Floating-Point Operation

The DSP block supports a fused multiply add (FMA) operation in BFLOAT16 format. BFLOAT16 is mapped to 16 bits as shown in the following figure.

Efinity synthesis does not support inferred floating-point operations in Titanium DSP blocks. Instead, To use this function you must instantiate the EFX_DSP48 primitive and set the MODE parameter to BFLOAT. In this mode, the DSP block inputs A, B, and C be in BFLOAT16 format, and the resulting output O is in 32-bit floating-point format (FP32). The accumulation happens internally in 32-bit floating-point. All pipeline registers (except O_REG) must be enabled for correct operation. The following code shows an example instantiation of the primitive:

EFX_DSP48 dut (.A(a), .B(b), .C(18'd0), .OP(op), .O(o), .CASCIN({48{1'b0}}),
        .CLK(clk), .RST(rst), .CE(1'b1), .SHIFT_ENA(1'b0));
    defparam dut.MODE  = "BFLOAT";
    defparam dut.M_SEL = "P";
    defparam dut.N_SEL = "W";
    defparam dut.W_SEL = "X";
    defparam dut.A_REG = 1;
    defparam dut.B_REG = 1;
    defparam dut.C_REG = 0;
    defparam dut.P_REG = 1;
    defparam dut.OP_REG = 1;
    defparam dut.W_REG = 1;
    defparam dut.O_REG = 0;
    defparam dut.RST_POLARITY = 1;