Skip to content

Commit

Permalink
Merge the logic of quantize and q_leaky_relu into one
Browse files Browse the repository at this point in the history
  • Loading branch information
Aba committed Sep 16, 2023
1 parent 3103116 commit 6e96871
Show file tree
Hide file tree
Showing 3 changed files with 39 additions and 39 deletions.
16 changes: 8 additions & 8 deletions c/model.h
View file
Original file line number Diff line number Diff line change
@@ -1,25 +1,25 @@
#define N_BUNDLES 7
Bundle_t bundles [N_BUNDLES] = {
{.n=8, .l=2, .kw=11, .coe=2, .coe_tl=2, .r_ll=8, .h=16, .w=8, .w_kw2=3, .t=8, .p=3, .cm=1, .cm_p0=1, .w_bpt=140, .w_bpt_p0=140, .x_bpt=840, .x_bpt_p0=840, .is_bias=1, .b_offset=0, .b_val_shift=9, .b_bias_shift=0, .x_header=414341061322735616, .x_header_p0=414341061322735616, .w_header=414587437826703360, .w_header_p0=414341061322735616 },
{.n=8, .l=2, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=16, .w=8, .w_kw2=8, .t=1, .p=1, .cm=20, .cm_p0=16, .w_bpt=200, .w_bpt_p0=200, .x_bpt=13320, .x_bpt_p0=13320, .is_bias=1, .b_offset=16, .b_val_shift=9, .b_bias_shift=0, .x_header=8700964375684448256, .x_header_p0=8700964375684448256, .w_header=8701210795138088960, .w_header_p0=8700964375684448256 },
{.n=8, .l=2, .kw=7, .coe=3, .coe_tl=4, .r_ll=8, .h=16, .w=8, .w_kw2=5, .t=6, .p=8, .cm=2, .cm_p0=2, .w_bpt=176, .w_bpt_p0=176, .x_bpt=1672, .x_bpt_p0=1672, .is_bias=1, .b_offset=40, .b_val_shift=9, .b_bias_shift=0, .x_header=846686625550303232, .x_header_p0=846686625550303232, .w_header=846933027824074752, .w_header_p0=846686625550303232 },
{.n=8, .l=2, .kw=5, .coe=4, .coe_tl=4, .r_ll=8, .h=16, .w=8, .w_kw2=6, .t=4, .p=4, .cm=4, .cm_p0=4, .w_bpt=248, .w_bpt_p0=248, .x_bpt=3336, .x_bpt_p0=3336, .is_bias=1, .b_offset=58, .b_val_shift=9, .b_bias_shift=0, .x_header=1927550536119222272, .x_header_p0=1927550536119222272, .w_header=1927796989932601344, .w_header_p0=1927550536119222272 },
{.n=8, .l=2, .kw=3, .coe=8, .coe_tl=8, .r_ll=8, .h=16, .w=8, .w_kw2=7, .t=3, .p=3, .cm=6, .cm_p0=4, .w_bpt=224, .w_bpt_p0=152, .x_bpt=5000, .x_bpt_p0=3336, .is_bias=1, .b_offset=74, .b_val_shift=9, .b_bias_shift=0, .x_header=3008414446688141312, .x_header_p0=1855492942081294336, .w_header=3008660883321651200, .w_header_p0=1855492942081294336 },
{.n=8, .l=2, .kw=1, .coe=24, .coe_tl=2, .r_ll=8, .h=16, .w=8, .w_kw2=8, .t=3, .p=2, .cm=20, .cm_p0=4, .w_bpt=248, .w_bpt_p0=56, .x_bpt=16648, .x_bpt_p0=3336, .is_bias=1, .b_offset=98, .b_val_shift=9, .b_bias_shift=0, .x_header=11006807384898142208, .x_header_p0=1783435348043366400, .w_header=11007053838711521280, .w_header_p0=1783435348043366400 },
{.n=1, .l=1, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=8, .w=1, .w_kw2=1, .t=1, .p=320, .cm=20, .cm_p0=20, .w_bpt=248, .w_bpt_p0=248, .x_bpt=138, .x_bpt_p0=138, .is_bias=1, .b_offset=170, .b_val_shift=9, .b_bias_shift=0, .x_header=10952754293765046272, .x_header_p0=10952754293765046272, .w_header=10952754456973803520, .w_header_p0=10952754293765046272 }
{.n=8, .l=2, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=16, .w=8, .w_kw2=8, .t=1, .p=1, .cm=20, .cm_p0=16, .w_bpt=200, .w_bpt_p0=200, .x_bpt=13320, .x_bpt_p0=13320, .is_bias=0, .b_offset=16, .b_val_shift=0, .b_bias_shift=0, .x_header=8700964375684448256, .x_header_p0=8700964375684448256, .w_header=8701210795138088960, .w_header_p0=8700964375684448256 },
{.n=8, .l=2, .kw=7, .coe=3, .coe_tl=4, .r_ll=8, .h=16, .w=8, .w_kw2=5, .t=6, .p=8, .cm=2, .cm_p0=2, .w_bpt=176, .w_bpt_p0=176, .x_bpt=1672, .x_bpt_p0=1672, .is_bias=1, .b_offset=16, .b_val_shift=9, .b_bias_shift=0, .x_header=846686625550303232, .x_header_p0=846686625550303232, .w_header=846933027824074752, .w_header_p0=846686625550303232 },
{.n=8, .l=2, .kw=5, .coe=4, .coe_tl=4, .r_ll=8, .h=16, .w=8, .w_kw2=6, .t=4, .p=4, .cm=4, .cm_p0=4, .w_bpt=248, .w_bpt_p0=248, .x_bpt=3336, .x_bpt_p0=3336, .is_bias=0, .b_offset=34, .b_val_shift=0, .b_bias_shift=0, .x_header=1927550536119222272, .x_header_p0=1927550536119222272, .w_header=1927796989932601344, .w_header_p0=1927550536119222272 },
{.n=8, .l=2, .kw=3, .coe=8, .coe_tl=8, .r_ll=8, .h=16, .w=8, .w_kw2=7, .t=3, .p=3, .cm=6, .cm_p0=4, .w_bpt=224, .w_bpt_p0=152, .x_bpt=5000, .x_bpt_p0=3336, .is_bias=1, .b_offset=34, .b_val_shift=9, .b_bias_shift=0, .x_header=3008414446688141312, .x_header_p0=1855492942081294336, .w_header=3008660883321651200, .w_header_p0=1855492942081294336 },
{.n=8, .l=2, .kw=1, .coe=24, .coe_tl=2, .r_ll=8, .h=16, .w=8, .w_kw2=8, .t=3, .p=2, .cm=20, .cm_p0=4, .w_bpt=248, .w_bpt_p0=56, .x_bpt=16648, .x_bpt_p0=3336, .is_bias=0, .b_offset=58, .b_val_shift=0, .b_bias_shift=0, .x_header=11006807384898142208, .x_header_p0=1783435348043366400, .w_header=11007053838711521280, .w_header_p0=1783435348043366400 },
{.n=1, .l=1, .kw=1, .coe=24, .coe_tl=0, .r_ll=8, .h=8, .w=1, .w_kw2=1, .t=1, .p=320, .cm=20, .cm_p0=20, .w_bpt=248, .w_bpt_p0=248, .x_bpt=138, .x_bpt_p0=138, .is_bias=1, .b_offset=58, .b_val_shift=9, .b_bias_shift=0, .x_header=10952754293765046272, .x_header_p0=10952754293765046272, .w_header=10952754456973803520, .w_header_p0=10952754293765046272 }
};

