def run(
act_dtype=ng.int16,
weight_dtype=ng.int8,
bias_dtype=ng.int32,
scale_dtype=ng.int8,
with_batchnorm=True,
disable_fusion=False,
conv2d_par_ich=1,
conv2d_par_och=1,
conv2d_par_col=1,
conv2d_par_row=1,
conv2d_concur_och=None,
conv2d_stationary='filter',
pool_par=1,
elem_par=1,
chunk_size=64,
axi_datawidth=32,
silent=False,
filename=None,
# simtype='iverilog',
# simtype='verilator',
simtype=None, # no RTL simulation
outputfile=None):
# input mean and standard deviation
imagenet_mean = np.array([0.485, 0.456, 0.406]).astype(np.float32)
imagenet_std = np.array([0.229, 0.224, 0.225]).astype(np.float32)
act_shape = (1, 224, 224, 3)
if not with_batchnorm:
raise ValueError('with_batchnorm must be True for ResNet18.')
# pytorch model
model = torchvision.models.resnet18(pretrained=True)
# Pytorch to ONNX
onnx_filename = 'resnet18_imagenet.onnx'
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['out']
model.eval()
torch.onnx.export(model,
dummy_input,
onnx_filename,
input_names=input_names,
output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
dtypes = {}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=dtypes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=weight_dtype,
default_constant_dtype=weight_dtype,
default_operator_dtype=act_dtype,
default_scale_dtype=scale_dtype,
default_bias_dtype=bias_dtype,
disable_fusion=disable_fusion)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
if act_dtype.width > 8:
act_scale_factor = 128
else:
act_scale_factor = int(round(2**(act_dtype.width - 1) * 0.5))
input_scale_factors = {'act': act_scale_factor}
input_means = {'act': imagenet_mean * act_scale_factor}
input_stds = {'act': imagenet_std * act_scale_factor}
ng.quantize(outputs, input_scale_factors, input_means, input_stds)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.conv2d):
op.attribute(par_ich=conv2d_par_ich,
par_och=conv2d_par_och,
par_col=conv2d_par_col,
par_row=conv2d_par_row,
concur_och=conv2d_concur_och,
stationary=conv2d_stationary)
if isinstance(op, (ng.avg_pool, ng.max_pool, ng.avg_pool_serial,
ng.max_pool_serial)):
op.attribute(par=pool_par)
if ng.is_elementwise_operator(op):
op.attribute(par=elem_par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
act = placeholders['act']
out = outputs['out']
# verification data
img = np.array(PIL.Image.open('car.png').convert('RGB')).astype(np.float32)
img = img.reshape([1] + list(img.shape))
img = img / 255
img = (img - imagenet_mean) / imagenet_std
# execution on pytorch
model_input = np.broadcast_to(img, act_shape)
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_out = model(torch.from_numpy(model_input)).detach().numpy()
if act.perm is not None and len(model_out.shape) == len(act.shape):
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
# software-based verification
vact = img * act_scale_factor
vact = np.clip(vact, -1.0 * (2**(act.dtype.width - 1) - 1),
1.0 * (2**(act.dtype.width - 1) - 1))
vact = np.round(vact).astype(np.int64)
vact = np.broadcast_to(vact, act_shape)
# compare outputs of hidden layers
relu_op = [
v for k, v in operators.items()
if isinstance(v, ng.conv2d) and not isinstance(v, ng.matmul)
][0]
maxpool_op = [
v for k, v in operators.items()
if isinstance(v, (ng.max_pool, ng.max_pool_serial))
][0]
relu_ops = [v for k, v in operators.items() if isinstance(v, ng.relu)]
layer1_0_op = relu_ops[0]
layer1_op = relu_ops[1]
layer2_0_op = relu_ops[2]
layer2_op = relu_ops[3]
layer3_0_op = relu_ops[4]
layer3_op = relu_ops[5]
layer4_0_op = relu_ops[6]
layer4_op = relu_ops[7]
avgpool_op = [
v for k, v in operators.items()
if isinstance(v, (ng.avg_pool, ng.avg_pool_serial))
][0]
fc_op = [v for k, v in operators.items() if isinstance(v, ng.matmul)][0]
sub_ops = [
relu_op, maxpool_op, layer1_0_op, layer1_op, layer2_0_op, layer2_op,
layer3_0_op, layer3_op, layer4_0_op, layer4_op, avgpool_op, fc_op
]
sub_outs = ng.eval(sub_ops, act=vact)
sub_outs = [sub_out.transpose([0, 3, 1, 2])
for sub_out in sub_outs[:-1]] + sub_outs[-1:]
sub_scale_factors = [sub_op.scale_factor for sub_op in sub_ops]
model.eval()
model_relu_out = nn.Sequential(model.conv1, model.bn1, model.relu)(
torch.from_numpy(model_input)).detach().numpy()
model_maxpool_out = nn.Sequential(
model.conv1, model.bn1, model.relu,
model.maxpool)(torch.from_numpy(model_input)).detach().numpy()
# class model_layer1_0(nn.Module):
# def __init__(self):
# super(model_layer1_0, self).__init__()
# self.conv1 = model.conv1
# self.bn1 = model.bn1
# self.relu = model.relu
# self.maxpool = model.maxpool
# self.layer1_0 = model.layer1[0]
#
# def forward(self, x):
# x = self.relu(self.bn1(self.conv1(x)))
# x = self.maxpool(x)
# x = self.layer1_0(x)
# return x
#
# model_layer1_0_out = model_layer1_0()(torch.from_numpy(model_input)).detach().numpy()
model_layer1_0_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool,
model.layer1[0])(torch.from_numpy(model_input)).detach().numpy()
model_layer1_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool,
model.layer1)(torch.from_numpy(model_input)).detach().numpy()
model_layer2_0_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2[0])(torch.from_numpy(model_input)).detach().numpy()
model_layer2_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2)(torch.from_numpy(model_input)).detach().numpy()
model_layer3_0_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2,
model.layer3[0])(torch.from_numpy(model_input)).detach().numpy()
model_layer3_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2,
model.layer3)(torch.from_numpy(model_input)).detach().numpy()
model_layer4_0_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2, model.layer3,
model.layer4[0])(torch.from_numpy(model_input)).detach().numpy()
model_layer4_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2, model.layer3,
model.layer4)(torch.from_numpy(model_input)).detach().numpy()
model_avgpool_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool, model.layer1,
model.layer2, model.layer3, model.layer4,
model.avgpool)(torch.from_numpy(model_input)).detach().numpy()
class Flatten(nn.Module):
def forward(self, input):
return input.view(input.size(0), -1)
model_fc_out = nn.Sequential(
model.conv1, model.bn1, model.relu, model.maxpool,
model.layer1, model.layer2, model.layer3, model.layer4, model.avgpool,
Flatten(), model.fc)(torch.from_numpy(model_input)).detach().numpy()
model_outs = [
model_relu_out, model_maxpool_out, model_layer1_0_out,
model_layer1_out, model_layer2_0_out, model_layer2_out,
model_layer3_0_out, model_layer3_out, model_layer4_0_out,
model_layer4_out, model_avgpool_out, model_fc_out
]
scaled_outs = [
model_out * scale_factor
for model_out, scale_factor in zip(model_outs, sub_scale_factors)
]
max_diffs = [
model_out.max() / sub_out.max()
for model_out, sub_out in zip(scaled_outs, sub_outs)
]
overflows = [
np.sum(np.abs(sub_out) >= abs(2**(sub_op.dtype.width - 1) - 1))
for sub_op, sub_out in zip(sub_ops, sub_outs)
]
mean_square_errors = [
np.sum((sub_out - model_out)**2) / sub_out.size
for model_out, sub_out in zip(scaled_outs, sub_outs)
]
corrcoefs = [
np.corrcoef(model_out.reshape([-1]), sub_out.reshape([-1]))
for model_out, sub_out in zip(model_outs, sub_outs)
]
# compare prediction results
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
mean_square_error = np.sum((vout - scaled_model_out)**2) / vout.size
corrcoef = np.corrcoef(model_out.reshape([-1]), vout.reshape([-1]))
class_index = json.load(open('imagenet_class_index.json', 'r'))
