commaai
/
tinygrad
mirror of https://github.com/commaai/tinygrad.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

qcom sleep when sync (#6785) 

* qcom sleep when sync

* linter

* short
This commit is contained in:
nimlgen 2024-09-27 19:14:10 +08:00 committed by GitHub
parent fb3fe6f39b
commit b95f47784a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tinygrad/device.py
View File

@ -522,7 +522,7 @@ class HCQCompiled(Compiled):
self.devices.append(self) self.devices.append(self)
def synchronize(self): def synchronize(self):
self.timeline_signal.wait(self.timeline_value - 1) self.timeline_signal.wait(self.timeline_value - 1) if not hasattr(self, '_syncdev') else self._syncdev()
if self.timeline_value > (1 << 31): self._wrap_timeline_signal() if self.timeline_value > (1 << 31): self._wrap_timeline_signal()
if PROFILE: if PROFILE:

6
tinygrad/runtime/ops_qcom.py
View File

@ -101,7 +101,7 @@ class QCOMComputeQueue(HWComputeQueue):
def _submit(self, device): def _submit(self, device):
if self.binded_device == device: submit_req = self.submit_req if self.binded_device == device: submit_req = self.submit_req
else: submit_req, _ = self._build_gpu_command(device) else: submit_req, _ = self._build_gpu_command(device)
kgsl.IOCTL_KGSL_GPU_COMMAND(device.fd, __payload=submit_req) device.last_cmd = kgsl.IOCTL_KGSL_GPU_COMMAND(device.fd, __payload=submit_req).timestamp
def _exec(self, prg, args_state, global_size, local_size): def _exec(self, prg, args_state, global_size, local_size):
global_size_mp = [int(g*l) for g,l in zip(global_size, local_size)] global_size_mp = [int(g*l) for g,l in zip(global_size, local_size)]
@ -345,7 +345,7 @@ class QCOMDevice(HCQCompiled):
QCOMDevice.dummy_addr = self._gpu_alloc(0x1000, map_to_cpu=False).va_addr QCOMDevice.dummy_addr = self._gpu_alloc(0x1000, map_to_cpu=False).va_addr
QCOMDevice.signals_page = self._gpu_alloc(16 * 65536, map_to_cpu=True, uncached=True) QCOMDevice.signals_page = self._gpu_alloc(16 * 65536, map_to_cpu=True, uncached=True)
QCOMDevice.signals_pool = [to_mv(self.signals_page.va_addr + off, 16).cast("Q") for off in range(0, self.signals_page.size, 16)] QCOMDevice.signals_pool = [to_mv(self.signals_page.va_addr + off, 16).cast("Q") for off in range(0, self.signals_page.size, 16)]
info, self.ctx, self.cmd_buf, self.cmd_buf_ptr = self._info(), self._ctx_create(), self._gpu_alloc(0x1000000, map_to_cpu=True), 0 info, self.ctx, self.cmd_buf, self.cmd_buf_ptr, self.last_cmd = self._info(), self._ctx_create(), self._gpu_alloc(0x1000000, map_to_cpu=True), 0,0
QCOMDevice.gpu_id = ((info.chip_id >> 24) & 0xFF) * 100 + ((info.chip_id >> 16) & 0xFF) * 10 + ((info.chip_id >> 8) & 0xFF) QCOMDevice.gpu_id = ((info.chip_id >> 24) & 0xFF) * 100 + ((info.chip_id >> 16) & 0xFF) * 10 + ((info.chip_id >> 8) & 0xFF)
if QCOMDevice.gpu_id >= 700: raise RuntimeError(f"Unsupported GPU: {QCOMDevice.gpu_id}") if QCOMDevice.gpu_id >= 700: raise RuntimeError(f"Unsupported GPU: {QCOMDevice.gpu_id}")
@ -398,3 +398,5 @@ class QCOMDevice(HCQCompiled):
self.synchronize() self.synchronize()
self._gpu_free(self._stack) self._gpu_free(self._stack)
self._stack = self._gpu_alloc(sz) self._stack = self._gpu_alloc(sz)
def _syncdev(self): kgsl.IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID(self.fd, context_id=self.ctx, timestamp=self.last_cmd, timeout=0xffffffff)