/src/serenity/AK/SIMDMath.h

Source (jump to first uncovered line)
/*
 * Copyright (c) 2021, Stephan Unverwerth <s.unverwerth@serenityos.org>
 *
 * SPDX-License-Identifier: BSD-2-Clause
 */

#pragma once

#include <AK/Math.h>
#include <AK/SIMD.h>
#include <AK/SIMDExtras.h>
#include <math.h>

namespace AK::SIMD {

// Functions ending in "_int_range" only accept arguments within range [INT_MIN, INT_MAX].
// Other inputs will generate unexpected results.

ALWAYS_INLINE static f32x4 truncate_int_range(f32x4 v)
{
    return simd_cast<f32x4>(simd_cast<i32x4>(v));
}

ALWAYS_INLINE static f32x4 floor_int_range(f32x4 v)
{
    auto t = truncate_int_range(v);
    return t > v ? t - 1.0f : t;
}

ALWAYS_INLINE static f32x4 ceil_int_range(f32x4 v)
{
    auto t = truncate_int_range(v);
    return t < v ? t + 1.0f : t;
}

ALWAYS_INLINE static f32x4 frac_int_range(f32x4 v)
{
    return v - floor_int_range(v);
}

template<SIMDVector T>
ALWAYS_INLINE T bitselect(T v1, T v2, T control_mask)
{
    return (v1 & control_mask) | (v2 & ~control_mask);
}

template<SIMDVector T>
requires(IsIntegral<ElementOf<T>>)
ALWAYS_INLINE T abs(T x)
{
    return bitselect(x, -x, x > 0);
}

ALWAYS_INLINE static f32x4 clamp(f32x4 v, f32x4 min, f32x4 max)
{
    return v < min ? min : (v > max ? max : v);
}

template<SIMDVector V, typename U = ElementOf<V>>
ALWAYS_INLINE static V clamp(V v, U min, U max)
{
    return v < min ? min : (v > max ? max : v);
}

ALWAYS_INLINE static f32x4 exp(f32x4 v)
{
    // FIXME: This should be replaced with a vectorized algorithm instead of calling the scalar expf 4 times
    return f32x4 {
        expf(v[0]),
        expf(v[1]),
        expf(v[2]),
        expf(v[3]),
    };
}

ALWAYS_INLINE static f32x4 exp_approximate(f32x4 v)
{
    static constexpr int number_of_iterations = 10;
    auto result = 1.f + v / (1 << number_of_iterations);
    for (int i = 0; i < number_of_iterations; ++i)
        result *= result;
    return result;
}

ALWAYS_INLINE static f32x4 sqrt(f32x4 v)
{
#if ARCH(X86_64)
    return __builtin_ia32_sqrtps(v);
#else
    return f32x4 {
        AK::sqrt(v[0]),
        AK::sqrt(v[1]),
        AK::sqrt(v[2]),
        AK::sqrt(v[3]),
    };
#endif
}

