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

Merge pull request #91 from AMDResearch/mi100-fix

Browse Source 
Custom metrics for mi100

[ROCm/rocprofiler-compute commit: ab30dd6a21]
This commit is contained in:
Cole Ramos
2023-02-21 10:10:24 -06:00
committed by GitHub
parent b0bcc846aa ecedfba0f0
commit a368456cd1
3 changed files with 933 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-compute/src/omniperf
+33 -16
View File
@@ -343,7 +343,7 @@ def characterize_app(path, cmd, verbose):
# Workload profiling
for fname in glob.glob(workload_dir + "/perfmon/*.txt"):
print(fname)
run_prof(fname, workload_dir, perfmon_dir, app_cmd, verbose)
run_prof(fname, workload_dir, perfmon_dir, app_cmd, target, verbose)
# run again with timestamps
run_subprocess(
[
@@ -395,7 +395,7 @@ def run_rocscope(args, fname):
sys.exit(1)
def run_prof(fname, workload_dir, perfmon_dir, cmd, verbose):
def run_prof(fname, workload_dir, perfmon_dir, cmd, target, verbose):
global rocprof_cmd
fbase = os.path.splitext(os.path.basename(fname))[0]
@@ -403,19 +403,36 @@ def run_prof(fname, workload_dir, perfmon_dir, cmd, verbose):
if verbose:
print("pmc file:", os.path.basename(fname))
# profile the app
run_subprocess(
[
rocprof_cmd,
"-i",
fname,
"--timestamp",
"on",
"-o",
workload_dir + "/" + fbase + ".csv",
'"' + cmd + '"',
]
)
# profile the app (run w/ custom config files for mi100)
if target == "mi100":
print("RUNNING WITH CUSTOM METRICS")
run_subprocess(
[
rocprof_cmd,
"-i",
fname,
"-m",
perfmon_dir + "/" + "metrics.xml",
"--timestamp",
"on",
"-o",
workload_dir + "/" + fbase + ".csv",
'"' + cmd + '"',
]
)
else:
run_subprocess(
[
rocprof_cmd,
"-i",
fname,
"--timestamp",
"on",
"-o",
workload_dir + "/" + fbase + ".csv",
'"' + cmd + '"',
]
)
def omniperf_profile(args, VER):
@@ -545,7 +562,7 @@ def omniperf_profile(args, VER):
if args.use_rocscope == True:
run_rocscope(args, fname)
else:
run_prof(fname, workload_dir, perfmon_dir, args.remaining, args.verbose)
run_prof(fname, workload_dir, perfmon_dir, args.remaining, args.target, args.verbose)
# run again with timestamps
run_subprocess(
projects/rocprofiler-compute/src/perfmon_pub/gfx908_metrics.xml
+737
View File
@@ -0,0 +1,737 @@
<gfx908>
# CPC counters
<metric
name="CPC_ME1_BUSY_FOR_PACKET_DECODE" block=CPC event=13 descr="Me1 busy for packet decode."
></metric>
<metric
name="CPC_UTCL1_STALL_ON_TRANSLATION" block=CPC event=24 descr="One of the UTCL1s is stalled waiting on translation, XNACK or PENDING response."
></metric>
<metric
name="CPC_CPC_STAT_BUSY" block=CPC event=25 descr="CPC Busy."
></metric>
<metric
name="CPC_CPC_STAT_IDLE" block=CPC event=26 descr="CPC Idle."
></metric>
<metric
name="CPC_CPC_STAT_STALL" block=CPC event=27 descr="CPC Stalled."
></metric>
<metric
name="CPC_CPC_TCIU_BUSY" block=CPC event=28 descr="CPC TCIU interface Busy."
></metric>
<metric
name="CPC_CPC_TCIU_IDLE" block=CPC event=29 descr="CPC TCIU interface Idle."
></metric>
<metric
name="CPC_CPC_UTCL2IU_BUSY" block=CPC event=30 descr="CPC UTCL2 interface Busy."
></metric>
<metric
name="CPC_CPC_UTCL2IU_IDLE" block=CPC event=31 descr="CPC UTCL2 interface Idle."
></metric>
<metric
name="CPC_CPC_UTCL2IU_STALL" block=CPC event=32 descr="CPC UTCL2 interface Stalled waiting on Free, Tags or Translation."
></metric>
<metric
name="CPC_ME1_DC0_SPI_BUSY" block=CPC event=33 descr="CPC Me1 Processor Busy."
></metric>
<metric
name="CPF_CMP_UTCL1_STALL_ON_TRANSLATION" block=CPF event=20 descr="One of the Compute UTCL1s is stalled waiting on translation, XNACK or PENDING response."
></metric>
<metric
name="CPF_CPF_STAT_BUSY" block=CPF event=23 descr="CPF Busy."
></metric>
<metric
name="CPF_CPF_STAT_IDLE" block=CPF event=24 descr="CPF Idle."
></metric>
<metric
name="CPF_CPF_STAT_STALL" block=CPF event=25 descr="CPF Stalled."
></metric>
<metric
name="CPF_CPF_TCIU_BUSY" block=CPF event=26 descr="CPF TCIU interface Busy."
></metric>
<metric
name="CPF_CPF_TCIU_IDLE" block=CPF event=27 descr="CPF TCIU interface Idle."
></metric>
<metric
name="CPF_CPF_TCIU_STALL" block=CPF event=28 descr="CPF TCIU interface Stalled waiting on Free, Tags."
></metric>
# GRBM counters
<metric
name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"
></metric>
<metric
name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"
></metric>
<metric
name="GRBM_CP_BUSY" block=GRBM event=3 descr="Any of the Command Processor (CPG/CPC/CPF) blocks are busy."
></metric>
<metric
name="GRBM_SPI_BUSY" block=GRBM event=11 descr="Any of the Shader Pipe Interpolators (SPI) are busy in the shader engine(s)."
></metric>
<metric
name="GRBM_TA_BUSY" block=GRBM event=13 descr="Any of the Texture Pipes (TA) are busy in the shader engine(s)."
></metric>
<metric
name="GRBM_TC_BUSY" block=GRBM event=28 descr="Any of the Texture Cache Blocks (TCP/TCI/TCA/TCC) are busy."
></metric>
<metric
name="GRBM_CPC_BUSY" block=GRBM event=30 descr="The Command Processor Compute (CPC) is busy."
></metric>
<metric
name="GRBM_CPF_BUSY" block=GRBM event=31 descr="The Command Processor Fetchers (CPF) is busy."
></metric>
<metric
name="GRBM_UTCL2_BUSY" block=GRBM event=34 descr="The Unified Translation Cache Level-2 (UTCL2) block is busy."
></metric>
<metric
name="GRBM_EA_BUSY" block=GRBM event=35 descr="The Efficiency Arbiter (EA) block is busy."
