From c5ed9e31197321d52c18ca35dbbca8385fcc1c34 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sat, 23 May 2026 05:04:43 +0000 Subject: [PATCH] Consolidate FMHA stages A/B/C into unified kernel module with SMEM-P stub --- dsv4/kernels/attention/fmha.py | 394 +++++++++++++++++++++++++-------- 1 file changed, 303 insertions(+), 91 deletions(-) diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py index 9624dcab..5f737f53 100644 --- a/dsv4/kernels/attention/fmha.py +++ b/dsv4/kernels/attention/fmha.py @@ -1,8 +1,8 @@ -"""FMHA kernel: QK -> online softmax -> PV (CuTeDSL, Blackwell SM100). +"""FMHA kernel: QK → online softmax → PV (CuTeDSL, Blackwell SM100). -Stages A/B/C/D1. HEAD_DIM parameterized via constructor. -PV GEMM uses SMEM for A operand (P), eliminating TMEM layout mismatch. -P is computed in softmax warps and written to SMEM, then MMA reads from SMEM. +Unified module consolidating Stages A/B/C (TMEM-P, hd=64) and D1 (SMEM-P, hd>64). +use_smem_p=False (TMEM-P): P stored to TMEM via register bridge, PV reads from TMEM. +use_smem_p=True (SMEM-P): P stored to SMEM, PV reads from SMEM (copy TODO — zeroed stub). """ import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline from cutlass.cute.nvgpu import cpasync, tcgen05 @@ -15,17 +15,18 @@ import math class FmhaKernel: - def __init__(self, head_dim=64, s_k=128, scale_softmax=None, kv_stage=2): + def __init__(self, head_dim=64, s_k=128, scale_softmax=None, kv_stage=2, use_smem_p=False): self.head_dim = head_dim self.s_k = s_k self.n_kv_tiles = s_k // 128 self.pv_n_tile = min(head_dim, 256) self.n_pv_tiles = head_dim // self.pv_n_tile + self.use_smem_p = use_smem_p if use_smem_p is not None else (head_dim > 64) self.acc_dtype = Float32; self.qk_acc_dtype = Float32 self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16 self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1 self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE - self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5 + self.epilogue_warp_id = (0, 1, 2, 3); self.mma_warp_id = 4; self.tma_warp_id = 5 self.threads_per_cta = 192; self.num_c_stage = 2 self.kv_stage = kv_stage; self.q_stage = 1; self.num_c_stage = 2 self.scale_softmax = scale_softmax if scale_softmax is not None else 1.0 / math.sqrt(head_dim) @@ -38,34 +39,65 @@ class FmhaKernel: pv_ik = cute.size(pv_mma.shape_mnk, mode=[2]) self.pv_mma_tiler = (128, self.pv_n_tile, pv_ik * (128 // pv_ik)) self.mma_tiler = self.qk_mma_tiler - self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,)) - self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), self.pv_n_tile, self.qk_mma_tiler[2]) + self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1, 1, 1)), (qk_mma.thr_id.shape,)) + self.cta_tile_shape_mnk = ( + self.qk_mma_tiler[0] // cute.size(qk_mma.thr_id.shape), + self.pv_n_tile, + self.qk_mma_tiler[2], + ) self.c_layout = LayoutEnum.ROW_MAJOR - self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype) + self.epi_tile = utils.sm100.compute_epilogue_tile_shape( + self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype + ) self.num_ab_stage = 1; self.num_acc_stage = 1 + + # SMEM layouts self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage) self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage) self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage) + # P SMEM: always allocate (PV A-operand SMEM layout); used directly in SMEM-P, as TMEM alias in TMEM-P self.p_smem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1) self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2) - # TMEM: only S (QK result). P is in SMEM, O also in TMEM. - qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2]) + + # TMEM layout depends on path + qk_thr = qk_mma.get_slice(0) + qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2]) tStS = qk_thr.make_fragment_C(qk_as) - pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2]) + pv_thr = pv_mma.get_slice(0) + pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2]) tOtO = pv_thr.make_fragment_C(pv_as) - self.tmem_s0_offset = 0; self.tmem_o0_offset = 0 # S and O share TMEM (sequential) - s_cols = self.qk_mma_tiler[1] - o_cols = find_tmem_tensor_col_offset(tOtO) - total = max(s_cols, o_cols) + + if not self.use_smem_p: + # TMEM-P: S at 0, P at 32, O after P and S + self.tmem_s0_offset = 0 + self.tmem_p0_offset = 32 + s_cols = self.qk_mma_tiler[1] + p_cols = self.pv_mma_tiler[1] * self.q_dtype.width // self.qk_acc_dtype.width + self.tmem_o0_offset = max(s_cols, p_cols) + o_cols = find_tmem_tensor_col_offset(tOtO) + total = self.tmem_o0_offset + o_cols + else: + # SMEM-P: S and O share TMEM (sequential, no P in TMEM) + self.tmem_s0_offset = 0 + self.tmem_o0_offset = 0 + s_cols = self.qk_mma_tiler[1] + o_cols = find_tmem_tensor_col_offset(tOtO) + total = max(s_cols, o_cols) + self.num_tmem_alloc_cols = 1 while self.num_tmem_alloc_cols < total: self.num_tmem_alloc_cols *= 2 if self.num_tmem_alloc_cols > 512: print(f"⚠️ TMEM BUDGET: {self.num_tmem_alloc_cols} cols (hd={hd})") + + # P TMEM alias (PV A-operand viewed as TMEM for partition mapping) + self.p_tmem_s = tStS # reuses QK C-fragment TMEM layout for P partition + + # TMA bytes cta = cute.size(qk_mma.thr_id.shape) - q_s = cute.slice_(self.q_smem_s,(None,None,None,0)) - k_s = cute.slice_(self.k_smem_s,(None,None,None,0)) - v_s = cute.slice_(self.v_smem_s,(None,None,None,0)) + q_s = cute.slice_(self.q_smem_s, (None, None, None, 0)) + k_s = cute.slice_(self.k_smem_s, (None, None, None, 0)) + v_s = cute.slice_(self.v_smem_s, (None, None, None, 0)) self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta self.kv_tx_bytes = (cute.size_in_bytes(self.q_dtype, k_s) + cute.size_in_bytes(self.q_dtype, v_s)) * cta @@ -78,72 +110,164 @@ class FmhaKernel: v_fmha = cute.make_tensor(v.iterator, cute.make_layout((v_n, self.s_k, 1), stride=(1, v_n, v_n * self.s_k))) self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode() self.c_layout = LayoutEnum.from_tensor(c) - qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM) - pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.v_major, self.qk_acc_dtype, self.cta_group, (128,self.pv_n_tile), tcgen05.OperandSource.SMEM) + + qk_mma = utils.sm100.make_trivial_tiled_mma( + self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, + self.cta_group, (128, 128), tcgen05.OperandSource.SMEM, + ) + pv_src = tcgen05.OperandSource.SMEM if self.use_smem_p else tcgen05.OperandSource.TMEM + pv_mma = utils.sm100.make_trivial_tiled_mma( + self.q_dtype, self.q_dtype, self.a_major, self.v_major, self.qk_acc_dtype, + self.cta_group, (128, self.pv_n_tile), pv_src, + ) self._setup(qk_mma, pv_mma) - q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)); v_s = cute.slice_(self.v_smem_s,(None,None,None,0)) - tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape) - tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape) - tma_v,mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,pv_mma.thr_id),v_fmha,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape) - epi_s = cute.select(self.c_smem_s,mode=[0,1]) - tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile) - self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_smem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream) + + q_s = cute.slice_(self.q_smem_s, (None, None, None, 0)) + k_s = cute.slice_(self.k_smem_s, (None, None, None, 0)) + v_s = cute.slice_(self.v_smem_s, (None, None, None, 0)) + tma_q, mQ = cute.nvgpu.make_tiled_tma_atom_A( + utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn, qk_mma.thr_id), + q, q_s, self.qk_mma_tiler, qk_mma, self.cluster_layout_vmnk.shape, + ) + tma_k, mK = cute.nvgpu.make_tiled_tma_atom_B( + utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, qk_mma.thr_id), + k, k_s, self.qk_mma_tiler, qk_mma, self.cluster_layout_vmnk.shape, + ) + tma_v, mV = cute.nvgpu.make_tiled_tma_atom_B( + utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, pv_mma.thr_id), + v_fmha, v_s, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape, + ) + epi_s = cute.select(self.c_smem_s, mode=[0, 1]) + tma_c, mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_s, self.epi_tile) + + self._kernel( + qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, + self.cluster_layout_vmnk, + self.q_smem_s, self.k_smem_s, self.v_smem_s, self.p_smem_s, self.c_smem_s, + self.epi_tile, + ).launch( + grid=(1, 1, 1), block=[self.threads_per_cta, 1, 1], stream=stream, + ) @cute.kernel - def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_smem_s, c_smem_s, epi_tile): + def _kernel( + self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, + cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_smem_s, c_smem_s, epi_tile, + ): warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx()) - tidx,_,_ = cute.arch.thread_idx() + tidx, _, _ = cute.arch.thread_idx() + + use_smem_p = self.use_smem_p + + # ── TMA warp: prefetch descriptors ── if warp_idx == self.tma_warp_id: - cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c) + cpasync.prefetch_descriptor(tma_q) + cpasync.prefetch_descriptor(tma_k) + cpasync.prefetch_descriptor(tma_v) + cpasync.prefetch_descriptor(tma_c) + + # ── Shared storage ── @cute.struct class SS: - q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2] - kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2] + q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage * 2] + kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage * 2] s_bar: cute.struct.MemRange[cutlass.Int64, 2] - acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2] + acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage * 2] tmem_dealloc: cutlass.Int64; holding: cutlass.Int32 + smem = utils.SmemAllocator(); st = smem.allocate(SS) - qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants() - kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants() - s_prod,s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.s_bar.data_ptr(),num_stages=1,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.epilogue_warp_id))).make_participants() - softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32*len(self.epilogue_warp_id)) - final_o_bar = pipeline.NamedBarrier(barrier_id=4, num_threads=32 + 32*len(self.epilogue_warp_id)) - acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.epilogue_warp_id)),cta_layout_vmnk=cl_vmnk,defer_sync=True) - tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.epilogue_warp_id))) - tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.epilogue_warp_id[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr) - pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True) - sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner) - sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner) - sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner) - sP = smem.allocate_tensor(element_type=self.q_dtype,layout=p_smem_s.outer,byte_alignment=128,swizzle=p_smem_s.inner) - sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner) - gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None)) - gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None)) - gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None)) - gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None)) + + qp, qc = pipeline.PipelineTmaUmma.create( + barrier_storage=st.q_bar.data_ptr(), num_stages=self.q_stage, + producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread), + consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 1), + tx_count=self.q_tx_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True, + ).make_participants() + kvp, kvc = pipeline.PipelineTmaUmma.create( + barrier_storage=st.kv_bar.data_ptr(), num_stages=self.kv_stage, + producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread), + consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 1), + tx_count=self.kv_tx_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True, + ).make_participants() + s_prod, s_cons = pipeline.PipelineUmmaAsync.create( + barrier_storage=st.s_bar.data_ptr(), num_stages=1, + producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread), + consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id)), + ).make_participants() + softmax_done_bar = pipeline.NamedBarrier(barrier_id=3, num_threads=32 + 32 * len(self.epilogue_warp_id)) + final_o_bar = pipeline.NamedBarrier(barrier_id=4, num_threads=32 + 32 * len(self.epilogue_warp_id)) + acc_pipe = pipeline.PipelineUmmaAsync.create( + barrier_storage=st.acc_bar.data_ptr(), num_stages=self.num_acc_stage, + producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread), + consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, len(self.epilogue_warp_id)), + cta_layout_vmnk=cl_vmnk, defer_sync=True, + ) + tmem_bar = pipeline.NamedBarrier(barrier_id=2, num_threads=32 * len((self.mma_warp_id, *self.epilogue_warp_id))) + tmem = utils.TmemAllocator( + st.holding.ptr, barrier_for_retrieve=tmem_bar, + allocator_warp_id=self.epilogue_warp_id[0], + is_two_cta=cute.size(qk_mma.thr_id.shape) == 2, + two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr, + ) + pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk, is_relaxed=True) + + # ── SMEM