def main(argv):
global threadnum
listimage = os.listdir(calib_image_dir)
threadAll = []
threadnum = int(argv[1])
i = 0
global runTotall
runTotall = len(listimage)
g = xir.graph.Graph.deserialize(pathlib.Path(argv[2]))
subgraphs = get_subgraph(g)
assert len(subgraphs) == 1 # only one DPU kernel
all_dpu_runners = []
for i in range(int(threadnum)):
all_dpu_runners.append(runner.Runner(subgraphs[0], "run"))
"""image list to be run """
xclbin_p = str("/mnt/dpu.xclbin")
kernelName_p = "pp_pipeline_accel"
deviceIdx_p = 0
fpga_pp = waa_rt.PreProcess(xclbin_p, kernelName_p, deviceIdx_p)
time1 = int(round(time.time() * 1000))
img = []
for i in range(runTotall):
path = os.path.join(calib_image_dir, listimage[i])
image = cv2.imread(path)
rows, cols, channels = image.shape
image = fpga_pp.preprocess_input(image, rows, cols)
img.append(image)
time_pre = int(round(time.time() * 1000))
start = 0
for i in range(int(threadnum)):
if (i == threadnum - 1):
end = len(img)
else:
end = start + (len(img) // threadnum)
t1 = threading.Thread(target=runResnet50,
args=(all_dpu_runners[i], img[start:end],
len(img[start:end])))
threadAll.append(t1)
start = end
for x in threadAll:
x.start()
for x in threadAll:
x.join()
time2 = int(round(time.time() * 1000))
timetotal = time2 - time1
fps = float(runTotall * 1000 / timetotal)
#print("Pre time: %d ms" %(time_pre - time1))
#print("DPU + post time: %d ms" %(time2 - time_pre))
#print("Total time : %d ms" %timetotal)
#print("Total frames : %d" %len(img))
print("Performance : %.2f FPS" % fps)
def pre_process(q, args):
xclbin_p = str(args['xclbin'] + "/xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin")
kernelName_p = "pp_pipeline_accel"
deviceIdx_p = args['deviceid']
fpga_pp = waa_rt.PreProcess(xclbin_p, kernelName_p, deviceIdx_p, 1)
batch_sz = args['batch_sz']
img_paths = xdnn_io.getFilePaths(args['images'])
for i in range(0, len(img_paths), batch_sz):
for j, p in enumerate(img_paths[i:i + batch_sz]):
arr, ht = fpga_pp.preprocess_input(p)
q.put(arr)
def main(argv):
global threadnum
listimage=os.listdir(calib_image_dir)
threadAll = []
threadnum = int(argv[1])
i = 0
global runTotall
runTotall = len(listimage)
g = xir.Graph.deserialize(argv[2])
subgraphs = get_child_subgraph_dpu(g)
assert len(subgraphs) == 1 # only one DPU kernel
all_dpu_runners = []
for i in range(int(threadnum)):
all_dpu_runners.append(vart.Runner.create_runner(subgraphs[0], "run"))
"""image list to be run """
xclbin_p=str("/usr/lib/dpu.xclbin")
kernelName_p="pp_pipeline_accel"
deviceIdx_p=0
fpga_pp = waa_rt.PreProcess(xclbin_p,kernelName_p,deviceIdx_p)
time1 = int(round(time.time() * 1000))
img = []
time_start = time.time()
for i in range(runTotall):
path = os.path.join(calib_image_dir,listimage[i])
img.append(fpga_pp.preprocess_input(path))
cnt = 1
"""run with batch """
for i in range(int(threadnum)):
t1 = threading.Thread(target=runResnet50, args=(all_dpu_runners[i], img, cnt))
threadAll.append(t1)
for x in threadAll:
x.start()
for x in threadAll:
x.join()
del all_dpu_runners
#print("Pre time: %d ms" %(time_pre - time1))
time_end = time.time()
timetotal = time_end - time_start
total_frames = runTotall
fps = float(total_frames / timetotal)
print(
"FPS=%.2f, total frames = %.2f , time=%.6f seconds"
% (fps, total_frames, timetotal)
)
def pre_process(q_img, q_shape,args):
xclbin_p=str(args['xclbin']+"/xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin")
kernelName_p="pp_pipeline_accel"
deviceIdx_p=args['deviceid']
handle_p = waa_rt.PreProcess(xclbin_p,kernelName_p,deviceIdx_p, 1)
if handle_p == -1:
print("Unable to Create handle for pre-processing kernel. Only U200 device is supported")
sys.exit()
batch_sz = args['batch_sz']
img_paths = xdnn_io.getFilePaths(args['images'])
print("Pre-processing handle created. Populating Queue")
for i in range(0, len(img_paths), batch_sz):
for j, p in enumerate(img_paths[i:i + batch_sz]):
arr, shape = handle_p.preprocess_input(p)
q_img.put(arr)
q_shape.put(shape)