nvfp4-megamoe-kernel/tests/archive/test_pv_mma_mn_major.py

"""
Isolated test for Bug 1: PV MMA with V MN-major.

Only tests the PV MMA (P@V) with V as MN-major B-operand.
No QK MMA, no identity softmax, no pipeline complexity.
P comes from TMEM (a_source=TMEM), V comes from SMEM (b from TMA load).

Architecture:
  - TMA load V into SMEM
  - P pre-populated in TMEM (via small QK MMA or direct write)
  - PV MMA: P @ V → O in TMEM
  - Epilogue: TMEM → GMEM

For simplicity, P is computed via a QK MMA first (Q@K^T → P in TMEM),
then PV MMA uses P from TMEM. No softmax — identity pass-through.
"""
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


class PvMmaTest:
    def __init__(self, mma_tiler_mn, use_2cta_instrs=False, use_tma_store=True):
        self.acc_dtype = Float32; self.qk_acc_dtype = Float32
        self.q_dtype = BFloat16; self.o_dtype = BFloat16
        self.use_2cta_instrs = use_2cta_instrs; self.use_tma_store = use_tma_store
        self.mma_tiler_mn = mma_tiler_mn; self.mma_tiler = (*mma_tiler_mn, 1)
        self.cluster_shape_mn = (1, 1)
        self.cta_group = tcgen05.CtaGroup.TWO if use_2cta_instrs else tcgen05.CtaGroup.ONE
        self.mma_warp_id = 0
        self.tma_warp_id = 1
        self.threads_per_cta = 64
        self.num_c_stage = 2

    def _setup(self, qk_mma, pv_mma):
        qk_inst_k = cute.size(qk_mma.shape_mnk, mode=[2])
        self.qk_mma_tiler = (*self.mma_tiler_mn, qk_inst_k * 4)
        self.pv_mma_tiler = (self.qk_mma_tiler[0], self.qk_mma_tiler[2], self.qk_mma_tiler[1])
        self.mma_tiler = self.qk_mma_tiler
        print(f"[pv_test] qk_mma_tiler = {self.qk_mma_tiler}")
        print(f"[pv_test] pv_mma_tiler = {self.pv_mma_tiler}")

        self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))

        # Compute epilogue tile from PV output (not QK)
        cta_tile_shape_mnk = (
            self.qk_mma_tiler[0] // cute.size(qk_mma.thr_id.shape),
            self.qk_mma_tiler[1],
            self.qk_mma_tiler[2],
        )
        self.epi_tile = utils.sm100.compute_epilogue_tile_shape(
            cta_tile_shape_mnk, self.use_2cta_instrs, self.c_layout, self.o_dtype)

        self.num_ab_stage = 1; self.num_acc_stage = 1

        self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.a_dtype, 1)
        self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.b_dtype, 1)
        self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.b_dtype, 1)
        self.p_tmem_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)

        qk_thr = qk_mma.get_slice(0)
        qk_acc_shape = qk_thr.partition_shape_C(self.mma_tiler[:2])
        tStS = qk_thr.make_fragment_C(qk_acc_shape)
        s_cols = find_tmem_tensor_col_offset(tStS)

        pv_thr = pv_mma.get_slice(0)
        pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
        tOtO = pv_thr.make_fragment_C(pv_acc_shape)
        o_cols = find_tmem_tensor_col_offset(tOtO)

        self.tmem_s0_offset = 0
        self.tmem_p0_offset = 0  # P = S (identity softmax, same TMEM)
        self.tmem_o0_offset = s_cols

        tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
        tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))
        self.tmem_alloc_cols = utils.get_num_tmem_alloc_cols([tCtS_fake, tCtO_fake], arch="sm_100")

        a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
        b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
        self.num_tma_load_bytes = (
            cute.size_in_bytes(self.a_dtype, a_smem) + cute.size_in_bytes(self.b_dtype, b_smem)
        ) * cute.size(qk_mma.thr_id.shape)

