Details:

• Updated create_data to use the actual data size calculated from the original element type and shape for u16/i16 types, ensuring memcpy only reads the valid memory region.

Description of the issue

• The create_data function used the layout's buffer size for memcpy, which is incorrect for u16/i16 types as they are converted to f32 in the layout. This caused bufSize (4 bytes/elem) to be larger than the actual source data (2 bytes/elem), leading to a buffer overread and SEH exception 0xc0000005.

The code and line that caused this issue (if it is not changed directly)

Reproduction step and snapshot

• Test command

timeout 600 valgrind --tool=memcheck --leak-check=full --show-leak-kinds=all --track-origins=yes --verbose --error-exitcode=1 --log-file=valgrind_i16_test.log ./bin/intel64/Debug/ov_gpu_func_tests '--gtest_filter=*ConstantResultSubgraphTest.Inference/SubgraphType=SEVERAL_COMPONENT_IS=[1,3,10,10]_IT=i16_Device=GPU' > valgrind_console.txt

• Output

==630497== Invalid read of size 8
==630497==    by 0x485293F: memmove (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==630497==    by 0x190A8A20: ov::intel_gpu::create_data(...) (constant.cpp:128)
...
==630497==  Address 0xed432e8 is 16 bytes after a block of size 600 alloc'd

Checklist

• [v] Is it a proper fix? (not a workaround)
• [v] Did you include test case for this fix, if necessary?
• [v] Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• CVS-172561