Add ssd_multibox_fullnet example

- To demonstrate running jdenet/jdetnet_voc on a single core,
  without paritioning the network.  This is useful for situations
  where SoC only has C66x cores but not EVE cores.
- MCT-1202

diff --git a/docs/source/changelog.rst b/docs/source/changelog.rst
index e69b6e44dbf1508cd5d1f76a9bc73280f16c5b4e..532a436c6dc7d23156d351f128d68e0b970476f9 100644 (file)
--- a/docs/source/changelog.rst
+++ b/docs/source/changelog.rst
@@ -30,7 +30,7 @@ Changelog
     Enable a two phase approach to generating execution graphs. Use the
     following API function to enable timestamp generation:
 
-    .. code::
+    .. code:: cpp
 
         bool EnableTimeStamps(const std::string& file = "timestamp.log", size_t num_frames=32);
 
@@ -38,7 +38,7 @@ Changelog
 
 #. Added Python 3 bindings for TIDL API. See the ``examples/pybind`` directory for examples of using the Python bindings. Set PYTHONPATH to the location of ``tidl.so``.
 
-   .. code::
+   .. code:: bash
 
         root@am57xx-evm:~# export PYTHONPATH=/home/root/tidl-api/tidl_api
         root@am57xx-evm:~# python3

diff --git a/docs/source/example.rst b/docs/source/example.rst
index ff70804267e5179aab6316088b66020e00a8a4a5..9c696ad8b6cc0f3bf301cc754c7f0747ec640db9 100644 (file)
--- a/docs/source/example.rst
+++ b/docs/source/example.rst
@@ -35,13 +35,13 @@ Examples
      - OpenCV used to read input image from file or capture from camera.
    * - ssd_multibox
      - Object detection
-     - EVE and C66x (network is split across both EVE and C66x)
+     - EVE and C66x (network is split across both EVE and C66x), EVE or C66x (full network on each core)
      - OpenCV used to read input image from file or capture from camera.
    * - mnist
      - handwritten digits recognition (MNIST).  This example illustrates
        low TIDL API overhead (~1.8%) for small networks with low compute
        requirements (<5ms).
-     - EVE
+     - EVE or C66x
      - Pre-processed white-on-black images read from file, with or without
        MNIST database file headers.
    * - classification
@@ -273,6 +273,10 @@ versus ExecutionObject level.
      - 1630 ms
      - 1408 ms
 
+When there is a requirement to run the SSD networks non-partitioned,
+for example, the SoC only has C66x cores but not EVE cores,
+use ``-e 0`` to run the full network only on C66x cores, without partitioning.
+
 .. _mnist-example:
 
 MNIST

diff --git a/docs/source/faq/out_of_memory.rst b/docs/source/faq/out_of_memory.rst
index ee97e58eb00d5b66cc7c3afda189aafee94934a6..d7c8e1f89c2a00da0f7fdee2cbf5bf833e9582e3 100644 (file)
--- a/docs/source/faq/out_of_memory.rst
+++ b/docs/source/faq/out_of_memory.rst
@@ -4,7 +4,7 @@ Why do I get an assertion failure from malloc_ddr?
 
 Application execution fails with the following error message:
 
-.. code:: shell
+.. code:: console
 
    tidl: device_alloc.h:31: T* tidl::malloc_ddr(size_t) [with T = char; size_t = unsigned int]: Assertion `val != nullptr' failed
 
@@ -28,7 +28,7 @@ Insufficient OpenCL global memory
 +++++++++++++++++++++++++++++++++
 Another possible reason is that total memory requirement specified in the ``Configuration`` using NETWORK_HEAP_SIZE and PARAM_HEAP_SIZE exceeds default memory available for OpenCL.  Follow the instructions below to increase the amount of CMEM (contiguous memory available for OpenCL) from 192MB (0xc000000) to 384MB (0x18000000):
 
-.. code:: bash
+.. code:: diff
 
    $ sudo apt-get install device-tree-compiler # In case dtc is not already installed
    $ scp root@am57:/boot/am57xx-evm-reva3.dtb .

diff --git a/examples/ssd_multibox/main.cpp b/examples/ssd_multibox/main.cpp
index 315834501227ccf166ce54cb29b96136d98fa6ea..93f4c89ab5b3aa94d4e36def357d3a7f4dbe1123 100644 (file)
--- a/examples/ssd_multibox/main.cpp
+++ b/examples/ssd_multibox/main.cpp
@@ -102,9 +102,9 @@ int main(int argc, char *argv[])
     // If there are no devices capable of offloading TIDL on the SoC, exit
     uint32_t num_eves = Executor::GetNumDevices(DeviceType::EVE);
     uint32_t num_dsps = Executor::GetNumDevices(DeviceType::DSP);
-    if (num_eves == 0 || num_dsps == 0)
+    if (num_eves == 0 && num_dsps == 0)
     {
-        cout << "ssd_multibox requires both EVE and DSP for execution." << endl;
+        cout << "ssd_multibox requires EVE or DSP for execution." << endl;
         return EXIT_SUCCESS;
     }
 
@@ -112,8 +112,8 @@ int main(int argc, char *argv[])
     cmdline_opts_t opts;
     opts.config = DEFAULT_CONFIG;
     opts.object_classes_list_file = DEFAULT_OBJECT_CLASSES_LIST_FILE;
-    opts.num_eves = 1;
-    opts.num_dsps = 1;
+    opts.num_eves = num_eves > 0 ? 1 : 0;
+    opts.num_dsps = num_dsps > 0 ? 1 : 0;
     opts.input_file = DEFAULT_INPUT;
     opts.output_prob_threshold = DEFAULT_OUTPUT_PROB_THRESHOLD;
     if (! ProcessArgs(argc, argv, opts))
@@ -121,7 +121,7 @@ int main(int argc, char *argv[])
         DisplayHelp();
         exit(EXIT_SUCCESS);
     }
-    assert(opts.num_dsps != 0 && opts.num_eves != 0);
+    assert(opts.num_dsps != 0 || opts.num_eves != 0);
     if (opts.num_frames == 0)
         opts.num_frames = (opts.is_camera_input || opts.is_video_input) ?
                           NUM_VIDEO_FRAMES :
@@ -164,6 +164,9 @@ bool RunConfiguration(const cmdline_opts_t& opts)
         return false;
     }
     c.enableApiTrace = opts.verbose;
+    if (opts.num_eves == 0 || opts.num_dsps == 0)
+        c.runFullNet = true;
+
     // setup camera/video input
     VideoCapture cap;
     if (! SetVideoInputOutput(cap, opts, "SSD_Multibox"))  return false;
@@ -198,43 +201,81 @@ bool RunConfiguration(const cmdline_opts_t& opts)
         // DSP will run layersGroupId 2 in the network
         Executor* e_eve = CreateExecutor(DeviceType::EVE, opts.num_eves, c, 1);
         Executor* e_dsp = CreateExecutor(DeviceType::DSP, opts.num_dsps, c, 2);
-
-        // Construct ExecutionObjectPipeline that utilizes multiple
-        // ExecutionObjects to process a single frame, each ExecutionObject
-        // processes one layerGroup of the network
-        //
-        // Pipeline depth can enable more optimized pipeline execution:
-        // Given one EVE and one DSP as an example, with different
-        //     pipeline_depth, we have different execution behavior:
-        // If pipeline_depth is set to 1,
-        //    we create one EOP: eop0 (eve0, dsp0)
-        //    pipeline execution of multiple frames over time is as follows:
-        //    --------------------- time ------------------->
-        //    eop0: [eve0...][dsp0]
-        //    eop0:                [eve0...][dsp0]
-        //    eop0:                               [eve0...][dsp0]
-        //    eop0:                                              [eve0...][dsp0]
-        // If pipeline_depth is set to 2,
-        //    we create two EOPs: eop0 (eve0, dsp0), eop1(eve0, dsp0)
-        //    pipeline execution of multiple frames over time is as follows:
-        //    --------------------- time ------------------->
-        //    eop0: [eve0...][dsp0]
-        //    eop1:          [eve0...][dsp0]
-        //    eop0:                   [eve0...][dsp0]
-        //    eop1:                            [eve0...][dsp0]
-        // Additional benefit of setting pipeline_depth to 2 is that
-        //    it can also overlap host ReadFrame() with device processing:
-        //    --------------------- time ------------------->
-        //    eop0: [RF][eve0...][dsp0]
-        //    eop1:     [RF]     [eve0...][dsp0]
-        //    eop0:                    [RF][eve0...][dsp0]
-        //    eop1:                             [RF][eve0...][dsp0]
         vector<ExecutionObjectPipeline *> eops;
-        uint32_t pipeline_depth = 2;  // 2 EOs in EOP -> depth 2
-        for (uint32_t j = 0; j < pipeline_depth; j++)
-            for (uint32_t i = 0; i < max(opts.num_eves, opts.num_dsps); i++)
-                eops.push_back(new ExecutionObjectPipeline(
+
+        if (e_eve != nullptr && e_dsp != nullptr)
+        {
+            // Construct ExecutionObjectPipeline that utilizes multiple
+            // ExecutionObjects to process a single frame, each ExecutionObject
+            // processes one layerGroup of the network
+            //
+            // Pipeline depth can enable more optimized pipeline execution:
+            // Given one EVE and one DSP as an example, with different
+            //     pipeline_depth, we have different execution behavior:
+            // If pipeline_depth is set to 1,
+            //    we create one EOP: eop0 (eve0, dsp0)
+            //    pipeline execution of multiple frames over time is as follows:
+            //    --------------------- time ------------------->
+            //    eop0: [eve0...][dsp0]
+            //    eop0:                [eve0...][dsp0]
