You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/10/09 20:56:07 UTC
[GitHub] [incubator-tvm] d-smirnov commented on a change in pull request #6532: [BYOC][ACL] Support add operation
d-smirnov commented on a change in pull request #6532:
URL: https://github.com/apache/incubator-tvm/pull/6532#discussion_r502667277
##########
File path: tests/python/contrib/test_arm_compute_lib/test_add.py
##########
@@ -0,0 +1,135 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you 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.
+"""Arm Compute Library integration reshape tests."""
+
+import numpy as np
+
+import tvm
+import tvm.testing
+from tvm import relay
+
+from test_arm_compute_lib.infrastructure import (
+ skip_runtime_test,
+ skip_codegen_test,
+ build_and_run,
+ verify,
+ verify_codegen,
+)
+from test_arm_compute_lib.infrastructure import Device
+
+_qnn_params = {
+ "lhs_scale": relay.const(0.0156863, "float32"),
+ "lhs_zero_point": relay.const(127, "int32"),
+ "rhs_scale": relay.const(0.0117647, "float32"),
+ "rhs_zero_point": relay.const(85, "int32"),
+ "output_scale": relay.const(0.0235294, "float32"),
+ "output_zero_point": relay.const(128, "int32"),
+}
+
+
+def _get_model(shape, dtype, var_names, op, op_params):
+ a = relay.var(next(var_names), shape=shape, dtype=dtype)
+ b = relay.var(next(var_names), shape=shape, dtype=dtype)
+ return op(a, b, **op_params)
+
+
+def _get_expected_codegen(shape, dtype, op_name, qnn_params):
+ input_a = {"op": "input", "name": "", "attrs": {"shape": [[list(shape)]], "dtype": [[dtype]]}}
+ input_b = {"op": "input", "name": "", "attrs": {"shape": [[list(shape)]], "dtype": [[dtype]]}}
+ input_qnn = [
+ {
+ "op": "const",
+ "name": "",
+ "attrs": {
+ "shape": [[list(qnn_params[_].data.shape)]],
+ "dtype": [[qnn_params[_].data.dtype]],
+ },
+ }
+ for _ in qnn_params
+ ]
+ inputs = [input_a, input_b, *input_qnn]
+ node = {
+ "op": "kernel",
+ "name": op_name,
+ "inputs": [[_, 0, 0] for _ in range(len(inputs))],
+ "attrs": {
+ "num_inputs": str(len(inputs)),
+ "num_outputs": "1",
+ "shape": [[list(shape)]],
+ "dtype": [[dtype]],
+ },
+ }
+
+ return [*inputs, node]
+
+
+def test_runtime_add():
+ Device.load("test_config.json")
+
+ if skip_runtime_test():
+ return
+
+ device = Device()
+ np.random.seed(0)
+
+ for dtype, low, high, atol, rtol, op, op_params in [
+ ("float32", -127, 128, 1e-7, 1e-7, relay.add, {}),
+ ("uint8", 0, 255, 0.0, 1.0, relay.qnn.op.add, _qnn_params),
+ ]:
+ shape = (2, 2)
+ for inputs in [
+ {
+ "a": tvm.nd.array(np.random.uniform(low, high, shape).astype(dtype)),
+ "b": tvm.nd.array(np.random.uniform(low, high, shape).astype(dtype)),
+ }
+ ]:
+ outputs = []
+ func = _get_model(shape, dtype, iter(inputs), op, op_params)
+ for acl in [True, False]:
+ outputs.append(build_and_run(func, inputs, 1, None, device, enable_acl=acl)[0])
+
+ config = {
+ "shape": shape,
+ "dtype": dtype,
+ "inputs": inputs,
+ "operation": op,
+ "op_params": op_params,
+ }
+
+ # verify_saturation=False as the result of add_QASYMM8_QASYMM8_QASYMM8
+ # is always saturated currently.
Review comment:
ACL's _add_QASYMM8_QASYMM8_QASYMM8_ implemented in a way that it always saturates result. It just cannot be verified as uint8 always cast to QASYMM8. As for the parameter combination which does not cause saturation for qnn.add could you elaborate why this is needed?
##########
File path: src/runtime/contrib/arm_compute_lib/acl_runtime.cc
##########
@@ -417,6 +420,45 @@ class ACLRuntime : public JSONRuntimeBase {
function->configure(&layer->inputs[0], &layer->inputs[1], &layer->outputs[0]);
layer->function = function;
}
+ /*!
+ * \brief Creates an add/qnn.add layer
+ *
+ * \param layer The ACL layer to build. Containing inputs, outputs and the ACL function.
+ * \param node The JSON representation of the operator.
