MOCKGPU on mac os (#8520)

* tweaks for macos

* fix

* fix

* typo

* remove nvidia changes

* remove nv related changes

* change address back
This commit is contained in:
patrini32
2025-01-07 20:27:43 +03:00
committed by GitHub
parent ab3ac2b58d
commit afef69a37d
4 changed files with 5 additions and 4 deletions
+1 -1
View File
@@ -69,7 +69,7 @@ generate_comgr() {
--clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so
fixup $BASE/comgr.py
sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/comgr.py
patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'"
patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'" "'/usr/local/lib/libamd_comgr.dylib'"
sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamd_comgr.so')\_try_dlopen_amd_comgr()\g" $BASE/comgr.py
python3 -c "import tinygrad.runtime.autogen.comgr"
}
+1 -1
View File
@@ -19,7 +19,7 @@ WAIT_REG_MEM_FUNCTION_ALWAYS = 0
WAIT_REG_MEM_FUNCTION_EQ = 3 # ==
WAIT_REG_MEM_FUNCTION_GEQ = 5 # >=
REMU_PATHS = ["libremu.so", "/usr/local/lib/libremu.so"]
REMU_PATHS = ["libremu.so", "/usr/local/lib/libremu.so", "libremu.dylib", "/usr/local/lib/libremu.dylib"]
def _try_dlopen_remu():
for path in REMU_PATHS:
try:
+1
View File
@@ -10,6 +10,7 @@ import ctypes, ctypes.util, os
PATHS_TO_TRY = [
'/opt/rocm/lib/libamd_comgr.so',
os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so',
'/usr/local/lib/libamd_comgr.dylib',
]
def _try_dlopen_amd_comgr():
library = ctypes.util.find_library("amd_comgr")
+2 -2
View File
@@ -6,7 +6,7 @@ from dataclasses import dataclass
from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, HWInterface
from tinygrad.ops import sint
from tinygrad.device import BufferSpec
from tinygrad.helpers import getenv, to_mv, round_up, data64_le, mv_address, DEBUG
from tinygrad.helpers import getenv, to_mv, round_up, data64_le, mv_address, DEBUG, OSX
from tinygrad.renderer.cstyle import AMDRenderer
from tinygrad.runtime.autogen import kfd, hsa, amd_gpu, libc, libpciaccess, vfio
from tinygrad.runtime.autogen.am import am
@@ -269,7 +269,7 @@ class AMDAllocator(HCQAllocator['AMDDevice']):
def map(self, buf:HCQBuffer): self.dev.dev_iface.map(buf._base if buf._base is not None else buf)
MAP_FIXED, MAP_NORESERVE, MAP_LOCKED = 0x10, 0x400, 0x2000
MAP_FIXED, MAP_NORESERVE, MAP_LOCKED = 0x10, 0x400, 0 if OSX else 0x2000
@dataclass
class AMDQueueDesc: