XeTLA v0.3.7

v0.3.7

• Enabled multi-SKUs support.
• Refined the project’s coding conventions.
• Cleaned up building warnings.
• Upgraded the toolchain to oneAPI-2024.0.0.
• Updated documents, tests, and examples.

XeTLA v0.3.6

v0.3.6

• Added GEMM new feature for any shapes support (odd shapes).
• Provided default configurations for GEMM API (users could get good performance by default configurations, only advanced users need to tune optimization options).
• Supported converting register layout between tiled and linear.
• Provided flexible large shape's APIs for other policy (e.g. splitk, improved mat_A & mat_B cache hit ratio).
• Refined mem_desc_t and payload_t to expose alignment parameter.
• Enabled epilogue to support D = alpha * A * B + beta * C.
• Replaced xetla_exec_item with sycl::nd_item.
• Refined some examples to invoke kernel level APIs, added fence and barrier to MLP example.
• Fixed some known issues, enhanced tests, and updated documents.

XeTLA v0.3.5

v0.3.5

• Enhanced limitation checking.
• Refined GEMM APIs’ name.
• Supported GEMM APIs load B from SLM.
• Supported GEMM of any (odd) shapes.
• Supported Streaming-K.
• Enhanced L3 K-slicing support.
• Improved GEMM's performance for large-N (M, N, K) shapes.
• Fixed tile load/store bugs.
• Enhanced examples, tests, and updated documents.

XeTLA v0.3.4

v0.3.4

• Enabled limitation checking.
• Provided "DEVICE_PRINTF" and "DEVICE_ASSERT" for kernels debugging.
• Refined fpu based GEMM.
• Refined tile reduce APIs, deprecated API "tile_row_reduce".
• Supported new data type int4 (experimental feature).
• Fixed tile load/store bugs.
• Enhanced examples, tests, and updated documents.

XeTLA v0.3.3

v0.3.3

• This is v0.3 golden release.
• Enabled debug build support.
• Updated documents, added some diagrams and details.
• Fixed some customer reported issues.
• Improved the project's quality.

XeTLA v0.3.2

v0.3.2

• This is v0.3 beta release.
• Added some kernel-level APIs' parameters check functions, users need to explicit call them before launch the kernel; will return fail and print error messages when detect unsupported scenarios, continue launching the kernel for unspported scenarios may lead to unpredictable result.
• Removed reduce_sum + tile_op epilogue policy.
• Added some unit test cases.
• Refined some examples code.
• Updated documents, added some diagrams and details.
• Fixed some customer reported issues.
• Improved the project's quality.

XeTLA v0.3.1

v0.3.1

• This is v0.3 alpha release.
• Initial open source release.
• Improved sub-group level fundamental features: core, tile, utils.
• Enhanced basic BRGEMM micro-kernels for different data types as well as different epilogue flavours.
• Unified the examples: basic BRGEMM, GEMM; fusion BRGEMM, GEMM; batched GEMM; MLP, GRU, MHA.
• Added some unit test cases.