[Runtime][PipelineExecutor] Tutorial of using pipeline executor.

gallery/how_to/work_with_relay/using_with_pipeline_executor.py

@@ -0,0 +1,262 @@
+# 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.
+"""
+Using Pipeline Executor in Relay
+=================================
+**Author**: `Hua Jiang <https://https://github.com/huajsj>`_
+
+This is a short tutorial on how to use "Pipeline Executor" with Relay.
+"""
+import tvm
+from tvm import te
+import numpy as np
+from tvm.contrib import graph_executor as runtime
+from tvm.relay.op.contrib.cutlass import partition_for_cutlass
+from tvm import relay
+from tvm.relay import testing
+import tvm.testing
+import time
+from tvm.contrib.cutlass import (
+    has_cutlass,
+    num_cutlass_partitions,
+    finalize_modules,
+    finalize_modules_vm,
+)
+
+img_size = 8
+#######################################################################
+# Create a simple network, this network can be a pre-trained model too.
+# ---------------------------------------------------------------------
+# Let's create a very simple network for demonstration.
+# It consists of convolution, batch normalization, and ReLU activation.
+def get_network():
+    out_channels = 16
+    batch_size = 1
+    data = relay.var("data", relay.TensorType((batch_size, 3, img_size, img_size), "float16"))
+    dense_weight = relay.var(
+        "dweight", relay.TensorType((batch_size, 16 * img_size * img_size), "float16")
+    )
+    weight = relay.var("weight")
+    second_weight = relay.var("second_weight")
+    bn_gamma = relay.var("bn_gamma")
+    bn_beta = relay.var("bn_beta")
+    bn_mmean = relay.var("bn_mean")
+    bn_mvar = relay.var("bn_var")
+    simple_net = relay.nn.conv2d(
+        data=data, weight=weight, kernel_size=(3, 3), channels=out_channels, padding=(1, 1)
+    )
+    simple_net = relay.nn.batch_norm(simple_net, bn_gamma, bn_beta, bn_mmean, bn_mvar)[0]
+    simple_net = relay.nn.relu(simple_net)
+    simple_net = relay.nn.batch_flatten(simple_net)
+    simple_net = relay.nn.dense(simple_net, dense_weight)
+    simple_net = relay.Function(relay.analysis.free_vars(simple_net), simple_net)
+    data_shape = (batch_size, 3, img_size, img_size)
+    net, params = testing.create_workload(simple_net)
+    return net, params, data_shape
+
+
+net, params, data_shape = get_network()
+###########################################
+# Splitting the network into two subgraphs.
+# -----------------------------------------
+# We use an testing linear graph splitting function as a example. User also can create their
+# own splitting function logic.
+import inspect
+import os
+
+test_path = os.path.dirname(inspect.getfile(lambda: None))
+os.sys.path.append(os.path.join(test_path, "../../../tests/python/relay"))
+from test_pipeline_executor import graph_split
+
+###########################################
+# Splitting the network into two subgraphs.
+split_config = [{"op_name": "nn.relu", "op_index": 0}]
+subgraphs = graph_split(net["main"], split_config, params)
+###########################################################
+# The generated subgraphs should look something like below.
+
+"""
+#subgraphs[0])
+
+ def @main(%data: Tensor[(1, 3, img_size, img_size), float16]) {
+  %0 = nn.conv2d(%data, meta[relay.Constant][0] /* ty=Tensor[(16, 3, 3, 3), float16] */, padding=[1, 1, 1, 1], channels=16, kernel_size=[3, 3]) /* ty=Tensor[(1, 16, img_size, img_size), float16] */;
+  %1 = nn.batch_norm(%0, meta[relay.Constant][1] /* ty=Tensor[(16), float16] */, meta[relay.Constant][2] /* ty=Tensor[(16), float16]*/, meta[relay.Constant][3] /* ty=Tensor[(16), float16] */, meta[relay.Constant][4] /* ty=Tensor[(16), float16] */) /* ty=(Tensor[(1,16, img_size, img_size), float16], Tensor[(16), float16], Tensor[(16), float16]) */;
+  %2 = %1.0;
+  nn.relu(%2) /* ty=Tensor[(1, 16, img_size, img_size), float16] */
+ }
+
+peline-tutorial
+
+#subgraphs[1]
+
+ def @main(%data_n_0: Tensor[(1, 16, 8, 8), float16] /* ty=Tensor[(1, 16, 8, 8), float16] */) {
+  %0 = nn.batch_flatten(%data_n_0) /* ty=Tensor[(1, 1024), float16] */;
+  nn.dense(%0, meta[relay.Constant][0] /* ty=Tensor[(1, 1024), float16] */, units=None) /* ty=Tensor[(1, 1), float16] */
+ }
+
+"""
+
+# sphinx_gallery_start_ignore
+from tvm import testing
+
+testing.utils.install_request_hook(depth=3)
+# sphinx_gallery_end_ignore
+
+#########################################
+# Build the subgraph with cutlass target.
+# ---------------------------------------
+
+cutlass = tvm.target.Target(
+    {
+        "kind": "cutlass",
+        "sm": int(tvm.target.Target("cuda").arch.split("_")[1]),
+        "use_3xtf32": True,
+        "split_k_slices": [1],
+        "profile_all_alignments": False,
+        "find_first_valid": True,
+        "use_multiprocessing": True,
+        "use_fast_math": False,
+        "tmp_dir": "./tmp",
+    },
+    host=tvm.target.Target("llvm"),
+)
+
+
+def cutlass_build(mod, target, params=None, target_host=None, mod_name="default"):
+    target = [target, cutlass]
+    lib = relay.build_module.build(
