Add performance analysis tool (#167)

Signed-off-by: Michael Mior <[email protected]>

contrib/CMakeLists.txt

@@ -22,4 +22,19 @@ if(BLAZE_COMPILER AND BLAZE_EVALUATOR)
     PRIVATE sourcemeta::blaze::compiler)
   target_link_libraries(sourcemeta_blaze_contrib_trace
     PRIVATE sourcemeta::blaze::evaluator)
+
+  if(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
+    add_executable(sourcemeta_blaze_contrib_perf perf.cc)
+    noa_add_default_options(PRIVATE sourcemeta_blaze_contrib_perf)
+    target_link_libraries(sourcemeta_blaze_contrib_perf
+      PRIVATE sourcemeta::jsontoolkit::json)
+    target_link_libraries(sourcemeta_blaze_contrib_perf
+      PRIVATE sourcemeta::jsontoolkit::jsonl)
+    target_link_libraries(sourcemeta_blaze_contrib_perf
+      PRIVATE sourcemeta::jsontoolkit::jsonschema)
+    target_link_libraries(sourcemeta_blaze_contrib_perf
+      PRIVATE sourcemeta::blaze::compiler)
+    target_link_libraries(sourcemeta_blaze_contrib_perf
+      PRIVATE sourcemeta::blaze::evaluator)
+  endif()
 endif()

