*   Implicit masks for `query`, `key` and `value` inputs will
            automatically be used to compute a correct attention mask for the
            layer. These padding masks will be combined with any
            `attention_mask` passed in directly when calling the layer. This can
            be used with
            [`tf.keras.layers.Embedding`](https://www.tensorflow.org/api_docs/python/tf/keras/layers/Embedding)

    * These `dependencies` are run before the normal parameter dependencies. And normal dependencies are run too. They can be combined.
    * Dependencies declared in a router are executed first, then the ones defined in *path operation decorators*, and then the ones declared in normal parameters. They are all combined and executed.

return Element{ rand.Uint64() & maskLow52Bits, rand.Uint64() & maskLow52Bits, rand.Uint64() & maskLow52Bits, rand.Uint64() & maskLow52Bits, rand.Uint64() & maskLow52Bits, } } // weirdLimbs can be combined to generate a range of edge-case field elements. // 0 and -1 are intentionally more weighted, as they combine well. var ( weirdLimbs51 = []uint64{ 0, 0, 0, 0, 1, 19 - 1, 19, 0x2aaaaaaaaaaaa, 0x5555555555555, (1 << 51) - 20, (1 << 51) - 19, (1 << 51) - 1, (1 << 51) - 1, (1 << 51) - 1, (1 << 51) - 1, }...