+        mod, target=target, params=params, target_host=target_host, mod_name=mod_name
+    )
+    return lib
+
+
+###########################################################
+# Run the two subgraphs in pipeline with pipeline executor.
+# ---------------------------------------------------------
+# Define a function to do all the codegen and pipeline executor works.
+# To run pipeline executor with dnnl, USE_PIPELINE_EXECUTOR need to get set as ON.
+# and the 'USE_CUTLASS' should set as ON in config.cmake.
+from tvm.contrib import graph_executor, pipeline_executor, pipeline_executor_build
+
+#########################################
+# Create subgraph pipeline configuration.
+# Associate the subgraph module with a target.
+# Using BYOC to set the codegen of the second subgraph module.
+# To use cutlass the 'USE_CUTLASS' should set as ON.
+mod0, mod1 = subgraphs[0], subgraphs[1]
+# Use cutlass as the codegen.
+mod1 = partition_for_cutlass(mod1)
+#################################################
+# Get the pipeline executor configuration object.
+pipe_config = pipeline_executor_build.PipelineConfig()
+###########################################################################
+# Set the compile target of the second subgraph module for example as LLVM.
+pipe_config[mod0].target = "llvm"
+pipe_config[mod0].dev = tvm.cpu(0)
+###############################################################################
+# Set the cpu afinity for control flow, for example using cpu 0 for control flow.
+pipe_config[mod1].cpu_affinity = "0"
+pipe_config[mod1].export_cc = None
+##############################################################
+# Set the compile target of the second subgraph module as LLVM.
+pipe_config[mod1].target = "cuda"
+pipe_config[mod1].dev = tvm.device("cuda", 0)
+pipe_config[mod1].build_func = cutlass_build
+pipe_config[mod1].export_cc = "nvcc"
+#################################################################################
+# Set the cpu afinity for control flow, for example using cpu 1 for control flow.
+pipe_config[mod1].cpu_affinity = "1"
+pipe_config["input"]["data"].connect(pipe_config[mod0]["input"]["data"])
+pipe_config[mod0]["output"][0].connect(pipe_config[mod1]["input"]["data_n_0"])
+pipe_config[mod1]["output"]["0"].connect(pipe_config["output"][0])
+######################################
+# The pipeline configuration as below.
+"""
+print(pipe_config)
+ Inputs
+  |data: mod0:data
+
+ output
+  |output(0) : mod1.output(0)
+
+ connections
+  |mod0.output(0)-> mod1.data_n_0
+"""
+
+# sphinx_gallery_start_ignore
+from tvm import testing
+
+# testing.utils.install_request_hook(depth=3)
+# sphinx_gallery_end_ignore
+##############################
+# Build the pipeline executor.
+# ----------------------------
+with tvm.transform.PassContext(opt_level=3):
+    pipeline_mod_factory = pipeline_executor_build.build(pipe_config)
+###############################################
+# Export the parameter configuration to a file.
+directory_path = tvm.contrib.utils.tempdir().temp_dir
+#############################################
+# If the directory does not exist, create it.
+if not os.path.exists(directory_path):
+    os.makedirs(directory_path)
+config_file_name = pipeline_mod_factory.export_library(directory_path)
+################################################################
+# Use the load function to create and initialize PipelineModule.
+# --------------------------------------------------------------
+pipeline_module = pipeline_executor.PipelineModule.load_library(config_file_name)
+
+############################
+# Run the pipeline executor.
+# --------------------------
+# Allocated a input data.
+data = np.random.uniform(-1, 1, size=data_shape).astype("float16")
+pipeline_module.set_input("data", tvm.nd.array(data))
+##########################################################################
+# Run the two subgraph in pipeline mode and get the output asynchronously.
+pipeline_module.run()
+outputs = []
+while not outputs:
+    outputs = pipeline_module.get_output()
+    time.sleep(0.001)
+######################################
+# Use graph_executor for verification.
+# ------------------------------------
+# Run these two subgraphs in sequence with graph_executor to get the output.
+target = "llvm"
+dev0 = tvm.device(target, 0)
+lib0 = relay.build_module.build(mod0, target, params=params)
+module0 = runtime.GraphModule(lib0["default"](dev0))
+cuda = tvm.target.Target("cuda", host=tvm.target.Target("llvm"))
+lib1 = relay.build_module.build(mod1, [cuda, cutlass], params=params)
+lib1 = finalize_modules(lib1, "compile.so", "./tmp")
+
+dev1 = tvm.device("cuda", 0)
+
+module1 = runtime.GraphModule(lib1["default"](dev1))
+
+module0.set_input("data", data)
+module0.run()
+out_shape = (1, 16, img_size, img_size)
+out = module0.get_output(0, tvm.nd.empty(out_shape, "float16"))
+module1.set_input("data_n_0", out)
+module1.run()
+out_shape = (1, 1)
+out = module1.get_output(0, tvm.nd.empty(out_shape, "float16"))
+####################
+# Verify the result.
+tvm.testing.assert_allclose(outputs[0].numpy(), out.numpy())

