Alveo V80 with SLASH
On the Alveo V80, TAPA targets the SLASH service shell instead of a Vitis
platform. The flow is a third target, xilinx-slash: tapa analyze selects
it, tapa synth --platform slash_v80_service resolves the V80 part and its
default clock, tapa pack writes a .zip package, and tapa link runs
SLASH's slashkit to produce a .vbin. The unchanged host program runs the
vbin through VRT, SLASH's runtime, on the board or in simulation.
The working example is tests/apps/vadd.
Prerequisites
- TAPA, plus the
cbcsolver onPATHif you floorplan. - Vivado and Vitis HLS 2025.1 on the tool host — the versions the pinned SLASH release builds the V80 shell with.
- SLASH at commit
26f9c55con the tool host, withslashkitonPATHand the V80 static shell installed.tapa linkrecords this pin inlink-metrics.json; revalidate the flow when you move to another commit. - For on-board runs: a V80 with the SLASH driver loaded and the
vrtddaemon running (/run/vrtd.sock), and SLASH'slibvrtdinstalled where the host executable runs. - For SLASH simulation runs:
libzmq(libzmq.so.5) installed where the host executable runs; no board,vrtd, orlibvrtdis involved.
tapa link resolves slashkit, vivado, and v++ by name on the tool host
(override each with --slashkit, --vivado, --vpp), so the flow works
unchanged under --remote-host.
The flow
tapa --work-dir work.out analyze --top VecAdd -f vadd.cpp \
--target xilinx-slash
tapa --work-dir work.out synth --platform slash_v80_service
tapa --work-dir work.out floorplan --connectivity v80.cfg # optional
tapa --work-dir work.out pack -o vadd.zip
tapa --work-dir work.out link --package vadd.zip \
--connectivity v80.cfg -o vadd.vbin
--target xilinx-slashselects the SLASH flow;--platform slash_v80_serviceand that target require each other, so a mismatch in either direction is an error.slash_v80_serviceresolves partxcv80-lsva4737-2MHP-e-Sand defaults the clock period to 3.33 ns (300 MHz). Pass--clock-periodto override the default; a--part-numthat disagrees with the shell is an error.tapa packwrites a.zippackage: the packaged IP directory (component.xml, HDL, GUI files),kernel.xml,hls_data.json, the generated RTL, and the build state. The package is also what fast cosim loads, so cosim runs without a board.- The SLASH target requires the
ap_ctrl_hscontrol protocol: a top task with stream ports (host streams) is rejected atpack, because SLASH has no host-stream channel yet. The runtime reports the same limit if it meets one. tapa floorplanis optional but recommended for cross-SLR timing; the V80 device table models the SLASH reconfigurable module, including its per-SLR URAM capacity (see Floorplan for Timing Closure). Every hardware link inserts Full AXI register slices on the HBM interfaces selected by the connectivity map, synchronizes their reset, and anchors the bridges and slices to SLR2. TAPA also preflights optimized URAM demand and requests clock-region replication for eligible high-fanout fabric controls. The generatedhbm-slices.tsvandcontrol-coverage.tsvrecord these actions.
The SLASH configuration
--connectivity takes slashkit's own configuration file. Its [connectivity]
section names the one kernel instance and binds each compiled M-AXI port to a
memory target:
[connectivity]
nk=VecAdd:1:VecAdd_0
sp=VecAdd_0.m_axi_a:HBM0
sp=VecAdd_0.m_axi_b:HBM1
sp=VecAdd_0.m_axi_c:HBM2
nk=<top>:1:<instance>declares the single kernel instance; the instance name (VecAdd_0here) prefixes everysp=port.- Each
sp=<instance>.<port>:<target>must name a compiled M-AXI interface of the top task — the packaged port name,m_axi_afor an argumenta, not the C++ spelling. Targets areHBM0–HBM63andDDR0–DDR3. - A
[clock]section is optional: TAPA appendskrnl=<instance>/freqhz=<requested>when it is absent. A[clock]you write yourself must name the solenkinstance at exactly the link's requested frequency; any other value is an error rather than a silent override.
The same file feeds tapa floorplan and tapa link.
