def profile_buf_size(low, high, step, scale):
arr = []
# systolic_arr_size, memory_bandwidth, buffer_size
for size in [0.25, 0.5, 1, 1.5, 2.0, 3.0, 6.0]:
config = hardware_constraints()
# print(config[2]*pow(1.2, size)/3.0)
config[2] = config[2] * size
# config[2] = config[2]*scale*pow(1.2, size)
setup_hardware(config)
setup_hardware3d(config)
res = opti_dnn()
arr.append([item[0:4] for item in res])
# gather results
res = {"sa_avg":[], "sa_std":[], "buf_avg":[], "buf_std":[], \
"cycle_avg": [], "traffic_avg": []}
for ls in arr:
res["sa_avg"].append(np.mean([i[2] for i in ls]))
res["sa_std"].append(np.std([i[2] for i in ls]))
res["buf_avg"].append(np.mean([i[3] for i in ls]))
res["buf_std"].append(np.std([i[3] for i in ls]))
res["cycle_avg"].append(np.mean([i[1] for i in ls]))
res["traffic_avg"].append(np.mean([i[0] for i in ls]))
print(res)
# plot_buf_size(res, [0.25, 0.5, 1, 1.5, 2.0, 3.0, 6.0], 1)
plot_buf_cycle(res, [0.25, 0.5, 1, 1.5, 2.0, 3.0, 6.0], 1)
def profile_bw_size(low, high):
arr = []
# systolic_arr_size, memory_bandwidth, buffer_size
for i in range(low, high):
config = hardware_constraints()
config[1] = float(pow(2, i))
setup_hardware(config)
setup_hardware3d(config)
res = opti_dnn()
arr.append([item[0:4] for item in res])
# gather results
res = {"sa_avg":[], "sa_std":[], "buf_avg":[], "buf_std":[], \
"cycle_avg": [], "traffic_avg": []}
for ls in arr:
res["sa_avg"].append(np.mean([i[2] for i in ls]))
res["sa_std"].append(np.std([i[2] for i in ls]))
res["buf_avg"].append(np.mean([i[3] for i in ls]))
res["buf_std"].append(np.std([i[3] for i in ls]))
res["cycle_avg"].append(np.mean([i[1] for i in ls]))
res["traffic_avg"].append(np.mean([i[0] for i in ls]))
plot_bw_size(res, low, high)
plot_bw_cycle(res, low, high)
def profile_sa_size(low, high, step):
for scale in range(1, 7):
arr = []
# systolic_arr_size, memory_bandwidth, buffer_size
for size in range(low, high + step, step):
config = hardware_constraints(sa_size=16.0,
mem_bw=32.0,
buf=1048576.0 * 0.5 * scale)
print("SIZE", size)
config[0] = float(size)
print(config)
setup_hardware(config)
setup_hardware3d(config)
res = opti_dnn()
arr.append([item[0:4] for item in res])
# gather results
res = {"sa_avg":[], "sa_std":[], "buf_avg":[], "buf_std":[], \
"cycle_avg": [], "traffic_avg": []}
for ls in arr:
res["sa_avg"].append(np.mean([i[2] for i in ls]))
res["sa_std"].append(np.std([i[2] for i in ls]))
res["buf_avg"].append(np.mean([i[3] for i in ls]))
res["buf_std"].append(np.std([i[3] for i in ls]))
res["cycle_avg"].append(np.mean([i[1] for i in ls]))
res["traffic_avg"].append(np.mean([i[0] for i in ls]))
# print >> sys.stderr, ('[%d]' % (scale))
print >> sys.stderr, str(res["traffic_avg"])
def profile_sa_size(low, high, step):
arr = []
# systolic_arr_size, memory_bandwidth, buffer_size
for size in range(low, high + step, step):
config = hardware_constraints()
print("SIZE", size)
config[0] = float(size)
print(config)
setup_hardware(config)
setup_hardware3d(config)
res = opti_dnn()
for j in range(len(res)):
if res[j] is None:
res[j] = arr[-1][j]
arr.append([item[0:4] for item in res])
# gather results
res = {"sa_avg":[], "sa_std":[], "buf_avg":[], "buf_std":[], \
"cycle_avg": [], "traffic_avg": []}
for ls in arr:
res["sa_avg"].append(np.mean([i[2] for i in ls]))
res["sa_std"].append(np.std([i[2] for i in ls]))
res["buf_avg"].append(np.mean([i[3] for i in ls]))
res["buf_std"].append(np.std([i[3] for i in ls]))
res["cycle_avg"].append(np.mean([i[1] for i in ls]))
res["traffic_avg"].append(np.mean([i[0] for i in ls]))
print(res)
plot_sa_size(res, low, high, step)
plot_sa_cycle(res, low, high, step)
# plot_buf_size(res, [0.25, 0.5, 1, 1.5, 2.0, 3.0, 6.0], 1)
plot_buf_cycle(res, [0.25, 0.5, 1, 1.5, 2.0, 3.0, 6.0], 1)
if __name__ == '__main__':
# import the dnn
import_dnn("dnns/GC_Net.txt")
# import_dnn("dnns/PSMNet.txt")
for ln in dnn3d:
print(ln)
# exit()
# check which characterization you want to proceed
if detail_profile:
# set up the hardware configuration
setup_hardware(hardware_constraints())
# set up the hardware configuration
setup_hardware3d(hardware_constraints())
# start the optimization main routine
res = opti_dnn()
# plot the result of each layer
plot_util_dnn(res, suffix)
profile_layer_cycle(res, suffix)
else:
# profile systolic array size impacts
profile_sa_size(8, 64, 8)
# profile bandwidth impacts
# profile_bw_size(-5, 12)
# profile buffer size impacts
# profile_buf_size(10, 200, 10, 0.01)
# profile_buf_size(3, 20, 2, 0.1)