xe: gemm: fix stride interface

[rjoursler]

Avoids incorrect offset calculations for batched GEMM. Fixes the following issue reported in MFDNN-14479

$ ./tests/benchdnn/benchdnn --matmul --engine=gpu --dt=f32:f32:f32 --stag=acbd --wtag=adbc --dtag=abcd --attr-post-ops=binary_mul:f32:0 --attr-scratchpad=user 4x3x16413x16:4x3x16x16413
Segmentation fault from GPU at 0xff0000000000f000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 1 (Write), banned: 1, aborting.
Segmentation fault from GPU at 0xff0000000000f000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 1 (Write), banned: 1, aborting.
Abort was called at 306 line in file:
./shared/source/os_interface/linux/drm_neo.cpp
Aborted (core dumped) 

[rjoursler]

make test
disable test_device_cpu
disable build_cpu_runtime_omp
disable build_cpu_runtime_sycl
disable build_cpu_runtime_tbb

[rjoursler]

make test_ov

[echeresh]

This is now 64-bit but the kernel interface uses 32-bit. I guess we don't have any type checks but theoretically that could lead to a similar overflow issue.

[h-sadia]

Could this be true for stride_binary? https://github.com/uxlfoundation/oneDNN/blob/main/src/gpu/intel/gemm/jit.cpp#L238

[rjoursler]

Good points, I will try updating these as well and check if there are any issues.

[Simonsays095]

Could we do something similar to what we do with the opencl dispatcher, and use 64-bit strides/dimensions only if required by the problem?

[rjoursler]

Could we do something similar to what we do with the opencl dispatcher, and use 64-bit strides/dimensions only if required by the problem?

We definitely can, but I only intend to do this if we encounter a performance regression. If the offset calculations is not important for performance (which is generally the case for GEMM), then there is no benefit to adding this control, it only creates an extra point of failure.

[rjoursler]

make test
disable test_device_cpu
disable build_cpu_runtime_omp
disable build_cpu_runtime_sycl
disable build_cpu_runtime_tbb

[rjoursler]

make test_ov