diff --git a/src/omniperf_soc/profile_configs/roofline/pmc_roof_perf.txt b/src/omniperf_soc/profile_configs/gfx90a/roofline/pmc_roof_perf.txt similarity index 100% rename from src/omniperf_soc/profile_configs/roofline/pmc_roof_perf.txt rename to src/omniperf_soc/profile_configs/gfx90a/roofline/pmc_roof_perf.txt diff --git a/src/omniperf_soc/profile_configs/gfx940/roofline/pmc_roof_perf.txt b/src/omniperf_soc/profile_configs/gfx940/roofline/pmc_roof_perf.txt new file mode 100644 index 0000000000..41ffdde344 --- /dev/null +++ b/src/omniperf_soc/profile_configs/gfx940/roofline/pmc_roof_perf.txt @@ -0,0 +1,9 @@ +pmc: SQ_INSTS_VALU_ADD_F16 SQ_INSTS_VALU_MUL_F16 SQ_INSTS_VALU_FMA_F16 SQ_INSTS_VALU_TRANS_F16 SQ_INSTS_VALU_ADD_F32 SQ_INSTS_VALU_MUL_F32 SQ_INSTS_VALU_FMA_F32 SQ_INSTS_VALU_TRANS_F32 + +pmc: SQ_INSTS_VALU_ADD_F64 SQ_INSTS_VALU_MUL_F64 SQ_INSTS_VALU_FMA_F64 SQ_INSTS_VALU_TRANS_F64 SQ_INSTS_VALU_MFMA_MOPS_F16 SQ_INSTS_VALU_MFMA_MOPS_BF16 SQ_INSTS_VALU_MFMA_MOPS_F32 SQ_INSTS_VALU_MFMA_MOPS_F64 TCP_TCC_READ_REQ_sum TCC_EA0_RDREQ_32B_sum TCC_EA0_RDREQ_sum TCC_EA0_WRREQ_64B_sum TCC_EA0_WRREQ_sum + +pmc: SQ_LDS_IDX_ACTIVE SQ_LDS_BANK_CONFLICT TCP_TOTAL_CACHE_ACCESSES_sum TCP_TCC_WRITE_REQ_sum TCP_TCC_ATOMIC_WITH_RET_REQ_sum TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum + +range: +gpu: +kernels: diff --git a/src/omniperf_soc/soc_gfx90a.py b/src/omniperf_soc/soc_gfx90a.py index 95b490f941..022404f74b 100644 --- a/src/omniperf_soc/soc_gfx90a.py +++ b/src/omniperf_soc/soc_gfx90a.py @@ -39,6 +39,7 @@ class gfx90a_soc(OmniSoC_Base): str(config.omniperf_home), "omniperf_soc", "profile_configs", + self.get_arch(), "roofline", ) ) diff --git a/src/omniperf_soc/soc_gfx940.py b/src/omniperf_soc/soc_gfx940.py index fd3d35f0d2..511e83710b 100644 --- a/src/omniperf_soc/soc_gfx940.py +++ b/src/omniperf_soc/soc_gfx940.py @@ -25,7 +25,7 @@ import os import config from omniperf_soc.soc_base import OmniSoC_Base -from utils.utils import demarcate, mibench, console_log +from utils.utils import demarcate, mibench, console_log, console_error from roofline import Roofline @@ -39,6 +39,7 @@ class gfx940_soc(OmniSoC_Base): str(config.omniperf_home), "omniperf_soc", "profile_configs", + "gfx940", "roofline", ) ) @@ -80,6 +81,8 @@ class gfx940_soc(OmniSoC_Base): def profiling_setup(self): """Perform any SoC-specific setup prior to profiling.""" super().profiling_setup() + if self.get_args().roof_only: + console_error("Roofline temporarily disabled in MI300") # Performance counter filtering self.perfmon_filter(self.get_args().roof_only) @@ -88,7 +91,7 @@ class gfx940_soc(OmniSoC_Base): """Perform any SoC-specific post profiling activities.""" super().post_profiling() - console_log("roofline", "Roofline temporarily disabled in Mi300") + console_log("roofline", "Roofline temporarily disabled in MI300") # if not self.get_args().no_roof: # logging.info("[roofline] Checking for roofline.csv in " + str(self.get_args().path)) # if not os.path.isfile(os.path.join(self.get_args().path, "roofline.csv")): diff --git a/src/omniperf_soc/soc_gfx941.py b/src/omniperf_soc/soc_gfx941.py index ba8f1f1702..fa26744634 100644 --- a/src/omniperf_soc/soc_gfx941.py +++ b/src/omniperf_soc/soc_gfx941.py @@ -25,7 +25,7 @@ import os import config from omniperf_soc.soc_base import OmniSoC_Base -from utils.utils import demarcate, mibench, console_log +from utils.utils import demarcate, mibench, console_log, console_error from roofline import Roofline @@ -39,6 +39,7 @@ class gfx941_soc(OmniSoC_Base): str(config.omniperf_home), "omniperf_soc", "profile_configs", + "gfx940", "roofline", ) ) @@ -80,6 +81,8 @@ class gfx941_soc(OmniSoC_Base): def profiling_setup(self): """Perform any SoC-specific setup prior to profiling.""" super().profiling_setup() + if self.get_args().roof_only: + console_error("Roofline temporarily disabled in MI300") # Performance counter filtering self.perfmon_filter(self.get_args().roof_only) @@ -88,7 +91,7 @@ class gfx941_soc(OmniSoC_Base): """Perform any SoC-specific post profiling activities.""" super().post_profiling() - console_log("roofline", "Roofline temporarily disabled in Mi300") + console_log("roofline", "Roofline temporarily disabled in MI300") # if not self.get_args().no_roof: # logging.info("[roofline] Checking for roofline.csv in " + str(self.get_args().path)) # if not os.path.isfile(os.path.join(self.get_args().path, "roofline.csv")): diff --git a/src/omniperf_soc/soc_gfx942.py b/src/omniperf_soc/soc_gfx942.py index 8fde4d8095..7b69c74193 100644 --- a/src/omniperf_soc/soc_gfx942.py +++ b/src/omniperf_soc/soc_gfx942.py @@ -25,7 +25,7 @@ import os import config from omniperf_soc.soc_base import OmniSoC_Base -from utils.utils import demarcate, mibench, console_log +from utils.utils import demarcate, mibench, console_log, console_error from roofline import Roofline @@ -39,6 +39,7 @@ class gfx942_soc(OmniSoC_Base): str(config.omniperf_home), "omniperf_soc", "profile_configs", + "gfx940", "roofline", ) ) @@ -80,6 +81,8 @@ class gfx942_soc(OmniSoC_Base): def profiling_setup(self): """Perform any SoC-specific setup prior to profiling.""" super().profiling_setup() + if self.get_args().roof_only: + console_error("Roofline temporarily disabled in MI300") # Performance counter filtering self.perfmon_filter(self.get_args().roof_only) @@ -88,7 +91,7 @@ class gfx942_soc(OmniSoC_Base): """Perform any SoC-specific post profiling activities.""" super().post_profiling() - console_log("roofline", "Roofline temporarily disabled in Mi300") + console_log("roofline", "Roofline temporarily disabled in MI300") # if not self.get_args().no_roof: # logging.info("[roofline] Checking for roofline.csv in " + str(self.get_args().path)) # if not os.path.isfile(os.path.join(self.get_args().path, "roofline.csv")):