masahi commented on code in PR #13450:
URL: https://github.com/apache/tvm/pull/13450#discussion_r1065125422
##########
src/relay/collage/gather_partition_specs.cc:
##########
@@ -103,6 +101,38 @@ BYOCStyle BYOCFusionStyleForCompiler(const String&
compiler) {
}
}
+/*!
+ * \brief Returns the fusion style for given compiler.
+ */
+BYOCStyle BYOCFusionStyleForCompiler(const String& compiler) {
+ tvm::transform::PassContext ctxt = tvm::transform::PassContext::Current();
+ std::string config_key = "relay.collage.byoc_fusion_style";
+ Optional<Array<String>> byoc_configs = ctxt->GetConfig(config_key,
Optional<Array<String>>());
+ BYOCStyle byoc_fusion_style = DefaultBYOCFusionStyleForCompiler(compiler);
+ if (!byoc_configs) {
+ LOG(INFO)<<"default fusion style";
+ return byoc_fusion_style;
+ }
+ for (auto config_ : byoc_configs.value()) {
+ std::vector<std::string> byoc_cfg = SplitString(config_, ".");
+ LOG(INFO) <<"byoc cfg: "<<byoc_cfg[0]<<" : "<<byoc_cfg[1];
+ if (byoc_cfg[0] == compiler) {
+ if (byoc_cfg[1] == "NoFusion") {
+ LOG(INFO)<<"NoFusion detected";
Review Comment:
Remove `LOG(INFO)`, three in this file
##########
src/relay/collage/utils.cc:
##########
@@ -134,6 +134,19 @@ bool MustBeLowered(const Expr& expr) {
return false;
}
+std::vector<std::string> SplitString(std::string stmt, const char* del) {
Review Comment:
Just define this in `gather_partition_specs.cc` where it's used, since it is
a one-off function
##########
tests/python/contrib/test_clml/test_adreno_collage_targets.py:
##########
@@ -0,0 +1,401 @@
+# 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.
+
+"""Compares Collage with various other baselines."""
+
+import tvm
+import logging
+import tempfile
+import os
+import shutil
+import numpy as np
+from tvm.relay import testing
+from tvm import rpc
+from tvm.contrib import utils, ndk
+from tvm.relay.build_module import bind_params_by_name
+
+# The following are necessary to force global functions or pattern tables to
be registered
+from tvm.relay.collage.collage import *
+from tvm.relay.op.contrib import clml
+import pytest
+
+logging.basicConfig(level=logging.INFO)
+
+
+########### Configuration ###########
+
+###
+### TVM Opencl AutoTvm log file name
+###
+TUNING_LOG = ""
+
+###
+### If true, run all models
+###
+ALL_MODELS = False
+
+###
+### If true, run all configurations
+###
+ALL_CONFIGS = False
+
+###
+### How aggressively to look for candidates?
+###
+TVM_MAX_DEPTH = 8
+BYOC_MAX_DEPTH = 8
+
+###
+### AutoTVM tuning parameters.
+###
+AUTOTVM_NUM_TRIALS = 1024
+AUTOTVM_EARLY_STOPPING = 600
+TIMEOUT = 10
+MEASURE_NUMBER = tvm.relay.collage.MEASURE_NUMBER
+MEASURE_REPEAT = tvm.relay.collage.MEASURE_REPEAT
+WARMUP_MIN_REPEAT_MS = tvm.relay.collage.WARMUP_MIN_REPEAT_MS
+
+##
+## RPC Build configuration
+##
+HOST = tvm.target.Target("llvm -mtriple=arm64-linux-android")
+OPENCL = tvm.target.Target("opencl", HOST)
+RPC_TRACKER_HOST = os.getenv("TVM_TRACKER_HOST", "localhost")
+RPC_TRACKER_PORT = int(os.getenv("TVM_TRACKER_PORT", 9090))
+RPC_KEY = os.getenv("RPC_DEVICE_KEY", "android")
+NDK_CROSS_COMPILER = os.getenv("TVM_NDK_CC", "aarch64-linux-android-g++")
+
+########### Runtime ###########
+
+# Code to run a model. The actual call to 'run' is appended at compile time.
+# We invoke the model as a sub-process so that we can wrap profiling tools
around it.
+runner_template = f"""
Review Comment:
Just make it a function instead of a template.
##########
tests/python/contrib/test_clml/test_adreno_collage_targets.py:
##########
@@ -0,0 +1,391 @@
+# 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.
+
+"""Compares Collage with various other baselines."""
