DensePanelKernels
Arithmetic over contiguous matrix columns and panels. Callers must supply non-negative lengths and valid array windows; implementations need not validate them. A zero length is legal and must not read an array. Unlike standalone BLAS AXPY, update operations in this family evaluate zero multipliers so matrix arithmetic preserves products such as 0 * infinity.
Functions
axpy4
Adds four scaled, equally spaced matrix columns into y. For every i in the run, the operation is y[yOff + i] += c0*a[aOff + i] + c1*a[aOff + stride + i] + c2*a[aOff + 2*stride + i] + c3*a[aOff + 3*stride + i], evaluated in coefficient order. Zero coefficients are evaluated.
axpyArithmetic
Adds alpha * x[xOff + i] into y[yOff + i] over len entries and always evaluates the product, including when alpha is zero. Equal input and destination runs are safe; callers must otherwise avoid overlap whose stores precede later input loads.
dot4
Writes four dot products against one shared right operand into out. For r in 0..3, out[outOff + r] becomes the dot product of the len entries beginning at aOff + r * stride and bOff. The four outputs are written only after their input reductions are complete.
dotAxpy
Returns the dot product of the a and x runs while adding alpha * a into y in the same pass. Each a and x element is loaded before the corresponding destination is stored, so an input may equal the destination run. Callers must otherwise avoid overlap whose stores precede later input loads. Zero alpha is evaluated.