am: move queues to gpus (tinygrad#8848)

* am: fix

* add flsg for thos

* do not depend on host parameter,

Author: nimlgen

tinygrad/runtime/ops_amd.py

@@ -353,7 +353,8 @@ def alloc(self, size:int, host=False, uncached=False, cpu_access=False) -> HCQBu
 
     if cpu_access or host: flags |= kfd.KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC
 
-    if host: buf = addr = HWInterface.anon_mmap(0, size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | mmap.MAP_ANONYMOUS, 0)
+    if flags & kfd.KFD_IOC_ALLOC_MEM_FLAGS_USERPTR:
+      buf = addr = HWInterface.anon_mmap(0, size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | mmap.MAP_ANONYMOUS, 0)
     else: buf, addr = 0, HWInterface.anon_mmap(0, size, 0, mmap.MAP_PRIVATE | mmap.MAP_ANONYMOUS | MAP_NORESERVE, 0)
     assert addr != 0xffffffffffffffff
 
@@ -365,7 +366,7 @@ def alloc(self, size:int, host=False, uncached=False, cpu_access=False) -> HCQBu
       if e.errno == errno.ENOMEM: raise MemoryError("Cannot allocate memory: no memory is available.") from e
       raise
 
-    if not host:
+    if not (flags & kfd.KFD_IOC_ALLOC_MEM_FLAGS_USERPTR):
       buf = self.drm_fd.mmap(mem.va_addr, mem.size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | MAP_FIXED, mem.mmap_offset)
       assert addr == buf == mem.va_addr
 
@@ -502,7 +503,7 @@ def _map_pci_range(self, bar, off=0, addr=0, size=None):
     return to_mv(fd.mmap(addr, sz, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | (MAP_FIXED if addr else 0), off), sz)
 
   def alloc(self, size:int, host=False, uncached=False, cpu_access=False):
-    if host or (uncached and cpu_access): # host or gtt-like memory.
+    if host or (not getenv("AMD_ALLOC_QUEUE_DEV_MEM", 1) and uncached and cpu_access): # host or gtt-like memory.
       vaddr = self.adev.mm.alloc_vaddr(size:=round_up(size, mmap.PAGESIZE), align=mmap.PAGESIZE)
       va = HWInterface.anon_mmap(vaddr, size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | mmap.MAP_ANONYMOUS | MAP_LOCKED | MAP_FIXED, 0)
 
@@ -564,7 +565,7 @@ def __init__(self, device:str=""):
     if self.target < 100300 or self.target >= 120000: raise RuntimeError(f"Unsupported arch: {self.arch}")
 
     if AMDDevice.signals_page is None:
-      AMDDevice.signals_page = self.dev_iface.alloc(16 * 65536, uncached=True, cpu_access=True)
+      AMDDevice.signals_page = self.dev_iface.alloc(16 * 65536, host=True, uncached=True, cpu_access=True)
       AMDDevice.signals_pool = [AMDDevice.signals_page.va_addr + off for off in range(0, AMDDevice.signals_page.size, 16)]
     else: self.dev_iface.map(AMDDevice.signals_page)