從cuda 7.5開始引入原生fp16（Tegra X1是第一塊支持的GPU https://gcc.gnu.org/onlinedocs/gcc/Half-Precision.html），實現了IEEE754標准中的半精度浮點型；
cuda使用half基本數據類型和half2結構體支持，需要引用cuda_fp16.h
Mixed Precision Performance on Pascal GPUs
The half precision (FP16) Format is not new to GPUs. In fact, FP16 has been supported as a storage format for many years on NVIDIA GPUs, mostly used for reduced precision floating point texture storage and filtering and other special-purpose operations. The Pascal GPU architecture implements general-purpose, IEEE 754 FP16 arithmetic. High performance FP16 is supported at full speed on Tesla P100 (GP100), and at lower throughput (similar to double precision) on other Pascal GPUs (GP102, GP104, and GP106), as the following table shows.
CUDA使用FP16進行半精度運算 - CSDN博客
Nvidia GPU的浮點計算能力(FP64/FP32/FP16) - CSDN博客
讓Faster R-CNN支持TX1的fp16(half float, float16)特性 - CSDN博客
CUDA Samples :: CUDA Toolkit Documentation
0_Simple/fp16ScalarProduct
caffe官方將會支持TX1的fp16的特性嗎？
https://www.zhihu.com/question/39715684
https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5
ARM支持fp16
https://blog.csdn.net/hunanchenxingyu/article/details/47003279
http://gcc.gnu.org/onlinedocs/gcc/Half-Precision.html
https://blog.csdn.net/tanli20090506/article/details/71435777
https://blog.csdn.net/soaringlee_fighting/article/details/78885394
https://developer.arm.com/technologies/neon/intrinsics
https://developer.arm.com/technologies/floating-point
https://blog.csdn.net/softee/article/details/79494335
https://blog.csdn.net/qq_18229381/article/details/71104059
http://half.sourceforge.net/
https://blog.csdn.net/cubesky/article/details/51793525
https://gcc.gnu.org/onlinedocs/gcc/Half-Precision.html
https://en.wikipedia.org/wiki/F16C
https://en.wikipedia.org/wiki/Half-precision_floating-point_format
https://docs.microsoft.com/en-us/windows/desktop/api/directxpackedvector/nf-directxpackedvector-xmconvertfloattohalf
https://software.intel.com/en-us/node/524287
https://software.intel.com/en-us/node/524286