Vivado Build

This page documents the Vivado build flow for the pccx v002 NPU core. All build scripts reside under hw/vivado/; build.sh is the single entry point. Source files are managed in the pccxai/pccx-FPGA-NPU-LLM-kv260 repository.

Build Flow

build.sh is a thin wrapper around Vivado batch-mode invocations. The first argument selects one of four stages.

./hw/vivado/build.sh project   # create project only
./hw/vivado/build.sh synth     # create project + OOC synthesis
./hw/vivado/build.sh impl      # full implementation + bitstream
./hw/vivado/build.sh clean     # remove build/ directory

The script first searches PATH for vivado; if not found it falls back through /tools/Xilinx/2025.2, 2024.1, 2023.2 in that order. Vivado 2023.2 or later is required.

Stage descriptions:

  1. create_project.tcl — Creates a Vivado project targeting part xck26-sfvc784-2LV-c (KV260 ZU5EV) at build/pccx_v002_kv260/. Parses filelist.f to populate the sources_1 fileset, then adds every *.xdc from hw/constraints/ to constrs_1.

  2. synth.tcl — Runs synth_design -mode out_of_context -flatten_hierarchy rebuilt. On completion, writes utilization_post_synth.rpt, clocks_post_synth.rpt, timing_summary_post_synth.rpt, and drc_post_synth.rpt to build/reports/. Check WNS in the timing summary before proceeding to implementation.

  3. impl.tcl — Verifies that synth_1 progress is 100%, then launches impl_1 -to_step write_bitstream -jobs 4. On success, the bitstream is copied to build/pccx_v002_kv260.bit. Implementation is an hour-scale job; run it only after OOC synthesis is clean.

The OOC mode is required because NPU_top uses SystemVerilog interface ports (axil_if / axis_if). Out-of-context synthesis leaves those ports unbound, allowing the core to be synthesised and checked in isolation before a block-design wrapper is available.

SV Interface Wrapper

The board-side SystemVerilog wrapper (npu_core_wrapper.sv) converts NPU_top’s SystemVerilog interface ports into plain AXI4-Lite and AXI4-Stream signal bundles so the core can be packaged as a Vivado IP and placed alongside the Zynq PS in a block design. The wrapper contains no registers and no CDC logic; it performs one-to-one signal expansion only.

The wrapper file is part of the KV260 board integration repository (pccxai/pccx-FPGA-NPU-LLM-kv260), not the reusable IP-core package. After the v002 extraction it remains in the board integration layer rather than pccx-v002, because it encodes board-side packaging concerns. The current authoritative source is therefore the board integration repo’s hw/vivado/npu_core_wrapper.sv.

The external interface exposed by the wrapper is as follows.

Table 13 npu_core_wrapper external interfaces

Port group

Direction

Width

Description

clk_core / rst_n_core

Input

1-bit

Core-domain clock and active-low reset

clk_axi / rst_axi_n

Input

1-bit

AXI-domain clock and active-low reset

s_axil_*

Slave

32-bit

AXI4-Lite control channel (CMD_IN / STAT_OUT)

s_axis_hp0 … hp3

Slave

128-bit

AXI4-Stream HP ports × 4 (weight streaming)

s_axis_acp_fmap

Slave

128-bit

AXI4-Stream ACP FMap input

m_axis_acp_result

Master

128-bit

AXI4-Stream ACP result output

The Vivado IP packager auto-infers AXI interfaces from plain signal ports, so this wrapper allows NPU_top to be placed directly into a block design alongside the Zynq PS.

Constraints

hw/constraints/pccx_timing.xdc is a timing-only constraint file. Pin and IO constraints are absent; they are delegated to the block design that wraps this core.

Two clock domains are declared.

Table 14 Clock domain definitions

Clock name

Period

Frequency

Scope

axi_clk

4.000 ns

250 MHz

AXI-Lite MMIO, CDC FIFO drain sides, DMA path

core_clk

2.500 ns

400 MHz

DSP48E2 array, GEMV lanes, CVO SFU

The two domains are genuinely asynchronous. set_clock_groups -asynchronous is applied; every domain crossing uses a CDC FIFO or a properly-staged reset synchroniser.

Additional path exceptions are set.

  • False paths — Reset bridge first-flop paths and XPM_FIFO_ASYNC gray-coded pointer crossings.

  • Multicycle path (setup 2, hold 1) — From DSP48E2 P-registers inside the GEMM systolic array to mat_result_normalizer. The controller stalls new MACs during accumulator flush, so the drain path tolerates two cycles.

File Manifest

hw/vivado/filelist.f is the source list for both OOC synthesis and xvlog lint. create_project.tcl parses this file to add sources to the Vivado project.

Compile order follows the declaration order in the file. Packages and interfaces must appear before the modules that import them.

The full file is organized into the following sections.

Table 15 filelist.f section structure

Section

Contents

A (comment only)

`include headers (*.svh); picked up via -i, not compiled

B

device_pkg.sv — device parameters

C

dtype_pkg.sv, mem_pkg.sv — data and memory types

D

vec_core_pkg.sv — Vector Core pipeline parameters

Library

BF16 math library, algorithms, QUEUE interface

ISA

isa_pkg.sv — 64-bit VLIW encoding package

MAT_CORE

GEMM systolic array and result packer

VEC_CORE

GEMV core (accumulate, LUT generation, reduction, top)

CVO_CORE

CORDIC unit, SFU, CVO top

PREPROCESS

BF16↔fixed-point pipeline, fmap cache

MEM_control

L2 cache, HP buffer, CVO bridge, dispatcher

NPU_Controller

AXI-Lite decoder, dispatcher, front-end, Global Scheduler, top

Top level

NPU_top.sv — always last

When adding a new .sv file, insert it in dependency order inside filelist.f.