[Issue]: hipblaslt offline tuning not picking up right kernels

[SKPsanjeevi]

Problem Description

Let's say we followed hipblaslt offline tuning. There are two kernels in the workload:

hipblaslt-bench --api_method c -m 4096 -n 24576 -k 1024 --lda 1024 --ldb 1024 --ldc 4096 --ldd 4096  --stride_a 0 --stride_b 0 --stride_c 0 --stride_d 0  --alpha 1.000000 --beta 0.000000 --transA T --transB N --batch_count 1 --scaleA 1 --scaleB 1 --bias_vector --bias_source d --a_type f8_r --b_type f8_r --c_type f16_r --d_type f16_r --scale_type f32_r --bias_type f16_r   --compute_type f32_r --activation_type none

hipblaslt-bench --api_method c -m 4096 -n 24576 -k 1024 --lda 1024 --ldb 1024 --ldc 4096 --ldd 4096  --stride_a 0 --stride_b 0 --stride_c 0 --stride_d 0  --alpha 1.000000 --beta 0.000000 --transA T --transB N --batch_count 1 --scaleA 1 --scaleB 1  --a_type f8_r --b_type bf8_r --c_type f16_r --d_type f16_r --scale_type f32_r --bias_type f32_r   --compute_type f32_r --activation_type none

These kernels nearly have dimensions and most parameters except for items highlighted in bold. Due to the similarity, the runtime does not pick the solution indices from the tuning.txt file. However, when one of the tuned solution is deleted, a right solution index is picked up.

Operating System

Ubuntu 22.04.5 LTS (Jammy Jellyfish)

CPU

Intel(R) Xeon(R) Platinum 8480C

GPU

Other

Other

MI325X

ROCm Version

ROCm 6.2.3

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response