ANGLE: Vulkan Back-end
June 25, 2026 ยท View on GitHub
ANGLE's Vulkan back-end implementation lives in this folder.
Vulkan is an explicit graphics API. Compared to APIs like OpenGL or D3D11 explicit APIs can offer a number of significant benefits:
- Lower API call CPU overhead.
- A smaller API surface with more direct hardware control.
- Better support for multi-core programming.
- Vulkan in particular has open-source tooling and tests.
Back-end Design
The vk::Renderer class represents an EGLDisplay. vk::Renderer owns shared global
resources like the VkDevice, VkQueue, the Vulkan format tables
and internal Vulkan shaders. The ContextVk class implements the back-end
of a front-end OpenGL Context. ContextVk processes state changes and handles action commands like
glDrawArrays and glDrawElements.
Command recording
A render pass has three states: unstarted, started and active (we call it active in short),
started but inactive (we call it inactive in short). The back-end records commands into command
buffers via the following ContextVk APIs:
beginNewRenderPass: Writes out (aka flushes) prior pending commands into a primary command buffer, then starts a new render pass. Returns a secondary command buffer inside a render pass instance.getOutsideRenderPassCommandBuffer: May flush prior command buffers and close the render pass if necessary, in addition to issuing the appropriate barriers. Returns a secondary command buffer outside a render pass instance.getStartedRenderPassCommands: Returns a reference to the currently open render pass' commands buffer.onRenderPassFinished: Puts render pass into inactive state where you can not record more commands into secondary command buffer, except in some special cases where ANGLE does some optimization internally.flushCommandsAndEndRenderPassWithoutSubmit: Marks the end of render pass. It flushes secondary command buffer into vulkan's primary command buffer, puts secondary command buffer back to unstarted state and then goes into recycler for reuse.
The back-end (mostly) records Image and Buffer barriers through additional CommandResources
APIs, the result of which is passed to getOutsideRenderPassCommandBuffer. Note that the barriers
are not actually recorded until getOutsideRenderPassCommandBuffer is called:
onBufferTransferReadandonBufferComputeShaderReadaccumulateVkBufferread barriers.onBufferTransferWriteandonBufferComputeShaderWriteaccumulateVkBufferwrite barriers.onBuffferSelfCopyis a special case forVkBufferself copies. It behaves the same as write.onImageTransferReadandonImageComputerShadeReadaccumulateVkImageread barriers.onImageTransferWriteandonImageComputerShadeWriteaccumulateVkImagewrite barriers.onImageRenderPassReadandonImageRenderPassWriteaccumulateVkImagebarriers inside a started RenderPass.
After the back-end records commands to the primary buffer and we flush (e.g. on swap) or when we call
vk::Renderer::finishQueueSerial, ANGLE submits the primary command buffer to a VkQueue.
See the code for more details.
Simple command recording example
In this example we'll be recording a buffer copy command:
// Ensure that ANGLE sets proper read and write barriers for the Buffers.
vk::CommandResources resources;
resources.onBufferTransferWrite(dstBuffer);
resources.onBufferTransferRead(srcBuffer);
// Get a pointer to a secondary command buffer for command recording.
vk::OutsideRenderPassCommandBuffer *commandBuffer;
ANGLE_TRY(contextVk->getOutsideRenderPassCommandBuffer(resources, &commandBuffer));
// Record the copy command into the secondary buffer. We're done!
commandBuffer->copyBuffer(srcBuffer->getBuffer(), dstBuffer->getBuffer(), copyCount, copies);
Handling EGL image offsets
A Texture can have a number of levels and layers. An EGL image may be created out of the texture,
viewing a specific level and layer specified by EGL_GL_TEXTURE_LEVEL_KHR and
EGL_GL_TEXTURE_ZOFFSET_KHR. The EGL image can then be imported into another texture whose own
state does not include these offsets (e.g. the texture is accessed at level 0). It may similarly be
imported into a renderbuffer that doesn't even have a level and layer state.
The TextureVk and RenderbufferVk classes use gl::OwnLevel, gl::OwnLayer, and
gl::OwnImageIndex to refer to the level and layer from the point of view of the
texture/renderbuffer itself, i.e. these values will be 0 for an EGL image target. When these values
are needed as integers, e.g. to access an array, they are retrieved via getUntranslated().
However, any references outside these classes to levels and layers, i.e. those that are paired with
VkImage and VkImageView, must use the source texture's original level and layer. This is done
via the mState.toSource*() helpers, which return a strongly typed gl::SourceLevel,
gl::SourceLayer and gl::SourceImageIndex. The level and layer can be retrieved from these types
with a get().
Ideally, all the code in the front-end would use gl::Own* types and all the code in the backend
outside these classes would use gl::Source* for maximum safety, where the compiler ensures there
are no mistakes. This is not yet done, but in the meantime, any reference to level and layer in the
front-end is implicitly from the point of view of the texture/renderbuffer itself (i.e. equivalent
to gl::Own*). Any such reference outside TextureVk and RenderbufferVk in the backend is
implicitly from the point of view of the owner of the VkImage (i.e. equivalent to gl::Source*).
The code in TextureVk and RenderbufferVk must always use mState.toSource* to convert the types
before using a level or layer in conjunction with an ImageHelper, ImageViewHelper,
RenderTargetVk, UtilsVk etc calls.
Additional Reading
More implementation details can be found in the doc directory: