github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-16 01:26:29 -05:00
Code Issues Packages Projects Releases Wiki Activity

BumpAllocator rename start -> base (#8415)

Browse Source
This commit is contained in:
nimlgen
2024-12-25 23:12:55 +03:00
committed by GitHub
parent 9ed064710a
commit a562ee2c6e
4 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tinygrad/runtime/graph/hcq.py
View File

@@ -31,7 +31,7 @@ class HCQGraph(MultiGraphRunner):
# Fill initial arguments.
self.ji_args: dict[int, HCQArgsState] = {}
kargs_alloc: dict[Compiled, BumpAllocator] = {dev:BumpAllocator(buf.size, start=cast(int, buf.va_addr)) for dev,buf in self.kernargs_bufs.items()}
kargs_alloc: dict[Compiled, BumpAllocator] = {dev:BumpAllocator(buf.size, base=cast(int, buf.va_addr)) for dev,buf in self.kernargs_bufs.items()}
for j,ji in enumerate(jit_cache):
if not isinstance(ji.prg, CompiledRunner): continue

6
tinygrad/runtime/ops_nv.py
View File

@@ -292,8 +292,8 @@ class NVDevice(HCQCompiled[NVSignal]):
# TODO: Need a proper allocator for va addresses
# 0x1000000000 - 0x2000000000, reserved for system/cpu mappings
# VA space is 48bits.
low_uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=0x1000000000, start=0x1000000000, wrap=False)
uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=(1 << 48) - 1, start=0x2000000000, wrap=False)
low_uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=0x1000000000, base=0x1000000000, wrap=False)
uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=(1 << 48) - 1, base=0x2000000000, wrap=False)
host_object_enumerator: int = 0x1000
def _new_gpu_fd(self):
@@ -450,7 +450,7 @@ class NVDevice(HCQCompiled[NVSignal]):
rmctrl.gpfifo_schedule(self.fd_ctl, self.root, channel_group, bEnable=1)
self.cmdq_page:HCQBuffer = self._gpu_alloc(0x200000, cpu_access=True, tag="cmdq")
self.cmdq_allocator = BumpAllocator(size=self.cmdq_page.size, start=cast(int, self.cmdq_page.va_addr), wrap=True)
self.cmdq_allocator = BumpAllocator(size=self.cmdq_page.size, base=cast(int, self.cmdq_page.va_addr), wrap=True)
self.cmdq: memoryview = to_mv(cast(int, self.cmdq_page.va_addr), 0x200000).cast("I")
self.num_gpcs, self.num_tpc_per_gpc, self.num_sm_per_tpc, self.max_warps_per_sm, self.sm_version = self._query_gpu_info('num_gpcs',

2
tinygrad/runtime/ops_qcom.py
View File

@@ -335,7 +335,7 @@ class QCOMDevice(HCQCompiled):
self.ctx = kgsl.IOCTL_KGSL_DRAWCTXT_CREATE(self.fd, flags=flags).drawctxt_id
self.cmd_buf = self._gpu_alloc(16 << 20)
self.cmd_buf_allocator = BumpAllocator(size=self.cmd_buf.size, start=cast(int, self.cmd_buf.va_addr), wrap=True)
self.cmd_buf_allocator = BumpAllocator(size=self.cmd_buf.size, base=cast(int, self.cmd_buf.va_addr), wrap=True)
self.border_color_buf = self._gpu_alloc(0x1000, fill_zeroes=True)

6
tinygrad/runtime/support/hcq.py
View File

@@ -15,13 +15,13 @@ ArgsStateType = TypeVar('ArgsStateType', bound='HCQArgsState')
QueueType = TypeVar('QueueType', bound='HWQueue')
class BumpAllocator:
def __init__(self, size:int, start:int=0, wrap:bool=True): self.size, self.ptr, self.start_off, self.wrap = size, 0, start, wrap
def __init__(self, size:int, base:int=0, wrap:bool=True): self.size, self.ptr, self.base, self.wrap = size, 0, base, wrap
def alloc(self, size:int, alignment:int=1) -> int:
if round_up(self.ptr, alignment) + size > self.size:
if not self.wrap: raise RuntimeError("Out of memory")
self.ptr = 0
self.ptr = (res:=round_up(self.ptr, alignment)) + size
return res + self.start_off
return res + self.base
class HWQueue(Generic[SignalType, DeviceType, ProgramType, ArgsStateType]):
"""
@@ -313,7 +313,7 @@ class HCQCompiled(Compiled, Generic[SignalType]):
super().__init__(device, allocator, renderer, compiler, runtime, HCQGraph)
self.kernargs_page:HCQBuffer = self.allocator.alloc(16 << 20, BufferSpec(cpu_access=True))
self.kernargs_allocator:BumpAllocator = BumpAllocator(self.kernargs_page.size, start=cast(int, self.kernargs_page.va_addr), wrap=True)
self.kernargs_allocator:BumpAllocator = BumpAllocator(self.kernargs_page.size, base=cast(int, self.kernargs_page.va_addr), wrap=True)
self.devices.append(self)
def synchronize(self):