Crate bhc_tensor_ir 

BHC Tensor IR

This crate defines the Tensor Intermediate Representation for the Basel Haskell Compiler. Tensor IR is specifically designed for numeric optimization and provides the foundation for guaranteed fusion and vectorization.

Overview

Tensor IR is the key to BHC’s numeric performance. It captures:

• Shape and stride information: For layout-aware optimization
• Element types (dtypes): For unboxed numeric computation
• Operation structure: For fusion analysis
• Aliasing information: For safe in-place updates

H26-SPEC Section 7 Compliance

Per the H26 specification, every tensor operation in Tensor IR must track:

Property	Description
dtype	Element type (Float32, Float64, Int32, etc.)
shape	Dimension sizes
strides	Byte strides per dimension
layout	Memory layout (contiguous, strided, tiled)
alias	Aliasing/ownership information

IR Pipeline Position

Source Code
    |
    v
[Parse/AST]
    |
    v
[HIR]
    |
    v
[Core IR]   <- General purpose optimizations
    |
    | (Numeric Profile only)
    v
[Tensor IR] <- This crate: shape-aware, fusion-ready
    |
    v
[Loop IR]   <- Explicit iteration

Guaranteed Fusion Patterns

Per H26-SPEC Section 8, these patterns MUST fuse:

1. map f (map g x) -> single traversal
2. zipWith f (map g a) (map h b) -> single traversal
3. sum (map f x) -> single traversal
4. foldl' op z (map f x) -> single traversal

Main Types

• TensorOp: Tensor operations
• TensorMeta: Metadata (shape, stride, dtype)
• Kernel: A fused computation unit
• Shape: Tensor dimensions
• DType: Element types

See Also

• bhc-core: Core IR that lowers to Tensor IR
• bhc-loop-ir: Loop IR for explicit iteration
• H26-SPEC Section 7: Tensor Model
• H26-SPEC Section 8: Fusion Laws

Modules

fusion

Tensor fusion pass.

lower

Core IR to Tensor IR lowering pass.

Structs

AllocInfo

Allocation information for a kernel.

Axis

An axis specification.

BufferId

A unique identifier for buffers (memory allocations).

ConvSpec

Convolution specification.

FoldFn

A fold function.

FusionInfo

Fusion information for debugging and reporting.

Kernel

A fused computation kernel.

KernelId

A unique identifier for kernels.

LoopInfo

Information about a single loop.

LoopNest

A simple loop nest representation.

MapFn

A map function (element-wise transformation).

Permutation

A permutation for transpose.

Shape

Tensor shape (list of dimensions).

SliceRange

A range within a slice.

SliceSpec

A slice specification.

Strides

Memory strides for each dimension.

TensorId

A unique identifier for tensor operations.

TensorMeta

Tensor metadata per H26-SPEC Section 7.3.

TensorRef

A reference to a tensor value.

TileInfo

Tiling information for cache-friendly layouts.

ZipFn

A zip function (combining two elements).

Enums

AllocRegion

Memory regions for allocation.

BinaryOp

Binary operations.

ConstantOp

A constant tensor operation.

DType

Tensor element types (data types).

DeviceTarget

Target device for GPU memory allocation.

Dim

A dimension size (may be static or dynamic).

FusionBlockReason

Why fusion was blocked.

FusionDecision

A fusion decision made by the compiler.

KernelBody

The body of a kernel.

Layout

Memory layout of a tensor.

MaterializeReason

Why a tensor was materialized.

ReduceOp

Reduction operations.

ScalarValue

A scalar value.

TensorIrError

Errors in Tensor IR operations.

TensorOp

Tensor operations in the IR.

UnaryOp

Unary operations.