def test_invalid_create_shm(self):
# Raises error since tried to create invalid cuda shared memory region
try:
shm_op0_handle = shm.create_shared_memory_region("dummy_data", "/dummy_data", -1)
shm.destroy_shared_memory_region(shm_op0_handle)
except Exception as ex:
self.assertTrue(str(ex) == "unable to initialize the size")
def test_unregister_before_register(self):
# Create a valid cuda shared memory region and unregister before register
shared_memory_ctx = SharedMemoryControlContext(_url, _protocol, verbose=True)
shm_op0_handle = shm.create_shared_memory_region("dummy_data", "/dummy_data", 8)
shared_memory_ctx.unregister(shm_op0_handle)
shm_status = shared_memory_ctx.get_shared_memory_status()
self.assertTrue(len(shm_status.shared_memory_region) == 0)
shm.destroy_shared_memory_region(shm_op0_handle)
def test_valid_create_set_register(self):
# Create a valid cuda shared memory region, fill data in it and register
shared_memory_ctx = SharedMemoryControlContext(_url, _protocol, verbose=True)
shm_op0_handle = shm.create_shared_memory_region("dummy_data", "/dummy_data", 8)
shm.set_shared_memory_region(shm_op0_handle, [np.array([1,2], dtype=np.float32)])
shared_memory_ctx.register(shm_op0_handle)
shm_status = shared_memory_ctx.get_shared_memory_status()
self.assertTrue(len(shm_status.shared_memory_region) == 1)
shm.destroy_shared_memory_region(shm_op0_handle)
def cleanup_shm_regions(self, shm_handles):
if len(shm_handles) != 0:
shared_memory_ctx = SharedMemoryControlContext("localhost:8000", ProtocolType.HTTP, verbose=True)
for shm_tmp_handle in shm_handles:
shared_memory_ctx.unregister(shm_tmp_handle)
if _test_system_shared_memory:
shm.destroy_shared_memory_region(shm_tmp_handle)
elif _test_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_tmp_handle)
def test_reregister_after_register(self):
# Create a valid cuda shared memory region and unregister after register
shared_memory_ctx = SharedMemoryControlContext(_url, _protocol, verbose=True)
shm_op0_handle = shm.create_shared_memory_region("dummy_data", "/dummy_data", 8)
shared_memory_ctx.register(shm_op0_handle)
try:
shared_memory_ctx.register(shm_op0_handle)
except Exception as ex:
self.assertTrue("shared memory region 'dummy_data' already in manager" in str(ex))
shm_status = shared_memory_ctx.get_shared_memory_status()
self.assertTrue(len(shm_status.shared_memory_region) == 1)
shm.destroy_shared_memory_region(shm_op0_handle)
def unregister_cleanup_shm_regions(shm_handles, precreated_shm_regions,
outputs):
shared_memory_ctx = SharedMemoryControlContext("localhost:8000",
ProtocolType.HTTP,
verbose=False)
shared_memory_ctx.unregister(shm_handles[0])
shared_memory_ctx.unregister(shm_handles[1])
if TEST_CUDA_SHARED_MEMORY:
cudashm.destroy_shared_memory_region(shm_handles[0])
cudashm.destroy_shared_memory_region(shm_handles[1])
else:
shm.destroy_shared_memory_region(shm_handles[0])
shm.destroy_shared_memory_region(shm_handles[1])
if precreated_shm_regions is None:
i = 0
if "OUTPUT0" in outputs:
shared_memory_ctx.unregister(shm_handles[2])
if TEST_CUDA_SHARED_MEMORY:
cudashm.destroy_shared_memory_region(shm_handles[2])
else:
shm.destroy_shared_memory_region(shm_handles[2])
i += 1
if "OUTPUT1" in outputs:
shared_memory_ctx.unregister(shm_handles[2 + i])
if TEST_CUDA_SHARED_MEMORY:
cudashm.destroy_shared_memory_region(shm_handles[2 + i])
else:
shm.destroy_shared_memory_region(shm_handles[2 + i])