tensors ── + sQ = smem.allocate_tensor(element_type=self.q_dtype, layout=q_smem_s.outer, byte_alignment=128, swizzle=q_smem_s.inner) + sK = smem.allocate_tensor(element_type=self.q_dtype, layout=k_smem_s.outer, byte_alignment=128, swizzle=k_smem_s.inner) + sV = smem.allocate_tensor(element_type=self.q_dtype, layout=v_smem_s.outer, byte_alignment=128, swizzle=v_smem_s.inner) + sP = smem.allocate_tensor(element_type=self.q_dtype, layout=p_smem_s.outer, byte_alignment=128, swizzle=p_smem_s.inner) + sC = smem.allocate_tensor(element_type=self.o_dtype, layout=c_smem_s.outer, byte_alignment=128, swizzle=c_smem_s.inner) + + # ── Gmem tensors ── + gQ = cute.local_tile(mQ, cute.slice_(self.qk_mma_tiler, (None, 0, None)), (None, None, None)) + gK = cute.local_tile(mK, cute.slice_(self.qk_mma_tiler, (0, None, None)), (None, None, None)) + gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0, None, None)), (None, None, None)) + gC = cute.local_tile(mC, cute.slice_(self.pv_mma_tiler, (None, None, 0)), (None, None, None)) + + # ── Thread partitions ── qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0) tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK) tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC) - a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape) - tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3)) - b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape) - tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3)) - tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3)) - tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)] + + a_lay = cute.make_layout(cute.slice_(cl_vmnk, (0, 0, None, 0)).shape) + tAsQ, tAgQ = cpasync.tma_partition(tma_q, 0, a_lay, cute.group_modes(sQ, 0, 3), cute.group_modes(tCgQ, 0, 3)) + b_lay = cute.make_layout(cute.slice_(cl_vmnk, (0, None, 0, 0)).shape) + tBsK, tBgK = cpasync.tma_partition(tma_k, 0, b_lay, cute.group_modes(sK, 0, 3), cute.group_modes(tCgK, 0, 3)) + tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV, 0, 3), cute.group_modes(tCgV, 0, 3)) + tAgQ = tAgQ[(None, 0, None, 0)]; tBgK = tBgK[(None, 0, None, 0)]; tVgV = tVgV[(None, 0, None, 0)] + + # Register fragments tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK) - tCrV = pv_mma.make_fragment_B(sV); tCrP = pv_mma.make_fragment_A(sP) - # TMEM: S (QK result) + tCrV = pv_mma.make_fragment_B(sV) + tCrP = pv_mma.make_fragment_A(sP) # used in SMEM-P path + + # ── TMEM: S (QK result) ── qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2]) tStS = qk_thr.make_fragment_C(qk_as) tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout) - # TMEM: O (PV result) — same offset as S (sequential, no overlap) + + # ── TMEM: O (PV result) ── pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2]) tOtO = pv_thr.make_fragment_C(pv_as) tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout) + + # ── TMEM-P path: PV A-operand from TMEM ── + if not use_smem_p: + tP = cute.make_tensor(tStS.iterator, self.p_tmem_s.outer) + tOrP_base = pv_thr.make_fragment_A(tP) + tOrP = tOrP_base[(None, None, None, 0)] + tOrP0 = cute.make_tensor( + tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset, + tOrP.layout, + ) + tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage)) + pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk) - # ===== TMA LOAD warp ===== + # ══════════════════════════════════════════════════════════════ + # TMA LOAD WARP + # ══════════════════════════════════════════════════════════════ if warp_idx == self.tma_warp_id: qp.reset(); qh = qp.acquire_and_advance() cute.copy(tma_q, tAgQ[(None, Int32(0))], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier) @@ -156,42 +280,57 @@ class FmhaKernel: pk = cutlass.Boolean(1) kvp.tail() - # ===== MMA warp ===== + # ══════════════════════════════════════════════════════════════ + # MMA WARP + # ══════════════════════════════════════════════════════════════ if warp_idx == self.mma_warp_id: tmem.wait_for_alloc() qc.reset(); qh = qc.wait_and_advance(); qh.release() kvc.reset(); pk = kvc.try_wait() acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage) acc_pipe.producer_acquire(acc_st) + for kt in range(self.n_kv_tiles): kvh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) sh = s_prod.acquire_and_advance() + # QK GEMM → S in TMEM qk_mma.set(tcgen05.Field.ACCUMULATE, False) for