changed version to 1.1.4

For the case that map or fun is not used in the msr-file, the corresponding
call to the setKernelArg still needs a 2nd argument > 0, otherwise macOS crashes.

CMakeLists.txt

@@ -2,7 +2,7 @@ CMAKE_MINIMUM_REQUIRED (VERSION 3.2)
 PROJECT (DKS)
 SET (DKS_VERSION_MAJOR 1)
 SET (DKS_VERSION_MINOR 1)
-SET (DKS_VERSION_PATCH 2)
+SET (DKS_VERSION_PATCH 4)
 set (DKS_VERSION ${DKS_VERSION_MAJOR}.${DKS_VERSION_MINOR}.${DKS_VERSION_PATCH})
 SET (PACKAGE \"dks\")
 SET (PACKAGE_BUGREPORT \"locans.uldis@psi.ch\")

ReadMe.first

@@ -1,7 +1,7 @@
 ##################################################################
 #
 # Name:		Dynamic Kernel Scheduler
-# Version:	1.0
+# Version:	1.1
 # Author: 	Uldis Locans
 # Contacts:	locans.uldis@psi.ch
 #

src/CUDA/CudaChiSquareRuntime.cu

@@ -87,14 +87,18 @@ int CudaChiSquareRuntime::compileProgram(std::string function, bool mlh) {
   //create program
   nvrtcProgram prog;
 //  std::cout << cudaProg.c_str() << std::endl;
-  nvrtcCreateProgram(&prog, cudaProg.c_str(), "chiSquareRuntime.cu", 0, NULL, NULL);
+  nvrtcResult createResult = nvrtcCreateProgram(&prog, cudaProg.c_str(), "chiSquareRuntime.cu", 0, NULL, NULL);
+  if (createResult != NVRTC_SUCCESS) {
+    DEBUG_MSG("Program creation failed!");
+    return DKS_ERROR;
+  }
 
   //compile program
-  const char *opts[] = {"-fmad=false", ""};
-  int numopts = 1;
+  const char *opts[] = {"-arch=compute_35", "-fmad=false", ""};
+  int numopts = 2;
   if (mlh) {
-    opts[1] = "-DMLH";
-    numopts = 2;
+    opts[2] = "-DMLH";
+    numopts = 3;
   }
 
   nvrtcResult compileResults = nvrtcCompileProgram(prog, numopts, opts);
@@ -118,7 +122,11 @@ int CudaChiSquareRuntime::compileProgram(std::string function, bool mlh) {
   if (ptx_m != NULL)
     delete[] ptx_m;
   size_t ptxSize; 
-  nvrtcGetPTXSize(prog, &ptxSize); 
+  nvrtcResult ptxSizeResult = nvrtcGetPTXSize(prog, &ptxSize);
+  if (ptxSizeResult != NVRTC_SUCCESS) {
+    DEBUG_MSG("PTX get size error!");
+    return DKS_ERROR;
+  }
   ptx_m = new char[ptxSize]; 
   nvrtcResult nvrtcPTXResult = nvrtcGetPTX(prog, ptx_m);  
   
@@ -127,10 +135,26 @@ int CudaChiSquareRuntime::compileProgram(std::string function, bool mlh) {
     return DKS_ERROR;
   }
 
+  // add some additional diagnostics
+  const int buffer_size = 8192;
+  CUjit_option options[3];
+  void* values[3];
+  char error_log[buffer_size];
+  int err;
+  options[0] = CU_JIT_ERROR_LOG_BUFFER;
+  values[0]  = (void*)error_log;
+  options[1] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
+  values[1]  = (void*)buffer_size;
+  options[2] = CU_JIT_TARGET_FROM_CUCONTEXT;
+  values[2]  = 0;
   //load module from ptx
-  CUresult loadResult = cuModuleLoadDataEx(&module_m, ptx_m, 0, 0, 0); 
+  CUresult loadResult = cuModuleLoadDataEx(&module_m, ptx_m, 3, options, values); 
   if (loadResult != CUDA_SUCCESS) {
-    DEBUG_MSG("Load module from ptx failed!");
+    const char *err_msg;
+    cuGetErrorString(loadResult, &err_msg);
+    std::string msg = "Load module from ptx failed! (" + std::to_string(loadResult) + ") : " + err_msg;
+    DEBUG_MSG(msg);
+    DEBUG_MSG(error_log);
     return DKS_ERROR;
   }
 

