AI/rocm-systems
2
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

MI350 Fix L2 cache to HBM read counters/metrics (#2501)

Browse Source 
* Fix rocprofiler-sdk metrics definition

* Use TCC_EA0_RDREQ_128B instead of TCC_BUBBLE counter for L2 cache to
  HBM counters and metrics

* Update MI350 counter definitions
    * FETCH_SIZE
    * BANDWIDTH_EA

* Update MI350 metrics definitions
    * System Speed of Light, L2-Fabric Read BW
    * Roofline Plot Points, AI (Arithmetic Intensity) HBM
    * Roofline Performance Rates, HBM Bandwidth

* Remove redundant definition for gfx950 and fix BANDWIDTH_EA definition

Test HBM bandwidth metric for memcopy workload

* Add memcopy.cpp workload

* Add metric validation test suite to validate HBM Bandwidth metric for
  memcopy workload

* Move gpu_soc() to test_utils.py for better re-usability

* Update TUI analysis config

* Fix hbm bandwidth formula for mi350 in calc_ai_profile

Co-authored-by: Alysa Liu <Alysa.Liu@amd.com>
This commit is contained in:
vedithal-amdandAlysa Liu authored and GitHub committed 2026-01-23 15:56:24 -05:00
1 parent 809eca7616
commit aa5dfb98f9
24 files changed
+528 -223
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

projects/rocprofiler-sdk/source/share/rocprofiler-sdk/counter_defs.yaml
+8 -2
View File
@@ -588,8 +588,10 @@ rocprofiler-sdk:
- gfx940
- gfx941
- gfx942
- gfx950
expression: (TCC_BUBBLE_sum*128 + (TCC_EA0_RDREQ_sum-TCC_BUBBLE_sum-TCC_EA0_RDREQ_32B_sum)*64 + TCC_EA0_RDREQ_32B_sum*32)/1024
- architectures:
- gfx950
expression: (TCC_EA0_RDREQ_128B_sum*128+TCC_EA0_RDREQ_64B_sum*64+TCC_EA0_RDREQ_32B_sum*32)/1024
- architectures:
- gfx10
- gfx1010
@@ -615,12 +617,16 @@ rocprofiler-sdk:
- architectures:
- gfx90a
expression: 1024*(WRITE_SIZE+FETCH_SIZE)/reduce(GRBM_GUI_ACTIVE,max)
# Approximation which excludes TCC_EA0_RDREQ_32B due to the hardware limitation on the number of counters that can be collected.
- architectures:
- gfx940
- gfx941
- gfx942
expression: (WRITE_SIZE*1024+TCC_BUBBLE_sum*128+(TCC_EA0_RDREQ_sum-TCC_BUBBLE_sum)*64)/reduce(GRBM_GUI_ACTIVE,max)
# Approximation which excludes TCC_EA0_RDREQ_32B due to the hardware limitation on the number of counters that can be collected.
- architectures:
- gfx950
expression: (WRITE_SIZE*1024+TCC_BUBBLE_sum*128+(TCC_BUBBLE_sum-TCC_EA0_RDREQ_sum)*64)/reduce(GRBM_GUI_ACTIVE,max)
expression: (WRITE_SIZE*1024+TCC_EA0_RDREQ_128B_sum*128+TCC_EA0_RDREQ_64B_sum*64)/reduce(GRBM_GUI_ACTIVE,max)
- name: FetchSize
description: The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache
or memory effects taken into account.