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

CUTLASS 3.1 (#915)

Browse Source 
Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
This commit is contained in:
ANIKET SHIVAM
2023-04-14 20:19:34 -07:00
committed by GitHub
parent 9b8166e3f0
commit d572cc1aab
482 changed files with 37184 additions and 16419 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
test/python/backend/gemm/__init__.py Normal file
View File

128
test/python/backend/gemm/gemm_bf16_sm80.py Normal file
View File

@ -0,0 +1,128 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmBF16TensorOpSm80(unittest.TestCase):
def SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32_64x128x64_32x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.bfloat16, element_b=cutlass_bindings.bfloat16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[64, 128, 64],
stages=4, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.bfloat16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.bfloat16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, cutlass_bindings.float32)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32_128x256x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.bfloat16, element_b=cutlass_bindings.bfloat16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[64, 128, 32],
stages=6, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.bfloat16, layout=cutlass_bindings.RowMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.bfloat16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.bfloat16, layout=cutlass_bindings.RowMajor,
alignment=8
)
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, cutlass_bindings.float32)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "multistage"))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

138
test/python/backend/gemm/gemm_bf16_sm90.py Normal file
View File

@ -0,0 +1,138 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
from functools import partial
import cutlass.backend
from cutlass.backend import *
from cutlass.backend import library
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
name_fn = partial(get_name, element_a=cutlass_bindings.bfloat16, element_b=cutlass_bindings.bfloat16, arch=90)
def add_test(cls, layouts, alignments, element_output, element_accumulator, element_epilogue,
cluster_shape, threadblock_shape, stages, opclass, persistent=False):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
:param cluster_shape: indexable container of dimensions of threadblock cluster to be launched
:param threadblock_shape: indexable container of dimensions of threadblock tiles
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass_bindings.OpClass
:param persistent: whether this is a persistent warp-specialized kernel
:type persistent: bool
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass_bindings.bfloat16
element_B = cutlass_bindings.bfloat16
inst_shape = [1, 1, 1] if opclass == cutlass_bindings.OpClass.Simt else None
warp_count = [2, 2, 1] if opclass == cutlass_bindings.OpClass.Simt else None
math_inst = MathInstruction(
instruction_shape=inst_shape,
element_a=element_A, element_b=element_B, element_accumulator=element_accumulator,
opcode_class=opclass, math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=threadblock_shape,
cluster_shape=cluster_shape,
stages=stages, warp_count=warp_count,
math_instruction=math_inst,
persistent=persistent
)
A = TensorDescription(element=element_A, layout=layouts[0], alignment=alignments[0])
B = TensorDescription(element=element_B, layout=layouts[1], alignment=alignments[1])
C = TensorDescription(element=element_output, layout=layouts[2], alignment=alignments[2])
epilogue_functor = LinearCombination(C.element, C.alignment, math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=90, tile_description=tile_description, A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor)
self.assertTrue(test_all_gemm(operation, "universal"))
if persistent:
suffix = "_persistent"
else:
suffix = ""
name = name_fn(layouts, alignments, element_output, element_accumulator,
element_epilogue, cluster_shape, threadblock_shape, stages, opclass=opclass, suffix=suffix)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < 90, "Device compute capability is insufficient for SM90 tests.")
class GemmBF16Sm90(unittest.TestCase):
"""
Wrapper class to which tests will be added dynamically in __main__
"""
pass
add_test_tensorop = partial(add_test, opclass=cutlass_bindings.OpClass.TensorOp)
add_test_simt = partial(add_test, opclass=cutlass_bindings.OpClass.Simt)
add_test_tensorop(GemmBF16Sm90, LayoutCombination.NNN, [8, 8, 8], cutlass_bindings.bfloat16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], 3)
add_test_tensorop(GemmBF16Sm90, LayoutCombination.NNN, [4, 4, 8], cutlass_bindings.bfloat16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], 5)
add_test_tensorop(GemmBF16Sm90, LayoutCombination.TNN, [8, 8, 8], cutlass_bindings.bfloat16, cutlass_bindings.float32, cutlass_bindings.float32, [2, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmBF16Sm90, LayoutCombination.TNN, [8, 8, 8], cutlass_bindings.bfloat16, cutlass_bindings.float32, cutlass_bindings.float32, [2, 1, 1], [128, 128, 32], None, persistent=True)
add_test_simt(GemmBF16Sm90, LayoutCombination.NNN, [1, 1, 1], cutlass_bindings.bfloat16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 8], 2)
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

