Get more consistent distributions in parallel scenarios test (#8451)

* Get more consistent distributions in parallel scenarios test

* Fix decile reporting

test/performance/parallel_scenarios.cpp

@@ -31,32 +31,36 @@ int main(int argc, char **argv) {
 
     int native_threads = Halide::Internal::JITSharedRuntime::get_num_threads();
 
+    std::map<std::tuple<bool, bool, int, int>, std::vector<float>> results;
+
     auto bench = [&](bool m, bool c, int i, int o) {
-        const int num_samples = 128;
         const int memory_limit = m ? max_memory : 128;
 
+        auto now = std::chrono::high_resolution_clock::now;
+        auto to_ns = [](auto delta) { return 1e9 * std::chrono::duration<float>(delta).count(); };
+
         auto bench_one = [&]() {
-            auto t1 = std::chrono::high_resolution_clock::now();
-            // Ignore error code because default halide_error() will abort on failure
+            auto t1 = now();
             (void)callable(i, o, memory_limit, in, out);
-            auto t2 = std::chrono::high_resolution_clock::now();
-            return 1e9 * std::chrono::duration<float>(t2 - t1).count() / (i * o);
+            auto t2 = now();
+            return to_ns(t2 - t1) / (i * o);
         };
 
-        std::vector<float> times(num_samples);
+        const int num_tasks = 8;
+        const int min_samples = 32;
+
+        std::vector<float> times[num_tasks];
         if (c) {
             Halide::Tools::ThreadPool<void> thread_pool;
-            const int num_tasks = 8;
-            const int samples_per_task = num_samples / num_tasks;
             Halide::Internal::JITSharedRuntime::set_num_threads(num_tasks * native_threads);
             std::vector<std::future<void>> futures(num_tasks);
             for (size_t t = 0; t < futures.size(); t++) {
                 futures[t] = thread_pool.async(
                     [&](size_t t) {
                         bench_one();
-                        for (int s = 0; s < samples_per_task; s++) {
-                            size_t idx = t * samples_per_task + s;
-                            times[idx] = bench_one();
+                        auto t_start = now();
+                        while (to_ns(now() - t_start) < 1e7 || times[t].size() < min_samples / num_tasks) {
+                            times[t].push_back(bench_one());
                         }
                     },
                     t);
@@ -67,32 +71,43 @@ int main(int argc, char **argv) {
         } else {
             Halide::Internal::JITSharedRuntime::set_num_threads(native_threads);
             bench_one();
-            for (int s = 0; s < num_samples; s++) {
-                times[s] = bench_one();
+            auto t_start = now();
+            while (to_ns(now() - t_start) < 1e7 || times[0].size() < min_samples) {
+                times[0].push_back(bench_one());
             }
         }
-        std::sort(times.begin(), times.end());
-        printf("%d %d %d %d ", m, c, i, o);
-        const int n = 8;
-        int off = (num_samples / n) / 2;
-        for (int i = 0; i < n; i++) {
-            printf("%g ", times[off + (num_samples * i) / n]);
+
+        std::vector<float> &r = results[{m, c, i, o}];
+        for (int i = 0; i < num_tasks; i++) {
+            r.insert(r.end(), times[i].begin(), times[i].end());
         }
-        printf("\n");
     };
 
     // The output is designed to be copy-pasted into a spreadsheet, not read by a human
-    printf("memory_bound contended inner outer t0 t1 t2 t3 t4 t5 t7\n");
-    for (bool contended : {false, true}) {
-        for (bool memory_bound : {false, true}) {
-            for (int i : {1 << 0, 1 << 6, 1 << 12, 1 << 18}) {
-                for (int o : {1, 2, 4, 8, 16, 32, 64, 128, 256}) {
-                    bench(memory_bound, contended, i, o);
+    printf("memory_bound contended inner outer num_samples 10%% 20%% 30%% 40%% 50%% 60%% 70%% 80%% 90%%\n");
+    for (int repeat = 0; repeat < 10; repeat++) {
+        for (bool contended : {false, true}) {
+            for (bool memory_bound : {false, true}) {
+                for (int i : {1 << 6, 1 << 9, 1 << 12, 1 << 15}) {
+                    for (int o : {1, 2, 4, 8, 16, 32, 64, 128, 256}) {
+                        bench(memory_bound, contended, i, o);
+                    }
                 }
             }
         }
     }
 
+    for (auto p : results) {
+        auto &times = p.second;
+        std::sort(times.begin(), times.end());
+        auto [m, c, i, o] = p.first;
+        printf("%d %d %d %d %d ", m, c, i, o, (int)times.size());
+        for (int decile = 10; decile <= 90; decile += 10) {
+            printf("%g ", times[(decile * times.size()) / 100]);
+        }
+        printf("\n");
+    }
+
     printf("Success!\n");
 
     return 0;