></metric>
# SPI counters
<metric
name="SPI_CSN_WINDOW_VALID" block=SPI event=47 descr="Clock count enabled by perfcounter_start event. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
<metric
name="SPI_CSN_BUSY" block=SPI event=48 descr="Number of clocks with outstanding waves (SPI or SH). Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
<metric
name="SPI_CSN_NUM_THREADGROUPS" block=SPI event=49 descr="Number of threadgroups launched. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
<metric
name="SPI_CSN_WAVE" block=SPI event=52 descr="Number of waves. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
<metric
name="SPI_RA_REQ_NO_ALLOC" block=SPI event=79 descr="Arb cycles with requests but no allocation. Source is RA0"
></metric>
<metric
name="SPI_RA_REQ_NO_ALLOC_CSN" block=SPI event=85 descr="Arb cycles with CSn req and no CSn alloc. Source is RA0"
></metric>
<metric
name="SPI_RA_RES_STALL_CSN" block=SPI event=91 descr="Arb cycles with CSn req and no CSn fits. Source is RA0"
></metric>
<metric
name="SPI_RA_TMP_STALL_CSN" block=SPI event=97 descr="Cycles where csn wants to req but does not fit in temp space."
></metric>
<metric
name="SPI_RA_WAVE_SIMD_FULL_CSN" block=SPI event=103 descr="Sum of SIMD where WAVE can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_VGPR_SIMD_FULL_CSN" block=SPI event=109 descr="Sum of SIMD where VGPR can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_SGPR_SIMD_FULL_CSN" block=SPI event=115 descr="Sum of SIMD where SGPR can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_LDS_CU_FULL_CSN" block=SPI event=120 descr="Sum of CU where LDS can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_BAR_CU_FULL_CSN" block=SPI event=123 descr="Sum of CU where BARRIER can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_BULKY_CU_FULL_CSN" block=SPI event=125 descr="Sum of CU where BULKY can't take csn wave when !fits. Source is RA0"
></metric>
<metric
name="SPI_RA_TGLIM_CU_FULL_CSN" block=SPI event=127 descr="Cycles where csn wants to req but all CU are at tg_limit"
></metric>
<metric
name="SPI_RA_WVLIM_STALL_CSN" block=SPI event=133 descr="Number of clocks csn is stalled due to WAVE LIMIT."
></metric>
<metric
name="SPI_SWC_CSC_WR" block=SPI event=189 descr="Number of clocks to write CSC waves to SGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
<metric
name="SPI_VWC_CSC_WR" block=SPI event=195 descr="Number of clocks to write CSC waves to VGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"
></metric>
# SQ counters
<metric
name="SQ_ACCUM_PREV" block=SQ event=1 descr="For counter N, increment by the value of counter N-1. Only accumulates once every 4 cycles."
></metric>
<metric
name="SQ_CYCLES" block=SQ event=2 descr="Clock cycles. (nondeterministic, per-simd, global)"
></metric>
<metric
name="SQ_BUSY_CYCLES" block=SQ event=3 descr="Clock cycles while SQ is reporting that it is busy. (nondeterministic, per-simd, global)"
></metric>
<metric
name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_LEVEL_WAVES" block=SQ event=5 descr="Track the number of waves. Set ACCUM_PREV for the next counter to use this. (level, per-simd, global)"
></metric>
<metric
name="SQ_WAVES_EQ_64" block=SQ event=6 descr="Count number of waves with exactly 64 active threads sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_WAVES_LT_64" block=SQ event=7 descr="Count number of waves with <64 active threads sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_WAVES_LT_48" block=SQ event=8 descr="Count number of waves with <48 active threads sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_WAVES_LT_32" block=SQ event=9 descr="Count number of waves sent <32 active threads sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_WAVES_LT_16" block=SQ event=10 descr="Count number of waves sent <16 active threads sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_BUSY_CU_CYCLES" block=SQ event=13 descr="Count quad-cycles each CU is busy. (nondeterministic, per-simd)"
></metric>
<metric
name="SQ_ITEMS" block=SQ event=14 descr="Number of valid items per wave. (per-simd, global)"
></metric>
<metric
name="SQ_INSTS" block=SQ event=25 descr="Number of instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_VALU" block=SQ event=26 descr="Number of VALU instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_MFMA" block=SQ event=27 descr="Number of MFMA instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_VMEM_WR" block=SQ event=28 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_VMEM_RD" block=SQ event=29 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_VMEM" block=SQ event=30 descr="Number of VMEM instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_SALU" block=SQ event=31 descr="Number of SALU instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_SMEM" block=SQ event=32 descr="Number of SMEM instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_FLAT" block=SQ event=33 descr="Number of FLAT instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_FLAT_LDS_ONLY" block=SQ event=34 descr="Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_LDS" block=SQ event=35 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_GDS" block=SQ event=36 descr="Number of GDS instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_EXP_GDS" block=SQ event=38 descr="Number of EXP and GDS instructions issued, excluding skipped export instructions. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_BRANCH" block=SQ event=39 descr="Number of Branch instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_SENDMSG" block=SQ event=40 descr="Number of Sendmsg instructions issued. (per-simd, emulated)"
></metric>
<metric
name="SQ_INSTS_VSKIPPED" block=SQ event=41 descr="Number of vector instructions skipped. (per-simd, emulated)"
></metric>
<metric
name="SQ_INST_LEVEL_VMEM" block=SQ event=42 descr="Number of in-flight VMEM instructions. Set next counter to ACCUM_PREV and divide by INSTS_VMEM for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"
></metric>
<metric
name="SQ_INST_LEVEL_SMEM" block=SQ event=43 descr="Number of in-flight SMEM instructions (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). Set next counter to ACCUM_PREV and divide by INSTS_SMEM for average latency per smem request. Falls slightly short of total request latency because some fetches are divided into two requests that may finish at different times and this counter collects the average latency of the two. (per-simd, level, nondeterministic)"
></metric>
<metric
name="SQ_INST_LEVEL_LDS" block=SQ event=44 descr="Number of in-flight LDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_LDS for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"
></metric>
<metric
name="SQ_WAVE_CYCLES" block=SQ event=47 descr="Number of wave-cycles spent by waves in the CUs (per-simd, nondeterministic)"
></metric>
<metric
name="SQ_WAIT_ANY" block=SQ event=58 descr="Number of wave-cycles spent waiting for anything (per-simd, nondeterministic)"
></metric>
<metric
name="SQ_WAIT_INST_ANY" block=SQ event=61 descr="Number of wave-cycles spent waiting for any instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"
></metric>
<metric
name="SQ_WAIT_INST_LDS" block=SQ event=64 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"
></metric>
<metric
name="SQ_ACTIVE_INST_ANY" block=SQ event=69 descr="Number of cycles each wave is working on an instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_VMEM" block=SQ event=70 descr="Number of cycles the SQ instruction arbiter is working on a VMEM instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_LDS" block=SQ event=71 descr="Number of cycles the SQ instruction arbiter is working on a LDS instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_VALU" block=SQ event=72 descr="Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_SCA" block=SQ event=73 descr="Number of cycles the SQ instruction arbiter is working on a SALU or SMEM instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_EXP_GDS" block=SQ event=74 descr="Number of cycles the SQ instruction arbiter is working on an EXPORT or GDS instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_MISC" block=SQ event=75 descr="Number of cycles the SQ instruction aribter is working on a BRANCH or SENDMSG instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACTIVE_INST_FLAT" block=SQ event=76 descr="Number of cycles the SQ instruction arbiter is working on a FLAT instruction. (per-simd, emulated)"
></metric>
<metric
name="SQ_INST_CYCLES_VMEM_WR" block=SQ event=77 descr="Number of cycles needed to send addr and cmd data for VMEM write instructions. (per-simd, emulated)"
></metric>
<metric
name="SQ_INST_CYCLES_VMEM_RD" block=SQ event=78 descr="Number of cycles needed to send addr and cmd data for VMEM read instructions. (per-simd, emulated)"
></metric>
<metric
name="SQ_INST_CYCLES_SMEM" block=SQ event=84 descr="Number of cycles needed to execute scalar memory reads. (per-simd, emulated)"
></metric>
<metric
name="SQ_INST_CYCLES_SALU" block=SQ event=85 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated)"
></metric>
<metric
name="SQ_THREAD_CYCLES_VALU" block=SQ event=86 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"
></metric>
<metric
name="SQ_IFETCH" block=SQ event=88 descr="Number of instruction fetch requests from cache. (per-simd, emulated)"
></metric>
<metric
name="SQ_IFETCH_LEVEL" block=SQ event=89 descr="Number of instruction fetch requests from cache. (per-simd, level)"
></metric>
<metric
name="SQ_LDS_BANK_CONFLICT" block=SQ event=94 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"
></metric>
<metric
name="SQ_LDS_ADDR_CONFLICT" block=SQ event=95 descr="Number of cycles LDS is stalled by address conflicts. (emulated,nondeterministic)"
></metric>
<metric
name="SQ_LDS_UNALIGNED_STALL" block=SQ event=96 descr="Number of cycles LDS is stalled processing flat unaligned load/store ops. (emulated)"
></metric>
<metric
name="SQ_LDS_MEM_VIOLATIONS" block=SQ event=97 descr="Number of threads that have a memory violation in the LDS.(emulated)"
></metric>
<metric
name="SQ_LDS_ATOMIC_RETURN" block=SQ event=98 descr="Number of atomic return cycles in LDS. (per-simd, emulated)"
></metric>
<metric
name="SQ_LDS_IDX_ACTIVE" block=SQ event=99 descr="Number of cycles LDS is used for indexed (non-direct,non-interpolation) operations. (per-simd, emulated)"
></metric>
<metric
name="SQ_ACCUM_PREV_HIRES" block=SQ event=158 descr="For counter N, increment by the value of counter N-1."
></metric>
<metric
name="SQ_WAVES_RESTORED" block=SQ event=159 descr="Count number of context-restored waves sent to SQs. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_WAVES_SAVED" block=SQ event=160 descr="Count number of context-saved waves. (per-simd, emulated, global)"
></metric>
<metric
name="SQ_INSTS_SMEM_NORM" block=SQ event=161 descr="Number of SMEM instructions issued normalized to match smem_level (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). (per-simd, emulated)"
></metric>
<metric
name="SQC_DCACHE_INPUT_VALID_READYB" block=SQ event=260 descr="Input stalled by SQC (per-SQ, nondeterministic, unwindowed)"
></metric>
<metric
name="SQC_TC_REQ" block=SQ event=262 descr="Total number of TC requests that were issued by instruction and constant caches. (No-Masking, nondeterministic)"
></metric>
<metric
name="SQC_TC_INST_REQ" block=SQ event=263 descr="Number of insruction requests to the TC (No-Masking, nondeterministic)"
></metric>
<metric
name="SQC_TC_DATA_READ_REQ" block=SQ event=264 descr="Number of data read requests to the TC (No-Masking, nondeterministic)"
></metric>
<metric
name="SQC_TC_DATA_WRITE_REQ" block=SQ event=265 descr="Number of data write requests to the TC (No-Masking, nondeterministic)"
></metric>
<metric
name="SQC_TC_DATA_ATOMIC_REQ" block=SQ event=266 descr="Number of data atomic requests to the TC (No-Masking, nondeterministic)"
></metric>
<metric
name="SQC_TC_STALL" block=SQ event=267 descr="Valid request stalled TC request interface (no-credits). (No-Masking, nondeterministic, unwindowed)"
></metric>
<metric
name="SQC_ICACHE_REQ" block=SQ event=270 descr="Number of requests. (per-SQ, per-Bank)"
></metric>
<metric
name="SQC_ICACHE_HITS" block=SQ event=271 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_ICACHE_MISSES" block=SQ event=272 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_ICACHE_MISSES_DUPLICATE" block=SQ event=273 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_DCACHE_REQ" block=SQ event=290 descr="Number of requests (post-bank-serialization). (per-SQ, per-Bank)"
></metric>
<metric
name="SQC_DCACHE_HITS" block=SQ event=291 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_DCACHE_MISSES" block=SQ event=292 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_DCACHE_MISSES_DUPLICATE" block=SQ event=293 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)"
></metric>
<metric
name="SQC_DCACHE_ATOMIC" block=SQ event=298 descr="Number of atomic requests. (per-SQ, per-Bank)"
></metric>
<metric
name="SQC_DCACHE_REQ_READ_1" block=SQ event=323 descr="Number of constant cache 1 dw read requests. (per-SQ)"
></metric>
<metric
name="SQC_DCACHE_REQ_READ_2" block=SQ event=324 descr="Number of constant cache 2 dw read requests. (per-SQ)"
></metric>
<metric
name="SQC_DCACHE_REQ_READ_4" block=SQ event=325 descr="Number of constant cache 4 dw read requests. (per-SQ)"
></metric>
<metric
name="SQC_DCACHE_REQ_READ_8" block=SQ event=326 descr="Number of constant cache 8 dw read requests. (per-SQ)"
></metric>
<metric
name="SQC_DCACHE_REQ_READ_16" block=SQ event=327 descr="Number of constant cache 16 dw read requests. (per-SQ)"
></metric>
# TA counters
<metric
name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."
></metric>
<metric
name="TA_TOTAL_WAVEFRONTS" block=TA event=32 descr="Total number of wavefronts processed by TA."
></metric>
<metric
name="TA_BUFFER_WAVEFRONTS" block=TA event=44 descr="Number of buffer wavefronts processed by TA."
></metric>
<metric
name="TA_BUFFER_READ_WAVEFRONTS" block=TA event=45 descr="Number of buffer read wavefronts processed by TA."
></metric>
<metric
name="TA_BUFFER_WRITE_WAVEFRONTS" block=TA event=46 descr="Number of buffer write wavefronts processed by TA."
></metric>
<metric
name="TA_BUFFER_ATOMIC_WAVEFRONTS" block=TA event=47 descr="Number of buffer atomic wavefronts processed by TA."
></metric>
<metric
name="TA_BUFFER_TOTAL_CYCLES" block=TA event=49 descr="Number of buffer cycles issued to TC."
></metric>
<metric
name="TA_BUFFER_COALESCED_READ_CYCLES" block=TA event=52 descr="Number of buffer coalesced read cycles issued to TC."
></metric>
<metric
name="TA_BUFFER_COALESCED_WRITE_CYCLES" block=TA event=53 descr="Number of buffer coalesced write cycles issued to TC."
></metric>
<metric
name="TA_ADDR_STALLED_BY_TC_CYCLES" block=TA event=54 descr="Number of cycles addr path stalled by TC. Perf_Windowing not supported for this counter."
></metric>
<metric
name="TA_ADDR_STALLED_BY_TD_CYCLES" block=TA event=55 descr="Number of cycles addr path stalled by TD. Perf_Windowing not supported for this counter."
></metric>
<metric
name="TA_DATA_STALLED_BY_TC_CYCLES" block=TA event=56 descr="Number of cycles data path stalled by TC. Perf_Windowing not supported for this counter."
></metric>
<metric
name="TA_FLAT_WAVEFRONTS" block=TA event=100 descr="Number of flat opcode wavfronts processed by the TA."
></metric>
<metric
name="TA_FLAT_READ_WAVEFRONTS" block=TA event=101 descr="Number of flat opcode reads processed by the TA."
></metric>
<metric
name="TA_FLAT_WRITE_WAVEFRONTS" block=TA event=102 descr="Number of flat opcode writes processed by the TA."
></metric>
<metric
name="TA_FLAT_ATOMIC_WAVEFRONTS" block=TA event=103 descr="Number of flat opcode atomics processed by the TA."
></metric>
# TCA counters
<metric
name="TCA_CYCLE" block=TCA event=1 descr="Number of cycles. Not windowable."
></metric>
<metric
name="TCA_BUSY" block=TCA event=2 descr="Number of cycles we have a request pending. Not windowable."
></metric>
# TCC counters
<metric
name="TCC_CYCLE" block=TCC event=1 descr="Number of cycles. Not windowable."
></metric>
<metric
name="TCC_BUSY" block=TCC event=2 descr="Number of cycles we have a request pending. Not windowable."
></metric>
<metric
name="TCC_REQ" block=TCC event=3 descr="Number of requests of all types. This is measured at the tag block. This may be more than the number of requests arriving at the TCC, but it is a good indication of the total amount of work that needs to be performed."
></metric>
<metric
name="TCC_STREAMING_REQ" block=TCC event=4 descr="Number of streaming requests. This is measured at the tag block."
></metric>
<metric
name="TCC_NC_REQ" block=TCC event=5 descr="The number of noncoherently cached requests. This is measured at the tag block."
></metric>
<metric
name="TCC_UC_REQ" block=TCC event=6 descr="The number of uncached requests. This is measured at the tag block."
></metric>
<metric
name="TCC_CC_REQ" block=TCC event=7 descr="The number of coherently cached requests. This is measured at the tag block."
></metric>
<metric
name="TCC_RW_REQ" block=TCC event=8 descr="The number of RW requests. This is measured at the tag block."
></metric>
<metric
name="TCC_PROBE" block=TCC event=9 descr="Number of probe requests. Not windowable."
></metric>
<metric
name="TCC_PROBE_ALL" block=TCC event=10 descr="Number of external probe requests with with EA_TCC_preq_all== 1. Not windowable."
></metric>
<metric
name="TCC_READ" block=TCC event=12 descr="Number of read requests. Compressed reads are included in this, but metadata reads are not included."
></metric>
<metric
name="TCC_WRITE" block=TCC event=13 descr="Number of write requests."
></metric>
<metric
name="TCC_ATOMIC" block=TCC event=14 descr="Number of atomic requests of all types."
></metric>
<metric
name="TCC_HIT" block=TCC event=17 descr="Number of cache hits."
></metric>
<metric
name="TCC_MISS" block=TCC event=19 descr="Number of cache misses. UC reads count as misses."
></metric>
<metric
name="TCC_WRITEBACK" block=TCC event=22 descr="Number of lines written back to main memory. This includes writebacks of dirty lines and uncached write/atomic requests."
></metric>
<metric
name="TCC_EA_WRREQ" block=TCC event=26 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."
></metric>
<metric
name="TCC_EA_WRREQ_64B" block=TCC event=27 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."
></metric>
<metric
name="TCC_EA_WR_UNCACHED_32B" block=TCC event=29 descr="Number of 32-byte write/atomic going over the TC_EA_wrreq interface due to uncached traffic. Note that CC mtypes can produce uncached requests, and those are included in this. A 64-byte request will be counted as 2"
></metric>
<metric
name="TCC_EA_WRREQ_STALL" block=TCC event=30 descr="Number of cycles a write request was stalled."
></metric>
<metric
name="TCC_EA_WRREQ_IO_CREDIT_STALL" block=TCC event=31 descr="Number of cycles a EA write request was stalled because the interface was out of IO credits."
></metric>
<metric
name="TCC_EA_WRREQ_GMI_CREDIT_STALL" block=TCC event=32 descr="Number of cycles a EA write request was stalled because the interface was out of GMI credits."
></metric>
<metric
name="TCC_EA_WRREQ_DRAM_CREDIT_STALL" block=TCC event=33 descr="Number of cycles a EA write request was stalled because the interface was out of DRAM credits."
></metric>
<metric
name="TCC_TOO_MANY_EA_WRREQS_STALL" block=TCC event=34 descr="Number of cycles the TCC could not send a EA write request because it already reached its maximum number of pending EA write requests."
></metric>
<metric
name="TCC_EA_WRREQ_LEVEL" block=TCC event=35 descr="The sum of the number of EA write requests in flight. This is primarily meant for measure average EA write latency. Average write latency = TCC_PERF_SEL_EA_WRREQ_LEVEL/TCC_PERF_SEL_EA_WRREQ."
></metric>
<metric
name="TCC_EA_ATOMIC" block=TCC event=36 descr="Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests."
></metric>
<metric
name="TCC_EA_ATOMIC_LEVEL" block=TCC event=37 descr="The sum of the number of EA atomics in flight. This is primarily meant for measure average EA atomic latency. Average atomic latency = TCC_PERF_SEL_EA_WRREQ_ATOMIC_LEVEL/TCC_PERF_SEL_EA_WRREQ_ATOMIC."
></metric>
<metric
name="TCC_EA_RDREQ" block=TCC event=38 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"
></metric>
<metric
name="TCC_EA_RDREQ_32B" block=TCC event=39 descr="Number of 32-byte TCC/EA read requests"
></metric>
<metric
name="TCC_EA_RD_UNCACHED_32B" block=TCC event=40 descr="Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2"
></metric>
<metric
name="TCC_EA_RDREQ_IO_CREDIT_STALL" block=TCC event=41 descr="Number of cycles there was a stall because the read request interface was out of IO credits. Stalls occur regardless of whether a read needed to be performed or not."
></metric>
<metric
name="TCC_EA_RDREQ_GMI_CREDIT_STALL" block=TCC event=42 descr="Number of cycles there was a stall because the read request interface was out of GMI credits. Stalls occur regardless of whether a read needed to be performed or not."
></metric>
<metric
name="TCC_EA_RDREQ_DRAM_CREDIT_STALL" block=TCC event=43 descr="Number of cycles there was a stall because the read request interface was out of DRAM credits. Stalls occur regardless of whether a read needed to be performed or not."
></metric>
<metric
name="TCC_EA_RDREQ_LEVEL" block=TCC event=44 descr="The sum of the number of TCC/EA read requests in flight. This is primarily meant for measure average EA read latency. Average read latency = TCC_PERF_SEL_EA_RDREQ_LEVEL/TCC_PERF_SEL_EA_RDREQ."
></metric>
<metric
name="TCC_TAG_STALL" block=TCC event=45 descr="Number of cycles the normal request pipeline in the tag was stalled for any reason. Normally, stalls of this nature are measured exactly from one point the pipeline, but that is not the case for this counter. Probes can stall the pipeline at a variety of places, and there is no single point that can reasonably measure the total stalls accurately."
></metric>
<metric
name="TCC_NORMAL_WRITEBACK" block=TCC event=68 descr="Number of writebacks due to requests that are not writeback requests."
></metric>
<metric
name="TCC_ALL_TC_OP_WB_WRITEBACK" block=TCC event=73 descr="Number of writebacks due to all TC_OP writeback requests."
></metric>
<metric
name="TCC_NORMAL_EVICT" block=TCC event=74 descr="Number of evictions due to requests that are not invalidate or probe requests."
></metric>
<metric
name="TCC_ALL_TC_OP_INV_EVICT" block=TCC event=80 descr="Number of evictions due to all TC_OP invalidate requests."
></metric>
<metric
name="TCC_EA_RDREQ_DRAM" block=TCC event=102 descr="Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC)."
></metric>
<metric
name="TCC_EA_WRREQ_DRAM" block=TCC event=103 descr="Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC)."
></metric>
<metric
name="TCC_CLIENT184_REQ" block=TCC event=312 descr=""
></metric>
<metric
name="TCC_CLIENT185_REQ" block=TCC event=313 descr=""
></metric>
<metric
name="TCC_CLIENT186_REQ" block=TCC event=314 descr=""
></metric>
<metric
name="TCC_CLIENT187_REQ" block=TCC event=315 descr=""
></metric>
<metric
name="TCC_CLIENT188_REQ" block=TCC event=316 descr=""
></metric>
<metric
name="TCC_CLIENT189_REQ" block=TCC event=317 descr=""
></metric>
<metric
name="TCC_CLIENT190_REQ" block=TCC event=318 descr=""
></metric>
<metric
name="TCC_CLIENT191_REQ" block=TCC event=319 descr=""
></metric>
# TCP counters
<metric
name="TCP_GATE_EN1" block=TCP event=0 descr="TCP interface clocks are turned on. Not Windowed."
></metric>
<metric
name="TCP_GATE_EN2" block=TCP event=1 descr="TCP core clocks are turned on. Not Windowed."
></metric>
<metric
name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=6 descr="TCP stalls TA data interface. Not Windowed."
></metric>
<metric
name="TCP_TD_TCP_STALL_CYCLES" block=TCP event=7 descr="TD stalls TCP"
></metric>
<metric
name="TCP_TCR_TCP_STALL_CYCLES" block=TCP event=8 descr="TCR stalls TCP_TCR_req interface"
></metric>
<metric
name="TCP_READ_TAGCONFLICT_STALL_CYCLES" block=TCP event=11 descr="Tagram conflict stall on a read"
></metric>
<metric
name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES" block=TCP event=12 descr="Tagram conflict stall on a write"
></metric>
<metric
name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES" block=TCP event=13 descr="Tagram conflict stall on an atomic"
></metric>
<metric
name="TCP_PENDING_STALL_CYCLES" block=TCP event=22 descr="Stall due to data pending from L2"
></metric>
<metric
name="TCP_TA_TCP_STATE_READ" block=TCP event=27 descr="Number of state reads"
></metric>
<metric
name="TCP_VOLATILE" block=TCP event=28 descr="Total number of L1 volatile pixels/buffers from TA"
></metric>
<metric
name="TCP_TOTAL_ACCESSES" block=TCP event=29 descr="Total number of pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_READ+TCP_PERF_SEL_TOTAL_NONREAD"
></metric>
<metric
name="TCP_TOTAL_READ" block=TCP event=30 descr="Total number of read pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_HIT_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_EVICT_READ"
></metric>
<metric
name="TCP_TOTAL_WRITE" block=TCP event=32 descr="Total number of local write pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_MISS_LRU_WRITE+ TCP_PERF_SEL_TOTAL_MISS_EVICT_WRITE"
></metric>
<metric
name="TCP_TOTAL_ATOMIC_WITH_RET" block=TCP event=38 descr="Total number of atomic with return pixels/buffers from TA"
></metric>
<metric
name="TCP_TOTAL_ATOMIC_WITHOUT_RET" block=TCP event=39 descr="Total number of atomic without return pixels/buffers from TA"
></metric>
<metric
name="TCP_TOTAL_WRITEBACK_INVALIDATES" block=TCP event=45 descr="Total number of cache invalidates. Equals TCP_PERF_SEL_TOTAL_WBINVL1+ TCP_PERF_SEL_TOTAL_WBINVL1_VOL+ TCP_PERF_SEL_CP_TCP_INVALIDATE+ TCP_PERF_SEL_SQ_TCP_INVALIDATE_VOL. Not Windowed."
