Refactor gfx950 A16W16 GEMM to use the universal kernel

.claude/skills/kernel-code-cleanup/SKILL.md

@@ -427,7 +427,7 @@ def _run_compiled(exe, *args):             # in-tree
 ```
 
 - Pass flat scalars/pointers (`data_ptr()`, `stride(i)`, sizes, `stream`) — it
-  bypasses DLPack. See `pa_decode_fp8.py`, `hgemm_splitk.py`.
+  bypasses DLPack. See `pa_decode_fp8.py`.
 - Reuse the shim; don't add a second copy.
 - Worth it for small kernels in tight loops, not cold one-shot launches. Arg
   order/types must match the compiled signature — verify.

README.md

@@ -365,7 +365,7 @@ See `examples/` for more examples including tiled copy (`02-tiledCopy.py`), tile
 | Category | Test File | Description |
 |----------|-----------|-------------|
 | **Preshuffle GEMM** | `test_preshuffle_gemm.py` | FP8, INT8, INT4, BF16, FP4 |
-| **HGEMM Split-K** | `test_hgemm_splitk.py` | FP16 GEMM split-K |
+| **GFX950 A16W16 GEMM** | `test_gemm_a16w16_gfx950.py` | FP16/BF16, NN/NT/TN/TT layouts, bias, FP32 output, slice-K/split-K |
 | **MoE GEMM** | `test_moe_gemm.py` | MoE 2-stage (gate/up + reduce) |
 | **MoE Reduce** | `test_moe_reduce.py` | MoE reduce kernel |
 | **PagedAttention** | `test_pa.py` | Paged attention decode (FP8) — *WIP perf tuning* |

docs/api/kernels.rst

@@ -12,7 +12,6 @@ GEMM Kernels
 - ``kernels.gemm.preshuffle_gemm`` -- MFMA-based GEMM with LDS pipeline and pre-shuffled weights (FP8, INT8, FP16, BF16)
 - ``kernels.gemm.mxfp4_preshuffle`` -- MXFP4 / FP4 (and f8f4) preshuffle GEMM
 - ``kernels.gemm.fp4_gemm_4wave`` -- 4-wave FP4 GEMM (gfx950)
-- ``kernels.gemm.hgemm_splitk`` -- FP16 split-K GEMM
 
 MoE (Mixture-of-Experts) Kernels
 ----------------------------------

docs/architecture_guide.md

@@ -82,7 +82,6 @@ FlyDSL/
 │   │   ├── fp4_gemm_4wave.py         # FP4 4-wave GEMM
 │   │   ├── fp8_gemm_4wave.py         # FP8 4-wave GEMM
 │   │   ├── fp8_gemm_8wave.py         # FP8 8-wave GEMM
-│   │   ├── hgemm_splitk.py           # FP16 GEMM split-K
 │   │   ├── rdna_f16_gemm.py          # RDNA FP16 GEMM
 │   │   ├── rdna_fp8_preshuffle_gemm.py # RDNA FP8 GEMM
 │   │   ├── gemm_common_gfx1250.py    # GFX1250 GEMM common

docs/prebuilt_kernels_guide.md

@@ -364,7 +364,6 @@ What operation do you need?
 | File | Description |
 |---|---|
 | `kernels/gemm/preshuffle_gemm.py` | GEMM (preshuffle layout) |
-| `kernels/gemm/hgemm_splitk.py` | FP16 GEMM split-K |
 | `kernels/moe/moe_gemm_2stage.py` | MoE GEMM 2-stage (gate/up + reduce) |
 | `kernels/moe/mxfp_moe/` | Fused a4w4/a8w4 MoE 2-stage GEMM (device fp4 re-quant) |
 | `kernels/attention/pa_decode_fp8.py` | Paged attention decode (FP8) |
@@ -392,7 +391,6 @@ What operation do you need?
 | File | Tests |
 |---|---|
 | `tests/kernels/test_preshuffle_gemm.py` | GEMM fp8/int8/fp16/bf16 |
-| `tests/kernels/test_hgemm_splitk.py` | FP16 GEMM split-K |
 | `tests/kernels/test_moe_gemm.py` | MoE GEMM |
 | `tests/kernels/test_moe_reduce.py` | MoE reduce kernel |
 | `tests/kernels/test_pa.py` | Paged attention decode |

kernels/common/mem_ops.py

@@ -56,8 +56,7 @@ def get_llvm_ptr(ptr, offset, dtype_bytes, ptr_type=None):
     """Build a global (address-space 1) ``!llvm.ptr`` at ``ptr + offset*dtype_bytes``.
 
     Shared home for the LLVM-ptr arithmetic used by atomic/global accesses
-    (previously duplicated in hgemm_splitk.py, small_m_hgemm.py, splitk_hgemm.py
-    and rmsnorm_kernel.py).
+    (previously duplicated in several kernels).
     """
     if ptr_type is None:
         ptr_type = ir.Type.parse("!llvm.ptr<1>")