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

CUTLASS 2.5

Browse Source
This commit is contained in:
Andrew Kerr
2021-02-26 09:58:26 -05:00
parent ccb697bac7
commit 0e13748649
771 changed files with 15476 additions and 1717 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
include/cutlass/transform/threadblock/predicated_tile_access_iterator.h
View File

@@ -1,40 +1,39 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
*modification, are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice,
*this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
*notice, this list of conditions and the following disclaimer in the
*documentation and/or other materials provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its
*contributors may be used to endorse or promote products derived from this
*software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
*AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
*IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
*DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY DIRECT,
*INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
*DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
*OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TOR (INCLUDING
*NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
*EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this list of
* conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright notice, this list of
* conditions and the following disclaimer in the documentation and/or other materials
* provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
* to endorse or promote products derived from this software without specific prior written
* permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
* FOR ANY DIRECT,INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*! \file
\brief Templates calculating the address and predicates to the load of tiles
from pitch-linear rank=2 tensors.
from pitch-linear rank=2 tensors.
This iterator uses masks to guard out-of-bounds accesses and visits the last
"residue" tile first, with the objective of minimizing predicate mask updates
during steady-state operation.
This iterator uses masks to guard out-of-bounds accesses. The first tile this
iterator visits maybe partial, then the remaining tiles are complete. So, we
only need to compute the predicates twice, once before the first tile and
once for the remaining full tiles which can share the same predicates.
A precomputed "Params" object minimizes the amount of state that must be
stored in registers, and integer addition is used to advance the pointer
through memory.
stored in registers, and integer addition is used to advance the pointer
through memory.
*/
#pragma once

44
include/cutlass/transform/threadblock/predicated_tile_access_iterator_2dthreadtile.h
View File

@@ -1,27 +1,25 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
*modification, are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice,
*this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
*notice, this list of conditions and the following disclaimer in the
*documentation and/or other materials provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its
*contributors may be used to endorse or promote products derived from this
*software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
*AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
*IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
*DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY DIRECT,
*INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
*DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
*OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TOR (INCLUDING
*NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
*EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this list of
* conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright notice, this list of
* conditions and the following disclaimer in the documentation and/or other materials
* provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
* to endorse or promote products derived from this software without specific prior written
* permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
* FOR ANY DIRECT,INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*! \file

8
include/cutlass/transform/threadblock/predicated_tile_iterator.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
@@ -25,8 +25,10 @@
/*! \file
\brief Templates implementing loading of tiles from pitch-linear rank=2 tensors.
This iterator uses masks to guard out-of-bounds accesses and visits the last "residue" tile
first, with the objective of minimizing predicate mask updates during steady-state operation.
This iterator uses masks to guard out-of-bounds accesses. The first tile this
iterator visits maybe partial, then the remaining tiles are complete. So, we
only need to compute the predicates twice, once before the first tile and
once for the remaining full tiles which can share the same predicates.
A precomputed "Params" object minimizes the amount of state that must be stored in registers,
and integer addition is used to advance the pointer through memory.

2
include/cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

2
include/cutlass/transform/threadblock/regular_tile_access_iterator.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
*modification, are permitted provided that the following conditions are met:

12
include/cutlass/transform/threadblock/regular_tile_access_iterator_pitch_linear.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
@@ -182,7 +182,15 @@ class RegularTileAccessIterator<
return prev;
}
/// Adds a tile offset
/// Adds a tile offset in the unit of tile.
/// In GEMM/Conv implementation, this is used to move in the k dimension in the shared memory.
/// Below layouts are the shared memory layouts. Current SM50 SIMT kernels only use col major A and row major B.
/// For row major A operand, k dimension is contiguous dimension;
/// For col major A operand, k dimension is strided dimension;
/// For row major B operand, k dimension is strided dimension;
/// For col major B operand, k dimension is contiguous dimension.
/// Below two classes map col/row major to the pitch linear coordinates used
/// in this base class.
CUTLASS_DEVICE
void add_tile_offset(TensorCoord const &coord) {
add_pointer_offset(coord.contiguous() * Shape::kContiguous +

2
include/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

2
include/cutlass/transform/threadblock/regular_tile_access_iterator_tensor_op_sm80.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

2
include/cutlass/transform/threadblock/regular_tile_iterator.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

13
include/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
@@ -227,14 +227,21 @@ public:
pointer_ += pointer_offset;
}
/// Adds a tile offset
/// Adds a tile offset in the unit of tile.
/// In GEMM/Conv implementation, this is used to move in the k dimension in the shared memory.
/// Below layouts are the shared memory layouts. Current SM50 SIMT kernels only use col major A and row major B.
/// For row major A operand, k dimension is contiguous dimension;
/// For col major A operand, k dimension is strided dimension;
/// For row major B operand, k dimension is strided dimension;
/// For col major B operand, k dimension is contiguous dimension.
/// Below two classes map col/row major to the pitch linear coordinates used
/// in this base class.
CUTLASS_DEVICE
void add_tile_offset(TensorCoord const &coord) {
int offset = sizeof_bits<Element>::value *
(coord.contiguous() * Shape::kContiguous + coord.strided() * Shape::kStrided * stride_) / 8;
add_pointer_offset(offset);
}
};
/////////////////////////////////////////////////////////////////////////////////////////////////

2
include/cutlass/transform/threadblock/regular_tile_iterator_pitch_linear_2dthreadtile.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

2
include/cutlass/transform/threadblock/regular_tile_iterator_tensor_op.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:

2
include/cutlass/transform/threadblock/regular_tile_iterator_tensor_op_sm70.h
View File

@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met: