Skip to content

Latest commit

 

History

History
59 lines (50 loc) · 2.74 KB

CompilerFeatures.md

File metadata and controls

59 lines (50 loc) · 2.74 KB

DLA Compiler

Layers and features support

Layer                                         Feature                                     FP16                   INT8                  
Convolution
Dilation
Winograd Not implemented in SW
Deconvolution
With padding Not implemented in SW Not implemented in SW
Winograd Not implemented in SW Not implemented in SW
Fully Connected
Winograd Not implemented in SW Not implemented in SW
Group Convolution Not implemented in SW
Winograd Not implemented in SW
Pooling
Max
Min
Avg
Inclusive padding
Exclusive padding Not supported in HW Not supported in HW
Activation
Bias
BatchNorm
Scale
Sigmoid Not implemented in SW
Tanh Not implemented in SW
EltWise SUM
EltWise SUB Not supported in HW Not supported in HW
EltWise MIN Not implemented in SW
EltWise MAX Not implemented in SW
LRN Not implemented in SW

Frameworks support

Framework                               Status                                      
Caffe
ONNX Future

Networks verification report

Network                                   Configuration                           fp16                   int8                  
MNIST nv_full,nv_large,nv_small Verified Verified
ResNet-18 nv_full,nv_large,nv_small Verified Verified
ResNet-50 nv_full,nv_large,nv_small Verified Verified

Known limitations

  • Not supported in HW
    • Dilation with Winograd
    • EltWise SUB
    • Pooling and convolution layers where pad size is greater than kernel size
  • Not implemented in SW
    • Deconvolution with strides > 32
    • Deconvolution with input/output padding