#define X_BITS_L2 2
#define W_BITS_L2 2
#define PE_ROWS 8
#define PE_COLS 24

#define WB_BYTES 98436
#define WB_BYTES 98212
#define W_BYTES 98048
#define X_BYTES 2520
#define X_BYTES_ALL 120040
#define Y_BYTES 294920
#define B_TYPE signed short
#define B_WORDS 194
#define B_WORDS 82
#define DATA_DIR "D:/dnn-engine/test/vectors"

38 changes: 18 additions & 20 deletions test/py/bundle.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -73,10 +73,18 @@ def extract_act(signature):
frac = d['bits']-int_bit-sign_bit

if isinstance(ilayer.quantizer, quantized_bits):
return { 'layer':ilayer, 'type':'quant', 'bits':d['bits'], 'frac':frac}
return { 'layer':ilayer, 'type':'quant', 'bits':d['bits'], 'frac':frac, 'plog_slope': 0, 'non_zero':1}
elif 'relu' in str(ilayer.quantizer.__class__) and ilayer.quantizer.negative_slope != 0:
return { 'layer':ilayer, 'type':'relu', 'slope':ilayer.quantizer.negative_slope, 'bits':d['bits'], 'frac':frac}
slope = ilayer.quantizer.negative_slope
if slope == 0:
assert ilayer.quantizer.bits != 1, "Error: Cannot use bits=1 with Relu. Use leaky_relu. Reason: Qkeras keeps relu signed"
ilayer.quantizer.bits -= 1
non_zero = 1*(slope != 0)
log_slope = np.log2(slope) if non_zero else 0
assert int(log_slope) == log_slope and log_slope <= 0, f"Error: negative_slope:{slope} of leaky_relu has to be a negative power of two. eg.0.125"
return { 'layer':ilayer, 'type':'relu', 'bits':d['bits'], 'frac':frac, 'slope':ilayer.quantizer.negative_slope, 'plog_slope':-int(log_slope), 'non_zero':non_zero}
else:
# TODO: support relu (slope=0). Qkeras uses different range for relu
raise Exception("Only leaky_relu (relu with negative_slope > 0) is suppported!")

