‘tt’ Dialect
The Triton IR in MLIR
Triton Dialect.
Dependent Dialects:
Arith:
addf, addi, andi, cmpf, cmpi, divf, fptosi, …
Math:
exp, sin, cos, log, …
StructuredControlFlow:
for, if, while, yield, condition
ControlFlow:
br, cond_br
[TOC]
Type constraints
ptr
Types
PackedFloat4E2M1Type
Two E2M1 floating-point values packed in one byte
Syntax: !tt.fp4e2m1x2
The low nibble holds the first value and the high nibble the second. Tensor dimensions and pointer offsets count packed pairs.
PointerType
Pointer type (::mlir::triton::PointerType) in Triton IR type system
Pointer type in Triton IR type system. Pointer types may only point to scalar element types.
Parameters:
Parameter |
C++ type |
Description |
|---|---|---|
pointeeType |
|
|
addressSpace |
|
TensorDescType
Tensor descriptor type (::mlir::triton::TensorDescType) in Triton IR type system
A portable abstraction for TMA descriptors. This is the base tensor descriptor type for tiled tensor memory access.
Shape and elementType describe the block dimensions and data type. The optional sharedLayout attribute carries the shared memory encoding (e.g. swizzle pattern) that is assigned during lowering.
For specialized access patterns like im2col, see TensorDescIm2ColType in the TritonNvidiaGPU dialect.
Parameters:
Parameter |
C++ type |
Description |
|---|---|---|
shape |
|
|
elementType |
|
|
sharedLayout |
|