit<FindUnit<T>, T, MakeSigned<T>> {
132132
using DI = RebindToSigned<D>;
133133
DI di;
134134

135-
FindUnit<T>(T find) : to_find(find) {}
135+
FindUnit(T find) : to_find(find) {}
136136

137137
hn::Vec<DI> Func(ptrdiff_t idx, const hn::Vec<D> x, const hn::Vec<DI> y) {
138138
const Mask<D> msk = hn::Eq(x, hn::Set(d, to_find));

‎avif-coder/src/main/cpp/hwy/highway.h‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -32,8 +32,8 @@ namespace hwy {
3232

3333
// API version (https://semver.org/); keep in sync with CMakeLists.txt.
3434
#define HWY_MAJOR 1
35-
#define HWY_MINOR 0
36-
#define HWY_PATCH 7
35+
#define HWY_MINOR 1
36+
#define HWY_PATCH 0
3737

3838
//------------------------------------------------------------------------------
3939
// Shorthand for tags (defined in shared-inl.h) used to select overloads.

‎avif-coder/src/main/cpp/hwy/ops/arm_neon-inl.h‎

Lines changed: 16 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -7568,22 +7568,35 @@ HWY_NEON_DEF_REDUCTION_CORE_TYPES(ReduceSum, vaddv)
75687568
HWY_NEON_DEF_REDUCTION_UI64(ReduceSum, vaddv)
75697569

75707570
// Emulate missing UI64 and partial N=2.
7571-
template <class D, HWY_IF_LANES_D(D, 2)>
7571+
template <class D, HWY_IF_LANES_D(D, 2),
7572+
HWY_IF_T_SIZE_ONE_OF_D(D, (1 << 1) | (1 << 2))>
75727573
HWY_API TFromD<D> ReduceSum(D /* tag */, VFromD<D> v10) {
75737574
return GetLane(v10) + ExtractLane(v10, 1);
75747575
}
75757576

7576-
template <class D, HWY_IF_LANES_D(D, 2)>
7577+
template <class D, HWY_IF_LANES_D(D, 2), HWY_IF_NOT_FLOAT_D(D),
7578+
HWY_IF_T_SIZE_ONE_OF_D(D, (1 << 1) | (1 << 2) | (1 << 8))>
75777579
HWY_API TFromD<D> ReduceMin(D /* tag */, VFromD<D> v10) {
75787580
return HWY_MIN(GetLane(v10), ExtractLane(v10, 1));
75797581
}
75807582

7581-
template <class D, HWY_IF_LANES_D(D, 2)>
7583+
template <class D, HWY_IF_LANES_D(D, 2), HWY_IF_NOT_FLOAT_D(D),
7584+
HWY_IF_T_SIZE_ONE_OF_D(D, (1 << 1) | (1 << 2) | (1 << 8))>
75827585
HWY_API TFromD<D> ReduceMax(D /* tag */, VFromD<D> v10) {
75837586
return HWY_MAX(GetLane(v10), ExtractLane(v10, 1));
75847587
}
75857588

75867589
#if HWY_HAVE_FLOAT16
7590+
template <class D, HWY_IF_LANES_D(D, 2), HWY_IF_F16_D(D)>
7591+
HWY_API float16_t ReduceMin(D d, VFromD<D> v10) {
7592+
return GetLane(Min(v10, Reverse2(d, v10)));
7593+
}
7594+
7595+
template <class D, HWY_IF_LANES_D(D, 2), HWY_IF_F16_D(D)>
7596+
HWY_API float16_t ReduceMax(D d, VFromD<D> v10) {
7597+
return GetLane(Max(v10, Reverse2(d, v10)));
7598+
}
7599+
75877600
template <class D, HWY_IF_F16_D(D), HWY_IF_V_SIZE_D(D, 8)>
75887601
HWY_API float16_t ReduceSum(D /* tag */, VFromD<D> v) {
75897602
const float16x4_t x2 = vpadd_f16(v.raw, v.raw);

‎avif-coder/src/main/cpp/hwy/ops/arm_sve-inl.h‎

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2089,6 +2089,22 @@ HWY_API svfloat32_t PromoteTo(Simd<float32_t, N, kPow2> /* d */,
20892089
return svcvt_f32_f16_x(detail::PTrue(Simd<float16_t, N, kPow2>()), vv);
20902090
}
20912091

2092+
#ifdef HWY_NATIVE_PROMOTE_F16_TO_F64
2093+
#undef HWY_NATIVE_PROMOTE_F16_TO_F64
2094+
#else
2095+
#define HWY_NATIVE_PROMOTE_F16_TO_F64
2096+
#endif
2097+
2098+
template <size_t N, int kPow2>
2099+
HWY_API svfloat64_t PromoteTo(Simd<float64_t, N, kPow2> /* d */,
2100+
const svfloat16_t v) {
2101+
// svcvt* expects inputs in even lanes, whereas Highway wants lower lanes, so
2102+
// first replicate each lane once.
2103+
const svfloat16_t vv = detail::ZipLowerSame(v, v);
2104+
return svcvt_f64_f16_x(detail::PTrue(Simd<float16_t, N, kPow2>()),
2105+
detail::ZipLowerSame(vv, vv));
2106+
}
2107+
20922108
template <size_t N, int kPow2>
20932109
HWY_API svfloat64_t PromoteTo(Simd<float64_t, N, kPow2> /* d */,
20942110
const svfloat32_t v) {
@@ -2546,6 +2562,20 @@ HWY_API svfloat16_t DemoteTo(Simd<float16_t, N, kPow2> d, const svfloat32_t v) {
25462562
in_even); // lower half
25472563
}
25482564

