Macro DEVICE_PICK_TIMED

#include <std/ml/device.h> source

#define DEVICE_PICK_TIMED 0

How a 16-bit precision picks per shape: DEVICE_PICK_TIMED (the default) times each new convolution / matmul shape in 16 bits against fp32 and keeps the faster; DEVICE_PICK_ALWAYS runs every one it can in 16 bits (tests, or to measure). Returns the previous mode.