Skip to content

CHANGELOG.md

Latest commit

 

History

History
91 lines (76 loc) · 4.08 KB

CHANGELOG.md

File metadata and controls

91 lines (76 loc) · 4.08 KB

Changelog

[0.5.0] - 2025-11-13

  • Add -o option for tinytc and tinytc-opt
  • (Experimental) add support for BMG G31, LNL, and PTL
  • Bugfix: group offset was applied incorrectly

[0.4.0] - 2025-10-06

Major changes

  • Support for complex numbers (c32 and c64) with optimized complex GEMM
  • Support for half precision and bfloat16 numbers
  • (Experimental) Support for accelerating low precision GEMMs with tensor cores (XMX)
  • Compiler directly generates SPIR-V binaries (instead of OpenCL-C); JIT compilation speed-up of about 2x
  • Removed dependency on clir and ocloc libraries
  • Added preprocessor

Tiny Tensor Language

  • Clarified execution model
  • Added align attribute
  • Added unroll attribute
  • Added boolean, f16 (half precision), bf16 (bfloat16), c32 (single complex), and c64 (double complex) type
  • Added address space to memref type
  • Added cooperative matrix type
  • Added cooperative matrix instructions (prefix "cooperative_matrix_"): apply, atomic_load, atomic_store, atomic_update, construct, extract, insert, load, mul_add, reduce, scale, store
  • Added new instructions: associated, barrier, constant, cumsum, foreach_tile, parallel, subgroup operations
  • Extended foreach for iteration spaces with dim > 1
  • For-loops can yield values
  • Removed immediate operands in favour of constant instruction
  • Instructions annotate the return type instead of operand types
  • Change 1d group id to 3d group id; introduce 2d subgroup id

Core API

  • Introduced compiler_context class for compilations flags and control of error reporting
  • Removed functions to create kernel bundle from source text
  • Removed source_context in favour of compiler_context
  • Added functions to compile tinytc prog to SPIR-V
  • Added API to run function passes
  • Added float <-> half and float <-> bfloat16 conversion functions
  • Added array_view class in C++-API

Builder API

  • Changed ownership model for data type, inst, region, and func
  • Normalized function names in builder API
  • Instruction creation function always include return type

Compiler

  • Overhauled infrastructure for writing advanced compiler passes
  • Introduced data flow analysis to properly insert barriers automatically when for-loops are present
  • Added constant folding, constant propagation, and dead code elimination passes
  • Added memref alignment analysis
  • Added analysis to check applicability of tensor core acceleration
  • Added SPIR-V support; Tiny Tensor Language to SPIR-V conversion and SPIR-V binary generation
  • Introduced GEMM to cooperative matrix conversion (instead of direct GEMM to OpenCL-C conversion)
  • Added fast cooperative matrix mul add implementation for complex numbers
  • Added attributes
  • Added "mochi" compiler-compiler to generate boilerplate for instructions, types, and enums

[0.3.1] - 2024-05-22

  • Bugfix: Add alias analysis for stack; needed to correctly insert barriers
  • Bugfix: Disable block writes as alignment analysis is missing

[0.3.0] - 2024-05-16

  • Added offset parameter to group type
  • Deploy FindOpenCL and FindLevelZero script for finding tinytc's dependencies via CMake
  • Replaced concept by SFINAE such that tinytc.hpp can be used in C++17 projects
  • Improved heuristic for subgroup size selection
  • Refactored kernel compilation workflow such that libocloc is only used by libtinytc_ze and libtinytc_cl uses the built-in compiler; libtinytc and libtinytc_cl can be built without libocloc
  • Added functions to query device support level
  • Added generic core info class for non-Intel devices
  • cl_mem object is passed by value instead of by pointer for recipes
  • Removed tinytc_mem_t; memory object type is part of set_args signature
  • The index type was changed to i64 (from u32); attention: the calling convention is updated
  • Add function to create tall and skinny recipe with additional specialization constants
  • Fix uninitialized value bug in GEMM generator
  • Use relaxed ordering for atomics
  • Remove unsigned integer types
  • Prefix arithmetic instructions with "arith."
  • Add binary arithmetic (and, or, xor, not)

[0.2.0] - 2024-04-26

  • Refactored API and added C-API

[0.1.0] - 2024-04-05

Initial release