github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-04-29 03:00:14 -04:00
Code Issues Packages Projects Releases Wiki Activity

add _alloc_signal/_free_signal to hcq (#5264)

Browse Source 
* add _alloc_signal/_free_signal api

* oops, revert this

* linter
This commit is contained in:
nimlgen
2024-07-02 23:35:39 +03:00
committed by GitHub
parent 9a25ee0b9a
commit 7be776f9af
7 changed files with 54 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
test/external/external_test_hcq.py vendored
View File

@@ -51,7 +51,7 @@ class TestHCQ(unittest.TestCase):
TestHCQ.d0.synchronize() # wait for copyins to complete
def test_run_1000_times_one_submit(self):
temp_signal, temp_value = TestHCQ.d0._get_signal(value=0), 0
temp_signal, temp_value = TestHCQ.d0._alloc_signal(value=0), 0
q = TestHCQ.compute_queue()
for _ in range(1000):
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
@@ -69,7 +69,7 @@ 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):
temp_signal = TestHCQ.d0._get_signal(value=0)
temp_signal = TestHCQ.d0._alloc_signal(value=0)
q = TestHCQ.compute_queue()
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
q.signal(temp_signal, 2).wait(temp_signal, 2)
@@ -84,7 +84,7 @@ class TestHCQ(unittest.TestCase):
assert (val:=TestHCQ.a.lazydata.buffer.as_buffer().cast("f")[0]) == 2000.0, f"got val {val}"
def test_run_to_3(self):
temp_signal = TestHCQ.d0._get_signal(value=0)
temp_signal = TestHCQ.d0._alloc_signal(value=0)
q = TestHCQ.compute_queue()
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
q.signal(temp_signal, 1).wait(temp_signal, 1)
@@ -109,7 +109,7 @@ class TestHCQ(unittest.TestCase):
@unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
def test_bind_run(self):
temp_signal = TestHCQ.d0._get_signal(value=0)
temp_signal = TestHCQ.d0._alloc_signal(value=0)
q = TestHCQ.compute_queue()
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
q.signal(temp_signal, 2).wait(temp_signal, 2)
@@ -141,7 +141,7 @@ class TestHCQ(unittest.TestCase):
@unittest.skipIf(CI, "Can't handle async update on CPU")
def test_wait_signal(self):
temp_signal = TestHCQ.d0._get_signal(value=0)
temp_signal = TestHCQ.d0._alloc_signal(value=0)
TestHCQ.compute_queue().wait(temp_signal, value=1).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
with self.assertRaises(RuntimeError):
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value, timeout=50)
@@ -152,7 +152,7 @@ class TestHCQ(unittest.TestCase):
@unittest.skipIf(CI, "Can't handle async update on CPU")
def test_wait_copy_signal(self):
temp_signal = TestHCQ.d0._get_signal(value=0)
temp_signal = TestHCQ.d0._alloc_signal(value=0)
TestHCQ.copy_queue().wait(temp_signal, value=1).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
with self.assertRaises(RuntimeError):
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value, timeout=50)
@@ -267,7 +267,7 @@ class TestHCQ(unittest.TestCase):
q = TestHCQ.compute_queue()
qc = TestHCQ.copy_queue()
q.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) # b = [1, 2]
q.signal(sig:=TestHCQ.d0._get_signal(value=0), value=1)
q.signal(sig:=TestHCQ.d0._alloc_signal(value=0), value=1)
qc.wait(sig, value=1)
qc.copy(TestHCQ.a.lazydata.buffer._buf.va_addr, TestHCQ.b.lazydata.buffer._buf.va_addr, 8)
qc.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
@@ -282,7 +282,7 @@ class TestHCQ(unittest.TestCase):
d1 = Device[f"{Device.DEFAULT}:1"]
q1 = TestHCQ.compute_queue()
q2 = TestHCQ.compute_queue()
q1.signal(sig:=TestHCQ.d0._get_signal(value=0), value=0xfff)
q1.signal(sig:=TestHCQ.d0._alloc_signal(value=0), value=0xfff)
q2.wait(sig, value=0xfff)
q2.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
q2.submit(TestHCQ.d0)

