kernbench2/src/kernbench/benches/_gqa_attention_decode_long.py

"""GQA fused-attention decode kernel — long context (ADR-0060).

Each rank holds an ``S_local = S_kv / (C·P)`` slice of K, V and the full
Q (replicated). The local attention is computed via an S_kv-axis tile
sweep (ADR-0063 §A.2) so per-rank scratch is bounded by ``TILE_S_KV``
regardless of ``S_local``. The partial ``(m, ℓ, O)`` is then reduced
to PE 0 of CUBE 0 via a 2-level chain (intra-CUBE row+col, then
inter-CUBE), and the root writes the final output.

Topology / SFR:
  - Requires ``configure_sfr_intercube_multisip`` when ``P > 1`` or
    ``C > 1`` (provides disjoint ``intra_*`` and ``E/W/N/S`` namespaces).
  - Intra-CUBE PEs are arranged as a 2×4 grid (no wrap).
  - Inter-CUBE CUBEs are arranged as a 1D row (no wrap).

Layout caveats:
  - GEMMs use ``tl.dot`` (no composite epilogue / ``softmax_scale``).
  - K is loaded as ``(d_head, S_local)`` via byte-conserving reshape of
    the deployed ``(S_local, h_kv·d_head)`` slice — correct for zero /
    symmetric inputs (ADR-0060 §3 reshape-not-transpose caveat).
"""
from __future__ import annotations


TILE_S_KV = 1024  # ADR-0063 §A.2 S_kv-axis tile sweep (per-tile width).


def _merge_running(m_local, l_local, O_local, m_other, l_other, O_other, *, tl):
    """Online-softmax merge of two partial ``(m, ℓ, O)`` triples."""
    m_new = tl.maximum(m_local, m_other)
    scale_old = tl.exp(m_local - m_new)
    scale_new = tl.exp(m_other - m_new)
    l_new = l_local * scale_old + l_other * scale_new
    O_new = O_local * scale_old + O_other * scale_new
    return m_new, l_new, O_new


def gqa_attention_decode_long_kernel(
    q_ptr: int,
    k_ptr: int,
    v_ptr: int,
    o_ptr: int,
    T_q: int,
    S_kv: int,
    h_q: int,
    h_kv: int,
    d_head: int,
    C: int,
    P: int,
    *,
    tl,
) -> None:
    """GQA decode with M-fold + S_kv tile sweep + 2-level chain reduce-to-root.

    Tensor layout:
      Q : (T_q, h_q · d_head) replicated on every rank; loaded as
          (G·T_q, d_head) — byte-conserving and math-correct for T_q=1.
      K : (S_kv, h_kv · d_head) sharded row_wise by (cube, pe); each rank
          loads its (d_head, S_local) slice via byte-conserving reshape.
      V : (S_kv, h_kv · d_head) sharded row_wise by (cube, pe); each rank
          loads its (S_local, d_head) slice.
      O : (T_q, h_q · d_head) — only PE 0 of CUBE 0 stores.
    """
    G = h_q // h_kv
    n_ranks = C * P
    S_local = S_kv // n_ranks
    pe_id = tl.program_id(axis=0)
    cube_id = tl.program_id(axis=1)

    # ── Local attention (S_kv-axis tile sweep, ADR-0063 §A.2) ──
    Q = tl.load(q_ptr, shape=(G * T_q, d_head), dtype="f16")
    n_tiles = (S_local + TILE_S_KV - 1) // TILE_S_KV
    KV_ROW_BYTES = d_head * 2  # f16

    # Tile 0: establishes persistent (m_local, l_local, O_local).
    #
    # Cannot be folded into the Tiles 1..N loop (kernbench-only limitation):
    #   - persistent (m, ℓ, O) must live OUTSIDE ``tl.scratch_scope``,
    #     otherwise scope teardown discards them before the next tile's
    #     merge can read them;
    #   - kernbench has no scratch-backed initializer — ``tl.zeros`` /
    #     ``tl.full`` return addr=0 handles with no backing storage, so
    #     they cannot be overwritten via ``tl.copy_to`` to seed (-inf, 0, 0).
    # So Tile 0 computes the initial running state directly; Tiles 1..N
    # fold into it. Triton port: limitation does not apply (SSA tensors
    # stay live across iterations) — a single unified loop suffices.
    tile_s0 = min(TILE_S_KV, S_local)
    K_T = tl.load(k_ptr, shape=(d_head, tile_s0), dtype="f16")
    V = tl.load(v_ptr, shape=(tile_s0, d_head), dtype="f16")
    scores = tl.dot(Q, K_T)
    m_local = tl.max(scores, axis=-1)
    centered = scores - m_local
    exp_scores = tl.exp(centered)
    l_local = tl.sum(exp_scores, axis=-1)
    O_local = tl.dot(exp_scores, V)

