I decided to build a FP with 8-bit exponent and 9-bit mantissa (and with no NANs, infinities, denorms or rounding).The sum of the bit-lengths (plus one sign bit) means that the FP number fits into a 18-bit M4K block word on the CycloneII FPGA. The 9-bit mantissa means that only one hardware multiplier (out of 70) is used for the floating multiplier. The exponent is represented in 8-bit offset binary form. For example 20=128, 22=130, and 2-2=126. The mantissa is represented as a 9-bit fraction with a range of [0 to 1.0-2-9]. A sign bit of zero implies positive. The Verilog representation is {sign,exp[7:0],mantissa[8:0]}. Denormalized numbers are not permitted, so the high-order bit (binary value 0.5) is always one, unless the value of the FP number underflows, then it is zero. No error detection is performed and there is no rounding. There are no NANs, infinities, denorms, or other special cases (which make little sense in a realtime system anyway). Some example representations are shown below.

Five operations are necessary for floating DSP. They are add, negate, multiply, integer-to-float, and float-to-integer. Negate is easy, just toggle the sign bit. The integer conversion algorithms are necessary because the audio and video codecs are integer-based. Outlines for the functions are below. Finally, the modules were tested by building IIR filters. The SOS filters shown below validated the performance of the floating point. An article written for Circuit Cellar Magazine describing this floating point is available.

Multiply algorithm:

  1. If either input number has a high-order bit of zero, then that input is zero and the product is zero.
  2. The output exponent is exp1+exp2-128 or exp1+exp2-129. If the sums of the input exponents is less than 129 then the exponent will underflow and the product is zero.
  3. If both inputs are nonzero and the exponents don't underflow:
    1. Then if (mantissa1)x(mantissa2) has the high order-bit set, the top 9-bits of the product are the output mantissa and the output exponent is exp1+exp2-128.
    2. Otherwise the second bit of the product will be set, and the output mantissa is the top 9-bits of (product)<<1 and the output exponent is exp1+exp2-129.
    3. The sign of the product is (sign1)xor(sign2)


Add algorithm:

  1. If both inputs are zero, the sum is zero.
  2. Determine which input is bigger, which smaller (absolute value) by first comparing the exponents, then the mantissas if necessary.
  3. Determine the difference in the exponents and shift the smaller input mantissa right by the difference. But if the exponent difference is greater than 8 then just output the bigger input.
  4. If the signs of the inputs are the same, add the bigger and (shifted) smaller mantissas. The result must be 0.5<sum<2.0. If the result is greater than one, shift the mantissa sum right one bit and increment the exponent. The sign is the sign of either input.
  5. If the signs of the inputs are different, subtract the bigger and (shifted) smaller mantissas so that the result is always positive. The result must be 0.0<difference<0.5. Shift the mantissa left until the high bit is set, while decrementing the exponent. The sign is the sign of the bigger input.


The multiplier takes about 60 logic elements plus one hardware multiplier on the CycloneII FPGA, while the adder takes about 220 logic elements. The timing analyser suggests that the purely combinatiorial multiplier should be able to run at 50 MHz and the adder at 30 MHz or so.

The integer-to-FP and FP-to-integer conversion routines allow you to specify a signed scale. Going from integer to float, the resulting floating point number is (integer_input)*2scale_input. This feature allows you to convert numbers less than one. Going from float back to integer, you choose the scale you want to bring the floating point number back into a small integer range. The signed integer inputs and outputs are 10-bit, 2's complement format.

Integer to FP:

I assumed 10-bit, 2’s complement, integers since the mantissa is only 9 bits, but the process generalizes to more bits.

  1. Save the sign bit of the input and take the absolute value of the input.
  2. Shift the input left until the high order bit is set and count the number of shifts required. This forms the floating mantissa.
  3. Form the floating exponent by subtracting the number of shifts from step 2 from the constant 137 or (0h89-(#of shifts)).
  4. Assemble the float from the sign, mantissa, and exponent.

FP to integer:

Converting back to integer is similarly simple, but no overflow is detected, so scale carefully.

  1. If the float exponent is less than 0h81, then the output is zero because the input is less than one.
  2. Otherwise shift the floating mantissa to the right by (0h89-(floating exponent)) to form the absolute value of the output integer.
  3. Form the 2’s complement signed integer.