+
+import tvm
+import logging
+import tempfile
+import os
+import shutil
+import numpy as np
+from test_clml import menangerie
+from tvm import rpc
+from tvm.contrib import utils, ndk
+
+# The following are necessary to force global functions or pattern tables to
be registered
+from tvm.relay.collage.collage import *
+from tvm.relay.op.contrib import clml
+import pytest
+
+logging.basicConfig(level=logging.INFO)
+
+
+########### Configuration ###########
+
+###
+### TVM Opencl AutoTvm log file name
+###
+TUNING_LOG = ""
+
+###
+### If true, run all models
+###
+ALL_MODELS = False
+
+###
+### If true, run all configurations
+###
+ALL_CONFIGS = False
+
+###
+### How aggressively to look for candidates?
+###
+TVM_MAX_DEPTH = 8
+BYOC_MAX_DEPTH = 8
+
+###
+### AutoTVM tuning parameters.
+###
+AUTOTVM_NUM_TRIALS = 1024
+AUTOTVM_EARLY_STOPPING = 600
+TIMEOUT = 10
+MEASURE_NUMBER = tvm.relay.collage.MEASURE_NUMBER
+MEASURE_REPEAT = tvm.relay.collage.MEASURE_REPEAT
+WARMUP_MIN_REPEAT_MS = tvm.relay.collage.WARMUP_MIN_REPEAT_MS
+
+##
+## RPC Build configuration
+##
+HOST = tvm.target.Target("llvm -mtriple=arm64-linux-android")
+OPENCL = tvm.target.Target("opencl", HOST)
+RPC_TRACKER_HOST = os.getenv("TVM_TRACKER_HOST", "localhost")
+RPC_TRACKER_PORT = int(os.getenv("TVM_TRACKER_PORT", 9090))
+RPC_KEY = os.getenv("RPC_DEVICE_KEY", "android")
+NDK_CROSS_COMPILER = os.getenv("TVM_NDK_CC", "aarch64-linux-android-g++")
+
+########### Runtime ###########
+
+# Code to run a model. The actual call to 'run' is appended at compile time.
+# We invoke the model as a sub-process so that we can wrap profiling tools
around it.
+runner_template = f"""
+import tvm
+import tvm.runtime.vm
+import numpy as np
+import logging
+from tvm import rpc
+import os
+logging.basicConfig(level=logging.INFO)
+
+RPC_TRACKER_HOST = os.environ["TVM_TRACKER_HOST"]
+RPC_TRACKER_PORT = int(os.environ["TVM_TRACKER_PORT"])
+RPC_KEY = "android"
+MEASURE_NUMBER = {MEASURE_NUMBER}
+MEASURE_REPEAT = {MEASURE_REPEAT}
+WARMUP_MIN_REPEAT_MS = {WARMUP_MIN_REPEAT_MS}
+
+def arg_for(shape, dtype, device):
+ return tvm.nd.array(
+ np.random.rand(*shape).astype(dtype), device=device)
+
+def vm_estimate_seconds(device, vm, args):
+ vm.benchmark(device, repeat=1, number=1,
min_repeat_ms=WARMUP_MIN_REPEAT_MS, **args)
+ return vm.benchmark(device, repeat=MEASURE_REPEAT, number=MEASURE_NUMBER,
min_repeat_ms=0,
+ **args)
+
+
+def run(label, name, lib_path, input_shapes, input_dtypes):
+ logging.info(f"Loading compiled code for {{name}} generated by {{label}}
from {{lib_path}}...")
+ tracker = rpc.connect_tracker(RPC_TRACKER_HOST, RPC_TRACKER_PORT)
+ remote = tracker.request(RPC_KEY, priority=0, session_timeout=600)
+ ctx = remote.cl(0)
+ remote_path = "/data/local/tmp/lib.so"
+ remote.upload(lib_path, target=remote_path)
+ lib = remote.load_module(remote_path)
+
+ vm_factory = tvm.runtime.vm.VirtualMachine(lib, ctx)
+ args = {{
+ input_name: arg_for(input_shapes[input_name],
input_dtypes[input_name], ctx)
+ for input_name in input_shapes.keys()
+ }}
+ logging.info(f"Benchmarking for {{name}} generated by {{label}}...")
+ profile = vm_estimate_seconds(ctx, vm_factory, args)
+ logging.info(f"Benchmarked for {{name}} generated by {{label}}:
{{profile}}")
+ logging.info(f"RESULT: {{label}} | {{name}} | {{profile.median * 1e3}}ms")
+
+if __name__ == "__main__":
+"""
+
+########### AutoTVM tuning helpers ###########
+
+
+def extract_autotvm_tasks(mod, target):
+ """Returns TVM kernels to tune for mod and target."""
+ return tvm.autotvm.task.extract_from_program(mod, target=target,
params=None)
+
+
+def optional_tuning_records(log_filename):
+ """Returns existing tuning records, if any."""
