From 17fac40bae72b32a32fa98aa5059189b08512fe2 Mon Sep 17 00:00:00 2001
From: Manjunath-Jakaraddi <manjunath.jakaraddi@amd.com>
Date: Mon, 14 Oct 2024 16:45:03 -0500
Subject: [PATCH] SWDEV-481162: Updating MfmaUtil metric RDC

Change-Id: I60efa183edc14b6f870f7b6a82f223ea2c9789e5


[ROCm/rocprofiler commit: f84ecfe99b0a3fd181c4fce38b66bfbb6523ebf4]
---
 .../src/core/counters/metrics/derived_counters.xml            | 4 ++--
 projects/rocprofiler/test/tool/metrics.xml                    | 3 ++-
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/projects/rocprofiler/src/core/counters/metrics/derived_counters.xml b/projects/rocprofiler/src/core/counters/metrics/derived_counters.xml
index 8ac94080d4..0b812ff2bc 100755
--- a/projects/rocprofiler/src/core/counters/metrics/derived_counters.xml
+++ b/projects/rocprofiler/src/core/counters/metrics/derived_counters.xml
@@ -259,7 +259,7 @@
   <metric name="ScaPipeIssueUtil"                                       expr=100*SQ_ACTIVE_INST_SCA/(GRBM_GUI_ACTIVE*CU_NUM)          descr="Unit: percent"></metric>
   <metric name="ValuPipeIssueUtil"                                      expr=100*SQ_ACTIVE_INST_VALU/(GRBM_GUI_ACTIVE*CU_NUM)         descr="Unit: percent"></metric>
   <metric name="VmemPipeIssueUtil"                                      expr=100*4*(SQ_ACTIVE_INST_VMEM+SQ_ACTIVE_INST_FLAT)/(GRBM_GUI_ACTIVE*CU_NUM)      descr="Unit: percent"></metric>
-  <metric name="MfmaUtil"                                               expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)                       descr="Unit: percent"></metric>
+  <metric name="MfmaUtil"                                               expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)               descr="The percentage of kernel's duration, the MFMA unit was busy executing instructions"></metric>
   <metric name="AvgNumActiveThreads"                                    expr=SQ_THREAD_CYCLES_VALU/SQ_ACTIVE_INST_VALU                descr="Unit: percent"></metric>
   <metric name="VmemLatency"                                            expr=SQ_ACCUM_PREV_HIRES/SQ_INSTS_VMEM                        descr="Unit: cycles"></metric>
   <metric name="SmemLatency"                                            expr=SQ_ACCUM_PREV_HIRES/SQ_INSTS_SMEM_NORM                   descr="Unit: cycles"></metric>
@@ -326,7 +326,7 @@
   <metric name="ScaPipeIssueUtil"                                       expr=100*SQ_ACTIVE_INST_SCA/(GRBM_GUI_ACTIVE*CU_NUM)          descr="Unit: percent"></metric>
   <metric name="ValuPipeIssueUtil"                                      expr=100*SQ_ACTIVE_INST_VALU/(GRBM_GUI_ACTIVE*CU_NUM)         descr="Unit: percent"></metric>
   <metric name="VmemPipeIssueUtil"                                      expr=100*4*(SQ_ACTIVE_INST_VMEM+SQ_ACTIVE_INST_FLAT)/(GRBM_GUI_ACTIVE*CU_NUM)      descr="Unit: percent"></metric>
-  <metric name="MfmaUtil"                                               expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)                       descr="Unit: percent"></metric>
+  <metric name="MfmaUtil"                                               expr=100*XCC_NUM*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)               descr="The percentage of kernel's duration, the MFMA unit was busy executing instructions"></metric>
   <metric name="AvgNumActiveThreads"                                    expr=SQ_THREAD_CYCLES_VALU/SQ_ACTIVE_INST_VALU                descr="Unit: percent"></metric>
 
   ## Local Data Share (LDS) Metrics
diff --git a/projects/rocprofiler/test/tool/metrics.xml b/projects/rocprofiler/test/tool/metrics.xml
index 3f3fa0cd93..a8e75d445d 100755
--- a/projects/rocprofiler/test/tool/metrics.xml
+++ b/projects/rocprofiler/test/tool/metrics.xml
@@ -224,6 +224,7 @@
   <metric name="ACTIVE_WAVES"                                           expr=SQ_WAVES descr="Active Waves"></metric>
   <metric name="BANDWIDTH_EA"                                           expr=1024*(FETCH_SIZE+WRITE_SIZE)/GRBM_GUI_ACTIVE descr="Memory Bandwidth measured at the TCC_EA interface. In units of bytes/cycle."></metric>
   <metric name="OccupancyPercent"                                       expr=400*SQ_WAVE_CYCLES/GRBM_GUI_ACTIVE/CU_NUM/32 descr="GPU occupancy as % of maximum."></metric>
+  <metric name="MfmaUtil"                                               expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)               descr="The percentage of kernel's duration, the MFMA unit was busy executing instructions"></metric>
 </gfx90a_expr>
 
 <gfx940_expr>
@@ -363,7 +364,7 @@
   <metric name="BANDWIDTH_EA"                                           expr=1024*(FETCH_SIZE+WRITE_SIZE)*XCC_NUM/GRBM_GUI_ACTIVE descr="Memory Bandwidth measured at the TCC_EA interface. In units of bytes/cycle."></metric>
   <metric name="OccupancyPercent"                                       expr=400*SQ_WAVE_CYCLES*XCC_NUM/GRBM_GUI_ACTIVE/CU_NUM/32 descr="GPU occupancy as % of maximum."></metric>
   <metric name="GPU_UTIL"                                               expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>  
-  <metric name="MfmaUtil"                                               expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4) descr="Unit: percent"></metric>
+  <metric name="MfmaUtil"                                               expr=100*XCC_NUM*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4)               descr="The percentage of kernel's duration, the MFMA unit was busy executing instructions"></metric>
   <metric name="FP64_ACTIVE"                                            expr=TOTAL_64_OPS/GRBM_GUI_ACTIVE descr="The ratio of total floating point 64 bit ops / total number of cycles across all XCCs."></metric>
   <metric name="ENGINE_ACTIVE"                                          expr=GPU_UTIL/100 descr="Ratio between 0-1 of the time the GPU is active"></metric>
   <metric name="TENSOR_ACTIVE"                                          expr=MfmaUtil descr="Tensor core active in percent, identical to MfmaUtil"></metric>