vulkan_support#

Shared Vulkan compute-pipeline helpers for kernelfoundry tasks.

A Vulkan task only differs from another in its compiled shader, buffer count, and push-constant layout; everything else (instance/device setup, descriptor sets, command buffer recording, submission) is identical boilerplate. This module builds that boilerplate once via create_vulkan_compute_context(), returning a VulkanCtx that a task’s vulkan_ctx fixture can expose directly.

Functions

create_vulkan_compute_context(spv_path, ...)

Builds a VkInstance/VkDevice/compute pipeline for a compiled SPIR-V module.

Classes

VulkanBuffer(buffer, memory, size)

A storage buffer plus its backing device memory and byte size.

VulkanCtx(device, queue, command_buffer, ...)

Holds the Vulkan objects (device, pipeline, descriptor set, command buffer) needed to dispatch a compute shader.

class kernelfoundry.eval_pipeline.utils.vulkan_support.VulkanBuffer(buffer: Any, memory: Any, size: int)[source]#

A storage buffer plus its backing device memory and byte size.

buffer: Any#
memory: Any#
size: int#
__init__(buffer: Any, memory: Any, size: int) → None#
class kernelfoundry.eval_pipeline.utils.vulkan_support.VulkanCtx(device: Any, queue: Any, command_buffer: Any, descriptor_set: Any, pipeline: Any, pipeline_layout: Any, mem_properties: Any)[source]#

Holds the Vulkan objects (device, pipeline, descriptor set, command buffer) needed to dispatch a compute shader. Built once per test session via create_vulkan_compute_context().

device: Any#
queue: Any#
command_buffer: Any#
descriptor_set: Any#
pipeline: Any#
pipeline_layout: Any#
mem_properties: Any#
find_memory_type(type_bits: int, flags: int) → int[source]#

Finds a device memory type index matching type_bits and property flags.

make_buffer(size_bytes: int) → VulkanBuffer[source]#

Allocates a host-visible/coherent storage buffer of the given byte size.

upload(vbuf: VulkanBuffer, data: ndarray) → None[source]#

Copies a numpy array into a storage buffer’s host-visible memory.

download(vbuf: ~kernelfoundry.eval_pipeline.utils.vulkan_support.VulkanBuffer, dtype=<class 'numpy.float32'>) → ndarray[source]#

Waits for all submitted work to complete, then reads a storage buffer’s host-visible memory back into a numpy array.

submit(buffers: Sequence[VulkanBuffer], n: int, push_constants: bytes = b'', local_size_x: int = 256) → None[source]#

Records (if needed) and submits ceil(n / local_size_x) workgroups to the queue, without waiting for completion. Call wait() (or download(), which waits internally) once submissions are batched, to synchronize.

wait() → None[source]#

Blocks until all work submitted to the queue has completed.

dispatch(buffers: Sequence[VulkanBuffer], n: int, push_constants: bytes = b'', local_size_x: int = 256) → None[source]#

Binds buffers to the pipeline’s descriptor set (in order, starting at binding 0), pushes push_constants if any, and dispatches ceil(n / local_size_x) workgroups, blocking until the GPU is done. Convenience for one-off calls; for tight loops, use submit() repeatedly and wait() once to avoid a full CPU-GPU sync on every call.

__init__(device: Any, queue: Any, command_buffer: Any, descriptor_set: Any, pipeline: Any, pipeline_layout: Any, mem_properties: Any) → None#
kernelfoundry.eval_pipeline.utils.vulkan_support.create_vulkan_compute_context(spv_path: str | Path, num_storage_buffers: int, push_constant_size: int = 0, app_name: str = 'kernelfoundry') → VulkanCtx[source]#

Builds a VkInstance/VkDevice/compute pipeline for a compiled SPIR-V module.

Picks the first integrated/discrete GPU with a compute-capable queue family, exactly as other kernelfoundry backends pick their first available device.

Parameters:
  • spv_path – Path to the compiled .spv shader module (see kernelfoundry.compiler.compile_glsl_to_spirv()).

  • num_storage_buffers – Number of SSBO bindings (0, 1, 2, …) the shader declares.

  • push_constant_size – Byte size of the shader’s push_constant block, 0 if none.

  • app_name – VkApplicationInfo application name (cosmetic).

Returns:

A VulkanCtx ready to receive make_buffer()/upload()/dispatch()/download() calls.