Add onnxruntime #19

smanolloff · 2025-11-23T09:53:26Z

This PR contains changes needed to migrate MMAI from libtorch to onnxruntime as @Laserlicht suggested here

It adds onnxruntime/1.18.1 requirement with several patches address android-arm64 and windows-x86 compilation errors, as well as conan recipe errors from transitive dependencies.

Please take a look and let me know if you have any comments.

The changes were tested with VCMI (develop branch) on MacOS 15 (arm), Windows 10 (x64), iOS 18 and Android 13 without noticing any issue.

This reverts commit 56f3512. Removing CreateFile2 calls was insufficient to enable win7 support. There are win8+ API calls which are not straightforward to replace: onnxruntime\core\platform\windows\env_time.cc(31): error C2065: 'GetSystemTimePreciseAsFileTime': undeclared identifier

smanolloff · 2025-11-25T00:46:23Z

Added conan patch which allows to keep macos-intel 10.13 and ios 12.0 min OS versions.
Unfortunately, my attempt at compiling onnxruntime on win7 was not successful: there are other calls such as GetSystemTimePreciseAsFileTime which are not available in the windows 7 API and are seem more difficult to replace. So the win7 / win8 build split remains.

GeorgeK1ng · 2025-11-25T08:00:23Z

I think there is issue with FILE_READ_ATTRIBUTES vs GENERIC_READ in W7 patch?

Also keeping 192 which means v142 toolset is needed to keep proper support for builded components for Windows 7.

smanolloff · 2025-11-26T00:22:22Z

I think there is issue with FILE_READ_ATTRIBUTES vs GENERIC_READ in W7 patch?

A hasty copy-paste on my side it seems, but that's irrelevant as I reverted the commit due to other compile errors.

Also keeping 192 which means v142 toolset is needed to keep proper support for builded components for Windows 7.

Thanks for pointing that out, I was under the impression "if it compiles, it runs" :) The compiler version is back to 192 for windows 7 now.

GeorgeK1ng · 2025-11-26T18:43:42Z

Seems your patches were almost close. This should be Windows 7 compatible patch https://github.com/GeorgeK1ng/vcmi-dependencies/actions/runs/19713613028 - compiled just fine. So you need to grab there Windows 7 patch and recipe ones to allow v142 toolset.

smanolloff · 2025-11-27T11:29:07Z

@GeorgeK1ng Thank you! I have added your suggested patch, finally the win7/win8 split is removed - the last remaining conan change in the PR has been dealt with.

kambala-decapitator

please also update PR description to match the current state

kambala-decapitator · 2025-11-29T09:56:23Z

.github/workflows/rebuildDependencies.yml


    - name: Build recipes with our patches
      run: |
+        set -x


don't forget to remove

Sure, I will remove it. But if I may I ask: why? Don't you find it helpful for troubleshooting?

kambala-decapitator · 2025-11-29T10:04:25Z

.github/workflows/rebuildDependencies.yml


+        # Patch conan recipe to prevent dragging libcurl as a dependency
+        # (can't be done via VCMI's conanfile as onnxruntime is built beforehand)
+        git apply --ignore-whitespace ../conan_patches/onnxruntime/recipe.diff


you can pass date/*:use_system_tz_db as an option to conan create or simply place it in our profile, to common I guess

The conan recipe also contains this change:

def _compilers_minimum_version(self): return { "Visual Studio": "17", - "msvc": "193", + "msvc": "192", "gcc": "9",

It can't be set through command-line options AFAIK, which means the patch remains anyway. Does it still make sense to extract the date/*:use_system_tz_db into a cmdline option and wouldn't that make it harder to follow?

kambala-decapitator · 2025-11-29T10:05:01Z

.github/workflows/rebuildDependencies.yml


          if [[ $package == qt ]] ; then
            packagePath="$recipePathQt/5.x.x"
+          elif [[ $package =~ ^onnx ]] && ${{ contains(matrix.conan_options, 'with_onnxruntime=False') }} ; then


but I don't see such option in the matrix?

Should I remove the qt check then as well?

kambala-decapitator · 2025-11-29T10:07:50Z

.github/workflows/rebuildDependencies.yml

+
+        # Some are both build and host requirements and must not be removed
+        # => exclude them from the list
+        cat $buildPackagesFile $hostPackagesFile | jq -s '


unclear why this is needed, no info in PR description

If a package is both a build and a host requirement, the old logic removes it. This leads to Missing prebuilt package error because conan is invoked with build=never in the VCMI build. In this case, zlib and protobuf were build requirements and were removed, which made them "missing" during VCMI build. This graph might help:
graph-msvc-x64.html

kambala-decapitator · 2025-11-29T10:10:23Z

conan_patches/onnxruntime/patches/msvc-x86.diff

+
+ #Set global compile flags for all the source code(including third_party code like protobuf)
+ #This section must be before any add_subdirectory, otherwise build may fail because /MD,/MT mismatch
+message(STATUS "MSVC=${MSVC} | CMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} | CMAKE_C_COMPILER_ARCHITECTURE_ID=${CMAKE_C_COMPILER_ARCHITECTURE_ID} | CMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}")


don't forget to remove these debug messages

These are not forgotten, I have left them on purpose. Please see my reply to your next comment regarding x86 vs X86.

