hipDNN backend descriptor type C API

2026-03-31

5 min read time

Applies to Linux and Windows

Descriptor type identifiers for hipDNN backend API.

This file defines the types of backend descriptors that can be created using hipdnnBackendCreateDescriptor(). Each descriptor type holds specific configuration for different aspects of graph execution.

Enums

enum hipdnnBackendDescriptorType_t

Types of backend descriptors.

Backend descriptors are opaque handles that store configuration data. Use hipdnnBackendCreateDescriptor() with one of these types to create a new descriptor, then configure it with hipdnnBackendSetAttribute().

See also

hipdnnBackendCreateDescriptor()

See also

hipdnnBackendSetAttribute()

See also

hipdnnBackendFinalize()

Values:

enumerator HIPDNN_INVALID_TYPE

Invalid descriptor type - used for error detection.

enumerator HIPDNN_BACKEND_ENGINE_DESCRIPTOR

Engine descriptor.

Represents a specific execution engine that can run an operation graph. Engines are discovered through heuristics and provide different performance/precision trade-offs.

enumerator HIPDNN_BACKEND_ENGINECFG_DESCRIPTOR

Engine configuration descriptor.

Holds the configuration for an engine, including knob settings and workspace requirements. Created from an engine descriptor.

enumerator HIPDNN_BACKEND_ENGINEHEUR_DESCRIPTOR

Engine heuristic descriptor.

Used to query available engines for an operation graph. Set the operation graph and heuristic mode, then query results to get ranked engine configurations.

enumerator HIPDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR

Execution plan descriptor.

The final compiled plan ready for execution. Created from an engine configuration and used with hipdnnBackendExecute().

enumerator HIPDNN_BACKEND_INTERMEDIATE_INFO_DESCRIPTOR

Intermediate tensor info descriptor.

Contains information about intermediate (virtual) tensors in a fused operation graph.

enumerator HIPDNN_BACKEND_KNOB_CHOICE_DESCRIPTOR

Knob choice descriptor.

Represents a specific value for an engine tuning knob. Used to configure engine behavior.

enumerator HIPDNN_BACKEND_KNOB_INFO_DESCRIPTOR

Knob info descriptor.

Contains metadata about an engine knob, including valid value ranges and default values.

enumerator HIPDNN_BACKEND_LAYOUT_INFO_DESCRIPTOR

Layout info descriptor.

Contains information about tensor layout requirements and supported formats for an engine.

enumerator HIPDNN_BACKEND_OPERATION_GEN_STATS_DESCRIPTOR

Generate statistics operation descriptor.

Represents an operation that generates statistics (e.g., for batch normalization running mean/variance).

enumerator HIPDNN_BACKEND_OPERATIONGRAPH_DESCRIPTOR

Operation graph descriptor.

The main descriptor representing a computational graph. Contains tensors and operations to be executed. Typically created from serialized FlatBuffer data.

enumerator HIPDNN_BACKEND_VARIANT_PACK_DESCRIPTOR

Variant pack descriptor.

Maps tensor UIDs to device memory pointers and workspace for execution. Passed to hipdnnBackendExecute().

enumerator HIPDNN_BACKEND_KERNEL_CACHE_DESCRIPTOR

Kernel cache descriptor.

Manages caching of compiled GPU kernels for faster subsequent execution plan creation.

enumerator HIPDNN_BACKEND_OPERATION_PAGED_CACHE_LOAD_DESCRIPTOR

Paged cache load operation descriptor.

Represents a paged memory cache loading operation.

enumerator HIPDNN_BACKEND_TENSOR_DESCRIPTOR

Tensor descriptor.

Represents a tensor with dimensions, strides, data type, and optional pass-by-value data.

enumerator HIPDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR

Convolution forward operation descriptor.

Represents a forward convolution operation with input (X), weight (W), and output (Y) tensors plus convolution parameters.

enumerator HIPDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR

Convolution backward filter operation descriptor.

