WIP: Start work on divide and conquer example with temporary buffers.

CMakeLists.txt

@@ -48,6 +48,7 @@ add_subdirectory(app/playground)
 add_subdirectory(app/benchmark_fft)
 add_subdirectory(app/benchmark_fft)
 add_subdirectory(app/benchmark_unbalanced)
 add_subdirectory(app/benchmark_unbalanced)
 add_subdirectory(app/benchmark_matrix)
 add_subdirectory(app/benchmark_matrix)
+add_subdirectory(app/benchmark_matrix_div_conquer)
 add_subdirectory(app/benchmark_fib)
 add_subdirectory(app/benchmark_fib)
 add_subdirectory(app/context_switch)
 add_subdirectory(app/context_switch)
 
 

app/benchmark_fib/main.cpp

 #include "pls/pls.h"
 #include "pls/pls.h"
 
 
-#include <iostream>
-
 #include "benchmark_runner.h"
 #include "benchmark_runner.h"
 #include "benchmark_base/fib.h"
 #include "benchmark_base/fib.h"
 
 
@@ -10,7 +8,7 @@ using namespace comparison_benchmarks::base;
 constexpr int MAX_NUM_TASKS = 32;
 constexpr int MAX_NUM_TASKS = 32;
 constexpr int MAX_STACK_SIZE = 4096 * 1;
 constexpr int MAX_STACK_SIZE = 4096 * 1;
 
 
-int pls_fib(int n) {
+int pls_fib(int n, int d) {
   if (n == 0) {
   if (n == 0) {
     return 0;
     return 0;
   }
   }
@@ -19,11 +17,11 @@ int pls_fib(int n) {
   }
   }
 
 
   int a, b;
   int a, b;
-  pls::spawn([n, &a]() {
-    a = pls_fib(n - 1);
+  pls::spawn([n, d, &a]() {
+    a = pls_fib(n - 1, d + 1);
   });
   });
-  pls::spawn([n, &b]() {
-    b = pls_fib(n - 2);
+  pls::spawn([n, d, &b]() {
+    b = pls_fib(n - 2, d + 1);
   });
   });
   pls::sync();
   pls::sync();
 
 
@@ -45,7 +43,7 @@ int main(int argc, char **argv) {
 //  scheduler.get_profiler().disable_memory_measure();
 //  scheduler.get_profiler().disable_memory_measure();
   runner.run_iterations(fib::NUM_ITERATIONS, [&]() {
   runner.run_iterations(fib::NUM_ITERATIONS, [&]() {
     scheduler.perform_work([&]() {
     scheduler.perform_work([&]() {
-      res = pls_fib(fib::INPUT_N);
+      res = pls_fib(fib::INPUT_N, 0);
     });
     });
   }, fib::NUM_WARMUP_ITERATIONS);
   }, fib::NUM_WARMUP_ITERATIONS);
 //  scheduler.get_profiler().current_run().print_dag(std::cout);
 //  scheduler.get_profiler().current_run().print_dag(std::cout);

app/benchmark_matrix_div_conquer/CMakeLists.txt

+add_executable(benchmark_matrix_div_conquer_pls_v3 main.cpp)
+target_link_libraries(benchmark_matrix_div_conquer_pls_v3 pls benchmark_runner benchmark_base)
+if (EASY_PROFILER)
+    target_link_libraries(benchmark_matrix_div_conquer_pls_v3 easy_profiler)
+endif ()

