本文主要是介绍fp16半精度浮点数转成float类型------C/C++,希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!
在深度学习算法模型推理时,会遇到fp16类型,但是我们的c语言中没有这种类型,直接转成unsigned short又会丧失精度,因此我们首先将FP16转成float类型,再进行计算。
方法1:
typedef unsigned short ushort;//占用2个字节
typedef unsigned int uint; //占用4个字节uint as_uint(const float x) {return *(uint*)&x;
}
float as_float(const uint x) {return *(float*)&x;
}float half_to_float(const ushort x) { // IEEE-754 16-bit floating-point format (without infinity): 1-5-10, exp-15, +-131008.0, +-6.1035156E-5, +-5.9604645E-8, 3.311 digitsconst uint e = (x&0x7C00)>>10; // exponentconst uint m = (x&0x03FF)<<13; // mantissaconst uint v = as_uint((float)m)>>23; // evil log2 bit hack to count leading zeros in denormalized formatreturn as_float((x&0x8000)<<16 | (e!=0)*((e+112)<<23|m) | ((e==0)&(m!=0))*((v-37)<<23|((m<<(150-v))&0x007FE000))); // sign : normalized : denormalized
}
ushort float_to_half(const float x) { // IEEE-754 16-bit floating-point format (without infinity): 1-5-10, exp-15, +-131008.0, +-6.1035156E-5, +-5.9604645E-8, 3.311 digitsconst uint b = as_uint(x)+0x00001000; // round-to-nearest-even: add last bit after truncated mantissaconst uint e = (b&0x7F800000)>>23; // exponentconst uint m = b&0x007FFFFF; // mantissa; in line below: 0x007FF000 = 0x00800000-0x00001000 = decimal indicator flag - initial roundingreturn (b&0x80000000)>>16 | (e>112)*((((e-112)<<10)&0x7C00)|m>>13) | ((e<113)&(e>101))*((((0x007FF000+m)>>(125-e))+1)>>1) | (e>143)*0x7FFF; // sign : normalized : denormalized : saturate
}
方法2:
float cpu_half2float(unsigned short x)
{unsigned sign = ((x >> 15) & 1);unsigned exponent = ((x >> 10) & 0x1f);unsigned mantissa = ((x & 0x3ff) << 13);if (exponent == 0x1f) { /* NaN or Inf */mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);exponent = 0xff;} else if (!exponent) { /* Denorm or Zero */if (mantissa) {unsigned int msb;exponent = 0x71;do {msb = (mantissa & 0x400000);mantissa <<= 1; /* normalize */--exponent;} while (!msb);mantissa &= 0x7fffff; /* 1.mantissa is implicit */}} else {exponent += 0x70;}int temp = ((sign << 31) | (exponent << 23) | mantissa);return *((float*)((void*)&temp));
}
3 demo
下面的demo中,yolov5_outputs[0].buf是void *类型的,void *类型不能++,因此先转换成ushort*类型。
...... float *data0 = (float*)malloc(4 * output_attrs[0].n_elems);float *data1 = (float*)malloc(4 * output_attrs[1].n_elems);float *data2 = (float*)malloc(4 * output_attrs[2].n_elems);unsigned short *temp0 = (ushort*)yolov5_outputs[0].buf;unsigned short *temp1 = (ushort*)yolov5_outputs[1].buf;unsigned short *temp2 = (ushort*)yolov5_outputs[2].buf;for(int i=0; i < output_attrs[0].n_elems;i++){data0[i] = half_to_float(temp0[i]);}for(int i=0; i < output_attrs[1].n_elems;i++){data1[i] = half_to_float(temp1[i]);}for(int i=0; i < output_attrs[2].n_elems;i++){data2[i] = half_to_float(temp2[i]);}......
参考文献:
https://github.com/PrincetonVision/marvin/blob/master/tools/tensorIO_matlab/half2float.cpp
这篇关于fp16半精度浮点数转成float类型------C/C++的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!