diff --git a/tests/unit/test_fmha_v3_stage_c_full.py b/tests/unit/test_fmha_v3_stage_c_full.py index 5ce99072..588cfebd 100644 --- a/tests/unit/test_fmha_v3_stage_c_full.py +++ b/tests/unit/test_fmha_v3_stage_c_full.py @@ -1,13 +1,8 @@ """ -FMHA v3 Stage-C Full: Production Blackwell pipeline with real softmax + correction. - -Architecture (12-warps, matches CUTLASS FMHA): - softmax warps 0-3 : S(TMEM) -> softmax -> P(TMEM), vec(TMEM) - correction warps 4-7 : vec(TMEM) + O(TMEM) -> corrected O(SMEM) - MMA warp 8 : QK and PV - TMA/load warp 9 : Q/K/V load - epilogue warp 10 : corrected O SMEM -> GMEM via TMA - empty warp 11 : tmem dealloc mbar init +FMHA v3 Stage-C: Real softmax + O normalization. +Builds on the 12w identity-softmax test by replacing identity softmax with +online softmax (row_max, exp2 scaling, P store) and adding O normalization +by row_sum before the epilogue writes to GMEM. """ import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline from cutlass.cute.nvgpu import cpasync, tcgen05 @@ -21,26 +16,16 @@ import math HEAD_DIM = 64 class FmhaV3StageC: - def __init__(self, s_k=128, scale_softmax=None): - self.s_k = s_k - self.acc_dtype = Float32; self.qk_acc_dtype = Float32; self.pv_acc_dtype = Float32 + def __init__(self, scale_softmax=None): + 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.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE - # 12-warp layout - self.softmax_warp_ids = (0, 1, 2, 3) - self.correction_warp_ids = (4, 5, 6, 7) - self.mma_warp_id = 8; self.tma_warp_id = 9 - self.epilogue_warp_id = 10; self.empty_warp_id = 11 - self.threads_per_cta = 32 * 12 - # Pipeline stages - self.mma_softmax_stage = 1; self.softmax_corr_stage = 1 - self.mma_corr_stage = 2; self.epi_stage = 2 - # TMA stages + 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.threads_per_cta = 192; self.num_c_stage = 2 self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2 - # Softmax scaling self.scale_softmax = scale_softmax if scale_softmax is not None else 1.0 / math.sqrt(HEAD_DIM) self.scale_softmax_log2 = self.scale_softmax * math.log2(math.e) - self.scale_output = 1.0 def _setup(self, qk_mma, pv_mma): qk_ik = cute.size(qk_mma.shape_mnk, mode=[2]) @@ -56,20 +41,27 @@ class FmhaV3StageC: 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) - self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, self.epi_stage) + self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2) self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1) 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]) tOtO = pv_thr.make_fragment_C(pv_as) - self.tmem_s0_offset = 0; self.tmem_vec0_offset = 0; self.tmem_p0_offset = 32 + self.tmem_s0_offset = 0; self.tmem_p0_offset = 32 + # P occupies [tmem_p0_offset, tmem_p0_offset + p_cols_fp32) + # S occupies [0, qk_mma_tiler[1]) = [0, 128) + # O must NOT overlap P. Place O after max(S end, P end), aligned to 32. p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width - p_end = self.tmem_p0_offset + p_cols_fp32; s_cols = self.qk_mma_tiler[1] - o_after = max(s_cols, p_end) - self.tmem_o0_offset = ((o_after + 31) // 32) * 32 - o_cols = find_tmem_tensor_col_offset(tOtO); total = self.tmem_o0_offset + o_cols + p_end = self.tmem_p0_offset + p_cols_fp32 # 32 + 64 = 96 + s_cols = self.qk_mma_tiler[1] # 128 + o_after = max(s_cols, p_end) # 128 + self.tmem_o0_offset = ((o_after + 31) // 32) * 32 # align to 32 = 128 + o_cols = find_tmem_tensor_col_offset(tOtO) # footprint of O + total = self.tmem_o0_offset + o_cols + # Must be multiple of 32 AND power of 2 self.num_tmem_alloc_cols = 1 - while self.num_tmem_alloc_cols < total: self.num_tmem_alloc_cols *= 2 + while self.num_tmem_alloc_cols < total: + self.num_tmem_alloc_cols *= 2 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)) self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta @@ -84,8 +76,8 @@ class FmhaV3StageC: v_fmha = cute.make_tensor( v.iterator, cute.make_layout( - (HEAD_DIM, self.s_k, 1), - stride=(1, HEAD_DIM, HEAD_DIM * self.s_k), + (HEAD_DIM, 128, 1), + stride=(1, HEAD_DIM, HEAD_DIM * 128), ), ) self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode() @@ -95,130 +87,143 @@ class FmhaV3StageC: 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_tmem_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_tmem_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() 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) @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] - mma_s_bar: cute.struct.MemRange[cutlass.Int64, self.mma_softmax_stage * 2] - s_corr_bar: cute.struct.MemRange[cutlass.Int64, self.softmax_corr_stage * 2] - mma_corr_bar: cute.struct.MemRange[cutlass.Int64, self.mma_corr_stage * 2] - corr_epi_bar: cute.struct.MemRange[cutlass.Int64, self.epi_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] tmem_dealloc: cutlass.Int64; holding: cutlass.Int32 - smem = utils.SmemAllocator(); st = smem.allocate(SS) - def cg(n): return pipeline.CooperativeGroup(pipeline.Agent.Thread, n) - qp, qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(), num_stages=self.q_stage, producer_group=cg(1), consumer_group=cg(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=cg(1), consumer_group=cg(1), tx_count=self.kv_tx_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True).make_participants() - mma_s_prod, mma_s_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_s_bar.data_ptr(), num_stages=self.mma_softmax_stage, producer_group=cg(1), consumer_group=cg(32 * len(self.softmax_warp_ids)), cta_layout_vmnk=cl_vmnk, defer_sync=True).make_participants() - s_corr_prod, s_corr_cons = pipeline.PipelineAsync.create(barrier_storage=st.s_corr_bar.data_ptr(), num_stages=self.softmax_corr_stage, producer_group=cg(32 * len(self.softmax_warp_ids)), consumer_group=cg(32 * len(self.correction_warp_ids))).make_participants() - mma_corr_prod, mma_corr_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_corr_bar.data_ptr(), num_stages=self.mma_corr_stage, producer_group=cg(1), consumer_group=cg(32 * len(self.correction_warp_ids)), cta_layout_vmnk=cl_vmnk, defer_sync=True).make_participants() - corr_epi_prod, corr_epi_cons = pipeline.PipelineAsync.create(barrier_storage=st.corr_epi_bar.data_ptr(), num_stages=self.epi_stage, producer_group=cg(32 * len(self.correction_warp_ids)), consumer_group=cg(32)).make_participants() - tmem_bar = pipeline.NamedBarrier(barrier_id=2, num_threads=32 * len((*self.softmax_warp_ids, *self.correction_warp_ids, self.mma_warp_id))) - tmem = utils.TmemAllocator(st.holding.ptr, barrier_for_retrieve=tmem_bar, allocator_warp_id=self.softmax_warp_ids[0], is_two_cta=cute.size(qk_mma.thr_id.shape) == 2, two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr) - if warp_idx == self.empty_warp_id: - cute.arch.mbarrier_init(st.tmem_dealloc, 32 * len((*self.softmax_warp_ids, *self.correction_warp_ids))) - cute.arch.mbarrier_init_fence() - pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk, is_relaxed=True) + 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)) + 