Stream stats #71
Stream stats #71
Conversation
code changes for adding per-stream status.
Added optional arg to cache_stats::print_stats, cache_stats::print_fail_stats and their upstream functions. When streamID is specified, print stats from that stream. When not specified, print all stats. NOTE: current implementation depending on streamid never equals -1
Per-stream stats tracking feature
Multi-Stream stats
|
@JRPan let me know if you need something from UW here |
|
@JRPan just checking again if you need anything from us? |
|
Sorry for the delay, this will be included in the next major release. accel-sim/accel-sim-framework#317 |
|
@FJShen Lets try get this merged asap |
|
I need some more description of this PR. E.g., What feature does it add? What bug does it fix? Does "stream" refer to CUDA stream? |
@FJShen : @JRPan took the original commit my students submitted to Accel-Sim and packaged it into this. Yes, "stream" means "CUDA stream". We have a full tech report that describes the support we added: https://arxiv.org/abs/2304.11136 Ultimately, what these commits do is it adds support to Accel-Sim/GPGPU-Sim to track statistics at a per-CUDA stream level. |
|
Thank you, Matt! @FJShen For each kernel, a stream id is always given (0 by default). The stats container is usually a This mainly affects cache stats (L1 and L2), and cycles. Other stats like occupancy, DRAM, and IPC are not changed. These stats does not make sense to be separated. I correlated GPU ubench and rodinia. Functionally, this is correct. I assigned you as the reviewer to check more about the C++ side of this :) I did not correlate this with any workload that runs with concurrent enabled, tho. Maybe we should write a ubench. |
I believe we did correlate with concurrent streams in the report I linked, FWIW. |
|
I found some possibly erroneous code (and sub-optimal coding practice). Please kindly address the reviews above. |
|
I only see one comment related to the config file. Is this correct? |
|
@JRPan I forgot to "submit" the reviews. I hope they are now visible. |
* Temp commit for Justin and Cassie to sync on code changes for adding per-stream status. * Resolved compile errors. * Removed redundant parameter * Passed cuda_stream_id from accelsim to gpgpusim * Cleaned up unused changes * Changed vector to map, having operator problems. * StreamID defaults to zero * Implemented streams to inc_stats and so on * Fixed TOTAL_ACCESS counts * Implemented GLOBAL_TIMER. * Fixed m_shader->get_kernel SEGFAULT issue in shader.cc. * Use warp_init to track streamID instead of issue_warp * Removed temp debug print * Modified cache_stats to only print data from latest finished stream Added optional arg to cache_stats::print_stats, cache_stats::print_fail_stats and their upstream functions. When streamID is specified, print stats from that stream. When not specified, print all stats. NOTE: current implementation depending on streamid never equals -1 * Removed default arg values of streamID * modified constructor of mem_fetch to pass in streamID * changed get_streamid to get_streamID * Added TODO to gpgpusim_entrypoint.cc and power_stat.cc * Only collect power stats when enabled * print last finished stream in PTX mode using last_streamID * take out additional printf * Add a field to baseline cache to indicate cache level * save gpu object in cache * Print stream ID only once per kernel * rm test print * use -1 for default stream id * cleanup debug prints * remove GLOABL_TIMER * Automated clang-format * Should be correct to print everything in power model * addressing concerns & errors * Automated clang-format * add m_stats_pw in operator+ * Automated Format --------- Co-authored-by: Justin Qiao <[email protected]> Co-authored-by: Justin Qiao <[email protected]> Co-authored-by: Tim Rogers <[email protected]> Co-authored-by: JRPan <[email protected]> Co-authored-by: purdue-jenkins <[email protected]>
Collect stats based on streams.
99% Ready. Creating a PR to trigger GitHub Actions to generate correlations.
Consult @JRPan before merging.