This option enables a hardware implementation of the softmax operator. The
hardware softmax accelerator is packaged inside the DPU IP wrapper, but is a
separate accelerator with its own interface and runtime and implements an int8 input
and floating-point output data formats. The hardware implementation of softmax can
be up to 160 times faster than a software implementation on MPSoC devices. Users can
enable this option if their networks/models include a softmax layer and they wish to
improve throughput.
Note: The hardware softmax
can support up to 1023 classes. If the number of classes is greater than 1023, a
software implementation of softmax may be considered. For more information,
refer to the
Vitis AI Library User Guide (UG1354)
.
When hardware softmax is enabled, an AXI master interface named SFM_M_AXI and
an interrupt port named sfm_interrupt are added to
the DPU IP wrapper. The softmax module uses m_axi_dpu_aclk as the source clock for SFM_M_AXI as well as for
computation.
The additional resources utilized when hardware softmax acceleration is enabled are listed in the following table.
| IP Name | Extra LUTs | Extra FFs | Extra BRAMs | Extra DSPs |
|---|---|---|---|---|
| Softmax | 9580 | 8019 | 4 | 14 |