github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-09 22:26:26 -05:00
Code Issues Packages Projects Releases Wiki Activity

amd: rptr/wptr in create_queue (#11817)

Browse Source
This commit is contained in:
nimlgen
2025-08-24 18:03:45 +03:00
committed by GitHub
parent d71444857e
commit e19f901330
1 changed files with 10 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tinygrad/runtime/ops_amd.py
View File

@@ -616,12 +616,12 @@ class KFDIface:
stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
assert stm.n_success == 1
def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
queue = kfd.AMDKFD_IOC_CREATE_QUEUE(KFDIface.kfd, ring_base_address=ring.va_addr, ring_size=ring.size, gpu_id=self.gpu_id,
queue_type=queue_type, queue_percentage=kfd.KFD_MAX_QUEUE_PERCENTAGE|(xcc_id<<8), queue_priority=kfd.KFD_MAX_QUEUE_PRIORITY,
eop_buffer_address=eop_buffer.va_addr if eop_buffer else 0, eop_buffer_size=eop_buffer.size if eop_buffer else 0, ctl_stack_size=ctl_stack_size,
ctx_save_restore_address=cwsr_buffer.va_addr if cwsr_buffer else 0, ctx_save_restore_size=ctx_save_restore_size,
write_pointer_address=gart.va_addr, read_pointer_address=gart.va_addr + 8 * (xcc_id + 1))
write_pointer_address=gart.va_addr+wptr, read_pointer_address=gart.va_addr+rptr+8*xcc_id)
if not hasattr(self, 'doorbells'):
self.doorbells_base = queue.doorbell_offset & (~0x1fff) # doorbell is two pages
@@ -666,18 +666,18 @@ class PCIIface(PCIIfaceBase):
'max_slots_scratch_cu': self.dev_impl.gc_info.gc_max_scratch_slots_per_cu, 'max_waves_per_simd': self.dev_impl.gc_info.gc_max_waves_per_simd,
'simd_arrays_per_engine': self.dev_impl.gc_info.gc_num_sa_per_se, 'lds_size_in_kb': self.dev_impl.gc_info.gc_lds_size}
def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
assert cwsr_buffer is None, "no cwsr buffer for am"
if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA:
self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_sDMA_ENGINE0), pipe=0, queue=0)
else:
self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0)
return AMDQueueDesc(ring=ring.cpu_view().view(fmt='I'), doorbells=[self.dev_impl.doorbell64.view(doorbell_index * 8, 8, fmt='Q')],
read_ptrs=[gart.cpu_view().view(size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=0x10, size=8, fmt='Q')])
read_ptrs=[gart.cpu_view().view(offset=rptr, size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=wptr, size=8, fmt='Q')])
def sleep(self, timeout):
if self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -719,9 +719,9 @@ class USBIface(PCIIface):
return HCQBuffer(am_mapping.va_addr, size, meta=PCIAllocationMeta(am_mapping, has_cpu_mapping=False),
view=USBMMIOInterface(self.usb, self.bars[0][0] + am_mapping.paddrs[0][0], size, fmt='B') if cpu_access else None, owner=self.dev)
def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE: self.usb._pci_cacheable += [(ring.cpu_view().addr, ring.size)]
return super().create_queue(queue_type, ring, gart, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
return super().create_queue(queue_type, ring, gart, rptr, wptr, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
def sleep(self, timeout): pass
@@ -812,8 +812,8 @@ class AMDDevice(HCQCompiled):
cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
eop_buffer = self.iface.alloc(eop_buffer_size) if eop_buffer_size else None
return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer, xcc_id=xcc_id,
ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, rptr=0, wptr=0x10, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
xcc_id=xcc_id, ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
for xcc_id in range(self.xccs if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE else 1)))
def _ensure_has_local_memory(self, required):