enable runtime cache for ascend dynamic scene

.gitignore

@@ -100,7 +100,7 @@ mindspore/python/mindspore/version.py
 mindspore/python/mindspore/default_config.py
 mindspore/python/mindspore/device_target.py
 mindspore/python/mindspore/package_name.py
-mindspore/.commit_id
+mindspore/python/mindspore/.commit_id
 
 # lite test file
 mindspore/lite/test/do_test/

mindspore/ccsrc/runtime/device/ascend/ascend_kernel_runtime.cc

@@ -113,6 +113,20 @@ void IntHandler(int, siginfo_t *, void *) {
   MS_LOG(WARNING) << "Process " << this_pid << " receive KeyboardInterrupt signal.";
   (void)kill(this_pid, SIGTERM);
 }
+
+void AscendEnableDynamicRuntimeCache(const KernelGraph *graph) {
+  const auto &node_list = graph->TopoSort(graph->get_return());
+  for (auto &node : node_list) {
+    auto kernel_info = node->kernel_info();
+    if (!kernel_info) {
+      continue;
+    }
+    MS_EXCEPTION_IF_NULL(kernel_info);
+    auto runtime_cache = kernel_info->runtime_cache();
+    MS_EXCEPTION_IF_NULL(runtime_cache);
+    runtime_cache->set_valid();
+  }
+}
 }  // namespace
 
 std::vector<rtExceptionInfo> AscendKernelRuntime::task_fail_infoes_ = {};
@@ -908,6 +922,7 @@ bool AscendKernelRuntime::RunDynamicKernelAsync(const session::KernelGraph &grap
     MS_LOG(ERROR) << "GraphId:" << graph.graph_id() << " Not Found! Please generator executor first";
     return false;
   }
+  AscendEnableDynamicRuntimeCache(&graph);
 
   auto dynamic_kernels = iter->second;
   for (const auto &dynamic_kernel : dynamic_kernels) {