+ if log_filename == "" or not os.path.exists(log_filename):
+ return tvm.autotvm.task.FallbackContext()
+ else:
+ return tvm.autotvm.task.ApplyHistoryBest(log_filename)
+
+
+def is_already_tuned(task, log_filename):
+ """Returns True if we already have a tuning record for task in turning
logs in log_filename"""
+ if not os.path.exists(log_filename):
+ return False
+
+ dispatch_context = tvm.autotvm.task.ApplyHistoryBest(log_filename)
+ return dispatch_context._query_inside(task.target, task.workload)
+
+
+def tune_autotvm_tasks(tasks, log_filename):
+ """Appends to log filename the best strategies for tasks"""
+ if len(tasks) == 0:
+ return
+
+ measure_option = tvm.autotvm.measure_option(
+ builder=tvm.autotvm.LocalBuilder(build_func=ndk.create_shared,
timeout=15),
+ runner=tvm.autotvm.RPCRunner(
+ RPC_KEY, host=RPC_TRACKER_HOST, port=RPC_TRACKER_PORT, number=100,
timeout=15
+ ),
+ )
+
+ logging.info(
+ f"Using autotvm tuning for {len(tasks)} tasks with
{AUTOTVM_NUM_TRIALS} trials, logging to {log_filename}"
+ )
+
+ # create tmp log file, starting with contents from existing log file
+ tmp_log_filename = log_filename + ".tmp"
+ if os.path.exists(tmp_log_filename):
+ os.remove(tmp_log_filename)
+ if os.path.exists(log_filename):
+ logging.info(f"Copying existing log {log_filename} to
{tmp_log_filename}")
+ shutil.copy(log_filename, tmp_log_filename)
+
+ for i, task in enumerate(reversed(tasks)):
+ prefix = "[Task %2d/%2d] " % (i + 1, len(tasks))
+ logging.info(f"Considering task {task.name} {prefix}")
+ if is_already_tuned(task, tmp_log_filename):
+ logging.info(f"Re-using existing record for {task.name}")
+ continue
+
+ logging.info(f"Using autotvm to tune {task.name}")
+ tuner_obj = tvm.autotvm.tuner.XGBTuner(task, loss_type="rank")
+ if os.path.exists(tmp_log_filename):
+
tuner_obj.load_history(tvm.autotvm.record.load_from_file(tmp_log_filename))
+
+ # do tuning
+ n_trial = min(AUTOTVM_NUM_TRIALS, len(task.config_space))
+ tuner_obj.tune(
+ n_trial=n_trial,
+ early_stopping=AUTOTVM_EARLY_STOPPING,
+ measure_option=measure_option,
+ callbacks=[
+ tvm.autotvm.callback.progress_bar(n_trial, prefix=prefix),
+ tvm.autotvm.callback.log_to_file(tmp_log_filename),
+ ],
+ )
+
+ # Pick best records and copy back to main log file
+ tvm.autotvm.record.pick_best(tmp_log_filename, log_filename)
+ os.remove(tmp_log_filename)
+
+ logging.info("Done with autotvm tuning")
+
+
+def autotvm_tune_module(mod, target, log_filename):
+ if log_filename == "":
+ logging.info("Not tuning with autotvm since disabled")
+ return
+ # Extract and tune any TVM kernels. BYOC partitions will have no tasks
extracted.
+ logging.info("Extracting tasks from overall module")
+ tasks = extract_autotvm_tasks(mod, target)
+ logging.info(f"Auto-tuning {len(tasks)} tasks from overall module")
+ tune_autotvm_tasks(tasks, log_filename)
+
+
+########### Drivers ###########
+
+
+def compile_and_benchmark(label, model, targets, tmp_dir):
+ """Compile model for target and run it with profiling."""
+ logging.info(f"Compiling {model['name']} using {label} with {targets}...")
+ model["mod"] = clml.preprocess_for_clml(model["mod"])
+ exe = tvm.relay.vm.compile(model["mod"], target=targets,
params=model["params"])
+ lib = exe.mod
+ lib_path = os.path.join(tmp_dir, "lib.so")
+ logging.info(f"Exporting library to {lib_path}...")
+ lib.export_library(lib_path, cc=NDK_CROSS_COMPILER)
+ runner = f"{runner_template} run('{label}', '{model['name']}',
'{lib_path}', {model['input_shapes']}, {model['input_dtypes']})\n"
+ runner_path = os.path.join(tmp_dir, "runner.py")
+ logging.info(f"Saving runner to {runner_path}...")
+ with open(runner_path, "w") as fo:
+ fo.write(runner)
+
+ logging.info(f"Invoking runner...")
+
+ os.system(f"python3 {runner_path}")
Review Comment:
That doesn't answer my question. Why not, instead of `runner_template` thing
and invoke the string with `os.system`, wrap the same functionality in a normal
function and call it however time you want.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
