SPIRVRunner: Investigate and handle multi-kernel execution (benchmark- gemm_streamk_benchmark.py)

[kballeda]

Describe the bug

Benchmarks such as - benchmark/triton_kernels_benchmark/gemm_streamk_benchmark.py uses two kernels as part of the execution. Investigate this scenario and if possible improve the SPIRVRunner to handle such cases.

     _, min_ms, max_ms, mean_ms, cv = benchmark_suit.do_bench(triton_fn, n_warmup=10, n_repeat=10,
                                                                 quantiles=quantiles,
                                                                 kernel_name=['first_wave', 'full_tiles'])

Environment details

This is an enhancement request. Triton: main, GPU: Max 1100.

[alexbaden]

Each kernel should dump to a different file, and the SPIRV runner can be run twice - once with the inputs from the first kernel and its args and SPIRV file, and once with the inputs from the second kernel and its args and SPIRV file. The runner was never designed to run both kernels at the same time and I don't see the need to add that functionality.