Index: MicroBenchmarks/ImageProcessing/CMakeLists.txt
===================================================================
--- MicroBenchmarks/ImageProcessing/CMakeLists.txt
+++ MicroBenchmarks/ImageProcessing/CMakeLists.txt
@@ -0,0 +1 @@
+add_subdirectory(Dilate)
Index: MicroBenchmarks/ImageProcessing/Dilate/CMakeLists.txt
===================================================================
--- MicroBenchmarks/ImageProcessing/Dilate/CMakeLists.txt
+++ MicroBenchmarks/ImageProcessing/Dilate/CMakeLists.txt
@@ -0,0 +1,11 @@
+set(IMAGEPROC_UTILS MicroBenchmarks/ImageProcessing/utils)
+list(APPEND CPPFLAGS -I ${CMAKE_SOURCE_DIR}/${IMAGEPROC_UTILS} -std=c++11)
+
+llvm_test_verify("${CMAKE_SOURCE_DIR}/HashProgramOutput.sh ${CMAKE_CURRENT_BINARY_DIR}/dilateOutput.txt")
+llvm_test_verify("${FPCMP} ${CMAKE_CURRENT_BINARY_DIR}/dilateOutput.txt ${CMAKE_CURRENT_SOURCE_DIR}/dilate.reference_output")
+
+llvm_test_run(WORKDIR ${CMAKE_CURRENT_BINARY_DIR})
+
+llvm_test_executable(Dilate ../utils/ImageHelper.cpp ../utils/glibc_compat_rand.cpp main.cpp dilateKernel.cpp)
+
+target_link_libraries(Dilate benchmark)
Index: MicroBenchmarks/ImageProcessing/Dilate/dilate.h
===================================================================
--- MicroBenchmarks/ImageProcessing/Dilate/dilate.h
+++ MicroBenchmarks/ImageProcessing/Dilate/dilate.h
@@ -0,0 +1,15 @@
+/**
+Pankaj Kukreja
+github.com/proton0001
+Indian Institute of Technology Hyderabad
+*/
+
+#ifndef _DILATE_H_
+#define _DILATE_H_
+
+#define HEIGHT 1024
+#define WIDTH 1024
+
+void dilateKernel(int *src, int *dst, int height, int width, int *temporary);
+
+#endif /* _DILATE_H_ */
Index: MicroBenchmarks/ImageProcessing/Dilate/dilate.reference_output
===================================================================
--- MicroBenchmarks/ImageProcessing/Dilate/dilate.reference_output
+++ MicroBenchmarks/ImageProcessing/Dilate/dilate.reference_output
@@ -0,0 +1 @@
+5f1e986b006f95a80a1bbf0d316fc41c
Index: MicroBenchmarks/ImageProcessing/Dilate/dilateKernel.cpp
===================================================================
--- MicroBenchmarks/ImageProcessing/Dilate/dilateKernel.cpp
+++ MicroBenchmarks/ImageProcessing/Dilate/dilateKernel.cpp
@@ -0,0 +1,52 @@
+/**
+  Source:
+  https://github.com/mompes/CUDA-dilation-and-erosion-filters/blob/master/erosionCPU.cpp
+  
+  Pankaj Kukreja
+  github.com/proton0001
+  Indian Institute of Technology Hyderabad
+*/
+#include "dilate.h"
+#define MAX(a, b) (a > b) ? a : b;
+#define MIN(a, b) (a < b) ? a : b;
+#include <cstdlib> // malloc, free
+
+// inpImage[HEIGHT * WIDTH]
+// outImage[height * width] 
+void dilateKernel(int *inpImage, int *outImage, int height, int width, int *temporary) {
+
+  typedef int inputMatrix[HEIGHT][WIDTH];
+  inputMatrix *inputImage;
+  inputImage = (inputMatrix *)inpImage;
+
+  typedef int outputMatrix[height][width];
+  outputMatrix *outputImage;
+  outputImage = (outputMatrix *)outImage;
+
+  outputMatrix *temp;
+  temp = (outputMatrix *)temporary;
+  
+  for (int i = 0; i < height; i++) {
+    for (int j = 0; j < width; j++) {
+      int value = 0;