></metric>
<metric
name="TCP_UTCL1_REQUEST" block=TCP event=47 descr="Total CLIENT_UTCL1 NORMAL requests"
></metric>
<metric
name="TCP_UTCL1_TRANSLATION_MISS" block=TCP event=48 descr="Total utcl1 translation misses"
></metric>
<metric
name="TCP_UTCL1_TRANSLATION_HIT" block=TCP event=49 descr="Total utcl1 translation hits"
></metric>
<metric
name="TCP_UTCL1_PERMISSION_MISS" block=TCP event=50 descr="Total utcl1 permission misses"
></metric>
<metric
name="TCP_TOTAL_CACHE_ACCESSES" block=TCP event=60 descr="Count of total cache line (tag) accesses (includes hits and misses)."
></metric>
<metric
name="TCP_TCP_LATENCY" block=TCP event=65 descr="Total TCP wave latency (from first clock of wave entering to first clock of wave leaving), divide by TA_TCP_STATE_READ to avg wave latency"
></metric>
<metric
name="TCP_TCC_READ_REQ_LATENCY" block=TCP event=66 descr="Total TCP->TCC request latency for reads and atomics with return. Not Windowed."
></metric>
<metric
name="TCP_TCC_WRITE_REQ_LATENCY" block=TCP event=67 descr="Total TCP->TCC request latency for writes and atomics without return. Not Windowed."
></metric>
<metric
name="TCP_TCC_READ_REQ" block=TCP event=69 descr="Total read requests from TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_WRITE_REQ" block=TCP event=70 descr="Total write requests from TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_ATOMIC_WITH_RET_REQ" block=TCP event=71 descr="Total atomic with return requests from TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ" block=TCP event=72 descr="Total atomic without return requests from TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_NC_READ_REQ" block=TCP event=75 descr="Total read requests with NC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_NC_WRITE_REQ" block=TCP event=76 descr="Total write requests with NC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_NC_ATOMIC_REQ" block=TCP event=77 descr="Total atomic requests with NC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_UC_READ_REQ" block=TCP event=78 descr="Total read requests with UC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_UC_WRITE_REQ" block=TCP event=79 descr="Total write requests with UC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_UC_ATOMIC_REQ" block=TCP event=80 descr="Total atomic requests with UC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_CC_READ_REQ" block=TCP event=81 descr="Total write requests with CC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_CC_WRITE_REQ" block=TCP event=82 descr="Total write requests with CC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_CC_ATOMIC_REQ" block=TCP event=83 descr="Total atomic requests with CC mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_RW_READ_REQ" block=TCP event=85 descr="Total write requests with RW mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_RW_WRITE_REQ" block=TCP event=86 descr="Total write requests with RW mtype from this TCP to all TCCs"
></metric>
<metric
name="TCP_TCC_RW_ATOMIC_REQ" block=TCP event=87 descr="Total atomic requests with RW mtype from this TCP to all TCCs"
></metric>
# TD counters
<metric
name="TD_TD_BUSY" block=TD event=1 descr="TD is processing or waiting for data. Perf_Windowing not supported for this counter."
></metric>
<metric
name="TD_TC_STALL" block=TD event=15 descr="TD is stalled waiting for TC data."
></metric>
<metric
name="TD_RESERVED_18" block=TD event=18 descr="RESERVED_18"
></metric>
<metric
name="TD_LOAD_WAVEFRONT" block=TD event=25 descr="Count the wavefronts with opcode = load, include atomics and store."
></metric>
<metric
name="TD_ATOMIC_WAVEFRONT" block=TD event=26 descr="Count the wavefronts with opcode = atomic."
></metric>
<metric
name="TD_STORE_WAVEFRONT" block=TD event=27 descr="Count the wavefronts with opcode = store."
></metric>
<metric
name="TD_COALESCABLE_WAVEFRONT" block=TD event=32 descr="Count wavefronts that TA finds coalescable."
></metric>
</gfx908>
projects/rocprofiler-compute/src/perfmon_pub/metrics.xml
+163
View File
@@ -0,0 +1,163 @@
#include "gfx908_metrics.xml"
<gfx9_expr>
<metric name="TA_BUSY_avr" expr=avr(TA_TA_BUSY,16) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=max(TA_TA_BUSY,16) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=min(TA_TA_BUSY,16) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=sum(TA_FLAT_READ_WAVEFRONTS,16) descr="Number of flat opcode reads processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=sum(TA_FLAT_WRITE_WAVEFRONTS,16) descr="Number of flat opcode writes processed by the TA. Sum over TA instances."></metric>
<metric name="TCC_BUSY_avr" expr=avr(TCC_BUSY,16) descr="TCC_BUSY avr over all memory channels."></metric>
<metric name="TCC_REQ_sum" expr=sum(TCC_REQ,16) descr="TCC_REQ sum over all memory channels."></metric>
<metric name="TCC_HIT_sum" expr=sum(TCC_HIT,16) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=sum(TCC_MISS,16) descr="Number of cache misses. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=sum(TCC_EA_RDREQ_32B,16) descr="Number of 32-byte TCC/EA read requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=sum(TCC_EA_RDREQ,16) descr="Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=sum(TCC_EA_WRREQ,16) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=sum(TCC_EA_WRREQ_64B,16) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=max(TCC_EA_WRREQ_STALL,16) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="FETCH_SIZE" expr=(TCC_EA_RDREQ_32B_sum*32+(TCC_EA_RDREQ_sum-TCC_EA_RDREQ_32B_sum)*64)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=((TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)*32+TCC_EA_WRREQ_64B_sum*64)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=TCC_EA_WRREQ_64B_sum*2+(TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum) descr="The total number of 32-byte effective memory writes."></metric>
#xlu - TA
<metric name="TA_TA_BUSY_sum" expr=sum(TA_TA_BUSY,16) descr="."></metric>
<metric name="TA_TOTAL_WAVEFRONTS_sum" expr=sum(TA_TOTAL_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_ADDR_STALLED_BY_TC_CYCLES_sum" expr=sum(TA_ADDR_STALLED_BY_TC_CYCLES,16) descr="."></metric>
<metric name="TA_ADDR_STALLED_BY_TD_CYCLES_sum" expr=sum(TA_ADDR_STALLED_BY_TD_CYCLES,16) descr="."></metric>
<metric name="TA_DATA_STALLED_BY_TC_CYCLES_sum" expr=sum(TA_DATA_STALLED_BY_TC_CYCLES,16) descr="."></metric>
<metric name="TA_FLAT_WAVEFRONTS_sum" expr=sum(TA_FLAT_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=sum(TA_FLAT_READ_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=sum(TA_FLAT_WRITE_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_FLAT_ATOMIC_WAVEFRONTS_sum" expr=sum(TA_FLAT_ATOMIC_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_BUFFER_WAVEFRONTS_sum" expr=sum(TA_BUFFER_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_BUFFER_READ_WAVEFRONTS_sum" expr=sum(TA_BUFFER_READ_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_BUFFER_WRITE_WAVEFRONTS_sum" expr=sum(TA_BUFFER_WRITE_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_BUFFER_ATOMIC_WAVEFRONTS_sum" expr=sum(TA_BUFFER_ATOMIC_WAVEFRONTS,16) descr="."></metric>
<metric name="TA_BUFFER_TOTAL_CYCLES_sum" expr=sum(TA_BUFFER_TOTAL_CYCLES,16) descr="."></metric>
<metric name="TA_BUFFER_COALESCED_READ_CYCLES_sum" expr=sum(TA_BUFFER_COALESCED_READ_CYCLES,16) descr="."></metric>
<metric name="TA_BUFFER_COALESCED_WRITE_CYCLES_sum" expr=sum(TA_BUFFER_COALESCED_WRITE_CYCLES,16) descr="."></metric>
#xlu -TD
<metric name="TD_TD_BUSY_sum" expr=sum(TD_TD_BUSY,16) descr="."></metric>
<metric name="TD_TC_STALL_sum" expr=sum(TD_TC_STALL,16) descr="."></metric>
<metric name="TD_LOAD_WAVEFRONT_sum" expr=sum(TD_LOAD_WAVEFRONT,16) descr="."></metric>
<metric name="TD_ATOMIC_WAVEFRONT_sum" expr=sum(TD_ATOMIC_WAVEFRONT,16) descr="."></metric>
<metric name="TD_STORE_WAVEFRONT_sum" expr=sum(TD_STORE_WAVEFRONT,16) descr="."></metric>
<metric name="TD_COALESCABLE_WAVEFRONT_sum" expr=sum(TD_COALESCABLE_WAVEFRONT,16) descr="."></metric>
#xlu -TCP
<metric name="TCP_GATE_EN1_sum" expr=sum(TCP_GATE_EN1,16) descr="."></metric>
<metric name="TCP_GATE_EN2_sum" expr=sum(TCP_GATE_EN2,16) descr="."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES_sum" expr=sum(TCP_TCP_TA_DATA_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_TD_TCP_STALL_CYCLES_sum" expr=sum(TCP_TD_TCP_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_TCR_TCP_STALL_CYCLES_sum" expr=sum(TCP_TCR_TCP_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_READ_TAGCONFLICT_STALL_CYCLES_sum" expr=sum(TCP_READ_TAGCONFLICT_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum" expr=sum(TCP_WRITE_TAGCONFLICT_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum" expr=sum(TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_PENDING_STALL_CYCLES_sum" expr=sum(TCP_PENDING_STALL_CYCLES,16) descr="."></metric>
<metric name="TCP_VOLATILE_sum" expr=sum(TCP_VOLATILE,16) descr="."></metric>
<metric name="TCP_TOTAL_ACCESSES_sum" expr=sum(TCP_TOTAL_ACCESSES,16) descr="."></metric>
<metric name="TCP_TOTAL_READ_sum" expr=sum(TCP_TOTAL_READ,16) descr="."></metric>
<metric name="TCP_TOTAL_WRITE_sum" expr=sum(TCP_TOTAL_WRITE,16) descr="."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITH_RET_sum" expr=sum(TCP_TOTAL_ATOMIC_WITH_RET,16) descr="."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITHOUT_RET_sum" expr=sum(TCP_TOTAL_ATOMIC_WITHOUT_RET,16) descr="."></metric>
<metric name="TCP_TOTAL_WRITEBACK_INVALIDATES_sum" expr=sum(TCP_TOTAL_WRITEBACK_INVALIDATES,16) descr="."></metric>
<metric name="TCP_UTCL1_REQUEST_sum" expr=sum(TCP_UTCL1_REQUEST,16) descr="."></metric>
<metric name="TCP_UTCL1_TRANSLATION_MISS_sum" expr=sum(TCP_UTCL1_TRANSLATION_MISS,16) descr="."></metric>
<metric name="TCP_UTCL1_TRANSLATION_HIT_sum" expr=sum(TCP_UTCL1_TRANSLATION_HIT,16) descr="."></metric>
<metric name="TCP_UTCL1_PERMISSION_MISS_sum" expr=sum(TCP_UTCL1_PERMISSION_MISS,16) descr="."></metric>
<metric name="TCP_TOTAL_CACHE_ACCESSES_sum" expr=sum(TCP_TOTAL_CACHE_ACCESSES,16) descr="."></metric>
<metric name="TCP_TCP_LATENCY_sum" expr=sum(TCP_TCP_LATENCY,16) descr="."></metric>
<metric name="TCP_TA_TCP_STATE_READ_sum" expr=sum(TCP_TA_TCP_STATE_READ,16) descr="."></metric>
<metric name="TCP_TCC_READ_REQ_LATENCY_sum" expr=sum(TCP_TCC_READ_REQ_LATENCY,16) descr="."></metric>
<metric name="TCP_TCC_WRITE_REQ_LATENCY_sum" expr=sum(TCP_TCC_WRITE_REQ_LATENCY,16) descr="."></metric>
<metric name="TCP_TCC_READ_REQ_sum" expr=sum(TCP_TCC_READ_REQ,16) descr="."></metric>
<metric name="TCP_TCC_WRITE_REQ_sum" expr=sum(TCP_TCC_WRITE_REQ,16) descr="."></metric>
<metric name="TCP_TCC_ATOMIC_WITH_RET_REQ_sum" expr=sum(TCP_TCC_ATOMIC_WITH_RET_REQ,16) descr="."></metric>
<metric name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum" expr=sum(TCP_TCC_ATOMIC_WITHOUT_RET_REQ,16) descr="."></metric>
<metric name="TCP_TCC_NC_READ_REQ_sum" expr=sum(TCP_TCC_NC_READ_REQ,16) descr="."></metric>
