ladybird/Userland/Libraries/LibSoftGPU/SIMD.h
Jelle Raaijmakers 69b94e4235 LibSoftGPU: Make blending simpler and more efficient
Previously, we would precalculate "alpha blend factors" on every
configuration update and then calculate the source and destination
blending factors in one go using all these factors. The idea here was
probably that we would get better performance by avoiding branching.

However, by measuring blending performance in Quake III, it seems that
this simpler version that only calculates the required factors reduces
the CPU time spent in `rasterize_triangle` by 3%.

As a bonus, `GL_SRC_ALPHA_SATURATE` is now also implemented.
2023-02-02 14:38:26 +01:00

147 lines
3.7 KiB
C++

/*
* Copyright (c) 2021, Stephan Unverwerth <s.unverwerth@serenityos.org>
* Copyright (c) 2023, Jelle Raaijmakers <jelle@gmta.nl>
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#pragma once
#include <AK/SIMDExtras.h>
#include <AK/SIMDMath.h>
#include <LibGfx/Vector2.h>
#include <LibGfx/Vector3.h>
#include <LibGfx/Vector4.h>
namespace SoftGPU {
ALWAYS_INLINE static constexpr Vector2<AK::SIMD::f32x4> expand4(Vector2<float> const& v)
{
return Vector2<AK::SIMD::f32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
};
}
ALWAYS_INLINE static constexpr Vector3<AK::SIMD::f32x4> expand4(Vector3<float> const& v)
{
return Vector3<AK::SIMD::f32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
AK::SIMD::expand4(v.z()),
};
}
ALWAYS_INLINE static constexpr Vector4<AK::SIMD::f32x4> expand4(Vector4<float> const& v)
{
return Vector4<AK::SIMD::f32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
AK::SIMD::expand4(v.z()),
AK::SIMD::expand4(v.w()),
};
}
ALWAYS_INLINE static constexpr Vector2<AK::SIMD::i32x4> expand4(Vector2<int> const& v)
{
return Vector2<AK::SIMD::i32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
};
}
ALWAYS_INLINE static constexpr Vector3<AK::SIMD::i32x4> expand4(Vector3<int> const& v)
{
return Vector3<AK::SIMD::i32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
AK::SIMD::expand4(v.z()),
};
}
ALWAYS_INLINE static constexpr Vector4<AK::SIMD::i32x4> expand4(Vector4<int> const& v)
{
return Vector4<AK::SIMD::i32x4> {
AK::SIMD::expand4(v.x()),
AK::SIMD::expand4(v.y()),
AK::SIMD::expand4(v.z()),
AK::SIMD::expand4(v.w()),
};
}
ALWAYS_INLINE static AK::SIMD::f32x4 ddx(AK::SIMD::f32x4 v)
{
return AK::SIMD::f32x4 {
v[1] - v[0],
v[1] - v[0],
v[3] - v[2],
v[3] - v[2],
};
}
ALWAYS_INLINE static AK::SIMD::f32x4 ddy(AK::SIMD::f32x4 v)
{
return AK::SIMD::f32x4 {
v[2] - v[0],
v[3] - v[1],
v[2] - v[0],
v[3] - v[1],
};
}
ALWAYS_INLINE static Vector2<AK::SIMD::f32x4> ddx(Vector2<AK::SIMD::f32x4> const& v)
{
return {
ddx(v.x()),
ddx(v.y()),
};
}
ALWAYS_INLINE static Vector2<AK::SIMD::f32x4> ddy(Vector2<AK::SIMD::f32x4> const& v)
{
return {
ddy(v.x()),
ddy(v.y()),
};
}
ALWAYS_INLINE static AK::SIMD::f32x4 length(Vector2<AK::SIMD::f32x4> const& v)
{
return AK::SIMD::sqrt(v.dot(v));
}
// Calculates a quadratic approximation of log2, exploiting the fact that IEEE754 floats are represented as mantissa * 2^exponent.
// See https://stackoverflow.com/questions/9411823/fast-log2float-x-implementation-c
ALWAYS_INLINE static AK::SIMD::f32x4 log2_approximate(AK::SIMD::f32x4 v)
{
union {
AK::SIMD::f32x4 float_val;
AK::SIMD::i32x4 int_val;
} u { v };
// Extract just the exponent minus 1, giving a lower integral bound for log2.
auto log = AK::SIMD::to_f32x4(((u.int_val >> 23) & 255) - 128);
// Replace the exponent with 0, giving a value between 1 and 2.
u.int_val &= ~(255 << 23);
u.int_val |= 127 << 23;
// Approximate log2 by adding a quadratic function of u to the integral part.
log += (-0.34484843f * u.float_val + 2.02466578f) * u.float_val - 0.67487759f;
return log;
}
ALWAYS_INLINE static Vector2<AK::SIMD::f32x4> to_vec2_f32x4(Vector2<AK::SIMD::i32x4> const& v)
{
return {
AK::SIMD::to_f32x4(v.x()),
AK::SIMD::to_f32x4(v.y()),
};
}
ALWAYS_INLINE static constexpr Vector4<AK::SIMD::f32x4> to_vec4(AK::SIMD::f32x4 v)
{
return { v, v, v, v };
}
}