Merge remote-tracking branch 'origin/develop' into default_to_per_kernel

ROCm · Feb 27, 2025 · ee842b4 · ee842b4
2 parents 152b26a + bec537e
commit ee842b4
Show file tree

Hide file tree

Showing 29 changed files with 358 additions and 134 deletions.
diff --git a/.azuredevops/rocm-ci.yml b/.azuredevops/rocm-ci.yml
@@ -13,6 +13,7 @@ trigger:
   batch: true
   branches:
     include:
+    - develop
     - amd-staging
     - amd-mainline
   paths:
@@ -29,6 +30,7 @@ pr:
   autoCancel: true
   branches:
     include:
+    - develop
     - amd-staging
     - amd-mainline
   paths:

diff --git a/.github/workflows/rhel-8.yml b/.github/workflows/rhel-8.yml
@@ -27,6 +27,8 @@ jobs:
     steps:
       - name: Install baseline OS dependencies
         run: |
+          yum clean all
+          yum makecache
           yum -y install git
           yum -y install python39
           yum -y install cmake3

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -381,6 +381,8 @@ add_custom_target(
     COMMAND ${Python3_EXECUTABLE} -m pip list | grep -i nuitka > /dev/null 2>&1
     # Check patchelf
     COMMAND ${Python3_EXECUTABLE} -m pip list | grep -i patchelf > /dev/null 2>&1
+    # Create VERSION.sha file
+    COMMAND git -C ${PROJECT_SOURCE_DIR} rev-parse HEAD > VERSION.sha
     # Build standalone binary
     COMMAND
         ${Python3_EXECUTABLE} -m nuitka --mode=onefile

diff --git a/docker/Dockerfile.standalone b/docker/Dockerfile.standalone
@@ -2,7 +2,10 @@ FROM redhat/ubi8:8.10-1184
 
 WORKDIR /app
 
-RUN yum install -y curl gcc cmake
+RUN yum install -y curl gcc cmake git
+
+# Allows running git commands in /app
+RUN git config --global --add safe.directory /app
 
 RUN yum install -y python38 python38-devel && \
     yum clean all && \

diff --git a/docker/Dockerfile.test b/docker/Dockerfile.test
@@ -6,10 +6,13 @@ WORKDIR /app
 
 # Update package list and install prerequisites
 RUN apt-get update && apt-get install -y \
-    software-properties-common cmake locales \
+    software-properties-common cmake locales git \
     && add-apt-repository ppa:deadsnakes/ppa \
     && apt-get update
 
+# Allows running git commands in /app
+RUN git config --global --add safe.directory /app
+
 # Generate the desired locale
 RUN locale-gen en_US.UTF-8
 

diff --git a/src/rocprof_compute_profile/profiler_base.py b/src/rocprof_compute_profile/profiler_base.py
@@ -26,6 +26,7 @@
 import logging
 import os
 import re
+import shutil
 import sys
 import time
 from abc import ABC, abstractmethod
@@ -77,6 +78,26 @@ def join_prof(self, out=None):
                 out = self.__args.path + "/pmc_perf.csv"
             files = glob.glob(self.__args.path + "/" + "pmc_perf_*.csv")
             files.extend(glob.glob(self.__args.path + "/" + "SQ_*.csv"))
+
+            if self.get_args().hip_trace:
+                # remove hip api trace ouputs from this list
+                files = [
+                    f
+                    for f in files
+                    if not re.compile(r"^.*_hip_api_trace\.csv$").match(
+                        os.path.basename(f)
+                    )
+                ]
+
+            if self.get_args().kokkos_trace:
+                # remove marker api trace ouputs from this list
+                files = [
+                    f
+                    for f in files
+                    if not re.compile(r"^.*_marker_api_trace\.csv$").match(
+                        os.path.basename(f)
+                    )
+                ]
         elif type(self.__args.path) == list:
             files = self.__args.path
         else:
@@ -266,7 +287,8 @@ def pre_processing(self):
         # verify correct formatting for application binary
         self.__args.remaining = self.__args.remaining[1:]
         if self.__args.remaining:
-            if not Path(self.__args.remaining[0]).is_file():
+            # Ensure that command points to an executable
+            if not shutil.which(self.__args.remaining[0]):
                 console_error(
                     "Your command %s doesn't point to a executable. Please verify."
                     % self.__args.remaining[0]

