def run(act_shape=(1, 7, 7, 3),
weight_shape=(9, 3, 3, 3),
act_dtype=ng.int32,
weight_dtype=ng.int32,
out_dtype=ng.int32,
stride=1,
padding=0,
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
layers = []
layers.append(
nn.Conv2d(weight_shape[3],
weight_shape[0],
weight_shape[1],
stride=stride,
padding=padding,
bias=not with_batchnorm))
if with_batchnorm:
layers.append(nn.BatchNorm2d(weight_shape[0]))
if act_func == 'relu':
layers.append(nn.ReLU(inplace=True))
elif act_func == 'leaky_relu':
layers.append(nn.LeakyReLU(inplace=True))
model = nn.Sequential(*layers)
# overwrite weight values for test
#model[0].weight.data = torch.from_numpy(np.ones_like(model[0].weight.data.numpy()))
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_conv2d.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=out_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.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.make_param_array(variables, constants, 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], 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, out_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / out_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,
out_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,
out_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_shape=(1, 7, 7, 15),
weight1_shape=(7, 3, 3, 15), bias1_shape=None, scale1_shape=None,
weight2_shape=(9, 3, 3, 7), bias2_shape=None, scale2_shape=None,
act_dtype=ng.int32,
weight1_dtype=ng.int32, bias1_dtype=ng.int32, scale1_dtype=ng.int32,
weight2_dtype=ng.int32, bias2_dtype=ng.int32, scale2_dtype=ng.int32,
tmp_dtype=ng.int32,
out_dtype=ng.int32,
stride1=(1, 1, 1, 1), stride2=(1, 1, 1, 1),
rshift_mul1=None, rshift_sum1=None, rshift_out1=None,
rshift_mul2=None, rshift_sum2=None, rshift_out2=None,
act_func1=None, act_func2=None,
par_ich1=1, par_och1=1, par_col1=1, par_row1=1,
concur_och1=None, stationary1='filter',
par_ich2=1, par_och2=1, par_col2=1, par_row2=1,
concur_och2=None, stationary2='filter',
input_ram_size1=None, filter_ram_size1=None,
bias_ram_size1=None, scale_ram_size1=None,
out_ram_size1=None,
input_ram_size2=None, filter_ram_size2=None,
bias_ram_size2=None, scale_ram_size2=None,
out_ram_size2=None,
chunk_size=64,
axi_datawidth=32, silent=False,
filename=None, simtype='iverilog', outputfile=None):
# create target hardware
act = ng.placeholder(act_dtype, shape=act_shape, name='act')
weight1 = ng.variable(weight1_dtype, shape=weight1_shape,
name='weight1')
if bias1_shape is not None:
bias1 = ng.variable(bias1_dtype, bias1_shape, name='bias1')
else:
bias1 = None
if scale1_shape is not None:
scale1 = ng.variable(scale1_dtype, scale1_shape, name='scale1')
else:
scale1 = None
weight2 = ng.variable(weight2_dtype, shape=weight2_shape,
name='weight2')
if bias2_shape is not None:
bias2 = ng.variable(bias2_dtype, bias2_shape, name='bias2')
else:
bias2 = None
if scale2_shape is not None:
scale2 = ng.variable(scale2_dtype, scale2_shape, name='scale2')
else:
scale2 = None
tmp = ng.conv2d(act, weight1, stride1,
bias1, scale1,
rshift_mul1, rshift_sum1, rshift_out1,
act_func1, 'SAME', False,
tmp_dtype, ng.int32, ng.int32,
'conv2d_1',
par_ich1, par_och1, par_col1, par_row1,
concur_och1, stationary1,
input_ram_size1, filter_ram_size1,
bias_ram_size1, scale_ram_size1,
None, None, None,
out_ram_size1)
out = ng.conv2d(tmp, weight2, stride2,
bias2, scale2,
rshift_mul2, rshift_sum2, rshift_out2,
act_func2, 'SAME', False,
out_dtype, ng.int32, ng.int32,
'conv2d_2',
par_ich2, par_och2, par_col2, par_row2,
concur_och2, stationary2,
input_ram_size2, filter_ram_size2,
bias_ram_size2, scale_ram_size2,
None, None, None,
out_ram_size2)
targ = ng.to_veriloggen([out], 'matrix_conv2d_conv2d_variable', silent=silent,
config={'maxi_datawidth': axi_datawidth,
'offchipram_chunk_bytes': chunk_size})
# verification data
vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [16]
vweight1 = np.arange(weight1.length,
dtype=np.int64).reshape(weight1_shape) % [32] - [16]
weight1.set_value(vweight1)
if bias1 is not None:
vbias1 = np.arange(bias1.length,
dtype=np.int64).reshape(bias1.shape) % [16]
bias1.set_value(vbias1)
else:
vbias1 = None
if scale1 is not None:
vscale1 = np.arange(scale1.length,
dtype=np.int64).reshape(scale1.shape) % [8]
scale1.set_value(vscale1)
else:
vscale1 = None
vweight2 = np.arange(weight2.length,
dtype=np.int64).reshape(weight2_shape) % [32] - [16]
weight2.set_value(vweight2)
if bias2 is not None:
vbias2 = np.arange(bias2.length,
dtype=np.int64).reshape(bias2.shape) % [16]
bias2.set_value(vbias2)
else:
vbias2 = None
if scale2 is not None:
vscale2 = np.arange(scale2.length,
dtype=np.int64).reshape(scale2.shape) % [8]
scale2.set_value(vscale2)
else:
vscale2 = None
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
# to memory image
size_max = int(math.ceil(max(act.memory_size, weight1.memory_size,
bias1.memory_size if bias1 is not None else 0,
scale1.memory_size if scale1 is not None else 0,
weight2.memory_size,
bias2.memory_size if bias2 is not None else 0,
scale2.memory_size if scale2 is not None else 0,
out.memory_size) / chunk_size)) * chunk_size
# assign custom addresses
variable_addr = max(act.addr, out.addr) + size_max
weight1_addr = variable_addr
bias1_addr = weight1_addr + int(math.ceil(weight1.memory_size / chunk_size)) * chunk_size
scale1_addr = (bias1_addr + int(math.ceil(bias1.memory_size / chunk_size)) * chunk_size
if bias1 is not None else weight1_addr)
weight2_addr = (scale1_addr + int(math.ceil(scale1.memory_size / chunk_size)) * chunk_size
if scale1 is not None else bias1_addr)
bias2_addr = weight2_addr + int(math.ceil(weight2.memory_size / chunk_size)) * chunk_size
scale2_addr = (bias2_addr + int(math.ceil(bias2.memory_size / chunk_size)) * chunk_size
if bias2 is not None else weight2_addr)
check_addr = scale2_addr + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)], dtype=np.int64)
mem = mem + [100]
axi.set_memory(mem, vact, memimg_datawidth,
act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_ich1))
axi.set_memory(mem, vweight1, memimg_datawidth,
weight1_dtype.width, weight1_addr,
max(int(math.ceil(axi_datawidth / weight1_dtype.width)), par_ich1))
if bias1_shape is not None:
axi.set_memory(mem, vbias1, memimg_datawidth,
bias1_dtype.width, bias1_addr,
max(int(math.ceil(axi_datawidth / bias1_dtype.width)), par_och1))
if scale1_shape is not None:
axi.set_memory(mem, vscale1, memimg_datawidth,
scale1_dtype.width, scale1_addr,
max(int(math.ceil(axi_datawidth / scale1_dtype.width)), par_och1))
axi.set_memory(mem, vweight2, memimg_datawidth,
weight2_dtype.width, weight2_addr,
max(int(math.ceil(axi_datawidth / weight2_dtype.width)), par_ich2))
if bias2_shape is not None:
axi.set_memory(mem, vbias2, memimg_datawidth,
bias2_dtype.width, bias2_addr,
max(int(math.ceil(axi_datawidth / bias2_dtype.width)), par_och2))
if scale2_shape is not None:
axi.set_memory(mem, vscale2, memimg_datawidth,
scale2_dtype.width, scale2_addr,
max(int(math.ceil(axi_datawidth / scale2_dtype.width)), par_och2))
axi.set_memory(mem, vout, memimg_datawidth,
out_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / out_dtype.width)), par_och2))
# 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
# set custom addresses
ng.sim.set_global_addrs(_saxi, tmp_addr, out.addr, act.addr, variable_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, out_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, out_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.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
def run(act_shape=(1, 7, 7, 7),
weight_shape=(3, 3, 3, 7),
bias_shape=None,
scale_shape=None,
act_dtype=ng.int32,
weight_dtype=ng.int4,
bias_dtype=ng.int32,
scale_dtype=ng.int32,
out_dtype=ng.int32,
stride=(1, 1, 1, 1),
rshift_mul=None,
rshift_sum=None,
rshift_out=0,
act_func=None,
par_ich=1,
par_och=1,
par_col=1,
par_row=1,
concur_och=None,
stationary='filter',
input_ram_size=None,
filter_ram_size=None,
bias_ram_size=None,
scale_ram_size=None,
out_ram_size=None,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# create target hardware
act = ng.placeholder(act_dtype, shape=act_shape, name='act')
weight = ng.variable(weight_dtype, shape=weight_shape, name='weight')
if bias_shape is not None:
bias = ng.variable(bias_dtype, bias_shape, name='bias')
else:
bias = None
if scale_shape is not None:
scale = ng.variable(scale_dtype, scale_shape, name='scale')
else:
scale = None
out = ng.log_weight_conv2d(act, weight, stride, bias, scale, rshift_mul,
rshift_sum, rshift_out, act_func, 'SAME',
out_dtype, ng.int32, ng.int32,
'log_weight_conv2d', par_ich, par_och, par_col,
par_row, concur_och, stationary, input_ram_size,
filter_ram_size, bias_ram_size, scale_ram_size,
None, None, None, out_ram_size)
targ = ng.to_veriloggen([out],
'matrix_log_weight_conv2d',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [5]
vweight = np.arange(weight.length, dtype=np.int64).reshape(
weight.shape) % [7] - [3]
if bias is not None:
vbias = np.arange(bias.length, dtype=np.int64).reshape(
bias.shape) % [4]
else:
vbias = None
if scale is not None:
vscale = np.arange(scale.length, dtype=np.int64).reshape(
scale.shape) % [6]
else:
vscale = None
vout = ng.verify.log_weight_conv2d(
vact, vweight, stride, vbias, vscale, rshift_mul, rshift_sum,
rshift_out, act_func, 'SAME', out_dtype, ng.int32, ng.int32, 'conv2d',
par_ich, par_och, par_col, par_row, concur_och, stationary,
input_ram_size, filter_ram_size, bias_ram_size, scale_ram_size, None,
None, None, out_ram_size, False, act_dtype, weight_dtype)
# to memory image
size_max = int(
math.ceil(
max(act.memory_size, weight.memory_size,
bias.memory_size if bias is not None else 0, scale.memory_size
if scale is not None else 0, out.memory_size) / 4096)) * 4096
check_addr = max(
act.addr, weight.addr, bias.addr if bias is not None else -1,
scale.addr if scale is not None else -1, out.addr) + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // memimg_datawidth], dtype=np.int64)
mem = mem + [100]
axi.set_memory(
mem, vact, memimg_datawidth, act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par_ich))
axi.set_memory(
mem, vweight, memimg_datawidth, weight_dtype.width, weight.addr,
max(int(math.ceil(axi_datawidth / weight_dtype.width)), par_ich))
if bias is not None:
axi.set_memory(
mem, vbias, memimg_datawidth, bias_dtype.width, bias.addr,
max(int(math.ceil(axi_datawidth / bias_dtype.width)), par_och))
if scale is not None:
axi.set_memory(
mem, vscale, memimg_datawidth, scale_dtype.width, scale.addr,
max(int(math.ceil(axi_datawidth / scale_dtype.width)), par_och))
axi.set_memory(
mem, vout, memimg_datawidth, out_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / out_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,
out_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,
out_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_shape=(15, 31),
act_dtype=ng.int32,
par=1,
chunk_size=64,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# pytorch model
model = MatrixExp()
# Pytorch to ONNX
onnx_filename = 'onnx_matrix_exp.onnx'
dummy_input = torch.randn(*act_shape)
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 = {'act': act_dtype, 'out': act_dtype}
(outputs, placeholders, variables, constants,
operators) = ng.from_onnx(onnx_filename,
value_dtypes=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
# --------------------
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.exp):
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
# random data
#std = 0.2
#mean = 0.5
std = 0.2
mean = 0.0
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_exp',
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])
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())
num_rep = functools.reduce(lambda x, y: x * y, out.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(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, 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(a_shape=(15, 15),
b_shape=(15, 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):
# create target hardware
a = ng.placeholder(a_dtype, shape=a_shape, name='a')
b = ng.placeholder(b_dtype, shape=b_shape, name='b')
d = ng.add(a, b, dtype=c_dtype, par=par)
e = ng.add(b, a, dtype=c_dtype, par=par)
# SW returns ng.add(x, y)
f = ng.extern([d, e], shape=a_shape, opcode=0x1, func=lambda x, y: x + y)
g = ng.sub(f, a)
# SW returns d as-is
h = ng.extern([g], shape=a_shape, opcode=0x2, func=lambda x: x)
c = ng.sub(h, b)
targ = ng.to_veriloggen([c],
'matrix_extern',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
va = np.arange(a.length, dtype=np.int64).reshape(a.shape) % [16]
vb = np.arange(b.length, dtype=np.int64).reshape(b.shape) % [32] + [16]
eval_outs = ng.eval([c], a=va, b=vb)
vc = eval_outs[0]
# to memory image
size_max = int(
math.ceil(
max(a.memory_size, b.memory_size, c.memory_size) / 4096)) * 4096
check_addr = max(a.addr, b.addr, c.addr) + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
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))
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_datawidth=memimg_datawidth,
memimg=mem,
memimg_name=memimg_name)
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('# 1st test start')
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - d.default_global_addr
y_offset = tmp_addr - e.default_global_addr
z_offset = tmp_addr - f.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
d.addr + x_offset, c_dtype.width)
y = memory.read_word(i * c.aligned_shape[-1] + j,
e.addr + y_offset, c_dtype.width)
z = x + y
memory.write_word(i * c.aligned_shape[-1] + j,
f.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
# software reset
ng.sim.sw_rst(_saxi)
print('# 1st software reset (before resume)')
# restart
ng.sim.start(_saxi)
print('# Restart')
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - d.default_global_addr
y_offset = tmp_addr - e.default_global_addr
z_offset = tmp_addr - f.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
d.addr + x_offset, c_dtype.width)
y = memory.read_word(i * c.aligned_shape[-1] + j,
e.addr + y_offset, c_dtype.width)
z = x + y
memory.write_word(i * c.aligned_shape[-1] + j,
f.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - g.default_global_addr
z_offset = tmp_addr - h.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
g.addr + x_offset, c_dtype.width)
z = x
memory.write_word(i * c.aligned_shape[-1] + j,
h.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok_1st = 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_1st = False
# else:
# print('OK', i, j, orig, check)
# 2nd test
# start
start_time = time_counter.value
ng.sim.start(_saxi)
print('# 2nd test start')
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - d.default_global_addr
y_offset = tmp_addr - e.default_global_addr
z_offset = tmp_addr - f.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
d.addr + x_offset, c_dtype.width)
y = memory.read_word(i * c.aligned_shape[-1] + j,
e.addr + y_offset, c_dtype.width)
z = x + y
memory.write_word(i * c.aligned_shape[-1] + j,
f.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
while (memory.waddr.awvalid) == 0:
pass
ng.sim.sw_rst(_saxi)
print('# 2nd software reset (during Master AXI transaction)')
# restart
ng.sim.start(_saxi)
print('# Restart')
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - d.default_global_addr
y_offset = tmp_addr - e.default_global_addr
z_offset = tmp_addr - f.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
d.addr + x_offset, c_dtype.width)
y = memory.read_word(i * c.aligned_shape[-1] + j,
e.addr + y_offset, c_dtype.width)
z = x + y
memory.write_word(i * c.aligned_shape[-1] + j,
f.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
# from extern-send
araddr = ng.control_reg_extern_send * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
print('# opcode = %d' % v)
for i in range(num_rep):
for j in range(c.shape[-1]):
x_offset = tmp_addr - g.default_global_addr
z_offset = tmp_addr - h.default_global_addr
x = memory.read_word(i * c.aligned_shape[-1] + j,
g.addr + x_offset, c_dtype.width)
z = x
memory.write_word(i * c.aligned_shape[-1] + j,
h.addr + z_offset, z, c_dtype.width)
# to extern-recv
awaddr = ng.control_reg_extern_recv * 4
_saxi.write(awaddr, 1)
ng.sim.wait(_saxi)
end_time = time_counter.value
print('# end')
print('# execution cycles: %d' % (end_time - start_time))
# verify
ok_2nd = 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_2nd = False
# else:
# print('OK', i, j, orig, check)
if ok_1st and ok_2nd:
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_dtype=ng.int16,
weight_dtype=ng.int16,
par_ich=2,
par_och=2,
chunk_size=64,
axi_datawidth=32,
silent=False,
weight_filename='mlp.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(act_dtype, shape=(1, 784), name='input_layer')
#w0 = ng.variable(weight_dtype, shape=(1024, 784), name='w0')
w0 = ng.variable(weight_dtype, shape=(128, 784), name='w0')
a0 = ng.matmul(input_layer,
w0,
transposed_b=True,
act_func=ng.relu,
sum_dtype=ng.dtype_int(64))
#w1 = ng.variable(weight_dtype, shape=(10, 1024), name='w1')
w1 = ng.variable(weight_dtype, shape=(10, 128), name='w1')
output_layer = ng.matmul(a0,
w1,
transposed_b=True,
name='output_layer',
sum_dtype=ng.dtype_int(64))
# --------------------
# (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)
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)
# Quantizing the floating-point weights by the NNgen quantizer.
# Alternatively, you can assign integer weights by yourself to each tensor.
img_mean = np.array([0.485]).astype(np.float32)
img_std = np.array([0.229]).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': img_mean * act_scale_factor}
input_stds = {'input_layer': img_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
# cshamt_out: right shift amount after applying bias/scale
a0.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)
# output_layer.attribute(cshamt_out=weight_dtype.width + 1)
# --------------------
# (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 * img_std + img_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], 'mlp', 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],
'mlp',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# to Verilog HDL RTL (the method returns a source code text)
# rtl = ng.to_verilog([output_layer], 'mlp', silent=silent,
# config={'maxi_datawidth': axi_datawidth})
# --------------------
# (6) Save the quantized weights
# --------------------
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 mkTest():
m = Module('test')
# target instance
led = mkLed()
# copy paras and ports
params = m.copy_params(led)
ports = m.copy_sim_ports(led)
clk = ports['CLK']
rst = ports['RST']
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# Timer
counter = m.Reg('counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(counter.inc())
def ctrl():
for i in range(100):
pass
awaddr = 4
sleep = 16
print('# sleep = %d' % sleep)
_saxi.write(awaddr, sleep)
awaddr = 8
size = 128
print('# size = %d' % size)
_saxi.write(awaddr, size)
awaddr = 0
start_time = counter
print('# start time = %d' % start_time)
_saxi.write(awaddr, 1)
araddr = 12
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
end_time = counter
print('# end time = %d' % end_time)
time = end_time - start_time
print('# exec time = %d' % time)
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(led,
'uut',
params=m.connect_params(led),
ports=m.connect_ports(led))
simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(100000),
Systask('finish'),
)
return m
def mkTest(memimg_name=None):
a_shape = (matrix_size, matrix_size)
b_shape = (matrix_size, matrix_size)
c_shape = (a_shape[0], b_shape[0])
n_raw_a = axi.shape_to_length(a_shape)
n_raw_b = axi.shape_to_length(b_shape)
n_a = axi.shape_to_memory_size(a_shape, datawidth)
n_b = axi.shape_to_memory_size(b_shape, datawidth)
a = np.zeros(a_shape, dtype=np.int64)
b = np.zeros(b_shape, dtype=np.int64)
value = 1
for y in range(a_shape[0]):
for x in range(a_shape[1]):
if x == y:
a[y][x] = value
value += 1
else:
a[y][x] = 0
for y in range(b_shape[0]):
for x in range(b_shape[1]):
if x == y:
b[y][x] = 2
else:
b[y][x] = 0
a_addr = a_offset
size_a = n_a * datawidth // 8
b_addr = b_offset
size_b = n_b * datawidth // 8
mem = np.zeros([1024 * 1024 * 8 // axi_datawidth], dtype=np.int64)
axi.set_memory(mem, a, axi_datawidth, datawidth, a_addr)
axi.set_memory(mem, b, axi_datawidth, datawidth, b_addr)
led = mkLed()
m = Module('test')
params = m.copy_params(led)
ports = m.copy_sim_ports(led)
clk = ports['CLK']
rst = ports['RST']
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
mem_datawidth=axi_datawidth,
memimg=mem,
memimg_name=memimg_name)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# Timer
counter = m.Reg('counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(counter.inc())
def ctrl():
for i in range(100):
pass
awaddr = 4
print('# matrix_size = %d' % matrix_size)
_saxi.write(awaddr, matrix_size)
awaddr = 8
print('# a_offset = %d' % a_offset)
_saxi.write(awaddr, a_offset)
awaddr = 12
print('# b_offset = %d' % b_offset)
_saxi.write(awaddr, b_offset)
awaddr = 16
print('# c_offset = %d' % c_offset)
_saxi.write(awaddr, c_offset)
awaddr = 0
start_time = counter
print('# start time = %d' % start_time)
_saxi.write(awaddr, 1)
araddr = 20
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
end_time = counter
print('# end time = %d' % end_time)
time = end_time - start_time
print('# exec time = %d' % time)
all_ok = True
for y in range(matrix_size):
for x in range(matrix_size):
v = memory.read(c_offset +
(y * matrix_size + x) * datawidth // 8)
if y == x and vthread.verilog.NotEql(v, (y + 1) * 2):
all_ok = False
print("NG [%d,%d] = %d" % (y, x, v))
if y != x and vthread.verilog.NotEql(v, 0):
all_ok = False
print("NG [%d,%d] = %d" % (y, x, v))
if all_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(led,
'uut',
params=m.connect_params(led),
ports=m.connect_ports(led))
simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(1000000),
Systask('finish'),
)
return m
def mkTest(memname='mymem.out'):
m = Module('test')
matrix_size = 16
# target instance
led = mkLed()
# copy paras and ports
params = m.copy_params(led)
ports = m.copy_sim_ports(led)
clk = ports['CLK']
rst = ports['RST']
# memory image
#memname = 'mymem.out'
def fwrite(f, value):
s = '%08x' % value
f.write('%s\n' % s[6:8])
f.write('%s\n' % s[4:6])
f.write('%s\n' % s[2:4])
f.write('%s\n' % s[0:2])
with open(memname, 'w') as f:
# ram_a
addr = 0
nv = 1
for x in range(matrix_size):
for y in range(matrix_size):
addr += 4
if x == y:
value = nv
nv += 1
else:
value = 0
fwrite(f, value)
for i in range(1024 - addr):
f.write('%s\n' % '00')
# ram_b
addr = 1024
for x in range(matrix_size):
for y in range(matrix_size):
addr += 4
if x == y:
value = 2
else:
value = 0
fwrite(f, value)
for i in range(2048 - addr):
f.write('%s\n' % '00')
# ram_c
addr = 2048
for x in range(matrix_size):
for y in range(matrix_size):
addr += 4
value = 100
fwrite(f, value)
for i in range(2**20 - addr):
f.write('%s\n' % '00')
memory = axi.AxiMemoryModel(m, 'memory', clk, rst, memimg=memname)
memory.connect(ports, 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
# Timer
counter = m.Reg('counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(counter.inc())
def ctrl():
for i in range(100):
pass
awaddr = 4
matrix_size = 16
print('# matrix_size = %d' % matrix_size)
_saxi.write(awaddr, matrix_size)
awaddr = 8
a_offset = 0
print('# a_offset = %d' % a_offset)
_saxi.write(awaddr, a_offset)
awaddr = 12
b_offset = 1024 * 1
print('# b_offset = %d' % b_offset)
_saxi.write(awaddr, b_offset)
awaddr = 16
c_offset = 1024 * 2
print('# c_offset = %d' % c_offset)
_saxi.write(awaddr, c_offset)
awaddr = 0
start_time = counter
print('# start time = %d' % start_time)
_saxi.write(awaddr, 1)
araddr = 20
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
end_time = counter
print('# end time = %d' % end_time)
time = end_time - start_time
print('# exec time = %d' % time)
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(led,
'uut',
params=m.connect_params(led),
ports=m.connect_ports(led))
simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(1000000),
Systask('finish'),
)
return m
def run(
act_dtype=ng.int8,
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)
# pytorch model
if with_batchnorm:
model = torchvision.models.vgg11_bn(pretrained=True)
else:
model = torchvision.models.vgg11(pretrained=True)
# Pytorch to ONNX
onnx_filename = 'vgg11_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 = 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)
# compare outputs of hidden layers
features_ops = [
v for k, v in operators.items()
if isinstance(v, ng.conv2d) and not isinstance(v, ng.matmul)
]
features_ops = list(sorted(set(features_ops), key=features_ops.index))
avgpool_op = [
v for k, v in operators.items()
if isinstance(v, (ng.avg_pool, ng.avg_pool_serial))
][0]
classifier_ops = [
v for k, v in operators.items() if isinstance(v, ng.matmul)
]
classifier_ops = list(sorted(set(classifier_ops),
key=classifier_ops.index))
sub_ops = features_ops + [avgpool_op] + classifier_ops
sub_outs = ng.eval(sub_ops, act=vact)
sub_outs = [
sub_out.transpose([0, 3, 1, 2])
for sub_out in sub_outs[:-len(classifier_ops)]
] + sub_outs[-len(classifier_ops):]
sub_scale_factors = (
[features_op.scale_factor
for features_op in features_ops] + [avgpool_op.scale_factor] +
[classifier_op.scale_factor for classifier_op in classifier_ops])
model.eval()
model_features_relu_layers = [
layer for layer in model.features if isinstance(layer, nn.ReLU)
]
model_features_relu_indexes = [
list(model.features).index(layer)
for layer in model_features_relu_layers
]
model_features_seqs = [
model.features[:i + 1] for i in model_features_relu_indexes
]
model_features_outs = [
seq(torch.from_numpy(model_input)).detach().numpy()
for seq in model_features_seqs
]
model_avgpool_out = nn.Sequential(model.features, 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_classifier_relu_layers = [
layer for layer in model.classifier if isinstance(layer, nn.ReLU)
]
model_classifier_relu_indexes = [
list(model.classifier).index(layer)
for layer in model_classifier_relu_layers
]
model_classifier_relu_indexes.append(len(model.classifier))
model_classifier_seqs = [
nn.Sequential(model.features, model.avgpool, Flatten(),
model.classifier[:i + 1])
for i in model_classifier_relu_indexes
]
model_classifier_outs = [
seq(torch.from_numpy(model_input)).detach().numpy()
for seq in model_classifier_seqs
]
model_outs = model_features_outs + [model_avgpool_out
] + model_classifier_outs
scaled_outs = [
model_out * scale_factor
for model_out, scale_factor in zip(model_outs, sub_scale_factors)
]
error_rates = [
np.sum(np.abs(sub_out - model_out)) / np.sum(np.abs(model_out))
for model_out, sub_out in zip(scaled_outs, sub_outs)
]
max_diffs = [
model_out.max() / sub_out.max()
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]
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]):
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],
'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 mkTest(n_input=784, n_classes=10):
# create target hardware
x = ng.placeholder(ng.int32, shape=[n_input])
w1 = ng.variable(ng.int32, shape=(n_input, n_input), name='h1')
w2 = ng.variable(ng.int32, shape=(n_input, n_input), name='h2')
w3 = ng.variable(ng.int32, shape=(n_classes, n_input), name='out')
l1 = ng.matmul(x, w1, transposed_b=True)
l1 = ng.relu(l1)
l2 = ng.matmul(l1, w2, transposed_b=True)
l2 = ng.relu(l2)
out = ng.matmul(l2, w3, transposed_b=True)
targ = ng.to_veriloggen([out], 'mlp')
#targ = ng.to_ipxact([model], 'mlp')
# 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
memory = axi.AxiMemoryModel(m, 'memory', clk, rst)
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
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))
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'),
)
return m
def run(
act_dtype=ng.int16,
weight_dtype=ng.int16,
bias_dtype=ng.int16,
scale_dtype=ng.int16,
par_ich=2,
par_och=2,
chunk_size=64,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
# simtype='verilator',
# simtype=None, # no RTL simulation
outputfile=None):
# --------------------
# (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,
sum_dtype=ng.int64)
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,
sum_dtype=ng.int64)
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,
sum_dtype=ng.int64)
# 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',
sum_dtype=ng.int64)
# --------------------
# (2) Assign quantized weights to the NNgen operators
# --------------------
# In this example, random integer values are assigned.
# In real cases, you should assign actual integer weight values
# obtianed by a training on DNN framework
w0_value = np.random.normal(size=w0.length).reshape(w0.shape)
w0_value = np.clip(w0_value, -5.0, 5.0)
w0_value = w0_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0
w0_value = np.round(w0_value).astype(np.int64)
w0.set_value(w0_value)
b0_value = np.random.normal(size=b0.length).reshape(b0.shape)
b0_value = np.clip(b0_value, -5.0, 5.0)
b0_value = b0_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0 / 100.0
b0_value = np.round(b0_value).astype(np.int64)
b0.set_value(b0_value)
s0_value = np.ones(s0.shape, dtype=np.int64)
s0.set_value(s0_value)
w1_value = np.random.normal(size=w1.length).reshape(w1.shape)
w1_value = np.clip(w1_value, -5.0, 5.0)
w1_value = w1_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0
w1_value = np.round(w1_value).astype(np.int64)
w1.set_value(w1_value)
b1_value = np.random.normal(size=b1.length).reshape(b1.shape)
b1_value = np.clip(b1_value, -5.0, 5.0)
b1_value = b1_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0 / 100.0
b1_value = np.round(b1_value).astype(np.int64)
b1.set_value(b1_value)
s1_value = np.ones(s1.shape, dtype=np.int64)
s1.set_value(s1_value)
w2_value = np.random.normal(size=w2.length).reshape(w2.shape)
w2_value = np.clip(w2_value, -5.0, 5.0)
w2_value = w2_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0
w2_value = np.round(w2_value).astype(np.int64)
w2.set_value(w2_value)
b2_value = np.random.normal(size=b2.length).reshape(b2.shape)
b2_value = np.clip(b2_value, -5.0, 5.0)
b2_value = b2_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0 / 100.0
b2_value = np.round(b2_value).astype(np.int64)
b2.set_value(b2_value)
s2_value = np.ones(s2.shape, dtype=np.int64)
s2.set_value(s2_value)
w3_value = np.random.normal(size=w3.length).reshape(w3.shape)
w3_value = np.clip(w3_value, -5.0, 5.0)
w3_value = w3_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0
w3_value = np.round(w3_value).astype(np.int64)
w3.set_value(w3_value)
b3_value = np.random.normal(size=b3.length).reshape(b3.shape)
b3_value = np.clip(b3_value, -5.0, 5.0)
b3_value = b3_value * (2.0**(weight_dtype.width - 1) - 1) / 5.0 / 100.0
b3_value = np.round(b3_value).astype(np.int64)
b3.set_value(b3_value)
s3_value = np.ones(s3.shape, dtype=np.int64)
s3.set_value(s3_value)
# --------------------
# (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
# cshamt_out: right shift amount after applying bias/scale
a0.attribute(par_ich=par_ich,
par_och=par_och,
cshamt_out=weight_dtype.width + 1)
a1.attribute(par_ich=par_ich,
par_och=par_och,
cshamt_out=weight_dtype.width + 1)
a2.attribute(par_ich=par_ich,
par_och=par_och,
cshamt_out=weight_dtype.width + 1)
output_layer.attribute(par_ich=par_ich,
par_och=par_och,
cshamt_out=weight_dtype.width + 1)
# max_pool
# par: parallelism in in/out channel
a0p.attribute(par=par_och)
# --------------------
# (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 = np.clip(input_layer_value, -5.0, 5.0)
input_layer_value = input_layer_value * (
2.0**(input_layer.dtype.width - 1) - 1) / 5.0
input_layer_value = np.round(input_layer_value).astype(np.int64)
#input_layer_value = np.ones(input_layer.shape).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) Simulate the generated hardware by Veriloggen and Verilog simulator
# --------------------
if simtype is None:
sys.exit()
# to memory image
param_data = ng.export_ndarray([output_layer], chunk_size)
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 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(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 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 mkTest():
m = Module('test')
copy_bytes = 1024 * 4
# target instance
memcpy = mkMemcpy()
uut = Submodule(m, memcpy, name='uut')
clk = uut['CLK']
rst = uut['RST']
memory = axi.AxiMemoryModel(m, 'memory', clk, rst)
memory.connect(uut.get_inst_ports(), 'maxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(uut.get_inst_ports(), 'saxi')
# Timer
counter = m.Reg('counter', 32, initval=0)
seq = Seq(m, 'seq', clk, rst)
seq(counter.inc())
def ctrl():
for i in range(100):
pass
awaddr = 4 * 1
print('# copy_bytes = %d' % copy_bytes)
_saxi.write(awaddr, copy_bytes)
awaddr = 4 * 2
src_offset = 0
print('# src_offset = %d' % src_offset)
_saxi.write(awaddr, src_offset)
awaddr = 4 * 3
dst_offset = 1024 * 8
print('# dst_offset = %d' % dst_offset)
_saxi.write(awaddr, dst_offset)
awaddr = 4 * 0
start_time = counter
print('# start time = %d' % start_time)
_saxi.write(awaddr, 1)
araddr = 4 * 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
end_time = counter
print('# end time = %d' % end_time)
time = end_time - start_time
print('# exec time = %d' % time)
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(1000000),
Systask('finish'),
)
return m
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)
def run(
act_dtype=ng.int16,
weight_dtype=ng.int16,
bias_dtype=ng.int32,
scale_dtype=ng.int16,
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
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)
if not with_batchnorm:
raise ValueError('with_batchnorm must be True for ResNet18.')
# pytorch model
model = torchvision.models.resnet18(pretrained=False)
model.conv1.in_channels = act_shape[-1]
model.fc = nn.Linear(in_features=model.fc.in_features,
out_features=10,
bias=True)
# Pytorch to ONNX
onnx_filename = 'resnet18.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.normal(size=act.length).astype(np.float32).reshape(
act.shape)
img = img * cifar10_std + cifar10_mean
# 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)
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]))
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], '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],
'onnx_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(a_shape=(15, 15),
b_shape=(15, 15),
bias_shape=None,
scale_shape=None,
a_dtype=ng.int32,
b_dtype=ng.int32,
bias_dtype=ng.int32,
scale_dtype=ng.int32,
c_dtype=ng.int32,
rshift_mul=None,
rshift_sum=None,
rshift_out=None,
act_func=None,
par_left_col=1,
par_left_row=1,
par_out_col=1,
concur_out_col=None,
stationary='right',
left_ram_size=None,
right_ram_size=None,
bias_ram_size=None,
scale_ram_size=None,
out_ram_size=None,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# create target hardware
a = ng.placeholder(a_dtype, shape=a_shape, name='a')
b = ng.placeholder(b_dtype, shape=b_shape, name='b')
if bias_shape is not None:
bias = ng.placeholder(bias_dtype, bias_shape, name='bias')
else:
bias = None
if scale_shape is not None:
scale = ng.placeholder(scale_dtype, scale_shape, name='scale')
else:
scale = None
transposed_a = False
transposed_b = True
c = ng.matmul(a, b, bias, scale, transposed_a, transposed_b, rshift_mul,
rshift_sum, rshift_out, act_func, False, c_dtype, ng.int32,
ng.int32, 'matmul', par_left_col, par_left_row, par_out_col,
concur_out_col, stationary, left_ram_size, right_ram_size,
bias_ram_size, scale_ram_size, None, None, None,
out_ram_size)
c = ng.relu(c)
targ = ng.to_veriloggen([c],
'matrix_matmul_relu',
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) % [5] - [3]
if bias is not None:
vbias = np.arange(bias.length, dtype=np.int64).reshape(
bias.shape) % [4]
else:
vbias = None
if scale is not None:
vscale = np.arange(scale.length, dtype=np.int64).reshape(
scale.shape) % [6]
else:
vscale = None
eval_outs = ng.eval([c], a=va, b=vb, bias=vbias, scale=vscale)
vc = eval_outs[0]
# to memory image
size_max = int(
math.ceil(
max(a.memory_size, b.memory_size,
bias.memory_size if bias is not None else 0, scale.memory_size
if scale is not None else 0, c.memory_size) / 4096)) * 4096
check_addr = max(a.addr, b.addr, bias.addr if bias is not None else -1,
scale.addr if scale is not None else -1,
c.addr) + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
axi.set_memory(
mem, va, memimg_datawidth, a_dtype.width, a.addr,
max(int(math.ceil(axi_datawidth / a_dtype.width)), par_left_col))
axi.set_memory(
mem, vb, memimg_datawidth, b_dtype.width, b.addr,
max(int(math.ceil(axi_datawidth / b_dtype.width)), par_left_col))
if bias is not None:
axi.set_memory(
mem, vbias, memimg_datawidth, bias_dtype.width, bias.addr,
max(int(math.ceil(axi_datawidth / bias_dtype.width)), par_out_col))
if scale is not None:
axi.set_memory(
mem, vscale, memimg_datawidth, scale_dtype.width, scale.addr,
max(int(math.ceil(axi_datawidth / scale_dtype.width)),
par_out_col))
axi.set_memory(
mem, vc, memimg_datawidth, c_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / c_dtype.width)), par_out_col))
# 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(c.shape[0]):
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(i, j, orig, check)
ok = False
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 mkTest(memimg_name=None, axi_datawidth=32, datawidth=4, addrwidth=10):
m = Module('test')
# target instance
led = mkLed(axi_datawidth, datawidth, addrwidth)
# copy paras and ports
params = m.copy_params(led)
ports = m.copy_sim_ports(led)
clk = ports['CLK']
rst = ports['RST']
memory = axi.AxiMemoryModel(m, 'memory', clk, rst, memimg_name=memimg_name)
memory.connect(ports, 'myaxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
def ctrl():
for i in range(100):
pass
for i in range(16):
# word addressing
r = memory.read_word(i, 0, datawidth)
print('read: mem[%d] -> %x' % (i, r))
# word addressing
w = (r + i + 100) % (2**datawidth - 1)
memory.write_word(i, 0, w, datawidth)
print('write: mem[%d] <- %x' % (i, w))
awaddr = 0
_saxi.write(awaddr, 1)
araddr = 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
araddr = 8
v = _saxi.read(araddr)
if v:
print('# verify: PASSED')
else:
print('# verify: FAILED')
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(led,
'uut',
params=m.connect_params(led),
ports=m.connect_ports(led))
# simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(1000000),
Systask('finish'),
)
return m
def mkTest(memimg_name=None, memimg_datawidth=32, datawidth=4):
m = Module('test')
# target instance
led = mkLed(datawidth=datawidth)
# copy paras and ports
params = m.copy_params(led)
ports = m.copy_sim_ports(led)
clk = ports['CLK']
rst = ports['RST']
length = 1024 * 1024 // (memimg_datawidth // 8)
mem = np.zeros([length], dtype=np.int64)
data = np.arange(length, dtype=np.int64) % [2**datawidth - 1] + [1]
addr = 0
axi.set_memory(mem, data, memimg_datawidth, datawidth, addr, None)
memory = axi.AxiMemoryModel(m,
'memory',
clk,
rst,
memimg=mem,
memimg_name=memimg_name,
memimg_datawidth=memimg_datawidth)
memory.connect(ports, 'myaxi')
# AXI-Slave controller
_saxi = vthread.AXIMLite(m, '_saxi', clk, rst, noio=True)
_saxi.connect(ports, 'saxi')
def ctrl():
for i in range(100):
pass
awaddr = 0
_saxi.write(awaddr, 1)
araddr = 4
v = _saxi.read(araddr)
while v == 0:
v = _saxi.read(araddr)
araddr = 8
v = _saxi.read(araddr)
if v:
print('# verify: PASSED')
else:
print('# verify: FAILED')
vthread.finish()
th = vthread.Thread(m, 'th_ctrl', clk, rst, ctrl)
fsm = th.start()
uut = m.Instance(led,
'uut',
params=m.connect_params(led),
ports=m.connect_ports(led))
#simulation.setup_waveform(m, uut)
simulation.setup_clock(m, clk, hperiod=5)
init = simulation.setup_reset(m, rst, m.make_reset(), period=100)
init.add(
Delay(1000000),
Systask('finish'),
)
return m
def run(act_shape=(1, 7, 7, 15),