479
test/python/backend/gemm/gemm_f16_sm80.py Normal file
View File

@ -0,0 +1,479 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmF16Sm80(unittest.TestCase):
def test_SM80_Device_Gemm_f32t_f32n_f32t_tensor_op_bf16_f32_128x128x32_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.BatchedIdentitySwizzle
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor,
direct_store=True
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32_128x128x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 64],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32_128x256x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 256, 64],
stages=3, warp_count=[2, 4, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32_256x128x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[256, 128, 64],
stages=3, warp_count=[4, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16_sliced_k_128x64x64_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float16, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 64, 64],
stages=3, warp_count=[2, 1, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float16
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_GemmUniversal_f16n_f16t_f32t_tensor_op_f32_64x64x32_32x32x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float16, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[64, 64, 32],
stages=10, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float16
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32_256x128x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[256, 128, 64],
stages=3, warp_count=[4, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_test_SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16_sliced_k_128x64x64_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 64, 64],
stages=3, warp_count=[2, 1, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32_128x256x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 256, 64],
stages=3, warp_count=[2, 4, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.RowMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32_128x256x64_64x64x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16],
element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 256, 64],
stages=3, warp_count=[2, 4, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

182
test/python/backend/gemm/gemm_f16_sm90.py Normal file
View File

@ -0,0 +1,182 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
from functools import partial
import cutlass.backend
from cutlass.backend import *
from cutlass.backend import library
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
# Partial specialziation for naming tests
name_fn = partial(get_name, element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16, arch=90)
def add_test(cls, layouts, alignments, element_output, element_accumulator, element_epilogue,
cluster_shape, threadblock_shape, stages, opclass, persistent=False):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
:param cluster_shape: indexable container of dimensions of threadblock cluster to be launched
:param threadblock_shape: indexable container of dimensions of threadblock tiles
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass_bindings.OpClass
:param persistent: whether this is a persistent warp-specialized kernel
:type persistent: bool
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass_bindings.float16
element_B = cutlass_bindings.float16
inst_shape = [1, 1, 1] if opclass == cutlass_bindings.OpClass.Simt else None
warp_count = [2, 2, 1] if opclass == cutlass_bindings.OpClass.Simt else None
math_inst = MathInstruction(
instruction_shape=inst_shape,
element_a=element_A, element_b=element_B, element_accumulator=element_accumulator,
opcode_class=opclass, math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=threadblock_shape,
cluster_shape=cluster_shape,
stages=stages, warp_count=warp_count,
math_instruction=math_inst,
persistent=persistent
)
A = TensorDescription(element=element_A, layout=layouts[0], alignment=alignments[0])
B = TensorDescription(element=element_B, layout=layouts[1], alignment=alignments[1])
C = TensorDescription(element=element_output, layout=layouts[2], alignment=alignments[2])
epilogue_functor = LinearCombination(C.element, C.alignment, math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=90, tile_description=tile_description, A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor)
self.assertTrue(test_all_gemm(operation, "universal"))
if persistent:
suffix = "_persistent"
else:
suffix = ""
name = name_fn(layouts, alignments, element_output, element_accumulator,
element_epilogue, cluster_shape, threadblock_shape, stages, opclass=opclass, suffix=suffix)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < 90, "Device compute capability is insufficient for SM90 tests.")
class GemmF16Sm90(unittest.TestCase):
"""
Wrapper class to which tests will be added dynamically in __main__
"""
pass
add_test_tensorop = partial(add_test, opclass=cutlass_bindings.OpClass.TensorOp)
add_test_simt = partial(add_test, opclass=cutlass_bindings.OpClass.Simt)
# Tests with 1x1x1 clusters
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNN, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], 3)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTN, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNN, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 64, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 64, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [4, 4, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [4, 4, 8], cutlass_bindings.float16, cutlass_bindings.float16, cutlass_bindings.float16, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float16, cutlass_bindings.float16, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 64, 64], 5)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [2, 2, 2], cutlass_bindings.float16, cutlass_bindings.float16, cutlass_bindings.float16, [1, 1, 1], [128, 128, 32], None)
# Tests with different cluster shapes
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [1, 4, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 4, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [4, 1, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [4, 2, 1], [64, 128, 64], None)
# Tests for persistent warp-specialized threadblocks
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 1, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 1, 1], [128, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [1, 2, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 2, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [1, 4, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [2, 4, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [4, 1, 1], [64, 128, 64], None, persistent=True)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass_bindings.float32, cutlass_bindings.float32, cutlass_bindings.float32, [4, 4, 1], [64, 128, 64], None, persistent=True)
# Tests using SIMT
add_test_simt(GemmF16Sm90, LayoutCombination.NNN, [1, 1, 1], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 128, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.TNN, [1, 1, 1], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 128, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.NTN, [1, 1, 1], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [128, 64, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.TTN, [1, 1, 1], cutlass_bindings.float16, cutlass_bindings.float32, cutlass_bindings.float32, [1, 1, 1], [64, 64, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.NNT, [1, 1, 1], cutlass_bindings.float16, cutlass_bindings.float16, cutlass_bindings.float16, [1, 1, 1], [128, 128, 8], 2)
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

