VML 64-bit API Bug Report

[meding]

The VML 64-bit API appears to internally dispatch to the VML 32-bit API. When the vector length parameter exceeds `INT_MAX`, it crashes. Defining `-DMKL_MKL_ILP64` does not help. Below is a minimum reproducible example.

#include <mkl_vml.h>
#include <mkl_service.h>
#include <vector>

int main()
{
    mkl_set_interface_layer(MKL_INTERFACE_ILP64); // Does not resolve crash

    // auto constexpr N = int64_t{INT_MAX}; // Works
    auto constexpr N = int64_t{INT_MAX} + int64_t{1}; // Crashes
    auto constexpr k = float{1};

    auto xs = std::vector<float>(N);
    vsAddI_64(N, xs.data(), 1, &k, 0, xs.data(), 1);
}

 `__INTEL_MKL_BUILD_DATE` is `20260401`
Linking against `mkl_rt.3.dll`

Stacktrace

1 vmldLinearFracI_26                           mkl_vml_avx2_3       0x7ff97a9eee14
2 mkl_vml_serv_threader_s_2iI_1oI mkl_intel_thread_3   0x7ff982641ec6
3 vsAddI_64                                            mkl_rt_3                      0x7ff98610536e
4 main                                                      main.cxx                37  0x7ff67466182a
5 invoke_main                                        exe_common.inl 79   0x7ff6746637f9
6 __scrt_common_main_seh              exe_common.inl 288 0x7ff6746636a2
7 __scrt_common_main                      exe_common.inl 331 0x7ff67466355e
8 mainCRTStartup                                exe_main.cpp       17   0x7ff67466388e
9 BaseThreadInitThunk                       KERNEL32                    0x7ffa5bf8e957
10 RtlUserThreadStart                         ntdll                               0x7ffa5d607c1c

Processor 13th Gen Intel(R) Core(TM) i5-13420H (2.10 GHz)
Installed RAM 16.0 GB (15.7 GB usable)
Graphics card NVIDIA GeForce RTX 3050 6GB Laptop GPU (6 GB)
Intel(R) UHD Graphics (128 MB)
Storage 258 GB of 477 GB used
System type 64-bit operating system, x64-based processor

[Fengrui]

Hi,

Thanks for reporting this bug.

I can reproduce it. It looks it only happens when incb=0. Let me escalate it internally and I will get back to you soon.

Thanks,

Fengrui

[davidwarner344]

You should report this to Intel. Your reproducer is minimal and clearly shows that vsAddI_64() crashes when N > INT_MAX, even with the ILP64 interface enabled, which shouldn't happen if 64-bit vector lengths are fully supported.

In the meantime, as a workaround, split the operation into chunks no larger than INT_MAX elements until Intel confirms whether this is a bug or provides a fix.

[AlHill]

@davidwarner344     "You should report this to Intel."   Is that not what the op did by posting here?  I love these non-answers/responses from new users who have established no credibility.

Doc (not an Intel employee or contractor)
[Success breeds complacency. Complacency breeds failure. (Andy Grove)]

[william1234]

This looks like an MKL bug. vsAddI_64 should support 64-bit lengths, so crashing at INT_MAX + 1 suggests an internal 32-bit limitation. I'd report it to Intel with your reproducer and stack trace.

[meding]

For those saying that I should report this to Intel, that is what the original post was for. If I try to use "Development Software Support" or "Contact Support", I get rejected for not being a paying customer. If there is a proper channel to report bugs, Intel has not designed their website in an intuitive manner to do so.

[cartergray70543]

This looks like a bug in MKL. Since vsAddI_64() crashes only when N > INT_MAX even with ILP64 enabled, it suggests an internal 32-bit limitation. I'd report it to Intel with your minimal reproducer and build details. For now, splitting the work into chunks smaller than INT_MAX is the safest workaround.

[masonreed11]

It looks like a bug in the current MKL release. Since the crash only happens when N > INT_MAX and ILP64 settings don't help, it suggests the call is still using a 32-bit path internally. Your minimal example is excellent—I'd submit it as a bug report so Intel can reproduce and investigate it quickly.