def __init__(self,
node_name,
cname,
params,
linear_q,
act_params,
act_q,
at_ver=3,
gen_ctrl=None):
if gen_ctrl is None:
gen_ctrl = GenCtrl(None, cname=cname)
else:
gen_ctrl.cname = cname
assert params is not None, "linear should always be included"
in_size = params.in_dims[0].size()
out_size = params.out_dims[0].size()
filter_q = linear_q.in_qs[1]
in_q = linear_q.in_qs[0]
out_q = linear_q.out_qs[0]
bias_q = linear_q.in_qs[2]
if act_params is not None:
act_op = gen_activation_op(act_params.activation)
out_q = act_q.out_qs[0]
if at_ver < 3:
if act_params.activation == "relu6" and out_q.q != 0:
gen_ctrl.ReluN = 6 << out_q.q
gen_ctrl.ReluNNoNorm = 1
else:
if act_params.activation == "relun":
gen_ctrl.ReluN = act_params.activation_params
else:
act_op = "KOP_NONE"
# attributes used to test equality - i.e. this kernel can be reused
attrs = {
'in_q': in_q,
'filter_q': filter_q,
'bias_q': bias_q,
'out_q': out_q,
'in_size': in_size,
'out_size': out_size,
'act_op': act_op
}
# other attributes
extra_attrs = {'cname': cname, 'node_name': node_name}
super().__init__(attrs, extra_attrs, gen_ctrl=gen_ctrl)
self.at_ver = at_ver
def __init__(self, node_name, cname, params, aparams, qrec, gen_ctrl=None):
if gen_ctrl is None:
gen_ctrl = GenCtrl(None, cname=cname)
else:
gen_ctrl.cname = cname
if isinstance(aparams,
ReluActivationParameters) and aparams.upper_bound:
gen_ctrl.ReluN = aparams.upper_bound
in_dims = params.in_dims
matop = "KOP_MATMUL" if len(in_dims) > 2 else "KOP_MATMUL_NOBIAS"
if isinstance(params, MatMulTransposedParameters):
matop += "_TRANSPOSED"
attrs = {
'ColM1':
in_dims[0][1],
'LineM1':
in_dims[0][0],
'ColM2':
in_dims[1][0] if isinstance(
params, MatMulTransposedParameters) else in_dims[1][1],
'LineM2':
in_dims[1][1] if isinstance(params, MatMulTransposedParameters)
else in_dims[1][0],
'Width':
0,
'Height':
0,
'Scx':
1,
'Scy':
1,
'kop_matmul':
matop,
'actoper':
gen_activation_op(aparams.activation if aparams else "none")
}
extra_attrs = {'cname': cname, 'node_name': node_name}
super().__init__(attrs, extra_attrs, gen_ctrl=gen_ctrl)
def __init__(self, node_name, cname, conv_params, conv_q,
pool_params, pool_q, act_params, act_q, at_ver=3, gen_ctrl=None):
if gen_ctrl is None:
gen_ctrl = GenCtrl(None, cname=cname)
else:
gen_ctrl.cname = cname
in_q = filter_q = out_q = bias_q = mul_biases_q = None
in_dim = out_dim = None
pad_compatibilities = []
if conv_params is not None:
at_conv_params = gen_conv_at_params(
conv_params, conv_q, pad_compatibilities)
in_dim = conv_params.in_dims[0]
out_dim = conv_params.out_dims[0]
filter_q = conv_q.in_qs[1]
in_q = conv_q.in_qs[0]
out_q = conv_q.out_qs[0]
bias_q = conv_q.in_qs[2]
if conv_params.has_mul_bias:
mul_biases_q = conv_q.mul_biases_q
else:
at_conv_params = NO_CONV
if pool_params is not None:
at_pool_params = gen_pool_at_params(
pool_params, pad_compatibilities)
if in_dim is None:
in_dim = pool_params.in_dims[0]
out_dim = pool_params.out_dims[0]
if in_q is None:
in_q = pool_q.in_qs[0]
out_q = pool_q.out_qs[0]
else:
at_pool_params = NO_POOL
if act_params is not None:
at_act_params = gen_active_at_params(act_params)
if in_dim is None:
in_dim = act_params.in_dims[0]
if out_dim is None:
out_dim = act_params.out_dims[0]
if in_q is None:
in_q = act_q.in_qs[0]
out_q = act_q.out_qs[0]
if at_ver < 3:
if act_params.activation == "relu6" and out_q.q != 0:
gen_ctrl.ReluN = 6 << out_q.q
gen_ctrl.ReluNNoNorm = 1
else:
if act_params.activation == "relun":
gen_ctrl.ReluN = act_params.activation_params
else:
at_act_params = NO_ACTIVATION
