New in CUTLASS SYCL 0.5

Major Architecture Changes

• Xe Rearchitecture (#477): Complete redesign of Xe CuTe atoms with new architecture
  • New MMA atoms for improved performance
  • Enhanced 2D copy atoms (loads, stores, prefetch with VNNI/transpose support)
  • New 2D copy helpers (low-level make_block_2d_copy and high-level make_block_2d_copy_{A,B,C})
  • Generic and optimized reorder atoms for {int4, uint4, int8, uint8, e2m1, e4m3, e5m2} -> {half, bfloat16}
  • Requires IGC version v2.18.5 or later

New Features

• G++ Host Compiler Support (#490): Support for G++ 13 as host compiler

• Migrated syclcompat to this repository as cutlasscompat for better compatibility

  • Fixed compilation issues when using G++ instead of clang++
  • Added new CI workflow for testing G++ host compiler builds
  • Enhanced build system to support -DDPCPP_HOST_COMPILER=g++ option

• Grouped GEMM for Mixed Dtype (#457): Extended grouped GEMM support to mixed precision operations

  • Added support for BF16 + S8 mixed dtype grouped GEMM
  • Added support for FP16 + U4 mixed dtype grouped GEMM
  • New examples: 10_bmg_grouped_gemm_bf16_f16_s8.cpp and 10_bmg_grouped_gemm_f16_u4.cpp

  See the CHANGELOG-SYCL for details of all past releases and updates.

What's Changed

Cutlass 3.9.2 SYCL backend Version 0.3 (2025-06-30)

• Add support for GEMM FP8 (E5M2 and E4M3)
• Add example for GEMM FP8 with support for channel-wise and group-wise quantization
• Add support for Grouped GEMM FP8
• Improve performance for FP8 to FP16 conversion
• Add support for epilogue data conversion
• Add support for FP16 GEMM with FP16 accumulator
• Add support for BF16 GEMM with BF16 accumulator
• Add support for mixed dtype GEMM with support for tensor-wise, channel-wise and group-wise quantization
• Add example of mixed dtype BF16 + INT8 using channel-wise and group-wise quantization
• Add example of mixed dtype FP16 + INT8 using tensor-wise quantization
• Add example of mixed dtype FP16 + INT4 using channel-wise and group-wise quantization
• Add support for zero-point quantization in INT4 and INT8 data types
• Add support for Flash Attention prefill FP8 with and without KV cache
• Add support for Flash Attention decode FP8 with and without KV cache

Full Changelog: v3.9-0.2...v3.9-0.3

Cutlass 3.9.2 SYCL backend Version 0.2 (2025-05-30)
Based on CUTLASS 3.9.2 - May 2025 release

Platforms

• Support for Intel GPU Data Center Max (1100 and 1550)
• Support for Intel Arc B580 ("Battlemage")

Features

• GEMM/StreamK/SplitK with support for FP16 data type
• Flash attention prefill with Paged KV cache with support for FP16 data type
• Performance improvements for flash attention prefill and decode

Full Changelog: v3.9-0.1...v3.9-0.2

Based on CUTLASS 3.9.0 March 2025 release

Platforms

• Support for Intel GPU Data Center Max (1100 and 1550)
• Support for Intel Arc B580 ("Battlemage")

Features

• GEMM/StreamK/SplitK with support for bfloat16 data type

• Flash attention prefill and decode with KV cache with support for bfloat16 data type

• Support for epilogue operations:

  • Element-wise, row-wise and column-wise bias
  • ReLU, SiLU, GELU activation fns
  • Softmax

• Mixed precision GEMM (bfloat16/int8, half/int4) with dequantization support

• Dual GEMM & Grouped GEMM

Full Changelog: https://github.com/codeplaysoftware/cutlass-sycl/commits/v3.9-0.1