一、支持的网络
Support most commonly used CNN network
支持大部分常用的 CNN 网络
Classical CNN Network: VGG AlexNet GoogleNet Inception …
Practical CNN Network: ResNet DenseNet SENet FPN …
Light-weight CNN Network: SqueezeNet MobileNetV1/V2 ShuffleNetV1/V2 MNasNet …
Detection Network: MTCNN facedetection …
Detection Network: VGG-SSD MobileNet-SSD SqueezeNet-SSD MobileNetV2-SSDLite …
Detection Network: Faster-RCNN R-FCN …
Detection Network: YOLOV2 YOLOV3 MobileNet-YOLOV3 …
Segmentation Network: FCN PSPNet …
二、加速技巧
1)使用低精度
半精度浮点数类型, 浮点数转化8bit量化
// convert float to half precision floating point
static unsigned short float2half(float value)
{
// 1 : 8 : 23
union
{
unsigned int u;
float f;
} tmp;
tmp.f = value;
// 1 : 8 : 23
unsigned short sign = (tmp.u & 0x80000000) >> 31;
unsigned short exponent = (tmp.u & 0x7F800000) >> 23;
unsigned int significand = tmp.u & 0x7FFFFF;
// fprintf(stderr, "%d %d %d\n", sign, exponent, significand);
// 1 : 5 : 10
unsigned short fp16;
if (exponent == 0)
{
// zero or denormal, always underflow
fp16 = (sign << 15) | (0x00 << 10) | 0x00;
}
else if (exponent == 0xFF)
{
// infinity or NaN
fp16 = (sign << 15) | (0x1F << 10) | (significand ? 0x200 : 0x00);
}
else
{
// normalized
short newexp = exponent + (- 127 + 15);
if (newexp >= 31)
{
// overflow, return infinity
fp16 = (sign << 15) | (0x1F << 10) | 0x00;
}
else if (newexp <= 0)
{
// underflow
if (newexp >= -10)
{
// denormal half-precision
unsigned short sig = (significand | 0x800000) >> (14 - newexp);
fp16 = (sign << 15) | (0x00 << 10) | sig;
}
else
{
// underflow
fp16 = (sign << 15) | (0x00 << 10) | 0x00;
}
}
else
{
fp16 = (sign << 15) | (newexp << 10) | (significand >> 13);
}
}
return fp16;
}
2)openmp多核多线程加速
#pragma omp parallel for num_threads(opt.num_threads)
for (int q=0; q<channels; q++)
{
float* ptr = bottom_top_blob.channel(q);
for (int i=0; i<size; i++)
{
if (ptr[i] < 0)
ptr[i] = -ptr[i];
}
}