Can you provide .so compiling setting? about as-mlp HOT 5 CLOSED

@laisimiao The shift op will be compiled on the fly and save to your system cache folder. So you can directly run without setup.py

from as-mlp.

@niujinshuchong Thanks. Recently I found a weird question:

x = torch.arange(5)
y = torch.arange(5)
mesh = torch.meshgrid([x, y])
xy = torch.stack([mesh[0], mesh[1]], dim=0)
a = torch.cat([xy, xy, xy], dim=0)[:5,:,:].unsqueeze(0).cuda()
print(a.shape)  # B,C,H,W

class AxialShift(nn.Module):
    def __init__(self, dim, shift_size):
        super().__init__()
        self.shift = Shift(shift_size, dim)

    def forward(self, x):
        return self.shift(x)

shift = AxialShift(2, 5).cuda()
y = shift(a)
print(y.shape)

it will give errors like this:

cupy.cuda.compiler.CompileException: /tmp/tmpigcfzd1m/617aebd77e7545b257dbbd01681f6aa29014226e.cubin.cu(6): error: identifier "None" is undefined

/tmp/tmpigcfzd1m/617aebd77e7545b257dbbd01681f6aa29014226e.cubin.cu(6): error: identifier "None" is undefined

/tmp/tmpigcfzd1m/617aebd77e7545b257dbbd01681f6aa29014226e.cubin.cu(15): error: identifier "None" is undefined

3 errors detected in the compilation of "/tmp/tmpigcfzd1m/617aebd77e7545b257dbbd01681f6aa29014226e.cubin.cu".

But if I provide shift with x = torch.randn(1,5,5,5).cuda() not a in above example, it's no error. Why would ths happen?

from as-mlp.

@laisimiao That's wired. Could you maybe try a = a.contiguous() ?

from as-mlp.

Oh I found the reason: it will give errors when input data type is int but fine with float data type.

from as-mlp.

Great.

from as-mlp.