28ms -> 8ms. Rendering board directly to a texture, instead of using …

…SDL pixel functions (#1) - moved from a 500x500 to 2560x1440 grid for more accurate scores - remove the SDL pixel position calculations and instead just map the memory address to RGB
Jumbub · Oct 2, 2021 · cda613a · cda613a
1 parent d55a8fe
commit cda613a
Show file tree

Hide file tree

Showing 9 changed files with 58 additions and 49 deletions.
diff --git a/results/benchmark.txt b/results/benchmark.txt
@@ -1,5 +1,5 @@
 ----------------------------------------------------------------------------
 Benchmark                                  Time             CPU   Iterations
 ----------------------------------------------------------------------------
-BM_NextBoard/min_time:5.000             23.1 ms         23.1 ms          301
-BM_RenderNextBoard/min_time:5.000       25.3 ms         25.2 ms          275
+BM_NextBoard/min_time:5.000              356 ms          356 ms           20
+BM_RenderNextBoard/min_time:5.000        362 ms          362 ms           19
diff --git a/src/board/board.h b/src/board/board.h
@@ -4,6 +4,11 @@
 #include <tuple>
 #include <array>
 #include <memory>
+#include <SDL2/SDL.h>
 
-using Board = std::tuple<std::shared_ptr<bool[]>, int, int>;
+using Cell = Uint32;
+const Cell ALIVE = SDL_MAX_UINT32;
+const Cell DEAD = 0;
+
+using Board = std::tuple<std::shared_ptr<Cell[]>, int, int>;
 using NeighbourPositions = std::array<int, 8>;
diff --git a/src/board/generate.cpp b/src/board/generate.cpp
@@ -7,9 +7,9 @@
 #include <tuple>
 
 Board randomBoard(int width, int height) {
-  auto board = std::shared_ptr<bool[]>(new bool[width * height]);
+  auto board = std::shared_ptr<Cell[]>(new Cell[width * height]);
   for (int i = 0; i < height * width; ++i)
-    board[i] = rand() % 2;
+    board[i] = rand() % 2 ? ALIVE : DEAD;
   return {board, width, height};
 }
 
@@ -25,22 +25,22 @@ Board benchmarkBoard(int width, int height) {
         "Did not meet minimum height required for the benchmark board");
 
   srand(0);
-  auto board = std::shared_ptr<bool[]>(new bool[width * height]);
+  auto board = std::shared_ptr<Cell[]>(new Cell[width * height]);
   for (int y = 0; y < height; ++y) {
     for (int x = 0; x < width; ++x) {
       const int i = y * width + x;
       if (y > height / 2) {
         if (x < width / 2)
-          board[i] = rand() % 2;
+          board[i] = rand() % 2 ? ALIVE : DEAD;
         if (x >= width / 2)
-          board[i] = (x / 8) % 2 != (y / 8) % 2;
+          board[i] = (x / 8) % 2 != (y / 8) % 2 ? ALIVE : DEAD;
       } else {
         const int breederMarginY = (height / 2 - BREEDER_HEIGHT) / 2;
         const int breederY = y - breederMarginY;
         if (breederY > 0 && breederY < BREEDER_HEIGHT && x < BREEDER_WIDTH) {
-          board[i] = BREEDER[breederY][x];
+          board[i] = BREEDER[breederY][x] ? ALIVE : DEAD;
         } else {
-          board[i] = false;
+          board[i] = DEAD;
         }
       }
     }

diff --git a/src/board/next.cpp b/src/board/next.cpp
@@ -13,13 +13,13 @@ Board nextBoard(Board board) {
   if (width == 0)
     return board;
 
-  auto output = std::shared_ptr<bool[]>(new bool[width*height]);
+  auto output = std::shared_ptr<Cell[]>(new Cell[width*height]);
 
   for (int y = 0; y < height; y++) {
     for (int x = 0; x < width; x++) {
       const int i = y * width + x;
 
-      auto alive = input[i];
+      auto state = input[i];
 
       int neighboursCount = 0;
       for (int y2 = -1; y2 <= 1; y2++)
@@ -43,12 +43,12 @@ Board nextBoard(Board board) {
           }
         }
 
-      if (alive && (neighboursCount < 2 || neighboursCount > 3))
-        output[i] = false;
-      else if (!alive && neighboursCount == 3)
-        output[i] = true;
+      if (state && (neighboursCount < 2 || neighboursCount > 3))
+        output[i] = DEAD;
+      else if (!state && neighboursCount == 3)
+        output[i] = ALIVE;
       else
-        output[i] = alive;
+        output[i] = state;
     }
   }
 

diff --git a/src/board/sdl.cpp b/src/board/sdl.cpp
@@ -1,4 +1,5 @@
 #include "sdl.h"
+#include "../util/profile.h"
 #include "generate.h"
 #include <vector>
 
@@ -8,22 +9,11 @@ Board boardForSdlWindow(SDL_Window *window) {
   return benchmarkBoard(width, height);
 }
 
-void renderBoardSdl(Board board, SDL_Renderer *renderer) {
+void renderBoardSdl(Board board, SDL_Renderer *renderer, SDL_Texture *texture) {
   const auto &[input, width, height] = board;
 
-  auto points = std::unique_ptr<SDL_Point[]>(new SDL_Point[width * height]);
-  int count = 0;
-  for (int i = 0; i < width * height; i++)
-    if (input[i]) {
-      points[count] = {i % width, i / width};
-      count++;
-    }
-
-  SDL_SetRenderDrawColor(renderer, 0, 0, 0, 255);
-  SDL_RenderClear(renderer);
-
-  SDL_SetRenderDrawColor(renderer, 255, 255, 255, 255);
-  SDL_RenderDrawPoints(renderer, &points[0], count);
+  SDL_UpdateTexture(texture, NULL, &input[0], width * sizeof(Uint32));
 
+  SDL_RenderCopy(renderer, texture, nullptr, nullptr);
   SDL_RenderPresent(renderer);
 }
diff --git a/src/board/sdl.h b/src/board/sdl.h
@@ -5,4 +5,4 @@
 
 Board boardForSdlWindow(SDL_Window *window);
 
-void renderBoardSdl(Board board, SDL_Renderer *renderer);
+void renderBoardSdl(Board board, SDL_Renderer *renderer, SDL_Texture *texture);
diff --git a/src/entrypoints/benchmark.cpp b/src/entrypoints/benchmark.cpp
@@ -4,35 +4,35 @@
 #include <SDL2/SDL.h>
 #include <benchmark/benchmark.h>
 
-static void BM_NextBoard(benchmark::State &state) {
-  const int width = 500;
-  const int height = 500;
+const int TEST_WIDTH = 2560;
+const int TEST_HEIGHT = 1440;
 
-  auto board = benchmarkBoard(width, height);
+static void BM_NextBoard(benchmark::State &state) {
+  auto board = benchmarkBoard(TEST_WIDTH, TEST_HEIGHT);
 
   for (auto _ : state) {
     board = nextBoard(board);
   }
 }
 
 static void BM_RenderNextBoard(benchmark::State &state) {
-  const int width = 500;
-  const int height = 500;
-
-  auto board = benchmarkBoard(width, height);
+  auto board = benchmarkBoard(TEST_WIDTH, TEST_HEIGHT);
 
   // Initialize graphics
   SDL_Init(SDL_INIT_VIDEO);
   SDL_Renderer *renderer;
   SDL_Window *window;
-  SDL_CreateWindowAndRenderer(width, height, 0, &window, &renderer);
+  SDL_CreateWindowAndRenderer(TEST_WIDTH, TEST_HEIGHT, 0, &window, &renderer);
+  auto texture = SDL_CreateTexture(renderer, SDL_PIXELFORMAT_ARGB8888,
+                                    SDL_TEXTUREACCESS_STATIC, TEST_WIDTH, TEST_HEIGHT);
 
   for (auto _ : state) {
     board = nextBoard(board);
-    renderBoardSdl(board, renderer);
+    renderBoardSdl(board, renderer, texture);
   }
 
