[SYCL][Docs] Add sycl_ext_oneapi_virtual_mem extension and implementation by steffenlarsen · Pull Request #8954 · intel/llvm

steffenlarsen · 2023-04-05T15:27:22Z

This commit adds the sycl_ext_oneapi_virtual_mem experimental extension for reserving and mapping virtual address ranges. Accompanying it is the implementation in the SYCL runtime, together with CUDA and Level Zero backend support for the corresponding features.

…tion This commit adds the sycl_ext_oneapi_virtual_mem experimental extension for reserving and mapping virtual address ranges. Accompanying it is the implementation in the SYCL runtime, together with CUDA and Level Zero backend support for the corresponding features. Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

sycl/plugins/level_zero/pi_level_zero.cpp

steffenlarsen · 2023-04-05T15:30:32Z

/testwin

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

cperkinsintel · 2023-04-05T15:38:42Z

sycl/test-e2e/VirtualMem/vector_with_virtual_mem.cpp

+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+// RUN: %ACC_RUN_PLACEHOLDER %t.out
+
+// TODO: Require ext_oneapi_virtual_mem aspect here when supported.


maybe use the feature test macro as well.

jandres742 · 2023-04-05T16:00:17Z

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

+Working with virtual address ranges and the underlying physical memory requires
+the user to align and adjust in accordance with a specified minimum granularity.
+In addition, devices can have a recommended granularity which may different from
+the minimum granularity and can be used instead of the minimum granularity.


In addition, devices can have a recommended granularity which may different from
the minimum granularity and can be used instead of the minimum granularity.

I would rephrase this to make it clearer.

I tried rephrasing it. Is this what you had in mind?

maybe:

-> In addition, adapters may return a recommended granularity to potentially achieve higher performance. Distinction between minimum and recommended is adapters-specific and may vary between devices.

I like that. I've changed it a little to refer to common SYCL concepts but otherwise I'm good with the message. Thanks!

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

EwanC

Some very minor spec comments from a quick read through to understand what this extension is about.

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

gmlueck · 2024-05-31T11:55:00Z

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

+mapped through a call to `physical_mem::map()` prior to calling this. The range
+must not be a proper sub-range of a previously mapped range. `syclContext` must
+be the same as the context returned by the `get_context()` member function on
+all the `physical_mem` the address ranges are currently mapped to.


The part about "all the physical_mem the address ranges are currently mapped to" is confusing. Under what circumstances could there be more than one physical_mem that maps this address range?

I do believe it should be "all the physical_mem the address range is currently mapped to." The test in sycl/test-e2e/VirtualMem/vector_with_virtual_mem.cpp shows a case like this, where we keep allocating new physical memory chunks, that are all used under a single virtual memory range. They cannot overlap though.

We have the following note two paragraphs below:

[Note: Unmapping ranges that span multiple contiguous mapped ranges is not supported. Doing so will result in undefined behavior. This restriction may be lifted in the future. — end note]

That seems to indicate that the range must correspond to exactly one physical_mem. Is the note wrong?

No, the note just states that if you have a virtual range that spans multiple physical memory allocations, you have to unmap the the sub-ranges that is mapped to each physical memory allocation in separation.

Discussion offline: The wording about all physical memory regions is the confusing part. It has been changed.

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

gmlueck · 2024-05-31T11:59:59Z

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

+
+|`void set_access_mode(const void *ptr, size_t numBytes, address_access_mode mode, const context &syclContext)` |
+Sets the access mode of a virtual memory range specified by `ptr` and
+`numBytes`.


Can you check your wording, then? The wording says:

ptr must be aligned to the minimum memory granularity of syclContext and numBytes must be a multiple of the minimum memory granularity of syclContext.

This would correspond to the host page size, right? Not the value returned by zeVirtualMemQueryPageSize.

gmlueck · 2024-05-31T12:03:50Z

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

+`access_mode::read` results in undefined behavior.
+
+An accessible pointer behaves the same as a pointer to device USM memory and can
+be used in place of a device USM pointer in any interface accepting one.


Can you map a new virtual address over physical memory that is already mapped to a different address?

I see nothing in both CUDA and L0 disallowing this.

That's interesting. What happens in this case? Is the physical memory automatically unmapped from the old virtual address range?