    @cute.jit
    def __call__(self, q: cute.Tensor, k: cute.Tensor, v: cute.Tensor, c: cute.Tensor, stream: cuda.CUstream):
        self.a_dtype = q.element_type; self.b_dtype = k.element_type; self.c_dtype = c.element_type
        self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()
        self.b_major = LayoutEnum.from_tensor(k).mma_major_mode()
        self.v_major = LayoutEnum.from_tensor(v).mma_major_mode()
        self.c_layout = LayoutEnum.from_tensor(c)

        print(f"[pv_test] a_major (Q) = {self.a_major}")
        print(f"[pv_test] b_major (K) = {self.b_major}")
        print(f"[pv_test] v_major (V) = {self.v_major}")

        qk_mma = utils.sm100.make_trivial_tiled_mma(
            self.a_dtype, self.b_dtype, self.a_major, self.b_major,
            self.qk_acc_dtype, self.cta_group, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
        # BUG 1 FIX: PV MMA uses V's MN-major mode
        pv_mma = utils.sm100.make_trivial_tiled_mma(
            self.q_dtype, self.b_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major,
            self.qk_acc_dtype, self.cta_group, self.mma_tiler_mn, tcgen05.OperandSource.TMEM)
        self._setup(qk_mma, pv_mma)

        q_smem = cute.slice_(self.a_smem_s, (None, None, None, 0))
        k_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
        v_smem = cute.slice_(self.v_smem_s, (None, None, None, 0))

        tma_q, tma_tq = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
        tma_k, tma_tk = 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_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
        tma_v, tma_tv = cute.nvgpu.make_tiled_tma_atom_B(
            utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, pv_mma.thr_id),
            v, v_smem, self.pv_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape)

        epi_smem = cute.select(self.c_smem_s, mode=[0, 1])
        tma_c, tma_tc = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(), c, epi_smem, self.epi_tile)

        self._kernel(qk_mma, pv_mma, tma_q, tma_tq, tma_k, tma_tk, tma_v, tma_tv,
            tma_c, tma_tc, self.cluster_layout_vmnk,
            self.a_smem_s, self.b_smem_s, self.v_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, a_smem_s, b_smem_s, v_smem_s, c_smem_s, epi_tile):
        warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
        tidx, _, _ = cute.arch.thread_idx()

        @cute.struct
        class SS:
            ab_bar: cute.struct.MemRange[cutlass.Int64, 2]
            acc_bar: cute.struct.MemRange[cutlass.Int64, 2]
            tmem_dealloc: cutlass.Int64
            holding: cutlass.Int32

        smem = utils.SmemAllocator(); st = smem.allocate(SS)

        ab_p, ab_c = pipeline.PipelineTmaUmma.create(
            barrier_storage=st.ab_bar.data_ptr(), num_stages=1,
            producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),
            consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32),
            tx_count=self.num_tmama_load_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True
        ).make_participants()

        acc_pipe = pipeline.PipelineUmmaAsync.create(
            barrier_storage=st.acc_bar.data_ptr(), num_stages=1,
            producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),
            consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32),
            cta_layout_vmnk=cl_vmnk, defer_sync=True)

        tmem_bar = pipeline.NamedBarrier(barrier_id=2, num_threads=64)
        tmem = utils.TmemAllocator(st.holding.ptr, barrier_for_retrieve=tmem_bar,
            allocator_warp_id=0, is_two_cta=False,
            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.a_dtype, layout=a_smem_s.outer, byte_alignment=128, swizzle=a_smem_s.inner)
        sK = smem.allocate_tensor(element_type=self.b_dtype, layout=b_smem_s.outer, byte_alignment=128, swizzle=b_smem_s.inner)
        sV = smem.allocate_tensor(element_type=self.b_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.mma_tiler, (None,0,None)), (None,None,None))
        gK = cute.local_tile(mK, cute.slice_(self.mma_tiler, (0,None,None)), (None,None,None))
        gC = cute.local_tile(mC, cute.slice_(self.mma_tiler, (None,None,0)), (None,None,None))
        k_cnt = cute.size(gQ, 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); 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)]

        gV = cute.local_tile(mV, cute.slice_(self.pv_mma_tiler, (0,None,None)), (None,None,None))
        tCgV = pv_thr.partition_B(gV)
        tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
        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)

        qk_acc_shape = qk_thr.partition_shape_C(self.mma_tiler[:2])
        tStS = qk_thr.make_fragment_C(qk_acc_shape)
        tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout)

        pv_acc_shape = pv_thr.partition_shape_C(self.pv_mma_tiler[:2])
        tOtO = pv_thr.make_fragment_C(pv_acc_shape)
        tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)

        # P from S TMEM — same location, MMA A-operand for PV
        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 = tOrP  # P is at same TMEM offset as S (identity softmax)

        tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, self.num_acc_stage))
        tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_acc_shape, self.num_acc_stage))

        pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)

