In enoki library v0.1.0, the operator [] of the DynamicArray<Packet<...>>, which stores Packet data, only accepts integer argument. Here is an example:

1
2
3
4
5
constexpr size_t PacketSize = enoki::max_packet_size / sizeof(float); // = 8
using FloatP   = Packet<float, PacketSize>;
using FloatX   = DynamicArray<FloatP>;

FloatX data(-1.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f, 9.f, 10.f);

Of course we can access the tenth element by data[9]. Also, we can access it using enoki::slice(data, 9) because it is accually the tenth slice in the dynamic array. Besides, since the PacketSize is 8, the tenth element is located in the second place of the second packet. We can access it with enoki::packet(data, 1)[1] as well.

But what if we want to access data using an index that is encapsulated in struct Packet?

1
2
3
4
using UInt32P = Packet<uint32_t, PacketSize>;
UInt32P index;
FloatP data_i = data[index]; // Wrong
FloatP data_i = enoki::gather(data, index); // Right

We cannot access it using data[index] because there is no overload of operator [] for struct Packet. In this case, we should use enoki::gather instead and get the result in FloatP which has the same size and the value type with index and data respectively. According to the official document, it is equivalent to the following scalar loop.

1
2
3
FloatP result;
for (size_t i = 0; i < index.size(); ++i)
    result[i] = ((float *) data.data())[index[i]];

We can get better performance with enoki::gather if the target hardware supports the specific instructions.

More information: Reference of enoki