labels = {int(key): value for (key, value) in class_index.items()}
mout = scaled_model_out
for bat in range(mout.shape[0]):
m_top10 = list(
sorted(enumerate(mout[bat]), key=lambda x: x[1],
reverse=True))[:10]
m_top10_indexes = [index for index, value in m_top10]
v_top10 = list(
sorted(enumerate(vout[bat]), key=lambda x: x[1],
reverse=True))[:10]
v_top10_indexes = [index for index, value in v_top10]
num_hit = 0
score = 0
for index, value in m_top10:
print("# mout: %s (%d) = %f" % (str(labels[index]), index, value))
for index, value in v_top10:
print("# vout: %s (%d) = %d" % (str(labels[index]), index, value))
if index in m_top10_indexes:
num_hit += 1
score += 10 - abs(
m_top10_indexes.index(index) -
v_top10_indexes.index(index))
print("# top-10 hit: %d" % num_hit)
print("# top-10 score: %d" % score)
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
# to Veriloggen object
# targ = ng.to_veriloggen([out], 'resnet18', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# to IP-XACT (the method returns Veriloggen object, as well as to_veriloggen)
targ = ng.to_ipxact([out],
'resnet18',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog([out], 'resnet18', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([out], chunk_size)
param_bytes = len(param_data)
variable_addr = int(math.ceil(
(act.addr + act.memory_size) / chunk_size)) * chunk_size
check_addr = int(math.ceil(
(variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil(
(check_addr + out.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
# mem = np.zeros([1024 * 1024 * 256 // (memimg_datawidth // 8)], dtype=np.int64)
mem = np.zeros([1024 * 1024 * 1024 // (memimg_datawidth // 8)],
dtype=np.int16)
mem = mem + [100]
# placeholder
axi.set_memory(
mem, vact, memimg_datawidth, act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(
mem, vout, memimg_datawidth, act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(out.shape[0]):
for x in range(out.shape[1]):
orig = memory.read_word(bat * out.aligned_shape[1] + x,
out.addr, act_dtype.width)
check = memory.read_word(bat * out.aligned_shape[1] + x,
check_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, x, ') orig: ', orig, ' check: ', check)
ok = False
else:
print('OK (', bat, x, ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(act_shape=(1, 4, 4, 3),
weight0_shape=(9, 3, 3, 3),
weight1_shape=(9, 36),
act_dtype=ng.int32,
weight_dtype=ng.int32,
stride0=1,
padding0=0,
with_batchnorm0=False,
with_batchnorm1=False,
act_func0='ReLU',
act_func1='relu',
disable_fusion=False,
par_ich=1,
par_och=1,
par_col=1,
par_row=1,
concur_och=None,
stationary='filter',
chunk_size=64,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# pytorch model
layers = []
layers.append(
nn.Conv2d(weight0_shape[3],
weight0_shape[0],
weight0_shape[1],
stride=stride0,
padding=padding0))
if with_batchnorm0:
layers.append(nn.BatchNorm2d(weight0_shape[0]))
if act_func0 is not None:
layers.append(getattr(nn, act_func0)())
class Transpose(nn.Module):
def __init__(self, perm):
super(Transpose, self).__init__()
self.perm = perm
def forward(self, input):
return input.permute(*self.perm)
layers.append(Transpose([0, 1, 3, 2]))
class Flatten(nn.Module):
def forward(self, input):
# return input.view(input.size(0), -1)
return torch.reshape(input, (input.size(0), -1))
layers.append(Flatten())
layers.append(nn.Linear(weight1_shape[1], weight1_shape[0]))
if with_batchnorm1:
layers.append(nn.BatchNorm2d(weight1_shape[0]))
if act_func1 is not None:
layers.append(getattr(nn, act_func1)())
model = nn.Sequential(*layers)
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_conv2d_transpose_linear.onnx'
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['out']
model.eval()
torch.onnx.export(model,
dummy_input,
onnx_filename,
input_names=input_names,
output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
value_dtypes = {
'act': act_dtype,
'0.weight': weight_dtype,
'3.weight': weight_dtype,
'out': act_dtype
}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=value_dtypes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=weight_dtype,
default_constant_dtype=weight_dtype,
default_operator_dtype=act_dtype,
default_scale_dtype=ng.int32,
default_bias_dtype=ng.int32,
disable_fusion=disable_fusion)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
if act_dtype.width > 8:
act_scale_factor = 128
else:
act_scale_factor = int(round(2**(act_dtype.width - 1) * 0.5))
input_scale_factors = {'act': act_scale_factor}
ng.quantize(outputs, input_scale_factors)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.conv2d):
op.attribute(par_ich=par_ich,
par_och=par_och,
par_row=par_row,
par_col=par_col,
concur_och=concur_och)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
act = placeholders['act']
out = outputs['out']
# verification data
# random data
std = 0.2
mean = 0.5
img = np.random.normal(size=act.length).astype(np.float32).reshape(
act.shape)
img = img * std + mean
# execution on pytorch
model_input = img
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_out = model(torch.from_numpy(model_input)).detach().numpy()
if act.perm is not None and len(model_out.shape) == len(act.shape):
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
# software-based verification
vact = img * act_scale_factor
vact = np.clip(vact, -1.0 * (2**(act.dtype.width - 1) - 1),
1.0 * (2**(act.dtype.width - 1) - 1))
vact = np.round(vact).astype(np.int64)
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
mean_square_error = np.sum((vout - scaled_model_out)**2) / vout.size
corrcoef = np.corrcoef(model_out.reshape([-1]), vout.reshape([-1]))
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
targ = ng.to_veriloggen([out],
'onnx_matrix_conv2d_transpose_linear',
silent=silent,
config={
'maxi_datawidth': axi_datawidth,
'chunk_size': chunk_size
})
# --------------------
# (6) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([out], chunk_size)
param_bytes = len(param_data)
variable_addr = int(math.ceil(
(act.addr + act.memory_size) / chunk_size)) * chunk_size
check_addr = int(math.ceil(
(variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil(
(check_addr + out.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(
mem, vact, memimg_datawidth, act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(
mem, vout, memimg_datawidth, act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for i in range(out.shape[0]):
for j in range(out.shape[1]):
orig = memory.read_word(i * out.aligned_shape[1] + j, out.addr,
act_dtype.width)
check = memory.read_word(i * out.aligned_shape[1] + j,
check_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', i, j, ') orig: ', orig, 'check: ', check)
ok = False
# else:
# print('OK (', i, j, ') orig: ', orig, 'check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(a_shape=(7, 15),
b_shape=(7, 15),
a_dtype=ng.int32,
b_dtype=ng.int32,
c_dtype=ng.int32,
par=1,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# pytorch model
model = MatrixMul()
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_mul.onnx'
dummy_a = torch.randn(*a_shape)
dummy_b = torch.randn(*b_shape)
dummy_inputs = (dummy_a, dummy_b)
input_names = ['a', 'b']
output_names = ['c']
model.eval()
torch.onnx.export(model,
dummy_inputs,
onnx_filename,
input_names=input_names,
output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
value_dtypes = {'a': a_dtype, 'b': b_dtype, 'c': c_dtype}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=value_dtypes,
