Bugfixes (#1971)

* Implement AMDGPU driver info and GPU VRAM attributes in system info.
  section of analysis report.

* Backward compatibility for rocprofiler-sdk avail module path migration

* Fix roofline calculation where AI data points are N/A

projects/rocprofiler-compute/CHANGELOG.md

@@ -19,6 +19,8 @@ Full documentation for ROCm Compute Profiler is available at [https://rocm.docs.
     * kernel: Counters are collected in a round robin fashion for unique kernels.
     * kernel_launch_params: Counters are collected in a round robin fashion for unique kernels having the exact same launch parameters.
 
+* Implement AMDGPU driver info and GPU VRAM attributes in system info. section of analysis report.
+
 ### Changed
 
 * Default output format for the underlying ROCprofiler-SDK tool has been changed from ``csv`` to ``rocpd``.

projects/rocprofiler-compute/src/rocprof_compute_soc/soc_base.py

@@ -414,13 +414,28 @@ class OmniSoC_Base:
         os.environ["ROCPROFILER_METRICS_PATH"] = str(
             config.rocprof_compute_home / "rocprof_compute_soc" / "profile_configs"
         )
-        sys.path.append(
-            str(
-                Path(args.rocprofiler_sdk_tool_path).parents[1]
-                / "python3/site-packages"
-            )
+
+        # Backward compatibility support for sdk avail module moved from
+        # <rocm_path>/bin/rocprofv3_avail_module/avail.py to
+        # <rocm_path>/lib/python3/site-packages/rocprofv3/avail.py
+        new_path = str(
+            Path(args.rocprofiler_sdk_tool_path).parents[1] / "python3/site-packages"
         )
-        from rocprofv3 import avail
+        old_path = str(Path(args.rocprofiler_sdk_tool_path).parents[2] / "bin")
+        try:
+            sys.path.append(new_path)
+            from rocprofv3 import avail
+        except ImportError:
+            console_debug(
+                f"Could not import rocprofiler-sdk avail module from {new_path}, "
+                f"trying {old_path}"
+            )
+            try:
+                sys.path.remove(new_path)
+                sys.path.append(old_path)
+                from rocprofv3_avail_module import avail
+            except ImportError:
+                console_error("Failed to import rocprofiler-sdk avail module.")
 
         avail.loadLibrary.libname = str(
             Path(args.rocprofiler_sdk_tool_path).parent / "librocprofv3-list-avail.so"

projects/rocprofiler-compute/src/utils/amdsmi_interface.py

@@ -134,3 +134,27 @@ def get_gpu_memory_partition() -> str:
     except Exception as e:
         console_warning(f"Error getting GPU memory partition: {e}")
         return "N/A"
+
+
+def get_amdgpu_driver_version() -> str:
+    """Get the AMDGPU driver version."""
+    try:
+        driver_info = amdsmi.amdsmi_get_gpu_driver_info(get_device_handle())
+        driver_version = driver_info["driver_version"]
+        console_debug(f"AMDGPU Driver Version: {driver_version}")
+        return driver_version
+    except Exception as e:
+        console_warning(f"Error getting AMDGPU driver version: {e}")
+        return "N/A"
+
+
+def get_gpu_vram_size() -> int:
+    """Get the GPU VRAM size in MB."""
+    try:
+        vram_info = amdsmi.amdsmi_get_gpu_vram_info(get_device_handle())
+        vram_size = str(int(vram_info["vram_size"]) * 1024)  # MB -> KB
+        console_debug(f"GPU VRAM Size: {vram_size} MB")
+        return vram_size
+    except Exception as e:
+        console_warning(f"Error getting GPU VRAM size: {e}")
+        return 0

