From 189fa9f027e1cb83e381517c01051f7cd3585f6e Mon Sep 17 00:00:00 2001 From: Jay Pipes Date: Tue, 28 Aug 2018 11:53:39 -0400 Subject: [PATCH] clean up CUDA traits There was a misunderstanding of what the CUDA traits really stood for. This wasn't helped by the fact that the NVIDIA CUDA documentation and versioning system is so awkward. Basically, there's the CUDA Compute Capability version, which represents the hardware GPU instruction set extensions supported by various NVIDIA GPUs. And there is the CUDA SDK version which represents a window of CUDA Compute Capability versions that the library is able to work with. This patch adds a new os_traits.hw.gpu.cuda module with these new traits. It removes the old os_traits.hw.gpu.api.CUDA* traits since those were misleading and inaccurate. It also adds a new utility function that maps CUDA SDK version to supported CUDA Compute Capability versions that the SDK supports. This support matrix information was taken from the CUDA wikipedia article which lists the supported version information. Change-Id: Id4adb6874615ad5f9f8aff780a83119e10a38c35 --- doc/source/reference/index.rst | 55 +++++++++++++++- os_traits/hw/gpu/api.py | 25 ------- os_traits/hw/gpu/cuda.py | 115 +++++++++++++++++++++++++++++++++ os_traits/tests/test_cuda.py | 42 ++++++++++++ 4 files changed, 211 insertions(+), 26 deletions(-) create mode 100644 os_traits/hw/gpu/cuda.py create mode 100644 os_traits/tests/test_cuda.py diff --git a/doc/source/reference/index.rst b/doc/source/reference/index.rst index 9218b26..b2ae1a7 100644 --- a/doc/source/reference/index.rst +++ b/doc/source/reference/index.rst @@ -2,4 +2,57 @@ Reference ========= -TODO. +CUDA +---- + +Applications that need to perform massively parallel operations, like +processing large arrays, may use the CUDA framework to accelerate their +processing on graphics processing units (GPUs). The CUDA framework has two +complementary pieces to it. + +There are a set of GPU instruction set extensions that are implemented by +various graphics cards. These instruction set extensions are known as the CUDA +Compute Capability. + +The second part of the framework is an SDK that allows developers to take +advantage of the hardware's instruction set extensions of a particular version +(a specific CUDA Compute Capability version, that is). + +An application will link with a version of the CUDA SDK, and the version of the +CUDA SDK controls which CUDA Compute Capability versions the application will +be able to work with. + +The ``os_traits.hw.gpu.cuda`` module contains traits for both the CUDA compute +capability version as well as the CUDA SDK version. For example, +``os_traits.hw.gpu.cuda.COMPUTE_CAPABILITY_V3_2`` and +``os_traits.hw.gpu.cuda.SDK_V6_5``. + +The ``os_traits.hw.gpu.cuda`` module contains a utility function called +``compute_capabilities_supported()`` that accepts a trait indicating the CUDA +SDK version and returns a ``set()`` containing the matching CUDA compute +capability traits that that version of the CUDA SDK knows how to utilize. + +Here is an example of listing the CUDA compute capability version traits that +the CUDA SDK 8.0 is capable of working with:: + + >>> from os_traits.hw.gpu import cuda + >>> import pprint + >>> + >>> sdk8_caps = cuda.compute_capabilities_supported(cuda.SDK_V8_0) + >>> pprint.pprint(sdk8_caps) + set(['HW_GPU_CUDA_COMPUTE_CAPABILITY_V2_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V2_1', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_2', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_5', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_7', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_2', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_3', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_1', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_2']) + +For more information on CUDA, see the `Wikipedia article`_. + +.. _Wikipedia article: https://en.wikipedia.org/wiki/CUDA diff --git a/os_traits/hw/gpu/api.py b/os_traits/hw/gpu/api.py index 4b41963..32764cb 100644 --- a/os_traits/hw/gpu/api.py +++ b/os_traits/hw/gpu/api.py @@ -65,29 +65,4 @@ TRAITS = [ 'OPENGL_V4_3', 'OPENGL_V4_4', 'OPENGL_V4_5', - # ref: https://en.wikipedia.org/wiki/CUDA - # ref: https://developer.nvidia.com/cuda-toolkit-archive - 'CUDA_V1_0', - 'CUDA_V1_1', - 'CUDA_V1_2', - 'CUDA_V1_3', - 'CUDA_V2_0', - 'CUDA_V2_1', - 'CUDA_V3_0', - 'CUDA_V3_2', - 'CUDA_V3_5', - 'CUDA_V3_7', - 