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

Replicate global counters across all derived counters (#936)

Browse Source 
Fix derived counters to have globals replicated across all architectures (that support them).
---------

Co-authored-by: Benjamin Welton <ben@amd.com>

[ROCm/rocprofiler-sdk commit: ab92bef1a4]
This commit is contained in:
Benjamin Welton
2024-06-18 23:40:19 -07:00
committed by GitHub
parent a7d7d47576
commit 88f277107e
4 changed files with 113 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.cpp
+1 -1
View File
@@ -101,7 +101,7 @@ loadXml(const std::string& filename, bool load_constants = false)
* respec the XML (which we should...).
*/
if(gfx_name.find("metric") == std::string::npos ||
gfx_name.find("top.") == std::string::npos)
gfx_name.find("top.") == std::string::npos || gfx_name.find("gfx") == std::string::npos)
continue;
auto& metricVec =
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/metrics_test.cpp
+1 -1
View File
@@ -171,7 +171,7 @@ TEST(metrics, check_agent_valid)
if(other_gfx == gfx) continue;
for(const auto& metric : other_counters)
{
if(common_metrics.count(metric.id())) continue;
if(common_metrics.count(metric.id()) || !metric.special().empty()) continue;
EXPECT_EQ(counters::checkValidMetric(gfx, metric), false)
<< fmt::format("GFX {} has Metric {} but shouldn't", gfx, metric);
}
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/metrics_test.h
+48 -2
View File
@@ -166,7 +166,52 @@ static const std::unordered_map<std::string, std::vector<std::vector<std::string
static const std::unordered_map<std::string, std::vector<std::vector<std::string>>> derived_gfx908 =
{{"gfx908",
{{"GPU_UTIL",
{{"GPUBusy",
"",
"",
"100*GRBM_GUI_ACTIVE/GRBM_COUNT",
"The percentage of time GPU was busy."},
{"Wavefronts", "", "", "SQ_WAVES", "Total wavefronts."},
{"VALUInsts",
"",
"",
"SQ_INSTS_VALU/SQ_WAVES",
"The average number of vector ALU instructions executed per work-item (affected by flow "
"control)."},
{"SALUInsts",
"",
"",
"SQ_INSTS_SALU/SQ_WAVES",
"The average number of scalar ALU instructions executed per work-item (affected by flow "
"control)."},
{"SFetchInsts",
"",
"",
"SQ_INSTS_SMEM/SQ_WAVES",
"The average number of scalar fetch instructions from the video memory executed per "
"work-item (affected by flow control)."},
{"GDSInsts",
"",
"",
"SQ_INSTS_GDS/SQ_WAVES",
"The average number of GDS read or GDS write instructions executed per work item "
"(affected by flow control)."},
{"MemUnitBusy",
"",
"",
"100*reduce(TA_TA_BUSY,max)/GRBM_GUI_ACTIVE/SE_NUM",
"The percentage of GPUTime the memory unit is active. The result includes the stall "
"time (MemUnitStalled). This is measured with all extra fetches and writes and any "
"cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound)."},
{"ALUStalledByLDS",
"",
"",
"400*SQ_WAIT_INST_LDS/SQ_WAVES/GRBM_GUI_ACTIVE",
"The percentage of GPUTime ALU units are stalled by the LDS input queue being full or "
"the output queue being not ready. If there are LDS bank conflicts, reduce them. "
"Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% "
"(optimal) to 100% (bad)."},
{"GPU_UTIL",
"",
"",
"100*GRBM_GUI_ACTIVE/GRBM_COUNT",
@@ -175,7 +220,8 @@ static const std::unordered_map<std::string, std::vector<std::vector<std::string
"",
"",
"reduce(SQ_WAVES,sum)",
"Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."},
"Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ "
"instances."},
{"TCC_HIT_sum",
"",
"",
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/xml/derived_counters.xml
+63 -64
View File
@@ -1,4 +1,63 @@
<gfx8>
<common_derived>
# GPUBusy The percentage of time GPU was busy.
<metric
name="GPUBusy"
descr="The percentage of time GPU was busy."
expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT
></metric>
# Wavefronts Total wavefronts.
<metric
name="Wavefronts"
descr="Total wavefronts."
expr=SQ_WAVES
></metric>
# VALUInsts The average number of vector ALU instructions executed per work-item (affected by flow control).
<metric
name="VALUInsts"
descr="The average number of vector ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_VALU/SQ_WAVES
></metric>
# SALUInsts The average number of scalar ALU instructions executed per work-item (affected by flow control).
<metric
