index 7036d8dff87c1169aaf00b37be9cab339e1bc433..7ba9bc7d3761368f0d9b5fae8b5b7a2eeb9472d6 100644 (file)
--- a/docs/source/example.rst
+++ b/docs/source/example.rst
We ship three end-to-end examples within the tidl-api package
to demonstrate three categories of deep learning networks. The first
-two examples can run on AM57x SoCs with either EVE or DSP. The last
+two examples can run on AM57x SoCs with either EVE or DSP devices. The last
example requires AM57x SoCs with both EVE and DSP. The performance
numbers that we present here were obtained on an AM5729 EVM, which
includes 2 ARM A15 cores running at 1.5GHz, 4 EVE cores at 535MHz, and
2 DSP cores at 750MHz.
+For each example, we report device processing time, host processing time,
+and TIDL API overhead. **Device processing time** is measured on the device,
+from the moment processing starts for a frame till processing finishes.
+**Host processing time** is measured on the host, from the moment
+``ProcessFrameStartAsync()`` is called till ``ProcessFrameWait()`` returns
+in user application. It includes the TIDL API overhead, the OpenCL runtime
+overhead, and the time to copy user input data into padded TIDL internal
+buffers.
+
Imagenet
--------
The imagenet example takes an image as input and outputs 1000 probabilities.
Each probability corresponds to one object in the 1000 objects that the
-network is pre-trained with. Our example outputs top 5 probabilities
+network is pre-trained with. Our example outputs top 5 predictions
as the most likely objects that the input image can be.
The following figure and tables shows an input image, top 5 predicted
.. table::
- ==== ============== ============
- Rank Object Classes Probability
- ==== ============== ============
- 1 tabby 0.996
- 2 Egyptian_cat 0.977
- 3 tiger_cat 0.973
- 4 lynx 0.941
- 5 Persian_cat 0.922
- ==== ============== ============
+ ==== ==============
+ Rank Object Classes
+ ==== ==============
+ 1 tabby
+ 2 Egyptian_cat
+ 3 tiger_cat
+ 4 lynx
+ 5 Persian_cat
+ ==== ==============
.. table::
The particular network that we ran in this category, jacintonet11v2,
has 14 layers. User can specify whether to run the network on EVE or DSP
for acceleration. We can see that EVE time is slightly higher than DSP time.
-Host time includes the OpenCL runtime overhead and the time to copy user
-input data into padded TIDL buffers. We can see that the overall overhead
-is less than 1.5%.
+We can also see that the overall overhead is less than 1.5%.
+
+.. note::
+ The predicitions reported here are based on the output of the softmax
+ layer in the network, which are not normalized to the real probabilities.
Segmentation
------------
DSP: 21.1 ms 22.3 ms 5.62 %
====================== ==================== ============
+Test
+----
+This example is used to test pre-converted networks included in the TIDL API package (``test/testvecs/config/tidl_models``). When run without any arguments, the program ``test_tidl`` will run all available networks on the C66x DSPs and EVEs available on the SoC. Use the ``-c`` option to specify a single network. Run ``test_tidl -h`` for details.
+
Running Examples
----------------
-The examples are located in ``/usr/share/ti/tidl-api/examples`` on
+The examples are located in ``/usr/share/ti/tidl/examples`` on
the EVM file system. Each example needs to be run its own directory.
Running an example with ``-h`` will show help message with option set.
The following code section shows how to run the examples, and