app/benchmark_matrix_div_conquer/main.cpp

+//#include "pls/pls.h"
+//using namespace pls;
+
+#include "benchmark_runner.h"
+
+#include <memory>
+#include <array>
+#include <math.h>
+
+// Helpers to directly index into blocked matrices
+const size_t MAX_BLOCK_LOOKUP = 256;
+std::array<std::array<size_t, MAX_BLOCK_LOOKUP>, MAX_BLOCK_LOOKUP> BLOCK_LOOKUP; // ROW, COLUMN
+void fill_block_lookup(size_t size) {
+  if (size <= 1) {
+    BLOCK_LOOKUP[0][0] = 0;
+    return;
+  }
+
+  fill_block_lookup(size / 2);
+
+  size_t elements_per_quarter = (size / 2) * (size / 2);
+  for (size_t row = 0; row < size / 2; row++) {
+    for (size_t column = 0; column < size / 2; column++) {
+      BLOCK_LOOKUP[row][size / 2 + column] = BLOCK_LOOKUP[row][column] + elements_per_quarter;
+      BLOCK_LOOKUP[size / 2 + row][column] = BLOCK_LOOKUP[row][column] + 2 * elements_per_quarter;
+      BLOCK_LOOKUP[size / 2 + row][size / 2 + column] = BLOCK_LOOKUP[row][column] + 3 * elements_per_quarter;
+    }
+  }
+}
+class blocked_matrix_view {
+ public:
+  blocked_matrix_view(double *data, size_t size) : data_{data}, size_{size} {}
+
+  blocked_matrix_view quadrant_1_1() {
+    size_t elements_per_quarter = (size_ / 2) * (size_ / 2);
+    return blocked_matrix_view(data_ + 0 * elements_per_quarter, size_ / 2);
+  }
+  blocked_matrix_view quadrant_1_2() {
+    size_t elements_per_quarter = (size_ / 2) * (size_ / 2);
+    return blocked_matrix_view(data_ + 1 * elements_per_quarter, size_ / 2);
+  }
+  blocked_matrix_view quadrant_2_1() {
+    size_t elements_per_quarter = (size_ / 2) * (size_ / 2);
+    return blocked_matrix_view(data_ + 2 * elements_per_quarter, size_ / 2);
+  }
+  blocked_matrix_view quadrant_2_2() {
+    size_t elements_per_quarter = (size_ / 2) * (size_ / 2);
+    return blocked_matrix_view(data_ + 3 * elements_per_quarter, size_ / 2);
+  }
+
+  double &at(size_t row, size_t column) {
+    return data_[BLOCK_LOOKUP[row][column]];
+  }
+
+  double *get_data() {
+    return data_;
+  }
+
+ private:
+  double *data_;
+  size_t size_;
+};
+
+void multiply_naive(size_t size, blocked_matrix_view &result, blocked_matrix_view &a, blocked_matrix_view &b) {
+  for (size_t i = 0; i < size; i++) {
+    for (size_t j = 0; j < size; j++) {
+      result.at(i, j) = 0;
+    }
+    for (size_t j = 0; j < size; j++) {
+      for (size_t k = 0; k < size; k++) {
+        result.at(i, j) += a.at(i, k) * b.at(k, j);
+      }
+    }
+  }
+}
+
+void multiply_div_conquer(size_t size, blocked_matrix_view &result, blocked_matrix_view &a, blocked_matrix_view &b) {
+  if (size <= 8) {
+    multiply_naive(size, result, a, b);
+    return;
+  }
+  // Temporary storage required for the intermediate results
+  std::unique_ptr<double[]> data_1_1_a{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_1_1_b{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_1_2_a{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_1_2_b{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_2_1_a{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_2_1_b{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_2_2_a{new double[(size / 2) * (size / 2)]};
+  std::unique_ptr<double[]> data_2_2_b{new double[(size / 2) * (size / 2)]};
+
+  // Handles to sub-matrices used
+  blocked_matrix_view result_1_1 = result.quadrant_1_1();
+  blocked_matrix_view result_1_2 = result.quadrant_1_2();
+  blocked_matrix_view result_2_1 = result.quadrant_2_1();
+  blocked_matrix_view result_2_2 = result.quadrant_2_2();
+
+  blocked_matrix_view result_1_1_a{data_1_1_a.get(), size / 2};
+  blocked_matrix_view result_1_1_b{data_1_1_b.get(), size / 2};
