Compiler Static Information Integration

thicket/compiler_static_info_reader.py

@@ -0,0 +1,178 @@
+# Copyright 2022 Lawrence Livermore National Security, LLC and other
+# Thicket Project Developers. See the top-level LICENSE file for details.
+#
+# SPDX-License-Identifier: MIT
+import os
+import json
+import math
+from functools import lru_cache
+from collections import defaultdict
+
+
+class CompilerStaticInfoReader:
+    MAX_IR_METRICS = {"maxLiveSSAValues", "maxLoopDepth"}
+    SUM_KRU_METRICS = {"SGPRs Spill", "VGPRs Spill"}
+
+    def __init__(self, extraction_plugin_out_dir, trace_log_file):
+        self.extraction_plugin_out_dir = extraction_plugin_out_dir
+        self.trace_log_file = trace_log_file
+
+    def read_trace_log(self):
+        with open(self.trace_log_file, "r") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    yield json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+
+    def build_path_to_node(self, graph):
+        path_to_node = {}
+
+        def dfs(node, path_parts, visited):
+            if node in visited:
+                return
+            visited.add(node)
+
+            name = node.frame["name"]
+            curr_path = path_parts + [name]
+            path_to_node["/".join(curr_path)] = node
+
+            for child in node.children:
+                dfs(child, curr_path, visited)
+
+        visited = set()
+        for root in graph.roots:
+            dfs(root, [], visited)
+
+        return path_to_node
+
+    @staticmethod
+    @lru_cache(maxsize=None)
+    def _load_module_functions_cached(extraction_plugin_out_dir, module_relative_path):
+        module_relative_path = module_relative_path.lstrip("/")
+        path = os.path.join(extraction_plugin_out_dir, module_relative_path)
+
+        if not os.path.isfile(path):
+            return {}
+
+        try:
+            with open(path, "r") as f:
+                return json.load(f).get("Functions", {})
+        except Exception:
+            return {}
+
+    def load_module_functions(self, module_relative_path):
+        return self._load_module_functions_cached(
+            self.extraction_plugin_out_dir,
+            module_relative_path,
+        )
+
+    def walk_call_tree(
+        self,
+        func_name,
+        module_relative_path,
+        ir_accum,
+        opt_accum,
+        visited,
+    ):
+        key = (func_name, module_relative_path)
+        if key in visited:
+            return
+        visited.add(key)
+
+        entry = self.load_module_functions(module_relative_path).get(func_name)
+        if not entry:
+            return
+
+        for metric, value in entry.get("IRStatistics", {}).items():
+            if value is None:
+                continue
+            if metric in self.MAX_IR_METRICS:
+                ir_accum[metric] = max(ir_accum[metric], value)
+            else:
+                ir_accum[metric] += value
+
+        for metric, value in (
+            entry.get("RemarkInformation", {}).get("RemarkStatistics", {}).items()
+        ):
+            if value is not None:
+                opt_accum[metric] += value
+
+        for callee in entry.get("CalledFunctions", []):
+            self.walk_call_tree(
+                callee,
+                module_relative_path,
+                ir_accum,
+                opt_accum,
+                visited,
+            )
+
+    def add_to_thicket(self, thicket):
+        region_index = self.build_path_to_node(thicket.graph)
+
+        region_roots = defaultdict(set)
+
+        for entry in self.read_trace_log():
+            region_path = entry.get("RegionPath")
+            func_name = entry.get("Function")
+            module_relative_path = entry.get("ModulePath")
+
+            if not region_path or not func_name or not module_relative_path:
+                continue
+
+            region_roots[region_path].add((func_name, module_relative_path))
+
+        for region_path, roots in region_roots.items():
+            th_node = region_index.get(region_path)
+            if th_node is None:
+                print(
+                    f"CompilerStaticInfoReader: Missing region path in thicket: {region_path}"
+                )
+                continue
+
+            ir_accum = defaultdict(int)
+            opt_accum = defaultdict(int)
+            kru_sum = defaultdict(int)
+            kru_min = defaultdict(lambda: math.inf)
+            kru_max = defaultdict(lambda: -math.inf)
+
+            for func_name, module_relative_path in roots:
+                entry = self.load_module_functions(module_relative_path).get(func_name)
+                if not entry:
+                    continue
+
+                # KRU: aggregate only root functions for this region
+                for metric, value in entry.get("KRUInformation", {}).items():
+                    if value is None:
+                        continue
+                    if metric in self.SUM_KRU_METRICS:
+                        kru_sum[f"{metric} (sum)"] += value
+                    else:
+                        kru_min[metric] = min(kru_min[metric], value)
+                        kru_max[metric] = max(kru_max[metric], value)
+
+                # IR + OPT: aggregate over full reachable call tree
+                self.walk_call_tree(
+                    func_name,
+                    module_relative_path,
+                    ir_accum,
+                    opt_accum,
+                    visited=set(),
+                )
+
+            for metric, value in ir_accum.items():
+                thicket.dataframe.loc[th_node, metric] = value
+
+            for metric, value in opt_accum.items():
+                thicket.dataframe.loc[th_node, metric] = value
+
+            for metric, value in kru_sum.items():
+                thicket.dataframe.loc[th_node, metric] = value
+
+            for metric in kru_min:
+                if kru_min[metric] != math.inf:
+                    thicket.dataframe.loc[th_node, f"{metric} (min)"] = kru_min[metric]
+                    thicket.dataframe.loc[th_node, f"{metric} (max)"] = kru_max[metric]

