biondizzle
467ade37b2
Key finding: C-fragment and A-fragment use different physical TMEM address mappings. St32x32bOp with C-fragment writes to C-layout addresses, but PV MMA reads from A-layout addresses. Forward FMHA recast validated FP16 only, not BF16. Working: FP32 ld/st roundtrip, BF16 elemwise, BF16 recast ld S0->st S1 (all cos 0.999999) Broken: C-frag st + A-frag read (NaN), A-frag store + PV MMA (cos -0.02) Next: Fix register data flow (128 FP16/thread load vs 64 BF16/thread store mismatch)
176 lines
13 KiB
Python
176 lines
13 KiB
Python
"""Test: ld FP32 from S0, st BF16 to A-fragment layout tdVrP,
|
|
ld BF16 back from tdVrP, epi the result.
|
|
If this works, the A-fragment store is correct and the issue is in the PV MMA."""
|
|
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 AFragRoundtrip:
|
|
def __init__(self, mma_tiler_mn):
|
|
self.qk_acc_dtype = Float32; self.q_dtype = BFloat16; self.o_dtype = BFloat16
|
|
self.c_dtype = BFloat16; self.acc_dtype = Float32
|
|
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.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.use_2cta_instrs = False; self.epilog_sync_bar_id = 1
|
|
|
|
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)
|
|
pv_inst_k = cute.size(pv_mma.shape_mnk, mode=[2])
|
|
self.pv_mma_tiler = (*self.mma_tiler_mn, pv_inst_k * 4)
|
|
self.mma_tiler = self.qk_mma_tiler
|
|
self.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.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))
|
|
self.a_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.mma_tiler, self.q_dtype, 1)
|
|
self.b_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.mma_tiler, self.q_dtype, 1)
|
|
self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1)
|
|
c_layout = LayoutEnum.ROW_MAJOR; self.c_layout = c_layout
|
|
self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, c_layout, self.o_dtype)
|
|
self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, c_layout, self.epi_tile, 2)
|
|
self.num_ab_stage = 1; self.num_acc_stage = 1
|
|
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); self.s_cols = find_tmem_tensor_col_offset(tStS)
|
|
self.tmem_alloc_cols = self.s_cols # Only need S region
|
|
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.q_dtype, a_smem) + cute.size_in_bytes(self.q_dtype, b_smem)) * cute.size(qk_mma.thr_id.shape)
|
|
|
|
@cute.jit
|
|
def __call__(self, a: cute.Tensor, b: cute.Tensor, c: cute.Tensor, stream: cuda.CUstream):
|
|
qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, LayoutEnum.from_tensor(a).mma_major_mode(), LayoutEnum.from_tensor(b).mma_major_mode(), self.qk_acc_dtype, self.cta_group, self.mma_tiler_mn, tcgen05.OperandSource.SMEM)
|
|
pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, LayoutEnum.from_tensor(b).mma_major_mode(), self.qk_acc_dtype, self.cta_group, self.mma_tiler_mn, tcgen05.OperandSource.TMEM)
|
|
self._setup(qk_mma, pv_mma)
|
|
a_smem = cute.slice_(self.a_smem_s, (None, None, None, 0)); b_smem = cute.slice_(self.b_smem_s, (None, None, None, 0))
|
|
tma_a, tma_ta = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn, qk_mma.thr_id), a, a_smem, self.mma_tiler, qk_mma, self.cluster_layout_vmnk.shape)
|
|
tma_b, tma_tb = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn, qk_mma.thr_id), b, b_smem, self.mma_tiler, qk_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_a, tma_ta, tma_b, tma_tb, tma_c, tma_tc, self.cluster_layout_vmnk, self.a_smem_s, self.b_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_a, mA, tma_b, mB, tma_c, mC, cl_vmnk, a_smem_s, b_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()
|
|
if warp_idx == self.tma_warp_id: cpasync.prefetch_descriptor(tma_a); cpasync.prefetch_descriptor(tma_b); cpasync.prefetch_descriptor(tma_c)
|
|
@cute.struct
|
|
class SS:
|
|
ab_bar: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage * 2]; mma_si_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)
|
|
ab_p, ab_c = pipeline.PipelineTmaUmma.create(barrier_storage=st.ab_bar.data_ptr(), num_stages=self.num_ab_stage, producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread), consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 1), tx_count=self.num_tma_load_bytes, cta_layout_vmnk=cl_vmnk, defer_sync=True).make_participants()
|
|
mma_si_prod, mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_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)), cta_layout_vmnk=cl_vmnk, defer_sync=True).make_participants()
|
|
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=False, two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr)
|
|
pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk, is_relaxed=True)
|
|
sA = smem.allocate_tensor(element_type=self.q_dtype, layout=a_smem_s.outer, byte_alignment=128, swizzle=a_smem_s.inner)
|
|
