Dirty Bits and State Changes

OpenGL render loops typically involve changing some render states followed by a draw call. For instance the app might change a few uniforms and invoke glDrawElements:

for (const auto &obj : scene) {
    for (const auto &uni : obj.uniforms) {
        glUniform4fv(uni.loc, uni.data);
    }
    glDrawElements(GL_TRIANGLES, obj.eleCount, GL_UNSIGNED_SHORT, obj.eleOffset);
}

Another update loop may change Texture and Vertex Array state before the draw:

for (const auto &obj : scene) {
    glBindBuffer(GL_ARRAY_BUFFER, obj.arrayBuffer);
    glBufferSubData(GL_ARRAY_BUFFER, obj.bufferOffset, obj.bufferSize, obj.bufferData);
    glVertexAttribPointer(obj.arrayIndex, obj.arraySize, GL_FLOAT, GL_FALSE, 0, nullptr);
    glBindTexture(GL_TEXTURE_2D, obj.texture);
    glDrawElements(GL_TRIANGLES, obj.eleCount, GL_UNSIGNED_SHORT, obj.eleOffset);
}

Other update loops may change render states like the blending modes, the depth test, or Framebuffer attachments. In each case ANGLE needs to validate, track, and translate these state changes to the back-end as efficiently as possible.

Dirty Bits

Each OpenGL Context state value is stored in gl::State. For instance the blending state, depth/stencil state, and current object bindings. Our problem is deciding how to notify the back-end when app changes front-end state. We decided to bundle changed state into bitsets. Each 1 bit indicates a specific changed state value. We call these bitsets “dirty bits”. See gl::State::DirtyBitType.

Each back-end handles state changes in a syncState implementation function that takes a dirty bitset. See examples in the GL back-end, D3D11 back-end and Vulkan back-end.

Container objects such as Vertex Array Objects and Framebuffers also have their own OpenGL front-end state. VAOs store vertex arrays and array buffer bindings. Framebuffers store attachment state and the active read and draw buffers. These containers also have internal dirty bits and syncState methods. See gl::Framebuffer::DirtyBitType and rx::FramebufferVk::syncState for example.

Dirty bits allow us to efficiently process groups of state updates. We use fast instrinsic functions to scan the bitsets for 1 bits. See bitset_utils.h for more information.

Cached Validation and State Change Notifications

To optimize validation we cache many checks. See gl::StateCache for examples. We need to refresh cached values on state changes. For instance, enabling a generic vertex array changes a cached mask of active vertex arrays. Changes to a texture‘s images could change a cached framebuffer’s completeness when the texture is bound as an attachment. And if the draw framebuffer becomes incomplete it changes a cached draw call validation check.

See a below example of a call to glTexImage2D that can affect draw call validation:

We use the Observer pattern to implement cache invalidation notifications. See Observer.h. In the example the Framebuffer observes Texture attachments via angle::ObserverBinding. Framebuffer implements angle::ObserverInterface::onSubjectStateChange to receive a notification to update its completeness cache. The STORAGE_CHANGED message triggers a call to gl::Context::onSubjectStateChange which in turn calls gl::StateCache::updateBasicDrawStatesError to re-validate the draw framebuffer's completeness. On subsequent draw calls we skip re-validation at minimal cost.

See the below diagram for the dependency relations between Subjects and Observers.

Back-end specific Optimizations

See Fast OpenGL State Transitions in Vulkan documents for additional information for how we implement state change optimization on the Vulkan back-end.