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) - sC = smem.allocate_tensor(element_type=self.o_dtype, layout=c_smem_s.outer, byte_alignment=128, swizzle=c_smem_s.inner) + 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) + 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)) + 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)) n_kv_tiles = cute.size(gK, mode=[3]) 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)] + + tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK) + tCrV = pv_mma.make_fragment_B(sV) - tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK); tCrV = pv_mma.make_fragment_B(sV) 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) 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) + + # --- PV read view (for MMA only, NOT for softmax store) --- tP = cute.make_tensor(tStS.iterator, 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, 1)) + 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) + + tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage)) + 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 WARP (9) ==================== + # TMA LOAD if warp_idx == self.tma_warp_id: qp.reset(); qh = qp.acquire_and_advance() - cute.copy(tma_q, tAgQ[(None, qh.count)], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier) + cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier) qp.tail() kvp.reset(); pk = kvp.try_acquire() - for kt in cutlass.range(n_kv_tiles, unroll=1): + for kt in cutlass.range(n_kv_tiles,unroll=1): kh = kvp.acquire_and_advance(pk) - cute.copy(tma_k, tBgK[(None, kh.count)], tBsK[(None, kh.index)], tma_bar_ptr=kh.barrier) + cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier) pk = cutlass.Boolean(1) vh = kvp.acquire_and_advance(pk) - cute.copy(tma_v, tVgV[(None, vh.count)], tVsV[(None, vh.index)], tma_bar_ptr=vh.barrier) + cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier) pk = cutlass.Boolean(1) kvp.tail() - # ==================== MMA WARP (8) ==================== + # MMA 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(n_kv_tiles): - # QK -> S kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) - sh = mma_s_prod.acquire_and_advance() + sh = s_prod.acquire_and_advance() 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, kh.index)], tStS0) + 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,kh.index)], tStS0) qk_mma.set(tcgen05.Field.ACCUMULATE, True) - cute.arch.fence_view_async_tmem_store(); sh.commit(); kh.release() - # PV -> O + cute.arch.fence_view_async_tmem_store() + sh.commit(); kh.release() + softmax_done_bar.arrive_and_wait() vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) - oh = mma_corr_prod.acquire_and_advance() pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0) - for kb in cutlass.range(cute.size(tOrP0, mode=[2]), unroll_full=True): - cute.gemm(pv_mma, tOtO0, tOrP0[(None, None, kb)], tCrV[(None, None, kb, vh.index)], tOtO0) + for kb in cutlass.range(cute.size(tOrP0,mode=[2]), unroll_full=True): + cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0) pv_mma.set(tcgen05.Field.ACCUMULATE, True) - cute.arch.fence_view_async_tmem_store(); oh.commit(); vh.release() - mma_s_prod.tail(); mma_corr_prod.tail() - cute.arch.relinquish_tmem_alloc_permit() - cute.arch.mbarrier_wait(st.tmem_dealloc, 0) - tmem_ptr = cute.arch.retrieve_tmem_ptr(self.qk_acc_dtype, alignment=16, ptr_to_buffer_holding_addr=st.holding) - cute.arch.dealloc_tmem(tmem_ptr, Int32(self.num_tmem_alloc_cols)) + cute.arch.fence_view_async_tmem_store() + vh.release() + acc_pipe.producer_commit(acc_st); acc_st.advance() + acc_pipe.producer_tail(acc_st) - # ==================== SOFTMAX