178
test/python/backend/gemm/gemm_f32_sm80.py Normal file
View File

@ -0,0 +1,178 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.memory_manager import get_allocated_size
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmF32nF32nF32nTensorOpF32Sm80(unittest.TestCase):
def test_SM80_Device_Gemm_f32t_f32n_f32t_tensor_op_bf16_f32_128x128x32_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 8],
element_a=cutlass_bindings.float32, element_b=cutlass_bindings.float32,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add_fast_bf16
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
B = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f32n_f32n_f32t_tensor_op_f32_128x128x32_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 8],
element_a=cutlass_bindings.float32, element_b=cutlass_bindings.float32,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
B = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f32n_f32n_f32t_tensor_op_fast_accurate_f32_64x64x32_32x32x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 8],
element_a=cutlass_bindings.float32, element_b=cutlass_bindings.float32,
element_accumulator=cutlass_bindings.float32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add_fast_f32
)
tile_description = TileDescription(
threadblock_shape=[64, 64, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
B = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**24, 2**24)
cutlass.backend.compiler.load_from_cache()
unittest.main()

134
test/python/backend/gemm/gemm_f64_sm80.py Normal file
View File

@ -0,0 +1,134 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmF64TensorOpSm80(unittest.TestCase):
def test_SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64_32x32x16_16x16x16(self):
math_inst = MathInstruction(
instruction_shape=[8, 8, 4],
element_a=cutlass_bindings.float64, element_b=cutlass_bindings.float64,
element_accumulator=cutlass_bindings.float64, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[32, 32, 16],
stages=4, warp_count=[2, 2, 1],
math_instruction=math_inst
)
# alignment 1 restricted for double
A = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.ColumnMajor,
alignment=1
)
B = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
C = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
element_epilogue = cutlass_bindings.float64
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
def test_SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64_64x64x16_32x32x16(self):
math_inst = MathInstruction(
instruction_shape=[8, 8, 4],
element_a=cutlass_bindings.float64, element_b=cutlass_bindings.float64,
element_accumulator=cutlass_bindings.float64, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[64, 64, 16],
stages=4, warp_count=[2, 2, 1],
math_instruction=math_inst
)
# alignment 1 restricted for double
A = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
B = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.ColumnMajor,
alignment=1
)
C = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
element_epilogue = cutlass_bindings.float64
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "universal"))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

