AMD CDNA 4

async_copy

buffer_load	AMD buffer load from global memory via a scalar base pointer and a tensor of offsets instead of a tensor of pointers.
buffer_store	AMD buffer store a tensor directly to global memory via a scalar base pointer and a tensor of offsets instead of a tensor of pointers.
buffer_atomic_add
buffer_atomic_and
buffer_atomic_max
buffer_atomic_min
buffer_atomic_or
buffer_atomic_xchg
buffer_atomic_xor
compute_efficient_padded_shared_layout	Compute an efficient padded shared layout that avoids bank conflicts.
get_mfma_scale_layout	Get the scale layout for MFMA scaled operands.
load_shared_fp4_repacked	Load M/N-packed fp4 bytes from shared memory into a K-packed MFMA dot operand layout.
mfma	Computes matrix multiplication a * b + acc using AMD native matrix core units.
mfma_scaled	AMD Scaled MFMA operation.
scaled_downcast	Scale and convert FP16, BF16, or FP32 values to a low-precision MX format, dividing by the raw E8M0 scale payload (int8 or uint8).
scaled_upcast	Upcast an fp4 or fp8 tensor and fold raw E8M0 scale payload into the CDNA4 scaled-upcast op.