def unregister_cleanup_shm_regions(shm_handles, precreated_shm_regions, outputs, use_system_shared_memory, use_cuda_shared_memory):
if not (use_system_shared_memory or use_cuda_shared_memory):
return None
shared_memory_ctx = SharedMemoryControlContext("localhost:8000", ProtocolType.HTTP, verbose=False)
shared_memory_ctx.unregister(shm_handles[0])
shared_memory_ctx.unregister(shm_handles[1])
if use_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_handles[0])
cudashm.destroy_shared_memory_region(shm_handles[1])
else:
shm.destroy_shared_memory_region(shm_handles[0])
shm.destroy_shared_memory_region(shm_handles[1])
if precreated_shm_regions is None:
i = 0
if "OUTPUT0" in outputs:
shared_memory_ctx.unregister(shm_handles[2])
if use_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_handles[2])
else:
shm.destroy_shared_memory_region(shm_handles[2])
i +=1
if "OUTPUT1" in outputs:
shared_memory_ctx.unregister(shm_handles[2+i])
if use_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_handles[2+i])
else:
shm.destroy_shared_memory_region(shm_handles[2+i])
# Fill data in shared memory region
shm.set_shared_memory_region(shm_op0_handle, [np.array([1, 2])])
# Unregister before register does not fail - does nothing
shared_memory_ctx.unregister(shm_op0_handle)
# Test if register is working
shared_memory_ctx.register(shm_op0_handle)
# Raises error if registering already registered region
try:
shared_memory_ctx.register(shm_op0_handle)
except Exception as ex:
assert "shared memory block 'dummy_data' already in manager" in str(ex)
# unregister after register
shared_memory_ctx.unregister(shm_op0_handle)
shm.destroy_shared_memory_region(shm_op0_handle)
shm_op0_handle = shm.create_shared_memory_region("output0_data", "/output0",
64)
shm_op1_handle = shm.create_shared_memory_region("output1_data", "/output1",
64)
shm_ip0_handle = shm.create_shared_memory_region("input0_data", "/input0", 64)
shm_ip1_handle = shm.create_shared_memory_region("input1_data", "/input1", 64)
input0_data = np.arange(start=0, stop=16, dtype=np.int32)
input1_data = np.ones(shape=16, dtype=np.int32)
shm.set_shared_memory_region(shm_ip0_handle, [input0_data])
shm.set_shared_memory_region(shm_ip1_handle, [input1_data])
shared_memory_ctx.register(shm_ip0_handle)
shared_memory_ctx.register(shm_ip1_handle)
shared_memory_ctx.register(shm_op0_handle)
shared_memory_ctx.register(shm_op1_handle)
def infer_exact(tester,
pf,
tensor_shape,
batch_size,
input_dtype,
output0_dtype,
output1_dtype,
output0_raw=True,
output1_raw=True,
model_version=None,
swap=False,
outputs=("OUTPUT0", "OUTPUT1"),
use_http=True,
use_grpc=True,
skip_request_id_check=False,
use_streaming=True,
correlation_id=0,
shm_region_names=None):
tester.assertTrue(use_http or use_grpc or use_streaming)
configs = []
if use_http:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8000", ProtocolType.HTTP, False, True))
else:
configs.append(("localhost:8000", ProtocolType.HTTP, False, False))
if use_grpc:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8001", ProtocolType.GRPC, False, True))
else:
configs.append(("localhost:8001", ProtocolType.GRPC, False, False))
if use_streaming:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8001", ProtocolType.GRPC, True, True))
else:
configs.append(("localhost:8001", ProtocolType.GRPC, True, False))
for config in configs:
model_name = tu.get_model_name(pf, input_dtype, output0_dtype,
output1_dtype)