    # Tiles 1..n_tiles-1: fold into running state via online-softmax merge.
    # Triton port: drop the ``with tl.scratch_scope():`` line and replace
    # each ``copy_to`` with a Python rebind.
    for tile_idx in range(1, n_tiles):
        tile_start = tile_idx * TILE_S_KV
        tile_s = min(TILE_S_KV, S_local - tile_start)
        with tl.scratch_scope():
            K_T_t = tl.load(k_ptr + tile_start * KV_ROW_BYTES,
                            shape=(d_head, tile_s), dtype="f16")
            V_t = tl.load(v_ptr + tile_start * KV_ROW_BYTES,
                          shape=(tile_s, d_head), dtype="f16")
            scores_t = tl.dot(Q, K_T_t)
            m_tile = tl.max(scores_t, axis=-1)
            centered_t = scores_t - m_tile
            exp_scores_t = tl.exp(centered_t)
            l_tile = tl.sum(exp_scores_t, axis=-1)
            O_tile = tl.dot(exp_scores_t, V_t)
            m_new, l_new, O_new = _merge_running(
                m_local, l_local, O_local, m_tile, l_tile, O_tile, tl=tl,
            )
            tl.copy_to(m_local, m_new)
            tl.copy_to(l_local, l_new)
            tl.copy_to(O_local, O_new)

    # ── Communication: chain reduce-to-root at PE 0 of CUBE 0 ──
    PE_GRID_COLS = 4
    pe_col = pe_id % PE_GRID_COLS
    pe_row = pe_id // PE_GRID_COLS
    pe_cols_used = min(PE_GRID_COLS, P)
    pe_rows_used = (P + PE_GRID_COLS - 1) // PE_GRID_COLS

    # Level-2 row chain (intra-CUBE, along intra_W, leftward).
    if pe_cols_used > 1:
        if pe_col < pe_cols_used - 1:
            with tl.scratch_scope():
                m_other = tl.recv(dir="intra_E", shape=m_local.shape, dtype="f16")
                l_other = tl.recv(dir="intra_E", shape=l_local.shape, dtype="f16")
                O_other = tl.recv(dir="intra_E", shape=O_local.shape, dtype="f16")
                m_new, l_new, O_new = _merge_running(
                    m_local, l_local, O_local, m_other, l_other, O_other, tl=tl,
                )
                tl.copy_to(m_local, m_new)
                tl.copy_to(l_local, l_new)
                tl.copy_to(O_local, O_new)
        if pe_col > 0:
            tl.send(dir="intra_W", src=m_local)
            tl.send(dir="intra_W", src=l_local)
            tl.send(dir="intra_W", src=O_local)

    # Level-2 col bridge (intra-CUBE, along intra_N, row-1 → row-0).
    if pe_col == 0 and pe_rows_used > 1:
        if pe_row < pe_rows_used - 1:
            with tl.scratch_scope():
                m_other = tl.recv(dir="intra_S", shape=m_local.shape, dtype="f16")
                l_other = tl.recv(dir="intra_S", shape=l_local.shape, dtype="f16")
                O_other = tl.recv(dir="intra_S", shape=O_local.shape, dtype="f16")
                m_new, l_new, O_new = _merge_running(
                    m_local, l_local, O_local, m_other, l_other, O_other, tl=tl,
                )
                tl.copy_to(m_local, m_new)
                tl.copy_to(l_local, l_new)
                tl.copy_to(O_local, O_new)
        if pe_row > 0:
            tl.send(dir="intra_N", src=m_local)
            tl.send(dir="intra_N", src=l_local)
            tl.send(dir="intra_N", src=O_local)

    # Level-1 inter-CUBE chain (along W, leftward; only PE 0 of each CUBE).
    if pe_id == 0 and C > 1:
        if cube_id < C - 1:
            with tl.scratch_scope():
                m_other = tl.recv(dir="E", shape=m_local.shape, dtype="f16")
                l_other = tl.recv(dir="E", shape=l_local.shape, dtype="f16")
                O_other = tl.recv(dir="E", shape=O_local.shape, dtype="f16")
                m_new, l_new, O_new = _merge_running(
                    m_local, l_local, O_local, m_other, l_other, O_other, tl=tl,
                )
                tl.copy_to(m_local, m_new)
                tl.copy_to(l_local, l_new)
                tl.copy_to(O_local, O_new)
        if cube_id > 0:
            tl.send(dir="W", src=m_local)
            tl.send(dir="W", src=l_local)
            tl.send(dir="W", src=O_local)

    # ── Final normalise + store (root only) ──
    if pe_id == 0 and cube_id == 0:
        O_final = O_local / l_local
        tl.store(o_ptr, O_final)