mirror of
https://jihulab.com/mr-one/openpilot.git
synced 2026-10-27 21:13:42 +08:00
Comma Device
21c4dd2b41
Applied upstream commit 94a32493e (modeld_v2 chestnut support). Skipped CI workflow files (deleted in this fork). Replaced vendored tinygrad_repo with submodule commit 66ee3cfb4. Rebuilt libparams_c.so for new ModelManager/USBGPU params.
58 lines
3.2 KiB
Python
58 lines
3.2 KiB
Python
import inspect, math
|
|
from tinygrad.device import Compiled, Allocator, ProfileGraphEntry, ProfileGraphEvent, Program, TinyELF
|
|
from tinygrad.engine.jit import MultiGraphRunner
|
|
from tinygrad.renderer import Renderer, cstyle, nir, ptx, llvmir, wgsl
|
|
from tinygrad.renderer.cstyle import CStyleLanguage
|
|
from tinygrad.uop.ops import UOp, Ops
|
|
from tinygrad.helpers import cpu_profile, getenv, dedup, NULL_ALLOW_COPYOUT, PROFILE, cpu_events, perf_counter_us
|
|
|
|
class NullRenderer(CStyleLanguage):
|
|
has_local = False
|
|
float4 = "float4"
|
|
barrier = "// BARRIER"
|
|
code_for_op = {**CStyleLanguage.code_for_op, Ops.THREEFRY: lambda a,b,dtype: f"threefry({a},{b})", Ops.MAX: lambda a,b,dtype: f"max({a},{b})"}
|
|
def asm(self, prg: UOp, lin: UOp) -> bytes:
|
|
assert self.target.arch.startswith("gfx"), "only amd supports assembly"
|
|
from tinygrad.renderer.amd.elf import assemble_linear
|
|
return assemble_linear(prg, lin, self.target.arch)
|
|
|
|
class NullProgram(Program['NullDevice']):
|
|
def __init__(self, dev:'NullDevice', obj:TinyELF): self.device, self.name = dev.device, obj.name
|
|
def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False, **kw):
|
|
with cpu_profile(self.name, self.device): return 1e-3
|
|
|
|
class NullAllocator(Allocator['NullDevice']):
|
|
def _alloc(self, size, options): pass
|
|
def _copyin(self, dest, src:memoryview): pass
|
|
def _copyout(self, dest:memoryview, src):
|
|
if not NULL_ALLOW_COPYOUT: raise RuntimeError("no copyout on NULL")
|
|
def _transfer(self, dest, src, sz:int, src_dev, dest_dev):
|
|
with cpu_profile(f"{src_dev.device} -> {dest_dev.device}", f"{src_dev.device}:SDMA:0"): pass
|
|
def _offset(self, buf, size:int, offset:int): pass
|
|
|
|
class NullGraph(MultiGraphRunner):
|
|
def __call__(self, input_uops:tuple[UOp, ...], var_vals:dict[str, int], wait=False) -> float|None:
|
|
if PROFILE:
|
|
st, descs = perf_counter_us(), []
|
|
event_count:dict[str, int] = {}
|
|
for (_,_,bufs,_),runtime in zip(self.calls, self.runtimes):
|
|
# description based on command, copied from HCQ graph
|
|
device = runtime.device if runtime is not None else f"{bufs[1].device}:SDMA:0"
|
|
descs.append((device, runtime.name if runtime is not None else f"{bufs[1].device} -> {bufs[0].device}", count:=event_count.get(device, 0)))
|
|
event_count[device] = count+1
|
|
# pack events evenly per device
|
|
dur, sigs, ents = max(1, math.ceil((perf_counter_us()-st)/max(event_count.values()))), [], []
|
|
for i,(device,name,count) in enumerate(descs):
|
|
sigs += [st+count*dur, st+(count+1)*dur]
|
|
ents.append(ProfileGraphEntry(device, name, 2*i, 2*i+1))
|
|
cpu_events.append(ProfileGraphEvent(ents, [], sigs))
|
|
return 1e-1
|
|
|
|
class NullDevice(Compiled):
|
|
def __init__(self, device:str):
|
|
assert (emu:=getenv("EMULATE", "")) == "", \
|
|
"EMULATE is deprecated, use DEV=NULL:HIP:"+{"AMD":"gfx1100", "AMD_RDNA4":"gfx1201", "AMD_CDNA4":"gfx950"}.get(emu, "<arch>")
|
|
renderers = [NullRenderer] + [r for m in [cstyle, nir, ptx, llvmir, wgsl] for r in m.__dict__.values()
|
|
if inspect.isclass(r) and issubclass(r, Renderer)]
|
|
super().__init__(device, NullAllocator(self), dedup(renderers), NullProgram, NullGraph)
|