        # WARP 1: TMA load
        if warp_idx == self.tma_warp_id:
            tmem.wait_for_alloc()
            ab_p.reset(); peek = ab_p.try_acquire()
            for kt in cutlass.range(k_cnt, unroll=1):
                h = ab_p.acquire_and_advance(peek)
                cute.copy(tma_q, tAgQ[(None,h.count)], tAsQ[(None,h.index)], tma_bar_ptr=h.barrier)
                cute.copy(tma_k, tBgK[(None,h.count)], tBsK[(None,h.index)], tma_bar_ptr=h.barrier)
                cute.copy(tma_v, tVgV[(None,h.count)], tVsV[(None,h.index)], tma_bar_ptr=h.barrier)
                peek = cutlass.Boolean(1)
                if h.count+1<k_cnt: peek = ab_p.try_acquire()
            ab_p.tail()

        # WARP 0: MMA (both QK and PV)
        if warp_idx == self.mma_warp_id:
            tmem.wait_for_alloc()
            ab_c.reset(); peek = ab_c.try_wait()

            acc_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
            acc_pipe.producer_acquire(acc_prod_st)

            # QK MMA: Q @ K^T → S in TMEM
            qk_mma.set(tcgen05.Field.ACCUMULATE, False)
            for kt in range(k_cnt):
                h = ab_c.wait_and_advance(peek)
                nblk = cute.size(tCrQ, mode=[2])
                for kb in cutlass.range(nblk, unroll_full=True):
                    cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,h.index)], tCrK[(None,None,kb,h.index)], tStS0)
                    qk_mma.set(tcgen05.Field.ACCUMULATE, True)
                h.release(); peek = cutlass.Boolean(1)
                if h.count+1<k_cnt: peek = ab_c.try_wait()

            cute.arch.fence_view_async_tmem_store()

            # PV MMA: P @ V → O in TMEM (identity softmax: P = S)
            pv_mma.set(tcgen05.Field.ACCUMULATE, True)
            tCrV_s = tCrV[(None, None, None, 0)]
            nblk_pv = cute.size(tOrP0, mode=[2])
            for kb in cutlass.range(nblk_pv, unroll_full=True):
                cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV_s[(None,None,kb)], tOtO0)

            acc_pipe.producer_commit(acc_prod_st)
            acc_prod_st.advance()
            acc_pipe.producer_tail(acc_prod_st)


def test():
    torch.manual_seed(42)
    m, n, head_dim = 128, 128, 64
    q = torch.randn(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')
    k = torch.randn(n, head_dim, 1, dtype=torch.bfloat16, device='cuda')
    # V: MN-major — (head_dim, seq) with strides (1, head_dim)
    v_base = torch.randn(head_dim, n, dtype=torch.bfloat16, device='cuda')
    v = v_base.as_strided((head_dim, n), (1, head_dim)).unsqueeze(-1)
    c = torch.zeros(m, head_dim, 1, dtype=torch.bfloat16, device='cuda')

    qf = q[:,:,0].float(); kf = k[:,:,0].float(); vf = v_base.float()
    # Q@K^T = (128,128), then P@V: (128,128) @ (64,128).T = (128,64)
    # Wait — with MN-major V, the MMA interprets V as (head_dim, seq) MN-major
    # which means the MMA computes P @ V (not P @ V^T)
    # So reference is: (Q @ K^T) @ V where V is (64, 128) row-major
    # But V has strides (1, 64), so V is NOT row-major.
    # The kernel sees V as MN-major B-operand, which means:
    # PV MMA computes: P[m, k] * V[n, k] -> O[m, n]
    # This is P @ V^T in matrix notation
    # So reference: Q@K^T @ V^T where V^T is (128, 64)
    ref = qf @ kf.T @ vf.T  # (128,128) @ (128,64) = (128,64)

    import cutlass.torch as ct
    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))
    mV = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
    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 = PvMmaTest(mma_tiler_mn=(128, 128), use_2cta_instrs=False, use_tma_store=True)
    print('Compiling...', flush=True)
    compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
    print('Running...', flush=True)
    compiled(mQ, mK, mV, mC, stream)
    torch.cuda.synchronize()
    out = c[:,:,0].float()
    cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
    max_err = (out - ref).abs().max().item()
    print('PV MMA test (V MN-major, no softmax):')
    print('  Cosine: {:.6f}, Max error: {:.6f}'.format(cos, max_err))
    print('  {}'.format('PASS' if cos >= 0.99 else 'FAIL'))

if __name__ == '__main__':
    test()