Merge branch 'dev' of github.com:AMDResearch/omniperf into mi300

Signed-off-by: coleramos425 <colramos@amd.com>

.github/workflows/rhel-8.yml

@@ -33,9 +33,6 @@ jobs:
           yum -y install which
       - name: Checkout
         uses: actions/checkout@v3
-        with:
-          submodules: recursive
-          token: ${{ secrets.GH_PAT }}
       - name: Install Python prereqs
         run: |
           python3.9 -m pip install -r requirements.txt

.github/workflows/ubuntu-focal.yml

@@ -32,9 +32,6 @@ jobs:
           sudo apt-get install -y cmake
       - name: Checkout
         uses: actions/checkout@v3
-        with:
-          submodules: recursive
-          token: ${{ secrets.GH_PAT }}
       - name: Install Python prereqs
         run: |
           python3 -m pip install -r requirements.txt

CHANGES

@@ -1,9 +1,31 @@
+Version 1.0.8 (30 May 2023)
+
+  * add `--kernel-names` option to toggle kernelName overlay in standalone roofline plot (#93)
+  * remove unused python modules (#96)
+  * fix empirical roofline calculation for single dispatch workloads (#97)
+  * match color of arithmetic intensity points to corresponding bw lines
+
+  * ux improvements in standalone GUI (#101)
+  * enhanced readability for filtering dropdowns in standalone GUI (#102)
+  * new logfile to capture rocprofiler output (#106)
+  * roofline support for sles15 sp4 and future service packs (#109)
+  * adding dockerfiles for all supported Linux distros
+  * new examples for `--roof-only` and `--kernel` options added to documentation
+
+  * enable cli analysis in Windows (#110)
+  * optional random port number in standalone GUI (#111)
+  * limit length of visible kernelName in `--kernel-names` option (#115)
+  * adjust metric definitions (#117, #130)
+  * manually merge rocprof runs, overriding default rocprofiler implementation (#125)
+  * fixed compatibility issues with Python 3.11 (#131)
+
 Version 1.0.8-PR2 (17 Apr 2023)
+
   * ux improvements in standalone GUI (#101)
   * enhanced readability for filtering dropdowns in standalone GUI (#102)
   * new logfile to capture rocprofiler output (#106)
   * roofline support for sles15 sp4 and future service packs (#109)
-  * adding dockerfiles for all supported Linux distos
+  * adding dockerfiles for all supported Linux distros
   * new examples for `--roof-only` and `--kernel` options added to documentation
 
 Version 1.0.8-PR1 (13 Mar 2023)

