tensors and such

Tensor primitives with arbitrary view, shape, stride. Cuda and CPU backends

Goal is high performance ML stack with minimal dependencies and maximal flexibility

todo

• Backend abstraction
• Nicer syntax. macro time tset!(tensor.view_mut(), v: 99, 1, 2) and tget!(tensor.view(), 1, 2)
• Basic GPU backend
• Slicing with ranges tensor.view().slice(0, 1..3) etc.
• Test more slicing syntaxes
• Slicing macro
• Elementwise broadcasting
• Basic linear algebra helpers
• Accelerated backends (GPU / parallel) in progress
• x86 SIMD paths (currently relying on llvm auto-vectorization for CPU which only works for contiguous memory)
• Multiple gpu devices allowed
• Do not lock thread on GPU dispatch
• strides and offset as bytes
• Broadcasting
• Idx should not be ref. makes it less ergonomic

to optimize

• collapse dims to allow backend to access contiguous fast paths
• view_to_owned can probably be optimized to copy larger chunks at once
• Restrict tensor values to require basic operations
• perf sucks for contiguous memory in unary CPU - fix
• perf sucks for contiguous memory in unary CUDA - fix
• perf sucks for non-contiguous memory in unary CPU - fix
• perf sucks for non-contiguous memory in unary CUDA - fix
• O(rank * size) instead of O(size) broadcasting ops is bad
• Broadcast cuda kernel puts a cap on tensor dim size - fix

Creating Tensors

let zeros = TensorBase::<f32, Cpu>::zeros((3, 4)); // 3x4 tensor of zeros

let cpu_ones = CpuTensor::<f32>::ones((2, 2)); // 2x2 tensor of ones
let gpu_max = CudaTensor::<f32>::max((5, 5)); // 5x5 tensor of maximum f32 values on GPU
let gpu_min = CudaTensor::<i32>::min((10, 1)); // 10x1 tensor of minimum i32 values on GPU

let cpu_tensor = CpuTensor::<f64>::from_buf(vec![1.0, 2.0, 3.0, 4.0], (2, 2));
let gpu_tensor = cpu_tensor.cuda();
let cpu2 = gpu_tensor.cpu();

let buf = vec![0.0f32; 16];
let tensor = TensorBase::<f32, Cuda>::from_buf(buf, (4, 4));

Arithmetic Operations

let mut a = CpuTensor::<f32>::ones((2, 2));
let view = a.view_mut();

a*= 3.0; 
a+= 2.0;
a-= 1.0;

let b: CpuTensor::<f32> = a.view() * 2.0; 

Slicing Syntax

// Basic range slicing
// slice(dimension, range), so modifications happen on the dimension specified
tensor.slice(0, 2..5)           // Select elements at indices 2, 3, 4
tensor.slice(0, 0..3)           // First 3 elements
tensor.slice(0, ..5)            // Elements from start to index 5 (exclusive)
tensor.slice(0, 2..)            // Elements from index 2 to end
tensor.slice(0, ..)             // All elements (full dimension)

// Inclusive ranges
tensor.slice(0, 2..=5)          // Select elements 2, 3, 4, 5 (inclusive end)
tensor.slice(0, 0..=2)          // First 3 elements (indices 0, 1, 2)

// Single index slicing
tensor.slice(0, 3)              // Select only element at index 3 (reduces dimension)

// Auto-reverse: reversed ranges automatically use negative step
tensor.slice(0, 5..2)           // Elements 5, 4, 3 (auto step=-1)
tensor.slice(0, 9..=0)          // Elements 9, 8, 7, ..., 1, 0 (auto step=-1)

// Explicit negative step using Slice builder
tensor.slice(0, Slice::from(..).step(-1))     // Reverse entire dimension
tensor.slice(0, Slice::from(8..).step(-1))    // From index 8 to start, reversed
tensor.slice(0, Slice::from(..5).step(-1))    // last element to index 5, reversed

// Custom step values
tensor.slice(0, Slice::from(..).step(2))      // Every other element (0, 2, 4, 6, ...)
tensor.slice(0, Slice::from(..).step(-2))     // Every other element, reversed
tensor.slice(0, Slice::from(1..8).step(3))    // Elements at 1, 4, 7