+            //    eop0:                               [eve0...][dsp0]
+            //    eop0:                                              [eve0...][dsp0]
+            // If pipeline_depth is set to 2,
+            //    we create two EOPs: eop0 (eve0, dsp0), eop1(eve0, dsp0)
+            //    pipeline execution of multiple frames over time is as follows:
+            //    --------------------- time ------------------->
+            //    eop0: [eve0...][dsp0]
+            //    eop1:          [eve0...][dsp0]
+            //    eop0:                   [eve0...][dsp0]
+            //    eop1:                            [eve0...][dsp0]
+            // Additional benefit of setting pipeline_depth to 2 is that
+            //    it can also overlap host ReadFrame() with device processing:
+            //    --------------------- time ------------------->
+            //    eop0: [RF][eve0...][dsp0]
+            //    eop1:     [RF]     [eve0...][dsp0]
+            //    eop0:                    [RF][eve0...][dsp0]
+            //    eop1:                             [RF][eve0...][dsp0]
+            uint32_t pipeline_depth = 2;  // 2 EOs in EOP -> depth 2
+            for (uint32_t j = 0; j < pipeline_depth; j++)
+                for (uint32_t i = 0; i < max(opts.num_eves, opts.num_dsps); i++)
+                    eops.push_back(new ExecutionObjectPipeline(
                       {(*e_eve)[i%opts.num_eves], (*e_dsp)[i%opts.num_dsps]}));
+        }
+        else
+        {
+            // Construct ExecutionObjectPipeline that utilizes a
+            // ExecutionObject to process a single frame, each ExecutionObject
+            // processes the full network
+            //
+            // Use duplicate EOPs to do double buffering on frame input/output
+            //    because each EOP has its own set of input/output buffers,
+            //    so that host ReadFrame() can overlap device processing
+            // Use one EO as an example, with different buffer_factor,
+            //    we have different execution behavior:
+            // If buffer_factor is set to 1 -> single buffering
+            //    we create one EOP: eop0 (eo0)
+            //    pipeline execution of multiple frames over time is as follows:
+            //    --------------------- time ------------------->
+            //    eop0: [RF][eo0.....][WF]
+            //    eop0:                   [RF][eo0.....][WF]
+            //    eop0:                                     [RF][eo0.....][WF]
+            // If buffer_factor is set to 2 -> double buffering
+            //    we create two EOPs: eop0 (eo0), eop1(eo0)
+            //    pipeline execution of multiple frames over time is as follows:
+            //    --------------------- time ------------------->
+            //    eop0: [RF][eo0.....][WF]
+            //    eop1:     [RF]      [eo0.....][WF]
+            //    eop0:                   [RF]  [eo0.....][WF]
+            //    eop1:                             [RF]  [eo0.....][WF]
+            uint32_t buffer_factor = 2;  // set to 1 for single buffering
+            for (uint32_t j = 0; j < buffer_factor; j++)
+            {
+                for (uint32_t i = 0; i < opts.num_eves; i++)
+                    eops.push_back(new ExecutionObjectPipeline({(*e_eve)[i]}));
+                for (uint32_t i = 0; i < opts.num_dsps; i++)
+                    eops.push_back(new ExecutionObjectPipeline({(*e_dsp)[i]}));
+            }
+        }
         uint32_t num_eops = eops.size();
 
         // Allocate input/output memory for each EOP
@@ -252,9 +293,7 @@ bool RunConfiguration(const cmdline_opts_t& opts)
 
             // Wait for previous frame on the same eop to finish processing
             if (eop->ProcessFrameWait())
-            {
                 WriteFrameOutput(*eop, c, opts, (float)prob_slider);
-            }
 
             // Read a frame and start processing it with current eo
             if (ReadFrame(*eop, frame_idx, c, opts, cap, ifs))
@@ -291,7 +330,9 @@ Executor* CreateExecutor(DeviceType dt, uint32_t num, const Configuration& c,
     for (uint32_t i = 0; i < num; i++)
         ids.insert(static_cast<DeviceId>(i));
 
-    return new Executor(dt, ids, c, layers_group_id);
+    Executor* e = new Executor(dt, ids, c, layers_group_id);
+    assert(e != nullptr);
+    return e;
 }
 
 bool ReadFrame(ExecutionObjectPipeline& eop, uint32_t frame_idx,