derek.holloway
7502bf64a4
2.7 KiB
Configuration
Simulation topologies are defined declaratively in JSON. The config specifies the component hierarchy, link connectivity, and simulation parameters.
Config files
Pre-built configs are in configs/:
| File | Description |
|---|---|
gfx950_cdna4.json |
Single CDNA4 GPU (standalone simulation) |
gfx950_cdna4_kmd.json |
Single CDNA4 GPU (daemon/KFD mode) |
gfx950_cdna4_kmd_2gpu.json |
Two CDNA4 GPUs (multi-GPU daemon mode) |
gfx942_cdna3.json |
Single CDNA3 GPU (standalone simulation) |
gfx942_cdna3_kmd.json |
Single CDNA3 GPU (daemon/KFD mode) |
gfx1250.json |
Single gfx1250 GPU (standalone simulation, no KMD) |
JSON structure
{
"max_ticks": 100000,
"num_threads": 1,
"exec_mode": "functional",
"vm": { "arch": "cdna4" },
"topology": {
"root": {
"name": "soc", "type": "soc",
"children": [
{ "name": "vram", "type": "gpu_memory" },
{ "name": "xcd[0:8]", "type": "xcd", "children": [...] }
]
},
"links": [
{
"pattern": "xcd[i].se[j].cu[k].req -> xcd[i].l2.cpl_[j*8+k]",
"for_ranges": [
{ "var_name": "i", "start": 0, "end": 8 },
{ "var_name": "j", "start": 0, "end": 4 },
{ "var_name": "k", "start": 0, "end": 8 }
],
"latency": 1, "weight": 10
}
]
}
}
The example above is intentionally minimal and single-threaded.
Top-level fields
| Field | Type | Description |
|---|---|---|
max_ticks |
int | Maximum simulation ticks (0 = unlimited) |
num_threads |
int | Number of PDES engine partitions/worker threads. |
exec_mode |
string | "functional" or "cycle" |
vm.arch |
string | Architecture: cdna3, cdna4, etc. |
Topology
Components are defined hierarchically under topology.root. Range
expansion (xcd[0:8]) creates multiple instances. Links connect
component ports using pattern expressions with loop variables.
KFD device section
KFD-mode configs include a vm.gpu.device section that defines the
properties reported through the simulated sysfs topology (GPU ID,
vendor/device IDs, CU counts, memory sizes, etc.). These must match
the component hierarchy defined in topology.
FlatBuffers schema
The JSON config is validated against FlatBuffers schemas in schemas/:
simulation_config.fbs— topology and simulation parameterscheckpoint.fbs— simulation state checkpointing
Multi-GPU
Multi-GPU configs define multiple SoCs with distinct GPU IDs and
location IDs. Each GPU gets its own command processor, memory, and
cache hierarchy. The daemon manages all GPUs and routes KFD ioctls
to the correct device based on gpu_id.
See configs/gfx950_cdna4_kmd_2gpu.json for a working two-GPU
configuration used by the RCCL collective tests.