Suggested perf improvements

[CoffeeBeforeArch]

GPGPU-Sim contains many asserts (and conditional checks) that are primarily used for debugging. As these checks never fail for most (if not all simulations run today), I propose a third mode of compilation be added to GPGPU-Sim (one that focuses on the speed of the simulation when benchmarks already pass).

Let's consider a brief hotspot analysis of the code from Linux perf tools from running some matrix multiplication code:

Samples: 350K of event 'cycles:ppp', Event count (approx.): 267406121728
Overhead  Command          Shared Object        Symbol
  13.27%  mmul             libcudart.so         [.] cache_stats::operator()
   7.36%  mmul             libcudart.so         [.] tag_array::probe
   5.90%  mmul             libcudart.so         [.] cache_stats::operator+=
   5.20%  mmul             libcudart.so         [.] sector_cache_block::is_reserved_line
   3.92%  mmul             libcudart.so         [.] ptx_thread_info::get_reg
   3.50%  mmul             libcudart.so         [.] Scoreboard::checkCollision
   2.86%  mmul             libcudart.so         [.] pipelined_simd_unit::cycle
   2.49%  mmul             libcudart.so         [.] simt_stack::get_pdom_stack_top_info
...

13.27% of the time is spent updating cache stats (this breakdown is consistent for tests running from 1 second to 10 minutes, and likely fully scaled sims). This is partially the result of branch-heavy code generated by:

  if (fail_outcome) {
    if (!check_fail_valid(access_type, access_outcome))
      assert(0 && "Unknown cache access type or fail outcome");

    return m_fail_stats[access_type][access_outcome];
  } else {
    if (!check_valid(access_type, access_outcome))
      assert(0 && "Unknown cache access type or access outcome");

    return m_stats[access_type][access_outcome];
  }

Where check_fail_valid and check_valid both contain additional branches.

Ignoring the fact that accesses to m_stats and m_fail_stats miss ~100% of the time, just removing the unnecessary branches resulted in a 5-15% speedup.

Perhaps it makes sense to have a performance build option that uses the preprocessor to select between versions of these functions (such as the modified version I have below). This PR just contains the modified function and would be edited to include the new build mode if people agree with this approach.

Food for thought.

[CoffeeBeforeArch]

Interestingly enough, -DNDEBUG removes the asserts, but the compiler still isn't smart enough to remove the call to the check_valid and check_fail_valid functions.

[CoffeeBeforeArch]

Some perf results for longer running apps:

• LavaMD
  • 10:35:45 -> 8:31:19
• Srad_v1
  • 8:46:32 -> 8:02:11
• Kmeans
  • 7:20:41 -> 6:15:42