Add broken mclk workaround for all arch (#453)

* Check for missing mclk, use hard-coded values if needed

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Fix typo

Signed-off-by: benrichard-amd <ben.richard@amd.com>

---------

Signed-off-by: benrichard-amd <ben.richard@amd.com>

[ROCm/rocprofiler-compute commit: 2de2faf944]
This commit is contained in:
Ben Richard
2024-10-25 18:13:10 -04:00
gecommit door GitHub
bovenliggende 3328cf8d6c
commit b0c7f78ca6
3 gewijzigde bestanden met toevoegingen van 17 en 6 verwijderingen
@@ -63,8 +63,10 @@ class gfx908_soc(OmniSoC_Base):
self._mspec.lds_banks_per_cu = 32
self._mspec.pipes_per_gpu = 4
# --showmclkrange is broken in Mi100, hardcode freq
self._mspec.max_mclk = 1200
self._mspec.cur_mclk = 1200
if self._mspec.gpu_model == "MI100":
if self._mspec.max_mclk == None or self._mspec.cur_mclk == None:
self._mspec.max_mclk = 1200
self._mspec.cur_mclk = 1200
@demarcate
def get_profiler_options(self):
@@ -71,6 +71,13 @@ class gfx90a_soc(OmniSoC_Base):
)
self.roofline_obj = Roofline(args, self._mspec)
# Workaround for broken --showmclkrange
# MI210/MI250/MI250X have 1600MHz mclk
if self._mspec.gpu_model == "MI200":
if self._mspec.max_mclk == None or self._mspec.cur_mclk == None:
self._mspec.max_mclk = 1600
self._mspec.cur_mclk = 1600
# Set arch specific specs
self._mspec._l2_banks = 32
self._mspec.lds_banks_per_cu = 32
@@ -69,10 +69,12 @@ class gfx942_soc(OmniSoC_Base):
)
# self.roofline_obj = Roofline(args, self._mspec)
# --showmclkrange is broken in MI308X, hardcode freq
if self._mspec.gpu_model == "MI308X":
self._mspec.max_mclk = 1300
self._mspec.cur_mclk = 1300
# Workaround for broken --showmclkrange
# MI300X/MI300A/MI308X have 1300MHz mclk
if self._mspec.gpu_model == "MI300":
if self._mspec.max_mclk == None or self._mspec.cur_mclk == None:
self._mspec.max_mclk = 1300
self._mspec.cur_mclk = 1300
# Set arch specific specs
self._mspec._l2_banks = 16