User Space Driver - DigitalMediaProfessionals/dv-sdk GitHub Wiki

This chapter describes low-level User-space API for working with AI processor, for high level API for inference of neural networks trained using Caffe or Keras see the Tool.

• Functions for interaction with the kernel module which issues hardware-specific commands to the device
• Functions for allocating memory accessible to the device

1. Create context, see dmp_dv_context_create.
2. Allocate memory for input/output and weights, see dmp_dv_mem_alloc.
3. Map memory to user virtual address space, see dmp_dv_mem_map.
4. Fill weights:
   • Call dmp_dv_mem_sync_start
   • Fill weights buffer with packed weights (pack them in advance with the tool or at runtime using Driver API)
   • Call dmp_dv_mem_sync_end.
5. Optionally unmap weights buffer from user virtual address space, see dmp_dv_mem_unmap.
6. Create command list, see dmp_dv_cmdlist_create.
7. Add the commands to command list with dmp_dv_cmdlist_add_raw.
8. Commit command list, see dmp_dv_cmdlist_commit.
9. Fill input:
   • Call dmp_dv_mem_sync_start
   • Fill the input in WHC8 format: {width, height, first 8 channels}, {width, height, second 8 channels}, ...
   • Call dmp_dv_mem_sync_end.
10. Execute command list, see dmp_dv_cmdlist_exec.
11. Wait for the command list completion, see dmp_dv_cmdlist_wait.
12. Read output:
    • Call dmp_dv_mem_sync_start
    • Examine an output (will be in WHC8 format)
    • Call dmp_dv_mem_sync_end.
13. Process next input if needed: step 9.
14. Release command list, see dmp_dv_cmdlist_release.
15. Release allocated memory, see dmp_dv_mem_release.
16. Release context, see dmp_dv_context_release.

1. #include <dmpdv.h>
2. Link with:
   • if using C++: -ldmpdv
   • if using C: -ldmpdv -lstdc++