نمذجة موتر في Mojo

A tensor is just an n-dimensional grid of numbers. A vector is 1D, a matrix is 2D, and an image batch is often 4D.

However many dimensions you imagine, a tensor really lives as one flat buffer of values in memory, laid out end to end.

The shape is a tuple of sizes, one per dimension. A shape of (2, 3) means two rows of three columns.

You name an element with one index per dimension. In a (2, 3) tensor, (1, 2) is the last column of the second row.

A stride tells you how many flat slots to jump to move one step along a dimension. Strides turn indices into a single offset.

Flatten any index by a dot product: offset equals the sum of each index times its stride. That offset reaches the right slot.

Mojo numeric code usually stores rows contiguously, called row-major. The last dimension has stride 1, so columns sit side by side.

With clever strides you can view, slice, or transpose a tensor without copying data. Only the strides change, not the buffer.

Every tensor has a fixed dtype, like Float32 or Int64. A uniform type lets Mojo pack values tightly and compute fast.

You can model a tensor as a struct that bundles a data pointer with its shape and strides, keeping layout info in one place.

A contiguous tensor has no gaps, so a linear scan is cheap. Strided views may skip around, which can slow memory access.

You have a (2, 3) row-major tensor. How do you reach element (1, 2)?

A tensor is a flat buffer plus shape and strides; the offset formula maps indices to memory, and row-major keeps columns contiguous. 🧮