+  blocked_matrix_view result_1_2_a{data_1_2_a.get(), size / 2};
+  blocked_matrix_view result_1_2_b{data_1_2_b.get(), size / 2};
+  blocked_matrix_view result_2_1_a{data_2_1_a.get(), size / 2};
+  blocked_matrix_view result_2_1_b{data_2_1_b.get(), size / 2};
+  blocked_matrix_view result_2_2_a{data_2_2_a.get(), size / 2};
+  blocked_matrix_view result_2_2_b{data_2_2_b.get(), size / 2};
+
+  blocked_matrix_view a_1_1 = a.quadrant_1_1();
+  blocked_matrix_view a_1_2 = a.quadrant_1_2();
+  blocked_matrix_view a_2_1 = a.quadrant_2_1();
+  blocked_matrix_view a_2_2 = a.quadrant_2_2();
+
+  blocked_matrix_view b_1_1 = b.quadrant_1_1();
+  blocked_matrix_view b_1_2 = b.quadrant_1_2();
+  blocked_matrix_view b_2_1 = b.quadrant_2_1();
+  blocked_matrix_view b_2_2 = b.quadrant_2_2();
+
+  // Divide Work Into Sub-Calls
+  multiply_div_conquer(size / 2, result_1_1_a, a_1_1, b_1_1);
+  multiply_div_conquer(size / 2, result_1_1_b, a_1_2, b_2_1);
+
+  multiply_div_conquer(size / 2, result_1_2_a, a_1_1, b_1_2);
+  multiply_div_conquer(size / 2, result_1_2_b, a_1_2, b_2_2);
+
+  multiply_div_conquer(size / 2, result_2_1_a, a_2_1, b_1_1);
+  multiply_div_conquer(size / 2, result_2_1_b, a_2_2, b_2_1);
+
+  multiply_div_conquer(size / 2, result_2_2_a, a_2_1, b_1_2);
+  multiply_div_conquer(size / 2, result_2_2_b, a_2_2, b_2_2);
+
+  // Combine results
+  for (size_t row = 0; row < size / 2; row++) {
+    for (size_t column = 0; column < size / 2; column++) {
+      result_1_1.at(row, column) = result_1_1_a.at(row, column) + result_1_1_b.at(row, column);
+      result_1_2.at(row, column) = result_1_2_a.at(row, column) + result_1_2_b.at(row, column);
+      result_2_1.at(row, column) = result_2_1_a.at(row, column) + result_2_1_b.at(row, column);
+      result_2_2.at(row, column) = result_2_2_a.at(row, column) + result_2_2_b.at(row, column);
+    }
+  }
+}
+
+constexpr int MAX_NUM_TASKS = 32;
+constexpr int MAX_STACK_SIZE = 4096 * 1;
+
+int main(int argc, char **argv) {
+  fill_block_lookup(MAX_BLOCK_LOOKUP);
+
+  size_t size = 64;
+  std::unique_ptr<double[]> result_data_naive{new double[size * size]};
+  std::unique_ptr<double[]> result_data_div{new double[size * size]};
+  std::unique_ptr<double[]> a_data{new double[size * size]};
+  std::unique_ptr<double[]> b_data{new double[size * size]};
+
+  blocked_matrix_view result_naive{result_data_naive.get(), size};
+  blocked_matrix_view result_div{result_data_div.get(), size};
+  blocked_matrix_view a{a_data.get(), size};
+  blocked_matrix_view b{b_data.get(), size};
+
+  for (size_t row = 0; row < size; row++) {
+    for (size_t column = 0; column < size; column++) {
+      a.at(row, column) = row;
+      b.at(row, column) = column;
+    }
+  }
+
+  multiply_div_conquer(size, result_div, a, b);
+  multiply_naive(size, result_naive, a, b);
+
+  size_t misses = 0;
+  for (size_t row = 0; row < size; row++) {
+    for (size_t column = 0; column < size; column++) {
+      if (result_div.at(row, column) != result_naive.at(row, column)) {
+        misses++;
+        printf("%5.5f\t\t", result_div.at(row, column) - result_naive.at(row, column));
+      }
+    }
+  }
+  printf("\n%d", misses);
+
+//  int num_threads;
+//  string directory;
+//  benchmark_runner::read_args(argc, argv, num_threads, directory);
+//
+//  string test_name = to_string(num_threads) + ".csv";
+//  string full_directory = directory + "/PLS_v3/";
+//  benchmark_runner runner{full_directory, test_name};
+//
+//  scheduler scheduler{(unsigned) num_threads, MAX_NUM_TASKS, MAX_STACK_SIZE};
+//
+//  runner.run_iterations(1000, [&]() {
+//    scheduler.perform_work([&]() {
+//    });
+//  }, 100);
+//  runner.commit_results(true);
+}

