QSB Manual

There are three major modes of operation:

• .qsb file generation.

• .qsb file inspection. For example, qsb -d myshader.frag.qsb will print the reflection metadata (in JSON form) and the included shaders.

• Extraction mode. This allows writing a given shader from an existing .qsb file into a separate file. For example, qsb -x spirv.100 -o myshader.spv myshader.frag.qsb writes the SPIR-V binary into myshader.spv.

Take the following fragment shader:

#version 440

layout(location = 0) in vec2 v_texcoord;
layout(location = 0) out vec4 fragColor;
layout(binding = 1) uniform sampler2D tex;

layout(std140, binding = 0) uniform buf {
    float uAlpha;
};

void main()
{
    vec4 c = texture(tex, v_texcoord);
    fragColor = vec4(c.rgb, uAlpha);
}

Executing qsb -o shader.frag.qsb shader.frag results in generating shader.frag.qsb. Inspecting this file with qsb -d shader.frag.qsb gives us:

Stage: Fragment
QSB_VERSION: 5
Has 1 shaders: (unordered list)
  Shader 0: SPIR-V 100 [Standard]

Reflection info: {
    "combinedImageSamplers": [
        {
            "binding": 1,
            "name": "tex",
            "set": 0,
            "type": "sampler2D"
        }
    ],
    "inputs": [
        {
            "location": 0,
            "name": "v_texcoord",
            "type": "vec2"
        }
    ],
    "localSize": [
        0,
        0,
        0
    ],
    "outputs": [
        {
            "location": 0,
            "name": "fragColor",
            "type": "vec4"
        }
    ],
    "uniformBlocks": [
        {
            "binding": 0,
            "blockName": "buf",
            "members": [
                {
                    "name": "uAlpha",
                    "offset": 0,
                    "size": 4,
                    "type": "float"
                }
            ],
            "set": 0,
            "size": 4,
            "structName": "_27"
        }
    ]
}


Shader 0: SPIR-V 100 [Standard]
Entry point: main
Contents:
Binary of 864 bytes

By default only SPIR-V is generated, so an application using this shader package would only be functional with Vulkan. Let's make it more useful:

qsb --glsl "100 es,120,150" --hlsl 50 --msl 12 -o shader.frag.qsb shader.frag

This leads to generating a shader package that makes it suitable for OpenGL, Direct 3D, and Metal as well. The features used in this shader are basic, and even GLSL ES 100 (the shading language of OpenGL ES 2.0) is suitable.

Inspecting the result shows:

Stage: Fragment
QSB_VERSION: 5
Has 6 shaders: (unordered list)
  Shader 0: GLSL 120 [Standard]
  Shader 1: HLSL 50 [Standard]
  Shader 2: GLSL 100 es [Standard]
  Shader 3: MSL 12 [Standard]
  Shader 4: SPIR-V 100 [Standard]
  Shader 5: GLSL 150 [Standard]

Reflection info: {
    ... <same as above>
}


Shader 0: GLSL 120 [Standard]
Entry point: main
Contents:
#version 120

struct buf
{
    float uAlpha;
};

uniform buf _27;

uniform sampler2D tex;

varying vec2 v_texcoord;

void main()
{
    vec4 c = texture2D(tex, v_texcoord);
    gl_FragData[0] = vec4(c.xyz, _27.uAlpha);
}

************************************

Shader 1: HLSL 50 [Standard]
Entry point: main
Native resource binding map:
0 -> [0, -1]
1 -> [0, 0]
Contents:
cbuffer buf : register(b0)
{
    float _27_uAlpha : packoffset(c0);
};

Texture2D<float4> tex : register(t0);
SamplerState _tex_sampler : register(s0);

static float2 v_texcoord;
static float4 fragColor;

struct SPIRV_Cross_Input
{
    float2 v_texcoord : TEXCOORD0;
};

struct SPIRV_Cross_Output
{
    float4 fragColor : SV_Target0;
};

void frag_main()
{
    float4 c = tex.Sample(_tex_sampler, v_texcoord);
    fragColor = float4(c.xyz, _27_uAlpha);
}

SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
{
    v_texcoord = stage_input.v_texcoord;
    frag_main();
    SPIRV_Cross_Output stage_output;
    stage_output.fragColor = fragColor;
    return stage_output;
}

************************************

...