thicket/tests/conftest.py

@@ -204,6 +204,41 @@ def rajaperf_unique_tunings(data_dir, tmpdir):
     return [os.path.join(str(tmpdir), f) for f in cali_files]
 
 
+@pytest.fixture
+def rajaperf_hip_O2_512M_cali(data_dir, tmpdir):
+    """All trials of for RAJA HIP optimization level O2 and total node problem size 536870912."""
+    cali_files = sorted(
+        glob(
+            f"{data_dir}/rajaperf-comp-static-data/tuolumne/runtime/O2/**/RAJA_HIP-block_256.cali",
+            recursive=True,
+        )
+    )
+    for cf in cali_files:
+        shutil.copy(cf, str(tmpdir))
+    return [os.path.join(str(tmpdir), f) for f in cali_files]
+
+
+@pytest.fixture
+def rajaperf_hip_O2_512M_compiler_static_info(data_dir, tmpdir):
+    """Trace log and extraction plugin output for RAJA HIP at optimization level O2 and problem size 512M"""
+    trace_log_file = os.path.join(
+        data_dir,
+        "rajaperf-comp-static-data/tuolumne/llvm_pass_plugins/association/trace_log.jsonl",
+    )
+    extraction_out_path = os.path.join(
+        data_dir,
+        "rajaperf-comp-static-data/tuolumne/llvm_pass_plugins/extraction/extraction_plugin_output",
+    )
+
+    dest_log = os.path.join(str(tmpdir), os.path.basename(trace_log_file))
+    dest_dir = os.path.join(str(tmpdir), "extraction_plugin_output")
+
+    shutil.copy(trace_log_file, dest_log)
+    shutil.copytree(extraction_out_path, dest_dir)
+
+    return [dest_log, dest_dir]
+
+
 @pytest.fixture
 def caliper_ordered(data_dir, tmpdir):
     """Builds a temporary directory containing the lulesh cali file."""

thicket/tests/data/rajaperf-comp-static-data/tuolumne/llvm_pass_plugins/association/trace_log.jsonl

@@ -0,0 +1,5 @@
+{"Function":"void RAJA::policy::hip::impl::forallp_hip_kernel<RAJA::policy::hip::hip_exec<RAJA::iteration_mapping::Direct, RAJA::hip::IndexGlobal<(RAJA::named_dim)0, 256, 0>, RAJA::hip::AvoidDeviceMaxThreadOccupancyConcretizer<RAJA::hip::FractionOffsetOccupancyConcretizer<RAJA::Fraction<unsigned long, 1ul, 1ul>, -1l>>, true>, RAJA::Iterators::numeric_iterator<long, long, long*>, void rajaperf::stream::TRIAD::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long), long, RAJA::expt::ForallParamPack<>, RAJA::iteration_mapping::Direct, RAJA::hip::IndexGlobal<(RAJA::named_dim)0, 256, 0>, 256ul>(void rajaperf::stream::TRIAD::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long), RAJA::Iterators::numeric_iterator<long, long, long*>, long, RAJA::expt::ForallParamPack<>)","ModulePath":"/src/stream/TRIAD-Hip.cpp.device.json","RegionPath":"RAJAPerf/Stream/Stream_TRIAD"}
+{"Function":"void RAJA::launch_new_reduce_global_fcn_fixed<void rajaperf::basic::MAT_MAT_SHARED::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(RAJA::LaunchContext), 256, RAJA::expt::ForallParamPack<>>(void rajaperf::basic::MAT_MAT_SHARED::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(RAJA::LaunchContext), RAJA::expt::ForallParamPack<>)","ModulePath":"/src/basic/MAT_MAT_SHARED-Hip.cpp.device.json","RegionPath":"RAJAPerf/Basic/Basic_MAT_MAT_SHARED"}
+{"Function":"void RAJA::internal::HipKernelLauncherFixed<256, RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, double&)>, RAJA::internal::HipStatementListExecutor<RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, double&)>, camp::list<RAJA::statement::For<0l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)1, 8, 0>>, RAJA::statement::For<1l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)0, 32, 0>>, RAJA::statement::Lambda<0l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>, RAJA::statement::For<2l, RAJA::policy::sequential::seq_exec, RAJA::statement::Lambda<1l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 2l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>, RAJA::statement::Lambda<2l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>>>, RAJA::internal::LoopTypes<camp::list<void, void, void>, camp::list<void, void, void>>>>(RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda'(long, long, double&)>)","ModulePath":"/src/polybench/POLYBENCH_3MM-Hip.cpp.device.json","RegionPath":"RAJAPerf/Polybench/Polybench_3MM"}
+{"Function":"void RAJA::internal::HipKernelLauncherFixed<256, RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, double&)>, RAJA::internal::HipStatementListExecutor<RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, double&)>, camp::list<RAJA::statement::For<0l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)1, 8, 0>>, RAJA::statement::For<1l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)0, 32, 0>>, RAJA::statement::Lambda<0l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>, RAJA::statement::For<2l, RAJA::policy::sequential::seq_exec, RAJA::statement::Lambda<1l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 2l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>, RAJA::statement::Lambda<2l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>>>, RAJA::internal::LoopTypes<camp::list<void, void, void>, camp::list<void, void, void>>>>(RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda0'(long, long, double&)>)","ModulePath":"/src/polybench/POLYBENCH_3MM-Hip.cpp.device.json","RegionPath":"RAJAPerf/Polybench/Polybench_3MM"}
+{"Function":"void RAJA::internal::HipKernelLauncherFixed<256, RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, double&)>, RAJA::internal::HipStatementListExecutor<RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, double&)>, camp::list<RAJA::statement::For<0l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)1, 8, 0>>, RAJA::statement::For<1l, RAJA::policy::hip::hip_indexer<RAJA::iteration_mapping::Direct, (RAJA::kernel_sync_requirement)0, RAJA::hip::IndexGlobal<(RAJA::named_dim)0, 32, 0>>, RAJA::statement::Lambda<0l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>, RAJA::statement::For<2l, RAJA::policy::sequential::seq_exec, RAJA::statement::Lambda<1l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 2l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>, RAJA::statement::Lambda<2l, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 0l>, RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_seg_t, 1l>>, camp::list<RAJA::internal::LambdaArg<RAJA::internal::lambda_arg_param_t, 0l>>>>>>, RAJA::internal::LoopTypes<camp::list<void, void, void>, camp::list<void, void, void>>>>(RAJA::internal::LoopData<camp::tuple<RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>, RAJA::Span<RAJA::Iterators::numeric_iterator<long, long, long*>, long>>, camp::tuple<double>, camp::resources::v1::Hip, void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, long, double&), void rajaperf::polybench::POLYBENCH_3MM::runHipVariantImpl<256ul>(rajaperf::VariantID)::'lambda1'(long, long, double&)>)","ModulePath":"/src/polybench/POLYBENCH_3MM-Hip.cpp.device.json","RegionPath":"RAJAPerf/Polybench/Polybench_3MM"}