Testing the FP routines using IIR filtering by Second-order-sections (SOS)
SOS filters have the advantage (over straight multipole filters) of smaller dynamic range on coefficients, so the numerical stability is better. SOS filters are also more straight forward to do with floating point. The downside is a few more state variables and a few more multiplies for each filter. A matlab program and function convert filter specifications to Verilog with 18-bit floating point. The top-level module defines filters of order 2, 4 and 6. The project is archived here.

Testing the FP routines using IIR filtering
The fpmult, fpadd, int2fp and fp2int routines were incorporated into the state machine filters described on the FPGA DSP page, example 4. The routines worked, implying that the logic is correct, however a 9-bit manitssa is apparently not accurate enough to implement high-order or narrow bandwidth filters. Second order filters work fine, but 4th and 6th order filters became inaccurate when the filter bandwidth was low. Use the SOS verions above for most actual filters. A matlab program (and associated function) were used to convert matlab-designed filter coefficients to floating point format. The top-level module defines three filters and connects them to the audio in/out. The entire project is zipped here.

FP reciprocal
The ability to take a reciprocal allows division to occur. Reciprocal was implemented Newton-Raphson interation on an initial linear estimate of the reciprocal. This design just tests for static correctness of the method by displaying values on the LEDs. The process is to take the input number, strip off the sign and exponent, compute the reciprocal of the remaining number between 0.5 and 1.0, form the new exponent as 0x81+(0x81-input_exponent) then merge together the input sign, new exponent and new mantissa from a Newton iteration process. The module will run at 14 MHz and uses 3 floating point adders and 4 floating point multipliers.

The algorithm (from http://en.wikipedia.org/wiki/Division_%28digital%29) is as follows, with all operations being floating arithmetic:

  1. Form In_reduced={1'b0, 8'h80, m1} where m1 is the mantissa of the normalized input float. This operation (with the sign set to + and the exponent set to 0x80), limits the range to 0.5 to 1.0.
  2. x0 = 2.9142 - 2*in_reduced ( 0.5<=in_reduced<=1.0)
  3. x1 = x0*(2-in_reduced*x0)
  4. x2 = x1*(2-in_reduced*x1)
  5. The reciprocal output mantissa is the mantissa of x2.
  6. The reciprocal output exponent =(in_reduced==9'b100000000)? 9'h102-e1 : 9'h101-e1
    because an input value of exactly a power of 2 adds one to the exponent.
    e1     is the exponent of the normalized input float.

FP reciprocal Square Root
A reciprocal square root function is useful when normalizing vectors (e.g. computer graphics) and con be converted to a square root with just one more multiply. This design just tests for static correctness of the method by displaying values on the LEDs. The process is to take the input number, strip off the sign and exponent, compute the reciprocal square root of the remaining number between 0.25 and 1.0, form the new exponent, then merge together the new exponent and new mantissa from a Newton iteration process. The module will run at 11 MHz and uses 3 floating point adders and 6 floating point multipliers.

The algorithm (from http://en.wikipedia.org/wiki/Methods_of_computing_square_roots) is as follows, with all operations being floating arithmetic. The x0 estimate is based on a linear approximation which I thought up.:

  1. Form input_exp = (e1[0]==1)? 8'h7f : 8'h80 and
    reduced_input = {1'b0, input_exp, m1}
    where m1 is the mantissa of the normalized input float. This operation, limits the range to 0.25 to 1.0.
  2. x0 = 2.05 - reduced_input with (input 0.25<=reduced_input<=1.0)
  3. x1 = x0/2 * (3 - reduced_input*x0*x0)
  4. x2 = x1/2 * (3 - reduced_input*x1*x1)
  5. The reciprocal sqrt output mantissa is the mantissa of x2.
  6. The reciprocal sqrt output exponent eout is given by
    ((m1==9'b100000000 && e1[0]==1) )? eout = 9'h82 + ((9'h80 - e1)>>1) : eout = 9'h81 + ((9'h80 - e1)>>1)
    because an input value of exactly a power of 2 with odd exponent adds one to the exponent. e1 is the exponent of the normalized input float.