mirror of
https://github.com/tinygrad/tinygrad.git
synced 2026-04-29 03:00:14 -04:00
WIP PM4 Support (#4110)
* pm4 kernel launch works * disable USE_THREAD_DIMENSIONS * add kernel code * work on real pm4 * pm4 signal * same * gate pm4 * hcq tests pass * ops passes * pm4 is closer * pm4 debug (#4165) * start debug tests passing * prg * smth * hdp flush * cleaner 1 * do not need this * logs not need * small things * linter * remove AQL * test hcq * fix tests * it's subtracting, it shouldn't be -1 * pm4 changes (#4251) * not need this anymore * sdma signal with non atomic --------- Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
This commit is contained in:
George Hotz
41
test/external/external_test_hcq.py
vendored
41
test/external/external_test_hcq.py
vendored
@@ -2,7 +2,7 @@ import unittest, ctypes, struct, time
|
||||
from tinygrad import Device, Tensor, dtypes
|
||||
from tinygrad.buffer import Buffer, BufferOptions
|
||||
from tinygrad.engine.schedule import create_schedule
|
||||
from tinygrad.runtime.ops_kfd import KFDDevice, HWCopyQueue, HWComputeQueue
|
||||
from tinygrad.runtime.ops_kfd import KFDDevice, HWCopyQueue, HWPM4Queue
|
||||
|
||||
def _time_queue(q, d):
|
||||
st = time.perf_counter()
|
||||
@@ -15,7 +15,7 @@ class TestHCQ(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(self):
|
||||
TestHCQ.d0: KFDDevice = Device["KFD"]
|
||||
TestHCQ.d1: KFDDevice = Device["KFD:1"]
|
||||
#TestHCQ.d1: KFDDevice = Device["KFD:1"]
|
||||
TestHCQ.a = Tensor([0.,1.], device="KFD").realize()
|
||||
TestHCQ.b = self.a + 1
|
||||
si = create_schedule([self.b.lazydata])[-1]
|
||||
@@ -26,13 +26,14 @@ class TestHCQ(unittest.TestCase):
|
||||
TestHCQ.addr2 = struct.pack("QQ", TestHCQ.a.lazydata.buffer._buf.va_addr, TestHCQ.b.lazydata.buffer._buf.va_addr)
|
||||
ctypes.memmove(TestHCQ.d0.kernargs_ptr, TestHCQ.addr, len(TestHCQ.addr))
|
||||
ctypes.memmove(TestHCQ.d0.kernargs_ptr+len(TestHCQ.addr), TestHCQ.addr2, len(TestHCQ.addr2))
|
||||
TestHCQ.compute_queue = HWPM4Queue
|
||||
|
||||
def setUp(self):
|
||||
TestHCQ.a.lazydata.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
|
||||
TestHCQ.b.lazydata.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 0))))
|
||||
|
||||
def test_run_1000_times_one_submit(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
for _ in range(1000):
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr+len(TestHCQ.addr), TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
@@ -42,20 +43,21 @@ class TestHCQ(unittest.TestCase):
|
||||
assert (val:=TestHCQ.a.lazydata.buffer.as_buffer().cast("f")[0]) == 2000.0, f"got val {val}"
|
||||
|
||||
def test_run_1000_times(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr+len(TestHCQ.addr), TestHCQ.runner.global_size,
|
||||
TestHCQ.runner.local_size, TestHCQ.d0.completion_signal)
|
||||
for _ in range(1000):
|
||||
q.submit(TestHCQ.d0)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
TestHCQ.d0.completion_signal.value = 1
|
||||
# confirm signal was reset
|
||||
with self.assertRaises(RuntimeError):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=50)
|
||||
assert (val:=TestHCQ.a.lazydata.buffer.as_buffer().cast("f")[0]) == 2000.0, f"got val {val}"
|
||||
|
||||
def test_run_to_3(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr+len(TestHCQ.addr), TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size, TestHCQ.d0.completion_signal)
|
||||
@@ -65,12 +67,12 @@ class TestHCQ(unittest.TestCase):
|
||||
|
||||
def test_wait_signal(self):
|
||||
TestHCQ.d0.completion_signal.value = 1
|
||||
HWComputeQueue().wait(TestHCQ.d0.completion_signal).signal(TestHCQ.d0.completion_signal).submit(TestHCQ.d0)
|
||||
TestHCQ.compute_queue().wait(TestHCQ.d0.completion_signal).signal(TestHCQ.d0.completion_signal).submit(TestHCQ.d0)
|
||||
with self.assertRaises(RuntimeError):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=50)
|
||||
# clean up
|
||||
TestHCQ.d0.completion_signal.value = 0
