Add chip specs (#681)

* Add perfmon config spec, enhance memory partition info.

* Add gfx950 perfmon config.

* Add High Freq variants in gfx942.

* Add backup detection methods for gpu model.

* Improve get_num_xcds logic by adding detection of 1to1 arch-to-compute_partition logic.

* Add default compute partition settings spx:8 for when gpu_model=None.

* Update gpu spec tests.

* Add backup compute partition detection.

---------

Signed-off-by: xuchen-amd <xuchen@amd.com>

Author: xuchen-amd

src/rocprof_compute_soc/soc_base.py

@@ -186,12 +186,73 @@ def populate_mspec(self):
         self._mspec.gpu_model = mi_gpu_specs.get_gpu_model(
             self._mspec.gpu_arch, self._mspec.gpu_chip_id
         )
+
+        if not self._mspec.gpu_model:
+            self._mspec.gpu_model = self.detect_gpu_model(self._mspec.gpu_arch)
+
         self._mspec.num_xcd = str(
             mi_gpu_specs.get_num_xcds(
-                self._mspec.gpu_model, self._mspec.compute_partition
+                self._mspec.gpu_arch, self._mspec.gpu_model, self._mspec.compute_partition
             )
         )
 
+    @demarcate
+    def detect_gpu_model(self, gpu_arch):
+        """
+        Detects the GPU model using various identifiers from 'amd-smi static'.
+        Falls back through multiple methods if the primary method fails.
+        """
+
+        from utils.specs import run, search
+
+        # TODO: use amd-smi python api when available
+        amd_smi_static = run(["amd-smi", "static", "--gpu=0"], exit_on_error=True)
+
+        # Purposely search for patterns without variants suffix to try and match a known GPU model.
+        detection_methods = [
+            {
+                "name": "Market Name",
+                "pattern": r"MARKET_NAME:\s*.*(mi|MI\d*[a-zA-Z]*)",
+            },
+            {
+                "name": "VBIOS Name",
+                "pattern": r"NAME:\s*.*(mi|MI\d*[a-zA-Z]*)",
+            },
+            {"name": "Product Name", "pattern": r"PRODUCT_NAME:\s*.*(mi|MI\d*[a-zA-Z]*)"},
+        ]
+
+        gpu_model = None
+        for method in detection_methods:
+            console_log(f"Determining GPU model using {method['name']}.")
+            gpu_model = search(method["pattern"], amd_smi_static)
+            if gpu_model:
+                break
+
+        if not gpu_model:
+            console_warning("Unable to determine the GPU model.")
+            return
+
+        gpu_model = self._adjust_mi300_model(gpu_model.lower(), gpu_arch.lower())
+
+        if gpu_model.lower() not in mi_gpu_specs.get_num_xcds_dict().keys():
+            console_warning(f"Unknown GPU model detected: '{gpu_model}'.")
+            return
+
+        return gpu_model.upper()
+
+    def _adjust_mi300_model(self, gpu_model, gpu_arch):
+        """
+        Applies specific adjustments for MI300 series GPU models based on architecture.
+        """
+
+        if gpu_model in ["mi300a", "mi300x"]:
+            if gpu_arch in ["gfx940", "gfx941"]:
+                gpu_model += "_a0"
+            elif gpu_arch == "gfx942":
+                gpu_model += "_a1"
+
+        return gpu_model
+
     @demarcate
     def detect_counters(self):
         """

src/rocprof_compute_soc/soc_gfx906.py

@@ -22,11 +22,9 @@
 # SOFTWARE.
 ##############################################################################el
 
-from pathlib import Path
-
-import config
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from utils.logger import console_error, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 
 
 class gfx906_soc(OmniSoC_Base):
@@ -35,20 +33,7 @@ def __init__(self, args, mspec):
         self.set_arch("gfx906")
         self.set_compatible_profilers(["rocprofv1", "rocscope"])
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config({mi_gpu_specs.get_perfmon_config("gfx906")})
 
