add function to generate list of random numbers with cuda and opencl on the device

src/OpenCL/OpenCLBase.cpp

@@ -7,21 +7,13 @@ cl_device_id OpenCLBase::m_device_id = NULL;
 cl_event OpenCLBase::m_last_event = NULL;
 
 OpenCLBase::OpenCLBase() {
-  //m_context = NULL;
-  //m_command_queue = NULL;
   m_program = NULL;
   m_kernel = NULL;
-  //m_device_id = NULL;
-  //m_platform_id = NULL;
   m_kernel_file = NULL;
 	
   m_last_event = NULL;
 	
-  //m_events = new cl_event[500];
-  //m_num_events = 0;
-
   defaultRndSet = 0;
-
 }
 
 OpenCLBase::~OpenCLBase() {
@@ -41,11 +33,11 @@ int OpenCLBase::ocl_createRndStates(int size) {
   strcat(kernel_file, "OpenCL/OpenCLKernels/OpenCLCollimatorPhysics.cl");
   ocl_loadKernel(kernel_file);
   delete[] kernel_file;
-
+  
   //allocate memory for rand states
   int ierr;
   defaultRndState = ocl_allocateMemory(sizeof(RNDState)*size, ierr);
-
+  
   //exec kernel
   int seed = 0;
   ocl_createKernel("initRand");
@@ -55,13 +47,34 @@ int OpenCLBase::ocl_createRndStates(int size) {
   
   size_t work_items = size;
   size_t work_group_size = 1;
-
   ocl_executeKernel(1, &work_items, &work_group_size);
-
   defaultRndSet = 1;
+  
+  return DKS_SUCCESS;
+}
 
-  return OCL_SUCCESS;
+int OpenCLBase::ocl_createRandomNumbers(void *mem_ptr, int size) {
+  //load kernel
+  char * kernel_file = new char[500];
+  kernel_file[0] = '\0';
+  strcat(kernel_file, OPENCL_KERNELS);
+  strcat(kernel_file, "OpenCL/OpenCLKernels/OpenCLCollimatorPhysics.cl");
+  ocl_loadKernel(kernel_file);
+  delete[] kernel_file;
 
+  //set kernel variables
+  cl_mem tmp_data = (cl_mem) mem_ptr;
+
+  ocl_createKernel("createRandoms");
+  ocl_setKernelArg(0, sizeof(cl_mem), &defaultRndState);
+  ocl_setKernelArg(1, sizeof(cl_mem), &tmp_data);
+  ocl_setKernelArg(2, sizeof(int), &size);
+
+  size_t work_size = 128;
+  size_t work_items = (size % work_size + 1) * work_size;
+  ocl_executeKernel(1, &work_items, &work_size);
+
+  return DKS_SUCCESS;
 }
 
 /* destroy rnd states */
@@ -70,7 +83,7 @@ int OpenCLBase::ocl_deleteRndStates() {
   ocl_freeMemory(defaultRndState);
   defaultRndSet = 0;
 
-  return OCL_SUCCESS;
+  return DKS_SUCCESS;
 
 }
 

src/OpenCL/OpenCLBase.h

@@ -30,11 +30,20 @@
 #include <CL/cl_ext.h>
 #endif
 
-#include "clRNG/clRNG.h"
-#include "clRNG/mrg31k3p.h"
-
 #include "../DKSDefinitions.h"
 
+/* struct for random number state */
+typedef struct {
+  double s10;
+  double s11;
+  double s12;
+  double s20;
+  double s21;
+  double s22;
+  double z;
+  bool gen;
+} RNDState;
+
 class OpenCLBase {
 
 private:
@@ -122,6 +131,11 @@ public:
   */
   int ocl_createRndStates(int size);
 
+  /* Create an array of random numbers on the device
+   *
+   */
+  int ocl_createRandomNumbers(void *mem_ptr, int size);
+
   /*
     Destroy rnd states
     Return: success or error code

src/OpenCL/OpenCLKernels/OpenCLCollimatorPhysics.cl

@@ -1,6 +1,4 @@
 #pragma OPENCL EXTENSION cl_khr_fp64 : enable
-#pragma OPENCL EXTENSION 
-
 
 /******Random numbers********/
 
@@ -89,13 +87,14 @@ __kernel void initRand(__global RNDState *s, unsigned int seed, int N) {
 
   if (id < N) {
     RNDState tmp;
-    int tmp_seed = id;// * 0x100000000ULL;
+    int tmp_seed = 2*id;// * 0x100000000ULL;
     tmp.s10 = 12345 + tmp_seed;
     tmp.s11 = 12345 + tmp_seed;
-    tmp.s12 = 123 + tmp_seed;
+    tmp.s12 = 12345 + tmp_seed;
     tmp.s20 = 12345 + tmp_seed;
     tmp.s21 = 12345 + tmp_seed;
-    tmp.s22 = 123 + tmp_seed;
+    tmp.s22 = 12345 + tmp_seed;
+
 
     tmp.z = 0;
     tmp.gen = true;
@@ -105,6 +104,19 @@ __kernel void initRand(__global RNDState *s, unsigned int seed, int N) {
 
 }
 
+/* create random numbers and fill an array */
+__kernel void createRandoms(__global RNDState *states, __global double *data, int size) {
+
+  int idx = get_global_id(0);
+
+  if (idx < size) {
+    RNDState s = states[idx];
+    data[idx] = rand_uniform(&s);
+    states[idx] = s;
+  }
+
+}
+
 
 /**********Degrader**********/
 enum PARAMS { POSITION,