contrib/perf.cc

@@ -0,0 +1,214 @@
+#include <sourcemeta/jsontoolkit/json.h>
+#include <sourcemeta/jsontoolkit/jsonl.h>
+#include <sourcemeta/jsontoolkit/jsonschema.h>
+
+#include <sourcemeta/blaze/compiler.h>
+#include <sourcemeta/blaze/evaluator.h>
+
+#include <linux/perf_event.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/syscall.h>
+
+#include <algorithm>  // std::sort, std::max
+#include <cstdlib>    // EXIT_SUCCESS, EXIT_FAILURE
+#include <filesystem> // std::filesystem::path
+#include <iostream>   // std::cerr, std::cout
+#include <map>        // std::map
+#include <utility>    // std::pair
+#include <vector>     // std::vector
+
+static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
+                            int cpu, long group_fd, unsigned long flags) {
+  long fd = syscall(SYS_perf_event_open, hw_event, pid, cpu, group_fd, flags);
+  if (fd == -1) {
+    fprintf(stderr, "Error creating event");
+    exit(EXIT_FAILURE);
+  }
+
+  return fd;
+}
+
+struct record_t {
+  uint64_t value;
+  uint64_t id;
+};
+
+struct read_format {
+  uint64_t nr;
+  record_t values[6];
+};
+
+class PerfEvents {
+private:
+  struct perf_event_attr pea;
+  unsigned long branch_predict_id;
+  long branch_predict_fd;
+  unsigned long branch_miss_id;
+  long branch_miss_fd;
+  unsigned long cache_access_id;
+  long cache_access_fd;
+  unsigned long cache_miss_id;
+  long cache_miss_fd;
+  unsigned long cycles_id;
+  long cycles_fd;
+  unsigned long instr_id;
+  long instr_fd;
+
+  void newHWEvent(struct perf_event_attr *newEvent, unsigned long type,
+                  long *fd, long group_fd, unsigned long *id) {
+    memset(newEvent, 0, sizeof(struct perf_event_attr));
+    newEvent->type = PERF_TYPE_HARDWARE;
+    newEvent->size = sizeof(struct perf_event_attr);
+    newEvent->config = type;
+    newEvent->disabled = 1;
+    newEvent->exclude_kernel = 1;
+    newEvent->exclude_hv = 1;
+    newEvent->read_format = PERF_FORMAT_GROUP | PERF_FORMAT_ID;
+    *fd = perf_event_open(newEvent, 0, -1, group_fd, 0);
+    ioctl(static_cast<int>(*fd), PERF_EVENT_IOC_ID, id);
+  }
+
+public:
+  PerfEvents() {
+    newHWEvent(&pea, PERF_COUNT_HW_BRANCH_INSTRUCTIONS, &branch_predict_fd, -1,
+               &branch_predict_id);
+    newHWEvent(&pea, PERF_COUNT_HW_BRANCH_MISSES, &branch_miss_fd,
+               branch_predict_fd, &branch_miss_id);
+    newHWEvent(&pea, PERF_COUNT_HW_CACHE_REFERENCES, &cache_access_fd,
+               branch_predict_fd, &cache_access_id);
+    newHWEvent(&pea, PERF_COUNT_HW_CACHE_MISSES, &cache_miss_fd,
+               branch_predict_fd, &cache_miss_id);
+    newHWEvent(&pea, PERF_COUNT_HW_CPU_CYCLES, &cycles_fd, branch_predict_fd,
+               &cycles_id);
+    newHWEvent(&pea, PERF_COUNT_HW_INSTRUCTIONS, &instr_fd, branch_predict_fd,
+               &instr_id);
+  }
+
+  void start() {
+    ioctl(static_cast<int>(branch_predict_fd), PERF_EVENT_IOC_RESET,
+          PERF_IOC_FLAG_GROUP);
+    ioctl(static_cast<int>(branch_predict_fd), PERF_EVENT_IOC_ENABLE,
+          PERF_IOC_FLAG_GROUP);
+  }
+
+  void stop() {
+    ioctl(static_cast<int>(branch_predict_fd), PERF_EVENT_IOC_DISABLE,
+          PERF_IOC_FLAG_GROUP);
+  }
+
+  void print() {
+    char buf[4096];
+    struct read_format *rf = (struct read_format *)buf;
+    unsigned long branch_predicts = 0;
+    unsigned long branch_misses = 0;
+    unsigned long cache_accesses = 0;
+    unsigned long cache_misses = 0;
+    unsigned long cycles = 0;
+    unsigned long instructions = 0;
+
+    if (read(static_cast<int>(branch_predict_fd), buf, sizeof(buf)) <= 0) {
+      std::cerr << "Error reading performance counters\n";
+      return;
+    }
+    for (unsigned long i = 0; i < rf->nr; i++) {
+      if (rf->values[i].id == branch_predict_id) {
+        branch_predicts = rf->values[i].value;
+      } else if (rf->values[i].id == branch_miss_id) {
+        branch_misses = rf->values[i].value;
+      } else if (rf->values[i].id == cache_access_id) {
+        cache_accesses = rf->values[i].value;
+      } else if (rf->values[i].id == cache_miss_id) {
+        cache_misses = rf->values[i].value;
+      } else if (rf->values[i].id == cycles_id) {
+        cycles = rf->values[i].value;
+      } else if (rf->values[i].id == instr_id) {
+        instructions = rf->values[i].value;
+      }
+    }
+
+    std::cerr << "       Cycles: " << cycles << "\n";
+    std::cerr << " Instructions: " << instructions << "\n";
+    std::cerr << " Cache misses: " << cache_misses << "/" << cache_accesses
+              << " = "
+              << (static_cast<double>(cache_misses) /
+                  static_cast<double>(cache_accesses) * 100)
+              << "%\n";
+    std::cerr << "Branch misses: " << branch_misses << "/" << branch_predicts
+              << " = "
+              << (static_cast<double>(branch_misses) /
+                  static_cast<double>(branch_predicts) * 100)
+              << "%\n";
+  }
+
+  ~PerfEvents() {
+    close(static_cast<int>(branch_miss_fd));
+    close(static_cast<int>(cache_access_fd));
+    close(static_cast<int>(cache_miss_fd));
+    close(static_cast<int>(cycles_fd));
+    close(static_cast<int>(instr_fd));
+    close(static_cast<int>(branch_predict_fd));
+  }
+};
+
+bool validate_all(auto &context, const auto &instances,
+                  const auto &schema_template) {
+  for (std::size_t num = 0; num < instances.size(); num++) {
+    context.prepare(instances[num]);
+    const auto result{sourcemeta::blaze::evaluate(schema_template, context)};
+    if (!result) {
+      std::cerr << "Error validating instance " << num << "\n";
+      return false;
+    }
+  }
+
+  return true;
+}
+
+auto main(int argc, char **argv) noexcept -> int {
+  if (argc < 3) {
+    std::cerr << "Usage: " << argv[0] << " <schema.json> <instances.jsonl>\n";
+    return EXIT_FAILURE;
+  }
+
+  const auto schema{sourcemeta::jsontoolkit::from_file(argv[1])};
+  const std::filesystem::path instance_path{argv[2]};
+  auto stream{sourcemeta::jsontoolkit::read_file(instance_path)};
+  std::vector<sourcemeta::jsontoolkit::JSON> instances;
+  for (const auto &instance : sourcemeta::jsontoolkit::JSONL{stream}) {
+    instances.push_back(instance);
+  }
+
+  const auto schema_template{sourcemeta::blaze::compile(
+      schema, sourcemeta::jsontoolkit::default_schema_walker,
+      sourcemeta::jsontoolkit::official_resolver,
+      sourcemeta::blaze::default_schema_compiler,
+      sourcemeta::blaze::Mode::FastValidation)};
+
+  sourcemeta::blaze::EvaluationContext context;
+  PerfEvents pe;
+
+  pe.start();
+  if (!validate_all(context, instances, schema_template)) {
+    return EXIT_FAILURE;
+  }
+  pe.stop();
+
+  std::cerr << "Cold\n==============================\n";
+  pe.print();
+  std::cerr << "\n";
+
+  for (int i = 0; i < 100; i++) {
+    validate_all(context, instances, schema_template);
+  }
+
+  pe.start();
+  validate_all(context, instances, schema_template);
+  pe.stop();
+
+  std::cerr << "Warm\n==============================\n";
+  pe.print();
+  std::cerr << "\n";
+
+  return EXIT_SUCCESS;
+}