name="SALUInsts"
descr="The average number of scalar ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_SALU/SQ_WAVES
></metric>
# SFetchInsts The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control).
<metric
name="SFetchInsts"
descr="The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control)."
expr=SQ_INSTS_SMEM/SQ_WAVES
></metric>
# GDSInsts The average number of GDS read or GDS write instructions executed per work item (affected by flow control).
<metric
name="GDSInsts"
descr="The average number of GDS read or GDS write instructions executed per work item (affected by flow control)."
expr=SQ_INSTS_GDS/SQ_WAVES
></metric>
# MemUnitBusy The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound).
<metric
name="MemUnitBusy"
descr="The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound)."
expr=100*reduce(TA_TA_BUSY,max)/GRBM_GUI_ACTIVE/SE_NUM
></metric>
# ALUStalledByLDS The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad).
<metric
name="ALUStalledByLDS"
descr="The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad)."
expr=400*SQ_WAIT_INST_LDS/SQ_WAVES/GRBM_GUI_ACTIVE
></metric>
</common_derived>
<gfx8 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
@@ -33,7 +92,7 @@
<metric name="LDSBankConflict" expr=100*SQ_LDS_BANK_CONFLICT/GRBM_GUI_ACTIVE/CU_NUM descr="The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad)."></metric>
</gfx8>
<gfx9>
<gfx9 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
@@ -428,7 +487,7 @@
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
</gfx940>
<gfx10>
<gfx10 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="MeanOccupancyPerCU" expr=GRBM_COUNT*0+SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV/GRBM_GUI_ACTIVE/CU_NUM descr="Mean occupancy per compute unit."></metric>
<metric name="MeanOccupancyPerActiveCU" expr=GRBM_COUNT*0+SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV*4/SQ_BUSY_CYCLES/CU_NUM descr="Mean occupancy per active compute unit."></metric>
@@ -476,7 +535,7 @@
<gfx1032 base="gfx10">
</gfx1032>
<gfx11>
<gfx11 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
<metric name="WAVE_DEP_WAIT" expr=100*SQ_WAIT_ANY/SQ_WAVE_CYCLES descr="Percentage of the SQ_WAVE_CYCLE time spent waiting for anything."></metric>
@@ -517,63 +576,3 @@
<gfx942 base="gfx940"></gfx942>
#Navi21
<gfx1032 base="gfx1032"></gfx1032>
<global>
# GPUBusy The percentage of time GPU was busy.
<metric
name="GPUBusy"
descr="The percentage of time GPU was busy."
expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT
></metric>
# Wavefronts Total wavefronts.
<metric
name="Wavefronts"
descr="Total wavefronts."
expr=SQ_WAVES
></metric>
# VALUInsts The average number of vector ALU instructions executed per work-item (affected by flow control).
<metric
name="VALUInsts"
descr="The average number of vector ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_VALU/SQ_WAVES
></metric>
# SALUInsts The average number of scalar ALU instructions executed per work-item (affected by flow control).
<metric
name="SALUInsts"
descr="The average number of scalar ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_SALU/SQ_WAVES
></metric>
# SFetchInsts The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control).
<metric
name="SFetchInsts"
descr="The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control)."
expr=SQ_INSTS_SMEM/SQ_WAVES
></metric>
# GDSInsts The average number of GDS read or GDS write instructions executed per work item (affected by flow control).
<metric
name="GDSInsts"
descr="The average number of GDS read or GDS write instructions executed per work item (affected by flow control)."
expr=SQ_INSTS_GDS/SQ_WAVES
></metric>
# MemUnitBusy The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound).
<metric
name="MemUnitBusy"
descr="The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound)."
expr=100*reduce(TA_TA_BUSY,max)/GRBM_GUI_ACTIVE/SE_NUM
></metric>
# ALUStalledByLDS The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad).
<metric
name="ALUStalledByLDS"
descr="The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad)."
expr=100*SQ_WAIT_INST_LDS*4/SQ_WAVES/GRBM_GUI_ACTIVE
></metric>
</global>