Macro DEVICE_FP32

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

#define DEVICE_FP32 0

Precisions (device_set_precision): what convolutions and matrix products read on a device. 16-bit modes keep values, weights, gradients and the optimizer state in fp32, feed these ops 16-bit copies and accumulate in fp32; fp16 also scales the loss (device_loss_scale).