Quartus® Prime Pro Edition User Guide: Design Recommendations

ID 683082
Date 7/08/2024
Public
Document Table of Contents

1.3.2. Inferring Multiply-Accumulator and Multiply-Adder Functions

Synthesis tools detect multiply-accumulator or multiply-adder functions, and either implement them as Intel FPGA IP cores or map them directly to device atoms. During placement and routing, the Quartus® Prime software places multiply-accumulator and multiply-adder functions in DSP blocks.
Note: Synthesis tools infer multiply-accumulator and multiply-adder functions only if the Intel device family has dedicated DSP blocks that support these functions.

A simple multiply-accumulator consists of a multiplier feeding an addition operator. The addition operator feeds a set of registers that then feeds the second input to the addition operator. A simple multiply-adder consists of two to four multipliers feeding one or two levels of addition, subtraction, or addition/subtraction operators. Addition is always the second-level operator, if it is used. In addition to the multiply-accumulator and multiply-adder, the Quartus® Prime Fitter also places input and output registers into the DSP blocks to pack registers and improve performance and area utilization.

Some device families offer additional advanced multiply-adder and accumulator functions, such as complex multiplication, input shift register, or larger multiplications.

The Verilog HDL and VHDL code samples infer multiply-accumulator and multiply-adder functions with input, output, and pipeline registers, as well as an optional asynchronous clear signal. Using the three sets of registers provides the best performance through the function, with a latency of three. To reduce latency, remove the registers in your design.
Note: To obtain high performance in DSP designs, use register pipelining and avoid unregistered DSP functions.

Verilog HDL Multiply-Accumulator

module sum_of_four_multiply_accumulate
   #(parameter INPUT_WIDTH=18, parameter OUTPUT_WIDTH=44)
   (
      input clk, ena,
      input [INPUT_WIDTH-1:0] dataa, datab, datac, datad,
      input [INPUT_WIDTH-1:0] datae, dataf, datag, datah,
      output reg [OUTPUT_WIDTH-1:0] dataout
   );
   // Each product can be up to 2*INPUT_WIDTH bits wide.
   // The sum of four of these products can be up to 2 bits wider.
   wire [2*INPUT_WIDTH+1:0] mult_sum;

   // Store the results of the operations on the current inputs
   assign mult_sum = (dataa * datab + datac * datad) +
                     (datae * dataf + datag * datah);

   // Store the value of the accumulation
   always @ (posedge clk)
   begin
      if (ena == 1)
         begin
            dataout <= dataout + mult_sum;
         end
   end
endmodule