[SYCL][DOC][CUDA][HIP] Update getStartedGuide.md #10669

JackAKirk · 2023-08-03T08:33:06Z

Some instructions for building/running/using HIP and CUDA backends was out of date. This PR updates them.

Note that I think there is scope for further improvement of this doc generally. This PR makes improvements only wrt the CUDA/HIP backends.

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

bader

The content looks good to me. Mostly formatting suggestions.

bader · 2023-08-14T17:27:39Z

sycl/doc/GetStartedGuide.md

+- [Getting Started with oneAPI DPC++](#getting-started-with-oneapi-dpc)
+  - [Table of contents](#table-of-contents)


I'm not sure if these two items are useful. I guess we intentionally skipped them.

What is the reason for changing * with -?

Makes sense, I've removed them.

I didn't notice the * -> - change, I think the lint did this. I've changed it back.

bader · 2023-08-14T17:34:33Z

sycl/doc/GetStartedGuide.md

 The CUDA backend should work on Windows or Linux operating systems with any
-GPU compatible with SM 50 or above. The default SM for the NVIDIA CUDA backend
-is 5.0. Users can specify lower values, but some features may not be supported.
+GPU with compute capability (SM version) sm_50 or above. The default
+SM version for the NVIDIA CUDA backend is sm_50. Users of sm_3X devices can
+attempt to specify the target architecture [ahead of time](#aot-target-architectures),
+provided that they use a 11.X CUDA Runtime version, but some features may
+not be supported. The CUDA backend has been tested with
+different Ubuntu linux distributions and a selection of supported CUDA toolkit versions
+and GPUs. The backend is tested by a relevant device/toolkit prior to a oneapi plugin release.
+Go to the plugin release [pages](https://developer.codeplay.com/products/oneapi/nvidia/)
+for further details.
+


Suggested change

The CUDA backend should work on Windows or Linux operating systems with any

GPU compatible with SM 50 or above. The default SM for the NVIDIA CUDA backend

is 5.0. Users can specify lower values, but some features may not be supported.

GPU with compute capability (SM version) sm_50 or above. The default

SM version for the NVIDIA CUDA backend is sm_50. Users of sm_3X devices can

attempt to specify the target architecture [ahead of time](#aot-target-architectures),

provided that they use a 11.X CUDA Runtime version, but some features may

not be supported. The CUDA backend has been tested with

different Ubuntu linux distributions and a selection of supported CUDA toolkit versions

and GPUs. The backend is tested by a relevant device/toolkit prior to a oneapi plugin release.

Go to the plugin release [pages](https://developer.codeplay.com/products/oneapi/nvidia/)

for further details.

The CUDA backend should work on Windows or Linux operating systems with any GPU

with compute capability (SM version) sm_50 or above. The default SM version for

the NVIDIA CUDA backend is sm_50. Users of sm_3X devices can attempt to specify

the target architecture [ahead of time](#aot-target-architectures), provided

that they use a 11.X CUDA Runtime version, but some features may not be

supported. The CUDA backend has been tested with different Ubuntu Linux

distributions and a selection of supported CUDA toolkit versions and GPUs. The

backend is tested by a relevant device/toolkit prior to a oneAPI plugin release.

Go to the plugin release

[pages](https://developer.codeplay.com/products/oneapi/nvidia/) for further

details.

Use 80-char line limit.

linux -> Linux

oneapi -> oneAPI

sycl/doc/GetStartedGuide.md

Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Use 80-char line limit. linux -> Linux oneapi -> oneAPI Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Use 80-char line limit. linux -> Linux oneapi -> oneAPI Co-authored-by: Alexey Bader <alexey.bader@intel.com>

- -> * Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

linux -> Linux oneapi -> oneAPI Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

hdelan · 2023-08-22T09:28:39Z

sycl/doc/GetStartedGuide.md

@@ -179,7 +179,7 @@ CUDA backend has Windows support; Windows Subsystem for Linux (WSL) is not
 needed to build and run the CUDA backend.

 Enabling this flag requires an installation of at least
-[CUDA 10.2](https://developer.nvidia.com/cuda-10.2-download-archive) on
+[CUDA 11.5](https://developer.nvidia.com/cuda-11-5-0-download-archive) on


Can we change this to 11.0? Our plugins are aimed to target 11.0 so would be good if we had the same support in open source

Yeah is it definitely working now for 11.0?