   // Destroy graphics
+  SDL_DestroyTexture(texture);
   SDL_DestroyRenderer(renderer);
   SDL_DestroyWindow(window);
   SDL_Quit();

diff --git a/src/entrypoints/interactive.cpp b/src/entrypoints/interactive.cpp
@@ -17,6 +17,12 @@ int main() {
   SDL_CreateWindowAndRenderer(500, 500, SDL_WINDOW_RESIZABLE, &window,
                               &renderer);
 
+  // Window texture
+  int width, height;
+  SDL_GetWindowSize(window, &width, &height);
+  auto texture = SDL_CreateTexture(renderer, SDL_PIXELFORMAT_ARGB8888,
+                                   SDL_TEXTUREACCESS_STATIC, width, height);
+
   // Generate initial board
   auto p1 = startProfiling();
   auto board = boardForSdlWindow(window);
@@ -33,19 +39,27 @@ int main() {
         running = false;
       // Re-create board when Enter is pressed
       else if (event.type == SDL_KEYDOWN &&
-               event.key.keysym.scancode == SDL_SCANCODE_RETURN)
+               event.key.keysym.scancode == SDL_SCANCODE_RETURN) {
+
         board = boardForSdlWindow(window);
+        int width, height;
+        SDL_GetWindowSize(window, &width, &height);
+        SDL_DestroyTexture(texture);
+        texture = SDL_CreateTexture(renderer, SDL_PIXELFORMAT_ARGB8888,
+                                    SDL_TEXTUREACCESS_STATIC, width, height);
+      }
     }
 
     auto p2 = startProfiling();
     board = nextBoard(board);
     stopProfiling(p2, "Calculated next board");
 
     auto p3 = startProfiling();
-    renderBoardSdl(board, renderer);
+    renderBoardSdl(board, renderer, texture);
     stopProfiling(p3, "Rendered next board");
   }
 
+  SDL_DestroyTexture(texture);
   SDL_DestroyRenderer(renderer);
   SDL_DestroyWindow(window);
   SDL_Quit();

diff --git a/src/entrypoints/test.cpp b/src/entrypoints/test.cpp
@@ -3,22 +3,22 @@
 #include "../board/next.h"
 #include <catch2/catch.hpp>
 
-Board generate(std::vector<std::vector<bool>> input) {
+Board generate(std::vector<std::vector<Cell>> input) {
   const auto height = (int)input.size();
   const auto width = (int)input[0].size();
 
-  auto board = std::shared_ptr<bool[]>(new bool[width * height]);
+  auto board = std::shared_ptr<Cell[]>(new Cell[width * height]);
   for (int y = 0; y < height; ++y)
     for (int x = 0; x < width; ++x)
-      board[y * width + x] = input[y][x];
+      board[y * width + x] = input[y][x] ? ALIVE : DEAD;
 
   return {board, width, height};
 }
 
-std::vector<std::vector<bool>> ungenerate(Board board) {
+std::vector<std::vector<Cell>> ungenerate(Board board) {
   const auto &[input, width, height] = board;
 
-  std::vector<std::vector<bool>> output(height, std::vector<bool>(width));
+  std::vector<std::vector<Cell>> output(height, std::vector<Cell>(width));
   for (int y = 0; y < height; ++y)
     for (int x = 0; x < width; ++x)
       output[y][x] = input[y * width + x];
Original file line number	Diff line number	Diff line change
Expand Up		@@ -5,4 +5,4 @@

		Board boardForSdlWindow(SDL_Window *window);

		void renderBoardSdl(Board board, SDL_Renderer *renderer);
		void renderBoardSdl(Board board, SDL_Renderer renderer, SDL_Texture texture);