def test_batch_norm():
input_shape = (1, 4, 4, 16)
target_host = "llvm"
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
inputs1 = nnvm.symbol.Variable("inputs1")
inputs2 = nnvm.symbol.Variable("inputs2")
z1 = nnvm.symbol.relu(inputs1)
# z2 = nnvm.symbol.relu(z1)
compute_graph = nnvm.graph.create(z1)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(compute_graph, target, shape =
{"inputs1" : input_shape}, dtype = "float32", target_host = target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type = 'S0')
deploy_graph, lib, params = nnvm.compiler.build(compute_graph, target, shape =
{"inputs1" : input_shape}, dtype = "float32", target_host = target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.uniform(size = (1, 4, 4, 16), low = -32, high = 32).astype(np.float32)
b_np = np.random.uniform(size = (1, 16), low = -32, high = 32).astype(np.float32)
print(a_np)
module.set_input(inputs1 = a_np)
module.run()
out = module.get_output(0, out = tvm.nd.empty((1, 4, 4, 16)))
print(out.asnumpy)
print(compute_graph.ir())
print(deploy_graph.ir())
def test_dense():
shape = (16, 1024)
weight_shape = (256, 1024)
bias_shape = (256, )
inputs = nnvm.symbol.Variable("inputs")
weights = nnvm.symbol.Variable("weights")
bias = nnvm.symbol.Variable("bias")
env = nnpu.get_env()
target_host = "llvm"
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
z = nnvm.symbol.dense(data=inputs, weight=weights, use_bias=0, units=256)
z1 = nnvm.symbol.relu(z)
compute_graph = nnvm.graph.create(z1)
with nnvm.compiler.build_config(opt_level=1):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={
"inputs": shape,
"weights": weight_shape
},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='SC')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={
"inputs": shape,
"weights": weight_shape
},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(size=shape)
b_np = np.random.random(size=weight_shape)
m.set_input(**{"inputs": a_np, "weights": b_np})
m.run()
gt = a_np.dot(b_np.transpose())
out = m.get_output(0, out=tvm.nd.empty((16, 256)))
np.testing.assert_allclose(out.asnumpy(), gt, rtol=5e-5)
print("tests")
print(out)
print(compute_graph.ir())
print(deploy_graph.ir())
def test_conv2d():
input_shape = (1, 16, 10, 64)
target_host = "llvm"
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
inputs = nnvm.symbol.Variable("inputs")
inputs1 = nnvm.symbol.Variable("inputs1")
z1 = nnvm.symbol.conv2d(data=inputs,
channels=64,
kernel_size=(3, 3),
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI')
z2 = nnvm.symbol.sigmoid(z1)
z = nnvm.symbol.elemwise_add(z2, inputs1)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=1):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={
"inputs": input_shape,
"inputs1": (1, 14, 8, 64)
},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='SC')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": input_shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.uniform(size=input_shape, low=-32,
high=32).astype(np.float32)
b_np = np.random.uniform(size=(1, 14, 8, 64), low=-32,
high=32).astype(np.float32)
module.set_input(inputs=a_np)
module.run()
print(deploy_graph.ir())
out = module.get_output(0, out=tvm.nd.empty((1, 14, 8, 64)))
def test_elemwise_mul():
env = nnpu.get_env()
device = "nnpu"
target_host = "llvm"
target = tvm.target.create("llvm -device={}".format(device))
inputs1 = nnvm.symbol.Variable("inputs1")
inputs2 = nnvm.symbol.Variable("inputs2")
shape = (16, 6, 16)
z = nnvm.symbol.elemwise_mul(inputs1, inputs2)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={
"inputs1": shape,
"inputs2": shape
},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={
"inputs1": shape,
"inputs2": shape
},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random((16, 6, 16))
b_np = np.random.random((16, 6, 16))
print("a_np : ")
print(a_np)
print("b_np : ")
print(b_np)
m.set_input(**{"inputs1": a_np, "inputs2": b_np})
gt = (a_np.astype("float32") *
b_np.astype("float32")).astype("float32")
m.run()
out = m.get_output(0, out=tvm.nd.empty((16, 6, 16)))
np.testing.assert_allclose(out.asnumpy(), gt)
print("elemwise_mul tests success")
print(out)
def test():
env = nnpu.get_env()
nnpu.set_device(env)
shape = (16, )
bigshape = (4, 64)
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
sph = ScheduleProcHelper()
a = tvm.placeholder(bigshape, dtype_n, 'a')
a_buf, a_dram = nnpu.utils.CopyHtoBuf(a, 'a', sph)
str_op = 'VAddMerge'
k = tvm.reduce_axis((0, 4), 'k')
c_buf = tvm.compute((64, ), lambda i: tvm.sum(a_buf[k, i], axis=k),
'c_buf')
sph.MarkScope(c_buf)
c_host, c_dram = nnpu.utils.CopyBufToH(c_buf, 'c', sph)
s = tvm.create_schedule(c_host.op)
sph.Transform(s)
#tensorize
ko, ki = s[c_buf].split(c_buf.op.reduce_axis[0], factor=1)
xo, xi = s[c_buf].split(c_buf.op.axis[0], factor=shape[0])
s[c_buf].reorder(xo, ko, ki, xi)
#s[c_buf].tensorize(ki, env.intrins.get(str_op, mode='n'))
print(nnpu.lower(s, [a, c_host], simple_mode=True))
exit()
func = nnpu.build(s, [a, c_host], 'nnpu', 'llvm', name='nnpu_func')
ctx = tvm.nd.TVMContext(13, 0)
a_np = np.random.randint(size=bigshape, dtype=a.dtype, low=-4, high=4)
a_nd = tvm.nd.array(a_np, ctx)
c_nd = tvm.nd.array(np.zeros((64, ), dtype=c_host.dtype), ctx)
func(a_nd, c_nd)
print(str_op)
print(c_nd.asnumpy())
gt = np.sum(a_np, axis=0, dtype=dtype_w)
print('ground truth=')
print(gt)
np.testing.assert_allclose(c_nd.asnumpy(), gt)
def test_max_pool2d():
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
target_host = "llvm"
inputs = nnvm.symbol.Variable("inputs")
shape = (1, 224, 224, 16)
kernels = nnvm.symbol.Variable("kernels")
kernel_shape = (2, 2)
z = nnvm.symbol.avg_pool2d(inputs,
pool_size=(2, 2),
strides=(1, 1),
layout="NHWC")
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32")
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32")
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(size=(1, 224, 224, 16))
m.set_input(**{"inputs": a_np})
m.run()
out = m.get_output(0, out=tvm.nd.empty((1, 223, 223, 16)))
gt = avg_pooling((1, 224, 224, 16), (1, 223, 223, 16), (2, 2), a_np,
(1, 1), "float32")
np.testing.assert_allclose(out.asnumpy(), gt, rtol=5e-7)
print("max_pool2d tests success")
print(gt)
print(out)
print("end")
def test_log():
env = nnpu.get_env()
shape = (1, 22, 22, 16)
device = "nnpu"
target_host = "llvm"
target = tvm.target.create("llvm -device={}".format(device))
inputs = nnvm.symbol.Variable("inputs")
z = nnvm.symbol.log(inputs)
z1 = nnvm.symbol.exp(z)
compute_graph = nnvm.graph.create(z1)
with nnvm.compiler.build_config(opt_level=1):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(shape)
print(a_np)
m.set_input(**{"inputs": a_np})
m.run()
out = m.get_output(0, out=tvm.nd.empty(shape))
gt = np.exp(np.log(
a_np.astype("float32")).astype("float32")).astype("float32")
print(out)
np.testing.assert_allclose(out.asnumpy(), gt)
print("log tests success")
print(compute_graph.ir())
print(deploy_graph.ir())
def test():
env = nnpu.get_env()
nnpu.set_device(env)
a = tvm.placeholder((4, 4, 16), 'int16', 'a')
#b = tvm.placeholder((16, ), 'int16', 'b')
sph = ScheduleProcHelper()
a_buf, a_dram = nnpu.utils.CopyHtoBuf(a, 'a', sph)
#b_buf, b_dram = nnpu.utils.CopyHtoBuf(b, 'b', sph)
k = tvm.reduce_axis((0, 4), 'k0')
c_buf = tvm.compute((4, 16), lambda i, j: tvm.sum(a_buf[k, i, j], axis=k),
'c_buf')
sph.MarkScope(c_buf)
c_host, c_dram = nnpu.utils.CopyBufToH(c_buf, 'c', sph)
s = tvm.create_schedule(c_host.op)
sph.Transform(s)
ko, ki = s[c_buf].split(c_buf.op.reduce_axis[0], factor=1)
s[c_buf].reorder(c_buf.op.axis[0], ko, ki, c_buf.op.axis[1])
s[c_buf].tensorize(ki, env.intrins.get('VAddMerge', mode='w', nDim=3))
print(nnpu.lower(s, [a, c_host], simple_mode=True))
func = nnpu.build(s, [a, c_host], 'nnpu', 'llvm', name='nnpu_exp')
ctx = tvm.nd.TVMContext(13, 0)
a_np = np.random.randint(size=(4, 4, 16),
dtype=a.dtype,
low=-4000,
high=4000)
#a_np = np.random.random(size=shape).astype(a_host.dtype)
a_nd = tvm.nd.array(a_np, ctx)
c_nd = tvm.nd.array(np.zeros((4, 16)).astype(c_host.dtype), ctx)
func(a_nd, c_nd)
print(c_nd.asnumpy())
print("numpy ground truth is")
gt = np.sum(a_np, axis=0)
print(gt)
def test_relu():
shape = (2, 16)
inputs = nnvm.symbol.Variable("inputs")
env = nnpu.get_env()
target_host = "llvm"
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
z = nnvm.symbol.relu(inputs)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(size=(2, 16)).astype("float32") - 0.5
m.set_input(**{'inputs': a_np})
m.run()
out = m.get_output(0, out=tvm.nd.empty((2, 16)))
print(a_np)
print(out.dtype)
print(out)
np.testing.assert_allclose(out.asnumpy(), np.maximum(a_np, 0))
print("tests")
print(compute_graph.ir())
print(deploy_graph.ir())
def test_onemore():
shape = (1, 32, 32, 16)
inputs = nnvm.symbol.Variable("inputs")
env = nnpu.get_env()
target_host = "llvm"
device = "nnpu"
target = tvm.target.create("llvm -device={}".format(device))
z1 = nnvm.symbol.relu(inputs)
z = nnvm.symbol.sqrt(z1)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"inputs": shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
m = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(size=(1, 32, 32, 16))
m.set_input(**{'inputs': a_np})
m.run()
out = m.get_output(0, out=tvm.nd.empty((1, 32, 32, 16)))
print(out)
print(compute_graph.ir())
print(deploy_graph.ir())
def test():
env = nnpu.get_env()
nnpu.set_device(env)
shape = (2, 16)
a_host = tvm.placeholder(shape, env.cfg['dtype_n'], 'a_host')
print('a host ' + str(a_host))
a = tvm.compute(shape, lambda *i: a_host(*i), name='a')
a_buf = tvm.compute(shape, lambda *i: a(*i), name='a_buf')
b_buf = tvm.compute(
shape,
lambda i, j: tvm.log(a_buf[i, j].astype(env.cfg['dtype_w'])),
name='b_buf')
b = tvm.compute(shape, lambda *i: b_buf(*i), name='b')
b_host = tvm.compute(shape, lambda *i: b(*i), name='b_host')
s = tvm.create_schedule(b_host.op)
# mark variable scopes
s[a].set_scope(env.dram_scope)
s[b].set_scope(env.dram_scope)
s[a_buf].set_scope(env.uni_scratchpad_scope)
s[b_buf].set_scope(env.uni_scratchpad_scope)
#print
# (dir(s[b].op.body))
# mark compiler pragmas
s[a].pragma(s[a].op.axis[0], env.dma_copy_pragma)
s[b_host].pragma(s[b_host].op.axis[0], env.dma_copy_pragma)
s[a_buf].pragma(s[a_buf].op.axis[0], env.scratchpad_ls)
s[b].pragma(s[b].op.axis[0], env.scratchpad_ls)
s[a_buf].compute_at(s[b_buf], b_buf.op.axis[0])
# tensorize
s[b_buf].tensorize(s[b_buf].op.axis[1], env.intrins.get('VLOG',
mode='inc'))
# build
print(tvm.lower(s, [a_host, b_host], simple_mode=True))
print(nnpu.lower(s, [a_host, b_host], simple_mode=True))
#exit()
func = nnpu.build(s, [a_host, b_host], 'nnpu', 'llvm', name='nnpu_log')
print('function built: ')
#print(func.get_source())
# prepare data
ctx = tvm.nd.TVMContext(13, 0) #???
print('i want to know:')
print(ctx.exist)
a_np = np.random.randint(size=shape, dtype=a_host.dtype, low=1, high=20)
a_nd = tvm.nd.array(a_np, ctx)
b_nd = tvm.nd.array(np.zeros(shape).astype(b_host.dtype), ctx)
# run
func(a_nd, b_nd)
print('run finished')
b_np = b_nd.asnumpy()
print('a=')
print(a_np)
print('b=')
print(b_np)
print('ground truth =')
gt = np.log(a_np, dtype=b_host.dtype)
print(gt)
np.testing.assert_allclose(b_np, gt)
def test():
env = nnpu.get_env()
nnpu.set_device(env)
shape = (2, 2, 16)
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
a = tvm.placeholder(shape, dtype_w, 'a')
sph = ScheduleProcHelper()
a_buf, a_dram = nnpu.utils.CopyHtoBuf(a, 'a', sph)
k = tvm.reduce_axis((0, 2), 'k')
add_buf = tvm.compute(
(2, 16), lambda i, j: tvm.sum(a_buf[k, i, j], axis=k), 'add_buf')
sph.MarkScope(add_buf)
add_host, add_dram = nnpu.utils.CopyBufToH(add_buf, 'add', sph)
k1 = tvm.reduce_axis((0, 2), 'k1')
mul_buf = tvm.compute(
(2, 16), lambda i, j: tvm.sum(a_buf[k1, i, j], axis=k1), 'mul_buf')
sph.MarkScope(mul_buf)
mul_host, mul_dram = nnpu.utils.CopyBufToH(mul_buf, 'mul', sph)
s = tvm.create_schedule([add_host.op, mul_host.op])
sph.Transform(s)
ko, ki = s[add_buf].split(add_buf.op.reduce_axis[0], factor=1)
s[add_buf].reorder(ko, ki, *(s[add_buf].op.axis))
s[add_buf].tensorize(ki, env.intrins.get('MAddMerge',
shape=shape,
mode='w'))
ko1, ki1 = s[mul_buf].split(mul_buf.op.reduce_axis[0], factor=1)
s[mul_buf].reorder(ko1, ki1, *(s[mul_buf].op.axis))
s[mul_buf].tensorize(ki1,
env.intrins.get('MMulMerge', shape=shape, mode='w'))
print(nnpu.lower(s, [a, add_host, mul_host], simple_mode=True))
func = nnpu.build(s, [a, add_host, mul_host],
'nnpu',
'llvm',
name='nnpu_func')
#exit()
ctx = tvm.nd.TVMContext(13, 0)
a_np = np.random.randint(size=(2, 2, 16), dtype=a.dtype, low=-16, high=16)
a_nd = tvm.nd.array(a_np, ctx)
add_nd = tvm.nd.array(np.zeros((2, 16)).astype(add_host.dtype), ctx)
mul_nd = tvm.nd.array(np.zeros((2, 16)).astype(mul_host.dtype), ctx)
func(a_nd, add_nd, mul_nd)
print('a = ')
print(a_np)
print('reduce sum row = ')
print(add_nd.asnumpy())
print('ground truth is: ')
gt = np.sum(a_np, axis=0)
print(gt)
np.testing.assert_allclose(add_nd.asnumpy(), gt)
print('reduce mul row = ')
print(mul_nd.asnumpy())
gt = np.multiply.reduce(a_np, axis=0, dtype=a.dtype)
print(gt)
np.testing.assert_allclose(mul_nd.asnumpy(), gt)
def test_Alexnet():
def Conv(data, kernel_size, filter_nums, stride=(1, 1), pad=(0, 0)):
if pad[0] != 0 or pad[1] != 0:
data = nnvm.symbol.pad(data=data,
pad_width=((0, 0), (pad[0], pad[0]),
(pad[1], pad[1]), (0, 0)))
datas = nnvm.symbol.conv2d(data=data,
kernel_size=kernel_size,
channels=filter_nums,
strides=stride,
layout='NHWC',
kernel_layout='HWOI')
datas = nnvm.symbol.relu(data=datas)
return datas
def get_symbol(datas, num_classes):
conv1 = Conv(data=datas,
kernel_size=(11, 11),
filter_nums=96,
stride=(4, 4))
pool1 = nnvm.symbol.max_pool2d(data=conv1,
pool_size=(3, 3),
strides=(2, 2),
layout='NHWC')
conv2 = Conv(data=pool1,
kernel_size=(5, 5),
filter_nums=256,
pad=(2, 2))
pool2 = nnvm.symbol.max_pool2d(data=conv2,
pool_size=(3, 3),
strides=(2, 2),
layout='NHWC')
conv3 = Conv(data=pool2,
kernel_size=(3, 3),
filter_nums=384,
pad=(1, 1))
conv4 = Conv(data=conv3,
kernel_size=(3, 3),
filter_nums=384,
pad=(1, 1))
conv5 = Conv(data=conv4,
kernel_size=(3, 3),
filter_nums=256,
pad=(1, 1))
pool3 = nnvm.symbol.max_pool2d(data=conv5,
pool_size=(3, 3),
strides=(2, 2),
layout='NHWC')
datas = nnvm.symbol.flatten(data=pool3)
fc1 = nnvm.symbol.dense(data=datas, units=1024)
relu1 = nnvm.symbol.relu(data=fc1)
drop1 = nnvm.symbol.dropout(data=relu1, rate=0.5)
fc2 = nnvm.symbol.dense(data=drop1, units=1024)
relu2 = nnvm.symbol.relu(data=fc2)
drop2 = nnvm.symbol.dropout(data=relu2, rate=0.5)
fc3 = nnvm.symbol.dense(data=drop2, units=16)
symbol = nnvm.symbol.softmax(fc3)
return symbol
input_shape = (1, 128, 128, 16)
target_host = "llvm"
device = "nnpu"
data = nnvm.symbol.Variable(name="data")
target = tvm.target.create("llvm -device={}".format(device))
print("ok")
num_runs = 1
z = get_symbol(datas=data, num_classes=16)
compute_graph = nnvm.graph.create(z)
print(compute_graph.ir())
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='SC')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.randint(size=input_shape, low=-32, high=32)
print(a_np)
module.set_input(data=a_np)
ftimer = module.module.time_evaluator("run",
ctx,
number=num_runs,
repeat=1)
# module.run()
out = module.get_output(0, out=tvm.nd.empty((1, 16)))
print(out.asnumpy)
print(deploy_graph.ir())
print(ftimer().mean * 10)
def test_densenet():
def Conv(datas, kernel_size, filter_nums, stride=(1, 1), pad=(0, 0)):
if pad[0] != 0 or pad[1] != 0:
datas = nnvm.symbol.pad(data=datas,
pad_width=((0, 0), (pad[0], pad[0]),
(pad[1], pad[1]), (0, 0)))
conv = nnvm.symbol.conv2d(data=datas,
kernel_size=kernel_size,
channels=filter_nums,
strides=stride,
layout='NHWC',
kernel_layout='HWOI')
return conv
def bottleneck_layer(datas, filters):
bn1 = nnvm.symbol.batch_norm(data=datas, epsilon=2e-5, axis=3)
relu1 = nnvm.symbol.relu(data=bn1)
conv1 = Conv(datas=relu1, kernel_size=(1, 1), filter_nums=4 * filters)