+ */
+ void CreateAddLayer(CachedLayer* layer, const JSONGraphNode& node) {
+ auto op_name = node.GetOpName();
+ if ("add" == op_name) {
+ layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[0]));
+ layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[1]));
+ layer->outputs.push_back(MakeACLTensorFromJSONNode(node));
+ } else if ("qnn.add" == op_name) {
+ layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[0], &node.GetInputs()[2],
+ &node.GetInputs()[3]));
+ layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[1], &node.GetInputs()[4],
+ &node.GetInputs()[5]));
+ layer->outputs.push_back(
+ MakeACLTensorFromJSONNode(node, &node.GetInputs()[6], &node.GetInputs()[7]));
+ } else {
+ throw std::runtime_error("Unsupported form of add op: " + op_name);
+ }
+
+ /** Initialise the kernel's inputs, output and conversion policy.
Review comment:
removed
##########
File path: tests/python/contrib/test_arm_compute_lib/test_add.py
##########
@@ -0,0 +1,135 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you 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.
+"""Arm Compute Library integration reshape tests."""
+
+import numpy as np
+
+import tvm
+import tvm.testing
+from tvm import relay
+
+from test_arm_compute_lib.infrastructure import (
+ skip_runtime_test,
+ skip_codegen_test,
+ build_and_run,
+ verify,
+ verify_codegen,
+)
+from test_arm_compute_lib.infrastructure import Device
+
+_qnn_params = {
+ "lhs_scale": relay.const(0.0156863, "float32"),
+ "lhs_zero_point": relay.const(127, "int32"),
+ "rhs_scale": relay.const(0.0117647, "float32"),
+ "rhs_zero_point": relay.const(85, "int32"),
+ "output_scale": relay.const(0.0235294, "float32"),
+ "output_zero_point": relay.const(128, "int32"),
+}
+
+
+def _get_model(shape, dtype, var_names, op, op_params):
+ a = relay.var(next(var_names), shape=shape, dtype=dtype)
+ b = relay.var(next(var_names), shape=shape, dtype=dtype)
+ return op(a, b, **op_params)
+
+
+def _get_expected_codegen(shape, dtype, op_name, qnn_params):
+ input_a = {"op": "input", "name": "", "attrs": {"shape": [[list(shape)]], "dtype": [[dtype]]}}
+ input_b = {"op": "input", "name": "", "attrs": {"shape": [[list(shape)]], "dtype": [[dtype]]}}
+ input_qnn = [
+ {
+ "op": "const",
+ "name": "",
+ "attrs": {
+ "shape": [[list(qnn_params[_].data.shape)]],
+ "dtype": [[qnn_params[_].data.dtype]],
+ },
+ }
+ for _ in qnn_params
+ ]
+ inputs = [input_a, input_b, *input_qnn]
+ node = {
+ "op": "kernel",
+ "name": op_name,
+ "inputs": [[_, 0, 0] for _ in range(len(inputs))],
+ "attrs": {
+ "num_inputs": str(len(inputs)),
+ "num_outputs": "1",
+ "shape": [[list(shape)]],
+ "dtype": [[dtype]],
+ },
+ }
+
+ return [*inputs, node]
+
+
+def test_runtime_add():
+ Device.load("test_config.json")
+
+ if skip_runtime_test():
+ return
+
+ device = Device()
+ np.random.seed(0)
+
+ for dtype, low, high, atol, rtol, op, op_params in [
+ ("float32", -127, 128, 1e-7, 1e-7, relay.add, {}),
+ ("uint8", 0, 255, 0.0, 1.0, relay.qnn.op.add, _qnn_params),
+ ]:
+ shape = (2, 2)
+ for inputs in [
+ {
+ "a": tvm.nd.array(np.random.uniform(low, high, shape).astype(dtype)),
+ "b": tvm.nd.array(np.random.uniform(low, high, shape).astype(dtype)),
+ }
+ ]:
+ outputs = []
+ func = _get_model(shape, dtype, iter(inputs), op, op_params)
+ for acl in [True, False]:
+ outputs.append(build_and_run(func, inputs, 1, None, device, enable_acl=acl)[0])
+
+ config = {
+ "shape": shape,
+ "dtype": dtype,
+ "inputs": inputs,
+ "operation": op,
+ "op_params": op_params,
+ }
+
+ # verify_saturation=False as the result of add_QASYMM8_QASYMM8_QASYMM8
+ # is always saturated currently.
+ verify(outputs, atol=atol, rtol=rtol, config=config, verify_saturation=False)
+
+
+def test_runtime_codegen_add():
Review comment:
operations added to arm_compute_lib.rst
method renamed
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org