         # Set arch specific specs
         self._mspec._l2_banks = 16

src/rocprof_compute_soc/soc_gfx908.py

@@ -27,6 +27,7 @@
 import config
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from utils.logger import console_error, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 
 
 class gfx908_soc(OmniSoC_Base):
@@ -37,20 +38,7 @@ def __init__(self, args, mspec):
             ["rocprofv1", "rocscope", "rocprofv3", "rocprofiler-sdk"]
         )
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx908"))
 
         # Set arch specific specs
         self._mspec._l2_banks = 32

src/rocprof_compute_soc/soc_gfx90a.py

@@ -28,6 +28,7 @@
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from roofline import Roofline
 from utils.logger import console_log, console_warning, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 from utils.utils import mibench
 
 
@@ -50,20 +51,7 @@ def __init__(self, args, mspec):
             ["rocprofv1", "rocscope", "rocprofv2", "rocprofv3", "rocprofiler-sdk"]
         )
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx90a"))
         # Create roofline object if mode is provided; skip for --specs
         if hasattr(self.get_args(), "mode") and self.get_args().mode:
             self.roofline_obj = Roofline(args, self._mspec)

src/rocprof_compute_soc/soc_gfx940.py

@@ -28,6 +28,7 @@
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from roofline import Roofline
 from utils.logger import console_error, console_log, console_warning, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 from utils.utils import mibench
 
 
@@ -50,20 +51,7 @@ def __init__(self, args, mspec):
             ["rocprofv1", "rocprofv2", "rocprofv3", "rocprofiler-sdk"]
         )
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx940"))
         # Create roofline object if mode is provided; skip for --specs
         if hasattr(self.get_args(), "mode") and self.get_args().mode:
             self.roofline_obj = Roofline(args, self._mspec)

src/rocprof_compute_soc/soc_gfx941.py

@@ -28,6 +28,7 @@
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from roofline import Roofline
 from utils.logger import console_error, console_log, console_warning, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 from utils.utils import mibench
 
 
@@ -50,20 +51,7 @@ def __init__(self, args, mspec):
             ["rocprofv1", "rocprofv2", "rocprofv3", "rocprofiler-sdk"]
         )
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx941"))
         # Create roofline object if mode is provided; skip for --specs
         if hasattr(self.get_args(), "mode") and self.get_args().mode:
             self.roofline_obj = Roofline(args, self._mspec)

src/rocprof_compute_soc/soc_gfx942.py

@@ -28,6 +28,7 @@
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from roofline import Roofline
 from utils.logger import console_error, console_log, console_warning, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
 from utils.utils import mibench
 
 
@@ -50,20 +51,7 @@ def __init__(self, args, mspec):
             ["rocprofv1", "rocprofv2", "rocprofv3", "rocprofiler-sdk"]
         )
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx942"))
         # Create roofline object if mode is provided; skip for --specs
         if hasattr(self.get_args(), "mode") and self.get_args().mode:
             self.roofline_obj = Roofline(args, self._mspec)

src/rocprof_compute_soc/soc_gfx950.py

@@ -27,8 +27,9 @@
 import config
 from rocprof_compute_soc.soc_base import OmniSoC_Base
 from roofline import Roofline
-from utils.logger import demarcate
-from utils.utils import console_error, console_log, console_warning, mibench
+from utils.logger import console_error, console_log, console_warning, demarcate
+from utils.mi_gpu_spec import mi_gpu_specs
+from utils.utils import mibench
 
 
 class gfx950_soc(OmniSoC_Base):
@@ -48,21 +49,7 @@ def __init__(self, args, mspec):
             )
         self.set_compatible_profilers(["rocprofv3", "rocprofiler-sdk"])
         # Per IP block max number of simultaneous counters. GFX IP Blocks
-        self.set_perfmon_config(
-            {
-                "SQ": 8,
-                "TA": 2,
-                "TD": 2,
-                "TCP": 4,
-                "TCC": 4,
-                "CPC": 2,
-                "CPF": 2,
-                "SPI": 2,
-                "GRBM": 2,
-                "GDS": 4,
-                "TCC_channels": 16,
-            }
-        )
+        self.set_perfmon_config(mi_gpu_specs.get_perfmon_config("gfx950"))
         # Create roofline object if mode is provided; skip for --specs
         if hasattr(self.get_args(), "mode") and self.get_args().mode:
             self.roofline_obj = Roofline(args, self._mspec)