'CUDA_V5_0', - 'CUDA_V5_2', - 'CUDA_V5_3', - 'CUDA_V6_0', - 'CUDA_V6_1', - 'CUDA_V6_2', - 'CUDA_V7_0', - 'CUDA_V7_1', - 'CUDA_V7_5', - 'CUDA_V8_0', - 'CUDA_V9_0', - 'CUDA_V9_1', - 'CUDA_V9_2', ] diff --git a/os_traits/hw/gpu/cuda.py b/os_traits/hw/gpu/cuda.py new file mode 100644 index 0000000..7e848fa --- /dev/null +++ b/os_traits/hw/gpu/cuda.py @@ -0,0 +1,115 @@ +# -*- coding: utf-8 -*- + +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT +# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the +# License for the specific language governing permissions and limitations +# under the License. + + +TRAITS = [ + # ref: https://en.wikipedia.org/wiki/CUDA + # ref: https://developer.nvidia.com/cuda-toolkit-archive + 'COMPUTE_CAPABILITY_V1_0', + 'COMPUTE_CAPABILITY_V1_1', + 'COMPUTE_CAPABILITY_V1_2', + 'COMPUTE_CAPABILITY_V1_3', + 'COMPUTE_CAPABILITY_V2_0', + 'COMPUTE_CAPABILITY_V2_1', + 'COMPUTE_CAPABILITY_V3_0', + 'COMPUTE_CAPABILITY_V3_2', + 'COMPUTE_CAPABILITY_V3_5', + 'COMPUTE_CAPABILITY_V3_7', + 'COMPUTE_CAPABILITY_V5_0', + 'COMPUTE_CAPABILITY_V5_2', + 'COMPUTE_CAPABILITY_V5_3', + 'COMPUTE_CAPABILITY_V6_0', + 'COMPUTE_CAPABILITY_V6_1', + 'COMPUTE_CAPABILITY_V6_2', + 'COMPUTE_CAPABILITY_V7_0', + 'COMPUTE_CAPABILITY_V7_1', + 'COMPUTE_CAPABILITY_V7_2', + 'SDK_V6_5', + 'SDK_V7_5', + 'SDK_V8_0', + 'SDK_V9_0', + 'SDK_V9_1', + 'SDK_V9_2', + 'SDK_V10_0', +] + +_CAPS_V1 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V1_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V1_1', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V1_2', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V1_3', +] + +_CAPS_V2 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V2_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V2_1', +] + +_CAPS_V3 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_2', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_5', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V3_7', +] + +_CAPS_V5 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_2', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V5_3', +] + +_CAPS_V6 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_1', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V6_2', +] + +_CAPS_V7 = [ + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V7_0', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V7_1', + 'HW_GPU_CUDA_COMPUTE_CAPABILITY_V7_2', +] + +_SDK_COMPUTE_CAP_SUPPORT = { + 'HW_GPU_CUDA_SDK_V6_5': set( + _CAPS_V1 + _CAPS_V2 + _CAPS_V3 + _CAPS_V5 + ), + 'HW_GPU_CUDA_SDK_V7_5': set( + _CAPS_V2 + _CAPS_V3 + _CAPS_V5 + ), + 'HW_GPU_CUDA_SDK_V8_0': set( + _CAPS_V2 + _CAPS_V3 + _CAPS_V5 + _CAPS_V6 + ), + 'HW_GPU_CUDA_SDK_V9_0': set( + _CAPS_V3 + _CAPS_V5 + _CAPS_V6 + _CAPS_V7 + ), + 'HW_GPU_CUDA_SDK_V9_1': set( + _CAPS_V3 + _CAPS_V5 + _CAPS_V6 + _CAPS_V7 + ), + 'HW_GPU_CUDA_SDK_V9_2': set( + _CAPS_V3 + _CAPS_V5 + _CAPS_V6 + _CAPS_V7 + ), + 'HW_GPU_CUDA_SDK_V10_0': set( + _CAPS_V3 + _CAPS_V5 + _CAPS_V6 + _CAPS_V7 + ), +} + + +def compute_capabilities_supported(sdk_trait): + """Given an SDK trait, returns a set of compute capability traits that the + version of the SDK supports. + + Returns None if no matches were found for the SDK trait. + """ + return _SDK_COMPUTE_CAP_SUPPORT.get(sdk_trait) diff --git a/os_traits/tests/test_cuda.py b/os_traits/tests/test_cuda.py new file mode 100644 index 0000000..eaa7984 --- /dev/null +++ b/os_traits/tests/test_cuda.py @@ -0,0 +1,42 @@ +# -*- coding: utf-8 -*- + +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT +# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the +# License for the specific language governing permissions and limitations +# under the License. + +from os_traits.hw.gpu import cuda +from os_traits.tests import base + + +class TestCUDA(base.TestCase): + + def test_unknown_sdk_support(self): + self.assertIsNone(cuda.compute_capabilities_supported('UNKNOWN')) + + def test_sdk6_5_support(self): + expected = set([ + cuda.COMPUTE_CAPABILITY_V1_0, + cuda.COMPUTE_CAPABILITY_V1_1, + cuda.COMPUTE_CAPABILITY_V1_2, + cuda.COMPUTE_CAPABILITY_V1_3, + cuda.COMPUTE_CAPABILITY_V2_0, + cuda.COMPUTE_CAPABILITY_V2_1, + cuda.COMPUTE_CAPABILITY_V3_0, + cuda.COMPUTE_CAPABILITY_V3_2, + cuda.COMPUTE_CAPABILITY_V3_5, + cuda.COMPUTE_CAPABILITY_V3_7, + cuda.COMPUTE_CAPABILITY_V5_0, + cuda.COMPUTE_CAPABILITY_V5_2, + cuda.COMPUTE_CAPABILITY_V5_3, + ]) + + actual = cuda.compute_capabilities_supported(cuda.SDK_V6_5) + self.assertEqual(expected, actual)