2.0.17

CUDA — WMMA / MMA performance overhaul

• Native mma.sync kernels — adds direct PTX mma.sync tiles for:
  • FP16 (m16n8k16) — replaces old WMMA fragments, reaches full-rate on Blackwell+
  • BF16 (m16n8k16) — native tile path for BF16 GEMM
  • TF32 (m16n8k8) — replaces the old WMMA m16n16k8 path, matches cuBLASLt peak (~20.6 TFLOPS on Blackwell)
• FP16 full-rate kernel (wmma_fp16_f16.cu) — fp16 input × fp16 accumulation for full tensor-core rate
• Dead-code fold fixes — WMMA accumulator chains no longer get optimized away, fixing throughput inflation across FP16, BF16, TF32, INT8 (k32), and FP64

oneAPI — Intel Arc A770 (DG2) joint_matrix

• Enabled joint_matrix on Intel Arc A770 (DG2) via dispatchJm mechanism
• Code review cleanups: dtype matching, fallback removal

Metal — simdgroup_matrix dead-code fix

• Fixed 4× throughput inflation in simdgroup_matrix kernels (accumulator dead-code folding) for FP16 and BF16

OpenCL — Work-group clamping fix

• Per-kernel work-group size clamping to device max, resolves #183

Android

• Fixed JNI leaks

Refactoring & infrastructure

• TextLogger moved to shared common module — LoggerText in include/common/logger_text.h / src/common/logger_text.cpp. Same formatted output now drives both the desktop CLI and Android (ostream injection), removed the old src/cli/logger_cli.cpp

What's Changed

• refactor: move TextLogger from CLI to shared common module by @krrishnarraj in #182
• fix(oneapi): dynamic joint_matrix tile shapes and sub-group size for Intel XMX GPUs by @savvadesogle in #185
• feat(oneapi): benchmark all device-reported joint_matrix combinations dynamically by @savvadesogle in #186
• Merge dev → master by @krrishnarraj in #184

New Contributors

• @savvadesogle made their first contribution in #185

Full Changelog: 2.0.16...2.0.17