def setup_env():
num_concurrent_streams = Config.get("compiler", "cuda",
"max_concurrent_streams")
if 'ORT_USE_STREAMS' in os.environ:
ONNXRuntimeCUDA.use_streams = _env2bool(os.environ["ORT_USE_STREAMS"])
if ONNXRuntimeCUDA.use_streams:
log.info("Using streams with ORT (experimental)")
if num_concurrent_streams == 0:
log.info("Setting compiler.cuda.max_concurrent_streams to 8")
Config.set("compiler",
"cuda",
"max_concurrent_streams",
value=8)
elif num_concurrent_streams == -1:
ONNXRuntimeCUDA.use_streams = False
else:
if num_concurrent_streams != -1:
log.info("Setting compiler.cuda.max_concurrent_streams to -1")
Config.set("compiler", "cuda", "max_concurrent_streams", value=-1)
ONNXRuntimeCUDA.use_streams = False
ONNXRuntimeCUDA.max_concurrent_streams = Config.get(
"compiler", "cuda", "max_concurrent_streams")
sdfg.add_edge(state2, copy_out_state,
dace.sdfg.sdfg.InterstateEdge())
sdfg.validate()
return sdfg
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("N", type=int, nargs="?", default=32)
args = vars(parser.parse_args())
size_n = args["N"]
from dace.config import Config
# set unique function to false to generate both sdfgs
Config.set("compiler", "unique_functions", value=False)
sdfg = make_nested_sdfg_fpga()
X = np.random.rand(size_n).astype(np.float32)
Y = np.random.rand(size_n).astype(np.float32)
sdfg(X=X, Y=Y, N=size_n)
ref = X+2
diff = np.linalg.norm(ref - Y) / size_n
if diff <= 1e-5 :
print("==== Program end ====")
else:
raise Exception("==== Program Error! ====")
@dace.program
def matmul_np(A: dace.float64[128, 64], B: dace.float64[64, 32],
C: dace.float64[128, 32]):
C[:] = A @ B
A = np.random.rand(128, 64).astype(np.float64)
B = np.random.rand(64, 32).astype(np.float64)
C = np.random.rand(128, 32).astype(np.float64)
sdfg = matmul_np.to_sdfg()
sdfg.apply_transformations([FPGATransformSDFG])
from dace.libraries.blas import Gemm
Gemm.default_implementation = "FPGA1DSystolic"
# We have to Inline
sdfg.expand_library_nodes()
sdfg.apply_transformations_repeated([InlineSDFG])
C_regression = A @ B
p = Process(target=evaluate, args=(sdfg, A, B, C, C_regression))
p.start()
p.join()
del sdfg
if __name__ == "__main__":
# These tests will be executed on seperate process. The tranfform_on_call DACE configuration must be set to false
Config.set('optimizer', 'transform_on_call', value=False)
test_gemm_vectorized()
test_gemm_size_not_multiples_of()
test_matmul_np()