sB = smem.allocate_tensor(element_type=self.q_dtype, layout=b_smem_s.outer, byte_alignment=128, swizzle=b_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)
|
|
gA = cute.local_tile(mA, cute.slice_(self.mma_tiler, (None,0,None)), (None,None,None))
|
|
gB = cute.local_tile(mB, 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(gA, mode=[3])
|
|
qk_thr = qk_mma.get_slice(0); tCgA = qk_thr.partition_A(gA); tCgB = qk_thr.partition_B(gB); tCgC = qk_thr.partition_C(gC)
|
|
a_lay = cute.make_layout(cute.slice_(cl_vmnk, (0,0,None,0)).shape)
|
|
tAsA, tAgA = cpasync.tma_partition(tma_a, 0, a_lay, cute.group_modes(sA,0,3), cute.group_modes(tCgA,0,3))
|
|
b_lay = cute.make_layout(cute.slice_(cl_vmnk, (0,None,0,0)).shape)
|
|
tBsB, tBgB = cpasync.tma_partition(tma_b, 0, b_lay, cute.group_modes(sB,0,3), cute.group_modes(tCgB,0,3))
|
|
tAgA = tAgA[(None,0,None,0)]; tBgB = tBgB[(None,0,None,0)]
|
|
tCrA = qk_mma.make_fragment_A(sA); tCrB = qk_mma.make_fragment_B(sB)
|
|
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, tStS.layout)
|
|
# A-fragment for pv_mma
|
|
pv_thr = pv_mma.get_slice(0)
|
|
tP_iter = cute.recast_ptr(tStS.iterator, dtype=self.q_dtype)
|
|
tP = cute.make_tensor(tP_iter, p_tmem_s.outer)
|
|
tOrP = pv_thr.make_fragment_A(tP)[None, None, None, 0]
|
|
tdVrP = cute.make_tensor(tOrP.iterator, tOrP.layout)
|
|
# TMEM ld (C-fragment)
|
|
tmem_ld = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype)
|
|
tiled_ld = tcgen05.make_tmem_copy(tmem_ld, tStS0)
|
|
sfw = tidx % (32 * len(self.epilogue_warp_id))
|
|
thr_ld = tiled_ld.get_slice(sfw)
|
|
tLdS = thr_ld.partition_S(tStS0)
|
|
cS_id = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1]))
|
|
tScS = qk_thr.partition_C(cS_id)
|
|
tLdcS = thr_ld.partition_D(tScS)
|
|
# TMEM st (A-fragment layout, BF16)
|
|
tmem_st = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(8)), self.q_dtype)
|
|
tiled_st = tcgen05.make_tmem_copy(tmem_st, tdVrP)
|
|
thr_st = tiled_st.get_slice(sfw)
|
|
tStP = thr_st.partition_D(tdVrP)
|
|
|
|
tCtS_fake = qk_mma.make_fragment_C(cute.append(qk_acc_shape, 1))
|
|
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
|
|
# TMA
|
|
if warp_idx == self.tma_warp_id:
|
|
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_a, tAgA[(None,h.count)], tAsA[(None,h.index)], tma_bar_ptr=h.barrier)
|
|
cute.copy(tma_b, tBgB[(None,h.count)], tBsB[(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()
|
|
# MMA
|
|
if warp_idx == self.mma_warp_id:
|
|
tmem.wait_for_alloc(); ab_c.reset(); peek = ab_c.try_wait()
|
|
s0_handle = mma_si_prod.acquire_and_advance()
|
|
acc_prod_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage)
|
|
acc_pipe.producer_acquire(acc_prod_st); qk_mma.set(tcgen05.Field.ACCUMULATE, False)
|
|
for kt in range(k_cnt):
|
|
h = ab_c.wait_and_advance(peek); nblk = cute.size(tCrA, mode=[2])
|
|
for kb in cutlass.range(nblk, unroll_full=True):
|
|
cute.gemm(qk_mma, tStS0, tCrA[(None,None,kb,h.index)], tCrB[(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(); s0_handle.commit()
|
|
acc_pipe.producer_commit(acc_prod_st); acc_prod_st.advance(); acc_pipe.producer_tail(acc_prod_st)
|
|
# EPILOGUE WARPS: ld FP32 → BF16 → st A-frag → ld A-frag BF16 → FP32 → epi
|
|
if warp_idx < self.mma_warp_id:
|
|
tmem.allocate(self.tmem_alloc_cols); tmem.wait_for_alloc(); tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype)
|
|
si_handle = mma_si_cons.wait_and_advance()
|
|
# 1. ld FP32 from S0
|
|
rLd = cute.make_rmem_tensor(tLdcS.shape, self.qk_acc_dtype); cute.copy(tiled_ld, tLdS, rLd)
|
|
cute.arch.fence_view_async_tmem_load()
|
|
# 2. Convert FP32 → BF16
|
|
rBf16 = cute.make_rmem_tensor(tLdcS.shape, self.q_dtype)
|
|
for i in cutlass.range(cute.size(rLd), vectorize=True):
|
|
rBf16[i] = rLd[i].to(self.q_dtype)
|
|
# 3. st BF16 to A-fragment layout
|
|
cute.copy(tiled_st, rBf16, tStP)
|
|
cute.arch.fence_view_async_tmem_store()
|
|
# 4. Store to A-frag done. Check if S0 epi still works.
|
|
si_handle.release()
|
|
tCtS0 = cute.make_tensor(tmem_ptr, 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, tCtS0, 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); m, n, k = 128, 128, 128
|
|
q = torch.randn(m, k, 1, dtype=torch.bfloat16, device='cuda')
|
|
kv = torch.randn(n, k, 1, dtype=torch.bfloat16, device='cuda')
|
|
c = torch.zeros(m, n, 1, dtype=torch.bfloat16, device='cuda')
|
|
ref = q[:,:,0].float() @ kv[:,:,0].float().T
|
|
import cutlass.torch as ct
|
|
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
|
|
mK = ct.from_dlpack(kv).mark_layout_dynamic(leading_dim=ct.get_leading_dim(kv))
|
|
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 = AFragRoundtrip(mma_tiler_mn=(128, 128))
|
|
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('A-frag roundtrip: cos={:.6f} (expect 0.999 from Stage A)'.format(cos))
|
|
|
|
if __name__ == '__main__':
|
|
test()
|