viva_tensor/config

Auto-Backend Selector - Smart tensor backend selection

Automatically chooses optimal backend (list vs strided) based on:

Operation type (sequential, random access, matrix)
Tensor size
User configuration

Benchmarked on RTX 4090 + BEAM VM

Types

OperationType

Operation type for backend selection

pub type OperationType {
  Sequential
  RandomAccess
  MatrixOp
}

Constructors

```
Sequential
```
Sequential operations - better with lists (dot, sum, reduce)
```
RandomAccess
```
Random access operations - better with strided (get, get2d, indexing)
```
MatrixOp
```
Matrix operations - strided for large, lists for small (matmul)

TensorConfig

</>

Configuration for automatic backend selection

pub type TensorConfig {
  TensorConfig(
    strided_threshold_random: Int,
    strided_threshold_matmul: Int,
    force_strided: Bool,
    force_list: Bool,
  )
}

Constructors

```
TensorConfig(
  strided_threshold_random: Int,
  strided_threshold_matmul: Int,
  force_strided: Bool,
  force_list: Bool,
)
```
Arguments

strided_threshold_random

Minimum size to use strided for random access (default: 500)

strided_threshold_matmul

Minimum size to use strided for matmul (default: 64 = 8x8)

force_strided

Force strided for all operations (override)

force_list

Force list for all operations (override)

Values

default_config

</>

pub fn default_config() -> TensorConfig

Default configuration based on benchmarks

ensure_optimal

</>

pub fn ensure_optimal(
  t: tensor.Tensor,
  op: OperationType,
  config: TensorConfig,
) -> tensor.Tensor

Ensure tensor is in optimal format for operation

for_indexing

</>

pub fn for_indexing(t: tensor.Tensor) -> tensor.Tensor

Auto-optimize tensor for indexing

for_matmul

</>

pub fn for_matmul(t: tensor.Tensor) -> tensor.Tensor

Auto-optimize tensor for matmul

for_reduction

</>

pub fn for_reduction(t: tensor.Tensor) -> tensor.Tensor

Auto-optimize tensor for reduction operations

gpu_config

</>

pub fn gpu_config() -> TensorConfig

GPU-optimized config (always strided for batched ops)

memory_config

</>

pub fn memory_config() -> TensorConfig

Memory-efficient config (prefer lists)

performance_config

</>

pub fn performance_config() -> TensorConfig

High-performance config (prefer strided for large tensors)

should_use_strided

</>

pub fn should_use_strided(
  t: tensor.Tensor,
  op: OperationType,
  config: TensorConfig,
) -> Bool

Check if should use strided backend for given operation

Constructors

Constructors

Arguments