bn2 = nnvm.symbol.batch_norm(data=conv1, epsilon=2e-5, axis=3)
relu2 = nnvm.symbol.relu(data=bn2)
conv2 = Conv(datas=relu2,
kernel_size=(3, 3),
filter_nums=filters,
pad=(1, 1))
return conv2
def transition_layer(datas, filters):
conv = Conv(datas=datas, kernel_size=(1, 1), filter_nums=filters)
pool = nnvm.symbol.avg_pool2d(data=conv,
pool_size=(2, 2),
strides=(2, 2),
layout='NHWC')
return pool
def dense_block(datas, filters, layers):
layers_concat = []
layers_concat.append(datas)
b_l = bottleneck_layer(datas, filters)
layers_concat.append(b_l)
for i in range(layers - 1):
x = nnvm.symbol.concatenate(*layers_concat, axis=3)
x = bottleneck_layer(x, filters)
layers_concat.append(x)
return x
def get_symbol(datas, num_classes=16):
x = Conv(datas, kernel_size=(7, 7), filter_nums=32, stride=(2, 2))
x = nnvm.symbol.max_pool2d(x,
pool_size=(3, 3),
strides=(2, 2),
layout='NHWC')
b1 = dense_block(x, 32, 6)
l1 = transition_layer(b1, 32)
b2 = dense_block(l1, 32, 12)
l2 = transition_layer(b2, 32)
b3 = dense_block(l2, 32, 48)
l3 = transition_layer(b3, 32)
b4 = dense_block(l3, 32, 32)
x = nnvm.symbol.global_avg_pool2d(data=b4, layout='NHWC')
x = nnvm.symbol.flatten(data=x)
fc = nnvm.symbol.dense(data=x, units=16)
symbol = nnvm.symbol.softmax(data=fc)
return symbol
input_shape = (1, 229, 229, 16)
target_host = "llvm"
device = "nnpu"
data = nnvm.symbol.Variable(name="data")
target = tvm.target.create("llvm -device={}".format(device))
print("ok")
num_runs = 3
z = get_symbol(datas=data, num_classes=16)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.random(size=input_shape)
print(a_np)
module.set_input(data=a_np)
ftimer = module.module.time_evaluator("run",
ctx,
number=num_runs,
repeat=1)
module.run()
out = module.get_output(0, out=tvm.nd.empty((1, 16)))
print(out.asnumpy)
print(deploy_graph.ir())
print(ftimer().mean)
import nnpu
import tvm
import topi
from nnpu.utils import ScheduleProcHelper
import numpy as np
with ScheduleProcHelper():
env = nnpu.get_env()
nnpu.set_device(env)
t = 10 # time
n = 16 # input depth
m = 16 # output depth
x_shape = (t, n)
h_shape = (t, m)
w_shape = (m, n)
u_shape = (m, m)
b_shape = (m, )
gemm_shape = (16, 16, 1)
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
x = tvm.placeholder(x_shape, dtype_n, 'x')
x_buf, _ = nnpu.utils.CopyHtoBuf(x, 'x')
w = tvm.placeholder(w_shape, dtype_n, 'w')
w_buf, _ = nnpu.utils.CopyHtoBuf(w, 'w')
u = tvm.placeholder(u_shape, dtype_w, 'u')
u_buf, _ = nnpu.utils.CopyHtoBuf(u, 'u')
b = tvm.placeholder(b_shape, dtype_w, 'b')
from nnpu.utils import ScheduleProcHelper
import numpy as np
import argparse
parser = argparse.ArgumentParser(description='test of NNPU Op')
parser.add_argument('--sim',
type=str,
help='the simulator to use',
default='S0',
choices=['S0', 'S1', 'SC'])
args = parser.parse_args()
env = nnpu.get_env()
assert env.cfg['multi_core'], 'multi core test need multi_core switch on'
nnpu.set_device(env, type=args.sim)
with ScheduleProcHelper():
env = nnpu.get_env()
shape1 = (8, 128) # (8, 128) reshaped & transpoased to (16, 8, 8)
shape2 = (128, 128) # (128, 128) tiled to (16, 128, 8)
gemm_shape = (8, 8, 8)
factor = gemm_shape[1]
assert shape1[1] == shape2[1], \
'gemm do dot product between rows, so the shape[1] of inputs should match'
assert shape1[0] % gemm_shape[
0] == 0, 'gemm insn require size of input 1 be x{0}'.format(
gemm_shape[0])
assert shape2[0] % gemm_shape[
2] == 0, 'gemm insn require size of input 2 be x{0}'.format(
gemm_shape[0])
def test_vgg():
def get_feature(internel_layer, layers, filters, batch_norm=False):
"""
Get VGG feature body as stacks of convoltions.
layers : [1, 1, 2, 2, 2]
filters : [64, 128, 256, 512, 512]
"""
for i, num in enumerate(layers):
"""
i = 0, num = 1
i = 1, num = 1
i = 2, num = 2
i = 3, num = 2
i = 4, num = 2
"""
for j in range(num):
internel_layer = sym.pad(data=internel_layer,
pad_width=((0, 0), (1, 1), (1, 1),
(0, 0)))
internel_layer = sym.conv2d(data=internel_layer,
kernel_size=(3, 3),
channels=filters[i],
layout='NHWC',
kernel_layout='HWOI',
name="conv%s_%s" % (i + 1, j + 1))
if batch_norm:
internel_layer = sym.batch_norm(data=internel_layer,
axis=3,
name="bn%s_%s" %
(i + 1, j + 1))
internel_layer = sym.relu(data=internel_layer,
name="relu%s_%s" % (i + 1, j + 1))
internel_layer = sym.max_pool2d(data=internel_layer,
pool_size=(2, 2),
strides=(2, 2),
layout="NHWC",
name="pool%s" % (i + 1))
return internel_layer
def get_classifier(input_data, num_classes):
"""
Get VGG classifier layers as fc layers.
"""
flatten = sym.flatten(data=input_data, name="flatten")
fc1 = sym.dense(data=flatten, units=32, name="fc1")
relu1 = sym.relu(data=fc1, name="relu1")
drop1 = sym.dropout(data=relu1, rate=0.5, name="drop1")
fc2 = sym.dense(data=drop1, units=32, name="fc2")
relu2 = sym.relu(data=fc2, name="relu2")
drop2 = sym.dropout(data=relu2, rate=0.5, name="drop2")
fc3 = sym.dense(data=drop2, units=num_classes, name="fc3")
return fc3
def get_symbol(datas, num_classes, num_layers=11, batch_norm=False):
"""
Parameters
------------
num_classes : int, default 16
Number of classification classes
num_layers : int
Number of layers for the variant of vgg. Options are 11, 13, 16, 19
batch_norm : bool, default False
Use batch normalization.
