def compute_ratings(self, device_infos):
devdt = {}
min_dt = {}
for desc, device_info in sorted(device_infos.items()):
krninfo = device_info.get("matrix_multiplication")
if krninfo is None:
continue
devdt[desc] = {}
for dtype, typeinfo in krninfo.items():
bs_vo_dt = typeinfo.get("0")
if bs_vo_dt is None or len(bs_vo_dt) < 3:
continue
devdt[desc][dtype] = bs_vo_dt[2]
min_dt[dtype] = min(min_dt.get(dtype, 1.0e30), bs_vo_dt[2])
table = prettytable.PrettyTable("device", " dtype", "rating")
table.align["device"] = "l"
table.align[" dtype"] = "l"
rating = {}
for desc, dtypedt in sorted(devdt.items()):
rating[desc] = {}
for dtype, dt in sorted(dtypedt.items()):
rating[desc][dtype] = min_dt[dtype] / dt
table.add_row(desc, dtype, "%.3f" % rating[desc][dtype])
self.debug("Device ratings:\n%s", str(table))
if self.device_info.desc in rating:
self.device_info.rating = rating[self.device_info.desc]
def __init__(self):
super(OpenCLDevice, self).__init__()
self._blas = None
self._id = None
# Workaround for NVIDIA
# (fixes incorrect behaviour with OpenCL binaries)
if os.getenv("CUDA_CACHE_DISABLE") is None:
os.putenv("CUDA_CACHE_DISABLE", "1")
# Workaround for AMD
# (fixes segmentation fault when accessed over ssh with X and
# no X is running or when accessing locally and integrated
# video device is used instead of AMD one)
d = os.getenv("DISPLAY")
if d is not None and d != os.getenv("COMPUTE"):
os.unsetenv("DISPLAY")
# Set 64-bit mode for AMD OpenCL by default
if os.getenv("GPU_FORCE_64BIT_PTR") is None:
os.putenv("GPU_FORCE_64BIT_PTR", "1")
# Get the device
res = self._get_some_device()
# Restore DISPLAY to enable drawing
if d is not None:
os.putenv("DISPLAY", d)
if not res:
return
self._fill_device_info_performance_values()
log_configs = "Selected the following OpenCL configuration:\n"
table = prettytable.PrettyTable("device", " dtype", "rating",
"BLOCK_SIZE", "VECTOR_OPT", "version")
table.align["device"] = "l"
table.align[" dtype"] = "l"
table.align["BLOCK_SIZE"] = "l"
table.align["VECTOR_OPT"] = "l"
for dtype in sorted(opencl_types.dtypes.keys()):
rating = self.device_info.rating.get(dtype)
if rating is None:
rating = ""
else:
rating = "%.3f" % rating
bs_vo = self.device_info.get_kernel_bs_vo(dtype=dtype)
table.add_row(self.device_info.desc, dtype, rating, bs_vo[0],
bs_vo[1], self.device_info.version)
self.info(log_configs + str(table))
def __init__(self):
super(CUDADevice, self).__init__()
self._context_ = None
self._id = None
self._blas_ = {}
# Get the device
self._get_some_device()
log_configs = "Selected the following CUDA device:\n"
table = prettytable.PrettyTable("device", "mem", "compute", "pci")
table.align["device"] = "l"
table.align["mem"] = "r"
table.align["pci"] = "l"
table.add_row(self.context.device.name,
self.context.device.total_mem // 1048576,
"%d.%d" % self.context.device.compute_capability,
self.context.device.pci_bus_id)
self.info(log_configs + str(table))
def _dump_unit_attributes(self, arrays=True):
import veles.external.prettytable as prettytable
from veles.workflow import Workflow
self.debug("Dumping unit attributes of %s...", str(self.workflow))
table = prettytable.PrettyTable("#", "unit", "attr", "value")
table.align["#"] = "r"
table.align["unit"] = "l"
table.align["attr"] = "l"
table.align["value"] = "l"
table.max_width["value"] = 100
for i, u in enumerate(self.workflow.units_in_dependency_order):
for k, v in sorted(u.__dict__.items()):
if k not in Workflow.HIDDEN_UNIT_ATTRS:
if (not arrays and hasattr(v, "__len__") and len(v) > 32
and not isinstance(v, str)
and not isinstance(v, bytes)):
strv = "object of class %s of length %d" % (repr(
v.__class__.__name__), len(v))
else:
strv = repr(v)
table.add_row(i, u.__class__.__name__, k, strv)
print(table)