default_placeholder_dtype=ng.int32,
default_variable_dtype=ng.int32,
default_constant_dtype=ng.int32,
default_operator_dtype=ng.int32,
default_scale_dtype=ng.int32,
default_bias_dtype=ng.int32,
disable_fusion=False)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
input_scale_factors = {'a': 10.0, 'b': 15.0}
ng.quantize(outputs, input_scale_factors)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.scaled_multiply):
op.attribute(par=par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
a = placeholders['a']
b = placeholders['b']
c = outputs['c']
# verification data
input_a = np.arange(a.length, dtype=np.int64).reshape(a.shape) % [17]
input_b = (np.arange(b.length, dtype=np.int64).reshape(b.shape) +
[100]) % [13]
# execution on pytorch
model_a = input_a.astype(np.float32)
if a.perm is not None:
model_a = np.transpose(model_a, a.reversed_perm)
model_b = input_b.astype(np.float32)
if b.perm is not None:
model_b = np.transpose(model_b, b.reversed_perm)
model.eval()
model_c = model(torch.from_numpy(model_a),
torch.from_numpy(model_b)).detach().numpy()
if a.perm is not None:
model_c = np.transpose(model_c, a.perm)
scaled_model_c = model_c * c.scale_factor
# software-based verification
va = input_a * input_scale_factors['a']
va = np.clip(va, -1.0 * (2**(a.dtype.width - 1) - 1),
1.0 * (2**(a.dtype.width - 1) - 1))
va = np.round(va).astype(np.int64)
vb = input_b * input_scale_factors['b']
vb = np.clip(vb, -1.0 * (2**(b.dtype.width - 1) - 1),
1.0 * (2**(b.dtype.width - 1) - 1))
vb = np.round(vb).astype(np.int64)
eval_outs = ng.eval([c], a=va, b=vb)
vc = eval_outs[0]
mean_square_error = np.sum((vc - scaled_model_c)**2) / vc.size
corrcoef = np.corrcoef(model_c.reshape([-1]), vc.reshape([-1]))
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
targ = ng.to_veriloggen([c],
'onnx_matrix_mul',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([c])
param_bytes = len(param_data)
variable_addr = int(
math.ceil(
max(a.addr + a.memory_size, b.addr + b.memory_size) / 4096)) * 4096
check_addr = int(math.ceil((variable_addr + param_bytes) / 4096)) * 4096
tmp_addr = int(math.ceil((check_addr + c.memory_size) / 4096)) * 4096
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(mem, va, memimg_datawidth, a_dtype.width, a.addr,
max(int(math.ceil(axi_datawidth / a_dtype.width)), par))
axi.set_memory(mem, vb, memimg_datawidth, b_dtype.width, b.addr,
max(int(math.ceil(axi_datawidth / b_dtype.width)), par))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(mem, vc, memimg_datawidth, c_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / c_dtype.width)), par))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
num_rep = functools.reduce(lambda x, y: x * y, c.shape[:-1], 1)
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for i in range(num_rep):
for j in range(c.shape[-1]):
orig = memory.read_word(i * c.aligned_shape[-1] + j, c.addr,
c_dtype.width)
check = memory.read_word(i * c.aligned_shape[-1] + j,
check_addr, c_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG', i, j, orig, check)
ok = False
# else:
# print('OK', i, j, orig, check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(1000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(act_shape=(1, 7, 7, 3),
act_dtype=ng.int32,
ksize=2, stride=2, padding=0,
par=1,
chunk_size=64,
axi_datawidth=32, silent=False,
filename=None, simtype='iverilog', outputfile=None):
# pytorch model
layers = []
layers.append(nn.AvgPool2d(ksize, stride=stride, padding=padding))
model = nn.Sequential(*layers)
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_avg_pool.onnx'
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['out']
torch.onnx.export(model, dummy_input, onnx_filename,
input_names=input_names, output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
value_dtypes = {'act': act_dtype,
'out': act_dtype}
(outputs, placeholders, variables,
constants, operators) = ng.from_onnx(onnx_filename,
value_dtypes=value_dtypes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=ng.int32,
default_constant_dtype=ng.int32,
default_operator_dtype=act_dtype,
default_scale_dtype=ng.int32,
default_bias_dtype=ng.int32,
disable_fusion=False)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
input_scale_factors = {'act': 1.0}
ng.quantize(outputs, input_scale_factors)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.avg_pool):
op.attribute(par=par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
act = placeholders['act']
out = outputs['out']
# verification data
if act_dtype.width > 4:
vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [11] + [1]
else:
vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [5] + [1]
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
# software-based verification
model_input = vact.astype(np.float32)
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_out = model(torch.from_numpy(model_input)).detach().numpy()
if act.perm is not None:
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
mean_square_error = np.sum((vout - scaled_model_out) ** 2) / vout.size
corrcoef = np.corrcoef(model_out.reshape([-1]), vout.reshape([-1]))
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
targ = ng.to_veriloggen([out], 'onnx_matrix_avg_pool', silent=silent,
config={'maxi_datawidth': axi_datawidth,
'chunk_size': chunk_size})
# --------------------
# (6) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([out])
param_bytes = len(param_data)
variable_addr = int(math.ceil((act.addr + act.memory_size) / chunk_size)) * chunk_size
check_addr = int(math.ceil((variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil((check_addr + out.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)], dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(mem, vact, memimg_datawidth,
act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth,
8, variable_addr)
# verification data
axi.set_memory(mem, vout, memimg_datawidth,
act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m, 'memory', clk, rst,
datawidth=axi_datawidth,
memimg=mem, memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(
time_counter.inc()
)
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(out.shape[0]):
for y in range(out.shape[1]):
for x in range(out.shape[2]):
for ch in range(out.shape[3]):
orig = memory.read_word(
bat * out.aligned_shape[1] * out.aligned_shape[2] * out.aligned_shape[3] +
y * out.aligned_shape[2] * out.aligned_shape[3] +
x * out.aligned_shape[3] + ch,
out.addr, act_dtype.width)
check = memory.read_word(
bat * out.aligned_shape[1] * out.aligned_shape[2] * out.aligned_shape[3] +
y * out.aligned_shape[2] * out.aligned_shape[3] +
x * out.aligned_shape[3] + ch,
check_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, y, x, ch,
') orig: ', orig, ' check: ', check)
ok = False
# else:
# print('OK (', bat, y, x, ch,
# ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ, 'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, resetn, m.make_reset(), period=100, polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog([output_layer], 'hello_nngen', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Save the quantized weights
# --------------------
# convert weight values to a memory image:
# on a real FPGA platform, this image will be used as a part of the model definition.
param_filename = 'hello_nngen.npy'
chunk_size = 64
param_data = ng.export_ndarray([output_layer], chunk_size)
np.save(param_filename, param_data)