act_dtype=ng.int32, out_dtype=ng.int32,
ksize=(1, 2, 2, 1), stride=(1, 2, 2, 1), padding='SAME',
par=1, value_ram_size=None, out_ram_size=None,
axi_datawidth=32, silent=False,
filename=None, simtype='iverilog', outputfile=None):
# create target hardware
act = ng.placeholder(act_dtype, shape=act_shape, name='act')
out = ng.max_pool(act, ksize=ksize,
strides=stride, padding=padding,
dtype=out_dtype, par=par,
value_ram_size=value_ram_size, out_ram_size=out_ram_size)
targ = ng.to_veriloggen([out], 'matrix_max_pool', silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
vact = np.arange(act.length, dtype=np.int64).reshape(act.shape) % [10]
eval_outs = ng.eval([out], act=vact)
vout = eval_outs[0]
# to memory image
size_max = int(math.ceil(max(act.memory_size, out.memory_size) / 4096)) * 4096
check_addr = max(act.addr, out.addr) + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)], dtype=np.int64)
mem = mem + [100]
axi.set_memory(mem, vact, memimg_datawidth,
act_dtype.width, act.addr,
max(int(math.ceil(axi_datawidth / act_dtype.width)), par))
axi.set_memory(mem, vout, memimg_datawidth,
out_dtype.width, check_addr,
max(int(math.ceil(axi_datawidth / out_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, out_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, out_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(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 mkTest(ich=3, och=10, ch=64, ksize=3, stride=1, col=28, row=28):
# create target hardware
# layer 0: conv2d, max_pool_serial, relu
input_layer = ng.placeholder(ng.int32,
shape=(1, row, col, ich),
name='input_layer')
w0 = ng.variable(ng.int32, shape=(ch, ksize, ksize, ich), name='w0')
a0 = ng.conv2d(input_layer, w0, strides=(1, stride, stride, 1))
a0 = ng.max_pool_serial(a0, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1))
a0 = ng.relu(a0)
# layer 1: conv2d, relu, reshape
w1 = ng.variable(ng.int32,
shape=(ch, ksize, ksize, a0.shape[-1]),
name='w1')
a1 = ng.conv2d(a0, w1, strides=(1, stride, stride, 1))
a1 = ng.relu(a1)
a1 = ng.reshape(a1, [-1])
# layer 2: full-connection
w2 = ng.variable(ng.int32, shape=(16, a1.shape[-1]), name='w2')
a2 = ng.matmul(a1, w2, transposed_b=True)
a2 = ng.relu(a2)
# layer 3: full-connection
w3 = ng.variable(ng.int32, shape=(och, a2.shape[-1]), name='w3')
output_layer = ng.matmul(a2, w3, transposed_b=True, name='output_layer')
targ = ng.to_veriloggen([output_layer], 'cnn')
#targ = ng.to_ipxact([output_layer], 'cnn')
# 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
memory = axi.AxiMemoryModel(m, 'memory', clk, rst, mem_addrwidth=23)
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
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))
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'),
)
return m
def run(a_shape=(15, 15),
axis=-1,
keep_dims=False,
a_dtype=ng.int32,
b_dtype=ng.int32,
par=1,
axi_datawidth=32,
silent=False,
filename=None,
simtype='iverilog',
outputfile=None):
# create target hardware
a = ng.placeholder(a_dtype, shape=a_shape, name='a')
b = ng.argmin(a, axis, keep_dims, par=par)
targ = ng.to_veriloggen([b],
'matrix_argmin',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
va = np.arange(a.length, dtype=np.int64).reshape(a.shape) % [150] - [100]
eval_outs = ng.eval([b], a=va)
vb = eval_outs[0]
# to memory image
size_max = int(math.ceil(max(a.memory_size, b.memory_size) / 4096)) * 4096
check_addr = max(a.addr, b.addr) + size_max
size_check = size_max
tmp_addr = check_addr + size_check
memimg_datawidth = 32
mem = np.zeros([1024 * 1024 * 8 // (memimg_datawidth // 8)],
dtype=np.int64)
mem = mem + [100]
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, check_addr,
max(int(math.ceil(axi_datawidth / b_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_datawidth=memimg_datawidth,
memimg=mem,
memimg_name=memimg_name)
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, b.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(b.shape[-1]):
orig = memory.read_word(i * b.aligned_shape[-1] + j, b.addr,
b_dtype.width)
check = memory.read_word(i * b.aligned_shape[-1] + j,
check_addr, b_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, 32, 32, 3),
act_dtype=ng.int32,
weight_dtype=ng.int32,
bias_dtype=ng.int32,
scale_dtype=ng.int32,
out_dtype=ng.int32,
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',
outputfile=None):
# 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)
# 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=out_dtype,
default_scale_dtype=scale_dtype,
default_bias_dtype=bias_dtype,
disable_fusion=disable_fusion)
# default linear quantization
value_ranges = {'act': (0, 255)}
ng.quantize(outputs, value_ranges=value_ranges)
# set attribute
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)
# create target hardware
act = placeholders['act']
out = outputs['out']
targ = ng.to_veriloggen([out],
'onnx_vgg11',
silent=silent,
config={'maxi_datawidth': axi_datawidth})
# verification data
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 and len(model_out.shape) == len(act.shape):
model_out = np.transpose(model_out, act.perm)
scaled_model_out = model_out * out.scale_factor
mout = scaled_model_out.astype(np.int64)
for bat in range(vout.shape[0]):
vout_max = np.max(vout[bat])
vout_max_index = list(vout[bat]).index(vout_max)
mout_max = np.max(mout[bat])
mout_max_index = list(mout[bat]).index(mout_max)
print("# vout[%d]: max = %d, index = %d" %
(bat, vout_max, vout_max_index))
print("# mout[%d]: max = %d, index = %d" %
(bat, mout_max, mout_max_index))
# out_diff = vout - scaled_model_out
# out_err = out_diff / (scaled_model_out + 0.00000001)
# max_out_err = np.max(np.abs(out_err))
# breakpoint()
# if max_out_err > 0.1:
# raise ValueError("too large output error: %f > 0.1" % max_out_err)
# to memory image
param_data = ng.make_param_array(variables, constants, 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], 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)), 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 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,
out_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,
out_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
