def run(batch_size,
height,
width,
channel,
kernel_size,
output_channel,
stride,
padding,
model_path,
epoch=5,
sample_size=16,
number=100,
test=False):
entities = []
func = FUNC_TABLE["conv2d_channel_batch"].func
args = (batch_size, height, width, channel, kernel_size, kernel_size,
output_channel, stride, padding)
entities.append(Entity("conv2d_channel_batch", args))
model_path = os.path.abspath(model_path)
if not test:
beg = time.time()
train_op_schedule_cpu_general_dx(entities, epoch, sample_size,
model_path)
end = time.time()
print("{}({}):".format("conv2d_channel_batch", args))
print("train done! use {}ms".format((end - beg) * 1e3))
test_graph_schedule_cpu_general_dx(func, args, model_path, number=number)
def transfer(pre_train, post_train):
entities = []
for i in pre_train:
args = arg_lst[i]
entities.append(Entity(func_names[i], args))
model_path = "gpu_transfer_pre{}_post{}.pkl".format(pre_train, post_train)
model_path = os.path.abspath(model_path)
train_op_schedule_gpu_general_dx(entities, 20, 50, model_path,
logfile="process_transfer_pre_{}_(post{})_gpu.txt".format(pre_train, post_train),
device="cuda:0")
entities = []
for i in post_train:
args = arg_lst[i]
entities.append(Entity(func_names[i], args))
model_path = "gpu_transfer0_pre{}_post{}.pkl".format(pre_train, post_train)
model_path = os.path.abspath(model_path)
train_op_schedule_gpu_general_dx(entities, 20, 50, model_path,
logfile="process_transfer_(pre_{})_post{}_gpu.txt".format(pre_train, post_train),
device="cuda:0")
def run(M, N, k, model_path, epoch=5, sample_size=16, number=100, test=False):
entities = []
func = FUNC_TABLE["gaussian_blur2d"].func
args = (M, N, k)
entities.append(Entity("gaussian_blur2d", args))
model_path = os.path.abspath(model_path)
if not test:
beg = time.time()
train_op_schedule_cpu_general_dx(entities, epoch, sample_size,
model_path)
end = time.time()
print("{}({}):".format("gaussian_blur2d", args))
print("train done! use {}ms".format((end - beg) * 1e3))
test_graph_schedule_cpu_general_dx(func, args, model_path, number=number)
def run(batch_size,
M,
N,
L,
model_path,
epoch=5,
sample=16,
number=100,
test=False):
entities = []
args = (batch_size, M, N, L)
entities.append(Entity("matmul_batch", args))
model_path = os.path.abspath(model_path)
if not test:
beg = time.time()
train_op_schedule_gpu_general_dx(entities, epoch, sample, model_path)
end = time.time()
print("{}({}):".format("matmul_batch", args))
print("train done! use {}ms".format((end - beg) * 1e3))
test_graph_schedule_gpu_general_dx(entities,
model_path,
sampling=True,
number=number)
(1, 1024, 1024, 1024),
(1, 2048, 2048, 2048),
(1, 4096, 4096, 4096),
(1, 1024, 32, 1024),
(1, 32, 1024, 32),
]
names = [
"128_128_128_b1",
"256_256_256_b1",
"512_512_512_b1",
"1024_1024_1024_b1",
"2048_2048_2048_b1",
"4096_4096_4096_b1",
"1024_32_1024_b1",
"32_1024_32_b1",
]
for i in range(len(arg_lst)):
model_path = "opt_gemm_" + names[i] + "_cpu.pkl"
entities = []
args = arg_lst[i]
entities.append(Entity("matmul_batch", args))
model_path = os.path.abspath(model_path)
train_op_schedule_cpu_general_dx(entities,
20,
50,
model_path,
logfile="process_gemm_" + names[i] +
"_cpu.txt",
device="cuda:0")
# params = [int(x) for x in args.params.split(",")]
# batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding = params
# if use_torch:
# pytorch_baseliine(batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding, args.number)
# if use_auto:
# run(batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding,
# args.model_file_path, args.epoch, args.sample, args.number, test)
# except Exception as e:
# raise ValueError("Bad parameters, please refer to usage")
arg_lst = [