# outputs are sum and difference of inputs so set max input
# values so that they will not overflow the output. This
# allows us to do an exact match. For float types use 8, 16,
# 32 int range for fp 16, 32, 64 respectively. When getting
# class outputs the result value/probability is returned as a
# float so must use fp32 range in that case.
rinput_dtype = _range_repr_dtype(input_dtype)
routput0_dtype = _range_repr_dtype(
output0_dtype if output0_raw else np.float32)
routput1_dtype = _range_repr_dtype(
output1_dtype if output1_raw else np.float32)
val_min = max(
np.iinfo(rinput_dtype).min,
np.iinfo(routput0_dtype).min,
np.iinfo(routput1_dtype).min) / 2
val_max = min(
np.iinfo(rinput_dtype).max,
np.iinfo(routput0_dtype).max,
np.iinfo(routput1_dtype).max) / 2
num_classes = 3
input0_list = list()
input1_list = list()
expected0_list = list()
expected1_list = list()
expected0_val_list = list()
expected1_val_list = list()
for b in range(batch_size):
in0 = np.random.randint(low=val_min,
high=val_max,
size=tensor_shape,
dtype=rinput_dtype)
in1 = np.random.randint(low=val_min,
high=val_max,
size=tensor_shape,
dtype=rinput_dtype)
if input_dtype != np.object:
in0 = in0.astype(input_dtype)
in1 = in1.astype(input_dtype)
if not swap:
op0 = in0 + in1
op1 = in0 - in1
else:
op0 = in0 - in1
op1 = in0 + in1
expected0_val_list.append(op0)
expected1_val_list.append(op1)
if output0_dtype == np.object:
expected0_list.append(
np.array([
unicode(str(x), encoding='utf-8')
for x in (op0.flatten())
],
dtype=object).reshape(op0.shape))
else:
expected0_list.append(op0.astype(output0_dtype))
if output1_dtype == np.object:
expected1_list.append(
np.array([
unicode(str(x), encoding='utf-8')
for x in (op1.flatten())
],
dtype=object).reshape(op1.shape))
else:
expected1_list.append(op1.astype(output1_dtype))
if input_dtype == np.object:
in0n = np.array([str(x) for x in in0.reshape(in0.size)],
dtype=object)
in0 = in0n.reshape(in0.shape)
in1n = np.array([str(x) for x in in1.reshape(in1.size)],
dtype=object)
in1 = in1n.reshape(in1.shape)
input0_list.append(in0)
input1_list.append(in1)
if config[3]:
input0_byte_size = input0_list[0].size * input0_list[
0].itemsize * batch_size
output0_byte_size = expected0_list[0].size * expected0_list[
0].itemsize * batch_size
output1_byte_size = expected1_list[0].size * expected1_list[
0].itemsize * batch_size
# create and register shared memory region for inputs and outputs
if shm_region_names is None:
shm_ip0_handle = shm.create_shared_memory_region(
"input0_data", "/input0", input0_byte_size)
shm_ip1_handle = shm.create_shared_memory_region(
"input1_data", "/input1", input0_byte_size)
if "OUTPUT0" in outputs:
shm_op0_handle = shm.create_shared_memory_region(
"output0_data", "/output0", output0_byte_size)
if "OUTPUT1" in outputs:
shm_op1_handle = shm.create_shared_memory_region(
"output1_data", "/output1", output1_byte_size)
else:
shm_ip0_handle = shm.create_shared_memory_region(
shm_region_names[0] + '_data', '/' + shm_region_names[0],
input0_byte_size)
shm_ip1_handle = shm.create_shared_memory_region(
shm_region_names[1] + '_data', '/' + shm_region_names[1],
input0_byte_size)
i = 0
if "OUTPUT0" in outputs:
shm_op0_handle = shm.create_shared_memory_region(
shm_region_names[2] + '_data',
'/' + shm_region_names[2], output0_byte_size)
i += 1
if "OUTPUT1" in outputs:
shm_op1_handle = shm.create_shared_memory_region(
shm_region_names[2 + i] + '_data',
'/' + shm_region_names[2 + i], output1_byte_size)
# copy data into shared memory region for input values
shm.set_shared_memory_region(shm_ip0_handle, input0_list)
shm.set_shared_memory_region(shm_ip1_handle, input1_list)
shared_memory_ctx = SharedMemoryControlContext(config[0],
config[1],
verbose=True)
shared_memory_ctx.register(shm_ip0_handle)
shared_memory_ctx.register(shm_ip1_handle)
if "OUTPUT0" in outputs:
shared_memory_ctx.register(shm_op0_handle)
if "OUTPUT1" in outputs:
shared_memory_ctx.register(shm_op1_handle)
expected0_sort_idx = [
np.flip(np.argsort(x.flatten()), 0) for x in expected0_val_list
]
expected1_sort_idx = [
np.flip(np.argsort(x.flatten()), 0) for x in expected1_val_list
]
output_req = {}
OUTPUT0 = "OUTPUT0"
OUTPUT1 = "OUTPUT1"
INPUT0 = "INPUT0"
INPUT1 = "INPUT1"