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx906/0200_system-speed-of-light.yaml b/src/rocprof_compute_soc/analysis_configs/gfx906/0200_system-speed-of-light.yaml
@@ -32,6 +32,12 @@ Panel Config:
             peak: (((($max_sclk * $cu_per_gpu) * 64) * 2) / 1000)
             pop: None # No perf counter
             tips:
+          MFMA FLOPs (F8):
+            value: None # No HW module
+            unit: GFLOP
+            peak: None # No HW module
+            pop: None # No HW module
+            tips:
           MFMA FLOPs (BF16):
             value: None # No perf counter
             unit: GFLOPs

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx906/1000_compute-unit-instruction-mix.yaml b/src/rocprof_compute_soc/analysis_configs/gfx906/1000_compute-unit-instruction-mix.yaml
@@ -241,6 +241,12 @@ Panel Config:
             max: None # No HW module
             unit: (instr + $normUnit)
             tips:
+          MFMA-F8:
+            avg: None # No HW module
+            min: None # No HW module
+            max: None # No HW module None # No HW module
+            unit: (instr + $normUnit)
+            tips:
           MFMA-F16:
             avg: None # No HW module
             min: None # No HW module

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx906/1100_compute-unit-compute-pipeline.yaml b/src/rocprof_compute_soc/analysis_configs/gfx906/1100_compute-unit-compute-pipeline.yaml
@@ -21,16 +21,22 @@ Panel Config:
         metric:
           VALU FLOPs:
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           VALU IOPs:
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
+          MFMA FLOPs (F8):
+            value: None # No perf counter
+            unit: GFLOP
+            peak: None # No perf counter
+            pop: None # No perf counter
+            tips:
           MFMA FLOPs (BF16):
             value: None # No perf counter
             Unit: None
@@ -39,25 +45,25 @@ Panel Config:
             tips:
           MFMA FLOPs (F16):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA FLOPs (F32):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA FLOPs (F64):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA IOPs (INT8):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
@@ -174,6 +180,12 @@ Panel Config:
             max: None # No perf counter
             unit: (OPs  + $normUnit)
             tips:
+          F8 OPs:
+            avg: None # No HW module
+            min: None # No HW module
+            max: None # No HW module
+            unit: (OPs  + $normUnit)
+            tips:
           F16 OPs:
             avg: None # No perf counter
             min: None # No perf counter

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx908/0200_system-speed-of-light.yaml b/src/rocprof_compute_soc/analysis_configs/gfx908/0200_system-speed-of-light.yaml
@@ -32,6 +32,12 @@ Panel Config:
             peak: (((($max_sclk * $cu_per_gpu) * 64) * 2) / 1000)
             pop: None # No perf counter
             tips:
+          MFMA FLOPs (F8):
+            value: None # No HW module
+            unit: GFLOP
+            peak: None # No HW module
+            pop: None # No HW module
+            tips:
           MFMA FLOPs (BF16):
             value: None # No perf counter
             unit: GFLOPs

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx908/1000_compute-unit-instruction-mix.yaml b/src/rocprof_compute_soc/analysis_configs/gfx908/1000_compute-unit-instruction-mix.yaml
@@ -73,13 +73,13 @@ Panel Config:
           unit: Unit
           tips: Tips
         metric:
-          INT-32:
+          INT32:
             avg: None # No perf counter
             min: None # No perf counter
             max: None # No perf counter
             unit: (instr + $normUnit)
             tips:
-          INT-64:
+          INT64:
             avg: None # No perf counter
             min: None # No perf counter
             max: None # No perf counter
@@ -241,6 +241,12 @@ Panel Config:
             max: None # No HW module
             unit: (instr + $normUnit)
             tips:
+          MFMA-F8:
+            avg: None # No HW module
+            min: None # No HW module
+            max: None # No HW module None # No HW module
+            unit: (instr + $normUnit)
+            tips:
           MFMA-F16:
             avg: None # No HW module
             min: None # No HW module

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx908/1100_compute-unit-compute-pipeline.yaml b/src/rocprof_compute_soc/analysis_configs/gfx908/1100_compute-unit-compute-pipeline.yaml
@@ -21,16 +21,22 @@ Panel Config:
         metric:
           VALU FLOPs:
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           VALU IOPs:
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
+          MFMA FLOPs (F8):
+            value: None # No perf counter
+            unit: GFLOP
+            peak: None # No perf counter
+            pop: None # No perf counter
+            tips:
           MFMA FLOPs (BF16):
             value: None # No perf counter
             Unit: None
@@ -39,25 +45,25 @@ Panel Config:
             tips:
           MFMA FLOPs (F16):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA FLOPs (F32):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA FLOPs (F64):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
           MFMA IOPs (INT8):
             value: None # No perf counter
-            Unit: None
+            unit: None
             peak: None
             pop: None
             tips:
@@ -174,6 +180,12 @@ Panel Config:
             max: None # No perf counter
             unit: (OPs  + $normUnit)
             tips:
+          F8 OPs:
+            avg: None # No HW module
+            min: None # No HW module
+            max: None # No HW module
+            unit: (OPs  + $normUnit)
+            tips:
           F16 OPs:
             avg: None # No perf counter
             min: None # No perf counter

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx90a/0200_system-speed-of-light.yaml b/src/rocprof_compute_soc/analysis_configs/gfx90a/0200_system-speed-of-light.yaml
@@ -42,6 +42,12 @@ Panel Config:
             pop: ((100 * AVG(((64 * (SQ_INSTS_VALU_INT32 + SQ_INSTS_VALU_INT64)) / (End_Timestamp
               - Start_Timestamp)))) / (((($max_sclk * $cu_per_gpu) * 64) * 2) / 1000))
             tips:
+          MFMA FLOPs (F8):
+            value: None
+            unit: GFLOP
+            peak: None
+            pop: None
+            tips:
           MFMA FLOPs (BF16):
             value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_BF16 * 512) / (End_Timestamp - Start_Timestamp)))
             unit: GFLOP

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx90a/1000_compute-unit-instruction-mix.yaml b/src/rocprof_compute_soc/analysis_configs/gfx90a/1000_compute-unit-instruction-mix.yaml
@@ -241,6 +241,12 @@ Panel Config:
             max: MAX((SQ_INSTS_VALU_MFMA_I8 / $denom))
             unit: (instr + $normUnit)
             tips:
+          MFMA-F8:
+            avg: None
+            min: None
+            max: None
+            unit: (instr + $normUnit)
+            tips:
           MFMA-F16:
             avg: AVG((SQ_INSTS_VALU_MFMA_F16 / $denom))
             min: MIN((SQ_INSTS_VALU_MFMA_F16 / $denom))