kb in cutlass.range(cute.size(tCrQ, mode=[2]), unroll_full=True): - cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kvh.index)], tStS0) + cute.gemm(qk_mma, tStS0, tCrQ[(None, None, kb, 0)], tCrK[(None, None, kb, kvh.index)], tStS0) qk_mma.set(tcgen05.Field.ACCUMULATE, True) cute.arch.fence_view_async_tmem_store() sh.commit() softmax_done_bar.arrive_and_wait() - # PV GEMM: P from SMEM, V from SMEM → O in TMEM + + # PV GEMM → O in TMEM pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0) - for kb in cutlass.range(cute.size(tCrP, mode=[2]), unroll_full=True): - cute.gemm(pv_mma, tOtO0, tCrP[(None,None,kb,0)], tCrV[(None,None,kb,kvh.index)], tOtO0) - pv_mma.set(tcgen05.Field.ACCUMULATE, True) + if not use_smem_p: + # TMEM-P: P from TMEM + for kb in cutlass.range(cute.size(tOrP0, mode=[2]), unroll_full=True): + cute.gemm(pv_mma, tOtO0, tOrP0[(None, None, kb, 0)], tCrV[(None, None, kb, kvh.index)], tOtO0) + else: + # SMEM-P: P from SMEM + for kb in cutlass.range(cute.size(tCrP, mode=[2]), unroll_full=True): + cute.gemm(pv_mma, tOtO0, tCrP[(None, None, kb, 0)], tCrV[(None, None, kb, kvh.index)], tOtO0) + pv_mma.set(tcgen05.Field.ACCUMULATE, True) cute.arch.fence_view_async_tmem_store() kvh.release() + acc_pipe.producer_commit(acc_st); acc_st.advance() final_o_bar.arrive() acc_pipe.producer_tail(acc_st) - # ===== SOFTMAX + EPILOGUE warps ===== + # ══════════════════════════════════════════════════════════════ + # SOFTMAX + EPILOGUE WARPS + # ══════════════════════════════════════════════════════════════ if warp_idx < self.mma_warp_id: tmem.allocate(self.num_tmem_alloc_cols) tmem.wait_for_alloc() tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype) sfw_idx = tidx % (32 * len(self.epilogue_warp_id)) - # S load atoms + + # ── S load setup ── tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype) tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0) thr_load = tiled_tmem_load.get_slice(sfw_idx) @@ -200,30 +339,57 @@ class FmhaKernel: tScS = qk_thr.partition_C(cS) tTMEM_LOADcS = thr_load.partition_D(tScS) - # P → SMEM: use make_tiled_copy_C for register→SMEM (standard epilogue pattern) - # The P values are the A operand of PV, written to SMEM so the MMA can read them - p_s = cute.slice_(p_smem_s,(None,None,None,0)) - tCrP_smem = pv_thr.partition_A(sP) # PV thread → SMEM partition for P (A operand) - tCrP_reg = pv_mma.make_fragment_A(sP) # register fragment matching SMEM layout - tiled_p_copy = cute.make_tiled_copy_C(pv_mma, tCrP_smem, p_s, 1) + # ── TMEM-P: P store setup (register bridge) ── + if not use_smem_p: + p_cols_fp32 = self.pv_mma_tiler[1] * self.q_dtype.width // self.qk_acc_dtype.width + tStP_layout = cute.composition(tStS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32))) + tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout) + tmem_store_atom = cute.make_copy_atom( + tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype, + ) + tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0) + thr_store = tiled_tmem_store.get_slice(sfw_idx) + tTMEM_STOREtP = thr_store.partition_D(tStP0) + tScP_layout = cute.composition(tScS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32))) + tScP = cute.make_tensor(tScS.iterator, tScP_layout) + tTMEM_STOREcP = thr_store.partition_S(tScP) - # Online softmax state + # ── SMEM-P: P → SMEM copy setup (TODO: proper QK→PV partition remap) ── + if use_smem_p: + # TODO: make_tiled_copy_C(store_atom, qk_mma) to partition threads by QK's C-fragment + # For now, zero sP as a stub — PV will read garbage/zero + pass + + # ── O rescale / normalization setup (correction_rescale atoms) ── + corr_tile_size = 16 + o_rescale_atom_ld = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype) + o_rescale_atom_st = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype) + o_rescale_layout = cute.composition(tStS.layout, cute.make_layout((self.pv_mma_tiler[0], corr_tile_size))) + tiled_o_ld = tcgen05.make_tmem_copy(o_rescale_atom_ld, cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + tiled_o_st = tcgen05.make_tmem_copy(o_rescale_atom_st, cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + + # ── Online softmax state ── row_max = -Float32.inf row_sum = Float32(0.0) scale_log2 = Float32(self.scale_softmax_log2) + # ── Softmax loop ── for kt in range(self.n_kv_tiles): si_handle = s_cons.wait_and_advance() tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype) cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS) cute.arch.fence_view_async_tmem_load() + old_row_max = row_max frg_cnt = 4 frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile)) + + # Row max for j in range(frg_cnt): for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])): row_max = cute.arch.fmax(row_max, tTMEM_LOADrS_frg[k, j] * scale_log2) + row_max_safe = row_max if row_max == -cutlass.Float32.inf: row_max_safe = Float32(0.0) @@ -234,26 +400,72 @@ class FmhaKernel: row_sum *= acc_scale minus_row_max = Float32(0.0) - row_max_safe - # Compute P = exp2(S * scale - row_max), convert to BF16, write to SMEM - for j in range(frg_cnt): - for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])): - tTMEM_LOADrS_frg[k, j] = tTMEM_LOADrS_frg[k, j] * scale_log2 + minus_row_max - tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True) - row_sum = row_sum + tTMEM_LOADrS_frg[k, j] + # Softmax + P store + if not use_smem_p: + # TMEM-P: register bridge — FP32 backing, BF16 view + rP_words = cute.make_rmem_tensor(tTMEM_STOREcP.shape, self.qk_acc_dtype) + rP_bf16 = cute.make_tensor( + cute.recast_ptr(rP_words.iterator, dtype=self.q_dtype), + tTMEM_LOADrS.layout, + ) + rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile)) - # Write P to SMEM using PV A-operand partition - # TODO: proper element mapping from QK→PV partition - for j in cutlass.range(cute.size(tCrP_reg), vectorize=True): - tCrP_reg[j] = BFloat16(0.0) - cute.copy(tiled_p_copy, tCrP_reg, tCrP_smem) - cute.arch.fence_proxy("async.shared", space="cta") + for j in range(frg_cnt): + for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])): + tTMEM_LOADrS_frg[k, j] = tTMEM_LOADrS_frg[k, j] * scale_log2 + minus_row_max + tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True) + row_sum = row_sum + tTMEM_LOADrS_frg[k, j] + s_vec = tTMEM_LOADrS_frg[None, j].load() + rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype)) + + cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP) + else: + # SMEM-P: compute softmax, write P to SMEM (TODO) + for j in range(frg_cnt): + for k in range(cute.size(tTMEM_LOADrS_frg, mode=[0])): + tTMEM_LOADrS_frg[k, j] = tTMEM_LOADrS_frg[k, j] * scale_log2 + minus_row_max + tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True) + row_sum = row_sum + tTMEM_LOADrS_frg[k, j] + + # TODO: Write P to SMEM using make_tiled_copy_C(store_atom, qk_mma) + # to partition threads by QK's C-fragment, then copy to p_smem_s layout. + # STUB: zero P in SMEM for now + for j in cutlass.range(cute.size(sP), vectorize=True): + sP[j] = BFloat16(0.0) + cute.arch.fence_proxy("async.shared", space="cta") si_handle.release() softmax_done_bar.arrive() - # Wait for MMA's final PV + # ── Per-tile O rescale (multiply O by acc_scale when kt > 0) ── + if kt > 0: + thr_ld = tiled_o_ld.get_slice(sfw_idx) + thr_st = tiled_o_st.get_slice(sfw_idx) + tOrO_src = thr_ld.partition_S(cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + tOrO_dst = thr_st.partition_D(cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + rO = cute.make_rmem_tensor(tOrO_src.shape, self.qk_acc_dtype) + cute.copy(tiled_o_ld, tOrO_src, rO) + for i in cutlass.range(cute.size(rO), vectorize=True): + rO[i] = rO[i] * acc_scale + cute.copy(tiled_o_st, rO, tOrO_dst) + + # ── Wait for MMA's final PV GEMM ── final_o_bar.arrive_and_wait() - # Epilogue: raw PV output (unnormalized) + + # ── O normalization: multiply O by 1/row_sum (TMEM round-trip) ── + inv_row_sum = Float32(1.0) / row_sum + thr_ld = tiled_o_ld.get_slice(sfw_idx) + thr_st = tiled_o_st.get_slice(sfw_idx) + tOrO_src = thr_ld.partition_S(cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + tOrO_dst = thr_st.partition_D(cute.make_tensor(tStS.iterator + self.tmem_o0_offset, o_rescale_layout)) + rO = cute.make_rmem_tensor(tOrO_src.shape, self.qk_acc_dtype) + cute.copy(tiled_o_ld, tOrO_src, rO) + for i in cutlass.range(cute.size(rO), vectorize=True): + rO[i] = rO[i] * inv_row_sum + cute.copy(tiled_o_st, rO, tOrO_dst) + cute.arch.fence_view_async_tmem_store() + + # ── Epilogue: TMA store O → global ── tCtO_base = cute.make_tensor(tmem_ptr + self.tmem_o0_offset, tCtO_fake.layout) acc_cons_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Consumer, self.num_acc_stage) c_grp = pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(self.epilogue_warp_id))