Detailed definition of tensorrt data format

TensorRT Detailed definition of data format

Click here to join NVIDIA Developer Program

TensorRT Support different data formats . There are two aspects to consider ： Data type and layout .

Data type format

The data type is the representation of each individual value . Its size determines the range of values and the accuracy of the representation , Namely FP32（32 Bit floating point , Or single precision ）,FP16（16 Bit floating point or half precision ）,INT32（32 Bit integers represent ）, and INT8（8 Who said ）.

Layout format

The layout format determines the order in which values are stored . Usually ,batch Dimension is the leftmost dimension , Other dimensions refer to aspects of each data item , For example, in the image C It's a channel , H It's height , W It's the width . Ignore the batch size always before these , C 、 H and W Usually sorted as CHW （ See chart 1 ） or HWC （ See chart 2 ）.

chart 1. CHW The layout of ： The image is divided into HxW matrix , One per channel , The matrix is stored sequentially ; All values of the channel are stored continuously .

chart 2. HWC The layout of ： Images are stored as a single file HxW matrix , The value is actually C Tuples , Each channel has a value ; One point （ Pixels ） All values of are stored continuously .

For faster computing , More formats are defined to package channel values together and use reduced precision . therefore ,TensorRT And support NC / 2HW2 and NHWC8 Equiform .

stay NC / 2HW2 ( TensorFormat::kCHW2 ) in , Channel value pairs are in each HxW Packed together in a matrix （ Null in case of odd channels ）. The result is a format , among ⌈ C/2 ⌉ HxW The value of the matrix is the value pair of two consecutive channels （ See chart 3 ）; Please note that , If they are in the same pair , Then this sort interleaves the dimensions into steps 1 Value of the channel of , Otherwise, the step size will be 2xHxW .

The following figure shows a pair of channel values in each HxW Packed together in a matrix . The result is a format , among [C/2] HxW The value of the matrix is the value pair of two consecutive channels

stay NHWC8 ( TensorFormat::kHWC8 ) in , HxW The entries of the matrix include the values of all channels （ See chart 4 ）. Besides , These values are packaged in ⌈ C/8 ⌉ 8 Tuple , also C Round up to the nearest 8 Multiple .

The following figure is in this NHWC8 In the format , HxW The entries of the matrix include the values of all channels .

other TensorFormat Follow the above mentioned TensorFormat::kCHW2 and TensorFormat::kHWC8 Similar rules .