kambala-decapitator · 2025-11-29T10:11:02Z

conan_patches/onnxruntime/patches/msvc-x86.diff

+     set(onnxruntime_target_platform "ARM")
+     enable_language(ASM_MARMASM)
+-  elseif (onnxruntime_target_platform STREQUAL "x64" OR onnxruntime_target_platform STREQUAL "x86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")
+  elseif (onnxruntime_target_platform MATCHES "[Xx]64" OR onnxruntime_target_platform MATCHES "[Xx]86_64" OR onnxruntime_target_platform STREQUAL "AMD64" OR CMAKE_GENERATOR MATCHES "Win64")


can it really start with capital X?

also not really clear why touching x64 if the patch is for x86

Apparently it can. Not sure where that comes from, but I find onnxruntime's target platform logic rather confusing. This is why I left the debug messages.

kambala-decapitator · 2025-11-29T10:14:22Z

conan_patches/onnxruntime/patches/win7.diff

+-      FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE,
+-      OPEN_EXISTING,
+-      &param)};
+  wil::unique_hfile file_handle{CreateFileW(


are these CreateFile2 -> CreateFileW changes worsening anything on Win 8+? If so, you could use runtime check for OS version or function availability

Nothing extra. Should be still same result.

kambala-decapitator · 2025-11-29T10:15:25Z

conan_patches/onnxruntime/conandata.yml

+    - patch_file: "patches/android-64.diff"
+      patch_description: |
+        Fixes error for arm64-v8a targets:
+        onnxruntime/core/mlas/lib/mlasi.h:366:11: error: unknown type name 'bfloat16_t'


and what's the reason? Is it about minimum API level?

I can't answer this question.
bfloat16 support seems introduced in Armv8.6-A in 2019, but onnxruntime/1.16.1 contains flags for armv8.2. This might cause issues on older chips where these optimized instructions are unavailable.

However, we should be safe. The MMAI models are not using bfloat16 arithmetic. I have tested the model with Samsung Galaxy A71 which has an ARM Cortex-A76 and Cortex-A55, both implementing Arm8.2-A and there have been no issues.

kambala-decapitator · 2025-11-29T10:16:00Z

conan_patches/onnxruntime/conandata.yml

+        onnxruntime/core/mlas/lib/mlasi.h:366:11: error: unknown type name 'bfloat16_t'
+    - patch_file: "patches/msvc-x86.diff"
+      patch_description: |
+        Fixes detection of x86 windows targets during cross-compilation.


does it make sense to upstream this patch? Or upstream dropped 32-bit support?

I would first wait to see if the models run on windows-x86 targets - I have no suitable hardware to test (honestly, I was surprised it even compiled. Hopefully it will run!)

IvanSavenko · 2025-12-04T18:44:26Z

@kambala-decapitator any further comments?

kambala-decapitator · 2025-12-05T21:10:54Z

will check over the weekend

onnxruntime

6506173

smanolloff mentioned this pull request Nov 23, 2025

MMAI: an ML-powered combat AI vcmi/vcmi#4788

Open

smanolloff force-pushed the mmai branch 4 times, most recently from 82dc03f to 6804927 Compare November 24, 2025 11:49

smanolloff added 3 commits November 25, 2025 02:38

macos-intel-10.13 and ios-12.0 compat

96f29dc

win7 compat

56f3512

smanolloff force-pushed the mmai branch from 6804927 to ef8b4b1 Compare November 25, 2025 00:41

lower compiler version for win7

75c2c1e

smanolloff force-pushed the mmai branch from 25f8e1c to 75c2c1e Compare November 26, 2025 00:15

smanolloff force-pushed the mmai branch 2 times, most recently from 0e42897 to 76afc17 Compare November 27, 2025 09:33

try suggested patch from PR review

80b722a

smanolloff force-pushed the mmai branch from 76afc17 to 80b722a Compare November 28, 2025 11:22

kambala-decapitator reviewed Nov 29, 2025

View reviewed changes

address PR comment

2e849f7

Add onnxruntime #19

Are you sure you want to change the base?

Add onnxruntime #19

Uh oh!

Conversation

smanolloff commented Nov 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

smanolloff commented Nov 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

GeorgeK1ng commented Nov 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

smanolloff commented Nov 26, 2025

Uh oh!

GeorgeK1ng commented Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

smanolloff commented Nov 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kambala-decapitator left a comment

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!

smanolloff Nov 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

smanolloff Nov 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

Choose a reason for hiding this comment

Uh oh!

IvanSavenko commented Dec 4, 2025

Uh oh!

kambala-decapitator commented Dec 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

smanolloff commented Nov 23, 2025 •

edited

Loading

smanolloff commented Nov 25, 2025 •

edited

Loading

GeorgeK1ng commented Nov 25, 2025 •

edited

Loading

GeorgeK1ng commented Nov 26, 2025 •

edited

Loading

smanolloff commented Nov 27, 2025 •

edited

Loading

smanolloff Nov 29, 2025 •

edited

Loading

smanolloff Nov 29, 2025 •

edited

Loading