SIGBUS error when run with build tag cuda

[datastream]

OS: debian 9.3
Kernal: 4.9.0-4-amd64 #1 SMP Debian 4.9.65-3 (2017-12-03) x86_64 GNU/Linux
CUDA: 9.0
cuDNN: 7.0.5
When i use LispMachine and build program with go build -tags='cuda', I got a "signal SIGBUS" error.
After some debug, it turn out memory alloc failed.

gorgonia/dual.go

Line 199 in afce988

mem, err := op.Get(op.Device, memsize)

+++ b/dual.go
@@ -196,11 +196,12 @@ func dvUnitManaged(v Value, op *ExternalOp) (*dualValue, error) {
        dt := v.Dtype()
        memsize := calcMemSize(dt, s)
        // allocate on device
+       fmt.Println(dt,s,memsize)
        mem, err := op.Get(op.Device, memsize)
        if err != nil {
                return nil, err
        }
+       fmt.Println("----", mem.MemSize())
        d, err := makeValueFromMem(TypeOf(v), s, mem)
        if err != nil {
                return nil, err

it turn out like this.

float32 (79) 316
---- 7897856
float32 (79) 316
2018/11/23 23:55:36 MEMSIZE ERR MemGetAddressRange: InvalidContext
---- 0
float32 (79) 316
2018/11/23 23:55:36 MEMSIZE ERR MemGetAddressRange: InvalidContext
---- 0
float32 () 4
2018/11/23 23:55:36 MEMSIZE ERR MemGetAddressRange: InvalidContext
---- 0
float32 () 4
2018/11/23 23:55:36 MEMSIZE ERR MemGetAddressRange: InvalidContext
---- 0
float32 (79) 316
2018/11/23 23:55:36 MEMSIZE ERR MemGetAddressRange: InvalidContext
---- 0

[chewxy]

Hm - do you have an example of your NN?

[chewxy]

(the allocator tries occasionally to do stupid "smart" things based on heuristics - so it may be just that)

[datastream]

I'm using github.com/owulveryck/lstm to handle system log info. The test program is very similar to https://github.com/owulveryck/lstm/blob/master/example/train/train.go.

[owulveryck]

Maybe a fix for issue #297 will fix this as well?