Video and Vision Processing Suite Intel® FPGA IP User Guide

ID 683329
Date 10/02/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents
1. About the Video and Vision Processing Suite 2. Getting Started with the Video and Vision Processing IPs 3. Video and Vision Processing IPs Functional Description 4. Video and Vision Processing IP Interfaces 5. Video and Vision Processing IP Registers 6. Video and Vision Processing IPs Software Programming Model 7. Protocol Converter Intel® FPGA IP 8. 3D LUT Intel® FPGA IP 9. AXI-Stream Broadcaster Intel® FPGA IP 10. Bits per Color Sample Adapter Intel FPGA IP 11. Chroma Key Intel® FPGA IP 12. Chroma Resampler Intel® FPGA IP 13. Clipper Intel® FPGA IP 14. Clocked Video Input Intel® FPGA IP 15. Clocked Video to Full-Raster Converter Intel® FPGA IP 16. Clocked Video Output Intel® FPGA IP 17. Color Space Converter Intel® FPGA IP 18. Deinterlacer Intel® FPGA IP 19. FIR Filter Intel® FPGA IP 20. Frame Cleaner Intel® FPGA IP 21. Full-Raster to Clocked Video Converter Intel® FPGA IP 22. Full-Raster to Streaming Converter Intel® FPGA IP 23. Genlock Controller Intel® FPGA IP 24. Generic Crosspoint Intel® FPGA IP 25. Genlock Signal Router Intel® FPGA IP 26. Guard Bands Intel® FPGA IP 27. Interlacer Intel® FPGA IP 28. Mixer Intel® FPGA IP 29. Pixels in Parallel Converter Intel® FPGA IP 30. Scaler Intel® FPGA IP 31. Stream Cleaner Intel® FPGA IP 32. Switch Intel® FPGA IP 33. Tone Mapping Operator Intel® FPGA IP 34. Test Pattern Generator Intel® FPGA IP 35. Video and Vision Monitor Intel FPGA IP 36. Video Frame Buffer Intel® FPGA IP 37. Video Frame Reader Intel FPGA IP 38. Video Frame Writer Intel FPGA IP 39. Video Streaming FIFO Intel® FPGA IP 40. Video Timing Generator Intel® FPGA IP 41. Warp Intel® FPGA IP 42. Design Security 43. Document Revision History for Video and Vision Processing Suite User Guide

38.3. Video Frame Writer IP Functional Description

The IP receives video fields or frames from its Intel FPGA streaming video input and writes them to external memory via its Avalon memory-mapped interface.
  1. Configure the IP with the number of buffers that you want to write to with the csr_num_buffers register.
  2. Program the base address in memory for the first buffer, the offset between buffers and the offset between individual lines within buffers via the csr_buffer_base, csr_inter_buffer_offset and csr_inter_line_offset registers.
  3. Set the csr_overwrite_broken register if you want broken frames to be overwritten and then csr_run to either free-running mode for continuous operation or single-shot mode.
  4. Commit these settings with a write to the csr_commit register.

If the IP receives any packets on its Intel FPGA streaming video input before you configure and commit it, the IP consumes the packets. When consuming, the IP raises axi4s_vid_in_tready.

Register Behavior

Bit [0] of the csr_status register goes high when the IP starts writing the first frame. It goes low after the IP finishes writing the last line of the frame. It returns high when the IP starts writing the next frame.

The IP sets the csr_buffer_available register after it writes the first frame and a buffer is available in memory. csr_buffer_start_address holds the base address of the first frame and csr_buffer_write_count increments with each frame the IP writes. If the IP receives an interlaced f1 field, it sets csr_buffer_f1_flag. The IP sets csr_buffer_field_width and csr_buffer_field_height to the dimensions of the new value. In full mode, the IP extracts the field count from each end of field packet and updatescsr_field_count.

Acknowledge the buffer by writing to the csr_buffer_acknowledge register, which resets csr_buffer_available. If you do not acknowledge the buffer, csr_buffer_available remains set, the IP keeps writing frames and csr_buffer_start_address and associated registers update as normal.

Latency

The frame writer latency depends on the availability of the external memory interfaces, which may create backpressure for the IP’s write interface via the av_mm_mem_read_host_waitrequest signal.

The worst-case latency figures are when the frame writer experiences no backpressure. Any backpressure increases these latencies by the same amount of cycles.

The latency figures are for when the host interfaces do not have a separate clock, so all interfaces operate from the same clock.

The latency figures are for a typical configuration. Reducing the burst target reduces latency but at the cost of decreasing efficiency and increasing bandwidth consumed on the bus.

Table 740.  Latency

The latency figures are the same for both full and lite variants of the IP.

Initiating event Resultant event Latency (measured in clock cycles)
axi4s_vid_in_tuser[0] strobe to indicate start of frame First av_mm_mem_write_host_write strobe of a frame write 44
Figure 86. Latency