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx90a/1100_compute-unit-compute-pipeline.yaml b/src/rocprof_compute_soc/analysis_configs/gfx90a/1100_compute-unit-compute-pipeline.yaml
@@ -41,6 +41,12 @@ Panel Config:
             pop: ((100 * AVG(((64 * (SQ_INSTS_VALU_INT32 + SQ_INSTS_VALU_INT64)) / (End_Timestamp
               - Start_Timestamp)))) / (((($max_sclk * $cu_per_gpu) * 64) * 2) / 1000))
             tips:
+          MFMA FLOPs (F8):
+            value: None
+            unit: GFLOP
+            peak: None
+            pop: None
+            tips:
           MFMA FLOPs (BF16):
             value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_BF16 * 512) / (End_Timestamp - Start_Timestamp)))
             unit: GFLOP
@@ -216,6 +222,12 @@ Panel Config:
             max: MAX(((64 * (SQ_INSTS_VALU_INT32 + SQ_INSTS_VALU_INT64)) + (SQ_INSTS_VALU_MFMA_MOPS_I8 * 512)) / $denom)
             unit: (OPs  + $normUnit)
             tips:
+          F8 OPs:
+            avg: None
+            min: None
+            max: None
+            unit: (OPs  + $normUnit)
+            tips:
           F16 OPs:
             avg: AVG(((((((64 * SQ_INSTS_VALU_ADD_F16) + (64 * SQ_INSTS_VALU_MUL_F16)) +
               (64 * SQ_INSTS_VALU_TRANS_F16)) + (128 * SQ_INSTS_VALU_FMA_F16)) + (512 *

diff --git a/src/rocprof_compute_soc/analysis_configs/gfx940/0200_system-speed-of-light.yaml b/src/rocprof_compute_soc/analysis_configs/gfx940/0200_system-speed-of-light.yaml
@@ -42,6 +42,13 @@ Panel Config:
             pop: ((100 * AVG(((64 * (SQ_INSTS_VALU_INT32 + SQ_INSTS_VALU_INT64)) / (End_Timestamp
               - Start_Timestamp)))) / (((($max_sclk * $cu_per_gpu) * 64) * 2) / 1000))
             tips:
+          MFMA FLOPs (F8):
+            value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_F8 * 512) / (End_Timestamp - Start_Timestamp)))
+            unit: GFLOP
+            peak: ((($max_sclk * $cu_per_gpu) * 4096) / 1000)
+            pop: ((100 * AVG(((SQ_INSTS_VALU_MFMA_MOPS_F8 * 512) / (End_Timestamp - Start_Timestamp))))
+              / ((($max_sclk * $cu_per_gpu) * 8192) / 1000))
+            tips:
           MFMA FLOPs (BF16):
             value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_BF16 * 512) / (End_Timestamp - Start_Timestamp)))
             unit: GFLOP
@@ -187,12 +194,14 @@ Panel Config:
               / ((($max_sclk / 1000) * 128) * TO_INT($total_l2_chan)))
             tips:
           L2-Fabric Read BW:
-            value: AVG((((TCC_EA0_RDREQ_32B_sum * 32) + ((TCC_EA0_RDREQ_sum - TCC_EA0_RDREQ_32B_sum)
-              * 64)) / (End_Timestamp - Start_Timestamp)))
+            value: AVG((128 * TCC_BUBBLE_sum +
+                        64 * (TCC_EA0_RDREQ_sum - TCC_BUBBLE_sum - TCC_EA0_RDREQ_32B_sum) +
+                        32 * TCC_EA0_RDREQ_32B_sum) / (End_Timestamp - Start_Timestamp))
             unit: GB/s
             peak: $hbm_bw
-            pop: ((100 * AVG((((TCC_EA0_RDREQ_32B_sum * 32) + ((TCC_EA0_RDREQ_sum - TCC_EA0_RDREQ_32B_sum)
-              * 64)) / (End_Timestamp - Start_Timestamp)))) / $hbm_bw)
+            pop: ((100 * (AVG((128 * TCC_BUBBLE_sum +
+                        64 * (TCC_EA0_RDREQ_sum - TCC_BUBBLE_sum - TCC_EA0_RDREQ_32B_sum) +
+                        32 * TCC_EA0_RDREQ_32B_sum) / (End_Timestamp - Start_Timestamp)))) / $hbm_bw)
             tips:
           L2-Fabric Write BW:
             value: AVG((((TCC_EA0_WRREQ_64B_sum * 64) + ((TCC_EA0_WRREQ_sum - TCC_EA0_WRREQ_64B_sum)