A fully open-source hardware AV1 intra encoder in SystemVerilog, verified bit-exact against two independent decoders (ffmpeg and dav1d). Implements 9 intra prediction modes with SAD-based mode decision, 4x4 DCT/quantization, entropy coding, and local reconstruction feedback. Synthesized to GDS-II on SkyWater 130nm — DRC and LVS clean.

Both versions target the SkyWater SKY130 process using OpenLane 2.3.1. The V2 critical path runs through the directional predictor's 16-pixel interpolation stage and 4-level SAD adder tree (~20.7 ns at the nominal corner). Synthesis is fully reproducible from the committed config.

Pixels ──► Mode Decision ──► Directional Predict ──► Residual ──► DCT ──► Quantize ──► Entropy ──► Bitstream
              │ (9 modes)       (16-pixel interp)                                          │
              │ SAD compare                                                                │
              ▼                                                                            ▼
           best_mode                                                                 Inv Quantize
                                                                                          │
           Recon pixels ◄──── Clamp ◄──── Add Prediction ◄──── Inverse DCT ◄──────────────┘

Prediction: 9 AV1 intra modes — DC, V, H, D45, D135, D113, D157, D203, D67. Mode decision evaluates all 9 via SAD against the original block and selects the best. Directional predictor computes sub-pixel interpolation using the AV1-specified dx/dy tables.

Transform + Quantization: Row-column 4x4 DCT with HEVC-convention shift values (SHIFT_ROW=1, SHIFT_COL=8). Quantizer uses multiply-by-reciprocal with 33-bit precision to avoid division. AV1-standard quantization tables (256 entries for DC/AC).

Entropy Coding: Bool encoder with a 1024-bit accumulation register (no streaming carry propagation). Coefficient encoder maps quantized coefficients to binary symbols (has_more, is_nz, sign, unary magnitude).

Reconstruction Feedback: Inverse quantize, inverse DCT, add prediction, and clamp to [0, 255]. Reconstructed pixels feed back as reference for subsequent blocks.

Bitstream Output: Full AV1 compliance for the supported feature set — IVF container, OBU framing (temporal delimiter + sequence header + frame OBU), tile encoding with Daala EC entropy coding and CDF-based context modeling.

• Python 3.9+
• Cocotb 2.0+ (pip install cocotb)
• Icarus Verilog 12.x (sudo apt install iverilog)
• ffmpeg (optional, for bitstream decode verification)

git clone https://github.com/Tgcohce/openAV1.git
cd openAV1
pip install -r requirements.txt

# V1 pipeline tests (57 tests across 8 suites)
cd tb
for mf in Makefile Makefile.quant Makefile.pipe Makefile.bool \
          Makefile.coeff_enc Makefile.inv Makefile.pred Makefile.intra; do
    make -f $mf SIM=icarus
done

# V2 directional prediction tests (34 tests across 4 suites)
for mf in Makefile.sad Makefile.dir_pred Makefile.mode_dec Makefile.intra_v2; do
    make -f $mf SIM=icarus
done

python -m golden.encode_frame       # Full encoder + ffmpeg decode test
python -m golden.av1_tile           # Tile encoder (53 tests)
python -m golden.bool_coder         # Bool encoder/decoder (994 tests)
python -m golden.coeff_encode       # Coefficient coding (1024 tests)

python scripts/encode_image.py --width 16 --height 16 --pattern gray output.ivf
ffmpeg -i output.ivf -f rawvideo -pix_fmt gray decoded.raw

Requires Docker with 12+ GB memory allocated.

docker run --rm \
    -v "$(pwd):/work" \
    -v openlane-pdk:/home/openlane/.volare \
    -w /work/synth \
    ghcr.io/efabless/openlane2:2.3.1 \
    python3 -m openlane /work/synth/config.json

Bitstreams validated against ffmpeg (libdav1d) and dav1d (VideoLAN) — both produce identical decoded output across all configurations.

Measured on a 16x16 random image (monochrome, DC + directional prediction):

golden/          Python golden models (DCT, quantizer, entropy, prediction, bitstream)
rtl/             SystemVerilog RTL (15 modules + 9 synthesis variants)
tb/              Cocotb testbenches (12 test files + 12 Makefiles)
synth/           OpenLane 2 config + synthesis reports
scripts/         Encoding tools, GDS renderer, test runner
sim/             Shell scripts for batch simulation

This is a proof-of-concept encoder targeting a specific AV1 subset:

• Monochrome only (no chroma planes)
• Still pictures (intra-only, no inter-frame prediction)
• 4x4 transform blocks (no 8x8, 16x16, or 32x32)
• 9 of 13 intra modes (no SMOOTH, SMOOTH_V, SMOOTH_H, or PAETH)
• Up to 64x64 pixels per frame
• Synthesis-verified (GDS-II generated and DRC/LVS clean, not fabricated)

• 4x4 DCT transform unit
• AV1 quantizer (multiply-by-reciprocal)
• Entropy coding (bool encoder + coefficient encoder)
• DC intra prediction with reconstruction feedback
• AV1 bitstream assembly (OBU framing, tile encoding, IVF container)
• V1 chip layout (SkyWater 130nm, 64 MHz)
• 9-mode directional prediction (golden model + RTL)
• V2 chip layout (SkyWater 130nm, 28.6 MHz)
• Multi-block streaming with neighbor feedback
• Larger transform sizes (8x8, 16x16, 32x32)
• Inter-frame prediction (integer-pel motion estimation)

MIT License. See LICENSE for details.