nvfp4-megamoe-kernel/tests/unit/test_qk_softmax.py

"""
Debug: QK + identity softmax, output P (BF16) to GMEM.
Tests the full QK -> softmax -> TMEM pipeline without PV.
Uses the QK C-fragment store (like FMHA).
Output = S.bf16() which is (Q@K^T).bfloat16(), shape (128, 128).
"""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32, Boolean, const_expr
from cutlass.utils import LayoutEnum
from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset
import cuda.bindings.driver as cuda
import cutlass.torch as ct

HEAD_DIM = 64

class QkSoftmaxTest:
    def __init__(self):
        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.threads_per_cta = 192; self.num_c_stage = 2
        self.kv_stage = 2; self.q_stage = 1

    def _setup(self, qk_mma):
        qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
        self.qk_mma_tiler = (128, 128, qk_ik * 4)
        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), 128, 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.num_ab_stage = 1; self.num_acc_stage = 1
        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.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2)
        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)
        self.tilePlikeFP32 = self.qk_mma_tiler[1] // Float32.width * self.o_dtype.width
        self.tmem_s0_offset = 0; self.tmem_p0_offset = 32; self.tmem_o0_offset = 0
        tCS = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
        self.num_tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCS], arch="sm_100")
        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
        self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta

    @cute.jit
    def __call__(self, q, k, c, stream):
        self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
        self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
        self.b_major = LayoutEnum.from_tensor(k).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)
        self._setup(qk_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))
        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)
        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,tma_q,mQ,tma_k,mK,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_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, tma_q, mQ, tma_k, mK, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, c_smem_s, epi_tile):
        warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_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_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]
            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)

        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()

        # P-ready: softmax -> MMA signal using NamedBarrier
        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)
        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))
        gC = cute.local_tile(mC,cute.slice_(self.qk_mma_tiler,(None,None,0)),(None,None,None))
        n_kv_tiles = cute.size(gK, mode=[3])

        qk_thr = qk_mma.get_slice(0)
        tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK); tCgC = qk_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))
        tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]

        tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK)

        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)

        tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_as, self.num_acc_stage))
        pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)

        # 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)
            qp.tail()
            kvp.reset(); pk = kvp.try_acquire()
            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)
                pk = cutlass.Boolean(1)
            kvp.tail()

        # MMA: QK, then wait for softmax, then signal done
        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):
                kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)

                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)
                    qk_mma.set(tcgen05.Field.ACCUMULATE, True)
                cute.arch.fence_view_async_tmem_store()
                sh.commit()
                kh.release()

                # Wait for softmax to finish
                softmax_done_bar.arrive_and_wait()

            # After all tiles: S contains the identity-softmax'd result
            # But we want to output P (written by softmax at p0 offset)
            # For this test, output what's at tmem_s0_offset (the final S accumulator)
            acc_pipe.producer_commit(acc_st); acc_st.advance()
            acc_pipe.producer_tail(acc_st)

        # EPILOGUE: identity softmax + output
        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
            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
            tStS_P_layout = cute.composition(tStS.layout, cute.make_layout((128, self.tilePlikeFP32)))
            tStS_P = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStS_P_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, tStS_P)
            thr_store = tiled_tmem_store.get_slice(sfw_idx)
            tTMEM_STOREtS_x4 = thr_store.partition_D(tStS_P)
            tScS_P_layout = cute.composition(tScS.layout, cute.make_layout((128, self.tilePlikeFP32)))
            tScS_P = cute.make_tensor(tScS.iterator, tScS_P_layout)
            tTMEM_STOREcS = thr_store.partition_S(tScS_P)

            for kt in range(n_kv_tiles):
                si_handle = s_cons.wait_and_advance()

                # Load S
                tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
                cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)

                # Identity softmax: FP32 S -> BF16 P, write to TMEM at p0 offset
                tTMEM_STORErS_x4 = cute.make_rmem_tensor(tTMEM_STOREcS.shape, self.qk_acc_dtype)
                tTMEM_STORErS_x4_e = cute.make_tensor(cute.recast_ptr(tTMEM_STORErS_x4.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout)

                frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
                tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
                tTMEM_STORErS_x4_e_frg = cute.logical_divide(tTMEM_STORErS_x4_e, cute.make_layout(frg_tile))
                for j in range(frg_cnt):
                    s_vec = tTMEM_LOADrS_frg[None, j].load()
                    tTMEM_STORErS_x4_e_frg[None, j].store(s_vec.to(self.q_dtype))
                cute.copy(tiled_tmem_store, tTMEM_STORErS_x4, tTMEM_STOREtS_x4)
                cute.arch.fence_view_async_tmem_store()

                si_handle.release()
                # Signal MMA
                softmax_done_bar.arrive()

            # Output: read from p0 offset (BF16 P values, but we read as FP32)
            # We need to output the BF16 P values. Read from p0 and convert.
            # Actually, output the S accumulator (at s0) for now to verify QK works
            tCtS_base = cute.make_tensor(tmem_ptr + self.tmem_s0_offset, tCtS_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, tCtS_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():
    torch.manual_seed(42)
    n = 128; m, hd = 128, HEAD_DIM
    q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
    k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device='cuda')
    c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
    qf = q[:,:,0].float(); kf = k[:,:,0].float()
    ref = (qf @ kf.T)
    mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
    mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
    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 = QkSoftmaxTest()
    print('Compiling...', flush=True)
    compiled = cute.compile(kernel, mQ, mK, mC, stream)
    print('Running...', flush=True)
    compiled(mQ, mK, mC, stream)
    torch.cuda.synchronize()
    out = c[:,:,0].float()
    cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
    print(f'QK+softmax n={n}: cosine {cos:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
    if cos < 0.99:
        print(f'  out[0,:4]={out[0,:4].tolist()} ref[0,:4]={ref[0,:4].tolist()}')

if __name__ == '__main__':
    test()