(1, 14, 14, 14, 1024, 1, 1, 1, 512, 1, 0),
(1, 28, 28, 28, 512, 1, 1, 1, 256, 1, 0),
(1, 56, 56, 56, 192, 1, 1, 1, 128, 1, 0),
]
names = [
"yolo19_b1",
"yolo7_b1",
"yolo3_b1",
]
for i in range(len(arg_lst)):
batch_size, height, width, depth, channel, kernel_size, _, _, output_channel, stride, padding = arg_lst[i]
model_path = "opt_conv3d_" + names[i] + "_cpu.pkl"
entities = []
func = FUNC_TABLE["conv3d_channel_batch"].func
args = arg_lst[i]
entities.append(Entity("conv3d_channel_batch", args))
model_path = os.path.abspath(model_path)
train_op_schedule_cpu_general_dx(entities, 20, 50, model_path, logfile="process_conv3d_" + names[i] + "_cpu.txt", device="cuda:2")
# batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding = params
# if use_torch:
# pytorch_baseliine(batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding, args.number)
# if use_auto:
# run(batch_size, height, width, depth, channel, kernel_size, output_channel, stride, padding,
# args.model_file_path, args.epoch, args.sample, args.number, test)
# except Exception as e:
# raise ValueError("Bad parameters, please refer to usage")
arg_lst = [
(512, 512),
(1024, 1024),
(2048, 2048),
]
names = ["512", "1024", "2048"]
for i in range(len(arg_lst)):
model_path = "opt_outer_" + names[i] + "_cpu.pkl"
entities = []
func = FUNC_TABLE["outer_product"].func
args = arg_lst[i]
entities.append(Entity("outer_product", args))
model_path = os.path.abspath(model_path)
train_op_schedule_cpu_general_dx(entities,
10,
20,
model_path,
logfile="process_outer_" + names[i] +
"_cpu.txt",
device="cuda:1")
# for i in range(len(arg_lst)):
# print(arg_lst[i])
# N, H, W, C, kernel_size, _, K, stride, padding = arg_lst[i]
# cost10 = pytorch_baseliine(N, H, W, C, kernel_size, K, stride, padding, 10)
# cost100 = pytorch_baseliine(N, H, W, C, kernel_size, K, stride, padding, 100)
# cost1000 = pytorch_baseliine(N, H, W, C, kernel_size, K, stride, padding, 1000)
# logs = "[10:{},100:{},1000:{}]\n".format(cost10, cost100, cost1000)
# with open("logs_pytorch_conv_gpu.txt", "a") as f:
# f.write(logs)
for i in range(len(arg_lst)):
model_path = "../models/opt_conv_" + names[i] + "_gpu_new.pkl"
entities = []
func = FUNC_TABLE["conv2d_channel_batch"].func
args = arg_lst[i]
entities.append(Entity("conv2d_channel_batch", args))
model_path = os.path.abspath(model_path)
beg = time.time()
train_op_schedule_gpu_general_dx(entities, 10, 32, model_path)
end = time.time()
logs = "{}({}):\n".format("conv2d_channel_batch",
args) + "train done! use {}ms\n".format(
(end - beg) * 1e3)
cost10 = test_graph_schedule_gpu_general_dx(entities,
model_path,
number=10)[0]
cost100 = test_graph_schedule_gpu_general_dx(entities,
model_path,
number=100)[0]
cost1000 = test_graph_schedule_gpu_general_dx(entities,
model_path,
use_auto = args.auto_schedule
try:
params = [int(x) for x in args.params.split(",")]
except Exception as e:
raise ValueError("Bad parameters, please refer to usage")
M, N, K, L = params
if use_numpy:
cost = numpy_baseline(M, N, K, L, args.number)
if use_auto:
run(M, N, K, L, args.model_file_path, args.epoch, args.sample,
args.number, test)
arg_lst = [(1024, 1024, 8, 8), (32, 32, 32, 32), (256, 256, 16, 32)]
names = ["1024_1024_8_8", "32_32_32_32", "256_256_16_32"]
for i in range(len(arg_lst)):
model_path = "../models/opt_mttkrp3_" + names[i] + "_cpu_process.pkl"
entities = []
func = FUNC_TABLE["dense_MTTKRP3"].func
args = arg_lst[i]
entities.append(Entity("dense_MTTKRP3", args))
model_path = os.path.abspath(model_path)
train_op_schedule_cpu_general_dx(entities,
5,
32,
model_path,
logfile="process_mttkrp_" + names[i] +
"_cpu.txt",
device="cuda:2")