Adddtrtri

[btracey]

No description provided.

[btracey]

This very reliably crashed in cgo in Dlansy

brendan:~/Documents/mygo/src/github.com/gonum/lapack/cgo$ go test
--- FAIL: TestDlantr (0.04s)
    dlantr.go:77: Norm mismatch. norm = O, unitdiag = true, upper = true, m = 5, n = 10, lda = 11, Want -7, got 6.371521178048223.
    dlantr.go:77: Norm mismatch. norm = O, unitdiag = true, upper = true, m = 5, n = 10, lda = 11, Want -7, got 4.167894107105489.
fatal error: unexpected signal during runtime execution
[signal 0xb code=0x1 addr=0x0 pc=0x4f06aae]

runtime stack:
runtime.throw(0x4253460, 0x2a)
    /Users/brendan/gover/go/src/runtime/panic.go:527 +0x90
runtime.sigpanic()
    /Users/brendan/gover/go/src/runtime/sigpanic_unix.go:12 +0x5a

goroutine 1102 [syscall, locked to thread]:
runtime.cgocall(0x400ce60, 0xc82003c9d8, 0xc800000000)
    /Users/brendan/gover/go/src/runtime/cgocall.go:120 +0x11b fp=0xc82003c990 sp=0xc82003c960
github.com/gonum/lapack/cgo/clapack._Cfunc_LAPACKE_dlansy(0x554f00000065, 0x3, 0xc82141a910, 0x3, 0x0)
    github.com/gonum/lapack/cgo/clapack/_obj/_cgo_gotypes.go:6605 +0x3a fp=0xc82003c9d8 sp=0xc82003c990
github.com/gonum/lapack/cgo/clapack.Dlansy(0x4f, 0x55, 0x3, 0xc82141a910, 0x9, 0x9, 0x3, 0x1)
    /Users/brendan/Documents/mygo/src/github.com/gonum/lapack/cgo/clapack/clapack.go:3044 +0x72 fp=0xc82003ca18 sp=0xc82003c9d8
github.com/gonum/lapack/cgo.Implementation.Dlansy(0x4f, 0x79, 0x3, 0xc82141a910, 0x9, 0x9, 0x3, 0xc82141a960, 0x9, 0x9, ...)
    /Users/brendan/Documents/mygo/src/github.com/gonum/lapack/cgo/lapack.go:107 +0x12e fp=0xc82003ca70 sp=0xc82003ca18
github.com/gonum/lapack/cgo.(*Implementation).Dlansy(0x4316d00, 0x4f, 0x79, 0x3, 0xc82141a910, 0x9, 0x9, 0x3, 0xc82141a960, 0x9, ...)
    <autogenerated>:2 +0x10c fp=0xc82003cad0 sp=0xc82003ca70
github.com/gonum/lapack/testlapack.DpoconTest(0xc8200957a0, 0xe9816f8, 0x4316d00)
    /Users/brendan/Documents/mygo/src/github.com/gonum/lapack/testlapack/dpocon.go:54 +0x3de fp=0xc82003cf28 sp=0xc82003cad0
github.com/gonum/lapack/cgo.TestDpocon(0xc8200957a0)
    /Users/brendan/Documents/mygo/src/github.com/gonum/lapack/cgo/lapack_test.go:91 +0x76 fp=0xc82003cf68 sp=0xc82003cf28
testing.tRunner(0xc8200957a0, 0x4305e98)
    /Users/brendan/gover/go/src/testing/testing.go:456 +0x98 fp=0xc82003cfa0 sp=0xc82003cf68
runtime.goexit()
    /Users/brendan/gover/go/src/runtime/asm_amd64.s:1696 +0x1 fp=0xc82003cfa8 sp=0xc82003cfa0
created by testing.RunTests
    /Users/brendan/gover/go/src/testing/testing.go:561 +0x86d

goroutine 1 [chan receive]:
testing.RunTests(0x426ad38, 0x4305d60, 0x10, 0x10, 0x0)
    /Users/brendan/gover/go/src/testing/testing.go:562 +0x8ad
testing.(*M).Run(0xc83735fef8, 0x4023188)
    /Users/brendan/gover/go/src/testing/testing.go:494 +0x70
main.main()
    github.com/gonum/lapack/cgo/_test/_testmain.go:84 +0x116

goroutine 17 [syscall, locked to thread]:
runtime.goexit()
    /Users/brendan/gover/go/src/runtime/asm_amd64.s:1696 +0x1
exit status 2
FAIL    github.com/gonum/lapack/cgo 4.132s
''''

Presumably these code modifications are not the cause.

[btracey]

Checking I have up to date OpenBLAS

[btracey]

After updating, the crashes are intermittant, and look like they may have Dlantr as the root cause.

[kortschak]

Try making the minor change I have here to see if the fault goes away.

[kortschak]

I don't see faults here in either state of OpenBLAS, but I do see this:

$ go test -cpu 2,2,2,2,2,2,2,2,2,2,2,2,2 -run Dtrtri -v github.com/gonum/lapack/cgo
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.05s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- FAIL: TestDtrtri (0.05s)
    dtrtri.go:84: inv(A) * A != I. Upper = true, unit = true, n = 200, lda = 205
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.05s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)
=== RUN   TestDtrtri
--- FAIL: TestDtrtri (0.06s)
    dtrtri.go:84: inv(A) * A != I. Upper = true, unit = true, n = 200, lda = 200
    dtrtri.go:84: inv(A) * A != I. Upper = true, unit = true, n = 200, lda = 205
=== RUN   TestDtrtri
--- PASS: TestDtrtri (0.06s)

The pattern of success/failure is consistent over runs which suggests to me that it's a function of the PRNG and so the failures are matrix element dependent.

[kortschak]

Do you have any idea about the failures above?

[kortschak]

Returns whether it's singular or whether it's not singular; ok is ambiguous here. I think we have words elsewhere that dealt with this better.

It's worth explaining briefly the difference here between the BLAS level 3 and 2 algorithms.

[kortschak]

LGTM, except minor comments and I'd like to now about the failures in cgo.

[btracey]

Do the failures go away when you decrease the tolerance (make it closer to 1)? The condition number of these randomly generated matrices is often pretty bad, so the error in the final solution can be high even when the algorithm is correct. You can see how high it has to be in Dgetri, so maybe it needs to be higher here as well.

[btracey]

(The failures don't happen on my machine)

[kortschak]

They go away at 1e-4.

[kortschak]

I think this could be better: "Dtrtri will not perform the inversion if the matrix is singular, and returns a boolean indicating whether the inversion was successful."

I think that covers the behaviour.