opt.use_int8_inference = True
opt.use_vulkan_compute = use_vulkan_compute
opt.use_fp16_packed = True
opt.use_fp16_storage = True
opt.use_fp16_arithmetic = True
opt.use_int8_storage = True
opt.use_int8_arithmetic = True
opt.use_packing_layout = True
ncnn.set_cpu_powersave(powersave)
ncnn.set_omp_dynamic(0)
ncnn.set_omp_num_threads(num_threads)
print("loop_count = %d" % (loop_count))
print("num_threads = %d" % (num_threads))
print("powersave = %d" % (ncnn.get_cpu_powersave()))
print("gpu_device = %d" % (gpu_device))
#must use named param w, h, c due to python has no size_t(unsigned int) to call the correct overload ncnn.Mat
benchmark("squeezenet", ncnn.Mat(w=227, h=227, c=3), opt)
benchmark("squeezenet_int8", ncnn.Mat(w=227, h=227, c=3), opt)
benchmark("mobilenet", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("mobilenet_int8", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("mobilenet_v2", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("mobilenet_v3", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("shufflenet", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("shufflenet_v2", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("mnasnet", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("proxylessnasnet", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("googlenet", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("googlenet_int8", ncnn.Mat(w=224, h=224, c=3), opt)
opt.use_fp16_packed = True
opt.use_fp16_storage = True
opt.use_fp16_arithmetic = True
opt.use_int8_storage = True
opt.use_int8_arithmetic = True
opt.use_packing_layout = True
opt.use_shader_pack8 = False
opt.use_image_storage = False
ncnn.set_cpu_powersave(powersave)
ncnn.set_omp_dynamic(0)
ncnn.set_omp_num_threads(num_threads)
print("loop_count =", loop_count)
print("num_threads =", num_threads)
print("powersave =", ncnn.get_cpu_powersave())
print("gpu_device =", gpu_device)
print("cooling_down =", g_enable_cooling_down)
benchmark("squeezenet", ncnn.Mat((227, 227, 3)), opt)
benchmark("squeezenet_int8", ncnn.Mat((227, 227, 3)), opt)
benchmark("mobilenet", ncnn.Mat((224, 224, 3)), opt)
benchmark("mobilenet_int8", ncnn.Mat((224, 224, 3)), opt)
benchmark("mobilenet_v2", ncnn.Mat((224, 224, 3)), opt)
# benchmark("mobilenet_v2_int8", ncnn.Mat(w=224, h=224, c=3), opt)
benchmark("mobilenet_v3", ncnn.Mat((224, 224, 3)), opt)
benchmark("shufflenet", ncnn.Mat((224, 224, 3)), opt)
benchmark("shufflenet_v2", ncnn.Mat((224, 224, 3)), opt)
benchmark("mnasnet", ncnn.Mat((224, 224, 3)), opt)
benchmark("proxylessnasnet", ncnn.Mat((224, 224, 3)), opt)
benchmark("efficientnet_b0", ncnn.Mat((224, 224, 3)), opt)