if pf == "libtorch" or pf == "libtorch_nobatch":
OUTPUT0 = "OUTPUT__0"
OUTPUT1 = "OUTPUT__1"
INPUT0 = "INPUT__0"
INPUT1 = "INPUT__1"
if "OUTPUT0" in outputs:
if config[3]:
output_req[OUTPUT0] = (InferContext.ResultFormat.RAW,
shm_op0_handle)
else:
if output0_raw:
output_req[OUTPUT0] = InferContext.ResultFormat.RAW
else:
output_req[OUTPUT0] = (InferContext.ResultFormat.CLASS,
num_classes)
if "OUTPUT1" in outputs:
if config[3]:
output_req[OUTPUT1] = (InferContext.ResultFormat.RAW,
shm_op1_handle)
else:
if output1_raw:
output_req[OUTPUT1] = InferContext.ResultFormat.RAW
else:
output_req[OUTPUT1] = (InferContext.ResultFormat.CLASS,
num_classes)
ctx = InferContext(config[0],
config[1],
model_name,
model_version,
correlation_id=correlation_id,
streaming=config[2],
verbose=True)
if config[3]:
results = ctx.run({
INPUT0: shm_ip0_handle,
INPUT1: shm_ip1_handle
}, output_req, batch_size)
else:
results = ctx.run({
INPUT0: input0_list,
INPUT1: input1_list
}, output_req, batch_size)
if not skip_request_id_check:
global _seen_request_ids
request_id = ctx.get_last_request_id()
tester.assertFalse(request_id in _seen_request_ids)
_seen_request_ids.add(request_id)
tester.assertEqual(ctx.get_last_request_model_name(), model_name)
if model_version is not None:
tester.assertEqual(ctx.get_last_request_model_version(),
model_version)
tester.assertEqual(len(results), len(outputs))
for (result_name, result_val) in iteritems(results):
for b in range(batch_size):
if ((result_name == OUTPUT0 and output0_raw)
or (result_name == OUTPUT1 and output1_raw)):
if result_name == OUTPUT0:
tester.assertTrue(
np.array_equal(result_val[b], expected0_list[b]),
"{}, {} expected: {}, got {}".format(
model_name, OUTPUT0, expected0_list[b],
result_val[b]))
elif result_name == OUTPUT1:
tester.assertTrue(
np.array_equal(result_val[b], expected1_list[b]),
"{}, {} expected: {}, got {}".format(
model_name, OUTPUT1, expected1_list[b],
result_val[b]))
else:
tester.assertTrue(
False,
"unexpected raw result {}".format(result_name))
else:
# num_classes values must be returned and must
# match expected top values
class_list = result_val[b]
tester.assertEqual(len(class_list), num_classes)
expected0_flatten = expected0_list[b].flatten()
expected1_flatten = expected1_list[b].flatten()
for idx, ctuple in enumerate(class_list):
if result_name == OUTPUT0:
# can't compare indices since could have
# different indices with the same
# value/prob, so compare that the value of
# each index equals the expected
# value. Can only compare labels when the
# indices are equal.
tester.assertEqual(ctuple[1],
expected0_flatten[ctuple[0]])
tester.assertEqual(
ctuple[1],
expected0_flatten[expected0_sort_idx[b][idx]])
if ctuple[0] == expected0_sort_idx[b][idx]:
tester.assertEqual(
ctuple[2], 'label{}'.format(
expected0_sort_idx[b][idx]))
elif result_name == OUTPUT1:
tester.assertEqual(ctuple[1],
expected1_flatten[ctuple[0]])
tester.assertEqual(
ctuple[1],
expected1_flatten[expected1_sort_idx[b][idx]])
else:
tester.assertTrue(
False, "unexpected class result {}".format(
result_name))
if config[3]:
shared_memory_ctx.unregister(shm_ip0_handle)
shm.destroy_shared_memory_region(shm_ip0_handle)
shared_memory_ctx.unregister(shm_ip1_handle)
shm.destroy_shared_memory_region(shm_ip1_handle)
if "OUTPUT0" in outputs:
shared_memory_ctx.unregister(shm_op0_handle)
shm.destroy_shared_memory_region(shm_op0_handle)
if "OUTPUT1" in outputs:
shared_memory_ctx.unregister(shm_op1_handle)
shm.destroy_shared_memory_region(shm_op1_handle)
return results
def infer_zero(tester,
pf,
batch_size,
tensor_dtype,
input_shapes,
output_shapes,
model_version=None,
use_http=True,
use_grpc=True,
use_streaming=True):
tester.assertTrue(use_http or use_grpc or use_streaming)
configs = []
if use_http:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8000", ProtocolType.HTTP, False, True))
else:
configs.append(("localhost:8000", ProtocolType.HTTP, False, False))
if use_grpc:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8001", ProtocolType.GRPC, False, True))
else:
configs.append(("localhost:8001", ProtocolType.GRPC, False, False))
if use_streaming:
if TEST_SHARED_MEMORY:
configs.append(("localhost:8001", ProtocolType.GRPC, True, True))
else:
configs.append(("localhost:8001", ProtocolType.GRPC, True, False))
tester.assertEqual(len(input_shapes), len(output_shapes))
io_cnt = len(input_shapes)
for config in configs:
model_name = tu.get_zero_model_name(pf, io_cnt, tensor_dtype)
input_dict = {}
output_dict = {}
expected_dict = {}
if config[3]:
# create and register shared memory region for inputs and outputs
shm_ip_handles = list()
shm_op_handles = list()
shared_memory_ctx = SharedMemoryControlContext(config[0],
config[1],
verbose=True)
for io_num in range(io_cnt):
input0_byte_size = tu.shape_element_count(input_shapes[io_num]) *\
np.dtype(tensor_dtype).itemsize * batch_size
output0_byte_size = tu.shape_element_count(output_shapes[io_num]) *\
np.dtype(tensor_dtype).itemsize * batch_size
shm_ip_handles.append(shm.create_shared_memory_region("input"+str(io_num)+"_data",\
"/input"+str(io_num), input0_byte_size))
shm_op_handles.append(shm.create_shared_memory_region("output"+str(io_num)+"_data",\
"/output"+str(io_num), output0_byte_size))
shm.register(shm_ip_handles[io_num])
shm.register(shm_op_handles[io_num])
offset_input = 0
offset_output = 0
for io_num in range(io_cnt):
if pf == "libtorch" or pf == "libtorch_nobatch":
input_name = "INPUT__{}".format(io_num)
output_name = "OUTPUT__{}".format(io_num)
else:
input_name = "INPUT{}".format(io_num)
output_name = "OUTPUT{}".format(io_num)
input_list = list()
expected_list = list()
for b in range(batch_size):
rtensor_dtype = _range_repr_dtype(tensor_dtype)
in0 = np.random.randint(low=np.iinfo(rtensor_dtype).min,
high=np.iinfo(rtensor_dtype).max,
size=input_shapes[io_num],
dtype=rtensor_dtype)
if tensor_dtype != np.object:
in0 = in0.astype(tensor_dtype)
expected0 = np.ndarray.copy(in0)
else:
expected0 = np.array([
unicode(str(x), encoding='utf-8')
for x in in0.flatten()
],
dtype=object)
in0 = np.array([str(x) for x in in0.flatten()],
dtype=object).reshape(in0.shape)
expected0 = expected0.reshape(output_shapes[io_num])
input_list.append(in0)
expected_list.append(expected0)
expected_dict[output_name] = expected_list
if config[3]:
# copy data into shared memory region for input values
shm.set_shared_memory_region(shm_ip_handles[io_num],
input_list)
input_dict[input_name] = shm_ip_handles[io_num]
output_dict[output_name] = (InferContext.ResultFormat.RAW,
shm_op_handles[io_num])
else:
input_dict[input_name] = input_list
output_dict[output_name] = InferContext.ResultFormat.RAW
ctx = InferContext(config[0],
config[1],
model_name,
model_version,
correlation_id=0,
streaming=config[2],
verbose=True)
results = ctx.run(input_dict, output_dict, batch_size)
tester.assertEqual(ctx.get_last_request_model_name(), model_name)
if model_version is not None:
tester.assertEqual(ctx.get_last_request_model_version(),
model_version)
tester.assertEqual(len(results), io_cnt)
for (result_name, result_val) in iteritems(results):
tester.assertTrue(result_name in output_dict)
tester.assertTrue(result_name in expected_dict)
for b in range(batch_size):
expected = expected_dict[result_name][b]
tester.assertEqual(result_val[b].shape, expected.shape)
tester.assertTrue(
np.array_equal(result_val[b], expected),
"{}, {}, slot {}, expected: {}, got {}".format(
model_name, result_name, b, expected, result_val[b]))
if config[3]:
for io_num in range(io_cnt):
shared_memory_ctx.unregister(shm_ip_handles[io_num])
shm.destroy_shared_memory_region(shm_ip_handles[io_num])
shared_memory_ctx.unregister(shm_op_handles[io_num])
shm.destroy_shared_memory_region(shm_op_handles[io_num])
return results
def infer_shape_tensor(tester,
pf,
batch_size,
tensor_dtype,
input_shape_values,
dummy_input_shapes,
model_version=None,
use_http=True,
use_grpc=True,
use_streaming=True,
shm_suffix="",
use_system_shared_memory=False,
use_cuda_shared_memory=False,
priority=0,
timeout_us=0):
tester.assertTrue(use_http or use_grpc or use_streaming)
configs = []
if use_http:
configs.append(("localhost:8000", ProtocolType.HTTP, False))
if use_grpc:
configs.append(("localhost:8001", ProtocolType.GRPC, False))
if use_streaming:
configs.append(("localhost:8001", ProtocolType.GRPC, True))
tester.assertEqual(len(input_shape_values), len(dummy_input_shapes))
io_cnt = len(input_shape_values)
if use_system_shared_memory and use_cuda_shared_memory:
raise ValueError(
"Cannot set both System and CUDA shared memory flags to 1")
input_dict = {}
output_dict = {}
expected_dict = {}
shm_ip_handles = list()
shm_op_handles = list()
shared_memory_ctx = SharedMemoryControlContext("localhost:8000",
ProtocolType.HTTP,
verbose=False)
for io_num in range(io_cnt):
tester.assertTrue(pf == "plan" or pf == "plan_nobatch")
input_name = "INPUT{}".format(io_num)
output_name = "OUTPUT{}".format(io_num)
dummy_input_name = "DUMMY_INPUT{}".format(io_num)
dummy_output_name = "DUMMY_OUTPUT{}".format(io_num)
input_list = list()
dummy_input_list = list()
expected_list = list()
for b in range(batch_size):
# Prepare the dummy tensor
rtensor_dtype = _range_repr_dtype(tensor_dtype)
if (rtensor_dtype != np.bool):
dummy_in0 = np.random.randint(low=np.iinfo(rtensor_dtype).min,
high=np.iinfo(rtensor_dtype).max,
size=dummy_input_shapes[io_num],
dtype=rtensor_dtype)
else:
dummy_in0 = np.random.choice(a=[False, True],
size=dummy_input_shapes[io_num])
if tensor_dtype != np.object:
dummy_in0 = dummy_in0.astype(tensor_dtype)
else:
dummy_in0 = np.array([str(x) for x in in0.flatten()],
dtype=object).reshape(in0.shape)
dummy_input_list.append(dummy_in0)
# Prepare shape input tensor. Only one tensor per batch
in0 = np.asarray(input_shape_values[io_num], dtype=np.int32)
input_list.append(in0)
# Prepare the expected list for the output
expected0 = np.ndarray.copy(in0)
expected_list.append(expected0)
expected_dict[output_name] = expected_list
input_byte_size = len(in0) * np.dtype(tensor_dtype).itemsize
output_byte_size = input_byte_size * batch_size
dummy_input_byte_size = tu.shape_element_count(dummy_input_shapes[io_num]) *\
np.dtype(tensor_dtype).itemsize * batch_size
# The dimension of this tensor will be the value of the shape tensor
dummy_output_byte_size = tu.shape_element_count(in0) *\
np.dtype(tensor_dtype).itemsize * batch_size
# create and register shared memory region for inputs and outputs
if use_cuda_shared_memory:
shm_ip_handles.append(
cudashm.create_shared_memory_region(
"input" + str(io_num) + "_data" + shm_suffix,
input_byte_size, 0))
shm_ip_handles.append(
cudashm.create_shared_memory_region(
"dummy_input" + str(io_num) + "_data" + shm_suffix,
dummy_input_byte_size, 0))
shm_op_handles.append(
cudashm.create_shared_memory_region(
"output" + str(io_num) + "_data" + shm_suffix,
output_byte_size, 0))
shm_op_handles.append(
cudashm.create_shared_memory_region(
"dummy_output" + str(io_num) + "_data" + shm_suffix,
dummy_output_byte_size, 0))
shared_memory_ctx.cuda_register(shm_ip_handles[2 * io_num])
shared_memory_ctx.cuda_register(shm_ip_handles[2 * io_num + 1])
shared_memory_ctx.cuda_register(shm_op_handles[2 * io_num])
shared_memory_ctx.cuda_register(shm_op_handles[2 * io_num + 1])
# copy data into shared memory region for input values
cudashm.set_shared_memory_region(shm_ip_handles[2 * io_num],
input_list)
cudashm.set_shared_memory_region(shm_ip_handles[2 * io_num + 1],
dummy_input_list)
elif use_system_shared_memory:
shm_ip_handles.append(shm.create_shared_memory_region("input"+str(io_num)+"_data"+shm_suffix,\
"/input"+str(io_num)+shm_suffix, input_byte_size))
shm_ip_handles.append(shm.create_shared_memory_region("dumy_input"+str(io_num)+"_data"+shm_suffix,\
"/dummy_input"+str(io_num)+shm_suffix, dummy_input_byte_size))
shm_op_handles.append(shm.create_shared_memory_region("output"+str(io_num)+"_data"+shm_suffix,\
"/output"+str(io_num)+shm_suffix, output_byte_size))
shm_op_handles.append(shm.create_shared_memory_region("dummy_output"+str(io_num)+"_data"+shm_suffix,\
"/dummy_output"+str(io_num)+shm_suffix, dummy_output_byte_size))
shared_memory_ctx.register(shm_ip_handles[2 * io_num])
shared_memory_ctx.register(shm_ip_handles[2 * io_num + 1])
shared_memory_ctx.register(shm_op_handles[2 * io_num])
shared_memory_ctx.register(shm_op_handles[2 * io_num + 1])
# copy data into shared memory region for input values
shm.set_shared_memory_region(shm_ip_handles[2 * io_num],
input_list)
shm.set_shared_memory_region(shm_ip_handles[2 * io_num + 1],
dummy_input_list)
if use_system_shared_memory or use_cuda_shared_memory:
input_dict[input_name] = (shm_ip_handles[2 * io_num],
[len(input_shape_values[0])])