2
test/external/fuzz_kfd.py vendored
View File

@@ -16,7 +16,7 @@ if __name__ == "__main__":
d2, b2 = random.choice(buffers)
d1._gpu_map(b2)
q = HWComputeQueue()
q.signal(sig:=AMDDevice._get_signal(10))
q.signal(sig:=AMDDevice._alloc_signal(10))
qc = HWCopyQueue()
qc.wait(sig)
qc.copy(b1.va_addr, b2.va_addr, min(b1.size, b2.size))

39
test/test_hcq.py
View File

@@ -48,7 +48,7 @@ class TestHCQ(unittest.TestCase):
def test_signal_update(self):
for queue_type in [TestHCQ.d0.hw_compute_queue_t]:
with self.subTest(name=str(queue_type)):
q = queue_type().signal(fake_signal := TestHCQ.d0._get_signal(), 0x1000)
q = queue_type().signal(fake_signal := TestHCQ.d0._alloc_signal(), 0x1000)
q.update_signal(0, signal=TestHCQ.d0.timeline_signal, value=TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
@@ -58,26 +58,26 @@ class TestHCQ(unittest.TestCase):
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
TestHCQ.d0.timeline_value += 1
TestHCQ.d0.signals_pool.append(fake_signal)
TestHCQ.d0._free_signal(fake_signal)
# Test wait
def test_wait(self):
for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
with self.subTest(name=str(queue_type)):
fake_signal = TestHCQ.d0._get_signal()
fake_signal = TestHCQ.d0._alloc_signal()
TestHCQ.d0._set_signal(fake_signal, 1)
queue_type().wait(fake_signal, 1) \
.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
TestHCQ.d0.timeline_value += 1
TestHCQ.d0.signals_pool.append(fake_signal)
TestHCQ.d0._free_signal(fake_signal)
@unittest.skipIf(MOCKGPU, "Can't handle async update on MOCKGPU for now")
def test_wait_late_set(self):
for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
with self.subTest(name=str(queue_type)):
fake_signal = TestHCQ.d0._get_signal()
fake_signal = TestHCQ.d0._alloc_signal()
queue_type().wait(fake_signal, 1) \
.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
@@ -89,12 +89,12 @@ class TestHCQ(unittest.TestCase):
TestHCQ.d0.timeline_value += 1
TestHCQ.d0.signals_pool.append(fake_signal)
TestHCQ.d0._free_signal(fake_signal)
def test_wait_update(self):
for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
with self.subTest(name=str(queue_type)):
fake_signal = TestHCQ.d0._get_signal()
fake_signal = TestHCQ.d0._alloc_signal()
q = queue_type().wait(TestHCQ.d0.timeline_signal, 0xffffffff).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
TestHCQ.d0._set_signal(fake_signal, 0x30)
@@ -103,7 +103,7 @@ class TestHCQ(unittest.TestCase):
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
TestHCQ.d0.timeline_value += 1
TestHCQ.d0.signals_pool.append(fake_signal)
TestHCQ.d0._free_signal(fake_signal)
# Test exec
def test_exec_one_kernel(self):
@@ -167,7 +167,7 @@ class TestHCQ(unittest.TestCase):
with self.subTest(name=str(queue_type)):
if not hasattr(queue_type(), 'bind'): self.skipTest("queue does not support bind api")
fake_signal = TestHCQ.d0._get_signal()
fake_signal = TestHCQ.d0._alloc_signal()
q = queue_type().wait(TestHCQ.d0.timeline_signal, 0xffffffff).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
q.bind(TestHCQ.d0)
@@ -177,13 +177,13 @@ class TestHCQ(unittest.TestCase):
TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