124
test/python/backend/gemm/gemm_f64_sm90.py Normal file
View File

@ -0,0 +1,124 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
from functools import partial
import cutlass.backend
from cutlass.backend import *
from cutlass.backend import library
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
name_fn = partial(get_name, element_a=cutlass_bindings.float64, element_b=cutlass_bindings.float64, arch=90)
def add_test(cls, layouts, alignments, element_output, element_accumulator, element_epilogue,
cluster_shape, threadblock_shape, stages, opclass):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
:param cluster_shape: indexable container of dimensions of threadblock cluster to be launched
:param threadblock_shape: indexable container of dimensions of threadblock tiles
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass_bindings.OpClass
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass_bindings.float64
element_B = cutlass_bindings.float64
inst_shape = [1, 1, 1] if opclass == cutlass_bindings.OpClass.Simt else None
warp_count = [2, 2, 1] if opclass == cutlass_bindings.OpClass.Simt else None
math_inst = MathInstruction(
instruction_shape=inst_shape,
element_a=element_A, element_b=element_B, element_accumulator=element_accumulator,
opcode_class=opclass, math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=threadblock_shape,
cluster_shape=cluster_shape,
stages=stages, warp_count=warp_count,
math_instruction=math_inst
)
A = TensorDescription(element=element_A, layout=layouts[0], alignment=alignments[0])
B = TensorDescription(element=element_B, layout=layouts[1], alignment=alignments[1])
C = TensorDescription(element=element_output, layout=layouts[2], alignment=alignments[2])
epilogue_functor = LinearCombination(C.element, C.alignment, math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=90, tile_description=tile_description, A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor)
self.assertTrue(test_all_gemm(operation, "universal"))
name = name_fn(layouts, alignments, element_output, element_accumulator,
element_epilogue, cluster_shape, threadblock_shape, stages, opclass=opclass)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < 90, "Device compute capability is insufficient for SM90 tests.")
class GemmF64Sm90(unittest.TestCase):
"""
Wrapper class to which tests will be added dynamically in __main__
"""
pass
add_test_simt = partial(add_test, opclass=cutlass_bindings.OpClass.Simt)
add_test_simt(GemmF64Sm90, LayoutCombination.NNN, [1, 1, 1], cutlass_bindings.float64, cutlass_bindings.float64, cutlass_bindings.float64, [1, 1, 1], [64, 64, 32], 2)
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

235
test/python/backend/gemm/gemm_grouped_sm80.py Normal file
View File

@ -0,0 +1,235 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_grouped_testbed import TestbedGrouped
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmGroupedSm80(unittest.TestCase):
def test_SM80_Device_GemmGrouped_f16n_f16t_f32n_tensor_op_f32_128x128x32_64x64x32(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16], element_a=cutlass_bindings.float16,
element_b=cutlass_bindings.float16, element_accumulator=cutlass_bindings.float32,
opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.BatchedIdentitySwizzle
for precompute_mode in [SchedulerMode.Device, SchedulerMode.Host]:
operation = GemmOperationGrouped(
80,
tile_description, A, B, C,
epilogue_functor, swizzling_functor,
precompute_mode=precompute_mode
)
testbed = TestbedGrouped(operation=operation)
self.assertTrue(testbed.run(24))
def test_SM80_Device_GemmGrouped_f64t_f64t_f64n_tensor_op_f64_64x64x16_32x32x16(self):
math_inst = MathInstruction(
instruction_shape=[8, 8, 4], element_a=cutlass_bindings.float64,
element_b=cutlass_bindings.float64, element_accumulator=cutlass_bindings.float64,
opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[64, 64, 16],
stages=4, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
B = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.RowMajor,
alignment=1
)
C = TensorDescription(
element=cutlass_bindings.float64, layout=cutlass_bindings.ColumnMajor,
alignment=1
)
element_epilogue = cutlass_bindings.float64
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.BatchedIdentitySwizzle
for precompute_mode in [SchedulerMode.Device, SchedulerMode.Host]:
operation = GemmOperationGrouped(
80,
tile_description, A, B, C,
epilogue_functor, swizzling_functor,
precompute_mode=precompute_mode
)
testbed = TestbedGrouped(operation=operation)
self.assertTrue(testbed.run(24))
def test_SM80_Device_GemmGrouped_f32t_f32t_f32t_simt_f32_128x64x8_64x32x1(self):
math_inst = MathInstruction(
instruction_shape=[1, 1, 1], element_a=cutlass_bindings.float32,
element_b=cutlass_bindings.float32, element_accumulator=cutlass_bindings.float32,
opcode_class=cutlass_bindings.OpClass.Simt,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 64, 8],
stages=4, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=1
)
B = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=1
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.RowMajor,
alignment=1
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.BatchedIdentitySwizzle
for precompute_mode in [SchedulerMode.Device, SchedulerMode.Host]:
operation = GemmOperationGrouped(
80,
tile_description, A, B, C,
epilogue_functor, swizzling_functor,
precompute_mode=precompute_mode
)
testbed = TestbedGrouped(operation=operation)
self.assertTrue(testbed.run(27))
def test_SM80_Device_GemmGrouped_f16n_f16t_f32n_tensor_op_f32_128x128x32_64x64x32_cache(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 16], element_a=cutlass_bindings.float16,
element_b=cutlass_bindings.float16, element_accumulator=cutlass_bindings.float32,
opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 32],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
B = TensorDescription(
element=cutlass_bindings.float16, layout=cutlass_bindings.ColumnMajor,
alignment=8
)
C = TensorDescription(
element=cutlass_bindings.float32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.float32
epilogue_functor = LinearCombination(
C.element, C.alignment,
math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.BatchedIdentitySwizzle
for precompute_mode in [SchedulerMode.Device, SchedulerMode.Host]:
operation = GemmOperationGrouped(
80,
tile_description, A, B, C,
epilogue_functor, swizzling_functor,
precompute_mode=precompute_mode
)
testbed = TestbedGrouped(operation=operation)
self.assertTrue(testbed.run(5))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