Shader 3: MSL 12 [Standard]
Entry point: main0
Native resource binding map:
0 -> [0, -1]
1 -> [0, 0]
Contents:
#include <metal_stdlib>
#include <simd/simd.h>

using namespace metal;

struct buf
{
    float uAlpha;
};

struct main0_out
{
    float4 fragColor [[color(0)]];
};

struct main0_in
{
    float2 v_texcoord [[user(locn0)]];
};

fragment main0_out main0(main0_in in [[stage_in]], constant buf& _27 [[buffer(0)]], texture2d<float> tex [[texture(0)]], sampler texSmplr [[sampler(0)]])
{
    main0_out out = {};
    float4 c = tex.sample(texSmplr, in.v_texcoord);
    out.fragColor = float4(c.xyz, _27.uAlpha);
    return out;
}

************************************

...

This package can now be used by Qt Quick with all supported graphics APIs: Vulkan, Direct 3D, Metal, OpenGL, and OpenGL ES. At run time the appropriate shader is picked up automatically by the Qt Rendering Hardware Interface that sits underneath Qt Quick and Qt Quick 3D.

Besides translating the SPIR-V bytecode back to higher level source code, the system takes care of additional problems, such as ensuring correct mapping of SPIR-V binding numbers onto native resources. For example, with HLSL we saw a section like this above:

Native resource binding map:
 0 -> [0, -1]
 1 -> [0, 0]

Internally, this allows mapping a SPIR-V style binding point 0 to the HLSL register b0 and binding 1 to t0 and s0. This helps making the differences in resource bindings between the various shading languages transparent to the users of the Rendering Hardware Interface, and allows everything in Qt to operate with Vulkan/SPIR-V style binding points as they are specified in the original Vulkan-style GLSL source code.

The type of shader is deduced from the input file extension. Thus, the extension must be one of the following:

• .vert - for vertex shaders

• .frag - for fragment (pixel) shaders

• .comp - for compute shaders

SPIR-V 1.0 is always generated. What gets generated in addition depends on the command-line arguments --glsl, --hlsl, and --msl.

These parameters are all followed by a comma-separated list. The list must include GLSL-style version numbers, with an optional suffix (es, indicating GLSL ES). The space between the suffix and the version is optional (not having the space can help avoiding the need for quoting).

For example, Qt Quick's built-in materials (the shaders backing items, such as Image, Text, Rectangle) all prepare their shaders with --glsl "100 es,120,150" --hlsl 50 --msl 12. This makes them compatible with OpenGL ES 2.0 and higher, OpenGL 2.1 and higher, and OpenGL core profile contexts of version 3.2 and higher.

If the shader uses functions or constructs that do not have an equivalent in the specified targets, qsb will fail. If that is the case, the targets will need to be adjusted, and this also means that the application's minimum system requirements get adjusted implicitly. As an example, take the textureLod GLSL function that is only available with OpenGL ES 3.0 and up (meaning GLSL ES 300 or higher). When requesting GLSL 300 es instead of 100 es, qsb will succeed, but the application using that .qsb file will now require OpenGL ES 3.0 or higher and will not be compatible with OpenGL ES 2.0 based systems.

Another obvious example of this is compute shaders: .comp shaders will need to specify --glsl 310es,430 as compute shaders are only available with OpenGL ES 3.1 or newer and OpenGL 4.3 or newer.

Adjusting the shader model version for HLSL, or the Metal Shading Language version is expected to be needed rarely. Shader Model 5.0 (--hlsl 50) and MSL 1.2 (--msl 12) will typically be sufficient.

The Qt Quick Scene Graph's renderer supports batching of geometry to reduce the number of draw calls. See the Scene Graph pages for details. This relies on injecting code to the vertex shader's main() function. In Qt 5.x this happened at run time, by modifying the supplied GLSL vertex shader code. In Qt 6 that is not an option. Instead, batchable variants of vertex shaders can be built by the qsb tool. This is requested by the -b argument. When the input is not a vertex shader with .vert extension, it has no effect. For vertex shaders, however, it will lead to generating two versions for each target. Qt Quick will then automatically pick the right variant (standard or batchable) at run time.

Take the following example vertex shader:

#version 440
layout(location = 0) in vec4 position;
layout(location = 1) in vec2 texcoord;
layout(location = 0) out vec2 v_texcoord;
layout(std140, binding = 0) uniform buf {
    mat4 mvp;
} ubuf;
out gl_PerVertex { vec4 gl_Position; };

void main()
{
    v_texcoord = texcoord;
    gl_Position = ubuf.mvp * position;
}

Running qsb -b --glsl 330 -o example.vert.qsb example.vert leads to:

Stage: Vertex
QSB_VERSION: 5
Has 4 shaders: (unordered list)
  Shader 0: GLSL 330 [Standard]
  Shader 1: GLSL 330 [Batchable]
  Shader 2: SPIR-V 100 [Standard]
  Shader 3: SPIR-V 100 [Batchable]

Reflection info: {
  ...

