Intel Intel Xpu Backend For Triton Intel Xpu Backend For Triton

Block Pointer → Tensor Descriptor (Intel XPU)

Translate kernels from the deprecated block-pointer API to the device-side tensor-descriptor API. tl.make_block_ptr is deprecated; its docstring points users to tl.make_tensor_descriptor. On Intel XPU, a well-formed descriptor lowers to the same hardware 2D-block-I/O path the block pointer used, so a careful translation is at worst neutral and removes the deprecation.

Prime directive: translate toward an efficiently-lowerable descriptor

Do not emit the first descriptor that compiles. There are usually several faithful ways to write a descriptor, and only some keep the fast 2D-block-I/O path on XPU. Always aim for a descriptor that is 2D, last-stride-1, locally-created, dot-feeding, PAD_ZERO. When that is achievable (the common GEMM/attention cases), produce it. When a particular load genuinely cannot be put on the fast path, still translate it (the user wants the modern API) but explicitly flag the caveat in the "Changes made" section so nobody is misled into thinking every load got block I/O.

See references/performance-notes.md for why (the verified backend gates) and references/api-reference.md for exact signatures and constraints. See references/examples.md for verbatim before→after pairs you can pattern-match against.

Workflow

1. Read the kernel (inline or from the given file path). Identify every block-pointer construct: tl.make_block_ptr, tl.advance, and the tl.load/tl.store calls that consume block pointers. Also note any interleaved tl.load(ptrs, mask=...) tensor-of-pointer loads (Rule 9) and the rank of each block pointer (Rule 11).
2. Apply the transformation rules below, honoring the prime directive.
3. Emit the output in the format specified at the bottom.

This is XPU-only. Never add triton.set_allocator or any host-launch change — device-side descriptors need no global-memory workspace on XPU (that hook is NVIDIA-TMA-specific). The migration is a pure kernel-body rewrite unless Rule 12 applies.

Transformation rules

1. tl.make_block_ptr(base, shape, strides, offsets, block_shape, order) → tl.make_tensor_descriptor(base, shape, strides, block_shape). Drop offsets and order from the constructor. The initial offsets values become the starting load/store offsets (Rules 2 & 7). Hoist the descriptor to the same scope as the old make_block_ptr — it is loop-invariant, create it once before the loop.

2. tl.advance(ptr, delta) → integer offset increment. Delete the tl.advance. Introduce a running integer for each advancing dimension, initialized to that dim's initial offset (off_k = 0 typically), and add the delta at the loop tail (off_k += BLOCK_K). Pass the running offset into desc.load([...]) / desc.store([...]).

3. tl.load(ptr, boundary_check=...) → desc.load([d0, d1]). Drop boundary_check — the descriptor's shape already masks out-of-bounds elements to the pad value, which is exactly what boundary_check did. Padding caveat: padding_option is set at descriptor-creation time, not at load time. If the original load used padding_option="nan", carry it onto the constructor: tl.make_tensor_descriptor(..., padding_option="nan"). "zero"/"" is the default and is dropped.

4. tl.store(ptr, val, boundary_check=...) → desc.store([d0, d1], val). Drop boundary_check; out-of-bounds writes are ignored automatically. Store has no padding_option.

5. Block-pointer loads/stores never carry mask=/other= (the API forbids it). Their only masking is boundary_check, handled by Rules 3–4. A mask= on the input only appears on legacy tensor-of-pointer loads — see Rule 9. (The skill itself only emits a mask= tensor-of-pointer load in the single sanctioned case of a non-unit-stride rank-1 block pointer that has no legal descriptor — see Rule 11.)

6. Transpose (order=(0, 1) in the block pointer). Do NOT emit a descriptor whose last stride ≠ 1 — that drops off the fast path. Instead describe the tensor in its native row-major layout (so the last stride is 1) and apply .T to the loaded block. Concretely, if B is stored (N, K) row-major but the dot wants (K, N): b_desc = tl.make_tensor_descriptor(base=b_ptr, shape=(N, K), strides=(stride_bn, stride_bk), block_shape=(BLOCK_N, BLOCK_K)) then b = b_desc.load([pid_n*BLOCK_N, off_k]).T.

7. Static offsets (a dim that never advances, e.g. pid_m * BLOCK_M) → pass the expression directly in .load() / .store(); no running variable needed.

8. Batched / 3D → pre-add the batch offset to base (base = a_ptr + bid.to(tl.int64) * stride_az) and keep the descriptor 2D over the M×K (or K×N) slice. This is the rank-3 special case of Rule 11.

9. Masked tensor-of-pointer loads (tl.load(ptrs, mask=..., other=...), not block pointers — may be interleaved in the same kernel). Classify the mask:

   • Boundary mask — a range check vs a dimension limit, e.g. offs_k[None,:] < K - k*BLOCK_K or offs_m[:,None] < M. This is a bounds check: fold it into the descriptor shape (e.g. shape=(M, K)), drop the mask, and let the descriptor zero-pad.
   • Data-dependent / non-boundary mask — causal masks, value-dependent predicates. Cannot be a rectangular shape. Load the full block with desc.load([...]), then re-apply in registers: v = tl.where(mask, v, other).
   • Mixed — split: boundary part → shape, residual predicate → tl.where.

