CUDA HAL driver

This document lists technical details regarding the CUDA implemenation of IREE's Hardware Abstraction Layer, called a CUDA HAL driver.

IREE provides a Hardware Abstraction Layer (HAL) as a common interface to different compute accelerators. IREE HAL's design draws inspiration from modern GPU architecture and APIs; so implementing a HAL driver using CUDA is mostly straightforward; though there are places we need emulation given no direct mapping concepts or mechanisms.

Overall design choices

CUDA driver vs runtime API

IREE HAL's design draws inspiration from modern GPU APIs--it provides explicit control of low-level GPU objects. The compiler is expected to plan the object lifetime and schedule workload and synchronization in an optimized way; IREE HAL implementation and the underlying GPU driver stack is expected to be a thin layer without much smarts and magic.

Therefore when implementing the IREE HAL using CUDA, we use the driver API instead of the runtime API. At runtime the HAL CUDA driver will load the libcuda.so/nvcuda.dll library dynamically and query a subset of the CUDA driver API used in HAL via the cuGetProcAddress() API.

GPU Objects

Driver

There is no direct CUDA construct that map to the IREE HAL iree_hal_driver_t abstraction. We use it to hold the dynamic symbols loaded for all devices, and device enumeration and creation.

Device

iree_hal_cuda_device_t implements iree_hal_device_t to provide the interface to CUDA GPU device by wrapping a CUdevice. For each device, right now we create two CUstreams--one for issuing commands for memory allocation and kernel lauches as instructed by the program; the other for issue host callback functions after dispatched command buffers completes. See synchronization section regarding the details.

Async allocation

The CUDA HAL driver supports async allocation (iree_hal_device_queue_alloca() and iree_hal_device_queue_dealloca()) via CUDA stream ordered memory allocation.

The async_allocations in the iree_hal_cuda_device_params_t struct allows to enable this feature.

Command buffer

iree_hal_command_buffer_t is a recording of commands to issue to the GPU; when the command buffer is submitted to the device it's then actually executed on the GPU asynchronously.

Two implementations of iree_hal_command_buffer_t exist in the CUDA HAL driver--one backed by CUgraph and the other backed by CUstream.

CUgraph conceptually matches iree_hal_command_buffer_t better given it's a recording of commands to issue to the GPU. Also using the CUgraph API allows to easily encode fine grain dependencies between dispatch without having to create multiple streams. Therefore, the CUgraph-backed implementation is a more natural one. Though note that CUgraph API is meant to be used for recording once and replying multiple times and there may be a performance penalty to using CUgraph API for one-shot command buffer.

The CUstream-backed implementation just issues commands directly to a CUstream when recording. Commands issued to CUstream can be immediately sent to the GPU for execution; there is no recording and replaying separation. In order to match the recording semantics of iree_hal_command_buffer_t, to use the CUstream-backed command buffer, we need to first record the command buffer into an in-memory iree_hal_deferred_command_buffer_t, and then when applying the command buffer, we create a new CUstream-backed implementation.

The command_buffer_mode in the iree_hal_cuda_device_params_t struct allows to select which implementation to use.

Allocator

The allocator will forward allocation requests to cuMemHostAlloc() for host local memory, cuMemAlloc() for device local and host invisible memory, and cuMemAllocManaged() for device local and host visible memory.

Buffer

CUDA buffers are represented either as a host pointer or a device pointer of type CUdeviceptr.

Executable

iree_hal_executable_t maps naturally to CUmodule.

The compiler generates a FlatBuffer containing a PTX image as well as a list of entry point functions and their associated metadata (names, workgroup size, dynamic shared memory size, etc.). At runtime, the CUDA HAL driver loads the PTX image and creates CUfunctions out of it for various entry points.

Synchronization

Event

iree_hal_event_t right now is not used in the compiler so it's not yet implemented in the CUDA HAL driver.

Semaphore

The IREE HAL uses semaphores to synchronize work between host CPU threads and device GPU streams. It's a unified primitive that covers all directions--host to host, host to device, device to host, and device to device, and allows flexible signal and wait ordering--signal before wait, or wait before signal. There is no limit on the number of waits of the same value too.

The core state of a HAL semaphore consists of a monotonically increasing 64-bit integer value, which forms a timeline--signaling the semaphore to a larger value advances the timeline and unblocks work waiting on some earlier values. The semantics closely mirrors Vulkan timeline semaphore.

In CUDA, there is no direct equivalent primitives providing all the capabilities needed by the HAL semaphore abstraction:

• Stream memory operations provides cuStreamWriteValue64() and cuStreamWaitValue64(), which can implment HAL semaphore 64-bit integer value signal and wait. Though these operations require device pointers and cannot accepts pointers to managed memory buffers, meaning no support for the host. Additionally, per the spec, "synchronization ordering established through these APIs is not visible to CUDA. CUDA tasks that are (even indirectly) ordered by these APIs should also have that order expressed with CUDA-visible dependencies such as events." So it's not suitable for integration with other CUDA components.
• For external resource interoperability, we have APIs like cuSignalExternalSemaphoresAsync() and cuWaitExternalSemaphoresAsync(), which can directly map to Vulkan timeline semaphores. Though these APIs are meant to handle exernal resources--there is no way to create CUexternalSemaphore objects directly other than cuImportExternalSemaphore().