"""
vgg_spec = {
11: ([1, 1, 2, 2, 2], [64, 128, 256, 512, 512]),
13: ([2, 2, 2, 2, 2], [64, 128, 256, 512, 512]),
16: ([2, 2, 3, 3, 3], [64, 128, 256, 512, 512]),
19: ([2, 2, 4, 4, 4], [64, 128, 256, 512, 512])
}
if num_layers not in vgg_spec:
raise ValueError(
"Invalide num_layers {}. Choices are 11, 13, 16, 19.".format(
num_layers))
layers, filters = vgg_spec[num_layers]
feature = get_feature(datas, layers, filters, batch_norm)
classifier = get_classifier(feature, num_classes)
symbol = sym.softmax(data=classifier, name="softmax")
return symbol
input_shape = (1, 224, 224, 16)
target_host = "llvm"
device = "nnpu"
data = nnvm.symbol.Variable(name="data")
target = tvm.target.create("llvm -device={}".format(device))
print("ok")
num_runs = 1
z = get_symbol(datas=data, num_classes=16)
compute_graph = nnvm.graph.create(z)
print(compute_graph.ir())
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
else:
nnpu.set_device(nnpu.get_env(), type='SC')
with ScheduleProcHelper():
with nnpu.build_config():
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.uniform(size=input_shape, low=-32,
high=32).astype(np.float32)
print(a_np)
module.set_input(data=a_np)
ftimer = module.module.time_evaluator("run",
ctx,
number=num_runs,
repeat=1)
# module.run()
out = module.get_output(0, out=tvm.nd.empty((1, 16)))
print(out.asnumpy)
print(deploy_graph.ir())
print(ftimer().mean * 10)
def test_inception_v3():
def Conv(data,
num_filter,
kernel=(1, 1),
stride=(1, 1),
pad=(0, 0),
name=None,
suffix=''):
if pad[0] != 0 or pad[1] != 0:
data = sym.pad(data=data,
pad_width=((0, 0), (pad[0], pad[0]),
(pad[1], pad[1]), (0, 0)))
conv = sym.conv2d(data=data,
channels=num_filter,
kernel_size=kernel,
strides=stride,
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name="%s%s_conv2d" % (name, suffix))
bn = sym.batch_norm(data=conv,
name="%s%s_batchnorm" % (name, suffix),
epsilon=2e-5,
axis=3)
act = sym.relu(data=bn, name="%s%s_relu" % (name, suffix))
return act
def Pooling(data, kernel, stride, pad, pool_type, name):
if pad[0] != 0 or pad[1] != 0:
data = sym.pad(data=data,
pad_width=((0, 0), (pad[0], pad[0]),
(pad[1], pad[1]), (0, 0)))
if pool_type == 'max':
return sym.max_pool2d(data=data,
pool_size=kernel,
strides=stride,
name=name,
layout='NHWC')
if pool_type == 'avg':
return sym.avg_pool2d(data=data,
pool_size=kernel,
strides=stride,
name=name,
layout='NHWC')
raise ValueError("Invalid pooling type: " + pool_type)
def Inception7A(data, num_1x1, num_3x3_red, num_3x3_1, num_3x3_2,
num_5x5_red, num_5x5, pool, proj, name):
# num_1x1 = 64
# num_3x3_red = 64
# num_3x3_1 = 96
# num_3x3_2 = 96
# num_5x5_red = 48
# num_5x5 = 64
tower_1x1 = Conv(data, num_1x1, name=('%s_conv' % name))
tower_5x5 = Conv(data,
num_5x5_red,
name=('%s_tower' % name),
suffix='_conv')
tower_5x5 = Conv(tower_5x5,
num_5x5,
kernel=(5, 5),
pad=(2, 2),
name=('%s_tower' % name),
suffix='_conv_1')
tower_3x3 = Conv(data,
num_3x3_red,
name=('%s_tower_1' % name),
suffix="_conv")
tower_3x3 = Conv(tower_3x3,
num_3x3_1,
kernel=(3, 3),
pad=(1, 1),
name=('%s_tower_1' % name),
suffix='_conv_1')
tower_3x3 = Conv(tower_3x3,
num_3x3_2,
kernel=(3, 3),
pad=(1, 1),
name=('%s_tower_1' % name),
suffix='_conv_2')
pooling = Pooling(data,
kernel=(3, 3),
stride=(1, 1),
pad=(1, 1),
pool_type=pool,
name=('%s_pool_%s_pool' % (pool, name)))
cproj = Conv(pooling, proj, name=('%s_tower_2' % name), suffix='_conv')
concat = sym.concatenate(*[tower_1x1, tower_5x5, tower_3x3, cproj],
axis=3,
name='ch_concat_%s_chconcat' % name)
return concat
def Inception7B(data, num_3x3, num_d3x3_red, num_d3x3_1, num_d3x3_2, pool,
name):
tower_3x3 = Conv(data,
num_3x3,
kernel=(3, 3),
pad=(0, 0),
stride=(2, 2),
name=('%s_conv' % name))
tower_d3x3 = Conv(data,
num_d3x3_red,
name=('%s_tower' % name),
suffix='_conv')
tower_d3x3 = Conv(tower_d3x3,
num_d3x3_1,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
name=('%s_tower' % name),
suffix='_conv_1')
tower_d3x3 = Conv(tower_d3x3,
num_d3x3_2,
kernel=(3, 3),
pad=(0, 0),
stride=(2, 2),
name=('%s_tower' % name),
suffix='_conv_2')
pooling = Pooling(data=data,
kernel=(3, 3),
stride=(2, 2),
pad=(0, 0),
pool_type="max",
name=('max_pool_%s_pool' % name))
concat = sym.concatenate(*[tower_3x3, tower_d3x3, pooling],
axis=3,
name='ch_concat_%s_chconcat' % name)
return concat
def Inception7C(data, num_1x1, num_d7_red, num_d7_1, num_d7_2, num_q7_red,
num_q7_1, num_q7_2, num_q7_3, num_q7_4, pool, proj, name):
tower_1x1 = Conv(data=data,
num_filter=num_1x1,
kernel=(1, 1),
name=('%s_conv' % name))
tower_d7 = Conv(data=data,
num_filter=num_d7_red,
name=('%s_tower' % name),
suffix='_conv')
tower_d7 = Conv(data=tower_d7,
num_filter=num_d7_1,
kernel=(1, 7),
pad=(0, 3),
name=('%s_tower' % name),
suffix='_conv_1')
tower_d7 = Conv(data=tower_d7,
num_filter=num_d7_2,
kernel=(7, 1),
pad=(3, 0),
name=('%s_tower' % name),
suffix='_conv_2')
tower_q7 = Conv(data=data,
num_filter=num_q7_red,
name=('%s_tower_1' % name),
suffix='_conv')
tower_q7 = Conv(data=tower_q7,
num_filter=num_q7_1,
kernel=(7, 1),
pad=(3, 0),
name=('%s_tower_1' % name),
suffix='_conv_1')
tower_q7 = Conv(data=tower_q7,
num_filter=num_q7_2,
kernel=(1, 7),
pad=(0, 3),
name=('%s_tower_1' % name),
suffix='_conv_2')
tower_q7 = Conv(data=tower_q7,
num_filter=num_q7_3,
kernel=(7, 1),
pad=(3, 0),
name=('%s_tower_1' % name),
suffix='_conv_3')
tower_q7 = Conv(data=tower_q7,
num_filter=num_q7_4,
kernel=(1, 7),
pad=(0, 3),
name=('%s_tower_1' % name),
suffix='_conv_4')
