Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA
    9 and CUDA 9.1 sometimes does not properly compute the carry bit when
    decomposing 64-bit address calculations with large offsets (e.g. `load [x +
    large_constant]`) into 32-bit arithmetic in SASS.

    As a result, these versions of `ptxas` miscompile most XLA programs which

11;\n}\n\n.order-12 {\n  -ms-flex-order: 12;\n  order: 12;\n}\n\n.offset-1 {\n  margin-left: 8.333333%;\n}\n\n.offset-2 {\n  margin-left: 16.666667%;\n}\n\n.offset-3 {\n  margin-left: 25%;\n}\n\n.offset-4 {\n  margin-left: 33.333333%;\n}\n\n.offset-5 {\n  margin-left: 41.666667%;\n}\n\n.offset-6 {\n  margin-left: 50%;\n}\n\n.offset-7 {\n  margin-left: 58.333333%;\n}\n\n.offset-8 {\n  margin-left: 66.666667%;\n}\n\n.offset-9 {\n  margin-left: 75%;\n}\n\n.offset-10 {\n  margin-left: 83.333333%;\n}\n\n.offset-11 {\n...

11;\n}\n\n.order-12 {\n  -ms-flex-order: 12;\n  order: 12;\n}\n\n.offset-1 {\n  margin-left: 8.333333%;\n}\n\n.offset-2 {\n  margin-left: 16.666667%;\n}\n\n.offset-3 {\n  margin-left: 25%;\n}\n\n.offset-4 {\n  margin-left: 33.333333%;\n}\n\n.offset-5 {\n  margin-left: 41.666667%;\n}\n\n.offset-6 {\n  margin-left: 50%;\n}\n\n.offset-7 {\n  margin-left: 58.333333%;\n}\n\n.offset-8 {\n  margin-left: 66.666667%;\n}\n\n.offset-9 {\n  margin-left: 75%;\n}\n\n.offset-10 {\n  margin-left: 83.333333%;\n}\n\n.offset-11 {\n...