tests/st/ops/ascend/test_dynamic_ops.py

@@ -0,0 +1,299 @@
+# Copyright 2021 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+import numpy as np
+import pytest
+
+from mindspore import ops, nn, ParameterTuple, context, set_seed
+from mindspore.train import DatasetHelper, connect_network_with_dataset
+import mindspore.dataset as ds
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+set_seed(2)
+
+
+def _exec_preprocess(network, is_train, dataset, dataset_sink_mode, epoch_num, sink_size):
+    if dataset_sink_mode and not is_train:
+        dataset.__loop_size__ = 1
+
+    dataset_helper = DatasetHelper(dataset, dataset_sink_mode, sink_size, epoch_num)
+
+    if dataset_sink_mode:
+        network = connect_network_with_dataset(network, dataset_helper)
+
+    return dataset_helper, network
+
+
+def dynamic_shape_sink_process(network, dataset, is_train=True):
+    # epoch_num=1 sink_size=1: exec one step
+    dataset_sink_mode = True
+    sink_size = 1
+    epoch_num = 1
+    dataset_helper, network = _exec_preprocess(network, is_train, dataset, dataset_sink_mode, epoch_num, sink_size)
+    network.set_train(is_train)
+    for inputs in dataset_helper:
+        outputs = network(*inputs)
+        return outputs
+
+
+def fixed_shape_process(network, dataset, is_train=True):
+    network.set_train(is_train)
+    for inputs in dataset.create_tuple_iterator():
+        outputs = network(*inputs)
+        return outputs
+
+
+def dataset_generator(data_list):
+    for data in data_list:
+        yield data
+
+
+class GradNetWrtX(nn.Cell):
+    def __init__(self, net):
+        super(GradNetWrtX, self).__init__()
+        self.net = net
+        self.grad_op = ops.GradOperation(get_all=True, get_by_list=True, sens_param=True)
+        self.params = ParameterTuple(net.trainable_params())
+
+    def construct(self, *inputs):
+        gradient_function = self.grad_op(self.net, self.params)
+        return gradient_function(*inputs)
+
+
+class LayerNormNet(nn.Cell):
+    def __init__(self, last_dim):
+        super(LayerNormNet, self).__init__()
+        self.layernorm = nn.LayerNorm([last_dim])
+
+    def construct(self, x):
+        return self.layernorm(x)
+
+
+class Conv2dNet(nn.Cell):
+    def __init__(self):
+        super(Conv2dNet, self).__init__()
+        self.conv = nn.Conv2d(3, 10, 4, pad_mode="valid", has_bias=False, weight_init='normal')
+
+    def construct(self, x):
+        return self.conv(x)
+
+
+class DropoutNet(nn.Cell):
+    def __init__(self):
+        super(DropoutNet, self).__init__()
+        self.drop = nn.Dropout(0.5)
+        self.relu = ops.ReLU()
+
+    def construct(self, x):
+        x = self.relu(x)
+        return self.relu(self.drop(x))
+
+
+class ReduceSumNet(nn.Cell):
+    def __init__(self, axis=()):
+        super(ReduceSumNet, self).__init__()
+        self.reduce = ops.ReduceSum()
+        self.axis = axis
+
+    def construct(self, x):
+        return self.reduce(x, self.axis)
+
+
+class AddNet(nn.Cell):
+    def construct(self, x, y):
+        return ops.add(x, y)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_layernorm():
+    """
+    Feature: Test LayerNorm and its backward. The input shape is dynamic.
+    Description: The second dim of input is unknown.
+    Expectation: Assert that results are consistent with fixed shape.
+    """
+    last_dim = 32
+    batch_size = 16
+    data_list = []
+    for i in range(20, 23):
+        data_list.append((np.random.rand(batch_size, i, last_dim).astype(np.float32),
+                          np.random.rand(batch_size, i, last_dim).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, None, last_dim], "data2": [batch_size, None, last_dim]})
+
+    net = GradNetWrtX(LayerNormNet(last_dim))
+
+    gradients = dynamic_shape_sink_process(net, dataset)
+    gradients_cmp = fixed_shape_process(net, dataset)
+    assert np.allclose(gradients[0][0].asnumpy(), gradients_cmp[0][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+    assert np.allclose(gradients[1][0].asnumpy(), gradients_cmp[1][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+    assert np.allclose(gradients[1][1].asnumpy(), gradients_cmp[1][1].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_conv2d():
+    """
+    Feature: Test Conv2d and its backward. The input shape is dynamic.
+    Description: Input dim of `H `or `W` is unknown. Conv2d's attr[pad] set to "valid".
+    Expectation: Assert that results are consistent with fixed shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(220, 224):
+        data_list.append((np.random.rand(batch_size, 3, i, 112).astype(np.float32),
+                          np.random.rand(batch_size, 10, 219, 109).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, 3, None, 112], "data2": [batch_size, 10, None, 109]})
+    net = GradNetWrtX(Conv2dNet())
+    gradients = dynamic_shape_sink_process(net, dataset)
+    gradients_cmp = fixed_shape_process(net, dataset)
+    assert np.allclose(gradients[0][0].asnumpy(), gradients_cmp[0][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+    assert np.allclose(gradients[1][0].asnumpy(), gradients_cmp[1][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_dropout():
+    """
+    Feature: Test Dropout and its backward.
+    Description: The input shape is dynamic.
+    Expectation: Dropout result is random, assert gradient shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(48, 50):
+        data_list.append((np.random.rand(batch_size, i, 256).astype(np.float32),
+                          np.random.rand(batch_size, i, 256).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, None, 256], "data2": [batch_size, None, 256]})
+    net = GradNetWrtX(DropoutNet())
+    net.set_train()
+    gradients = dynamic_shape_sink_process(net, dataset)
+    assert gradients[0][0].shape == (batch_size, 49, 256)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_reducesum1():
+    """
+    Feature: Test ReduceSum and its backward. The input shape is dynamic.
+    Description: axis=(), result of reduce sum is a scalar, gradient shape is the same as input, value is all one.
+    Expectation: Assert gradient shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(48, 50):
+        data_list.append((np.random.rand(batch_size, i, i + 2).astype(np.float32),
+                          np.array(1).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, None, None], "data2": []})
+    net = GradNetWrtX(ReduceSumNet())
+
+    gradients = dynamic_shape_sink_process(net, dataset)
+    assert gradients[0][0].shape == (batch_size, 49, 51)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_reducesum2():
+    """
+    Feature: Test ReduceSum and its backward. The input shape is dynamic.
+    Description: axis is a scalar, not tuple.
+    Expectation: Assert that results are consistent with fixed shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(48, 50):
+        data_list.append((np.random.rand(batch_size, i, i + 2).astype(np.float32),
+                          np.random.rand(batch_size, i + 2).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, None, None], "data2": [batch_size, None]})
+    net = GradNetWrtX(ReduceSumNet(1))
+
+    gradients = dynamic_shape_sink_process(net, dataset)
+    gradients_cmp = fixed_shape_process(net, dataset)
+    assert np.allclose(gradients[0][0].asnumpy(), gradients_cmp[0][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_add1():
+    """
+    Feature: Test Add and its backward. The input shape is dynamic.
+    Description: Second input is a scalar. Shape of forward result is the same as first input.
+    Expectation: Assert that results are consistent with fixed shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(48, 50):
+        data_list.append((np.random.rand(batch_size, i).astype(np.float32),
+                          np.array(1).astype(np.float32),
+                          np.random.rand(batch_size, i).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2", "data3"])
+    dataset.set_dynamic_columns(columns={"data1": [batch_size, None], "data2": [], "data3": [batch_size, None]})
+    net = GradNetWrtX(AddNet())
+
+    gradients = dynamic_shape_sink_process(net, dataset)
+    gradients_cmp = fixed_shape_process(net, dataset)
+    assert np.allclose(gradients[0][0].asnumpy(), gradients_cmp[0][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+    assert np.allclose(gradients[0][1].asnumpy(), gradients_cmp[0][1].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_add2():
+    """
+    Feature: Test Add and its backward. The input shape is dynamic.
+    Description: Shape of forward result is the same as first input. The axis of reduce_sum in add's bprop will be a
+                 empty Tensor.
+    Expectation: Assert that results are consistent with fixed shape.
+    """
+    batch_size = 16
+    data_list = []
+    for i in range(48, 50):
+        data_list.append((np.random.rand(batch_size, 2, i).astype(np.float32),
+                          np.random.rand(2, i).astype(np.float32),
+                          np.random.rand(batch_size, 2, i).astype(np.float32)))
+
+    dataset = ds.GeneratorDataset(data_list, ["data1", "data2", "data3"])
+    dataset.set_dynamic_columns(columns=
+                                {"data1": [batch_size, 2, None], "data2": [2, None], "data3": [batch_size, 2, None]})
+    net = GradNetWrtX(AddNet())
+
+    gradients = dynamic_shape_sink_process(net, dataset)
+    gradients_cmp = fixed_shape_process(net, dataset)
+    assert np.allclose(gradients[0][0].asnumpy(), gradients_cmp[0][0].asnumpy(), rtol=1.0e-4, atol=1.0e-4)
+    assert np.allclose(gradients[0][1].asnumpy(), gradients_cmp[0][1].asnumpy(), rtol=1.0e-4, atol=1.0e-4)