Module kernel 

GPU compute kernels.

WGSL-based compute shaders for tensor operations. All kernels operate on Buffer and require a GpuContext.

Categories

• activation — relu, sigmoid.
• arithmetic — add, sub, mul, div, rem, pow (and scalar variants).
• initializer — fill.
• linalg — gemm, transpose.
• reduction — sum.
• shape — broadcast_rows.

Functions

add

Adds two buffers element-wise.

add_scalar

Adds a scalar to each element of a buffer.

broadcast_rows

Broadcasts a vector to all rows of a matrix.

div

Divides two buffers element-wise.

div_scalar

Divides each element of a buffer by a scalar.

fill

Fills a buffer with a constant value.

gemm

Performs general matrix multiplication.

mul

Multiplies two buffers element-wise.

mul_scalar

Multiplies each element of a buffer by a scalar.

pow

Raises elements to a power element-wise.

pow_scalar

Raises each element of a buffer to a scalar power.

relu

Applies ReLU activation element-wise.

rem

Computes the remainder of division element-wise.

rem_scalar

Computes the remainder of division by a scalar.

sigmoid

Applies sigmoid activation element-wise.

sub

Subtracts two buffers element-wise.

sub_scalar

Subtracts a scalar from each element of a buffer.

sum

Computes the sum of all elements in a buffer.

sync

Synchronizes GPU operations.

transpose

Transposes a 2D matrix.