README.md

@@ -40,18 +40,19 @@ This software can be cited using a Zenodo
 style reference is provided below for convenience:
 
 ```
-@software{xiamin_lu_2022_7314631
+@software{xiaomin_lu_2022_7314631
   author       = {Xiaomin Lu and
                   Cole Ramos and
                   Fei Zheng and
                   Karl W. Schulz and
                   Jose Santos and
-                  Keith Lowery},
-  title        = {AMDResearch/omniperf: v1.0.8-PR2 (17 April 2023)},
-  month        = apr,
+                  Keith Lowery and
+                  Cristian Di Pietrantonio},
+  title        = {AMDResearch/omniperf: v1.0.8 (30 May 2023)},
+  month        = may,
   year         = 2023,
   publisher    = {Zenodo},
-  version      = {v1.0.8-PR1},
+  version      = {v1.0.8},
   doi          = {10.5281/zenodo.7314631},
   url          = {https://doi.org/10.5281/zenodo.7314631}
 }

VERSION

@@ -1 +1 @@
-1.0.8-PR2
+1.0.8

src/omniperf

@@ -38,7 +38,7 @@ from pathlib import Path as path
 
 from parser import parse
 from utils import specs
-from utils.perfagg import perfmon_filter, pmc_filter
+from utils.perfagg import perfmon_filter, pmc_filter, pmc_perf_split, join_prof
 from utils import remove_workload
 from utils import csv_converter  # Import workload
 from omniperf_analyze.omniperf_analyze import roofline_only  # Standalone roofline
@@ -169,7 +169,7 @@ def gen_sysinfo(workload_name, workload_dir, ip_blocks, app_cmd, skip_roof):
     header += "command,"
     header += "host_name,host_cpu,host_distro,host_kernel,host_rocmver,date,"
     header += "gpu_soc,numSE,numCU,numSIMD,waveSize,maxWavesPerCU,maxWorkgroupSize,"
-    header += "L1,L2,sclk,mclk,cur_sclk,cur_mclk,L2Banks,name,numSQC,hbmBW,"
+    header += "L1,L2,sclk,mclk,cur_sclk,cur_mclk,L2Banks,LDSBanks,name,numSQC,hbmBW,"
     header += "ip_blocks\n"
     sysinfo.write(header)
 
@@ -213,11 +213,11 @@ def gen_sysinfo(workload_name, workload_dir, ip_blocks, app_cmd, skip_roof):
     blocks = []
     hbmBW = int(mspec.cur_MCLK) / 1000 * 4096 / 8 * 2
     if mspec.GPU == "gfx906":
-        param += ["16", "mi50", str(int(mspec.CU) // 4), str(hbmBW)]
+        param += ["16", "32", "mi50", str(int(mspec.CU) // 4), str(hbmBW)]
     elif mspec.GPU == "gfx908":
-        param += ["32", "mi100", "48", str(hbmBW)]
+        param += ["32", "32", "mi100", "48", str(hbmBW)]
     elif mspec.GPU == "gfx90a":
-        param += ["32", "mi200", "56", str(hbmBW)]
+        param += ["32", "32", "mi200", "56", str(hbmBW)]
         if not skip_roof:
             blocks.append("roofline")
 
@@ -380,6 +380,9 @@ def characterize_app(args, VER):
     # Perfmon filtering
     pmc_filter(workload_dir, perfmon_dir, args.target)
 
+    # Separate pmc_perf runs
+    pmc_perf_split(workload_dir)
+
     # Set up a log file
     log = open(workload_dir + "/log.txt", "w")
     print("Log: ", workload_dir + "/log.txt\n")
@@ -388,7 +391,7 @@ def characterize_app(args, VER):
     for fname in glob.glob(workload_dir + "/perfmon/*.txt"):
         # Kernel filtering (in-place replacement)
         if not args.kernel == None:
-            run_subprocess(
+            success, output = capture_subprocess_output(
                 [
                     "sed",
                     "-i",
@@ -397,10 +400,11 @@ def characterize_app(args, VER):
                     fname,
                 ]
             )
+            log.write(output)
 
         # Dispatch filtering (inplace replacement)
         if not args.dispatch == None:
-            run_subprocess(
+            success, output = capture_subprocess_output(
                 [
                     "sed",
                     "-i",
@@ -409,14 +413,14 @@ def characterize_app(args, VER):
                     fname,
                 ]
             )
+            log.write(output)
         print(fname)
         if args.use_rocscope == True:
             run_rocscope(args, fname)
         else:
             run_prof(fname, workload_dir, perfmon_dir, app_cmd, args.target, log, args.verbose)
 
-    # Close log
-    log.close()
+
 
 
 ################################################
@@ -521,6 +525,9 @@ def omniperf_profile(args, VER):
     # Perfmon filtering
     perfmon_filter(workload_dir, perfmon_dir, args)
 
+    # Separate pmc_perf runs
+    pmc_perf_split(workload_dir)
+
     # Set up a log file
     log = open(workload_dir + "/log.txt", "w")
     print("Log: ", workload_dir + "/log.txt\n")
