419 LayerNorm
419 : LayerNorm
- Author: Jakubie
- Description: 8-element LayerNorm accelerator
- GitHub repository
- Open in 3D viewer
- Open in Tiny Tapeout Explorer
- Clock: 1000000 Hz
How it works
This project implements a simplified LayerNorm accelerator for eight signed 8-bit integer samples. It calculates an integer mean and variance, then scales each sample's deviation from the mean using a reciprocal square-root lookup table.
The reciprocal square-root lookup table supports integer variances from 1 to 15 and uses Q0.9 fixed-point coefficients. A variance of 0 or a variance greater than 15 produces zero outputs. Results are signed 8-bit integers truncated toward zero.
The design does not implement epsilon, learnable scale or bias, or overflow protection.
Interface
- ui_in[7:0]: signed input sample (two’s complement).
- uio_in[0] — VALID: accepts one sample per rising clock edge during input collection.
- uio_in[1] — START: starts a new operation while idle.
- uo_out[7:0]: signed output sample.
All bidirectional pins are configured as inputs. There is no output-valid or busy signal.
How to test
- Hold rst_n low for at least one rising clock edge, then release it.
- Pulse START high for one clock cycle.
- Starting on the next rising edge, supply eight samples with VALID high. Set VALID low afterward.
- Taking the eighth sample’s capture edge as cycle 0, read outputs after rising edges 35, 39, 43, 47, 51, 55, 59 and 63, allowing signals to settle.
- The design returns to idle at cycle 64. A new START can be accepted at cycle 65.
Outputs appear in input order and remain available for one clock cycle each. The output bus is zero between results; zero is also a valid result.
Example input: [0, 1, 2, 3, 4, 5, 6, 7] Expected output: [-1, 0, 0, 0, 0, 0, 1, 1]
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | DATA[0] | OUTPUT[0] | VALID |
| 1 | DATA[1] | OUTPUT[1] | START |
| 2 | DATA[2] | OUTPUT[2] | |
| 3 | DATA[3] | OUTPUT[3] | |
| 4 | DATA[4] | OUTPUT[4] | |
| 5 | DATA[5] | OUTPUT[5] | |
| 6 | DATA[6] | OUTPUT[6] | |
| 7 | DATA[7] | OUTPUT[7] |