projects/rocprofiler-compute/src/utils/roofline_calc.py

@@ -412,13 +412,13 @@ def calc_ai_analyze(
                 metric = row.get("Metric", "")
                 value = row.get("Value", 0)
                 if metric == "AI HBM":
-                    ai_hbm = value if value and value != "" else 0
+                    ai_hbm = value if value and value not in ("", "N/A") else 0
                 elif metric == "AI L2":
-                    ai_l2 = value if value and value != "" else 0
+                    ai_l2 = value if value and value not in ("", "N/A") else 0
                 elif metric == "AI L1":
-                    ai_l1 = value if value and value != "" else 0
+                    ai_l1 = value if value and value not in ("", "N/A") else 0
                 elif metric == "Performance (GFLOPs)":
-                    performance = value if value and value != "" else 0
+                    performance = value if value and value not in ("", "N/A") else 0
 
         console_debug(
             "roofline",

projects/rocprofiler-compute/src/utils/specs.py

@@ -43,9 +43,11 @@ import pandas as pd
 import config
 from utils.amdsmi_interface import (
     amdsmi_ctx,
+    get_amdgpu_driver_version,
     get_gpu_compute_partition,
     get_gpu_memory_partition,
     get_gpu_vbios_part_number,
+    get_gpu_vram_size,
 )
 from utils.logger import (
     console_debug,
@@ -182,25 +184,26 @@ def generate_machine_specs(
     soc_info = extract_soc_info()
 
     # Combine all specifications
-    specs = MachineSpecs(
-        version=specs_version,
-        timestamp=timestamp,
-        rocminfo_lines=soc_info["rocminfo_lines"],
-        hostname=socket.gethostname(),
-        cpu_model=machine_info["cpu_model"],
-        sbios=machine_info["sbios"],
-        linux_kernel_version=machine_info["linux_kernel_version"],
-        amd_gpu_kernel_version="",
-        cpu_memory=machine_info["cpu_memory"],
-        gpu_memory="",
-        linux_distro=machine_info["linux_distro"],
-        rocm_version=get_rocm_ver().strip(),
-        vbios=gpu_info["vbios"],
-        compute_partition=gpu_info["compute_partition"],
-        memory_partition=gpu_info["memory_partition"],
-        gpu_arch=soc_info["gpu_arch"],
-        gpu_chip_id=soc_info["gpu_chip_id"],
-    )
+    with amdsmi_ctx():
+        specs = MachineSpecs(
+            version=specs_version,
+            timestamp=timestamp,
+            rocminfo_lines=soc_info["rocminfo_lines"],
+            hostname=socket.gethostname(),
+            cpu_model=machine_info["cpu_model"],
+            sbios=machine_info["sbios"],
+            linux_kernel_version=machine_info["linux_kernel_version"],
+            amd_gpu_kernel_version=get_amdgpu_driver_version(),
+            cpu_memory=machine_info["cpu_memory"],
+            gpu_memory=get_gpu_vram_size(),
+            linux_distro=machine_info["linux_distro"],
+            rocm_version=get_rocm_ver().strip(),
+            vbios=gpu_info["vbios"],
+            compute_partition=gpu_info["compute_partition"],
+            memory_partition=gpu_info["memory_partition"],
+            gpu_arch=soc_info["gpu_arch"],
+            gpu_chip_id=soc_info["gpu_chip_id"],
+        )
 
     # Load above SoC specs via module import
     try:
@@ -436,10 +439,7 @@ class MachineSpecs:
     amd_gpu_kernel_version: Optional[str] = field(
         default=None,
         metadata={
-            "doc": (
-                "[RESERVED] The version of the AMDGPU driver installed on the machine. "
-                "Unimplemented."
-            ),
+            "doc": ("The version of the AMDGPU driver installed on the machine."),
             "name": "AMD GPU Kernel Version",
             "show_in_table": True,
         },
@@ -457,8 +457,8 @@ class MachineSpecs:
         default=None,
         metadata={
             "doc": (
-                "[RESERVED] The total amount of memory available to accelerators/GPUs "
-                "in the system. Unimplemented."
+                "The total amount of memory available to accelerators/GPUs "
+                "in the system."
             ),
             "unit": "KB",
             "name": "GPU Memory",