leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

Make Python interface work for non-SM80 targets (#726)

Browse Source 
* Make Python interface work for non-SM80 targets

* Remove line in README
This commit is contained in:
Jack Kosaian
2022-12-07 21:53:33 -05:00
committed by GitHub
parent d6117ca362
commit df81d847d7
33 changed files with 149 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/40_cutlass_py/README.md
-1
View File
@@ -2,7 +2,6 @@
This directory contains examples of using CUTLASS's Python interface. It consists of two types of examples:
* _Basic examples_: minimal examples that illustrate how to set up GEMMs, convolutions, and grouped GEMM operations
* [_Customizable examples_](customizable): examples that allow one to specify a variety of template parameters for the given kernel
>>>>>>> Add simplified examples
## Setting up the Python interface
Please follow the instructions [here](/tools/library/scripts/pycutlass/README.md#installation) to set up the Python API.
examples/40_cutlass_py/conv2d.py
+12 -3
View File
@@ -41,7 +41,7 @@ import sys
import cutlass
import pycutlass
from pycutlass import *
import util
from pycutlass.utils.device import device_cc
parser = argparse.ArgumentParser(
@@ -62,7 +62,7 @@ except:
sys.exit(0)
# Check that the device is of a sufficient compute capability
cc = util.get_device_cc()
cc = device_cc()
assert cc >= 70, "The CUTLASS Python Conv2d example requires compute capability greater than or equal to 70."
alignment = 1
@@ -82,8 +82,17 @@ C = TensorDescription(cutlass.float32, cutlass.TensorNHWC, alignment)
element_acc = cutlass.float32
element_epilogue = cutlass.float32
# Select instruction shape based on the Tensor Core instructions supported
# by the device on which we are running
if cc == 70:
instruction_shape = [8, 8, 4]
elif cc == 75:
instruction_shape = [16, 8, 8]
else:
instruction_shape = [16, 8, 16]
math_inst = MathInstruction(
[16, 8, 8], # Shape of the Tensor Core instruction
instruction_shape,
A.element, B.element, element_acc,
cutlass.OpClass.TensorOp,
MathOperation.multiply_add
examples/40_cutlass_py/customizable/conv2d.py
+6
View File
@@ -34,6 +34,7 @@ import pycutlass
from pycutlass import *
from pycutlass.conv2d_operation import *
from pycutlass.utils import reference_model
from pycutlass.utils.device import device_cc
import sys
import torch.nn.functional as F
@@ -146,6 +147,11 @@ try:
except:
sys.exit(0)
cc = device_cc()
if args.compute_capability != cc:
raise Exception(("Parameter --compute-capability of {} "
"does not match that of the device of {}.").format(args.compute_capability, cc))
pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
np.random.seed(0)
examples/40_cutlass_py/customizable/gemm.py
+6 -1
View File
@@ -34,6 +34,7 @@ import pycutlass
from pycutlass import *
import cutlass
from bfloat16 import bfloat16
from pycutlass.utils.device import device_cc
import sys
import argparse
@@ -131,12 +132,16 @@ parser.add_argument("-activ_arg", "--activation_args", default=[], nargs="+", ty
parser.add_argument('--print_cuda', action="store_true",
help="print the underlying CUDA kernel")
try:
args = parser.parse_args()
except:
sys.exit(0)
cc = device_cc()
if args.compute_capability != cc:
raise Exception(("Parameter --compute-capability of {} "
"does not match that of the device of {}.").format(args.compute_capability, cc))
pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
pycutlass.compiler.nvcc()
examples/40_cutlass_py/customizable/gemm_grouped.py
+6
View File
@@ -32,6 +32,7 @@
import numpy as np
import pycutlass
from pycutlass import *
from pycutlass.utils.device import device_cc
import csv
import sys
@@ -129,6 +130,11 @@ try:
except:
sys.exit(0)
cc = device_cc()
if args.compute_capability != cc:
raise Exception(("Parameter --compute-capability of {} "
"does not match that of the device of {}.").format(args.compute_capability, cc))
pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
np.random.seed(0)
examples/40_cutlass_py/gemm.py
+13 -3
View File
@@ -40,7 +40,7 @@ import sys
import cutlass
import pycutlass
from pycutlass import *
import util
from pycutlass.utils.device import device_cc
parser = argparse.ArgumentParser(description="Launch a GEMM kernel from Python: 'D = alpha * A * B + beta * C'")
@@ -55,7 +55,7 @@ except:
sys.exit(0)
# Check that the device is of a sufficient compute capability
cc = util.get_device_cc()
cc = device_cc()
assert cc >= 70, "The CUTLASS Python GEMM example requires compute capability greater than or equal to 70."
alignment = 8
@@ -78,13 +78,23 @@ C = TensorDescription(cutlass.float32, cutlass.ColumnMajor, alignment)
element_acc = cutlass.float32
element_epilogue = cutlass.float32
# Select instruction shape based on the Tensor Core instructions supported
# by the device on which we are running
if cc == 70:
instruction_shape = [8, 8, 4]
elif cc == 75:
instruction_shape = [16, 8, 8]
else:
instruction_shape = [16, 8, 16]
math_inst = MathInstruction(
[16, 8, 8], # Shape of the Tensor Core instruction
instruction_shape,
A.element, B.element, element_acc,
cutlass.OpClass.TensorOp,
MathOperation.multiply_add
)
tile_description = TileDescription(
[128, 128, 32], # Threadblock shape
2, # Number of stages
examples/40_cutlass_py/gemm_grouped.py
+12 -3
View File
@@ -40,7 +40,7 @@ import sys
import cutlass
import pycutlass
from pycutlass import *
import util
from pycutlass.utils.device import device_cc
parser = argparse.ArgumentParser(description="Launch a grouped GEMM kernel from Python")
@@ -52,7 +52,7 @@ except:
sys.exit(0)
# Check that the device is of a sufficient compute capability
cc = util.get_device_cc()
cc = device_cc()
assert cc >= 70, "The CUTLASS Python grouped GEMM example requires compute capability greater than or equal to 70."
np.random.seed(0)
@@ -71,8 +71,17 @@ C = TensorDescription(cutlass.float32, cutlass.ColumnMajor, alignment)
element_acc = cutlass.float32
element_epilogue = cutlass.float32
# Select instruction shape based on the Tensor Core instructions supported
# by the device on which we are running
if cc == 70:
instruction_shape = [8, 8, 4]
elif cc == 75:
instruction_shape = [16, 8, 8]
else:
instruction_shape = [16, 8, 16]
math_inst = MathInstruction(
[16, 8, 8], # Shape of the Tensor Core instruction
instruction_shape,
A.element, B.element, element_acc,
cutlass.OpClass.TensorOp,
MathOperation.multiply_add
examples/40_cutlass_py/util.py
-60
View File
@@ -1,60 +0,0 @@
#################################################################################################
#
# Copyright (c) 2017 - 2022 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.
#
#################################################################################################
"""
Utility functions for interacting with device
"""
from cuda import cudart
# Raises an exception if `result` returned an error. Otherwise returns the result.
def check_cuda_errors(result: list):
# `result` is of the format : (cudaError_t, result...)
err = result[0]
if err.value:
raise RuntimeError("CUDA error: {}".format(cudart.cudaGetErrorName(err)))
if len(result) == 1:
return None
elif len(result) == 2:
return result[1]
else:
return result[1:]
# Returns the integer representation of the device compute capability
def get_device_cc(device: int = 0):
deviceProp = check_cuda_errors(cudart.cudaGetDeviceProperties(device))
major = str(deviceProp.major)
minor = str(deviceProp.minor)
return int(major + minor)