pooling = Pooling(data=data,
kernel=(3, 3),
stride=(1, 1),
pad=(1, 1),
pool_type=pool,
name=('%s_pool_%s_pool' % (pool, name)))
cproj = Conv(data=pooling,
num_filter=proj,
kernel=(1, 1),
name=('%s_tower_2' % name),
suffix='_conv')
concat = sym.concatenate(*[tower_1x1, tower_d7, tower_q7, cproj],
axis=3,
name='ch_concat_%s_chconcat' % name)
return concat
def Inception7D(data, num_3x3_red, num_3x3, num_d7_3x3_red, num_d7_1,
num_d7_2, num_d7_3x3, pool, name):
tower_3x3 = Conv(data=data,
num_filter=num_3x3_red,
name=('%s_tower' % name),
suffix='_conv')
tower_3x3 = Conv(data=tower_3x3,
num_filter=num_3x3,
kernel=(3, 3),
pad=(0, 0),
stride=(2, 2),
name=('%s_tower' % name),
suffix='_conv_1')
tower_d7_3x3 = Conv(data=data,
num_filter=num_d7_3x3_red,
name=('%s_tower_1' % name),
suffix='_conv')
tower_d7_3x3 = Conv(data=tower_d7_3x3,
num_filter=num_d7_1,
kernel=(1, 7),
pad=(0, 3),
name=('%s_tower_1' % name),
suffix='_conv_1')
tower_d7_3x3 = Conv(data=tower_d7_3x3,
num_filter=num_d7_2,
kernel=(7, 1),
pad=(3, 0),
name=('%s_tower_1' % name),
suffix='_conv_2')
tower_d7_3x3 = Conv(data=tower_d7_3x3,
num_filter=num_d7_3x3,
kernel=(3, 3),
stride=(2, 2),
name=('%s_tower_1' % name),
suffix='_conv_3')
pooling = Pooling(data=data,
kernel=(3, 3),
stride=(2, 2),
pool_type=pool,
pad=(0, 0),
name=('%s_pool_%s_pool' % (pool, name)))
concat = sym.concatenate(*[tower_3x3, tower_d7_3x3, pooling],
axis=3,
name='ch_concat_%s_chconcat' % name)
return concat
def Inception7E(data, num_1x1, num_d3_red, num_d3_1, num_d3_2,
num_3x3_d3_red, num_3x3, num_3x3_d3_1, num_3x3_d3_2, pool,
proj, name):
tower_1x1 = Conv(data=data,
num_filter=num_1x1,
kernel=(1, 1),
name=('%s_conv' % name))
tower_d3 = Conv(data=data,
num_filter=num_d3_red,
name=('%s_tower' % name),
suffix='_conv')
tower_d3_a = Conv(data=tower_d3,
num_filter=num_d3_1,
kernel=(1, 3),
pad=(0, 1),
name=('%s_tower' % name),
suffix='_mixed_conv')
tower_d3_b = Conv(data=tower_d3,
num_filter=num_d3_2,
kernel=(3, 1),
pad=(1, 0),
name=('%s_tower' % name),
suffix='_mixed_conv_1')
tower_3x3_d3 = Conv(data=data,
num_filter=num_3x3_d3_red,
name=('%s_tower_1' % name),
suffix='_conv')
tower_3x3_d3 = Conv(data=tower_3x3_d3,
num_filter=num_3x3,
kernel=(3, 3),
pad=(1, 1),
name=('%s_tower_1' % name),
suffix='_conv_1')
tower_3x3_d3_a = Conv(data=tower_3x3_d3,
num_filter=num_3x3_d3_1,
kernel=(1, 3),
pad=(0, 1),
name=('%s_tower_1' % name),
suffix='_mixed_conv')
tower_3x3_d3_b = Conv(data=tower_3x3_d3,
num_filter=num_3x3_d3_2,
kernel=(3, 1),
pad=(1, 0),
name=('%s_tower_1' % name),
suffix='_mixed_conv_1')
pooling = Pooling(data=data,
kernel=(3, 3),
stride=(1, 1),
pad=(1, 1),
pool_type=pool,
name=('%s_pool_%s_pool' % (pool, name)))
cproj = Conv(data=pooling,
num_filter=proj,
kernel=(1, 1),
name=('%s_tower_2' % name),
suffix='_conv')
concat = sym.concatenate(*[
tower_1x1, tower_d3_a, tower_d3_b, tower_3x3_d3_a, tower_3x3_d3_b,
cproj
],
axis=3,
name='ch_concat_%s_chconcat' % name)
return concat
def get_symbol(data, num_classes=16, **kwargs):
conv = Conv(data, 32, kernel=(3, 3), stride=(2, 2), name="conv")
conv_1 = Conv(conv, 32, kernel=(3, 3), name="conv_1")
conv_2 = Conv(conv_1, 64, kernel=(3, 3), pad=(1, 1), name="conv_2")
pool = Pooling(data=conv_2,
kernel=(3, 3),
stride=(2, 2),
pool_type="max",
pad=(0, 0),
name="pool")
conv_3 = Conv(pool, 80, kernel=(1, 1), name="conv_3")
conv_4 = Conv(conv_3, 192, kernel=(3, 3), name="conv_4")
pool1 = Pooling(data=conv_4,
kernel=(3, 3),
stride=(2, 2),
pool_type="max",
pad=(0, 0),
name="pool1")
in3a = Inception7A(pool1, 64, 64, 96, 96, 48, 64, "avg", 32, "mixed")
in3b = Inception7A(in3a, 64, 64, 96, 96, 48, 64, "avg", 64, "mixed_1")
in3c = Inception7A(in3b, 64, 64, 96, 96, 48, 64, "avg", 64, "mixed_2")
in3d = Inception7B(in3c, 384, 64, 96, 96, "max", "mixed_3")
in4a = Inception7C(in3d, 192, 128, 128, 192, 128, 128, 128, 128, 192,
"avg", 192, "mixed_4")
in4b = Inception7C(in4a, 192, 160, 160, 192, 160, 160, 160, 160, 192,
"avg", 192, "mixed_5")
in4c = Inception7C(in4b, 192, 160, 160, 192, 160, 160, 160, 160, 192,
"avg", 192, "mixed_6")
in4d = Inception7C(in4c, 192, 192, 192, 192, 192, 192, 192, 192, 192,
"avg", 192, "mixed_7")
in4e = Inception7D(in4d, 192, 320, 192, 192, 192, 192, "max",
"mixed_8")
in5a = Inception7E(in4e, 320, 384, 384, 384, 448, 384, 384, 384, "avg",
192, "mixed_9")
in5b = Inception7E(in5a, 320, 384, 384, 384, 448, 384, 384, 384, "max",
192, "mixed_10")
pool = Pooling(data=in5b,
kernel=(8, 8),
stride=(1, 1),
pool_type="avg",
pad=(0, 0),
name="global_pool")
flatten = sym.flatten(data=pool, name="flatten")
fc1 = sym.dense(data=flatten, units=num_classes, name="fc1")
softmax = sym.softmax(data=fc1, name="softmax")
return softmax
input_shape = (1, 299, 299, 16)
target_host = "llvm"
device = "nnpu"
data = nnvm.symbol.Variable(name="data")
target = tvm.target.create("llvm -device={}".format(device))
print("ok")
num_runs = 3
z = get_symbol(data=data, num_classes=16)
compute_graph = nnvm.graph.create(z)
with nnvm.compiler.build_config(opt_level=1):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.uniform(size=(1, 299, 299, 16), low=-32,
high=32).astype(np.float32)
print(a_np)
module.set_input(data=a_np)
ftimer = module.module.time_evaluator("run",
ctx,
number=num_runs,
repeat=1)
module.run()
out = module.get_output(0, out=tvm.nd.empty((1, 16)))
print(out.asnumpy)
print(deploy_graph.ir())
print(ftimer().mean * 10)
import nnpu
import tvm
import topi
from nnpu.utils import ScheduleProcHelper
