Documentation
¶
Overview ¶
Package provider defines the ExecutionProvider interface for routing ONNX operator execution to different compute backends (CPU, GPU). Providers are queried in priority order; the first provider that supports a given op type handles execution. CPUProvider serves as the universal fallback and should always be last in the chain.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type CPUProvider ¶
type CPUProvider struct{}
CPUProvider executes all operators registered in the ops registry on the CPU. It is stateless and serves as the universal fallback provider.
func (*CPUProvider) Close ¶
func (p *CPUProvider) Close() error
Close is a no-op for the CPU provider since it holds no resources.
func (*CPUProvider) Execute ¶
func (p *CPUProvider) Execute(opType string, inputs []*tensor.Tensor, _, _ []string, attributes map[string]*onnx.Attribute, alloc tensor.Allocator) ([]*tensor.Tensor, error)
Execute retrieves the operator from the ops registry and delegates execution to it. The inputNames and outputNames parameters are accepted for interface compliance but ignored by the CPU provider.
func (*CPUProvider) Supports ¶
func (p *CPUProvider) Supports(opType string) bool
Supports returns true if the ops registry contains an operator for the given op type.
type ExecutionProvider ¶
type ExecutionProvider interface {
// Supports reports whether this provider handles the given ONNX op type.
Supports(opType string) bool
// Execute runs the operator identified by opType on the given inputs.
// inputNames and outputNames carry the ONNX graph tensor names so that
// providers (e.g., GPUProvider) can register output buffers by name.
Execute(opType string, inputs []*tensor.Tensor, inputNames, outputNames []string, attributes map[string]*onnx.Attribute, alloc tensor.Allocator) ([]*tensor.Tensor, error)
// Close releases any resources held by the provider (e.g., GPU device, buffers).
Close() error
}
ExecutionProvider abstracts operator execution across compute backends. The engine queries providers in priority order and dispatches to the first provider that reports support for the requested op type.
Source Files