Docs / nablatensor-engine-cpu / com.nablatensor.engine.jit
interface
JitKernelF32
Single-precision counterpart of JitKernel. Arithmetic runs in float; the transcendentals still go through Math in double and narrow, as the JVM has no float exp/ log. Per-scenario value and gradient totals are accumulated by JitReplay in double.
Methods
float forward(float[] v, float[] in, float[] draws, float[] scratch)
void reverse(float[] v, float[] d, float[] scratch, float[] draws)