"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

1 is computed during the next loop. This is // possible because each iteration only uses T[i] in Step 2 and then // discards it in Step 6. d := bLimbs[i] c1 := addMulVVW(T[i:n+i], aLimbs, d) // Step 6 is replaced by shifting the virtual window we operate // over: T of the algorithm is T[i:] for us. That means that T1 in // Step 2 (T mod 2^_W) is simply T[i]. k0 in Step 3 is our m0inv. Y := T[i] * m.m0inv // Step 4 and 5 add Y × m to T, which as mentioned above is stored // at T[i:]. The two carries...

1 is computed during the next loop. This is // possible because each iteration only uses T[i] in Step 2 and then // discards it in Step 6. d := bLimbs[i] c1 := addMulVVW(T[i:n+i], aLimbs, d) // Step 6 is replaced by shifting the virtual window we operate // over: T of the algorithm is T[i:] for us. That means that T1 in // Step 2 (T mod 2^_W) is simply T[i]. k0 in Step 3 is our m0inv. Y := T[i] * m.m0inv // Step 4 and 5 add Y × m to T, which as mentioned above is stored // at T[i:]. The two carries...