# --------------------
# (7) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
# If you don't check the RTL behavior, exit here.
# print('# Skipping RTL simulation. If you simulate the RTL behavior, comment out the next line.')
# sys.exit()
import math
from veriloggen import *
import veriloggen.thread as vthread
import veriloggen.types.axi as axi
def run(a_shape=(7, 15),
b_shape=(7, 15),
a_dtype=ng.int32,
b_dtype=ng.int32,
c_dtype=ng.int32,
par=1,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# model definition
model = MatrixAdd()
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_add.onnx'
dummy_a = torch.randn(*a_shape)
dummy_b = torch.randn(*b_shape)
dummy_inputs = (dummy_a, dummy_b)
input_names = ['a', 'b']
output_names = ['c']
model.eval()
torch.onnx.export(model,
dummy_inputs,
onnx_filename,
input_names=input_names,
output_names=output_names)
# ONNX to NNgen
value_dtypes = {'a': a_dtype, 'b': b_dtype, 'c': c_dtype}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=value_dtypes,
default_placeholder_dtype=ng.int32,
default_variable_dtype=ng.int32,
default_constant_dtype=ng.int32,
default_operator_dtype=ng.int32,
default_scale_dtype=ng.int32,
default_bias_dtype=ng.int32,
disable_fusion=False)
# set attribute
for op in operators.values():
if isinstance(op, ng.add):
op.attribute(par=par)
# create target hardware
a = placeholders['a']
b = placeholders['b']
c = outputs['c']
targ = ng.to_veriloggen([c],
'onnx_matrix_add',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
va = np.arange(a.length, dtype=np.int64).reshape(a.shape) % [5]
vb = (np.arange(b.length, dtype=np.int64).reshape(b.shape) + [100]) % [6]
eval_outs = ng.eval([c], a=va, b=vb)
vc = eval_outs[0]
# exec on pytorch
model_a = va.astype(np.float32)
model_b = vb.astype(np.float32)
if a.perm is not None:
model_a = np.transpose(model_a, a.reversed_perm)
if b.perm is not None:
model_b = np.transpose(model_b, b.reversed_perm)
model.eval()
model_c = model(torch.from_numpy(model_a),
torch.from_numpy(model_b)).detach().numpy()
if a.perm is not None:
model_c = np.transpose(model_c, a.perm)
scaled_model_c = model_c * c.scale_factor
c_diff = vc - scaled_model_c
c_err = c_diff / (scaled_model_c + 0.00000001)
max_c_err = np.max(np.abs(c_err))
# if max_c_err > 0.1:
# raise ValueError("too large output error: %f > 0.1" % max_c_err)
# to memory image
param_data = ng.export_ndarray([c])
param_bytes = len(param_data)
variable_addr = int(
math.ceil(
max(a.addr + a.memory_size, b.addr + b.memory_size) / 4096)) * 4096
check_addr = int(math.ceil((variable_addr + param_bytes) / 4096)) * 4096
tmp_addr = int(math.ceil((check_addr + c.memory_size) / 4096)) * 4096
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(mem, va, memimg_datawidth, a_dtype.width, a.addr,
max(int(math.ceil(axi_datawidth / a_dtype.width)), par))
axi.set_memory(mem, vb, memimg_datawidth, b_dtype.width, b.addr,
max(int(math.ceil(axi_datawidth / b_dtype.width)), par))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(mem, vc, memimg_datawidth, c_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / c_dtype.width)), par))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
num_rep = functools.reduce(lambda x, y: x * y, c.shape[:-1], 1)
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for i in range(num_rep):
for j in range(c.shape[-1]):
orig = memory.read_word(i * c.aligned_shape[-1] + j, c.addr,
c_dtype.width)
check = memory.read_word(i * c.aligned_shape[-1] + j,
check_addr, c_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG', i, j, orig, check)
ok = False
# else:
# print('OK', i, j, orig, check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(1000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(act_shape=(1, 7, 7, 15),
weight_shape=(15, 3, 3, 15),
act_dtype=ng.int32,
weight_dtype=ng.int32,
stride=1,
padding=1,
with_batchnorm=False,
act_func='relu',
disable_fusion=False,
par_ich=1,
par_och=1,
par_col=1,
par_row=1,
concur_och=None,
stationary='filter',
chunk_size=64,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# model definition
model = MatrixConv2dConcat(weight_shape, stride, padding, with_batchnorm,
act_func)
# overwrite weight values for test
# model.conv.weight.data = torch.from_numpy(np.ones_like(model.conv.weight.data.numpy()))
# model.conv.bias.data = torch.from_numpy(np.zeros_like(model.conv.bias.data.numpy()))
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_conv2d_concat.onnx'
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['out']
model.eval()
torch.onnx.export(model,
dummy_input,
onnx_filename,
input_names=input_names,
output_names=output_names)
# ONNX to NNgen
value_dtypes = {
'act': act_dtype,
'0.weight': weight_dtype,
'out': act_dtype
}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=value_dtypes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=weight_dtype,
default_constant_dtype=weight_dtype,
default_operator_dtype=act_dtype,
default_scale_dtype=ng.int32,
default_bias_dtype=ng.int32,
disable_fusion=disable_fusion)
# default linear quantization
if act_dtype.width >= 8:
value_ranges = {'act': (-120, 120)}
else:
value_ranges = {
'act': (-(2**(act_dtype.width - 1)), (2**(act_dtype.width - 1)))
}
ng.quantize(outputs, value_ranges=value_ranges)
# set attribute
for op in operators.values():
if isinstance(op, ng.conv2d):
op.attribute(par_ich=par_ich,
par_och=par_och,
par_row=par_row,
par_col=par_col,
concur_och=concur_och)
# create target hardware
act = placeholders['act']
out = outputs['out']
targ = ng.to_veriloggen([out],
'onnx_matrix_conv2d',
silent=silent,
config={
'maxi_datawidth': axi_datawidth,
'chunk_size': chunk_size
})
# verification data
# if act_dtype.width > 4:
# vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [11] + [1]
# else:
# vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [5] + [1]
#vact = np.ones(act.shape)
#vact = np.zeros(act.shape)
vact = np.random.normal(size=act.length).reshape(act.shape)
vact = np.clip(vact, -3.0, 3.0)
vact_min_val, vact_max_val = value_ranges['act']
vact_max_abs_range = max(abs(vact_min_val), abs(vact_max_val))
vact_width = vact_max_abs_range.bit_length() + 1
vact = vact * (1.0 * (2**(vact_width - 1) - 1)) / 3.0
vact = np.round(vact).astype(np.int64)
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
# exec on pytorch
model_input = vact.astype(np.float32)
