[iq136boy]

Motivation

MIOpen has been incrementally retiring its OpenCL tensor operation solvers in favour of HIP equivalents to reduce maintenance overhead, eliminate the OpenCL compiler dependency, and align with ROCm's HIP-first strategy. This is the second PR in that series (following PR #6935 which retired Op2dTensorLite, Op4dTensorGeneric, and OpTensorFwdBias).

This PR covers the two remaining OCL tensor solvers that had no HIP kernel yet:

• OpTensorLeadingOnes – 4-D tensor ops for packed and non-packed (generic stride) paths
• Op5dTensorGeneric – generic fallback for all 5-D tensor element-wise ops

Technical Details

Three new HIP kernels are appended to src/kernels/MIOpenTensorKernelsHip.cpp:

• OpTensorLeadingOnes – ports OCL OpTensorLeadingOnes (line ~218), compiled under USE_LEADING_ONES
• OpTensorLeadingOnesGeneric – ports OCL OpTensorLeadingOnesGeneric (line ~310), compiled under USE_LEADING_ONES_GENERIC
• Op5dTensorGeneric – ports OCL Op5dTensorGeneric (line ~544), compiled under USE_5D_TENSOR_GENERIC

All kernels are faithful ports of the OCL originals with standard syntax substitutions (__kernel → global, get_global_id(0) → blockIdx.x * blockDim.x + threadIdx.x, barrier(CLK_LOCAL_MEM_FENCE) →
__syncthreads(), (long) → (int64_t), etc.).

Solver-side changes are minimal (2–3 lines each):

• src/solver/tensorOp/OpTensorLeadingOnes.cpp – switch kbp::OpenCL{} → kbp::HIP{}, kernel file to MIOpenTensorKernelsHip.cpp, and flip the Get4dParams is_hip flag to true.
• src/solver/tensorOp/Op5dTensorGeneric.cpp – switch kbp::OpenCL{} → kbp::HIP{} and kernel file to MIOpenTensorKernelsHip.cpp.
  No header, solver registration, or CMakeLists.txt changes are required. MIOpenTensorKernels.cl is not deleted as it is still referenced by Op1dTensorGeneric, Op2dTensorGeneric, Op2dTensorSquash, and Op3dTensorGeneric.

Test Plan

• Build MIOpen to confirm no compile errors.
• Verify the three new HIP kernels compile cleanly with their respective #ifdef guards (USE_LEADING_ONES, USE_LEADING_ONES_GENERIC, USE_5D_TENSOR_GENERIC).
• Confirm neither modified solver file references MIOpenTensorKernels.cl after the change.
• Run existing MIOpen tensor op unit tests covering 4-D non-packed layouts and 5-D cases.
• Full CI pipeline.

Test Result

• Build passes with no compile errors.
• All existing tensor op tests pass (4-D packed, 4-D non-packed/generic stride, 5-D generic).
• CI pending.

Submission Checklist

• New HIP kernels are logically equivalent to the OCL originals (same index arithmetic, same bitmap logic, same beta==0 / beta!=0 branches).
• No unintended changes to solver headers, CMakeLists, or solver registration.
• MIOpenTensorKernels.cl retained (still needed by other solvers not in scope for this PR).