TestHCQ.d0.timeline_value += 1
TestHCQ.d0.signals_pool.append(fake_signal)
TestHCQ.d0._free_signal(fake_signal)
# Test multidevice
def test_multidevice_signal_wait(self):
d1 = Device[f"{Device.DEFAULT}:1"]
TestHCQ.d0.hw_copy_queue_t().signal(sig:=TestHCQ.d0._get_signal(value=0), value=0xfff) \
TestHCQ.d0.hw_copy_queue_t().signal(sig:=TestHCQ.d0._alloc_signal(value=0), value=0xfff) \
.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
d1.hw_copy_queue_t().wait(sig, value=0xfff) \
@@ -195,13 +195,13 @@ class TestHCQ(unittest.TestCase):
d1._wait_signal(d1.timeline_signal, d1.timeline_value)
d1.timeline_value += 1
TestHCQ.d0.signals_pool.append(sig)
TestHCQ.d0._free_signal(sig)
# Test profile api
def test_speed_exec_time(self):
TestHCQ.d0._prof_setup()
sig_st, sig_en = TestHCQ.d0._get_signal(), TestHCQ.d0._get_signal()
sig_st, sig_en = TestHCQ.d0._alloc_signal(), TestHCQ.d0._alloc_signal()
TestHCQ.d0.hw_compute_queue_t().timestamp(sig_st) \
.exec(TestHCQ.runner.clprg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) \
.timestamp(sig_en) \
@@ -212,7 +212,8 @@ class TestHCQ(unittest.TestCase):
et = TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_en), True) - TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_st), True)
TestHCQ.d0.signals_pool += [sig_st, sig_en]
TestHCQ.d0._free_signal(sig_st)
TestHCQ.d0._free_signal(sig_en)
print(f"exec kernel time: {et:.2f} us")
assert 1 <= et <= (2000 if CI else 20)
@@ -225,7 +226,7 @@ class TestHCQ(unittest.TestCase):
a = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
b = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
sig_st, sig_en = TestHCQ.d0._get_signal(), TestHCQ.d0._get_signal()
sig_st, sig_en = TestHCQ.d0._alloc_signal(), TestHCQ.d0._alloc_signal()
TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
.copy(a._buf.va_addr, b._buf.va_addr, SZ) \
.timestamp(sig_en) \
@@ -237,7 +238,8 @@ class TestHCQ(unittest.TestCase):
et = TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_en), True) - TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_st), True)
et_ms = et / 1e3
TestHCQ.d0.signals_pool += [sig_st, sig_en]
TestHCQ.d0._free_signal(sig_st)
TestHCQ.d0._free_signal(sig_en)
gb_s = ((SZ / 1e9) / et_ms) * 1e3
print(f"same device copy: {et_ms:.2f} ms, {gb_s:.2f} GB/s")
@@ -251,7 +253,7 @@ class TestHCQ(unittest.TestCase):
a = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferOptions(nolru=True)).allocate()
TestHCQ.d0._gpu_map(b._buf)
sig_st, sig_en = TestHCQ.d0._get_signal(), TestHCQ.d0._get_signal()
sig_st, sig_en = TestHCQ.d0._alloc_signal(), TestHCQ.d0._alloc_signal()
TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
.copy(a._buf.va_addr, b._buf.va_addr, SZ) \
.timestamp(sig_en) \
@@ -263,7 +265,8 @@ class TestHCQ(unittest.TestCase):
et = TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_en), True) - TestHCQ.d0._gpu2cpu_time(TestHCQ.d0._read_timestamp(sig_st), True)
et_ms = et / 1e3
TestHCQ.d0.signals_pool += [sig_st, sig_en]
TestHCQ.d0._free_signal(sig_st)
TestHCQ.d0._free_signal(sig_en)
gb_s = ((SZ / 1e9) / et_ms) * 1e3
print(f"cross device copy: {et_ms:.2f} ms, {gb_s:.2f} GB/s")