if pad_compatibilities:
reduction = PadDim.pad_compatibility_reduce(*pad_compatibilities,
"convolution padding is not compatible with pool padding")
if not reduction[2]: # default is balanced pad left
at_pad_ctrl = next(i for i, v in enumerate(reduction) if v)
LOG.debug("%s: generating pad control block", node_name)
gen_ctrl.PadType = at_pad_ctrl
attrs = {
'in_qtype': in_q,
'out_qtype': out_q,
'filter_qtype': filter_q,
'bias_qtype': bias_q,
'mul_biases_qtype': mul_biases_q,
'relu_oper': at_act_params.ReLUOper
}
if at_pool_params.PoolOper != 'KOP_NONE':
attrs.update({
'pool_oper': at_pool_params.PoolOper,
'pool_w': at_pool_params.Fpx,
'pool_h': at_pool_params.Fpy,
'pool_d_w': at_pool_params.Dpx,
'pool_d_h': at_pool_params.Dpy,
'pool_s_w': at_pool_params.Spx,
'pool_s_h': at_pool_params.Spy,
'pool_pad': at_pool_params.PoolPad
})
else:
attrs.update({
'pool_oper': 'KOP_NONE',
'pool_w': 0,
'pool_h': 0,
'pool_d_w': 0,
'pool_d_h': 0,
'pool_s_w': 0,
'pool_s_h': 0,
'pool_pad': 0
})
if at_conv_params == NO_CONV:
if in_q.dtype_bits != out_q.dtype_bits:
raise NotImplementedError(
"only homogenious operations are supported at present")
LOG.debug("%s: pool relu inq %s outq %s control block",
node_name, in_q, out_q)
if at_pool_params.PoolOper == 'KOP_NONE' and (not in_dim.is_named or not in_dim.has_keys(['c', 'w', 'h'])):
in_shape = in_dim.shape + ([1] * (3 - len(in_dim.shape)))
in_c, in_h, in_w = in_shape[0], in_shape[1], in_shape[2]
else:
in_c, in_h, in_w = in_dim.c, in_dim.h, in_dim.w
if out_dim.is_named and out_dim.has_key('c'):
out_c = out_dim.c
else:
out_c = in_c
attrs.update({
'in_c': in_c,
'in_h': in_h,
'in_w': in_w,
'out_c': out_c,
'conv_oper': 'KOP_NONE'
})
self.template = 'CALL_TEMPLATE_POOL_RELU'
else:
# swap w and h if w and filter w is 1 so generator sees 1D conv
if in_dim.w == 1 and at_conv_params.Fcx == 1:
attrs.update({
'in_c': in_dim.c,
'in_h': 1,
'in_w': in_dim.h,
'out_c': out_dim.c,
'conv_oper': at_conv_params.ConvOper,
'conv_w': at_conv_params.Fcy,
'conv_h': 1,
'conv_d_w': at_conv_params.Dcy,
'conv_d_h': at_conv_params.Dcx,
'conv_s_w': at_conv_params.Scy,
'conv_s_h': at_conv_params.Scx,
'conv_pad': at_conv_params.ConvPad
})
else:
attrs.update({
'in_c': in_dim.c,
'in_h': in_dim.h,
'in_w': in_dim.w,
'out_c': out_dim.c,
'conv_oper': at_conv_params.ConvOper,
'conv_w': at_conv_params.Fcx,
'conv_h': at_conv_params.Fcy,
'conv_d_w': at_conv_params.Dcx,
'conv_d_h': at_conv_params.Dcy,
'conv_s_w': at_conv_params.Scx,
'conv_s_h': at_conv_params.Scy,
'conv_pad': at_conv_params.ConvPad
})
if isinstance(at_conv_params, ConvATParam):
if mul_biases_q is not None:
LOG.debug("%s: mulconv pool relu inq %s outq %s control block",
node_name, in_q, out_q)
self.template = 'CALL_TEMPLATE_MULBIAS_CONV_POOL_RELU'
else:
LOG.debug("%s: conv pool relu inq %s outq %s control block",
node_name, in_q, out_q)
self.template = 'CALL_TEMPLATE_CONV_POOL_RELU'
elif isinstance(at_conv_params, GroupedConvATParam):
attrs.update({
'group_in': at_conv_params.GroupIn,
'group_out': at_conv_params.GroupOut
})
if mul_biases_q is not None:
LOG.debug("%s: grouped conv pool relu inq %s outq %s control block",
node_name, in_q, out_q)
self.template = 'CALL_TEMPLATE_GROUPED_MULBIAS_CONV_POOL_RELU'
else:
LOG.debug("%s: grouped mulconv pool relu inq %s outq %s control block",
node_name, in_q, out_q)
self.template = 'CALL_TEMPLATE_GROUPED_CONV_POOL_RELU'
else:
raise ValueError('Internal error')
# other attributes
extra_attrs = {
'cname': cname,
'node_name': node_name
}
super().__init__(attrs, extra_attrs, gen_ctrl=gen_ctrl)