import numpy as np
with (ScheduleProcHelper()):
env = nnpu.get_env()
nnpu.set_device(env, type='S0')
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
a = tvm.placeholder((2, 4, 16), dtype_n, 'a')
a_buf, a_dram = nnpu.utils.CopyHtoBuf(a, 'a')
pad_buf = tvm.compute((2, 6, 16), lambda i, j, k: tvm.expr.Select(
j >= 2, a_buf[i, j - 2, k], tvm.const(0, dtype_n)), 'pad')
nnpu.utils.MarkScope(pad_buf)
nnpu.utils.PragmaCopy(pad_buf)
tile_host, _ = nnpu.utils.CopyBufToH(pad_buf, 'tile')
s = nnpu.create_schedule([tile_host.op])
print(tvm.lower(s, [a, tile_host], simple_mode=True))
print(nnpu.lower(s, [a, tile_host], simple_mode=True))
# exit(0)
func = nnpu.build(s, [a, tile_host], 'nnpu', 'llvm', name='nnpu_func')
ctx = tvm.nd.TVMContext(13, 0)
a_np = np.random.randint(size=(2, 4, 16),
dtype=a.dtype,
low=-128,
def test_ib():
print('aaaa')
env = nnpu.get_env()
nnpu.set_device(env)
shape = (16, )
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
a = tvm.placeholder(shape, dtype_w, name='a')
w = shape[0]
e = 16
def build_nms_ir(ten_in, ten_out):
ib = tvm.ir_builder.create()
imm_value = 10
ib.scope_attr(env.nnpu_axis, "coproc_scope", 0)
p_in = ib.buffer_ptr(ten_in[0])
p_out = ib.buffer_ptr(ten_out[0])
#with ib.for_range(0,w, name="k") as k:
with ib.for_range(0, w / e, name="i") as i:
ib.emit(
make_intrin_call(
"void", 'VAddI', ten_out[0].access_ptr("w", 'uint32') +
i * dtype_bytes(dtype_w),
ten_in[0].access_ptr("r", 'uint32') +
i * dtype_bytes(dtype_w), tvm.const(imm_value, 'float64'),
env.cfg['vector_unit']['size'], 3))
stmt = ib.get()
return stmt
sph = ScheduleProcHelper()
a_buf, a_dram = nnpu.utils.CopyHtoBuf(a, 'a', sph)
sph.MarkScope(a_buf)
out = tvm.extern(a_buf.shape, [a_buf],
build_nms_ir,
in_buffers=[
tvm.decl_buffer(a_buf.shape,
dtype_w,
data_alignment=dtype_bytes(dtype_w),
scope='local.nnpu_scratchpad0')
],
out_buffers=[
tvm.decl_buffer(a_buf.shape,
dtype_w,
data_alignment=dtype_bytes(dtype_w),
scope='local.nnpu_scratchpad0')
],
dtype=dtype_w,
name="test_ir")
sph.MarkScope(out)
out_host, out_dram = nnpu.utils.CopyBufToH(out, 'out', sph)
s = tvm.create_schedule([out_host.op])
sph.Transform(s)
print(tvm.lower(s, [a, out_host], simple_mode=True))
print(nnpu.lower(s, [a, out_host], simple_mode=True))
# exit(0)
func = nnpu.build(s, [a, out_host], 'nnpu', 'llvm', name='nnpu_test')
ctx = tvm.nd.TVMContext(13, 0)
a_np = np.random.randint(size=(16, ), dtype=a.dtype, low=0, high=127)
a_nd = tvm.nd.array(a_np, ctx)
b_nd = tvm.nd.array(np.zeros(16, ).astype(out_host.dtype), ctx)
func(a_nd, b_nd)
print('a = ')
print(a_np)
print('xjb sum = ')
print(b_nd.asnumpy())
return
def test_resnets():
def residual_unit(data,
num_filter,
stride,
dim_match,
name,
bottle_neck=True):
# bottle_neck : False
"""
Return Resnet Unit symbol for building Resnet
Parameters
------------
data : str
Input data
num_filter : int
Number of output channels
stride : tuple
Stride used in convolution
dim_match : Boolean
True means channel number between input and output is the same
otherwise means differ
"""
if bottle_neck:
bn1 = nnvm.symbol.batch_norm(data=data,
axis=3,
epsilon=2e-5,
name=name + '_bn1')
act1 = nnvm.symbol.relu(data=bn1, name=name + '_relu1')
conv1 = nnvm.symbol.conv2d(data=act1,
channels=int(num_filter * 0.25),
kernel_size=(1, 1),
strides=stride,
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_conv1')
bn2 = nnvm.symbol.batch_norm(data=conv1,
axis=3,
epsilon=2e-5,
name=name + '_bn2')
act2 = nnvm.symbol.relu(data=bn2, name=name + '_relu2')
pad = nnvm.symbol.pad(data=act2,
pad_width=((0, 0), (1, 1), (1, 1), (0, 0)))
conv2 = nnvm.symbol.conv2d(data=pad,
channels=int(num_filter * 0.25),
kernel_size=(3, 3),
strides=(1, 1),
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_conv2')
bn3 = nnvm.symbol.batch_norm(data=conv2,
axis=3,
epsilon=2e-5,
name=name + '_bn3')
act3 = nnvm.symbol.relu(data=bn3, name=name + '_relu3')
conv3 = nnvm.symbol.conv2d(data=act3,
channels=num_filter,
kernel_size=(1, 1),
strides=(1, 1),
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_conv3')
if dim_match:
shortcut = data
else:
shortcut = nnvm.symbol.conv2d(data=act1,
channels=num_filter,
kernel_size=(1, 1),
strides=stride,
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_sc')
return nnvm.symbol.elemwise_add(conv3, shortcut)
else:
# bottle_neck = False
# i = 0 : filter_list[1] = 64, (1, 1), False
# i = 1 : filter_list[2] = 128, (2, 2), False
# i = 2 : filter_list[3] = 256, (2, 2), False
# i = 3 : filter_list[4] = 512, (2, 2), False
# bn1 = nnvm.symbol.batch_norm(data = data, axis = 3, epsilon = 2e-5, name = name + '_bn1')
act1 = nnvm.symbol.relu(data=data, name=name + '_relu1')
# (56, 56, 64)
# num_filter = filter_list[1] = 64
# strides = (1, 1)
pad1 = nnvm.symbol.pad(data=act1,
pad_width=((0, 0), (1, 1), (1, 1), (0, 0)))
conv1 = nnvm.symbol.conv2d(data=pad1,
channels=num_filter,
kernel_size=(3, 3),
strides=stride,
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_bn2')
# i = 0 : (56, 56, 64)
# i = 1 : (28, 28, 128)
# i = 2 : (14, 14, 256)
# i = 3 : (7, 7, 512)
# bn2 = nnvm.symbol.batch_norm(data = conv1, axis = 3, epsilon = 2e-5, name = name + '_bn2')
act2 = nnvm.symbol.relu(data=conv1, name=name + '_relu2')
# (56, 56, 64)
pad2 = nnvm.symbol.pad(data=act2,
pad_width=((0, 0), (1, 1), (1, 1), (0, 0)))
conv2 = nnvm.symbol.conv2d(data=pad2,
channels=num_filter,
kernel_size=(3, 3),
strides=(1, 1),
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_conv2')
# i = 0 : (56, 56, 64)
# i = 1 : (28, 28, 128)
# i = 2 : (14, 14, 256)
if dim_match:
shortcut = data
else:
shortcut = nnvm.symbol.conv2d(data=act1,
channels=num_filter,
kernel_size=(1, 1),
strides=stride,
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name=name + '_sc')
return nnvm.symbol.elemwise_add(conv2, shortcut)
def resnet(datas,
units,
num_stages,
filter_list,
num_classes,
image_shape,
bottle_neck=True):
# units = [2, 2, 2, 2]
# num_stages = 4
# filter_list = [64, 64, 128, 256, 512]
# num_classes = 1000
# image_shape = (224, 224, 16)
# bottle_neck = False
"""
Return Resnet symbol of
Parameters
------------
units : list
Number of units in each stage
num_stage : int
Number of stage
filter_list : list
Channel size of each stage
num_classes : int
Output size of symbol
"""
num_unit = len(units)
assert num_unit == num_stages
data = nnvm.symbol.batch_norm(data=datas,
axis=3,
epsilon=2e-5,
scale=False,
name="bn_data")
(_, height, _, _) = image_shape
if height <= 32:
pad = nnvm.symbol.pad(data=data,
pad_width=((0, 0), (1, 1), (1, 1), (0, 0)))
body = nnvm.symbol.conv2d(data=pad,
channels=filter_list[0],
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name="conv0")
else:
pad = nnvm.symbol.pad(data=data,
pad_width=((0, 0), (3, 3), (3, 3), (0, 0)))
body = nnvm.symbol.conv2d(data=pad,
channels=filter_list[0],
kernel_size=(7, 7),
strides=(2, 2),
padding=(0, 0),
use_bias=False,
layout='NHWC',
kernel_layout='HWOI',
name="conv0")
# body.shape = (112, 112, 64)
# body = nnvm.symbol.batch_norm(data = body, axis = 3, epsilon = 2e-5, name = "bn0")
body = nnvm.symbol.relu(data=body, name="relu0")
# body = nnvm.symbol.pad(data = body, pad_width = ((0, 0), (1, 1), (1, 1), (0, 0)))
body = nnvm.symbol.max_pool2d(data=body,
pool_size=(3, 3),
strides=(2, 2),
layout='NHWC')
# body.shape = (56, 56, 64)
for i in range(num_stages):
# num_stages == 4
# i = 0: (56, 56, 64)
# i = 1: filter_list[2] = 128, (2, 2), False (28, 28, 128)
# i = 2: filter_list[3] = 256, (2, 2), False
# i = 3: filter_list[4] = 512, (2, 2), False
body = residual_unit(body,
filter_list[i + 1],
(1 if i == 0 else 2, 1 if i == 0 else 2),
False,
name='stage%d_unit%d' % (i + 1, 1),
bottle_neck=bottle_neck)
# (56, 56, 64)
# units[0] - 1 = 1
for j in range(units[i] - 1):
body = residual_unit(body,
filter_list[i + 1], (1, 1),
True,
name="stage%d_unit%d" % (i + 1, j + 2),
bottle_neck=bottle_neck)
# (56, 56, 64)
# (7, 7, 512)
# bn1 = nnvm.symbol.batch_norm(data = body, axis = 3, epsilon = 2e-5, name = "bn1")
relu1 = nnvm.symbol.relu(data=body, name="relu1")
pool1 = nnvm.symbol.global_avg_pool2d(data=relu1,
layout='NHWC',
name="pool1")
# (1, 1, 512)
flat = nnvm.symbol.flatten(data=pool1)
# (512)
fc1 = nnvm.symbol.dense(data=flat, units=num_classes, name='fc1')
return nnvm.symbol.softmax(data=fc1, name='softmax')
def get_symbol(datas,
num_classes,
num_layers=50,
image_shape=(1, 224, 224, 16),
**kwargs):
(_, height, _, _) = image_shape
if height <= 28:
num_stages = 3
if (num_layers - 2) % 9 == 0 and num_layers >= 164:
per_unit = [(num_layers - 2) // 9]
filter_list = [16, 64, 128, 256]
bottle_neck = True
elif (num_layers - 2) % 6 == 0 and num_layers < 164:
per_unit = [(num_layers - 2) // 6]
filter_list = [16, 16, 32, 64]
bottle_neck = False
else:
raise ValueError(
"no experiments done on num_layers {}".format(num_layers))
units = per_unit * num_stages
else:
print("height = 224 > 28")
# height = 224 > 28
if num_layers >= 50:
filter_list = [64, 256, 512, 1024, 2048]
bottle_neck = True
else:
print("num_layers = 18 < 50")
# num_layers = 18 < 50
filter_list = [64, 64, 128, 256, 512]
bottle_neck = False
num_stages = 4
if num_layers == 18:
units = [2, 2, 2, 2]
elif num_layers == 34:
units = [3, 4, 6, 3]
elif num_layers == 50:
units = [3, 4, 6, 3]
elif num_layers == 101:
units = [3, 4, 23, 3]
elif num_layers == 152:
units = [3, 8, 36, 3]
elif num_layers == 200:
units = [3, 24, 36, 3]
elif num_layers == 269:
units = [3, 30, 48, 8]
else:
raise ValueError(
"no experiments done on num_layers {}".format(num_layers))
return resnet(datas=datas,
units=units,
num_stages=num_stages,
filter_list=filter_list,
num_classes=num_classes,
image_shape=image_shape,
bottle_neck=bottle_neck)
input_shape = (1, 224, 224, 16)
target_host = "llvm"
device = "nnpu"
data = nnvm.symbol.Variable(name="data")
target = tvm.target.create("llvm -device={}".format(device))
print("ok")
num_runs = 3
z = get_symbol(datas=data,
num_classes=16,
num_layers=18,
image_shape=(1, 224, 224, 16))
compute_graph = nnvm.graph.create(z)
print(compute_graph.ir())
with nnvm.compiler.build_config(opt_level=0):
if target.device_name != "nnpu":
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
else:
with ScheduleProcHelper():
with nnpu.build_config():
nnpu.set_device(nnpu.get_env(), type='S0')
deploy_graph, lib, params = nnvm.compiler.build(
compute_graph,
target,
shape={"data": input_shape},
dtype="float32",
target_host=target_host)
print(deploy_graph.ir())
ctx = tvm.context(str("nnpu"), 0) if device == "nnpu" else tvm.context(
str("llvm"), 0)
module = runtime.create(deploy_graph, lib, ctx)
a_np = np.random.uniform(size=(1, 224, 224, 16), low=-32,
high=32).astype(np.float32)
print(a_np)
module.set_input(data=a_np)
ftimer = module.module.time_evaluator("run",
ctx,
number=num_runs,
repeat=1)
module.run()
out = module.get_output(0, out=tvm.nd.empty((1, 16)))
print(out.asnumpy)
print(deploy_graph.ir())
print(ftimer().mean * 10)