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_out = model(torch.from_numpy(model_input)).detach().numpy()
if act.perm is not None:
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
out_diff = vout - scaled_model_out
out_err = out_diff / (scaled_model_out + 0.00000001)
max_out_err = np.max(np.abs(out_err))
# if max_out_err > 0.1:
# raise ValueError("too large output error: %f > 0.1" % max_out_err)
# to memory image
param_data = ng.export_ndarray([out], chunk_size)
param_bytes = len(param_data)
variable_addr = int(math.ceil(
(act.addr + act.memory_size) / chunk_size)) * chunk_size
check_addr = int(math.ceil(
(variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil(
(check_addr + out.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(
mem, vact, memimg_datawidth, act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(
mem, vout, memimg_datawidth, act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(out.shape[0]):
for y in range(out.shape[1]):
for x in range(out.shape[2]):
for ch in range(out.shape[3]):
orig = memory.read_word(
bat * out.aligned_shape[1] * out.aligned_shape[2] *
out.aligned_shape[3] +
y * out.aligned_shape[2] * out.aligned_shape[3] +
x * out.aligned_shape[3] + ch, out.addr,
act_dtype.width)
check = memory.read_word(
bat * out.aligned_shape[1] * out.aligned_shape[2] *
out.aligned_shape[3] +
y * out.aligned_shape[2] * out.aligned_shape[3] +
x * out.aligned_shape[3] + ch, check_addr,
act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, y, x, ch, ') orig: ', orig,
' check: ', check)
ok = False
# else:
# print('OK (', bat, y, x, ch,
# ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(act_dtype=ng.int16, weight_dtype=ng.int16,
bias_dtype=ng.int32, scale_dtype=ng.int16,
with_batchnorm=False, disable_fusion=False,
conv2d_par_ich=1, conv2d_par_och=1, conv2d_par_col=1, conv2d_par_row=1,
conv2d_concur_och=None, conv2d_stationary='filter',
pool_par=1, elem_par=1,
chunk_size=64,
axi_datawidth=32, silent=False,
filename=None,
simtype='iverilog',
# simtype='verilator',
# simtype=None, # no RTL simulation
outputfile=None):
# input mean and standard deviation
cifar10_mean = np.array([0.4914, 0.4822, 0.4465]).astype(np.float32)
cifar10_std = np.array([0.247, 0.243, 0.261]).astype(np.float32)
act_shape = (1, 32, 32, 3)
# pytorch model
if with_batchnorm:
model = torchvision.models.vgg11_bn(pretrained=False)
else:
model = torchvision.models.vgg11(pretrained=False)
model.features[0].in_channels = act_shape[-1]
model.avgpool = nn.Identity()
#model.classifier[0] = nn.Linear(512, 4096)
#model.classifier[6] = nn.Linear(4096, 10)
model.classifier = nn.Sequential(
nn.Linear(in_features=512, out_features=1024, bias=True),
nn.ReLU(inplace=True),
nn.Dropout(p=0.5),
nn.Linear(in_features=1024, out_features=1024, bias=True),
nn.ReLU(inplace=True),
nn.Dropout(p=0.5),
nn.Linear(in_features=1024, out_features=10, bias=True),
)
# Pytorch to ONNX
onnx_filename = 'vgg11.onnx'
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['out']
model.eval()
torch.onnx.export(model, dummy_input, onnx_filename,
input_names=input_names, output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
dtypes = {}
(outputs, placeholders, variables,
constants, operators) = ng.from_onnx(onnx_filename,
value_dtypes=dtypes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=weight_dtype,
default_constant_dtype=weight_dtype,
default_operator_dtype=act_dtype,
default_scale_dtype=scale_dtype,
default_bias_dtype=bias_dtype,
disable_fusion=disable_fusion)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
if act_dtype.width > 8:
act_scale_factor = 128
else:
act_scale_factor = int(round(2 ** (act_dtype.width - 1) * 0.5))
input_scale_factors = {'act': act_scale_factor}
input_means = {'act': cifar10_mean * act_scale_factor}
input_stds = {'act': cifar10_std * act_scale_factor}
ng.quantize(outputs, input_scale_factors, input_means, input_stds)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.conv2d):
op.attribute(par_ich=conv2d_par_ich,
par_och=conv2d_par_och,
par_col=conv2d_par_col,
par_row=conv2d_par_row,
concur_och=conv2d_concur_och,
stationary=conv2d_stationary)
if isinstance(op, (ng.avg_pool, ng.max_pool,
ng.avg_pool_serial, ng.max_pool_serial)):
op.attribute(par=pool_par)
if ng.is_elementwise_operator(op):
op.attribute(par=elem_par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
act = placeholders['act']
out = outputs['out']
# verification data
# random data
img = np.random.uniform(size=act.length).astype(np.float32).reshape(act.shape)
img = img * 12.0 * cifar10_std + cifar10_mean
# img = np.random.normal(size=act.length).astype(np.float32).reshape(act.shape)
# img = img * cifar10_std + cifar10_mean
# execution on pytorch
model_input = img
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_out = model(torch.from_numpy(model_input)).detach().numpy()
if act.perm is not None and len(model_out.shape) == len(act.shape):
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
# software-based verification
vact = img * act_scale_factor
vact = np.clip(vact,
-1.0 * (2 ** (act.dtype.width - 1) - 1),
1.0 * (2 ** (act.dtype.width - 1) - 1))
vact = np.round(vact).astype(np.int64)
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
labels = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
mout = scaled_model_out
for bat in range(mout.shape[0]):
for index, value in list(sorted(enumerate(mout[bat]),
key=lambda x: x[1], reverse=True))[:10]:
print("# mout: %s (%d) = %f" % (str(labels[index]), index, value))
for index, value in list(sorted(enumerate(vout[bat]),
key=lambda x: x[1], reverse=True))[:10]:
print("# vout: %s (%d) = %d" % (str(labels[index]), index, value))
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
# to Veriloggen object
# targ = ng.to_veriloggen([out], 'vgg11', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# to IP-XACT (the method returns Veriloggen object, as well as to_veriloggen)
targ = ng.to_ipxact([out], 'onnx_vgg11', silent=silent,
config={'maxi_datawidth': axi_datawidth})
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog([out], 'vgg11', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([out], chunk_size)
param_bytes = len(param_data)
variable_addr = int(math.ceil((act.addr + act.memory_size) / chunk_size)) * chunk_size
