Added collecting and printing of scores for the kernel-selection tuner

2 files changed, 61 insertions, 9 deletions

diff --git a/src/tuning/routines/xgemm.cpp b/src/tuning/routines/xgemm.cpp
index 9590323a..1ccaa0ca 100644
--- a/src/tuning/routines/xgemm.cpp
+++ b/src/tuning/routines/xgemm.cpp
@@ -90,15 +90,36 @@ void TuneXgemm(int argc, char* argv[]) {
   ForceSelectIndirectFrom<T>(to * to * to + 1, device);
   const auto direct = TimeRoutine(from, to, step, num_runs, queue, buffers, RunGemmRoutine<T>);
 
-  // Results
-  printf("[----------] Collecting results\n");
+  // Determining final score and best kernel selection point
   assert(indirect.size() == direct.size());
+  printf("[----------] Collecting results\n");
+  auto ratios = std::vector<double>(indirect.size());
+  for (auto i = size_t{0}; i < indirect.size(); ++i) {
+    ratios[i] = indirect[i].second / direct[i].second;
+  }
+  auto scores = std::vector<TuningResult>(ratios.size());
+  for (auto i = size_t{0}; i < scores.size(); ++i) {
+    auto score = 0;
+    for (auto j = size_t{0}; j < i; ++j) { score += (ratios[j] <= 1.0); }
+    for (auto j = i + 1; j < ratios.size(); ++j) { score += (ratios[j] > 1.0); }
+    const auto epsilon = (scores.size() - i) / 1e3; // favour later results over earlier ones
+    scores[i] = TuningResult{
+        "gemm_kernel_selection",
+        static_cast<double>(score) / static_cast<double>(scores.size() - 1) + epsilon,
+        TuningParameters{TuningParameter{"XGEMM_MIN_INDIRECT_SIZE", indirect[i].first}}
+    };
+  }
+
+  // Displaying results
   for (auto i = size_t{0}; i < indirect.size(); ++i) {
     assert(indirect[i].first == direct[i].first);
     const auto value = indirect[i].first;
-    const auto gflops_indirect = (2 * value * value * value) / (indirect[i].second * 1.0e6);
-    const auto gflops_direct = (2 * value * value * value) / (direct[i].second * 1.0e6);
-    printf("[ -------> ] %7zu %8.2lf %8.2lf\n", value, gflops_indirect, gflops_direct);
+    if (indirect[i].second != -1 && direct[i].second != -1) {
+      const auto gflops_indirect = (2 * value * value * value) / (indirect[i].second * 1.0e6);
+      const auto gflops_direct = (2 * value * value * value) / (direct[i].second * 1.0e6);
+      printf("[ -------> ] %7zu %8.2lf %8.2lf %8.2lf\n",
+             value, gflops_indirect, gflops_direct, scores[i].score);
+    }
   }
 
   // Outputs the results as JSON to disk, including some meta-data
@@ -108,7 +129,7 @@ void TuneXgemm(int argc, char* argv[]) {
       {"precision", precision_string},
   };
   PrintTimingsToFileAsJSON("clblast_routine_gemm_" + precision_string + ".json",
-                           device, platform, metadata);
+                           device, platform, metadata, scores);
 
 }
 
diff --git a/src/utilities/timing.hpp b/src/utilities/timing.hpp
index 4622aa99..423e6e2b 100644
--- a/src/utilities/timing.hpp
+++ b/src/utilities/timing.hpp
@@ -65,13 +65,18 @@ std::vector<Timing> TimeRoutine(const size_t from, const size_t to, const size_t
 
 // =================================================================================================
 
+using TuningParameter = std::pair<std::string, size_t>;
+using TuningParameters = std::vector<TuningParameter>;
+struct TuningResult { std::string name; double score; TuningParameters parameters; };
+
 void PrintTimingsToFileAsJSON(const std::string &filename,
                               const Device& device, const Platform& platform,
-                              const std::vector<std::pair<std::string,std::string>> &descriptions) {
+                              const std::vector<std::pair<std::string,std::string>> &metadata,
+                              const std::vector<TuningResult>& tuning_results) {
   auto file = fopen(filename.c_str(), "w");
   fprintf(file, "{\n");
-  for (auto &description: descriptions) {
-    fprintf(file, "  \"%s\": \"%s\",\n", description.first.c_str(), description.second.c_str());
+  for (auto &datum: metadata) {
+    fprintf(file, "  \"%s\": \"%s\",\n", datum.first.c_str(), datum.second.c_str());
   }
   fprintf(file, "  \"platform_version\": \"%s\",\n", platform.Version().c_str());
   fprintf(file, "  \"device_name\": \"%s\",\n", GetDeviceName(device).c_str());
@@ -80,6 +85,32 @@ void PrintTimingsToFileAsJSON(const std::string &filename,
   fprintf(file, "  \"device_architecture\": \"%s\",\n", GetDeviceArchitecture(device).c_str());
   fprintf(file, "  \"device_core_clock\": \"%zu\",\n", device.CoreClock());
   fprintf(file, "  \"device_compute_units\": \"%zu\",\n", device.ComputeUnits());
+  fprintf(file, "  \"results\": [\n");
+
+  // Loops over all results
+  auto num_results = tuning_results.size();
+  for (auto r = size_t{0}; r < num_results; ++r) {
+    auto result = tuning_results[r];
+    fprintf(file, "    {\n");
+    fprintf(file, "      \"kernel\": \"%s\",\n", result.name.c_str());
+    fprintf(file, "      \"time\": %.3lf,\n", result.score);
+
+    // Loops over all the parameters for this result
+    fprintf(file, "      \"parameters\": {");
+    auto num_configs = result.parameters.size();
+    for (auto p=size_t{0}; p<num_configs; ++p) {
+      auto config = result.parameters[p];
+      fprintf(file, "\"%s\": %zu", config.first.c_str(), config.second);
+      if (p < num_configs-1) { fprintf(file, ","); }
+    }
+    fprintf(file, "}\n");
+
+    // The footer
+    fprintf(file, "    }");
+    if (r < num_results - 1) { fprintf(file, ","); }
+    fprintf(file, "\n");
+  }
+  fprintf(file, "  ]\n");
   fprintf(file, "}\n");
   fclose(file);
 }