python/tvm/contrib/pipeline_executor.py

@@ -302,11 +302,16 @@ def export_library(self, directory_path):
                 self.pipeline_mods[lib_index]["dev"].device_type,
                 self.pipeline_mods[lib_index]["dev"].device_id,
             )
-
             # Get the graph, lib, and parameters from GraphExecutorFactoryModule.
             lib = self.pipeline_mods[lib_index]["lib"]
             # Export the lib, graph, and parameters to disk.
-            lib.export_library(mconfig["lib_name"])
+            if self.pipeline_mods[lib_index]["export_cc"]:
+                lib.export_library(
+                    mconfig["lib_name"], cc=self.pipeline_mods[lib_index]["export_cc"]
+                )
+            else:
+                lib.export_library(mconfig["lib_name"])
+
             with open(mconfig["json_name"], "w") as file_handle:
                 file_handle.write(lib.graph_json)
             with open(mconfig["params_name"], "wb") as file_handle:

python/tvm/contrib/pipeline_executor_build.py

@@ -86,7 +86,12 @@ def build(pipe_configs):
         # Use "mod_idx" as the key to create a "module_connection" map which is not only
         # for the module index but also for the module connection used to build the pipeline.
         module_string_config[mod_idx] = pipe_config
-        libs[mod_idx] = {"lib": lib, "dev": dev, "fcompile": mod_config["fcompile"]}
+        libs[mod_idx] = {
+            "lib": lib,
+            "dev": dev,
+            "fcompile": mod_config["fcompile"],
+            "export_cc": mod_config["export_cc"],
+        }
 
     # Creating a text form configuration to record the "input_connection" and the
     # "module_connection" information. The "input_connection" is used to record the
@@ -132,10 +137,7 @@ def export_library(factory, directory_path):
         mconfig["json_name"] = "{}/json{}".format(directory_path, lib_index)
         mconfig["params_name"] = "{}/params{}".format(directory_path, lib_index)
         lib_config = factory.pipeline_mods[lib_index]
-        mconfig["dev"] = "{},{}".format(
-            lib_config["dev"].device_type,
-            lib_config["dev"].device_id,
-        )
+        mconfig["dev"] = "{},{}".format(lib_config["dev"].device_type, lib_config["dev"].device_id)
         fcompile = lib_config["fcompile"]
         if not fcompile:
             fcompile = False
@@ -413,6 +415,7 @@ def __init__(self, mod=None):
             self.fcompile = None
             self.name = None
             self.dev = None
+            self.export_cc = None
             self.cpu_affinity = ""
             self.idx = None
             self.mod = mod
@@ -601,6 +604,7 @@ def get_config(self):
                 "target": module.target,
                 "fcompile": module.fcompile,
                 "dev": module.dev,
+                "export_cc": module.export_cc,
             }
 
         # Creating a map including pipeline inputs and subgraph inputs.

tests/scripts/task_config_build_gpu.sh

@@ -47,3 +47,5 @@ echo set\(USE_LIBBACKTRACE AUTO\) >> config.cmake
 echo set\(USE_CCACHE OFF\) >> config.cmake
 echo set\(SUMMARIZE ON\) >> config.cmake
 echo set\(HIDE_PRIVATE_SYMBOLS ON\) >> config.cmake
+echo set\(USE_PIPELINE_EXECUTOR ON\) >> config.cmake
+echo set\(USE_CUTLASS ON\) >> config.cmake