check_addr = int(math.ceil((variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil((check_addr + out.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
# mem = np.zeros([1024 * 1024 * 256 // (memimg_datawidth // 8)], dtype=np.int64)
mem = np.zeros([1024 * 1024 * 1024 // (memimg_datawidth // 8)], dtype=np.int16)
mem = mem + [100]
# placeholder
axi.set_memory(mem, vact, memimg_datawidth,
act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth,
8, variable_addr)
# verification data
axi.set_memory(mem, vout, memimg_datawidth,
act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if outputfile is None:
outputfile = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + outputfile
memory = axi.AxiMemoryModel(m, 'memory', clk, rst,
datawidth=axi_datawidth,
memimg=mem, memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(
time_counter.inc()
)
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(out.shape[0]):
for x in range(out.shape[1]):
orig = memory.read_word(bat * out.aligned_shape[1] + x,
out.addr, act_dtype.width)
check = memory.read_word(bat * out.aligned_shape[1] + x,
check_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, x,
') orig: ', orig, ' check: ', check)
ok = False
# else:
# print('OK (', bat, x,
# ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ, 'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, resetn, m.make_reset(), period=100, polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if filename is not None:
m.to_verilog(filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=outputfile)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(
act_dtype=ng.int8,
weight_dtype=ng.int8,
bias_dtype=ng.int32,
scale_dtype=ng.int8,
par_ich=2,
par_och=2,
chunk_size=64,
axi_datawidth=32,
silent=False,
weight_filename='cnn.npy',
verilog_filename=None,
sim_filename=None,
# simtype='iverilog',
simtype='verilator',
# simtype=None, # no RTL simulation
):
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# input
input_layer = ng.placeholder(
dtype=act_dtype,
shape=(1, 32, 32, 3), # N, H, W, C
name='input_layer')
# layer 0: conv2d (with bias and scale (= batchnorm)), relu, max_pool
w0 = ng.variable(
dtype=weight_dtype,
shape=(64, 3, 3, 3), # Och, Ky, Kx, Ich
name='w0')
b0 = ng.variable(dtype=bias_dtype, shape=(w0.shape[0], ), name='b0')
s0 = ng.variable(dtype=scale_dtype, shape=(w0.shape[0], ), name='s0')
a0 = ng.conv2d(input_layer,
w0,
strides=(1, 1, 1, 1),
bias=b0,
scale=s0,
act_func=ng.relu,
dtype=act_dtype,
sum_dtype=ng.int32)
a0p = ng.max_pool_serial(a0, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1))
# layer 1: conv2d, relu, reshape
w1 = ng.variable(weight_dtype, shape=(64, 3, 3, a0.shape[-1]), name='w1')
b1 = ng.variable(bias_dtype, shape=(w1.shape[0], ), name='b1')
s1 = ng.variable(scale_dtype, shape=(w1.shape[0], ), name='s1')
a1 = ng.conv2d(a0p,
w1,
strides=(1, 1, 1, 1),
bias=b1,
scale=s1,
act_func=ng.relu,
dtype=act_dtype,
sum_dtype=ng.int32)
a1r = ng.reshape(a1, [1, -1])
# layer 2: full-connection, relu
w2 = ng.variable(weight_dtype, shape=(256, a1r.shape[-1]), name='w2')
b2 = ng.variable(bias_dtype, shape=(w2.shape[0], ), name='b2')
s2 = ng.variable(scale_dtype, shape=(w2.shape[0], ), name='s2')
a2 = ng.matmul(a1r,
w2,
bias=b2,
scale=s2,
transposed_b=True,
act_func=ng.relu,
dtype=act_dtype,
sum_dtype=ng.int32)
# layer 3: full-connection, relu
w3 = ng.variable(weight_dtype, shape=(10, a2.shape[-1]), name='w3')
b3 = ng.variable(bias_dtype, shape=(w3.shape[0], ), name='b3')
s3 = ng.variable(scale_dtype, shape=(w3.shape[0], ), name='s3')
# output
output_layer = ng.matmul(a2,
w3,
bias=b3,
scale=s3,
transposed_b=True,
name='output_layer',
dtype=act_dtype,
sum_dtype=ng.int32)
# --------------------
# (2) Assign weights to the NNgen operators
# --------------------
# In this example, random floating-point values are assigned.
# In a real case, you should assign actual weight values
# obtianed by a training on DNN framework.
# If you don't you NNgen's quantizer, you can assign integer weights to each tensor.
w0_value = np.random.normal(size=w0.length).reshape(w0.shape)
w0_value = np.clip(w0_value, -3.0, 3.0)
w0.set_value(w0_value)
b0_value = np.random.normal(size=b0.length).reshape(b0.shape)
b0_value = np.clip(b0_value, -3.0, 3.0)
b0.set_value(b0_value)
s0_value = np.ones(s0.shape)
s0.set_value(s0_value)
w1_value = np.random.normal(size=w1.length).reshape(w1.shape)
w1_value = np.clip(w1_value, -3.0, 3.0)
w1.set_value(w1_value)
b1_value = np.random.normal(size=b1.length).reshape(b1.shape)
b1_value = np.clip(b1_value, -3.0, 3.0)
b1.set_value(b1_value)
s1_value = np.ones(s1.shape)
s1.set_value(s1_value)
w2_value = np.random.normal(size=w2.length).reshape(w2.shape)
w2_value = np.clip(w2_value, -3.0, 3.0)
w2.set_value(w2_value)
b2_value = np.random.normal(size=b2.length).reshape(b2.shape)
b2_value = np.clip(b2_value, -3.0, 3.0)
b2.set_value(b2_value)
s2_value = np.ones(s2.shape)
s2.set_value(s2_value)
w3_value = np.random.normal(size=w3.length).reshape(w3.shape)
w3_value = np.clip(w3_value, -3.0, 3.0)
w3.set_value(w3_value)
b3_value = np.random.normal(size=b3.length).reshape(b3.shape)
b3_value = np.clip(b3_value, -3.0, 3.0)
b3.set_value(b3_value)
s3_value = np.ones(s3.shape)
s3.set_value(s3_value)
# Quantizing the floating-point weights by the NNgen quantizer.
# Alternatively, you can assign integer weights by yourself to each tensor.
imagenet_mean = np.array([0.485, 0.456, 0.406]).astype(np.float32)
imagenet_std = np.array([0.229, 0.224, 0.225]).astype(np.float32)
if act_dtype.width > 8:
act_scale_factor = 128
else:
act_scale_factor = int(round(2**(act_dtype.width - 1) * 0.5))
input_scale_factors = {'input_layer': act_scale_factor}
input_means = {'input_layer': imagenet_mean * act_scale_factor}
input_stds = {'input_layer': imagenet_std * act_scale_factor}
ng.quantize([output_layer], input_scale_factors, input_means, input_stds)
# --------------------
# (3) Assign hardware attributes
# --------------------
# conv2d, matmul
# par_ich: parallelism in input-channel
# par_och: parallelism in output-channel
# par_col: parallelism in pixel column
# par_row: parallelism in pixel row
a0.attribute(par_ich=par_ich, par_och=par_och)
a1.attribute(par_ich=par_ich, par_och=par_och)
a2.attribute(par_ich=par_ich, par_och=par_och)
output_layer.attribute(par_ich=par_ich, par_och=par_och)