input_dict[dummy_input_name] = (shm_ip_handles[2 * io_num + 1],
dummy_input_shapes[io_num])
output_dict[output_name] = (InferContext.ResultFormat.RAW,
shm_op_handles[2 * io_num])
output_dict[dummy_output_name] = (InferContext.ResultFormat.RAW,
shm_op_handles[2 * io_num + 1])
else:
input_dict[input_name] = input_list
input_dict[dummy_input_name] = dummy_input_list
output_dict[output_name] = InferContext.ResultFormat.RAW
output_dict[dummy_output_name] = InferContext.ResultFormat.RAW
# Run inference and check results for each config
for config in configs:
model_name = tu.get_zero_model_name(pf, io_cnt, tensor_dtype)
ctx = InferContext(config[0],
config[1],
model_name,
model_version,
correlation_id=0,
streaming=config[2],
verbose=True)
results = ctx.run(input_dict,
output_dict,
batch_size,
priority=priority,
timeout_us=timeout_us)
tester.assertEqual(ctx.get_last_request_model_name(), model_name)
if model_version is not None:
tester.assertEqual(ctx.get_last_request_model_version(),
model_version)
tester.assertEqual(len(results), 2 * io_cnt)
for (result_name, result_val) in iteritems(results):
tester.assertTrue(result_name in output_dict)
expected = expected_dict[output_name][0]
for b in range(batch_size):
if result_name == output_name:
tester.assertEqual(result_val[b].shape, expected.shape)
tester.assertTrue(
np.array_equal(result_val[b], expected),
"{}, {}, slot {}, expected: {}, got {}".format(
model_name, result_name, b, expected,
result_val[b]))
elif result_name == dummy_output_name:
# The shape of the dummy output should be equal to the shape values
# specified in the shape tensor
tester.assertTrue(
np.array_equal(result_val[b].shape, expected),
"{}, {}, slot {}, expected: {}, got {}".format(
model_name, result_name, b, expected,
result_val[b]))
if use_cuda_shared_memory or use_system_shared_memory:
for io_num in range(2 * io_cnt):
shared_memory_ctx.unregister(shm_ip_handles[io_num])
shared_memory_ctx.unregister(shm_op_handles[io_num])
if use_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_ip_handles[io_num])
cudashm.destroy_shared_memory_region(shm_op_handles[io_num])
else:
shm.destroy_shared_memory_region(shm_ip_handles[io_num])
shm.destroy_shared_memory_region(shm_op_handles[io_num])
return results
def check_sequence(self,
trial,
model_name,
input_dtype,
correlation_id,
sequence_thresholds,
values,
expected_result,
protocol,
batch_size=1,
sequence_name="<unknown>",
tensor_shape=(1, )):
"""Perform sequence of inferences. The 'values' holds a list of
tuples, one for each inference with format:
(flag_str, value, (ls_ms, gt_ms), (pre_delay_ms, post_delay_ms)
"""
if (("savedmodel" not in trial) and ("graphdef" not in trial)
and ("netdef" not in trial) and ("custom" not in trial)
and ("onnx" not in trial) and ("libtorch" not in trial)
and ("plan" not in trial)):
self.assertFalse(True, "unknown trial type: " + trial)
# Can only send the request exactly once since it is a
# sequence model with state, so can have only a single config.
configs = []
if protocol == "http":
configs.append(("localhost:8000", ProtocolType.HTTP, False))
if protocol == "grpc":
configs.append(("localhost:8001", ProtocolType.GRPC, False))
if protocol == "streaming":
configs.append(("localhost:8001", ProtocolType.GRPC, True))
self.assertFalse(
_test_system_shared_memory and _test_cuda_shared_memory,
"Cannot set both System and CUDA shared memory flags to 1")
self.assertEqual(len(configs), 1)
# create and register shared memory output region in advance
if _test_system_shared_memory or _test_cuda_shared_memory:
shared_memory_ctx = SharedMemoryControlContext("localhost:8000",
ProtocolType.HTTP,
verbose=True)
output_byte_size = 512
if _test_system_shared_memory:
shm_op_handle = shm.create_shared_memory_region(
"output_data", "/output", output_byte_size)
shared_memory_ctx.unregister(shm_op_handle)
shared_memory_ctx.register(shm_op_handle)
elif _test_cuda_shared_memory:
shm_op_handle = cudashm.create_shared_memory_region(
"output_data", output_byte_size, 0)
shared_memory_ctx.unregister(shm_op_handle)
shared_memory_ctx.cuda_register(shm_op_handle)
for config in configs:
ctx = InferContext(config[0],
config[1],
model_name,
correlation_id=correlation_id,
streaming=config[2],
verbose=True)