Pending https://github.com/intel/llvm/pull/10922/files it should be

Will it work for e.g. 10.2 or even earlier? Ideally we can support as many cuda versions as possible. It's been pointed out many times that clusters don't update their versions for long periods e.g. #10858 (comment)

I think as long as we keep intermittently checking that CUDA 10.2 is working we can keep this support. But the problem is that we don't have any CI for CUDA 10.2 whereas we do have CI for CUDA 11.0 (to test the releases internally)

There are version checks for above/below 10.2 here:

llvm/sycl/plugins/unified_runtime/ur/adapters/cuda/usm.cpp

Line 242 in f69dae7

// CU_POINTER_ATTRIBUTE_RANGE_START_ADDR was introduced in CUDA 10.2

Maybe with the right checks it < 10.2 could work fine. Or maybe it already works.

I would say that it already works. But we don't have CI to check every commit

intel/llvm CI started with 11.4 version.
We switched to 11.7 due to build issues of SYCL-CTS. #6545
Today we validate each commit with 12.1 version, but I unfortunately, I didn't record the reason for uplift from 11.7.

I would say that it already works. But we don't have CI to check every commit

Do we need to use 10.2 in CI instead of 12.1?
IIRC, all previous uplifts were agreed with @AerialMantis.

I believe that from the next release we are meant to be supporting 12.1, so it makes sense the CI tests that as the priority. It would be nice to keep track of whether older releases still work, so that systems that don't regularly update their toolkit can still stay supported; but I think that is more on codeplay to test from time to time, and when introducing new features that might break older versions.

I've dug through some older commits and I think that versions older than 10.2 should work for appropriate devices.
We can't easily test it since we don't have a system where 10.x or older versions are installed.
I've updated the doc accordingly.

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

JackAKirk · 2023-08-28T09:21:44Z

@intel/llvm-gatekeepers could you please merge this?

JackAKirk added 2 commits August 2, 2023 17:54

Updated nvidia and hip docs in getStartedGuide.md

17633f4

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

Moved text to appropriate location.

b176f08

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

JackAKirk marked this pull request as ready for review August 3, 2023 14:01

JackAKirk requested a review from a team as a code owner August 3, 2023 14:01

Emphasize that amd requires target arch.

7be2d75

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

bader approved these changes Aug 14, 2023

View reviewed changes

jchlanda approved these changes Aug 15, 2023

View reviewed changes

JackAKirk and others added 8 commits August 21, 2023 15:04

Update sycl/doc/GetStartedGuide.md

01f0ba5

Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Update sycl/doc/GetStartedGuide.md

83fb9bc

Use 80-char line limit. linux -> Linux oneapi -> oneAPI Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Update sycl/doc/GetStartedGuide.md

41dd918

Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Update sycl/doc/GetStartedGuide.md

fdf9ec5

Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Update sycl/doc/GetStartedGuide.md

f969e43

Use 80-char line limit. linux -> Linux oneapi -> oneAPI Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Merge branch 'sycl' into update-getStarted-cuda

f521051

Remove toc and get started headers

dc409b5

- -> * Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

Use 80-char line limit.

c015aa8

linux -> Linux oneapi -> oneAPI Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

hdelan reviewed Aug 22, 2023

View reviewed changes

Updated discussion of cuda toolkit support.

ada674c

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

JackAKirk requested a review from hdelan August 25, 2023 14:02

hdelan approved these changes Aug 25, 2023

View reviewed changes

dm-vodopyanov merged commit 3cec71f into intel:sycl Aug 28, 2023
1 check passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SYCL][DOC][CUDA][HIP] Update getStartedGuide.md #10669

[SYCL][DOC][CUDA][HIP] Update getStartedGuide.md #10669

JackAKirk commented Aug 3, 2023

bader left a comment

bader Aug 14, 2023

JackAKirk Aug 21, 2023

bader Aug 14, 2023

hdelan Aug 22, 2023

JackAKirk Aug 22, 2023

hdelan Aug 22, 2023

JackAKirk Aug 22, 2023

hdelan Aug 22, 2023

JackAKirk Aug 22, 2023

hdelan Aug 22, 2023 •

edited

Loading

bader Aug 22, 2023

JackAKirk Aug 22, 2023 •

edited

Loading

JackAKirk Aug 24, 2023

JackAKirk commented Aug 28, 2023

		- [Getting Started with oneAPI DPC++](#getting-started-with-oneapi-dpc)
		- [Table of contents](#table-of-contents)

[SYCL][DOC][CUDA][HIP] Update getStartedGuide.md #10669

[SYCL][DOC][CUDA][HIP] Update getStartedGuide.md #10669

Conversation

JackAKirk commented Aug 3, 2023

bader left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hdelan Aug 22, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JackAKirk Aug 22, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JackAKirk commented Aug 28, 2023

hdelan Aug 22, 2023 •

edited

Loading

JackAKirk Aug 22, 2023 •

edited

Loading