Carry over fixed numSQCs calc from #247

Co-authored-by: Nick Curtis <nicholas.curtis@amd.com>
Signed-off-by: colramos-amd <colramos@amd.com>

src/omniperf_soc/soc_base.py

@@ -90,7 +90,7 @@ class OmniSoC_Base():
 
     @demarcate
     def populate_mspec(self):
-        from utils.specs import search, run
+        from utils.specs import search, run, total_sqc
 
         if not hasattr(self._mspec, "_rocminfo"):
             return
@@ -144,6 +144,13 @@ class OmniSoC_Base():
                 self._mspec.max_waves_per_cu = key
                 break
         
+        self._mspec.sqc_per_gpu = str(
+            total_sqc(
+                self._mspec.gpu_arch, 
+                self._mspec.cu_per_gpu,
+                self._mspec.se_per_gpu
+            ))
+        
         # we get the max mclk from rocm-smi --showmclkrange
         rocm_smi_mclk = run(["rocm-smi", "--showmclkrange"], exit_on_error=True)
         self._mspec.max_mclk = search(r'(\d+)Mhz\s*$', rocm_smi_mclk)

src/omniperf_soc/soc_gfx906.py

@@ -53,7 +53,6 @@ class gfx906_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 16
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 15
         self._mspec.pipes_per_gpu = 4
 
     # -----------------------

src/omniperf_soc/soc_gfx908.py

@@ -53,7 +53,6 @@ class gfx908_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 32
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 30
         self._mspec.pipes_per_gpu = 4
         # --showmclkrange is broken in Mi100, hardcode freq
         self._mspec.max_mclk = 1200

src/omniperf_soc/soc_gfx90a.py

@@ -59,7 +59,6 @@ class gfx90a_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 32
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 56
         self._mspec.pipes_per_gpu = 4
 
     # -----------------------

src/omniperf_soc/soc_gfx940.py

@@ -64,7 +64,6 @@ class gfx940_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 16
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 56
         self._mspec.pipes_per_gpu = 4
 
     # -----------------------

src/omniperf_soc/soc_gfx941.py

@@ -64,7 +64,6 @@ class gfx941_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 16
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 56
         self._mspec.pipes_per_gpu = 4
 
     # -----------------------

src/omniperf_soc/soc_gfx942.py

@@ -64,7 +64,6 @@ class gfx942_soc (OmniSoC_Base):
         # Set arch specific specs
         self._mspec.L2Banks = 16
         self._mspec.lds_banks_per_cu = 32
-        self._mspec.sqc_per_gpu = 56
         self._mspec.pipes_per_gpu = 4
 
     # -----------------------

src/utils/specs.py

@@ -34,6 +34,7 @@ import logging
 import pandas as pd
 
 from datetime import datetime
+from math import ceil
 from dataclasses import dataclass
 from pathlib import Path as path
 from textwrap import dedent
@@ -290,6 +291,14 @@ def total_l2_banks(archname, L2Banks, memory_partition):
             archname, memory_partition)
     return str(totalL2Banks)
 
+def total_sqc(archname, numCUs, numSEs):
+    cu_per_se = float(numCUs) / float(numSEs)
+    sq_per_se = cu_per_se / 2
+    if archname.lower() in ['mi50', 'mi100']:
+        sq_per_se = cu_per_se / 3
+    sq_per_se = ceil(sq_per_se)
+    return int(sq_per_se) * int(numSEs)
+
 
 if __name__ == "__main__":
     print(MachineSpecs())