Cube Maps (🟠WIP)

With webgpu.hpp

Resulting code: step117

Vanilla webgpu.h

Resulting code: step117-vanilla

The computation of the ibl_uv coordinates at which we sampled the environment lighting in the previous chapter is a bit costly, due to the acos and atan2 operations. A more efficient way to store the environment map is as a cube map.

TODO

Cube maps are more efficient to sample and hardware accelerated.

Multi-layer textures

We will see in the Cubemap Conversion chapter how to convert an equirectangular environment map into a cubemap and vice versa.

All we need to know for now is that a cubemap is a special type of texture. It is stored as a 2D array texture with 6 layers, which means that when creating the texture, with specify a dimension of 2D but the size has 3 dimensions:

TextureDescriptor textureDesc;
textureDesc.dimension = TextureDimension::_2D;
textureDesc.size = { size, size, 6 };
// [...]

By convention, the face of the cube are stored in the following order:

Layer	Cube Map Face	S	T
0	Positive X	-Z	-Y
1	Negative X	+Z	-Y
2	Positive Y	+X	+Z
3	Negative Y	+X	-Z
4	Positive Z	+X	-Y
5	Negative Z	-X	-Y

As you can see, the convention also specifies the world-space direction to which the local texture axes S and T correspond.

CubeMaps are represented as 2D array textures.

In practice, we load the faces one by one, from individual files. The computations of MIP levels is also done face by face. The texture sampler will take care of mixing faces together appropriately.

Extent3D singleLayerSize = { size, size, 1 };
for (uint32_t layer = 0; layer < 6; ++layer) {
    destination.origin = { 0, 0, layer };
    m_queue.writeTexture(destination, pixelData[layer], (size_t)(4 * size * size), source, singleLayerSize);
}

Each face of a cube map is loaded from a different image file.

Note

Images appear upside down because the convention was designed by people who use \(Y\) as the vertical axis, and in this guide we use \(Z\) as the vertical. Anyways even when using \(Y\)-up it is better to stick to the convention table above than to try to intuitively guess the correct S and T texture axes.

TODO

Implementation

TODO

Unzip autumn_park_4k.zip in your resource directory.

// In Application.h
bool initTexture(const std::filesystem::path& path, bool isCubemap = false);

// In onInit()
if (!initTexture(RESOURCE_DIR "/autumn_park_4k"), true /* isCubemap */) return false;

// In Application.cpp
bool Application::initTexture(const std::filesystem::path& path, bool isCubemap) {
    TextureView textureView = nullptr;
    Texture texture =
        isCubemap
        ? ResourceManager::loadCubemapTexture(path, m_device, &textureView)
        : ResourceManager::loadTexture(path, m_device, &textureView);

    // [...]

    bindingLayout.texture.viewDimension =
        isCubemap
        ? TextureViewDimension::Cube
        : TextureViewDimension::_2D;

    // [...]
}

// In ResourceManager.h
static wgpu::Texture loadCubemapTexture(const path& path, wgpu::Device device, wgpu::TextureView* pTextureView = nullptr);

// In ResourceManager.cpp
Texture ResourceManager::loadCubemapTexture(const path& path, Device device, TextureView* pTextureView) {
    const char* cubemapPaths[] = {
        "cubemap-posX.png",
        "cubemap-negX.png",
        "cubemap-posY.png",
        "cubemap-negY.png",
        "cubemap-posZ.png",
        "cubemap-negZ.png",
    };

    // Load image data for each of the 6 layers
    Extent3D cubemapSize = { 0, 0, 6 };
    std::array<uint8_t*, 6> pixelData;
    for (uint32_t layer = 0; layer < 6; ++layer) {
        int width, height, channels;
        auto p = path / cubemapPaths[layer];
        pixelData[layer] = stbi_load(p.string().c_str(), &width, &height, &channels, 4 /* force 4 channels */);
        if (nullptr == pixelData[layer]) throw std::runtime_error("Could not load input texture!");
        if (layer == 0) {
            cubemapSize.width = (uint32_t)width;
            cubemapSize.height = (uint32_t)height;
        }
        else {
            if (cubemapSize.width != (uint32_t)width || cubemapSize.height != (uint32_t)height)
                throw std::runtime_error("All cubemap faces must have the same size!");
        }
    }

    // [...]
    textureDesc.size = cubemapSize;

    // [...]
    Extent3D cubemapLayerSize = { cubemapSize.width , cubemapSize.height , 1 };
    for (uint32_t layer = 0; layer < 6; ++layer) {
        Extent3D origin = { 0, 0, layer };

        writeMipMaps(device, texture, cubemapLayerSize, textureDesc.mipLevelCount, pixelData[layer], origin);

        // Free CPU-side data
        stbi_image_free(pixelData[layer]);
    }

    // [...]
    textureViewDesc.arrayLayerCount = 6;
    //                                ^ This was 1
    textureViewDesc.dimension = TextureViewDimension::Cube;
    //                                                ^ This was 2D

Note that we also add a new extra argument to writeMipMaps to specify which layer to upload to:

template<typename component_t>
static void writeMipMaps(
    /* [...] */
    Origin3D origin = { 0, 0, 0 }
) {
    // [...]
    destination.origin = origin;
    // ^                 ^ This was { 0, 0, 0 }

// In shader
@group(0) @binding(4) var cubemapTexture: texture_cube<f32>;

// [...]

let ibl_sample = textureSample(cubemapTexture, textureSampler, ibl_direction).rgb;
//                                                             ^ This was ibl_uv

TODO

Conclusion

With webgpu.hpp

Resulting code: step117

Vanilla webgpu.h

Resulting code: step117-vanilla