G1.4.1 The Mixer That Clipped

MediumFixIcarus · browser samples / server checksReview the topic

Fix a fixed-point gain stage that keeps the wrong bits of its product and never saturates.

Hardware context: A fixed-point gain stage of an audio mixer and the display rounding of a digital scale. Repair the gain stage that keeps the low byte of a Q8.8 product and wraps when it should saturate.

An audio mixer scales each sample by a gain. The sample and the result are signed Q4.4 numbers, and the gain is an unsigned Q4.4 number. On the bench the output is noise, even at a gain of 1.0.

This is a Fix problem. The multiplier in the block is right. The block then takes the wrong bits out of the product. Find the bits that hold a Q4.4 result.

There is a second fault. A product can be larger than Q4.4 can hold. The result must then stop at the nearest end of the range: 7.9375 at the top and -8.0 at the bottom.

The check applies ten pairs of sample and gain, one per nanosecond, and records the output after each.

Signals in the Timeline

SignalWidthRole
script160input of each check
Full specification
The module (keep the name and the ports)
module mix_gain (
  input  logic signed [7:0] sample,   // signed Q4.4: -8.0 to 7.9375
  input  logic        [7:0] gain,     // unsigned Q4.4: 0.0 to 15.9375
  output logic signed [7:0] out       // signed Q4.4: sample x gain, saturated
);
  // the logic to repair
endmodule
StepRule
MultiplyThe product of the two stored integers has eight fraction bits.
Drop bitsDrop the four lowest fraction bits. Do not round: a dropped part always makes the result smaller.
SaturateA result above 7.9375 becomes 7.9375 (stored 127). A result below -8.0 becomes -8.0 (stored -128).

The result fields are named out_step0 to out_step9. Each shows the stored integer of the output as a signed number: the value times 16.

Examples

Example 1

Small Samples, Small Gains
Input
2.5 x 1.5; 1 x 1; 0.5 x 0.5; 3 x 2; 1.25 x 4; 0 x 15; 7.9375 x 1; 2 x 0; 0.0625 x 1; 4 x 1.5
Output
out_step0 = 60, out_step1 = 16, out_step2 = 4, out_step3 = 96, out_step4 = 80, out_step5 = 0, out_step6 = 127, out_step7 = 0, out_step8 = 1, out_step9 = 96

2.5 times 1.5 is 3.75, stored as 60. A gain of 1.0 must return the sample unchanged, and a gain of 0 must return 0.

Example 2

Negative Samples That Fit
Input
-2.5 x 1.5; -1 x 1; -0.5 x 0.5; -3 x 2; -1.25 x 4; -8 x 1; -4 x 2; -0.0625 x 1; -7.9375 x 1; -2 x 3.5
Output
out_step0 = -60, out_step1 = -16, out_step2 = -4, out_step3 = -96, out_step4 = -80, out_step5 = -128, out_step6 = -128, out_step7 = -1, out_step8 = -127, out_step9 = -112

-2.5 times 1.5 is -3.75, stored as -60. The sample -8.0 at a gain of 1.0 is the most negative result that fits.

Example 3

Results Above the Top
Input
4 x 2; 7.9375 x 2; 1 x 8; 0.5 x 15.9375; 7.9375 x 15.9375; 3 x 3; 2 x 4; 5.5 x 1.5; 1 x 7.9375; 1 x 8.0625
Output
out_step0 = 127, out_step1 = 127, out_step2 = 127, out_step3 = 127, out_step4 = 127, out_step5 = 127, out_step6 = 127, out_step7 = 127, out_step8 = 127, out_step9 = 127

4 times 2 is 8.0, one step above the largest Q4.4 value. The output stops at 7.9375, stored as 127. The broken block wraps round instead.

Constraints and Contract

  • Inputs are known 0/1 values. Each pair is held for one nanosecond and out is sampled at its end.
  • Keep the module name and the ports, including their signed and unsigned declarations.

Hints

Hint 1

Count the fraction bits. Each input has four, so the product has eight. The output has four again. Which bit of the product has the weight 1/16?

Hint 2

After you drop four bits, the value still has more integer bits than the output. Keep them all in a wider signed signal first.

Hint 3

Compare that wide value with the largest and the smallest stored integer of a signed byte. Pick one of three values for the output.

What the grader checks: Six directed and eight seeded cases of ten steps each run on the block. Every output is compared with a model written in JavaScript. Rounding modes other than dropping the low bits are outside this problem.

14 checks on Icarus · browser samples / server checks. Run the examples first, then check the full regression. Full checks use the server judge. Downloads contain public samples only. A passing result establishes the tested cases, not a proof for all possible inputs.

See all 2 steps and the final capstone

Report a problem with this page

Your code SystemVerilogSaved locally
Ln 1, Col 1Tab indents · Ctrl/⌘ + Enter runs samples
Not run yet

Small Samples, Small Gains

Input
2.5 x 1.5; 1 x 1; 0.5 x 0.5; 3 x 2; 1.25 x 4; 0 x 15; 7.9375 x 1; 2 x 0; 0.0625 x 1; 4 x 1.5
Expected output
out_step0 = 60, out_step1 = 16, out_step2 = 4, out_step3 = 96, out_step4 = 80, out_step5 = 0, out_step6 = 127, out_step7 = 0, out_step8 = 1, out_step9 = 96

2.5 times 1.5 is 3.75, stored as 60. A gain of 1.0 must return the sample unchanged, and a gain of 0 must return 0.

Run samples executes all 3 public checks. Check all tests runs 14 checks on the server.

Icarus · browser samples / server checks