Interface GPUBackend

All Superinterfaces:

Backend, ComputeBackend

All Known Implementing Classes:

NativeCUDADenseLinearAlgebraDoubleBackend, NativeCUDADenseLinearAlgebraFloatBackend, NativeCUDANBodyBackend, NativeCUDASparseLinearAlgebraDoubleBackend, NativeCUDASparseLinearAlgebraFloatBackend, NativeCUDATensorBackend, NativeCUDAVisionBackend, NativeOpenCLDenseLinearAlgebraDoubleBackend, NativeOpenCLDenseLinearAlgebraFloatBackend, NativeOpenCLFFTBackend, NativeOpenCLLatticeBoltzmannBackend, NativeOpenCLNBodyBackend, NativeOpenCLSparseLinearAlgebraDoubleBackend, NativeOpenCLSparseLinearAlgebraFloatBackend, NativeOpenCLVisionBackend

public interface GPUBackend
extends ComputeBackend

GPU device management backend using CUDA or OpenCL.

This interface provides GPU device discovery, selection, and memory management. Algorithm-specific GPU operations (FFT, matrix multiply, reductions) are provided by their respective AlgorithmProvider implementations.

Implementation Strategy:

• Use JCuda for NVIDIA GPUs: http://www.jcuda.org/
• Use JOCL for OpenCL devices: http://www.jocl.org/
• Use Project Panama for direct native GPU library access

Since:

1.2

Author:

Silvere Martin-Michiellot, Gemini AI (Google DeepMind)

Nested Class Summary

Nested Classes

Modifier and Type	Interface	Description
static class	GPUBackend.DeviceInfo	GPU device information.

Method Summary

Modifier and Type	Method	Description
long	allocateGPUMemory(long sizeBytes)	Allocates GPU memory and returns a handle.
void	copyFromGPU(long gpuHandle, DoubleBuffer hostBuffer, long sizeBytes)	Copies data from GPU to host.
void	copyToGPU(long gpuHandle, DoubleBuffer hostBuffer, long sizeBytes)	Copies data from host to GPU.
void	freeGPUMemory(long gpuHandle)	Frees GPU memory.
GPUBackend.DeviceInfo[]	getDevices()	Returns information about available GPU devices.
void	matrixMultiply(DoubleBuffer A, DoubleBuffer B, DoubleBuffer C, int m, int n, int k)	Performs matrix multiplication C = A * B on the GPU.
void	selectDevice(int deviceId)	Selects a GPU device for computation.
void	synchronize()	Synchronizes GPU execution (waits for all kernels to complete).

Methods inherited from interface Backend

getAlgorithmProviders, getName, getPriority, getStatusMessage, isAvailable, isExplicitlyDisabled, shutdown

Methods inherited from interface ComputeBackend

createBackend, createContext, getAcceleratorType, getDescription, getId, getType, supportsComplexNumbers, supportsFloatingPoint, supportsParallelOps

Method Details

getDevices

GPUBackend.DeviceInfo[] getDevices()

Returns information about available GPU devices.

selectDevice

void selectDevice(int deviceId)

Selects a GPU device for computation.

Parameters:

deviceId - Device index (0-based)

allocateGPUMemory

long allocateGPUMemory(long sizeBytes)

Allocates GPU memory and returns a handle.

Parameters:

sizeBytes - Size in bytes

Returns:

GPU memory handle (opaque pointer)

copyToGPU

void copyToGPU(long gpuHandle,
 DoubleBuffer hostBuffer,
 long sizeBytes)

Copies data from host to GPU.

Parameters:

gpuHandle - GPU memory handle

hostBuffer - Host buffer

sizeBytes - Number of bytes to copy

copyFromGPU

void copyFromGPU(long gpuHandle,
 DoubleBuffer hostBuffer,
 long sizeBytes)

Copies data from GPU to host.

Parameters:

gpuHandle - GPU memory handle

hostBuffer - Host buffer

sizeBytes - Number of bytes to copy

freeGPUMemory

void freeGPUMemory(long gpuHandle)

Frees GPU memory.

Parameters:

gpuHandle - GPU memory handle

synchronize

void synchronize()

Synchronizes GPU execution (waits for all kernels to complete).

matrixMultiply

void matrixMultiply(DoubleBuffer A,
 DoubleBuffer B,
 DoubleBuffer C,
 int m,
 int n,
 int k)

Performs matrix multiplication C = A * B on the GPU. Dimensions: A(m x k), B(k x n), C(m x n).