'''
Expand Down Expand Up @@ -282,25 +290,15 @@ def div_round(n,d):
return (n + (d//2) - (~(d|n//d) &1)) // d

def apply_act(act_dict):
assert act_dict['type'] in ['quant', 'relu'], 'Error: Only quant & relu are supported yet'

x = self.proc['int'].astype(np.int32)
frac, bits = act_dict['frac'], act_dict['bits']

if act_dict['type'] == 'quant':
shift_bits = self.proc['frac']-frac

x = shift_round(x, shift_bits) # = np.around(x/2**shift_bits)
x = np.clip(x, -2**(bits-1), 2**(bits-1)-1).astype(int)

elif act_dict['type'] == 'relu':
log_act = -int(np.log2(act_dict['slope']))
assert log_act == -np.log2(act_dict['slope']), f"Leaky Relu slope: {act_dict['slope']} should be a power of two (eg:0.125)"
shift_bits = log_act + self.proc['frac']-frac

x = (x<0)*x + (((x>0)*x) << log_act)
x = shift_round(x, shift_bits) # = np.around(x/2**shift_bits)
x = np.clip(x, -2**(bits-log_act-1), 2**(bits-1)-1).astype(int)
else:
raise Exception('Only relu is supported yet')
frac, bits, plog_slope, non_zero = act_dict['frac'], act_dict['bits'], act_dict['plog_slope'], act_dict['non_zero']
shift_bits = plog_slope + self.proc['frac']-frac

x = ((x<0)*x)*non_zero + (((x>0)*x) << plog_slope)
x = shift_round(x, shift_bits) # = np.around(x/2**shift_bits)
x = np.clip(x, -2**(bits-plog_slope-1), 2**(bits-1)-1).astype(int)

self.proc['int'], self.proc['bits'], self.proc['frac'] = x, bits, frac

Expand Down
24 changes: 13 additions & 11 deletions test/py/param_test.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -159,6 +159,8 @@ def compile(c):
class Config:
K : int
CO: int
is_bias: bool
act_q: str
flatten: bool = False
dense: bool = False

Expand All @@ -185,34 +187,34 @@ class Config:
READY_PROB = [1],
)))
def test_dnn_engine(COMPILE):
c = make_compile_params(COMPILE)

input_shape = (8,16,8,3) # (XN, XH, XW, CI)
model_config = [
Config(11, 16),
Config(1, 16),
Config(7, 16),
Config(5, 16),
Config(3, 24),
Config(1, 50, flatten=True),
Config(1, 10, dense= True),
Config(11, 16, True , f'quantized_relu({c.X_BITS},0,negative_slope=0.125)'),
Config(1 , 16, False, f'quantized_bits({c.K_BITS},0,False,True,1)'),
Config(7 , 16, True , f'quantized_bits({c.K_BITS},0,False,True,1)'),
Config(5 , 16, False, f'quantized_relu({c.X_BITS},0,negative_slope=0.125)'),
Config(3 , 24, True , f'quantized_relu({c.X_BITS},0,negative_slope=0.125)'),
Config(1 , 50, False, f'quantized_relu({c.X_BITS},0,negative_slope=0.125)', flatten=True),
Config(1 , 10, True , f'quantized_relu({c.X_BITS},0,negative_slope=0.125)', dense= True),
]

'''
Build Model
'''
c = make_compile_params(COMPILE)
assert c.X_BITS in [1,2,4,8] and c.K_BITS in [1,2,4,8], "X_BITS and K_BITS should be in [1,2,4,8]"
assert c.B_BITS in [8,16,32], "B_BITS should be in [8,16,32]"
xq, kq, bq, aq = f'quantized_bits({c.X_BITS},0,False,True,1)', f'quantized_bits({c.K_BITS},0,False,True,1)', f'quantized_bits({c.B_BITS},0,False,True,1)', f'quantized_relu({c.X_BITS},0,negative_slope=0.125)'
xq, kq, bq = f'quantized_bits({c.X_BITS},0,False,True,1)', f'quantized_bits({c.K_BITS},0,False,True,1)', f'quantized_bits({c.B_BITS},0,False,True,1)'
inp = {'bits':c.X_BITS, 'frac':c.X_BITS-1}

x = x_in = Input(input_shape[1:], name='input')
x = QActivation(xq)(x)
for i, g in enumerate(model_config):
if g.dense:
d = {'core': {'type':'dense', 'units':g.CO, 'kernel_quantizer':kq, 'bias_quantizer':bq, 'use_bias':True, 'act_str':aq}}
d = {'core': {'type':'dense', 'units':g.CO, 'kernel_quantizer':kq, 'bias_quantizer':bq, 'use_bias':g.is_bias, 'act_str':g.act_q}}
else:
d = {'core': {'type':'conv', 'filters':g.CO, 'kernel_size':(g.K,g.K), 'strides':(1,1), 'padding':'same', 'kernel_quantizer':kq, 'bias_quantizer':bq, 'use_bias':True, 'act_str':aq}, 'flatten':g.flatten,}
d = {'core': {'type':'conv', 'filters':g.CO, 'kernel_size':(g.K,g.K), 'strides':(1,1), 'padding':'same', 'kernel_quantizer':kq, 'bias_quantizer':bq, 'use_bias':g.is_bias, 'act_str':g.act_q}, 'flatten':g.flatten,}
x = Bundle(**d)(x)

model = Model(inputs=x_in, outputs=x)
Expand Down

0 comments on commit 6e96871

Please sign in to comment.