+      for (int k=-1; k<=1; k++) {
+        if((j+k)>0 && (j+k)<width) {
+          value = MAX((*inputImage)[i][j+k], value);
+        }
+      }
+      (*temp)[i][j] = value;
+    }
+  }
+
+  for (int i = 0; i < height; i++) {
+    for (int j = 0; j < width; j++) {
+      int value = 0;
+      for (int k=-1; k<=1; k++) {
+        if((i+k)>0 && (i+k)<height) {
+          value = MAX((*temp)[i+k][j], value);
+        }
+      }
+      (*outputImage)[i][j] = value;
+    }
+  }
+}
Index: MicroBenchmarks/ImageProcessing/Dilate/main.cpp
===================================================================
--- MicroBenchmarks/ImageProcessing/Dilate/main.cpp
+++ MicroBenchmarks/ImageProcessing/Dilate/main.cpp
@@ -0,0 +1,97 @@
+/**
+  Pankaj Kukreja
+  github.com/proton0001
+  Indian Institute of Technology Hyderabad
+*/
+#include "ImageHelper.h"
+#include "dilate.h"
+#include <iostream>
+
+#define BENCHMARK_LIB
+#ifdef BENCHMARK_LIB
+#include "benchmark/benchmark.h"
+#endif
+
+int *inputImage;
+
+int main(int argc, char *argv[]) {
+#ifdef BENCHMARK_LIB
+  ::benchmark::Initialize(&argc, argv);
+#endif
+
+  const char *dilateOutputFileName = (const char *)"./dilateOutput.txt";
+  inputImage = (int *)malloc(sizeof(int) * (HEIGHT) * (WIDTH));
+
+  if (inputImage == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(EXIT_FAILURE);
+  }
+
+  initializeRandomImage(inputImage, HEIGHT, WIDTH);
+
+#ifdef BENCHMARK_LIB
+  ::benchmark::RunSpecifiedBenchmarks();
+#endif
+
+  int *outputImage;
+  outputImage = (int *)malloc(sizeof(int) * (HEIGHT) * (WIDTH));
+  int *temp;
+  temp = (int *)malloc(sizeof(int) * (HEIGHT) * (WIDTH));
+
+  if (outputImage == NULL || temp == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(EXIT_FAILURE);
+  }
+
+  dilateKernel(inputImage, outputImage, HEIGHT, WIDTH, temp);
+
+  for (int j = 0; j < WIDTH; j++) {
+    outputImage[0 * WIDTH + j] = 0;
+    outputImage[(HEIGHT - 1) * WIDTH + j] = 0;
+  }
+
+  for (int i = 0; i < HEIGHT; i++) {
+    outputImage[i * WIDTH + 0] = 0;
+    outputImage[i * WIDTH + (WIDTH - 1)] = 0;
+  }
+
+  saveImage(outputImage, dilateOutputFileName, HEIGHT, WIDTH);
+  free(temp);
+  free(outputImage);
+  free(inputImage);
+  return (EXIT_SUCCESS);
+}
+
+#ifdef BENCHMARK_LIB
+void BENCHMARK_DILATE(benchmark::State &state) {
+
+  int height = state.range(0);
+  int width = state.range(1);
+  int *outputImage;
+  outputImage = (int *)malloc(sizeof(int) * (height) * (width));
+  int *temp;
+  temp = (int *)malloc(sizeof(int) * (HEIGHT) * (WIDTH));
+  if (outputImage == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(EXIT_FAILURE);
+  }
+
+  dilateKernel(inputImage, outputImage, height, width, temp);
+
+  for (auto _ : state) {
+    dilateKernel(inputImage, outputImage, height, width, temp);
+  }
+
+  if (state.range(0) == 20) {
+    saveImage(outputImage, (const char *)"testFailed.txt", height, width);
+  }
+  free(temp);
+  free(outputImage);
+}
+BENCHMARK(BENCHMARK_DILATE)
+    ->Args({128, 128})
+    ->Args({256, 256})
+    ->Args({512, 512})
+    ->Args({1024, 1024})
+    ->Unit(benchmark::kMicrosecond);
+#endif