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=1000)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=1000, skip_check=True)
|
||||
|
||||
def test_wait_copy_signal(self):
|
||||
TestHCQ.d0.completion_signal.value = 1
|
||||
@@ -79,25 +81,26 @@ class TestHCQ(unittest.TestCase):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=50)
|
||||
# clean up
|
||||
TestHCQ.d0.completion_signal.value = 0
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=1000)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=1000, skip_check=True)
|
||||
|
||||
def test_run_normal(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size, TestHCQ.d0.completion_signal)
|
||||
q.submit(TestHCQ.d0)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
assert (val:=TestHCQ.b.lazydata.buffer.as_buffer().cast("f")[0]) == 1.0, f"got val {val}"
|
||||
|
||||
def test_submit_empty_queues(self):
|
||||
HWComputeQueue().submit(TestHCQ.d0)
|
||||
TestHCQ.compute_queue().submit(TestHCQ.d0)
|
||||
HWCopyQueue().submit(TestHCQ.d0)
|
||||
|
||||
def test_signal_timeout(self):
|
||||
TestHCQ.d0.completion_signal.value = 1
|
||||
with self.assertRaises(RuntimeError):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=50)
|
||||
|
||||
def test_signal(self):
|
||||
HWComputeQueue().signal(TestHCQ.d0.completion_signal).submit(TestHCQ.d0)
|
||||
TestHCQ.compute_queue().signal(TestHCQ.d0.completion_signal).submit(TestHCQ.d0)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
|
||||
def test_copy_signal(self):
|
||||
@@ -105,7 +108,7 @@ class TestHCQ(unittest.TestCase):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
|
||||
def test_run_signal(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size)
|
||||
q.signal(TestHCQ.d0.completion_signal)
|
||||
q.submit(TestHCQ.d0)
|
||||
@@ -117,9 +120,10 @@ class TestHCQ(unittest.TestCase):
|
||||
q.copy(TestHCQ.a.lazydata.buffer._buf.va_addr, TestHCQ.b.lazydata.buffer._buf.va_addr, 8)
|
||||
q.copy(TestHCQ.b.lazydata.buffer._buf.va_addr, TestHCQ.a.lazydata.buffer._buf.va_addr, 8)
|
||||
q.signal(TestHCQ.d0.completion_signal)
|
||||
for _ in range(1000):
|
||||
for i in range(1000):
|
||||
q.submit(TestHCQ.d0)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
TestHCQ.d0.completion_signal.value = 1
|
||||
# confirm signal was reset
|
||||
with self.assertRaises(RuntimeError):
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal, timeout=50)
|
||||
@@ -147,8 +151,8 @@ class TestHCQ(unittest.TestCase):
|
||||
|
||||
def test_cross_device_copy_bandwidth(self):
|
||||
SZ = 2_000_000_000
|
||||
a = Buffer("KFD", SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
|
||||
b = Buffer("KFD:1", SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
|
||||
a = Buffer("KFD", SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
|
||||
TestHCQ.d0._gpu_map(b._buf)
|
||||
q = HWCopyQueue()
|
||||
q.copy(a._buf.va_addr, b._buf.va_addr, SZ)
|
||||
@@ -158,7 +162,7 @@ class TestHCQ(unittest.TestCase):
|
||||
assert gb_s > 2 and gb_s < 50
|
||||
|
||||
def test_interleave_compute_and_copy(self):
|
||||
q = HWComputeQueue()
|
||||
q = TestHCQ.compute_queue()
|
||||
qc = HWCopyQueue()
|
||||
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.global_size, TestHCQ.runner.local_size) # b = [1, 2]
|
||||
q.signal(sig:=KFDDevice._get_signal(10))
|
||||
@@ -173,12 +177,13 @@ class TestHCQ(unittest.TestCase):
|
||||
assert (val:=TestHCQ.a.lazydata.buffer.as_buffer().cast("f")[0]) == 1.0, f"got val {val}"
|
||||
|
||||
def test_cross_device_signal(self):
|
||||
q1 = HWComputeQueue()
|
||||
q2 = HWComputeQueue()
|
||||
d1 = Device["KFD:1"]
|
||||
q1 = TestHCQ.compute_queue()
|
||||
q2 = TestHCQ.compute_queue()
|
||||
q1.signal(TestHCQ.d0.completion_signal)
|
||||
q2.wait(TestHCQ.d0.completion_signal)
|
||||
q2.submit(TestHCQ.d0)
|
||||
q1.submit(TestHCQ.d1)
|
||||
q1.submit(d1)
|
||||
TestHCQ.d0._wait_signal(TestHCQ.d0.completion_signal)
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
Reference in New Issue
Block a user