# cshamt_out: right shift amount after applying bias/scale
# If you assign integer weights by yourself to each tensor,
# cshamt (constant shift amount) must be assigned to each operator.
# a0.attribute(cshamt_out=weight_dtype.width + 1)
# a1.attribute(cshamt_out=weight_dtype.width + 1)
# a2.attribute(cshamt_out=weight_dtype.width + 1)
# output_layer.attribute(cshamt_out=weight_dtype.width + 1)
# max_pool
# par: parallelism in in/out channel
par = par_och
a0p.attribute(par=par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
# In this example, random integer values are assigned.
# In real case, you should assign actual integer activation values, such as an image.
input_layer_value = np.random.normal(size=input_layer.length).reshape(
input_layer.shape)
input_layer_value = input_layer_value * imagenet_std + imagenet_mean
input_layer_value = np.clip(input_layer_value, -5.0, 5.0)
input_layer_value = input_layer_value * act_scale_factor
input_layer_value = np.clip(input_layer_value,
-1 * 2**(act_dtype.width - 1) - 1,
2**(act_dtype.width - 1))
input_layer_value = np.round(input_layer_value).astype(np.int64)
eval_outs = ng.eval([output_layer], input_layer=input_layer_value)
output_layer_value = eval_outs[0]
# print(output_layer_value)
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
# to Veriloggen object
# targ = ng.to_veriloggen([output_layer], 'cnn', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# to IP-XACT (the method returns Veriloggen object, as well as to_veriloggen)
targ = ng.to_ipxact([output_layer],
'cnn',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog([output_layer], 'cnn', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Save the quantized weights
# --------------------
# convert weight values to a memory image:
# on a real FPGA platform, this image will be used as a part of the model definition.
param_filename = 'hello_nngen.npy'
chunk_size = 64
param_data = ng.export_ndarray([output_layer], chunk_size)
np.save(weight_filename, param_data)
# --------------------
# (7) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
param_bytes = len(param_data)
variable_addr = int(
math.ceil((input_layer.addr + input_layer.memory_size) /
chunk_size)) * chunk_size
check_addr = int(math.ceil(
(variable_addr + param_bytes) / chunk_size)) * chunk_size
tmp_addr = int(
math.ceil(
(check_addr + output_layer.memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 256 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
# placeholder
axi.set_memory(
mem, input_layer_value, memimg_datawidth, act_dtype.width,
input_layer.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(
mem, output_layer_value, memimg_datawidth, act_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if sim_filename is None:
sim_filename = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + sim_filename
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(output_layer.shape[0]):
for x in range(output_layer.shape[1]):
orig = memory.read_word(
bat * output_layer.aligned_shape[1] + x, output_layer.addr,
act_dtype.width)
check = memory.read_word(
bat * output_layer.aligned_shape[1] + x, check_addr,
act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, x, ') orig: ', orig, ' check: ', check)
ok = False
else:
print('OK (', bat, x, ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if verilog_filename is not None:
m.to_verilog(verilog_filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=sim_filename)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
def run(
act_dtype=ng.int16,
weight_dtype=ng.int8,
bias_dtype=ng.int32,
scale_dtype=ng.int8,
disable_fusion=False,
conv2d_par_ich=1,
conv2d_par_och=1,
conv2d_par_col=1,
conv2d_par_row=1,
conv2d_concur_och=None,
conv2d_stationary='filter',
pool_par=1,
elem_par=1,
chunk_size=64,
axi_datawidth=32,
silent=False,
onnx_filename='yolov3-tiny.onnx',
weight_filename='yolov3-tiny.npy',
verilog_filename=None,
sim_filename=None,
# simtype=None, # no RTL simulation
# simtype='iverilog',
simtype='verilator',
cfg_filename='yolov3-tiny.cfg',
weights_filename='yolov3-tiny.weights',
model_path='yolov3'):
# input mean and standard deviation
imagenet_mean = np.array([0.485, 0.456, 0.406]).astype(np.float32)
imagenet_std = np.array([0.229, 0.224, 0.225]).astype(np.float32)
img_size = (416, 416)
act_shape = (1, img_size[0], img_size[1], 3)
# pytorch model
model_url = "https://github.com/ultralytics/yolov3"
if not os.path.isdir(model_path):
raise FileNotFoundError(
"Download the YOLOv3 model using Pytorch, such as "
"'%s'. Then extract it, and rename it as '%s'" %
(model_url, model_path))
# Darknet model configuration and pretrained weights
cfg_url = "https://github.com/pjreddie/darknet/blob/master/cfg/yolov3-tiny.cfg"
if not os.path.isfile(cfg_filename):
urllib.request.urlretrieve(cfg_url, cfg_filename)
weights_url = "https://pjreddie.com/media/files/yolov3-tiny.weights"
if not os.path.isfile(weights_filename):
urllib.request.urlretrieve(weights_url, weights_filename)
sys.path.insert(0, model_path)
import models
models.ONNX_EXPORT = True
model = models.Darknet(cfg_filename, img_size).to('cpu')
models.load_darknet_weights(model, weights_filename)
# Pytorch to ONNX
dummy_input = torch.randn(*act_shape).transpose(1, 3)
input_names = ['act']
output_names = ['scores', 'boxes']
model.eval()
torch.onnx.export(model,
dummy_input,
onnx_filename,
input_names=input_names,
output_names=output_names)
# --------------------
# (1) Represent a DNN model as a dataflow by NNgen operators
# --------------------
# ONNX to NNgen
dtypes = {}
shapes = {}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=dtypes,
value_shapes=shapes,
default_placeholder_dtype=act_dtype,
default_variable_dtype=weight_dtype,
default_constant_dtype=weight_dtype,
default_operator_dtype=act_dtype,
default_scale_dtype=scale_dtype,
default_bias_dtype=bias_dtype,
disable_fusion=disable_fusion,
verbose=False)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
if act_dtype.width > 8:
act_scale_factor = 128
else:
act_scale_factor = int(round(2**(act_dtype.width - 1) * 0.5))
input_scale_factors = {'act': act_scale_factor}
input_means = {'act': imagenet_mean * act_scale_factor}
input_stds = {'act': imagenet_std * act_scale_factor}
ng.quantize(outputs, input_scale_factors, input_means, input_stds)
# --------------------
# (3) Assign hardware attributes
# --------------------
for op in operators.values():
if isinstance(op, ng.conv2d):
op.attribute(par_ich=conv2d_par_ich,
par_och=conv2d_par_och,
par_col=conv2d_par_col,
par_row=conv2d_par_row,
concur_och=conv2d_concur_och,
stationary=conv2d_stationary)
if isinstance(op, (ng.avg_pool, ng.max_pool, ng.avg_pool_serial,
ng.max_pool_serial)):
op.attribute(par=pool_par)
if ng.is_elementwise_operator(op):
op.attribute(par=elem_par)
# --------------------
# (4) Verify the DNN model behavior by executing the NNgen dataflow as a software
# --------------------
act = placeholders['act']
outs = (outputs['scores'], outputs['boxes'])
# verification data
img = np.array(PIL.Image.open('car416x416.png').convert('RGB')).astype(
np.float32)
img = img.reshape([1] + list(img.shape))
img = img / 255
img = (img - imagenet_mean) / imagenet_std
# execution on pytorch
model_input = img
if act.perm is not None:
model_input = np.transpose(model_input, act.reversed_perm)
model.eval()
model_rslts = model(torch.from_numpy(model_input))
model_outs = [rslt.detach().numpy() for rslt in model_rslts]
model_outs = [(np.transpose(model_out, act.perm) if act.perm is not None
and len(model_out.shape) == len(act.shape) else model_out)
for model_out in model_outs]
scaled_model_outs = [
model_out * out.scale_factor
for model_out, out in zip(model_outs, outs)
]
# software-based verification
vact = img * act_scale_factor
vact = np.clip(vact, -1.0 * (2**(act.dtype.width - 1) - 1),
1.0 * (2**(act.dtype.width - 1) - 1))
vact = np.round(vact).astype(np.int64)
# compare outputs of hidden layers
leaky_relu_ops = [
v for k, v in operators.items()
if (isinstance(v, ng.conv2d)
and isinstance(v.act_func, ng.leaky_relu_base))
]
leaky_relu_ops = list(sorted(set(leaky_relu_ops),
key=leaky_relu_ops.index))
conv2d_ops = [
v for k, v in operators.items()
if (isinstance(v, ng.conv2d) and v.act_func is None)
]
conv2d_ops = list(sorted(set(conv2d_ops), key=conv2d_ops.index))
# only 1st output
sub_ops = leaky_relu_ops[:9] + conv2d_ops[:1]
sub_outs = ng.eval(sub_ops, act=vact)
sub_outs = [sub_out.transpose([0, 3, 1, 2]) for sub_out in sub_outs]
sub_scale_factors = [sub_op.scale_factor for sub_op in sub_ops]
model.eval()
mouts = []
# all Conv2d-LeakyReLU layers before YOLOLayer
mouts.append(
nn.Sequential(model.module_list[0])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:3])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:5])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:7])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:9])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:11])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:13])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:14])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:15])(
torch.from_numpy(model_input)).detach().numpy())
mouts.append(
nn.Sequential(*model.module_list[0:16])(
torch.from_numpy(model_input)).detach().numpy())
scaled_mouts = [
mout * scale_factor
for mout, scale_factor in zip(mouts, sub_scale_factors)
]
sub_mean_square_errors = [
np.sum((sub_out - mout)**2) / sub_out.size
for mout, sub_out in zip(scaled_mouts, sub_outs)
]
sub_corrcoefs = [
np.corrcoef(mout.reshape([-1]), sub_out.reshape([-1]))
for mout, sub_out in zip(mouts, sub_outs)
]
# compare prediction results
vouts = ng.eval(outs, act=vact)
mean_square_errors = [
np.sum((vout - scaled_model_out)**2) / vout.size
for vout, scaled_model_out in zip(vouts, scaled_model_outs)
]
corrcoefs = [
np.corrcoef(model_out.reshape([-1]), vout.reshape([-1]))
for model_out, vout in zip(model_outs, vouts)
]
# breakpoint()
# --------------------
# (5) Convert the NNgen dataflow to a hardware description (Verilog HDL and IP-XACT)
# --------------------
# to Veriloggen object
# targ = ng.to_veriloggen(outs, 'yolov3tiny', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# to IP-XACT (the method returns Veriloggen object, as well as to_veriloggen)
targ = ng.to_ipxact(outs,
'yolov3tiny',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog(outs, 'yolov3tiny', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Save the quantized weights
# --------------------
param_data = ng.export_ndarray(outs, chunk_size)
param_bytes = len(param_data)
np.save(weight_filename, param_data)
# --------------------
# (7) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
variable_addr = int(math.ceil(
(act.addr + act.memory_size) / chunk_size)) * chunk_size
check0_addr = int(math.ceil(
(variable_addr + param_bytes) / chunk_size)) * chunk_size
check1_addr = int(
math.ceil(
(check0_addr + outs[0].memory_size) / chunk_size)) * chunk_size
tmp_addr = int(math.ceil(
(check1_addr + outs[1].memory_size) / chunk_size)) * chunk_size
memimg_datawidth = 32
# mem = np.zeros([1024 * 1024 * 256 // (memimg_datawidth // 8)], dtype=np.int64)
mem = np.zeros([1024 * 1024 * 256 // (memimg_datawidth // 8)],
dtype=np.int16)
mem = mem + [100]
# placeholder
axi.set_memory(
mem, vact, memimg_datawidth, act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_ich))
# parameters (variable and constant)
axi.set_memory(mem, param_data, memimg_datawidth, 8, variable_addr)
# verification data
axi.set_memory(
mem, vouts[0], memimg_datawidth, act_dtype.width, check0_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_och))
axi.set_memory(
mem, vouts[1], memimg_datawidth, act_dtype.width, check1_addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), conv2d_par_och))
# test controller
m = Module('test')
params = m.copy_params(targ)
ports = m.copy_sim_ports(targ)
clk = ports['CLK']
resetn = ports['RESETN']
rst = m.Wire('RST')
rst.assign(Not(resetn))
# AXI memory model
if sim_filename is None:
sim_filename = os.path.splitext(os.path.basename(__file__))[0] + '.out'
memimg_name = 'memimg_' + sim_filename
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# timer
time_counter = m.Reg('time_counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(time_counter.inc())
def ctrl():
for i in range(100):
pass
ng.sim.set_global_addrs(_saxi, tmp_addr)
start_time = time_counter.value
ng.sim.start(_saxi)
print('# start')
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok = True
for bat in range(outs[0].shape[0]):
for x in range(outs[0].shape[1]):
orig = memory.read_word(bat * outs[0].aligned_shape[1] + x,
outs[0].addr, act_dtype.width)
check = memory.read_word(bat * outs[0].aligned_shape[1] + x,
check0_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, x, ') orig: ', orig, ' check: ', check)
ok = False
# else:
# print('OK (', bat, x,
# ') orig: ', orig, ' check: ', check)
for bat in range(outs[1].shape[0]):
for x in range(outs[1].shape[1]):
orig = memory.read_word(bat * outs[1].aligned_shape[1] + x,
outs[1].addr, act_dtype.width)
check = memory.read_word(bat * outs[1].aligned_shape[1] + x,
check1_addr, act_dtype.width)
if vthread.verilog.NotEql(orig, check):
print('NG (', bat, x, ') orig: ', orig, ' check: ', check)
ok = False
# else:
# print('OK (', bat, x,
# ') orig: ', orig, ' check: ', check)
if ok:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(targ,
'uut',
params=m.connect_params(targ),
ports=m.connect_ports(targ))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m,
resetn,
m.make_reset(),
period=100,
polarity='low')
init.add(
Delay(10000000),
Systask('finish'),
)
# output source code
if verilog_filename is not None:
m.to_verilog(verilog_filename)
# run simulation
sim = simulation.Simulator(m, sim=simtype)
rslt = sim.run(outputfile=sim_filename)
lines = rslt.splitlines()
if simtype == 'verilator' and lines[-1].startswith('-'):
rslt = '\n'.join(lines[:-1])
return rslt
