Fix Mi300 clk timings using GRBM_GUI_ACTIVE from #255

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


[ROCm/rocprofiler-compute commit: bd794abb12]

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

@@ -81,6 +81,9 @@ build_in_vars = {
               0) / $max_waves_per_cu) * 8) + MIN(MOD(ROUND(AVG(((4 * SQ_BUSY_CU_CYCLES) \
               / GRBM_GUI_ACTIVE)), 0), $max_waves_per_cu), 8)), $cu_per_gpu))",
     "kernelBusyCycles": "ROUND(AVG((((End_Timestamp - Start_Timestamp) / 1000) * $max_sclk)), 0)",
+    "GRBM_GUI_ACTIVE_PER_XCD": "(GRBM_GUI_ACTIVE / $num_xcd)",
+    "GRBM_COUNT_PER_XCD": "(GRBM_COUNT / $num_xcd)",
+    "GRBM_SPI_BUSY_PER_XCD" : "(GRBM_SPI_BUSY / $num_xcd)"
 }
 
 supported_call = {
@@ -691,6 +694,7 @@ def eval_metric(dfs, dfs_type, sys_info, raw_pmc_df, debug):
     ammolite__max_waves_per_cu = sys_info.max_waves_per_cu
     ammolite__hbm_bw = sys_info.hbm_bw
     ammolite__total_l2_chan = calc_builtin_var("$total_l2_chan", sys_info)
+    ammolite__num_xcd = sys_info.num_xcd
 
     # TODO: fix all $normUnit in Unit column or title
 
@@ -709,6 +713,9 @@ def eval_metric(dfs, dfs_type, sys_info, raw_pmc_df, debug):
 
     ammolite__numActiveCUs = ammolite__build_in["numActiveCUs"]
     ammolite__kernelBusyCycles = ammolite__build_in["kernelBusyCycles"]
+    ammolite__GRBM_GUI_ACTIVE_PER_XCD = ammolite__build_in["GRBM_GUI_ACTIVE_PER_XCD"]
+    ammolite__GRBM_COUNT_PER_XCD = ammolite__build_in["GRBM_COUNT_PER_XCD"]
+    ammolite__GRBM_SPI_BUSY_PER_XCD = ammolite__build_in["GRBM_SPI_BUSY_PER_XCD"]
 
     # Hmmm... apply + lambda should just work
     # df['Value'] = df['Value'].apply(lambda s: eval(compile(str(s), '<string>', 'eval')))
@@ -1003,6 +1010,7 @@ def correct_sys_info(df, specs_correction):
         "hbmBW": "hbmBW",
         "compute_partition": "compute_partition",
         "memory_partition": "memory_partition",
+        "num_xcd": "num_xcd"
     }
 
     # todo: more err checking for string specs_correction

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

@@ -132,6 +132,7 @@ class MachineSpecs:
         self.pipes_per_gpu: str = None
         self.total_l2_chan: str = None
         self.hbm_bw: str = None
+        self.num_xcd: str = None
         # Load above SoC specs via module import
         try:
             soc_module = importlib.import_module('omniperf_soc.soc_'+ self.gpu_arch)
@@ -226,6 +227,7 @@ class MachineSpecs:
             HBM BW:                     {self.hbm_bw} MB/s
             Compute Partition:          {self.compute_partition}
             Memory Partition:           {self.memory_partition}
+            Num XCDs:                   {self.num_xcd}
         """
         )
 
@@ -299,6 +301,37 @@ def total_sqc(archname, numCUs, numSEs):
     sq_per_se = ceil(sq_per_se)
     return int(sq_per_se) * int(numSEs)
 
+def total_xcds(archname, compute_partition):
+    # check MI300 has a valid compute partition
+    mi300a_archs = ["mi300a_a0", "mi300a_a1"]
+    mi300x_archs = ["mi300x_a0", "mi300x_a1"]
+    if archname.lower() in mi300a_archs + mi300x_archs \
+        and compute_partition == "NA":
+            error("Invalid compute partition found for {}".format(archname))
+    if archname.lower() not in mi300a_archs + mi300x_archs:
+        return 1
+    # from the whitepaper
+    # https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/white-papers/amd-cdna-3-white-paper.pdf
+    if compute_partition.lower() == "spx":
+        if archname.lower() in mi300a_archs:
+            return 6
+        if archname.lower() in mi300x_archs:
+            return 8
+    if compute_partition.lower() == "tpx":
+        if archname.lower() in mi300a_archs:
+            return 2
+    if compute_partition.lower() == "dpx":
+        if archname.lower() in mi300x_archs:
+            return 4
+    if compute_partition.lower() == "qpx":
+        if archname.lower() in mi300x_archs:
+            return 2
+    if compute_partition.lower() == "cpx":
+        if archname.lower() in mi300x_archs:
+            return 2
+    error("Unknown compute partition / arch found for {} / {}".format(
+        compute_partition, archname))
+
 
 if __name__ == "__main__":
     print(MachineSpecs())