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

gmlueck

spec changes LGTM

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

sycl/source/detail/device_impl.cpp

sycl/source/detail/physical_mem_impl.hpp

sycl/source/virtual_mem.cpp

sycl/doc/extensions/experimental/sycl_ext_oneapi_virtual_mem.asciidoc

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

AllanZyne · 2025-07-16T01:46:17Z

Hi, I found "read-only" mapped VM doesn't work #19189, do we have plan to fix it?

steffenlarsen requested a review from a team April 5, 2023 15:27

steffenlarsen commented Apr 5, 2023

View reviewed changes

sycl/plugins/level_zero/pi_level_zero.cpp Outdated Show resolved Hide resolved

steffenlarsen temporarily deployed to aws April 5, 2023 15:33 — with GitHub Actions Inactive

Fix declaration in ESIMD emu

81703b6

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

cperkinsintel reviewed Apr 5, 2023

View reviewed changes

steffenlarsen temporarily deployed to aws April 5, 2023 16:06 — with GitHub Actions Inactive

jandres742 reviewed Apr 5, 2023

View reviewed changes

steffenlarsen temporarily deployed to aws April 5, 2023 17:34 — with GitHub Actions Inactive

steffenlarsen temporarily deployed to aws April 5, 2023 20:17 — with GitHub Actions Inactive

steffenlarsen temporarily deployed to aws April 5, 2023 21:31 — with GitHub Actions Inactive

Larsen, Steffen added 4 commits April 6, 2023 00:25

Initialize CUDA property structs

ce0829c

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Add Windows symbols

0fef874

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Clarify need for unmapped free

294dd96

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Change std::min to std::max

053cbdc

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

steffenlarsen temporarily deployed to aws April 6, 2023 08:13 — with GitHub Actions Inactive

EwanC reviewed Apr 6, 2023

View reviewed changes

steffenlarsen temporarily deployed to aws April 6, 2023 08:45 — with GitHub Actions Inactive

Larsen, Steffen added 2 commits April 6, 2023 01:49

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

1129750

Fix formatting

7b1b5ec

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

steffenlarsen temporarily deployed to aws April 6, 2023 09:25 — with GitHub Actions Inactive

steffenlarsen temporarily deployed to aws April 6, 2023 09:57 — with GitHub Actions Inactive

Larsen, Steffen added 3 commits April 6, 2023 04:01

Reduce work-item count

ac8843f

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Fix formatting

f8fad93

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Rephrase recommended granularity

0a6e5f9

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

steffenlarsen temporarily deployed to aws April 6, 2023 13:29 — with GitHub Actions Inactive

steffenlarsen temporarily deployed to aws April 6, 2023 14:00 — with GitHub Actions Inactive

Larsen, Steffen added 2 commits April 11, 2023 06:21

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

d0b3229

Address comments

27c7200

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Larsen, Steffen added 11 commits May 29, 2024 05:55

Change back to 2023

1fa82c0

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

997cc39

Remove windows symbol

3be01f4

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Add new context granularity query

743eb2e

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Amend granularity reqs

b34d068

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Fix windows build

93db42b

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Fix wrong name

eeaaad0

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

e8c98b5

Add missing Windows symbol

e5483be

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Clarify map and access functions

290401c

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Change from host page size to minimum granularity

6630795

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

gmlueck reviewed May 31, 2024

View reviewed changes

Larsen, Steffen added 5 commits June 20, 2024 03:01

Address comments

5c7330d

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

187b0f7

Change requirement for access mode functions

8dd8cb3

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Reword and add restriction

327687b

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Fix formatting

639aab3

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

gmlueck approved these changes Jun 20, 2024

View reviewed changes

KseniyaTikhomirova reviewed Jun 27, 2024

View reviewed changes

Larsen, Steffen added 3 commits June 27, 2024 05:57

Address comments

42cbcde

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Merge remote-tracking branch 'intel/sycl' into steffen/virtual_mem_ext

95cdcbb

Fix typo

56ff29c

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

gmlueck approved these changes Jun 27, 2024

View reviewed changes

Larsen, Steffen added 2 commits June 27, 2024 06:09

Add noexcept(false) to dtors

24dea7a

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

Merge branch 'sycl' into steffen/virtual_mem_ext

f85146f

KseniyaTikhomirova approved these changes Jun 27, 2024

View reviewed changes

Steffen Larsen added 2 commits July 1, 2024 12:28

Merge branch 'sycl' into steffen/virtual_mem_ext

6df0773

Add physical_mem dtor body to avoid RHEL issues

0bd2ce3

Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>

gmlueck mentioned this pull request Jul 15, 2025

[SYCL][Graph] async_malloc use allocation size for zeVirtualMemQueryPageSize #19402

Merged

Conversation

steffenlarsen commented Apr 5, 2023

Uh oh!

Uh oh!

steffenlarsen commented Apr 5, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

EwanC left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gmlueck left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AllanZyne commented Jul 16, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

16 participants