def bind_flex_scales(self):
scaleAB = self.flex_entry_A.scale * self.flex_entry_B.scale
scaleC = self.flex_entry_C.scale
alpha = self.alpha * scaleAB
beta = self.beta * scaleC
# TODO: hardcoding these sucks
self.params[6] = alpha
self.params[7] = beta
self.params[20] = 1. / scaleC
FlexPtrDescription.bind_ptr(self.params)
def bind_flex_scales(self):
scaleAB = self.flex_entry_I.scale * self.flex_entry_F.scale
scaleC = self.flex_entry_O.scale
alpha = self.alpha * scaleAB
beta = self.beta * scaleC
for kernel in self.kernels:
kernel[8] = alpha
kernel[9] = beta
kernel[-1] = 1. / scaleC
for kernel in self.kernels:
FlexPtrDescription.bind_ptr(kernel)
def bind_flex_scales(self):
scaleAB = self.flex_entry_E.scale * self.flex_entry_F.scale
scaleC = self.flex_entry_O.scale
alpha = self.alpha * scaleAB
beta = self.beta * scaleC
for kernel in self.kernels[1:1 + len(self.bprop_kernels)]:
kernel[8] = alpha
kernel[9] = beta
kernel[-1] = 1. / scaleC
if self.convert_out:
self.kernels[-1][-2] = 1. / scaleC
for kernel in self.kernels:
FlexPtrDescription.bind_ptr(kernel)
def _ew_bind_flex_scales(kernel):
for index, flex_scale_desc in kernel.flex_scale_info:
scale = flex_scale_desc.flex_entry.scale
scale = 1.0 / scale if flex_scale_desc.is_output else scale
kernel.params[index] = scale
FlexPtrDescription.bind_ptr(kernel.params)
def bind_flex_scales(self):
for k in self.kernels:
kernel, params = k
FlexPtrDescription.bind_ptr(params)
def bind_flex_scales(self):
self.transformer.get_op_tensor(self.op).flex_entry.allocate()
FlexPtrDescription.bind_ptr(self.params)