Note how all target languages and versions now exist in two variants: Standard and a slightly modified Batchable.

qsb can invoke certain external tools. These fall into two categories: tools for performing optimizations on the shader bytecode (SPIR-V), and platform-specific tools to perform the first phase of shader compilation, from source to some intermediate bytecode format.

These are enabled by the following command-line options:

• -O - invokes spirv-opt as a post-processing step on the SPIR-V binary. The .qsb file will include the optimized version. This assumes that spirv-opt is available on the system (for example, from the Vulkan SDK) and ready to be invoked.

• -c or --fxc - invokes fxc.exe, the Direct 3D shader compiler. The resulting DXBC (DirectX Byte Code) data is what gets stored in the .qsb file instead of HLSL. Qt will automatically pick it up at run time, so it is up to the .qsb file's creator to decide what to include, HLSL source or the intermediate format. Whenever possible, prefer the latter since it eliminates the need for parsing an HLSL source at run time, leading to potentially significant performance gains upon graphics pipeline creation. The downside is that this argument can only be used when qsb runs on Windows.

• -t or --metallib - invokves the appropriate XCode Metal tools to generate a .metallib file and includes that in the .qsb package instead of the MSL source code. This option is only available when qsb is running on macOS.

• -D - defines a macro. This allows using #ifdef and similar in the GLSL source code.

• -g - enables generating full debug information for SPIR-V, thus enabling tools like RenderDoc to display the full source when inspecting a pipeline or when performing vertex or fragment debugging. Also has an effect for Direct 3D when -c is specified, because fxc is then instructed to include debug information in the generated intermediate bytecode.

It can sometimes be necessary to use shading language constructs that are specific to OpenGL and GLSL, and are not applicable to other shading languages, intermediate formats, and graphics APIs.

The prime examples of this are the external textures and samplers of OpenGL ES. Implementing video playback or showing a camera viewfinder may involve, depending on the platform, working with OpenGL texture objects that are not meant to be used as regular 2D textures, but are rather usable, with a limited feature set, via the GL_TEXTURE_EXTERNAL_OES binding point in the OpenGL API and the samplerExternalOES sampler type in shaders. The latter presents a potential showstopper when using the SPIR-V based shader pipeline of Qt: running such a shader through qsb will result in a failure due to samplerExternalOES not being accepted as a valid type due to not being mappable to SPIR-V and other target shading languages.

To overcome this, qsb offers the option to replace the contents of any given shader variant in a .qsb file with user-provided data that is read from a file, fully replacing the original qsb-generated shader source or byte code.

Take the following fragment shader. Note the type of tex. What if the type needs to be samplerExternalOES when running with OpenGL ES?

#version 440

layout(location = 0) in vec2 texCoord;
layout(location = 0) out vec4 fragColor;

layout(std140, binding = 0) uniform buf {
    float opacity;
} ubuf;

layout(binding = 1) uniform sampler2D tex;

void main()
{
    fragColor = texture(tex, texCoord).rgba * ubuf.opacity;
}

Just changing the type of samplerExternalOES is not feasible. That leads to compilation errors right away.

There is a simple solution, however: writing a separate, pure OpenGL ES targeted version of the shader, and injecting it into the .qsb file. The following shader is only compatible with GLSL ES and cannot be run through qsb. However, we know that it can be processed by OpenGL ES at run time.

precision highp float;
#extension GL_OES_EGL_image_external : require
varying vec2 texCoord;

struct buf
{
    float opacity;
};

uniform buf ubuf;
uniform samplerExternalOES tex;

void main()
{
    gl_FragColor = texture2D(tex, texCoord).rgba * ubuf.opacity;
}

Let's call this shader_gles.frag. Once qsb --glsl 100es -o shader.frag.qsb shader.frag completes, giving us a (half-ready) .qsb file, we can do qsb -r glsl,100es,shader_gles.frag shader.frag.qsb to update shader.frag.qsb by substituting the shader for GLSL 100 es with the contents of the specified file (shader_gles.frag). Now shader.frag.qsb is ready to be used at run time with OpenGL ES.