@@ -598,7 +605,7 @@ def omniperf_profile(args, VER):
         for fname in glob.glob(workload_dir + "/perfmon/*.txt"):
             # Kernel filtering (in-place replacement)
             if not args.kernel == None:
-                run_subprocess(
+                success, output = capture_subprocess_output(
                     [
                         "sed",
                         "-i",
@@ -607,10 +614,11 @@ def omniperf_profile(args, VER):
                         fname,
                     ]
                 )
+                log.write(output)
 
             # Dispatch filtering (inplace replacement)
             if not args.dispatch == None:
-                run_subprocess(
+                success, output = capture_subprocess_output(
                     [
                         "sed",
                         "-i",
@@ -619,12 +627,17 @@ def omniperf_profile(args, VER):
                         fname,
                     ]
                 )
+                log.write(output)
             print(fname)
             if args.use_rocscope == True:
                 run_rocscope(args, fname)
             else:
                 run_prof(fname, workload_dir, perfmon_dir, args.remaining, args.target, log, args.verbose)
 
+        # Manually join each pmc_perf*.csv output
+        if args.use_rocscope == False:
+            join_prof(workload_dir, args.join_type, log, args.verbose)
+
     # Generate sysinfo
     gen_sysinfo(args.name, workload_dir, args.ipblocks, args.remaining, args.no_roof)
 

src/omniperf_analyze/configs/gfx906/0200_system-speed-of-light.yaml

@@ -104,11 +104,11 @@ Panel Config:
               / SQ_ACTIVE_INST_ANY))) / 5)
             tips: 
           LDS BW:
-            value: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            value: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / (EndNs - BeginNs)))
             unit: GB/sec
             peak: (($sclk * $numCU) * 0.128)
-            pop: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            pop: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / (EndNs - BeginNs)) / (($sclk * $numCU) * 0.00128)))
             tips: 
           LDS Bank Conflict:

src/omniperf_analyze/configs/gfx906/0700_wavefront-launch.yaml

@@ -50,9 +50,15 @@ Panel Config:
             unit: Wavefronts
             tips: 
           VGPRs:
-            avg: AVG(vgpr)
-            min: MIN(vgpr)
-            max: MAX(vgpr)
+            avg: AVG(arch_vgpr)
+            min: MIN(arch_vgpr)
+            max: MAX(arch_vgpr)
+            unit: Registers
+            tips: 
+          AGPRs:
+            avg: AVG(accum_vgpr)
+            min: MIN(accum_vgpr)
+            max: MAX(accum_vgpr)
             unit: Registers
             tips: 
           SGPRs:

src/omniperf_analyze/configs/gfx906/1200_lds.yaml

@@ -23,7 +23,7 @@ Panel Config:
             value: AVG(((200 * SQ_ACTIVE_INST_LDS) / (GRBM_GUI_ACTIVE * $numCU)))
             tips: 
           Bandwidth (Pct-of-Peak):
-            value: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            value: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / (EndNs - BeginNs)) / (($sclk * $numCU) * 0.00128)))
             tips: 
           Bank Conflict Rate:
@@ -42,24 +42,18 @@ Panel Config:
           unit: Unit
           tips: Tips
         metric:
-          Wave Cycles:
-            avg: AVG(((4 * SQ_WAVE_CYCLES) / SQ_WAVES))
-            min: MIN(((4 * SQ_WAVE_CYCLES) / SQ_WAVES))
-            max: MAX(((4 * SQ_WAVE_CYCLES) / SQ_WAVES))
-            unit: Cycles/Wave
-            tips: 
           LDS Instrs:
             avg: AVG((SQ_INSTS_LDS / $denom))
             min: MIN((SQ_INSTS_LDS / $denom))
             max: MAX((SQ_INSTS_LDS / $denom))
             unit: (Instr  + $normUnit)
             tips: 
           Bandwidth:
-            avg: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            avg: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / $denom))
-            min: MIN(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            min: MIN(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / $denom))
-            max: MAX(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            max: MAX(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / $denom))
             unit: (Bytes  + $normUnit)
             tips: 