<metric name="TCP_TCC_NC_WRITE_REQ_sum" expr=sum(TCP_TCC_NC_WRITE_REQ,16) descr="."></metric>
<metric name="TCP_TCC_NC_ATOMIC_REQ_sum" expr=sum(TCP_TCC_NC_ATOMIC_REQ,16) descr="."></metric>
<metric name="TCP_TCC_UC_READ_REQ_sum" expr=sum(TCP_TCC_UC_READ_REQ,16) descr="."></metric>
<metric name="TCP_TCC_UC_WRITE_REQ_sum" expr=sum(TCP_TCC_UC_WRITE_REQ,16) descr="."></metric>
<metric name="TCP_TCC_UC_ATOMIC_REQ_sum" expr=sum(TCP_TCC_UC_ATOMIC_REQ,16) descr="."></metric>
<metric name="TCP_TCC_CC_READ_REQ_sum" expr=sum(TCP_TCC_CC_READ_REQ,16) descr="."></metric>
<metric name="TCP_TCC_CC_WRITE_REQ_sum" expr=sum(TCP_TCC_CC_WRITE_REQ,16) descr="."></metric>
<metric name="TCP_TCC_CC_ATOMIC_REQ_sum" expr=sum(TCP_TCC_CC_ATOMIC_REQ,16) descr="."></metric>
</gfx9_expr>
<gfx908_expr base="gfx9_expr">
<metric name="TCC_BUSY_avr" expr=avr(TCC_BUSY,32) descr="TCC_BUSY avr over all memory channels."></metric>
<metric name="TCC_REQ_sum" expr=sum(TCC_REQ,32) descr="TCC_REQ sum over all memory channels."></metric>
<metric name="TCC_HIT_sum" expr=sum(TCC_HIT,32) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=sum(TCC_MISS,32) descr="Number of cache misses. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=sum(TCC_EA_RDREQ_32B,32) descr="Number of 32-byte TCC/EA read requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=sum(TCC_EA_RDREQ,32) descr="Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=sum(TCC_EA_WRREQ,32) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=sum(TCC_EA_WRREQ_64B,32) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=max(TCC_EA_WRREQ_STALL,32) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
#xlu - TCP
<metric name="TCP_TCC_RW_READ_REQ_sum" expr=sum(TCP_TCC_RW_READ_REQ,16) descr="."></metric>
<metric name="TCP_TCC_RW_WRITE_REQ_sum" expr=sum(TCP_TCC_RW_WRITE_REQ,16) descr="."></metric>
<metric name="TCP_TCC_RW_ATOMIC_REQ_sum" expr=sum(TCP_TCC_RW_ATOMIC_REQ,16) descr="."></metric>
#xlu - TCC
<metric name="TCC_CYCLE_sum" expr=sum(TCC_CYCLE,32) descr="."></metric>
<metric name="TCC_BUSY_sum" expr=sum(TCC_BUSY,32) descr="."></metric>
<metric name="TCC_REQ_sum" expr=sum(TCC_REQ,32) descr="."></metric>
<metric name="TCC_STREAMING_REQ_sum" expr=sum(TCC_STREAMING_REQ,32) descr="."></metric>
<metric name="TCC_NC_REQ_sum" expr=sum(TCC_NC_REQ,32) descr="."></metric>
<metric name="TCC_UC_REQ_sum" expr=sum(TCC_UC_REQ,32) descr="."></metric>
<metric name="TCC_CC_REQ_sum" expr=sum(TCC_CC_REQ,32) descr="."></metric>
<metric name="TCC_RW_REQ_sum" expr=sum(TCC_RW_REQ,32) descr="."></metric>
<metric name="TCC_PROBE_sum" expr=sum(TCC_PROBE,32) descr="."></metric>
<metric name="TCC_PROBE_ALL_sum" expr=sum(TCC_PROBE_ALL,32) descr="."></metric>
<metric name="TCC_READ_sum" expr=sum(TCC_READ,32) descr="."></metric>
<metric name="TCC_WRITE_sum" expr=sum(TCC_WRITE,32) descr="."></metric>
<metric name="TCC_ATOMIC_sum" expr=sum(TCC_ATOMIC,32) descr="."></metric>
<metric name="TCC_HIT_sum" expr=sum(TCC_HIT,32) descr="."></metric>
<metric name="TCC_MISS_sum" expr=sum(TCC_MISS,32) descr="."></metric>
<metric name="TCC_TAG_STALL_sum" expr=sum(TCC_TAG_STALL,32) descr="."></metric>
<metric name="TCC_WRITEBACK_sum" expr=sum(TCC_WRITEBACK,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=sum(TCC_EA_WRREQ,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=sum(TCC_EA_WRREQ_64B,32) descr="."></metric>
<metric name="TCC_EA_WR_UNCACHED_32B_sum" expr=sum(TCC_EA_WR_UNCACHED_32B,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_STALL_sum" expr=sum(TCC_EA_WRREQ_STALL,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_IO_CREDIT_STALL_sum" expr=sum(TCC_EA_WRREQ_IO_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_GMI_CREDIT_STALL_sum" expr=sum(TCC_EA_WRREQ_GMI_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_DRAM_CREDIT_STALL_sum" expr=sum(TCC_EA_WRREQ_DRAM_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_TOO_MANY_EA_WRREQS_STALL_sum" expr=sum(TCC_TOO_MANY_EA_WRREQS_STALL,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_LEVEL_sum" expr=sum(TCC_EA_WRREQ_LEVEL,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_LEVEL_sum" expr=sum(TCC_EA_RDREQ_LEVEL,32) descr="."></metric>
<metric name="TCC_EA_ATOMIC_sum" expr=sum(TCC_EA_ATOMIC,32) descr="."></metric>
<metric name="TCC_EA_ATOMIC_LEVEL_sum" expr=sum(TCC_EA_ATOMIC_LEVEL,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=sum(TCC_EA_RDREQ,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=sum(TCC_EA_RDREQ_32B,32) descr="."></metric>
<metric name="TCC_EA_RD_UNCACHED_32B_sum" expr=sum(TCC_EA_RD_UNCACHED_32B,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_IO_CREDIT_STALL_sum" expr=sum(TCC_EA_RDREQ_IO_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_GMI_CREDIT_STALL_sum" expr=sum(TCC_EA_RDREQ_GMI_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum" expr=sum(TCC_EA_RDREQ_DRAM_CREDIT_STALL,32) descr="."></metric>
<metric name="TCC_NORMAL_WRITEBACK_sum" expr=sum(TCC_NORMAL_WRITEBACK,32) descr="."></metric>
<metric name="TCC_ALL_TC_OP_WB_WRITEBACK_sum" expr=sum(TCC_ALL_TC_OP_WB_WRITEBACK,32) descr="."></metric>
<metric name="TCC_NORMAL_EVICT_sum" expr=sum(TCC_NORMAL_EVICT,32) descr="."></metric>
<metric name="TCC_ALL_TC_OP_INV_EVICT_sum" expr=sum(TCC_ALL_TC_OP_INV_EVICT,32) descr="."></metric>
<metric name="TCC_EA_RDREQ_DRAM_sum" expr=sum(TCC_EA_RDREQ_DRAM,32) descr="."></metric>
<metric name="TCC_EA_WRREQ_DRAM_sum" expr=sum(TCC_EA_WRREQ_DRAM,32) descr="."></metric>
</gfx908_expr>
<gfx9 base="gfx9_expr"></gfx9>
<gfx908 base="gfx908_expr"> </gfx908>