Therefore, to implement the support, we need to leverage multiple native CPU or CUDA primitives under the hood.

CUevent capabilities

The main synchronization mechanism is CUDA event--CUevent. As a functionality and integration baseline, we use CUevent to implement the IREE HAL semaphore abstraction.

CUevent natively supports the following capabilities:

• State: binary; either unsignaled or signaled. There can exist multiple waits (e.g., via cuEventSynchronize() or cuGraphAddEventWaitNode()) for the same CUevent signal (e.g., via cuEventRecord() or cuGraphAddEventRecordNode()).
• Ordering: must be signal before wait. Waiting before signal would mean waiting an empty set of work, or previously recorded work.
• Direction: device to device, device to host.

We need to fill the remaining capability gaps. Before going into details, the overall approach would be to:

• State: we need a 64-bit integer value timeline. Given the binary state of a CUevent, each CUevent would just be a "timepoint" on the timeline.
• Ordering: we need to defer releasing the workload to the GPU until the semaphore waits are reached on the host, or we can have some device CUevent to wait on.
• Direction: host to host and host to device is missing; we can support that with host synchronization mechanisms.

Signal to wait analysis

Concretely, for a given HAL semaphore, looking at the four directions:

CPU signal

A CPU thread signals the semaphore timeline to a new value.

If there are CPU waits, it is purely on the CPU side. We just need to use common CPU notification mechanisms. In IREE we have iree_event_t wrapping various low-level OS primitives for it. So we can just use that to represent a wait timepoint. We need to keep track of all CPU wait timepoints in the timeline. After a new signaled value, go through the timeline and notify all those waiting on earlier values.

If there are GPU waits, given that there are no way we can signal a CUevent on CPU, one way to handle this is to cache and defer the submission batches by ourselves until CPU signals past the desired value. To support this, we would need to implement a deferred/pending actions queue.

GPU signal

GPU signals can only be through a CUevent object, which has a binary state. We need to advance the timeline too. One way is to use cuLaunchHostFunc() to advance from the CPU side with iree_hal_semaphore_list_signal(). This additionally would mean we can reuse the logic form CPU signaling to unblock CPU waits.

After advancing the timeline from the CPU side with cuLaunchHostFunc(), we can release more workload from the deferred/pending actions queue to the GPU. Though, per the documentation of cuLaunchHostFunc(), "the host function must not make any CUDA API calls." So we cannot do that directly inside cuLaunchHostFunc(); we need to notify another separate thread to call CUDA APIs to push more work to the GPU. So the deferred/pending action queue should have an associcated thread.

For GPU waits, we can also leverage the same logic--using CPU signaling to unblock deferred GPU queue actions. Though this is performant, given that the CPU is involved for GPU internal synchronization. We want to use CUevent instead:

• We keep track of all GPU signals in the timeline. Once we see a GPU wait request, try to scan the timeline to find a GPU signal that advances the timeline past the desired value, and use that for waiting instead. (This actually applies to CPU waits too, and it's an optimization over pure CPU side iree_event_t polling.)
• We may not see GPU signal before seeing GPU wait requests, then we can also keep track of all GPU waits in the timeline. Later once see either a CPU signal or GPU signal advancing past the waited value, we can handle them accordingly--submitting immediately or associating the CUevent. This would also guarantee the requirement of CUevent--recording should happen before waiting.
• We can use the same CUevent to unblock multiple GPU waits. That's allowed, though it would mean we need to be careful regarding CUevent lifetime management. Here we can use reference counting to see how many timepoints are using it and automatically return to a pool once done.

Another problem is that per the cuLaunchHostFunc() doc, "the function will be called after currently enqueued work and will block work added after it." We don't want the blocking behavior involving host. So we can use a dedicated CUstream for launching the host function, waiting on the CUevent from the original stream too. We can also handle resource deallocation together there.

Data structures

To summarize, we need the following data structures to implement HAL semaphore:

• iree_event_t: CPU notification mechanism wrapping low-level OS primitives. Used by host wait timepoints.
• iree_event_pool_t: a pool for CPU iree_event_t objects to recycle.
• iree_hal_cuda_event_t: GPU notification mechanism wrapping a CUevent and a reference count. Used by device signal and wait timepoints. Associates with a iree_hal_cuda_event_pool_t pool--returns to the pool directly on once reference count goes to 0.
• iree_hal_cuda_event_pool_t: a pool for GPU iree_hal_cuda_event_t objects to recycle.
• iree_hal_cuda_timepoint_t: an object that wraps a CPU iree_event_t or GPU iree_hal_cuda_event_t to represent wait/signal of a timepoint on a timeline.
• iree_hal_cuda_timepoint_pool_t: a pool for iree_hal_cuda_timepoint_t objects to recycle. This pool builds upon the CPU and GPU event pool--it acquires CPU/GPU event objects there.
• iree_hal_cuda_timeline_semaphore_t: contains a list of CPU wait and GPU wait/signal timepoints.
• iree_hal_cuda_queue_action_t: a pending queue action (kernel launch or stream-ordered allocation).
• iree_hal_cuda_pending_queue_actions_t: a data structure to manage pending queue actions. It provides APIs to enqueue actions, and advance the queue on demand--queue actions are released to the GPU when all their wait semaphores are signaled past the desired value, or we can have a CUevent object already recorded to some CUstream to wait on.