Segmentation fault (core dumped): hcl.compute with placeholders whose dim more than 2

[tonyjie]

I'm trying to write a conv_layer using HeteroCL, but a memory error can't be solved.

input = hcl.placeholder((4,32,10,10), "input")
filter = hcl.placeholder((32,32,3,3), "filter")
bias = hcl.placeholder((32,), "bias")
print(bias.shape)

def conv_layer(input, filter, bias):
    def compare_with_zero(a):
        with hcl.if_(a > 0):
            hcl.return_(a)
        with hcl.else_():
            hcl.return_(0)
    output = hcl.compute(input.shape, lambda n,c,h,w: compare_with_zero(input[n,c,h,w]), "output")
    return output
s = hcl.create_schedule([input, filter, bias], conv_layer)
print(hcl.lower(s))
f = hcl.build(s)
print('done')

Above program went error when doing with f = hcl.build(s) However, when the input placeholder's dimension is 2, this program works. And when the dimension becomes 3 or 4, the error of Segmentation Fault (core dumped) came.

[seanlatias]

HeteroCL does not support return_ statement inside a normal Python function call. In this case, you have several options.

1. Use hcl.select

   def compare_with_zero(a):
   return hcl.select(a > 0, a ,0)

2. Use an intermediate variable to hold the result

   def compare_with_zero(a):
   b = hcl.local(0)
   with hcl.if_(a > 0):
       b[0] = a
   with hcl.else_():
       b[0] = 0
   return b[0]

Please close this issue if this solves the problem.

[tonyjie]

HeteroCL does not support return_ statement inside a normal Python function call. In this case, you have several options.

1. Use hcl.select

def compare_with_zero(a):
    return hcl.select(a > 0, a ,0)

1. Use an intermediate variable to hold the result

def compare_with_zero(a):
    b = hcl.local(0)
    with hcl.if_(a > 0):
        b[0] = a
    with hcl.else_():
        b[0] = 0
    return b[0]

Please close this issue if this solves the problem.

Thanks! Both of the methods can work. In this case, I think the second option is a general solution for now, using a local intermediate variable to hold the result. By the way, another question is that when using return instead of hcl.return_(), the normal Python function didn't work. To be specific:

def conv_layer(input, filter, bias):
    def compare_with_zero(a):
        with hcl.if_(a > 0):
            return a
        with hcl.else_():
            return 0

when printing hcl.lower(s), this function only became an output = input. That's the reason why I'm using hcl.return_() . I think the reason is that when we use some dsl function of hcl, such as hcl.if_() and hcl.else_(), HeteroCL doesn't support normal return. I'm not sure is that correct?

Anyway, the respond helps. Thanks

[seanlatias]

Yes, you are right. It would be a good idea to support hcl.return_ within an if/else statement in the future.

[seanlatias]

Fixed by #116.