IP Definition

Convolution Operation

The accelerator performs a simple 2D-convolution. Specifically, given an input array x[n0,n1] and kernel ker[k0,k1], the function computes the 2D convolution:

y[n0,n1] = \sum_{i0,i1} ker[i0,i1]*x[n0+i1,n0+i2]

For simplicity in this example, the accelerator makes the following restrictions:

• Input x and kernel ker are fixed point with constant total bits and fractional bits
• Only a single input and output channel are supported
• Only same mode is supported so that y has the same shape as the x.
• In the current implementation, the hardware supports images up to ` max_nrow x max_ncols = 512 x 512 pixels and kernels up to max_kernel_size x max_kernel_size = 4 x 4`. Kernels are restricted to be square.

Interface Definition

The kernel definion can be found in conv2d/conv2d.cpp and has three

• mem: A shared memory interface to retrieve the kernel and input data, and to retrieve the output data
• in_stream: An AXI4 stream for getting the command data
• out_stream: An AXI4 strem for transmitting back the response at the end of the operation
• debug_stream: An AXI4 stream for outputting events during the processing. This stream is used for debugging and profiling

Interface protocol

The operation with a test bench (TB) is as follows:

• TB sends a Conv2DCmd message with:
  • a transaction ID,
  • read pointer in the shared memory for the kernel and data
  • size information such nrows and ncols
  • write pointer in the shared memory for the output
• Kernel fetches the input and kernel data, computes the output and writes the output data to shared memory.
  • As the kernel is processing, it sends Conv2DDebug messages indicating which stage of processing (load, compute or store) the kernel is executing and the row index. This information can be used for profiling
• When complete, the kernels sends a Conv2DResp message to the TB that echoes the transaction ID and also provides error codes, if any.

All messages are defined with the PySilicon DataList class to enable auto-generation of the related Vitis include files. See conv2d_demo.py for details of the schema definitions.