# Execute the sequence of inference...
try:
seq_start_ms = int(round(time.time() * 1000))
for flag_str, value, thresholds, delay_ms in values:
if delay_ms is not None:
time.sleep(delay_ms[0] / 1000.0)
flags = InferRequestHeader.FLAG_NONE
if flag_str is not None:
if "start" in flag_str:
flags = flags | InferRequestHeader.FLAG_SEQUENCE_START
if "end" in flag_str:
flags = flags | InferRequestHeader.FLAG_SEQUENCE_END
input_list = list()
for b in range(batch_size):
if input_dtype == np.object:
in0 = np.full(tensor_shape, value, dtype=np.int32)
in0n = np.array(
[str(x) for x in in0.reshape(in0.size)],
dtype=object)
in0 = in0n.reshape(tensor_shape)
else:
in0 = np.full(tensor_shape,
value,
dtype=input_dtype)
input_list.append(in0)
# create input shared memory and copy input data values into it
if _test_system_shared_memory or _test_cuda_shared_memory:
input_list_tmp = iu._prepend_string_size(
input_list) if (input_dtype
== np.object) else input_list
input_byte_size = sum(
[i0.nbytes for i0 in input_list_tmp])
if _test_system_shared_memory:
shm_ip_handle = shm.create_shared_memory_region(
"input_data", "/input", input_byte_size)
shm.set_shared_memory_region(
shm_ip_handle, input_list_tmp)
shared_memory_ctx.unregister(shm_ip_handle)
shared_memory_ctx.register(shm_ip_handle)
elif _test_cuda_shared_memory:
shm_ip_handle = cudashm.create_shared_memory_region(
"input_data", input_byte_size, 0)
cudashm.set_shared_memory_region(
shm_ip_handle, input_list_tmp)
shared_memory_ctx.unregister(shm_ip_handle)
shared_memory_ctx.cuda_register(shm_ip_handle)
input_info = (shm_ip_handle, tensor_shape)
output_info = (InferContext.ResultFormat.RAW,
shm_op_handle)
else:
input_info = input_list
output_info = InferContext.ResultFormat.RAW
start_ms = int(round(time.time() * 1000))
INPUT = "INPUT__0" if trial.startswith(
"libtorch") else "INPUT"
OUTPUT = "OUTPUT__0" if trial.startswith(
"libtorch") else "OUTPUT"
results = ctx.run({INPUT: input_info},
{OUTPUT: output_info},
batch_size=batch_size,
flags=flags)
end_ms = int(round(time.time() * 1000))
self.assertEqual(len(results), 1)
self.assertTrue(OUTPUT in results)
result = results[OUTPUT][0][0]
print("{}: {}".format(sequence_name, result))
if thresholds is not None:
lt_ms = thresholds[0]
gt_ms = thresholds[1]
if lt_ms is not None:
self.assertTrue(
(end_ms - start_ms) < lt_ms,
"expected less than " + str(lt_ms) +
"ms response time, got " +
str(end_ms - start_ms) + " ms")
if gt_ms is not None:
self.assertTrue(
(end_ms - start_ms) > gt_ms,
"expected greater than " + str(gt_ms) +
"ms response time, got " +
str(end_ms - start_ms) + " ms")
if delay_ms is not None:
time.sleep(delay_ms[1] / 1000.0)
seq_end_ms = int(round(time.time() * 1000))
if input_dtype == np.object:
self.assertEqual(int(result), expected_result)
else:
self.assertEqual(result, expected_result)
if sequence_thresholds is not None:
lt_ms = sequence_thresholds[0]
gt_ms = sequence_thresholds[1]
if lt_ms is not None:
self.assertTrue((seq_end_ms - seq_start_ms) < lt_ms,
"sequence expected less than " +
str(lt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
if gt_ms is not None:
self.assertTrue((seq_end_ms - seq_start_ms) > gt_ms,
"sequence expected greater than " +
str(gt_ms) + "ms response time, got " +
str(seq_end_ms - seq_start_ms) + " ms")
except Exception as ex:
self.add_deferred_exception(ex)
if _test_system_shared_memory or _test_cuda_shared_memory:
shared_memory_ctx.unregister(shm_op_handle)
if _test_system_shared_memory:
shm.destroy_shared_memory_region(shm_op_handle)
elif _test_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_op_handle)
def _cleanup_server(self, shm_handles):
for shm_handle in shm_handles:
shm.destroy_shared_memory_region(shm_handle)
def destroy_either_shm_region(shm_handle, use_system_shared_memory, use_cuda_shared_memory):
if use_cuda_shared_memory:
cudashm.destroy_shared_memory_region(shm_handle)
else:
shm.destroy_shared_memory_region(shm_handle)
