def test_occupancy_32_9_21_threads(self):
mat_a = DenseMatrix(num_rows=32,
num_cols=56,
addressing="strided",
bbox=[0, 0, 31, 20],
transpose=False)
mat_b = DenseMatrix(num_rows=32,
num_cols=56,
addressing="strided",
bbox=[0, 0, 20, 8],
transpose=False)
mat_c = DenseMatrix(num_rows=32,
num_cols=9,
bbox=[0, 0, 31, 8],
addressing="strided",
transpose=False)
self.gen.generate(mat_a, mat_b, mat_c, alpha=1.1, beta=1.1)
# one multiplication per block because the kernel requires too much of
# shared memory
self.assertEqual(self.gen.num_active_threads, 32)
self.assertEqual(self.gen.num_mult_per_block, 2)
def test_exact_transpose_loader(self):
# load a column in one go
loader = shm_mem_factory(vm=self._vm,
matrix=DenseMatrix(num_rows=33,
num_cols=56,
addressing='none',
bbox=[0, 0, 15, 20],
transpose=True),
num_active_threads=32,
load_and_transpose=True)
self.assertIsInstance(loader, ExactTransposePatchLoader)
# multiple hops to load a column
loader = shm_mem_factory(vm=self._vm,
matrix=DenseMatrix(num_rows=65,
num_cols=56,
addressing='none',
bbox=[0, 0, 34, 20],
transpose=True),
num_active_threads=32,
load_and_transpose=True)
self.assertIsInstance(loader, ExactTransposePatchLoader)
def test_occupancy_56_9_9_threads(self):
mat_a = DenseMatrix(num_rows=56,
num_cols=9,
addressing="strided",
bbox=[0, 0, 55, 8],
transpose=False)
mat_b = DenseMatrix(num_rows=9,
num_cols=9,
addressing="strided",
bbox=[0, 0, 8, 8],
transpose=False)
mat_c = DenseMatrix(num_rows=56,
num_cols=9,
bbox=[0, 0, 55, 8],
addressing="strided",
transpose=False)
self.gen.generate(mat_a, mat_b, mat_c, alpha=1.1, beta=1.1)
self.assertEqual(self.gen.num_active_threads, 64)
self.assertEqual(self.gen.num_mult_per_block, 1)
def produce_matrix(spec):
return DenseMatrix(num_rows=spec["num_rows"],
num_cols=spec["num_cols"],
addressing=spec["addressing"],
bbox=spec["bbox"],
transpose=spec["trans"])
def _produce_matrix(self, matrix_spec):
return DenseMatrix(num_rows=matrix_spec["rows"],
num_cols=matrix_spec["cols"],
addressing=matrix_spec["addressing"],
bbox=matrix_spec["bbox"],
transpose=matrix_spec["trans"])
"--manufacturer",
action="store",
help="Name of the Manufacturer, currently nvidia and amd are supported",
default="nvidia")
parser.add_argument(
"-s",
"--sub_arch",
action="store",
help="Sub_arch of the GPU, e.g sm_60 for Nvidia or gfx906 for AMD",
default="sm_60")
args = parser.parse_args()
mat_a = DenseMatrix(num_rows=9,
num_cols=9,
addressing="strided",
bbox=[0, 0, 8, 8],
transpose=False)
mat_b = DenseMatrix(num_rows=9,
num_cols=9,
addressing="strided",
bbox=[0, 0, 8, 8],
transpose=False)
try:
vm = vm_factory(name=args.manufacturer,
sub_name=args.sub_arch,
fp_type="float")
gen = CsaGenerator(vm)