2565+
#ifdef HWY_NATIVE_DEMOTE_F64_TO_F16
2566+
#undef HWY_NATIVE_DEMOTE_F64_TO_F16
2567+
#else
2568+
#define HWY_NATIVE_DEMOTE_F64_TO_F16
2569+
#endif
2570+
2571+
template <size_t N, int kPow2>
2572+
HWY_API svfloat16_t DemoteTo(Simd<float16_t, N, kPow2> d, const svfloat64_t v) {
2573+
const svfloat16_t in_lo16 = svcvt_f16_f64_x(detail::PTrue(d), v);
2574+
const svfloat16_t in_even = detail::ConcatEvenFull(in_lo16, in_lo16);
2575+
return detail::ConcatEvenFull(in_even,
2576+
in_even); // lower half
2577+
}
2578+
25492579
template <size_t N, int kPow2>
25502580
HWY_API VBF16 DemoteTo(Simd<bfloat16_t, N, kPow2> dbf16, svfloat32_t v) {
25512581
const svuint16_t in_even = BitCast(ScalableTag<uint16_t>(), v);

‎avif-coder/src/main/cpp/hwy/ops/emu128-inl.h‎

Lines changed: 10 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -382,16 +382,11 @@ HWY_API MFromD<DTo> RebindMask(DTo /* tag */, MFrom mask) {
382382
return to;
383383
}
384384

385-
template <typename T, size_t N>
386-
Vec128<T, N> VecFromMask(Mask128<T, N> mask) {
387-
Vec128<T, N> v;
388-
CopySameSize(&mask.bits, &v.raw);
389-
return v;
390-
}
391-
392385
template <class D>
393386
VFromD<D> VecFromMask(D /* tag */, MFromD<D> mask) {
394-
return VecFromMask(mask);
387+
VFromD<D> v;
388+
CopySameSize(&mask.bits, &v.raw);
389+
return v;
395390
}
396391

397392
template <class D>
@@ -407,19 +402,20 @@ HWY_API MFromD<D> FirstN(D d, size_t n) {
407402
template <typename T, size_t N>
408403
HWY_API Vec128<T, N> IfThenElse(Mask128<T, N> mask, Vec128<T, N> yes,
409404
Vec128<T, N> no) {
410-
return IfVecThenElse(VecFromMask(mask), yes, no);
405+
const DFromV<decltype(yes)> d;
406+
return IfVecThenElse(VecFromMask(d, mask), yes, no);
411407
}
412408

413409
template <typename T, size_t N>
414410
HWY_API Vec128<T, N> IfThenElseZero(Mask128<T, N> mask, Vec128<T, N> yes) {
415411
const DFromV<decltype(yes)> d;
416-
return IfVecThenElse(VecFromMask(mask), yes, Zero(d));
412+
return IfVecThenElse(VecFromMask(d, mask), yes, Zero(d));
417413
}
418414

419415
template <typename T, size_t N>
420416
HWY_API Vec128<T, N> IfThenZeroElse(Mask128<T, N> mask, Vec128<T, N> no) {
421417
const DFromV<decltype(no)> d;
422-
return IfVecThenElse(VecFromMask(mask), Zero(d), no);
418+
return IfVecThenElse(VecFromMask(d, mask), Zero(d), no);
423419
}
424420

425421
template <typename T, size_t N>
@@ -445,7 +441,8 @@ HWY_API Vec128<T, N> ZeroIfNegative(Vec128<T, N> v) {
445441

446442
template <typename T, size_t N>
447443
HWY_API Mask128<T, N> Not(Mask128<T, N> m) {
448-
return MaskFromVec(Not(VecFromMask(Simd<T, N, 0>(), m)));
444+
const Simd<T, N, 0> d;
445+
return MaskFromVec(Not(VecFromMask(d, m)));
449446
}
450447

451448
template <typename T, size_t N>
@@ -1610,7 +1607,7 @@ HWY_INLINE ToT CastValueForF2IConv(FromT val) {
16101607

16111608
template <class ToT, class ToTypeTag, class FromT>
16121609
HWY_INLINE ToT CastValueForPromoteTo(ToTypeTag /* to_type_tag */, FromT val) {
1613-
return static_cast<ToT>(val);
1610+
return ConvertScalarTo<ToT>(val);
16141611
}
16151612

16161613
template <class ToT>

0 commit comments

Comments
 (0)