Kore
(Kha's C++ sister project)

krafix
(Shader cross-compiler)

kraffiti
(Image compression tool)

g2/g4
(2D/3D Graphics API)

Kha

Returns

Speed

1. Runtime Performance
   1. Rendering
   2. Load Times
   3. ...
2. Development Workflow
   1. Debugging
   2. Compiling
   3. ...

Rendering Performance

Bunnymark

Big caveats:
Only draws the same image again and again
No browser tests yet

Why is NME/OpenFL fast?

for (i in 0...numBunnies) {
	// ...
	drawList[index] = bunny.position.x;
	drawList[index + 1] = bunny.position.y;
}
tilesheet.drawTiles(graphics, drawList, false);
					

Why is Kha fast?

for (i in 0...bunnies.length) {
	g.drawImage(bunnies[i].texture, bunnies[i].x, bunnies[i].y);
}	
					

Why is Kha fast?

I use profilers

Profilers

• hxScout
  • Shows data on Haxe level
• Visual Studio, Xcode, ...
  • Shows data on C++ level
• Intel VTune, ...
  • Shows data on CPU level

Why is Kha fast?

• Inline Constructors
• 32 bit floats
  • kha.FastFloat
• SIMD
  • kha.simd

SIMD

• Single Instruction, Multiple Data
• SSE was introduced in 1999 (MMX in 1997)
• Support in C# added in 2014
• Support in JavaScript in development
• Support in C/C++ not in sight (use compiler intrinsics)

kha.simd

• SIMD api similar to simd.js
• Maps to SSE and Neon where available
• Maps to scalar code everywhere else

Next steps in runtime performance?

A comprehensive preformance test suite

Loading Data

Typically the biggest data set and also the biggest optimization potential:

Images

Loading Images

Worst possible strategy for a local application:

Load PNG files

• Has to be uncompressed before used
• Uncompression is slower than disk read speeds

Loading Images

• Use light compression which decompressed faster than disk speeds (Snappy)
• Use compressed texture formats when possible
• Preprocess as much as possible

khafiles

// uses snappy for non-web targets
project.addAssets('Assets/**');
					

khafiles

// tries to use compressed textures
project.addAssets('Assets/**', { quality: 0.7 });
					

khafiles

project.addAssets('Assets/**', { quality: 0.7, size: 0.5 });
if (target == Target.Windows)
	project.addAssets('Assets/**', { name: 'hq-{name}' });
					

khafiles

project.addTarget('windows-steam', ['Windows', 'Direct3D11']);
if (target == 'windows-steam') {
	// ...
}
					

khafiles

project.addShaders('Shaders/**', {
	name: 'bla-{name}', defines: ['bla=1']
});
					

Shader News

• Tesselation Shaders
• Compute Shaders
• Automatic compilation and use of shader variants
  • Specialized shaders for number of supported textures
  • Specialized shaders for instanced rendering backwards compatibility

G5

• Vulkan/Direct3D12 style API
• G4 on G5 (first version in Kore)
• G5 on G4 (coming soon)
• Lower level control, no magic performance boost compared to G4
• Considerable, magic performance boost in G2 coming up
  • Texture atlas usage outdated when using G5

http://luboslenco.com/notes/

Audio

• Object based surround audio
• Dolby Atmos/DTS:X support in research
• Surround simulation on headphones available
  (only htm5 currently, based on hrtf-panner-js)

Native Performance

It's a lie

Your Flash and OpenFL experience does not apply

Haxe/JavaScript regularly outperforms Haxe/C++

Chrome

• Fast code execution for Haxe thanks to V8
• WebGL's feature set is severely outdated
• WebGL is not as fast as it could be because of security
• Still no compute shader support

Krom

• Fast code execution for Haxe thanks to V8
• No WebGL restrictions
• Speed is priority 1
• Compute shaders coming

Krom

• Alternative for deployment
• Typically faster for unoptimized code
• Exception: iOS

Krom

• Special development features
• Soon to be the new default debugger in Kode Studio

Kode Studio

• Fork of Visual Studio Code
• Uses the vshaxe extension
  • Developed in tandem with Haxe 3.3
• Uses JavaScript for Haxe debugging
  • Much faster compilation than C++
  • Much better feature set and runtime performance than Flash (especially with Krom)
  • Improvements in Haxe 3.3

Kode Studio

Demo