WARPS (0-3) ==================== - if warp_idx < len(self.softmax_warp_ids): - tmem.allocate(self.num_tmem_alloc_cols); tmem.wait_for_alloc() - sfw_idx = tidx % (32 * len(self.softmax_warp_ids)) - # S load setup + # SOFTMAX + EPILOGUE (warps 0-3) + # Step 1: Real online softmax: load S, compute row_max, exp2 scale, store P + # Step 2: After all KV tiles, normalize O in TMEM by row_sum + # Step 3: Epilogue writes normalized O from TMEM to GMEM + 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 (QK C-fragment layout) --- 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) tTMEM_LOADtS = thr_load.partition_S(tStS0) + cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1])) tScS = qk_thr.partition_C(cS) tTMEM_LOADcS = thr_load.partition_D(tScS) - # P store setup (QK C-fragment layout composition, FMHA pattern) + + # --- P store (QK C-fragment layout composition, FMHA pattern) --- p_cols_fp32 = self.pv_mma_tiler[2] * 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) @@ -227,48 +232,38 @@ class FmhaV3StageC: 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))) - tTMEM_STOREcP = thr_store.partition_S(cute.make_tensor(tScS.iterator, tScP_layout)) - # Vec store setup - tStS_vec_layout = cute.composition(tStS.layout, cute.make_layout((128, 2))) - tStS_vec = cute.make_tensor(tStS.iterator + self.tmem_vec0_offset, tStS_vec_layout) - tmem_store_vec_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(2)), self.qk_acc_dtype) - tiled_tmem_store_vec = tcgen05.make_tmem_copy(tmem_store_vec_atom, tStS_vec) - thr_store_vec = tiled_tmem_store_vec.get_slice(sfw_idx) - tTMEM_STORE_VECtS = thr_store_vec.partition_D(tStS_vec) - tScS_vec_layout = cute.composition(tScS.layout, cute.make_layout((128, 2))) - tScS_vec = cute.make_tensor(tScS.iterator, tScS_vec_layout) - tTMEM_STORE_VECcS = thr_store_vec.partition_S(tScS_vec) + tScP = cute.make_tensor(tScS.iterator, tScP_layout) + tTMEM_STOREcP = thr_store.partition_S(tScP) + # --- Online softmax loop --- row_max = -Float32.inf; row_sum = Float32(0.0) - vec_handle = s_corr_prod.acquire_and_advance() scale_log2 = Float32(self.scale_softmax_log2) for kt in range(n_kv_tiles): - si_handle = mma_s_cons.wait_and_advance() + si_handle = s_cons.wait_and_advance() + # Load S from TMEM 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() - # Row max + + # Update row_max old_row_max = row_max row_max = tTMEM_LOADrS.load().reduce(cute.ReductionOp.MAX, row_max, 0) row_max_safe = row_max if row_max == -cutlass.Float32.inf: row_max_safe = Float32(0.0) - # Vec = [old_max, new_max] - tTMEM_STORE_VECrS = cute.make_rmem_tensor(tTMEM_STORE_VECcS.shape, self.qk_acc_dtype) - tTMEM_STORE_VECrS[0] = old_row_max; tTMEM_STORE_VECrS[1] = row_max_safe - cute.copy(tiled_tmem_store_vec, tTMEM_STORE_VECrS, tTMEM_STORE_VECtS) - cute.arch.fence_view_async_tmem_store() - vec_handle.commit() - # P = exp2((S - new_max) * scale_log2) via register bridge - 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) - minus_row_max_scale = (Float32(0.0) - row_max_safe) * scale_log2 - # Scale existing row_sum + + # Scale existing row_sum: row_sum *= exp2((old_max - new_max) * scale_log2) acc_scale_ = scale_log2 * (old_row_max - row_max_safe) acc_scale = cute.math.exp2(acc_scale_, fastmath=True) if old_row_max == -cutlass.Float32.inf: acc_scale = Float32(0.0) row_sum *= acc_scale + + # P = exp2((S - new_max) * scale_log2) via register bridge + 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) + minus_row_max_scale = (Float32(0.0) - row_max_safe) * scale_log2 + frg_cnt = 4 frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile)) @@ -279,39 +274,19 @@ class FmhaV3StageC: tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True) s_vec = tTMEM_LOADrS_frg[None, j].load() rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype)) + # Accumulate row_sum from P values for k in cutlass.range(cute.size(tTMEM_LOADrS_frg, mode=[0])): row_sum = row_sum + tTMEM_LOADrS_frg[k, j] + cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP) cute.arch.fence_view_async_tmem_store() si_handle.release() - vec_handle = s_corr_prod.acquire_and_advance() + softmax_done_bar.arrive() - # Final vec = [row_sum, row_max] for correction epilog - tTMEM_STORE_VECrS = cute.make_rmem_tensor(tTMEM_STORE_VECcS.shape, self.qk_acc_dtype) - tTMEM_STORE_VECrS[0] = row_sum; tTMEM_STORE_VECrS[1] = row_max - cute.copy(tiled_tmem_store_vec, tTMEM_STORE_VECrS, tTMEM_STORE_VECtS) - cute.arch.fence_view_async_tmem_store() - vec_handle.commit() - s_corr_prod.acquire() # balance final pipe step - s_corr_prod.tail() - cute.arch.mbarrier_arrive(st.tmem_dealloc) - tmem.relinquish_alloc_permit() - - # ==================== CORRECTION WARPS (4-7) ==================== - if warp_idx >= len(self.softmax_warp_ids) and warp_idx < len(self.softmax_warp_ids) + len(self.correction_warp_ids): - tmem.wait_for_alloc() - corr_idx = tidx % (32 * len(self.correction_warp_ids)) - # Vec load - tStS_vec_layout = cute.composition(tStS.layout, cute.make_layout((128, 2))) - tStS_vec = cute.make_tensor(tStS.iterator + self.tmem_vec0_offset, tStS_vec_layout) - tmem_load_vec_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(2)), self.qk_acc_dtype) - tiled_tmem_load_vec = tcgen05.make_tmem_copy(tmem_load_vec_atom, tStS_vec) - thr_load_vec = tiled_tmem_load_vec.get_slice(corr_idx) - tTMEM_LOAD_VECtS = thr_load_vec.partition_S(tStS_vec) - tScS_vec_layout = cute.composition(tScS.layout, cute.make_layout((128, 2))) - tScS_vec = cute.make_tensor(tScS.iterator, tScS_vec_layout) - tTMEM_LOAD_VECcS = thr_load_vec.partition_D(tScS_vec) - # O load/store for correction_rescale (matching CUTLASS pattern) + # --- Normalize O in TMEM by row_sum --- + # O is at tmem_o0_offset in TMEM. Load each element, divide by row_sum, store back. + # Use the O TMEM layout (pv_thr C-fragment) for the load/store. + # We need a tiled TMEM copy for O. cO = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1])) tOcO = pv_thr.partition_C(cO) corr_tile_size = 16 @@ -319,98 +294,38 @@ class FmhaV3StageC: tOcO_i_layout = cute.composition(tOcO.layout, cute.make_layout((128, corr_tile_size))) tOtO_i = cute.make_tensor(tOtO.iterator, tOtO_i_layout) tOcO_i = cute.make_tensor(tOcO.iterator, tOcO_i_layout) - tmem_load_o_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.pv_acc_dtype) - tmem_store_o_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.pv_acc_dtype) + tmem_load_o_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype) + tmem_store_o_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.acc_dtype) tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_o_atom, tOtO_i) tiled_tmem_store_o = tcgen05.make_tmem_copy(tmem_store_o_atom, tOtO_i) - thr_load_o = tiled_tmem_load_o.get_slice(corr_idx) - thr_store_o = tiled_tmem_store_o.get_slice(corr_idx) + thr_load_o = tiled_tmem_load_o.get_slice(sfw_idx) + thr_store_o = tiled_tmem_store_o.get_slice(sfw_idx) tTMEM_LOAD_OtO = thr_load_o.partition_S(tOtO_i) tTMEM_LOAD_OcO = thr_load_o.partition_D(tOcO_i) tTMEM_STORE_OtO = thr_store_o.partition_D(tOtO_i) - scale_log2 = Float32(self.scale_softmax_log2) - - # First vec has no previous O to rescale - first_vec = s_corr_cons.wait_and_advance(); first_vec.release() - for kt in range(n_kv_tiles - 1): - vec = s_corr_cons.wait_and_advance() - # Read vec = [old_max, new_max] - tTMEM_LOAD_VECrS = cute.make_rmem_tensor(tTMEM_LOAD_VECcS.shape, self.qk_acc_dtype) - cute.copy(tiled_tmem_load_vec, tTMEM_LOAD_VECtS, tTMEM_LOAD_VECrS) - cute.arch.fence_view_async_tmem_load() - old_max = tTMEM_LOAD_VECrS[0]; new_max = tTMEM_LOAD_VECrS[1] - # scale = exp2((old_max - new_max) * scale_log2) - corr_scale = cute.math.exp2(scale_log2 * (old_max - new_max), fastmath=True) - # Wait for O from MMA, rescale O in TMEM - o_handle = mma_corr_cons.wait_and_advance() - o_col_tiles = self.pv_mma_tiler[1] // corr_tile_size - for i in range(o_col_tiles): - tTMEM_LOAD_O_i = cute.make_tensor(tTMEM_LOAD_OtO.iterator + i * corr_tile_size, tTMEM_LOAD_OtO.layout) - tTMEM_STORE_O_i = cute.make_tensor(tTMEM_STORE_OtO.iterator + i * corr_tile_size, tTMEM_STORE_OtO.layout) - tTMrO_i_ = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.pv_acc_dtype) - tTMrO_i_layout = cute.composition(tTMrO_i_.layout, cute.make_layout(tTMEM_LOAD_OcO.shape[0])) - tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout) - cute.copy(tiled_tmem_load_o, tTMEM_LOAD_O_i, tTMrO_i) - for k in cutlass.range(cute.size(tTMrO_i), vectorize=True): - tTMrO_i[k] = tTMrO_i[k] * corr_scale - cute.copy(tiled_tmem_store_o, tTMrO_i, tTMEM_STORE_O_i) - cute.arch.fence_view_async_tmem_store() - o_handle.release(); vec.release() - - # Final: read [row_sum, row_max], normalize O, write to SMEM - final_vec = s_corr_cons.wait_and_advance() - tTMEM_LOAD_VECrS = cute.make_rmem_tensor(tTMEM_LOAD_VECcS.shape, self.qk_acc_dtype) - cute.copy(tiled_tmem_load_vec, tTMEM_LOAD_VECtS, tTMEM_LOAD_VECrS) - cute.arch.fence_view_async_tmem_load() - row_sum = tTMEM_LOAD_VECrS[0]; row_max = tTMEM_LOAD_VECrS[1] - final_vec.release() - - final_o = mma_corr_cons.wait_and_advance() - epi_handle = corr_epi_prod.acquire_and_advance() - - # Correction epilog: load O from TMEM, normalize, convert to BF16, write SMEM - # Following CUTLASS correction_epilog pattern - corr_tile_size_epi = 32 * 8 // self.o_dtype.width - tOsO = pv_thr.partition_C(sC) - tOcO_epi = pv_thr.partition_C(cO) - tOtO_i_epi = cute.logical_divide(tOtO, cute.make_layout((128, corr_tile_size_epi))) - tOcO_i_epi = cute.logical_divide(tOcO_epi, cute.make_layout((128, corr_tile_size_epi))) - tOsO_i = cute.logical_divide(tOsO, cute.make_layout((128, corr_tile_size_epi))) - - epi_subtile = (self.epi_tile[0], corr_tile_size_epi) - tmem_copy_atom = utils.sm100.get_tmem_load_op(self.pv_mma_tiler, self.c_layout, self.o_dtype, self.pv_acc_dtype, epi_subtile, use_2cta_instrs=False) - tiled_tmem_load_epi = tcgen05.make_tmem_copy(tmem_copy_atom, tOtO_i_epi[(None, None), 0]) - thr_tmem_load_epi = tiled_tmem_load_epi.get_slice(corr_idx) - smem_copy_atom = utils.sm100.get_smem_store_op(self.c_layout, self.o_dtype, self.pv_acc_dtype, tiled_tmem_load_epi) - tiled_smem_store = cute.make_tiled_copy_D(smem_copy_atom, tiled_tmem_load_epi) - - tTMEM_LOAD_EPItO = thr_tmem_load_epi.partition_S(tOtO_i_epi[(None, None), None]) - tTMEM_LOAD_EPIdS = thr_tmem_load_epi.partition_D(tOsO_i[(None, None), None]) - tTMEM_LOAD_EPIdO = thr_tmem_load_epi.partition_D(tOcO_i_epi[(None, None), None]) - inv_row_sum = Float32(1.0) / row_sum - for i in range(self.pv_mma_tiler[1] // corr_tile_size_epi): - tTMrO = cute.make_rmem_tensor(tTMEM_LOAD_EPIdO[None, 0, 0, i].shape, self.pv_acc_dtype) - cute.copy(tiled_tmem_load_epi, tTMEM_LOAD_EPItO[None, 0, 0, i], tTMrO) - for k in cutlass.range(cute.size(tTMrO), vectorize=True): - tTMrO[k] = tTMrO[k] * inv_row_sum - tSMrO = cute.make_rmem_tensor(tTMrO.shape, self.o_dtype) - tSMrO.store(tTMrO.load().to(self.o_dtype)) - cute.copy(tiled_smem_store, tSMrO, tTMEM_LOAD_EPIdS[None, 0, 0, i]) + o_col_tiles = self.pv_mma_tiler[1] // corr_tile_size + for i in range(o_col_tiles): + tTMEM_LOAD_O_i = cute.make_tensor(tTMEM_LOAD_OtO.iterator + i * corr_tile_size, tTMEM_LOAD_OtO.layout) + tTMEM_STORE_O_i = cute.make_tensor(tTMEM_STORE_OtO.iterator + i * corr_tile_size, tTMEM_STORE_OtO.layout) + tTMrO_i_ = cute.make_rmem_tensor(tTMEM_LOAD_OcO.shape, self.acc_dtype) + tTMrO_i_layout = cute.composition(tTMrO_i_.layout, cute.make_layout(tTMEM_LOAD_OcO.shape[0])) + tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout) + cute.copy(tiled_tmem_load_o, tTMEM_LOAD_O_i, tTMrO_i) + for k in cutlass.range(cute.size(tTMrO_i), vectorize=True): + tTMrO_i[k] = tTMrO_i[k] * inv_row_sum + cute.copy(tiled_tmem_store_o, tTMrO_i, tTMEM_STORE_O_i) + cute.arch.fence_view_async_tmem_store() - cute.arch.fence_proxy("async.shared", space="cta") - final_o.release() - epi_handle.commit() - cute.arch.mbarrier_arrive(st.tmem_dealloc) - - # ==================== EPILOGUE WARP (10) ==================== - if warp_idx == self.epilogue_warp_id: - epi_handle = corr_epi_cons.wait_and_advance() - # TMA store O from SMEM to GMEM - cute.copy(tma_c, sC, tCgC[(None, 0)]) - cute.arch.cp_async_bulk_commit_group() - cute.arch.cp_async_bulk_wait_group(0, read=True) - epi_handle.release() + # --- Epilogue: write O from TMEM to GMEM --- + 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)) + c_pipe = pipeline.PipelineTmaStore.create(num_stages=self.num_c_stage, producer_group=c_grp) + acc_cons_st = utils.gemm.sm100.epilogue_tma_store(self, tidx, warp_idx, tma_c, tCtO_base, sC, tCgC, epi_tile, 0, const_expr(lambda x: x), (0,0,0), acc_cons_st, acc_pipe, c_pipe) + c_pipe.producer_tail() + tmem.relinquish_alloc_permit() + tmem.free(tmem_ptr) def test(): @@ -422,7 +337,7 @@ def test(): v = torch.randn(n, hd, dtype=torch.bfloat16, device='cuda') v_kernel = v.unsqueeze(-1) c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda') - # Reference: softmax(Q @ K^T) @ V + # Reference: softmax(Q @ K^T / sqrt(d)) @ V qf = q[:,:,0].float(); kf = k[:,:,0].float() scale = 1.0 / math.sqrt(hd) attn = qf @ kf.T * scale @@ -433,10 +348,10 @@ def test(): mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel)) mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c)) stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream) - kernel = FmhaV3StageC(s_k=n) + kernel = FmhaV3StageC() print(f'n={n}: Compiling...', flush=True) compiled = cute.compile(kernel, mQ, mK, mV, mC, stream) - print(f'n={n}: tmem_offsets: s0={kernel.tmem_s0_offset} vec0={kernel.tmem_vec0_offset} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} alloc={kernel.num_tmem_alloc_cols}', flush=True) + print(f'n={n}: tmem_offsets: s0={kernel.tmem_s0_offset} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} alloc={kernel.num_tmem_alloc_cols}', flush=True) print(f'n={n}: Running...', flush=True) compiled(mQ, mK, mV, mC, stream) torch.cuda.synchronize()