geohot/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests Actions 7 Activity

amd: fix ALL2ALL speed on amdgpu (gpt) (#17518)

Browse Source
This commit is contained in:
qazal
2026-08-13 16:24:45 +09:00
committed by GitHub
parent 16c5ff2490
commit cc6d33bde7
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tinygrad/runtime/ops_amd.py
+5 -1
View File
@@ -790,10 +790,14 @@ class KFDIface:
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, idx=0):
if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA and idx and self.kfd_ver >= (1,17):
queue_type = kfd.KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID
sdma_engine_id = idx % (self.props['num_sdma_engines'] + self.props.get('num_sdma_xgmi_engines', 0))
else: sdma_engine_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=getenv("AMD_KFD_QUEUE_PRIORITY", 7),
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,
ctx_save_restore_address=cwsr_buffer.va_addr if cwsr_buffer else 0, ctx_save_restore_size=ctx_save_restore_size, sdma_engine_id=sdma_engine_id,
write_pointer_address=gart.va_addr+wptr, read_pointer_address=gart.va_addr+rptr+8*xcc_id)
if not hasattr(self, 'doorbells'):