src/omniperf_analyze/configs/gfx906/1800_L2_cache_per_channel.yaml

@@ -204,7 +204,7 @@ Panel Config:
               + TO_INT(TCC_REQ[22])) + TO_INT(TCC_REQ[23])) + TO_INT(TCC_REQ[24])) + TO_INT(TCC_REQ[25]))
               + TO_INT(TCC_REQ[26])) + TO_INT(TCC_REQ[27])) + TO_INT(TCC_REQ[28])) + TO_INT(TCC_REQ[29]))
               + TO_INT(TCC_REQ[30])) + TO_INT(TCC_REQ[31])) / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L1 - L2 Read Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_READ[0]) + TO_INT(TCC_READ[1]))
@@ -247,7 +247,7 @@ Panel Config:
               + TO_INT(TCC_READ[24])) + TO_INT(TCC_READ[25])) + TO_INT(TCC_READ[26])) +
               TO_INT(TCC_READ[27])) + TO_INT(TCC_READ[28])) + TO_INT(TCC_READ[29])) + TO_INT(TCC_READ[30]))
               + TO_INT(TCC_READ[31])) / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L1 - L2 Write Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_WRITE[0]) + TO_INT(TCC_WRITE[1]))
@@ -294,7 +294,7 @@ Panel Config:
               + TO_INT(TCC_WRITE[24])) + TO_INT(TCC_WRITE[25])) + TO_INT(TCC_WRITE[26]))
               + TO_INT(TCC_WRITE[27])) + TO_INT(TCC_WRITE[28])) + TO_INT(TCC_WRITE[29]))
               + TO_INT(TCC_WRITE[30])) + TO_INT(TCC_WRITE[31])) / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L1 - L2 Atomic Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_ATOMIC[0]) + TO_INT(TCC_ATOMIC[1]))
@@ -345,7 +345,7 @@ Panel Config:
               + TO_INT(TCC_ATOMIC[26])) + TO_INT(TCC_ATOMIC[27])) + TO_INT(TCC_ATOMIC[28]))
               + TO_INT(TCC_ATOMIC[29])) + TO_INT(TCC_ATOMIC[30])) + TO_INT(TCC_ATOMIC[31]))
               / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L2 - EA Read Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_EA_RDREQ[0]) + TO_INT(TCC_EA_RDREQ[1]))
@@ -396,7 +396,7 @@ Panel Config:
               + TO_INT(TCC_EA_RDREQ[26])) + TO_INT(TCC_EA_RDREQ[27])) + TO_INT(TCC_EA_RDREQ[28]))
               + TO_INT(TCC_EA_RDREQ[29])) + TO_INT(TCC_EA_RDREQ[30])) + TO_INT(TCC_EA_RDREQ[31]))
               / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L2 - EA Write Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_EA_WRREQ[0]) + TO_INT(TCC_EA_WRREQ[1]))
@@ -447,7 +447,7 @@ Panel Config:
               + TO_INT(TCC_EA_WRREQ[26])) + TO_INT(TCC_EA_WRREQ[27])) + TO_INT(TCC_EA_WRREQ[28]))
               + TO_INT(TCC_EA_WRREQ[29])) + TO_INT(TCC_EA_WRREQ[30])) + TO_INT(TCC_EA_WRREQ[31]))
               / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L2 - EA Atomic Req:
             mean: AVG((((((((((((((((((((((((((((((((((TO_INT(TCC_EA_ATOMIC[0]) + TO_INT(TCC_EA_ATOMIC[1]))
@@ -498,7 +498,7 @@ Panel Config:
               + TO_INT(TCC_EA_ATOMIC[26])) + TO_INT(TCC_EA_ATOMIC[27])) + TO_INT(TCC_EA_ATOMIC[28]))
               + TO_INT(TCC_EA_ATOMIC[29])) + TO_INT(TCC_EA_ATOMIC[30])) + TO_INT(TCC_EA_ATOMIC[31]))
               / 32) / $denom))
-            units: ( $normUnit )
+            units: ( + $normUnit)
             tips: 
           L2 - EA Read Lat:
             mean: AVG((((((((((((((((((((((((((((((((((TCC_EA_RDREQ_LEVEL[0] + TCC_EA_RDREQ_LEVEL[1])

src/omniperf_analyze/configs/gfx908/0200_system-speed-of-light.yaml

@@ -104,11 +104,11 @@ Panel Config:
               / SQ_ACTIVE_INST_ANY))) / 5)
             tips: 
           LDS BW:
-            value: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            value: AVG(((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / (EndNs - BeginNs)))
             unit: GB/sec
             peak: (($sclk * $numCU) * 0.128)
-            pop: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($L2Banks))
+            pop: AVG((((((SQ_LDS_IDX_ACTIVE - SQ_LDS_BANK_CONFLICT) * 4) * TO_INT($LDSBanks))
               / (EndNs - BeginNs)) / (($sclk * $numCU) * 0.00128)))
             tips: 
           LDS Bank Conflict:

src/omniperf_analyze/configs/gfx908/0700_wavefront-launch.yaml

@@ -50,9 +50,15 @@ Panel Config:
             unit: Wavefronts
             tips: 
           VGPRs:
-            avg: AVG(vgpr)
-            min: MIN(vgpr)
-            max: MAX(vgpr)
+            avg: AVG(arch_vgpr)
+            min: MIN(arch_vgpr)
+            max: MAX(arch_vgpr)
+            unit: Registers
+            tips: 
+          AGPRs:
+            avg: AVG(accum_vgpr)
+            min: MIN(accum_vgpr)
+            max: MAX(accum_vgpr)
             unit: Registers
             tips: 
           SGPRs: