Expand description
Cost model and latency prediction for kernel operations.
Provides per-op FLOP/memory estimation, roofline analysis, and hardware profile–based latency prediction.
Structs§
- Hardware
Profile - Hardware profile for roofline analysis.
- OpCost
- Per-operation cost in terms of compute and memory traffic.
Enums§
- Bottleneck
- Whether an operation is bottlenecked by compute or memory bandwidth.
Functions§
- default_
profiles - Return default hardware profiles for known backends.
- estimate_
activation_ cost - Estimate cost for an activation operation with known element count.
- estimate_
conv2d_ cost - Estimate cost with explicit numeric dimensions for a Conv2D pattern.
- estimate_
elementwise_ cost - Estimate cost for an element-wise operation with known element count.
- estimate_
kernel_ cost - Estimate the compute and memory cost of a classified kernel pattern.
- estimate_
matmul_ cost - Estimate cost with explicit numeric dimensions for a MatMul pattern.