// manual slice construction
tensor.slice(0, Slice::new(Some(8), Some(2), Some(-2)))  // From 8 to 2, step -2: [8, 6, 4]
tensor.slice(0, Slice::new(None, None, Some(-1)))        // Full reverse
tensor.slice(0, Slice::new(Some(5), None, Some(1)))      // From 5 to end

// Edge cases
tensor.slice(0, Slice::from(1..3).step(-1))   // Empty slice (start < end with negative step)
tensor.slice(0, Slice::from(5..5))            // Empty slice (start == end)

Broadcasting

Broadcasting Rules

1. If tensors have different ranks, prepend 1s to the shape of the smaller rank tensor
2. For each dimension, the sizes must either:
   • Be equal, OR
   • One of them must be 1 (which gets broadcasted)

Basic Broadcasting Examples

// Scalar to any shape - broadcasts everywhere
let scalar = CpuTensor::<f32>::from_buf(vec![5.0], vec![]).unwrap();
let matrix = CpuTensor::<f32>::ones((3, 4));
let result = scalar.view() + matrix.view();  // Shape: (3, 4), all values are 6.0

// Vector to matrix - broadcasts along rows
let vector = CpuTensor::<f32>::from_buf(vec![1.0, 2.0, 3.0], vec![3]).unwrap();
let matrix = CpuTensor::<f32>::ones((4, 3));
let result = vector.view() + matrix.view();  // Shape: (4, 3), each row is [2, 3, 4]

// Row vector + Column vector = Matrix
let row = CpuTensor::<f32>::ones((1, 4));    // Shape: (1, 4)
let col = CpuTensor::<f32>::ones((3, 1));    // Shape: (3, 1)
let result = row.view() + col.view();         // Shape: (3, 4), all 2.0

// Matrix broadcasts to 3D tensor
let matrix = CpuTensor::<f32>::ones((3, 4));      // Shape: (3, 4)
let tensor_3d = CpuTensor::<f32>::ones((2, 3, 4)); // Shape: (2, 3, 4)
let result = matrix.view() + tensor_3d.view();     // Shape: (2, 3, 4)

// Singleton dimensions broadcast
let a = CpuTensor::<f32>::ones((1, 3, 1));    // Shape: (1, 3, 1)
let b = CpuTensor::<f32>::ones((2, 3, 4));    // Shape: (2, 3, 4)
let result = a.view() + b.view();              // Shape: (2, 3, 4)

// 4D Broadcasting
let vector = CpuTensor::<f32>::from_buf(vec![10.0, 20.0], vec![2]).unwrap();
let tensor_4d = CpuTensor::<f32>::ones((3, 4, 5, 2));
let result = vector.view() + tensor_4d.view();  // Shape: (3, 4, 5, 2)

// Complex singleton pattern
let a = CpuTensor::<f32>::ones((1, 1, 5, 1));  // Shape: (1, 1, 5, 1)
let b = CpuTensor::<f32>::ones((2, 3, 5, 4));  // Shape: (2, 3, 5, 4)
let result = a.view() + b.view();               // Shape: (2, 3, 5, 4)

// Classic: row vector + column vector
let row = CpuTensor::<f32>::ones((1, 4));      // Shape: (1, 4)
let col = CpuTensor::<f32>::ones((3, 1));      // Shape: (3, 1)
let result = row.view() + col.view();           // Shape: (3, 4)

// 3D mutual broadcasting
let a = CpuTensor::<f32>::ones((1, 4, 5));     // Shape: (1, 4, 5)
let b = CpuTensor::<f32>::ones((3, 1, 5));     // Shape: (3, 1, 5)
let result = a.view() + b.view();               // Shape: (3, 4, 5)

// Complex 4D mutual broadcasting
let a = CpuTensor::<f32>::ones((1, 3, 1, 5));  // Shape: (1, 3, 1, 5)
let b = CpuTensor::<f32>::ones((2, 1, 4, 1));  // Shape: (2, 1, 4, 1)
let result = a.view() + b.view();               // Shape: (2, 3, 4, 5)