def test_throughput_hw(self, topology, wbits, abits, kind):
prev_chkpt_name = get_checkpoint_name(topology, wbits, abits, "deploy_" + kind)
end2end_example = "%s_w%da%d_%s" % (topology, wbits, abits, kind)
model = load_test_checkpoint_or_skip(prev_chkpt_name) # NOQA
cfg = get_build_env(kind, target_clk_ns)
if cfg["ip"] == "":
pytest.skip("PYNQ board IP address not specified")
ret = dict()
# try a range of batch sizes, some may fail due to insufficient DMA
# buffers
bsize_range_in = [8 ** i for i in range(5)]
bsize_range = []
for bsize in bsize_range_in:
res = throughput_test_remote(model, bsize)
if res is not None:
ret[bsize] = res
bsize_range.append(bsize)
else:
# assume we reached largest possible N
break
y = [ret[key]["runtime[ms]"] for key in bsize_range]
lrret = linregress(bsize_range, y)
ret_str = ""
ret_str += "\n" + "%s Throughput Test Results" % end2end_example
ret_str += "\n" + "-----------------------------"
ret_str += "\n" + "From linear regression:"
ret_str += "\n" + "Invocation overhead: %f ms" % lrret.intercept
ret_str += "\n" + "Time per sample: %f ms" % lrret.slope
ret_str += "\n" + "Raw data:"
ret_str += "\n" + "{:<8} {:<16} {:<16} {:<16} {:<16} {:<16}".format(
"N", "runtime[ms]", "fclk[mhz]", "fps", "DRAM rd[Mb/s]", "DRAM wr[Mb/s]"
)
for k in bsize_range:
v = ret[k]
ret_str += "\n" + "{:<8} {:<16} {:<16} {:<16} {:<16} {:<16}".format(
k,
np.round(v["runtime[ms]"], 4),
v["fclk[mhz]"],
np.round(v["throughput[images/s]"], 2),
np.round(v["DRAM_in_bandwidth[Mb/s]"], 2),
np.round(v["DRAM_out_bandwidth[Mb/s]"], 2),
)
ret_str += "\n" + "-----------------------------"
warnings.warn(ret_str)
largest_bsize = bsize_range[-1]
update_dashboard_data(
topology, wbits, abits, "fclk[mhz]", ret[largest_bsize]["fclk[mhz]"]
)
update_dashboard_data(
topology,
wbits,
abits,
"throughput[images/s]",
ret[largest_bsize]["throughput[images/s]"],
)
import shutil
import os
# image = Image.open('/workspace/finn/onnx_experiments/img_MNIST_grayscale.png')
# x = TF.to_tensor(image)
# x.unsqueeze_(0)
#
# parent_model = load("5_dataflow_parent")
# sdp_node = parent_model.graph.node[2]
# remote_exec_model = build_dir + "8_deploy.onnx"
# getCustomOp(sdp_node).set_nodeattr("model", remote_exec_model)
# save(parent_model,"9_dataflow_parent_with_remote_bitfile_exec")
# THROUGHPUT TESTS
from finn.core.throughput_test import throughput_test_remote
res = throughput_test_remote(model, batchsize=10000)
print("Network metrics:")
with open(build_dir + "res.txt", "w") as f:
for key in res:
print(str(key) + ": " + str(res[key]))
f.write(str(key) + ": " + str(res[key]) + "\n")
II = 64
# frequency in MHz
f_MHz = 100
# expected throughput in MFPS
expected_throughput = f_MHz / II
# measured throughput (FPS) from throughput test, converted to MFPS
measured_throughput = res["throughput[images/s]"] * 0.000001
# peformance
print("We reach approximately " +