Frequency: requested, recorded, and the closing loop
One clock period — the one tapa synth used — drives the whole flow. tapa link derives its request from the packaged flow state, or takes an explicit
--clock-mhz. The link constrains the reconfigurable module at that period
(SLASH alone would implement it at the static shell's 200 MHz default) and
writes min(requested, achieved) into the vbin as the recorded
ClockFrequency.
A link succeeds when slashkit produces a vbin. When routing misses the
requested period, tapa link warns with both numbers in SLASH's terms — the
requested clock was derated, not failed:
SLASH routed kernel timing did not meet 300 MHz; recording 257.431 MHz in `vadd.vbin`
Hold violations on the kernel clock and violations in the shell's fixed 400 MHz domain are also warnings; the shell's own timing belongs to SLASH, not to your link.
Every link writes link-metrics.json into the work directory's link/
subdirectory:
{
"slash_commit": "26f9c55c",
"requested_mhz": 300,
"recorded_mhz": 257.431,
"kernel_wns_ns": -0.522,
"kernel_fmax_mhz": 257.431,
"timing_closed": false,
"fixed_domain_wns_ns": -0.186,
"timing_report": ".../vadd.vbin.prj/slash_rm/report_timing_vadd.txt"
}
If the recorded frequency is lower than you wanted, rerun tapa link with a
lower --clock-mhz (and the matching [clock], if you wrote one) and keep
the highest value that closes timing:
tapa --work-dir work.out link --package vadd.zip \
--connectivity v80.cfg -o vadd-280.vbin --clock-mhz 280
Over-constraining routes worse, so step down rather than asking for more than
the design achieves. At run time the host runtime programs the board's user
clock to the recorded frequency itself — no manual v80-smi clock step is
needed — and logs the rate vrtd reports back.
Fast cosim on the package
The .zip package loads like an XO: pass it as the host executable's
--bitstream and TAPA runs its usual fast cosimulation, no board and no
SLASH install:
./vadd --bitstream=vadd.zip 1024
SLASH simulation
tapa link --sim asks slashkit for a simulation vbin instead of a hardware
image. The vbin embeds its own vpp_sim simulator model, which the host
runtime starts and drives over ZeroMQ — no board, vrtd, or libvrtd is
involved:
At the pinned SLASH commit, the shipped vpp_sim still ignores its endpoint
argument and listens on tcp://*:5555. TAPA implements the documented IPC
protocol only, so that stock binary cannot be driven by this runtime. Use a
SLASH build whose vpp_sim accepts the IPC endpoint; TAPA has no TCP fallback.
tapa --work-dir work.out link --package vadd.zip \
--connectivity v80.cfg -o vadd.sim.vbin --sim
./vadd --bitstream=vadd.sim.vbin 1024
--sim exists only for the xilinx-slash target. The link itself still runs
slashkit and Vivado on the tool host. Slashkit emits no routed timing report
in sim mode, so the link records the requested clock unchanged and
link-metrics.json omits the kernel and fixed-shell timing fields and the
report path.
On the board
With vrtd running on the board host, select the card by its PCIe address
and run the same host program against the hardware vbin:
./vadd --bitstream=vadd.vbin --vrt_bdf=0000:a1:00 1024
--vrt_bdf (or FRT_VRT_BDF) is required for a hardware vbin. The runtime
loads the vbin's PDI through vrtd, waits for the shell to settle, programs
the recorded user clock, and runs the kernel through BAR-mapped registers and
vrtd device buffers. If libvrtd is not on the loader's search path, point
FRT_VRTD_LIBRARY at libvrtd.so.1; for simulation, FRT_ZMQ_LIBRARY
overrides the libzmq.so.5 lookup the same way.
What is not supported
| Feature | Status |
|---|---|
| Host streams on the top task | Rejected at pack; the VRT backend has no host-stream channel. Use shared-memory streams between tasks and an mmap port to the host. |
kill (kernel cancellation) | VRT has no cancellation. finish is bounded and reports a timeout on a hung kernel; its buffers are retained on close. |
SLASH emulation (-p emu) | Not applicable: a TAPA package is RTL, and SLASH emulation runs HLS C models. Use fast cosim or SLASH simulation. |