"true" or "1", this environment variable makes `tf.nn.bias_add` operate
    deterministically (i.e. reproducibly), but currently only when XLA JIT
    compilation is *not* enabled. Setting `TF_DETERMINISTIC_OPS` to "true" or
    "1" also makes cuDNN convolution and max-pooling operate deterministically.
    This makes Keras Conv\*D and MaxPool\*D layers operate deterministically in

fieldMul(a, b fieldElement) fieldElement { x := uint32(a) * uint32(b) return fieldReduce(x) } // fieldMulSub returns a * (b - c). This operation is fused to save a // fieldReduceOnce after the subtraction. func fieldMulSub(a, b, c fieldElement) fieldElement { x := uint32(a) * uint32(b-c+q) return fieldReduce(x) } // fieldAddMul returns a * b + c * d. This operation is fused to save a // fieldReduceOnce and a fieldReduce. func fieldAddMul(a, b, c, d fieldElement) fieldElement { x := uint32(a) * uint32(b)...

qNegInv u := (x + uint64(t)*q) >> 32 return fieldReduceOnce(uint32(u)) } // fieldMontgomeryMulSu returns a * (b - c). This operation is fused to save a // fieldReduceOnce after the subtraction. func fieldMontgomeryMulSu(a, b, c fieldElement) fieldElement { x := uint64(a) * uint64(b-c+q) return fieldMontgomeryReduc(x) } // fieldMontgomeryAddMu returns a * b + c * d. This operation is fused to save // a fieldReduceOnce and a fieldReduce. func fieldMontgomeryAddMu(a, b, c, d fieldElement) fieldElement...