Represents a backward filter convolution operation with input (X), output gradient (DY), and weight gradient (DW) tensors plus convolution parameters.

enumerator HIPDNN_BACKEND_OPERATION_BATCHNORM_INFERENCE_DESCRIPTOR_EXT

Batch normalization inference operation descriptor.

Represents a batch normalization inference operation with input (X), mean, inverse variance, scale, bias, and output (Y) tensors.

enumerator HIPDNN_BACKEND_OPERATION_POINTWISE_DESCRIPTOR

Pointwise operation descriptor.

Represents a pointwise (element-wise) operation with 1-3 input tensors and activation parameters. Supports unary, binary, and ternary operations.

enumerator HIPDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DESCRIPTOR

Convolution backward data (Dgrad) operation descriptor.

Represents a backward data convolution operation with output gradient (DY), weight (W), and input gradient (DX) tensors plus convolution parameters.

enumerator HIPDNN_BACKEND_OPERATION_BATCHNORM_BACKWARD_DESCRIPTOR_EXT

Batchnorm backward operation descriptor (extension)

Represents a backward batch normalization operation computing gradients with respect to input (DX), scale (DScale), and bias (DBias).

enumerator HIPDNN_BACKEND_OPERATION_BATCHNORM_INFERENCE_VARIANCE_DESCRIPTOR_EXT

Batchnorm inference variance ext operation descriptor.

Represents a batch normalization inference operation (variance ext variant) with input (X), mean, variance, scale, bias, epsilon, and output (Y) tensors.

enumerator HIPDNN_BACKEND_OPERATION_MATMUL_DESCRIPTOR_EXT

Matrix multiplication operation descriptor.

Represents a matrix multiplication operation with input (A), input (B), and output (C) tensors plus a compute data type.

enumerator HIPDNN_BACKEND_OPERATION_RMSNORM_DESCRIPTOR_EXT

RMSNorm operation descriptor.

Represents an RMS normalization forward operation with input (X), scale, epsilon, bias (optional), output (Y), and inverse RMS (optional, training only) tensors.

enumerator HIPDNN_BACKEND_OPERATION_SDPA_FPROP_DESCRIPTOR_EXT

SDPA forward propagation operation descriptor (extension)

Represents a scaled dot-product attention forward operation with query (Q), key (K), value (V), and output (O) tensors plus attention parameters.

enumerator HIPDNN_BACKEND_OPERATION_LAYERNORM_DESCRIPTOR_EXT

Layer normalization operation descriptor.

Represents a layer normalization operation with input (X), scale, bias, epsilon tensors, output (Y), and optional mean and inverse variance outputs plus a compute data type.

enumerator HIPDNN_BACKEND_OPERATION_BLOCK_SCALE_QUANTIZE_DESCRIPTOR_EXT

Block scale quantize operation descriptor.

Represents a block scale quantize operation with input (X), output (Y), and scale tensors plus block_size, axis, and transpose parameters.

enumerator HIPDNN_BACKEND_OPERATION_BATCHNORM_DESCRIPTOR_EXT

Batchnorm training forward operation descriptor (extension)

Represents a batch normalization training forward operation with input (X), scale, bias, epsilon, output (Y), optional mean and inverse variance outputs, optional running statistics, and peer stats.

enumerator HIPDNN_BACKEND_OPERATION_BLOCK_SCALE_DEQUANTIZE_DESCRIPTOR_EXT

Block scale dequantize operation descriptor.

Represents a block scale dequantize operation with input (X), scale, and output (Y) tensors plus block size parameters.

enumerator HIPDNN_BACKEND_OPERATION_CUSTOM_OP_DESCRIPTOR_EXT

Custom operation descriptor.

Represents an opaque custom (plugin-provided) operation with variable-length input/output tensor arrays, a plugin identifier string, and an opaque byte payload interpreted by the plugin.

enumerator HIPDNN_BACKEND_OPERATION_SDPA_BPROP_DESCRIPTOR_EXT

SDPA backward propagation operation descriptor (extension)

Represents a scaled dot-product attention backward operation with query (Q), key (K), value (V), output (O), gradient output (dO), and stats tensors as inputs, producing gradients dQ, dK, dV.