10. Type-annotation hygiene → if the last stride is annotated tl.int64/tl.int32, flag it (best practice: never annotate the last stride; prefer tl.constexpr or no annotation). Do not silently change a signature.

11. Rank normalization → always aim for 2D (XPU optimizes only 2D; block pointers may be any rank):

    • Rank 2 → map 1:1.
    • Rank ≥ 3 → fold the leading dims into base (one term per outer dim, base + z*stride_z + h*stride_h) and build a 2D descriptor over the inner two dims; .T/reshape the result if the consumer needs it. Each folded outer index must be a per-program scalar (e.g. program_id, a scalar load). If an outer index is itself a vector/range it cannot be folded — flag it and keep the original.
    • Rank 1, unit stride (strides=(1,)) → translate to a 1D descriptor: tl.make_tensor_descriptor(base, shape=(N,), strides=(1,), block_shape=(BLOCK,)), desc.load([off]) / desc.store([off], val). Note that XPU does not give 1D descriptors 2D block I/O today (the 1D fast path is pending), so flag that this load lowers via the pointer path — the form is correct and forward-compatible.
    • Rank 1, non-unit stride (a strided 1D view, e.g. SGLang FLA's p_beta = tl.make_block_ptr(beta + bos*H + i_h, (T,), (H,), (i_t*BT,), (BT,), (0,)) with stride H != 1) → no legal descriptor exists: a descriptor's last (here only) stride must be 1, so strides=(H,) fails to compile (Tensor descriptor last dim must be 1 but got H). This is the one case where a boundary_check cannot be modernized into a descriptor at all — so do not emit a descriptor. Reproduce the access as an explicit tensor-of-pointer load/store whose mask re-expresses the original boundary_check (it stays on the pointer path — no 2D block I/O — which is the only correct option here, unlike Rule 9 which removes such masks by folding them into a descriptor shape):

      o = i_t * BT + tl.arange(0, BT)
      b_beta = tl.load(beta + bos*H + i_h + o*H, mask=o < T, other=0.0)
      # strided rank-1 store: tl.store(g + bos*H + i_h + o*H, b_val, mask=o < T)
      

      If the offset advances in a loop, fold the running offset into o exactly as Rule 2 does for descriptors. Report this under "Changes made" (it emits no descriptor), and flag it off the 2D-block fast path.

12. Block pointer received as a helper argument. A block pointer can never arrive from host launch (there is no host-side block-pointer object); it can only be a @triton.jit → @triton.jit parameter — and arriving that way already defeats the fast path (a descriptor passed across tt.call loses its identity). If the kernel body uses a block pointer that was not created locally, the translation is not body-only:

    • Change the helper's signature to take raw base, shape, strides, offsets (and block_shape/order if not already tl.constexpr there).
    • Build the descriptor inside the helper with tl.make_tensor_descriptor.
    • Update every call site to pass the raw ingredients. Report this as a caller-affecting change and list the touched call sites. No-op for the common create-and-use-in-the-same-kernel case.

Output format

1. The fully translated @triton.jit kernel — every block-pointer construct replaced. Preserve surrounding code, comments, and signatures (except Rule 12, which legitimately changes a helper signature).
2. "Changes made" — a short list: counts and each transformation applied, e.g. "Replaced 3 make_block_ptr with make_tensor_descriptor; removed 2 tl.advance, introduced off_k; B loaded transposed via .T; flagged stride_ak: tl.int64". State plainly that no host-launch changes are required on XPU.
3. Performance note (per descriptor) — for each descriptor, say whether it is on the efficient path (2D, last-stride-1, locally-created, dot-feeding, OWord-pitched → same 2D-block-I/O the block pointer used). For any descriptor that could not be put on the fast path, flag it with the reason (untraceable across tt.call, non-OWord pitch, not feeding a tl.dot, rank>2 unfoldable, rank-1 unit-stride 1D descriptor). A non-unit-stride rank-1 block pointer (Rule 11) emits no descriptor — report it under "Changes made" as a masked tensor-of-pointer load on the pointer path, not as a per-descriptor note. Do not over-promise: the README's ">2x" figure is vs tensor-of-pointers, not vs block pointers.
4. Caller-affecting changes (only if Rule 12 fired) — the changed helper signature and the list of call sites to update.

Self-check before returning

• Zero tl.make_block_ptr / tl.advance remain (unless Rule 12 deliberately keeps a block-ptr arg pending a caller fix you've flagged).
• Number of make_tensor_descriptor ≈ number of distinct block pointers, minus any non-unit-stride rank-1 pointers that became masked tensor-of-pointer loads (Rule 11).
• Every descriptor has last stride == 1 (transposes via .T, not last-stride≠1).
• No descriptor was emitted for a non-unit-stride rank-1 block pointer — each such pointer became a masked tl.load/tl.store whose mask= reproduces the original boundary_check, flagged off the fast path (Rule 11). (A strides=(S,) descriptor with S != 1 would not compile.)
• Descriptors are created in the function that loads from them (Rule 12).
• padding_option="nan" carried to the constructor iff the source used it.
• No triton.set_allocator / host-launch change anywhere in the output.

Source: intel/intel-xpu-backend-for-triton — distributed by TomeVault.