leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 26 Packages Projects Releases Wiki Activity

v4.5 dev update. (#3153)

Browse Source
This commit is contained in:
Junkai-Wu
2026-04-07 12:16:05 -04:00
committed by GitHub
parent 418d38a5de
commit a221da7ccf
265 changed files with 4913 additions and 1478 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/python/CuTeDSL/blackwell/blockwise_gemm/blockwise_gemm.py
+3 -7
View File
@@ -549,7 +549,7 @@ class BlockwiseGemmKernel:
cutlass.Int64, self.num_tile_stage * 2
]
epi_mbar_ptr: cute.struct.MemRange[cutlass.Int64, 1 * 2]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
# (EPI_TILE_M, EPI_TILE_N, STAGE)
sC: cute.struct.Align[
@@ -614,7 +614,6 @@ class BlockwiseGemmKernel:
grid=grid,
block=[self.threads_per_cta, 1, 1],
cluster=(*self.cluster_shape_mn, 1),
smem=self.shared_storage.size_in_bytes(),
stream=stream,
min_blocks_per_mp=1,
)
@@ -682,9 +681,6 @@ class BlockwiseGemmKernel:
smem = utils.SmemAllocator()
storage = smem.allocate(self.shared_storage)
tmem_dealloc_mbar_ptr = storage.tmem_dealloc_mbar_ptr
tmem_holding_buf = storage.tmem_holding_buf
# Initialize mainloop ab_pipeline (barrier) and states
ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
num_tma_producer = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
@@ -771,11 +767,11 @@ class BlockwiseGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_alloc_barrier,
allocator_warp_id=self.epilog_warp_id[0],
is_two_cta=use_2cta_instrs,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar.ptr,
)
# Cluster arrive after barrier init
examples/python/CuTeDSL/blackwell/blockwise_gemm/contiguous_grouped_gemm.py
+3 -7
View File
@@ -568,7 +568,7 @@ class BlockwiseContiguousGroupedGemmKernel:
cutlass.Int64, self.num_tile_stage * 2
]
epi_mbar_ptr: cute.struct.MemRange[cutlass.Int64, 1 * 2]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
# (EPI_TILE_M, EPI_TILE_N, STAGE)
sC: cute.struct.Align[
@@ -634,7 +634,6 @@ class BlockwiseContiguousGroupedGemmKernel:
grid=grid,
block=[self.threads_per_cta, 1, 1],
cluster=(*self.cluster_shape_mn, 1),
smem=self.shared_storage.size_in_bytes(),
stream=stream,
min_blocks_per_mp=1,
)
@@ -703,9 +702,6 @@ class BlockwiseContiguousGroupedGemmKernel:
smem = utils.SmemAllocator()
storage = smem.allocate(self.shared_storage)
tmem_dealloc_mbar_ptr = storage.tmem_dealloc_mbar_ptr
tmem_holding_buf = storage.tmem_holding_buf
# Initialize mainloop ab_pipeline (barrier) and states
ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
num_tma_producer = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
@@ -792,11 +788,11 @@ class BlockwiseContiguousGroupedGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_alloc_barrier,
allocator_warp_id=self.epilog_warp_id[0],
is_two_cta=use_2cta_instrs,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar.ptr,
)
# Cluster arrive after barrier init
examples/python/CuTeDSL/blackwell/blockwise_gemm/masked_grouped_gemm.py
+3 -7
View File
@@ -567,7 +567,7 @@ class BlockwiseMaskedGroupedGemmKernel:
cutlass.Int64, self.num_tile_stage * 2
]
epi_mbar_ptr: cute.struct.MemRange[cutlass.Int64, 1 * 2]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
# (EPI_TILE_M, EPI_TILE_N, STAGE)
sC: cute.struct.Align[
@@ -633,7 +633,6 @@ class BlockwiseMaskedGroupedGemmKernel:
grid=grid,
block=[self.threads_per_cta, 1, 1],
cluster=(*self.cluster_shape_mn, 1),
smem=self.shared_storage.size_in_bytes(),
stream=stream,
min_blocks_per_mp=1,
)
@@ -702,9 +701,6 @@ class BlockwiseMaskedGroupedGemmKernel:
smem = utils.SmemAllocator()
storage = smem.allocate(self.shared_storage)
tmem_dealloc_mbar_ptr = storage.tmem_dealloc_mbar_ptr
tmem_holding_buf = storage.tmem_holding_buf
# Initialize mainloop ab_pipeline (barrier) and states
ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
num_tma_producer = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
@@ -791,11 +787,11 @@ class BlockwiseMaskedGroupedGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_alloc_barrier,
allocator_warp_id=self.epilog_warp_id[0],
is_two_cta=use_2cta_instrs,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar.ptr,
)
# Cluster arrive after barrier init