github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-10 23:48:01 -05:00
Code Issues Packages Projects Releases Wiki Activity

amd doorbell size is 64bits (#4448)

Browse Source 
* amd doorbell size ids 64bits

* add test

* test to pass 32bit boundary is more correct

* no need to round there
This commit is contained in:
nimlgen
2024-05-06 16:59:59 +03:00
committed by GitHub
parent fc995d4446
commit 113c2f00b9
2 changed files with 29 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
test/external/external_test_amd.py vendored Normal file
View File

@@ -0,0 +1,27 @@
import unittest
from tinygrad import Device, Tensor
from tinygrad.engine.schedule import create_schedule
from tinygrad.runtime.ops_amd import AMDDevice
class TestAMD(unittest.TestCase):
@classmethod
def setUpClass(self):
TestAMD.d0: AMDDevice = Device["AMD"]
TestAMD.a = Tensor([0.,1.], device="AMD").realize()
TestAMD.b = self.a + 1
si = create_schedule([self.b.lazydata])[-1]
TestAMD.d0_runner = TestAMD.d0.get_runner(*si.ast)
TestAMD.b.lazydata.buffer.allocate()
def test_amd_ring_64bit_doorbell(self):
TestAMD.d0.pm4_write_pointer[0] = TestAMD.d0.pm4_write_pointer[0] + (2 << 32) - TestAMD.d0.pm4_ring.size // 4
for _ in range(2000):
TestAMD.d0_runner.clprg(TestAMD.b.lazydata.buffer._buf, TestAMD.a.lazydata.buffer._buf,
global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
TestAMD.d0_runner.clprg(TestAMD.a.lazydata.buffer._buf, TestAMD.b.lazydata.buffer._buf,
global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
assert (val:=TestAMD.a.lazydata.buffer.as_buffer().cast("f")[0]) == 4000.0, f"got val {val}"
if __name__ == "__main__":
unittest.main()

4
tinygrad/runtime/ops_amd.py
View File

@@ -554,7 +554,7 @@ class AMDDevice(Compiled):
self.sdma_read_pointer = to_mv(self.sdma_queue.read_pointer_address, 8).cast("Q")
self.sdma_write_pointer = to_mv(self.sdma_queue.write_pointer_address, 8).cast("Q")
self.sdma_doorbell = to_mv(self.doorbells + self.sdma_queue.doorbell_offset - self.doorbells_base, 4).cast("I")
self.sdma_doorbell = to_mv(self.doorbells + self.sdma_queue.doorbell_offset - self.doorbells_base, 8).cast("Q")
self.sdma_doorbell_value = 0
# PM4 Queue
@@ -572,7 +572,7 @@ class AMDDevice(Compiled):
self.pm4_read_pointer = to_mv(self.pm4_queue.read_pointer_address, 8).cast("Q")
self.pm4_write_pointer = to_mv(self.pm4_queue.write_pointer_address, 8).cast("Q")
self.pm4_doorbell = to_mv(self.doorbells + self.pm4_queue.doorbell_offset - self.doorbells_base, 4).cast("I")
self.pm4_doorbell = to_mv(self.doorbells + self.pm4_queue.doorbell_offset - self.doorbells_base, 8).cast("Q")
super().__init__(device, AMDAllocator(self), AMDCompiler(self.arch), functools.partial(AMDProgram, self))