lib/pls/CMakeLists.txt

@@ -48,7 +48,7 @@ add_library(pls STATIC
 
 
         include/pls/internal/profiling/dag_node.h src/internal/profiling/dag_node.cpp
         include/pls/internal/profiling/dag_node.h src/internal/profiling/dag_node.cpp
         include/pls/internal/profiling/profiler.h src/internal/profiling/profiler.cpp
         include/pls/internal/profiling/profiler.h src/internal/profiling/profiler.cpp
-        include/pls/internal/profiling/thread_stats.h src/internal/profiling/thread_stats.cpp)
+        include/pls/internal/profiling/thread_stats.h src/internal/profiling/thread_stats.cpp include/pls/algorithms/divide_and_conquer_buffers.h)
 
 
 # Dependencies for pls
 # Dependencies for pls
 target_link_libraries(pls Threads::Threads)
 target_link_libraries(pls Threads::Threads)

lib/pls/include/pls/algorithms/divide_and_conquer_buffers.h

+
+#ifndef PLS_ALGORITHMS_DIVIDE_AND_CONQUER_BUFFERS_H_
+#define PLS_ALGORITHMS_DIVIDE_AND_CONQUER_BUFFERS_H_
+
+#endif //PLS_ALGORITHMS_DIVIDE_AND_CONQUER_BUFFERS_H_

lib/pls/include/pls/internal/base/system_details.h

@@ -47,7 +47,7 @@ constexpr size_t CACHE_LINE_SIZE = 64;
 /**
 /**
  * Helper to align types/values on cache lines.
  * Helper to align types/values on cache lines.
  */
  */
-#define PLS_CACHE_ALIGN alignas(base::system_details::CACHE_LINE_SIZE)
+#define PLS_CACHE_ALIGN alignas(::pls::internal::base::system_details::CACHE_LINE_SIZE)
 
 
 /**
 /**
  * Helper to find mmap page size. Either set constant or rely on system specific getter function.
  * Helper to find mmap page size. Either set constant or rely on system specific getter function.

lib/pls/include/pls/internal/scheduling/strain_local_resource.h

@@ -24,7 +24,7 @@ struct PLS_CACHE_ALIGN strain_resource {
   unsigned const index_;
   unsigned const index_;
   unsigned const depth_;
   unsigned const depth_;
 
 
-  bool used_{false};
+  std::atomic<int> used_{0};
   std::atomic<strain_resource *> next_{nullptr};
   std::atomic<strain_resource *> next_{nullptr};
 };
 };
 
 
@@ -48,7 +48,7 @@ class strain_local_resource {
     item_handle(item_handle &&) = delete;
     item_handle(item_handle &&) = delete;
     item_handle &operator=(item_handle &&) = delete;
     item_handle &operator=(item_handle &&) = delete;
 
 
-    explicit item_handle(strain_resource *resource) : resource_{resource} {}
+    explicit item_handle(strain_resource *resource);
     ~item_handle();
     ~item_handle();
 
 
     unsigned get_strain_index() {
     unsigned get_strain_index() {
@@ -56,7 +56,7 @@ class strain_local_resource {
     }
     }
 
 
    private:
    private:
-    strain_resource *resource_;
+    strain_resource *const resource_;
   };
   };
 
 
   strain_local_resource(unsigned num_threads,
   strain_local_resource(unsigned num_threads,

lib/pls/src/internal/scheduling/strain_local_resource.cpp

@@ -23,11 +23,14 @@ strain_local_resource::item_handle strain_local_resource::get_item(unsigned dept
   active_task->attached_resources_.store(result, std::memory_order_relaxed);
   active_task->attached_resources_.store(result, std::memory_order_relaxed);
 
 
   // Wrap it for RAII usage on stack
   // Wrap it for RAII usage on stack
-  PLS_ASSERT(!result->used_, "Must not try to allocate an already used resource!");
-  result->used_ = true;
   return strain_local_resource::item_handle{result};
   return strain_local_resource::item_handle{result};
 }
 }
 
 
+strain_local_resource::item_handle::item_handle(strain_resource *resource) : resource_{resource} {
+  PLS_ASSERT(resource_->used_.fetch_add(1, std::memory_order_relaxed) == 0,
+             "Must not create a handle of a already used resource!");
+}
+
 // Return item to locally owned items
 // Return item to locally owned items
 strain_local_resource::item_handle::~item_handle() {
 strain_local_resource::item_handle::~item_handle() {
   // Only change our resource usage when synced.
   // Only change our resource usage when synced.
@@ -54,8 +57,8 @@ strain_local_resource::item_handle::~item_handle() {
   // Give the resource handle back to our local resource array
   // Give the resource handle back to our local resource array
   auto &local_resource = resource_->strain_local_resource_->local_items_[my_state.get_thread_id()][resource_->depth_];
   auto &local_resource = resource_->strain_local_resource_->local_items_[my_state.get_thread_id()][resource_->depth_];
   local_resource.resource_ = resource_;
   local_resource.resource_ = resource_;
-  PLS_ASSERT(resource_->used_, "Must only release used resources!");
-  resource_->used_ = false;
+  PLS_ASSERT(resource_->used_.fetch_sub(1, std::memory_order_relaxed) == 1,
+             "Accidentally freed resource that was accessed multiple times!");
 }
 }
 
 
 strain_resource *strain_local_resource::get_local_copy(strain_resource *other_resources, unsigned thread_id) {
 strain_resource *strain_local_resource::get_local_copy(strain_resource *other_resources, unsigned thread_id) {