q3::execution::gpu::device::GPUDeviceBackend Class Reference

#include <init.hpp>

Inheritance diagram for q3::execution::gpu::device::GPUDeviceBackend:

Public Member Functions
virtual	~GPUDeviceBackend ()=default
virtual void *	allocate_unified (size_t size_bytes)=0
virtual void	deallocate (void *pointer)=0
virtual std::string	device_name () const =0
virtual bool	supports_double_precision () const =0
virtual void	synchronize ()=0
virtual size_t	total_memory_bytes () const =0

Detailed Description

Definition at line 22 of file init.hpp.

Constructor & Destructor Documentation

~GPUDeviceBackend()

virtual q3::execution::gpu::device::GPUDeviceBackend::~GPUDeviceBackend ( )

virtualdefault

Member Function Documentation

allocate_unified()

virtual void * q3::execution::gpu::device::GPUDeviceBackend::allocate_unified ( size_t  size_bytes )

pure virtual

Allocate unified (managed) memory accessible from both host and device.

On CUDA, this maps to cudaMallocManaged(). On Metal, this maps to MTLDevice::newBuffer() with MTLResourceStorageModeShared.

The returned memory is NOT zero-initialized.

Parameters

size_bytes

Number of bytes to allocate.

Returns

Pointer to the allocated memory.

Exceptions

std::runtime_error

If allocation fails.

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

Referenced by q3::cli::validate_gpu_device_backend().

deallocate()

virtual void q3::execution::gpu::device::GPUDeviceBackend::deallocate ( void *  pointer )

pure virtual

Deallocate memory previously allocated via allocate_unified().

Parameters

pointer

Pointer returned by allocate_unified(). Must not be null.

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

Referenced by q3::cli::validate_gpu_device_backend().

device_name()

virtual std::string q3::execution::gpu::device::GPUDeviceBackend::device_name ( ) const

pure virtual

Human-readable name of the GPU device (e.g., "Apple M2 Max", "NVIDIA A100").

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

Referenced by q3::cli::validate_gpu_device_backend().

supports_double_precision()

virtual bool q3::execution::gpu::device::GPUDeviceBackend::supports_double_precision ( ) const

pure virtual

Whether the device supports native double precision (FP64) computation.

Returns false for Apple Silicon (Metal has no FP64 compute shaders). Returns true for CUDA GPUs (all supported compute capabilities).

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

Referenced by q3::cli::validate_gpu_device_backend().

synchronize()

virtual void q3::execution::gpu::device::GPUDeviceBackend::synchronize ( )

pure virtual

Block until all pending GPU operations have completed.

Must be called before reading GPU-produced results on the host, and before precision casts or MPI collectives.

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

total_memory_bytes()

virtual size_t q3::execution::gpu::device::GPUDeviceBackend::total_memory_bytes ( ) const

pure virtual

Total device memory in bytes.

Implemented in q3::execution::gpu::device::CudaGPUDeviceBackend, q3::execution::gpu::device::HIPGPUDeviceBackend, and q3::execution::gpu::device::SYCLGPUDeviceBackendImpl.

Referenced by q3::cli::validate_gpu_device_backend().

---

The documentation for this class was generated from the following file:

• src/layer_0_core/execution/gpu/device/init.hpp