}

Line	Count	Source (jump to first uncovered line)
1		/*
2		* Copyright (c) 2021, Stephan Unverwerth <s.unverwerth@serenityos.org>
3		*
4		* SPDX-License-Identifier: BSD-2-Clause
5		*/
6
7		#pragma once
8
9		#include <AK/Math.h>
10		#include <AK/SIMD.h>
11		#include <AK/SIMDExtras.h>
12		#include <math.h>
13
14		namespace AK::SIMD {
15
16		// Functions ending in "_int_range" only accept arguments within range [INT_MIN, INT_MAX].
17		// Other inputs will generate unexpected results.
18
19		ALWAYS_INLINE static f32x4 truncate_int_range(f32x4 v)
20	0	{
21	0	return simd_cast<f32x4>(simd_cast<i32x4>(v));
22	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::truncate_int_range(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::truncate_int_range(float __vector(4))
23
24		ALWAYS_INLINE static f32x4 floor_int_range(f32x4 v)
25	0	{
26	0	auto t = truncate_int_range(v);
27	0	return t > v ? t - 1.0f : t;
28	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::floor_int_range(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::floor_int_range(float __vector(4))
29
30		ALWAYS_INLINE static f32x4 ceil_int_range(f32x4 v)
31	0	{
32	0	auto t = truncate_int_range(v);
33	0	return t < v ? t + 1.0f : t;
34	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::ceil_int_range(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::ceil_int_range(float __vector(4))
35
36		ALWAYS_INLINE static f32x4 frac_int_range(f32x4 v)
37	0	{
38	0	return v - floor_int_range(v);
39	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::frac_int_range(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::frac_int_range(float __vector(4))
40
41		template<SIMDVector T>
42		ALWAYS_INLINE T bitselect(T v1, T v2, T control_mask)
43	0	{
44	0	return (v1 & control_mask) \| (v2 & ~control_mask);
45	0	} Unexecuted instantiation: _ZN2AK4SIMD9bitselectITkNS0_10SIMDVectorEDv4_sEET_S3_S3_S3_ Unexecuted instantiation: _ZN2AK4SIMD9bitselectITkNS0_10SIMDVectorEDv4_iEET_S3_S3_S3_
46
47		template<SIMDVector T>
48		requires(IsIntegral<ElementOf<T>>)
49		ALWAYS_INLINE T abs(T x)
50	0	{
51	0	return bitselect(x, -x, x > 0);
52	0	} Unexecuted instantiation: _ZN2AK4SIMD3absITkNS0_10SIMDVectorEDv4_sQ10IsIntegralINS_6Detail17__RemoveReferenceIDTixcl7declvalIT_EELi0EEE4TypeEEEES5_S5_ Unexecuted instantiation: _ZN2AK4SIMD3absITkNS0_10SIMDVectorEDv4_iQ10IsIntegralINS_6Detail17__RemoveReferenceIDTixcl7declvalIT_EELi0EEE4TypeEEEES5_S5_
53
54		ALWAYS_INLINE static f32x4 clamp(f32x4 v, f32x4 min, f32x4 max)
55	0	{
56	0	return v < min ? min : (v > max ? max : v);
57	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::clamp(float __vector(4), float __vector(4), float __vector(4))
58
59		template<SIMDVector V, typename U = ElementOf<V>>
60		ALWAYS_INLINE static V clamp(V v, U min, U max)
61		{
62		return v < min ? min : (v > max ? max : v);
63		}
64
65		ALWAYS_INLINE static f32x4 exp(f32x4 v)
66	0	{
67	0	// FIXME: This should be replaced with a vectorized algorithm instead of calling the scalar expf 4 times
68	0	return f32x4 {
69	0	expf(v[0]),
70	0	expf(v[1]),
71	0	expf(v[2]),
72	0	expf(v[3]),
73	0	};
74	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::exp(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::exp(float __vector(4))
75
76		ALWAYS_INLINE static f32x4 exp_approximate(f32x4 v)
77	0	{
78	0	static constexpr int number_of_iterations = 10;
79	0	auto result = 1.f + v / (1 << number_of_iterations);
80	0	for (int i = 0; i < number_of_iterations; ++i)
81	0	result *= result;
82	0	return result;
83	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::exp_approximate(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::exp_approximate(float __vector(4))
84
85		ALWAYS_INLINE static f32x4 sqrt(f32x4 v)
86	0	{
87	0	#if ARCH(X86_64)
88	0	return __builtin_ia32_sqrtps(v);
89	0	#else
90	0	return f32x4 {
91	0	AK::sqrt(v[0]),
92	0	AK::sqrt(v[1]),
93	0	AK::sqrt(v[2]),
94	0	AK::sqrt(v[3]),
95	0	};
96	0	#endif
97	0	} Unexecuted instantiation: ImageDecoder.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: ICOLoader.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: PNGLoader.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: Font.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: FuzzPNGLoader.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: Filter.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: HTMLCanvasElement.cpp:AK::SIMD::sqrt(float __vector(4)) Unexecuted instantiation: PNGWriter.cpp:AK::SIMD::sqrt(float __vector(4))
98
99		}

Coverage Report

Created: 2025-09-05 06:52