261
test/python/backend/gemm/gemm_s8_sm80.py Normal file
View File

@ -0,0 +1,261 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
from cutlass.backend import *
from cutlass.backend.epilogue import LinearCombinationClamp
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
class GemmS8TensorOpF32Sm80(unittest.TestCase):
def test_SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32_64x64x64_32x32x64(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 32],
element_a=cutlass_bindings.int8, element_b=cutlass_bindings.int8,
element_accumulator=cutlass_bindings.int32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add_saturate
)
tile_description = TileDescription(
threadblock_shape=[64, 64, 64],
stages=6, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajorInterleaved32,
alignment=16
)
B = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajorInterleaved32,
alignment=16
)
C = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajorInterleaved32,
alignment=8
)
epilogue_functor = FastLinearCombinationClamp(
C.element, C.alignment
)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "interleaved"))
def test_SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32_256x128x128_64x64x128(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 32],
element_a=cutlass_bindings.int8, element_b=cutlass_bindings.int8,
element_accumulator=cutlass_bindings.int32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 128],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajor,
alignment=16
)
B = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajor,
alignment=16
)
C = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajor,
alignment=16
)
epilogue_functor = FastLinearCombinationClamp(
C.element, C.alignment
)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "multistage"))
def test_SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32_128x128x128_64x64x128(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 32],
element_a=cutlass_bindings.int8, element_b=cutlass_bindings.int8,
element_accumulator=cutlass_bindings.int32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 128],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajor,
alignment=16
)
B = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajor,
alignment=16
)
C = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajor,
alignment=16
)
epilogue_functor = FastLinearCombinationClamp(
C.element, C.alignment
)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "multistage"))
def test_SM80_Device_Gemm_s8t_s8n_s32n_tensor_op_s32_128x128x128_64x64x128(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 32],
element_a=cutlass_bindings.int8, element_b=cutlass_bindings.int8,
element_accumulator=cutlass_bindings.int32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 128],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajor,
alignment=16
)
B = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajor,
alignment=16
)
C = TensorDescription(
element=cutlass_bindings.int32, layout=cutlass_bindings.ColumnMajor,
alignment=4
)
element_epilogue = cutlass_bindings.int32
epilogue_functor = LinearCombinationClamp(
C.element, C.alignment, math_inst.element_accumulator,
element_epilogue
)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "multistage"))
def test_SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32_128x128x128_64x64x128(self):
math_inst = MathInstruction(
instruction_shape=[16, 8, 32],
element_a=cutlass_bindings.int8, element_b=cutlass_bindings.int8,
element_accumulator=cutlass_bindings.int32, opcode_class=cutlass_bindings.OpClass.TensorOp,
math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=[128, 128, 128],
stages=3, warp_count=[2, 2, 1],
math_instruction=math_inst
)
A = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.RowMajor,
alignment=16
)
B = TensorDescription(
element=cutlass_bindings.int8, layout=cutlass_bindings.ColumnMajor,
alignment=16
)
C = TensorDescription(
element=cutlass_bindings.int32, layout=cutlass_bindings.RowMajor,
alignment=4
)
element_epilogue = cutlass_bindings.int32
epilogue_functor = LinearCombinationClamp(
C.element, C.alignment, math_inst.element_accumulator,
element_epilogue
)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=80, tile_description=tile_description,
A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
)
self.assertTrue(test_all_gemm(operation, "multistage"))
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

154
test/python/backend/gemm/gemm_s8_sm90.py Normal file
View File

@ -0,0 +1,154 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
from functools import partial
import cutlass.backend
from cutlass.backend import *
from cutlass.backend import library
from cutlass.backend.test import *
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.gemm_testbed import test_all_gemm
from cutlass.backend.utils.device import device_cc
name_fn = partial(get_name, element_a=cutlass_bindings.float16, element_b=cutlass_bindings.float16, arch=90)
def add_test(cls, layouts, alignments, element_output, element_accumulator, element_epilogue,
cluster_shape, threadblock_shape, stages, opclass, persistent=False):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
:param cluster_shape: indexable container of dimensions of threadblock cluster to be launched
:param threadblock_shape: indexable container of dimensions of threadblock tiles
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass_bindings.OpClass
:param persistent: whether this is a persistent warp-specialized kernel
:type persistent: bool
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass_bindings.int8
element_B = cutlass_bindings.int8
inst_shape = [1, 1, 1] if opclass == cutlass_bindings.OpClass.Simt else None
warp_count = [2, 2, 1] if opclass == cutlass_bindings.OpClass.Simt else None
math_inst = MathInstruction(
instruction_shape=inst_shape,
element_a=element_A, element_b=element_B, element_accumulator=element_accumulator,
opcode_class=opclass, math_operation=MathOperation.multiply_add
)
tile_description = TileDescription(
threadblock_shape=threadblock_shape,
cluster_shape=cluster_shape,
stages=stages, warp_count=warp_count,
math_instruction=math_inst,
persistent=persistent
)
A = TensorDescription(element=element_A, layout=layouts[0], alignment=alignments[0])
B = TensorDescription(element=element_B, layout=layouts[1], alignment=alignments[1])
C = TensorDescription(element=element_output, layout=layouts[2], alignment=alignments[2])
if opclass == cutlass_bindings.OpClass.Simt:
epilogue_functor_cls = LinearCombinationClamp
else:
epilogue_functor_cls = LinearCombination
epilogue_functor = epilogue_functor_cls(C.element, C.alignment, math_inst.element_accumulator, element_epilogue)
swizzling_functor = cutlass_bindings.IdentitySwizzle1
operation = GemmOperationUniversal(
arch=90, tile_description=tile_description, A=A, B=B, C=C,
epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor)
self.assertTrue(test_all_gemm(operation, "universal"))
if persistent:
suffix = "_persistent"
else:
suffix = ""
name = name_fn(layouts, alignments, element_output, element_accumulator,
element_epilogue, cluster_shape, threadblock_shape, stages, opclass=opclass, suffix=suffix)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < 90, "Device compute capability is insufficient for SM90 tests.")
class GemmS8Sm90(unittest.TestCase):
"""
Wrapper class to which tests will be added dynamically in __main__
"""
pass
add_test_tensorop = partial(add_test, opclass=cutlass_bindings.OpClass.TensorOp)
add_test_simt = partial(add_test, opclass=cutlass_bindings.OpClass.Simt)
# Tests with 1x1x1 clusters
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNN, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [128, 128, 128], 3)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 8], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [64, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [128, 64, 32], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [4, 4, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [128, 128, 128], None)
# Tests with different cluster shapes
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [2, 2, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 4, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [4, 4, 1], [128, 128, 128], None)
# Tests with persistent warp-specialized threadblocks
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [2, 1, 1], [128, 128, 128], None, persistent=True)
# Tests for SIMT
add_test_simt(GemmS8Sm90, LayoutCombination.TNN, [1, 1, 1], cutlass_bindings.int8, cutlass_bindings.int32, cutlass_bindings.int32, [1, 1, 1], [64, 32, 8], 2)
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
unittest.main()

41
test/python/backend/gemm/run_all_tests.py Normal file
View File

@ -0,0 +1,41 @@
#################################################################################################
#
# Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. 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.
#
# 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
import cutlass.backend
import unittest
if __name__ == '__main__':
cutlass.backend.get_memory_pool(2**30, 2**30)
loader = unittest.TestLoader()
tests = loader.discover('./', 'gemm_*.py')
testRunner = unittest.runner.TextTestRunner()
testRunner.run(tests)