src/CUDA/NVRTCKernels/CudaChiSquareKernel.cu

@@ -83,6 +83,56 @@ __device__ double ifld(double t, double alpha, double phi, double nu, double lam
   return alpha*cos(wt+ph)*exp(-lambdaT*t) + (1.0-alpha)*exp(-lambdaL*t);
 }
 
+__device__ double ifgk(double t, double alpha, double nu, double sigma, double lambda, double beta) {
+  double wt = TWO_PI*nu*t;
+  double rate2 = sigma*sigma*t*t;
+  double rateL = 0.0;
+  double result = 0.0;
+
+  // make sure lambda > 0
+  if (lambda < 0.0)
+    return 0.0;
+
+  if (beta < 0.001) {
+    rateL = 1.0;
+  } else {
+    rateL = pow(lambda*t, beta);
+  }
+
+  if (nu < 0.01) {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(1.0-rate2)*exp(-0.5*rate2);
+  } else {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(cos(wt)-sigma*sigma*t*t/(wt)*sin(wt))*exp(-0.5*rate2);
+  }
+
+  return result;
+}
+
+__device__ double ifll(double t, double alpha, double nu, double a, double lambda, double beta) {
+  double wt = TWO_PI*nu*t;
+  double at = a*t;
+  double rateL = 0.0;
+  double result = 0.0;
+
+  // make sure lambda > 0
+  if (lambda < 0.0)
+    return 0.0;
+
+  if (beta < 0.001) {
+    rateL = 1.0;
+  } else {
+    rateL = pow(lambda*t, beta);
+  }
+
+  if (nu < 0.01) {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(1.0-at)*exp(-at);
+  } else {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(cos(wt)-a/(TWO_PI*nu)*sin(wt))*exp(-at);
+  }
+
+  return result;
+}
+
 __device__ double b(double t, double phi, double nu) {
   return j0(TWO_PI*nu*t + DEG_TO_RAD*phi);
 }

src/OpenCL/OpenCLChiSquareRuntime.cpp

@@ -76,7 +76,6 @@ int OpenCLChiSquareRuntime::compileProgram(std::string function, bool mlh) {
 
 double OpenCLChiSquareRuntime::calculateSum(cl_mem data, int length) {
 
-  
   int ierr;
   //calc number of threads per workgroup and nr of work groups
   size_t work_size_sum = (size_t)blockSize_m;
@@ -105,7 +104,7 @@ double OpenCLChiSquareRuntime::calculateSum(cl_mem data, int length) {
   m_oclbase->ocl_setKernelArg(3, sizeof(int), &length);
   m_oclbase->ocl_executeKernel(1, &work_items, &work_size_sum);
   
-  //read partial sums and free temp mempry
+  //read partial sums and free temp memory
   m_oclbase->ocl_readData(tmp_ptr, partial_sums, sizeof(double)*work_groups);
   m_oclbase->ocl_freeMemory(tmp_ptr);
 
@@ -157,6 +156,7 @@ int OpenCLChiSquareRuntime::launchChiSquare(int fitType,
       return ierr;
 
     //set kernel args
+    size_t num=1;
     m_oclbase->ocl_setKernelArg(0, sizeof(cl_mem), &cl_mem_data);
     m_oclbase->ocl_setKernelArg(1, sizeof(cl_mem), &cl_mem_err);
     m_oclbase->ocl_setKernelArg(2, sizeof(cl_mem), &cl_param);
@@ -172,20 +172,23 @@ int OpenCLChiSquareRuntime::launchChiSquare(int fitType,
     m_oclbase->ocl_setKernelArg(12, sizeof(double), &tau_m);
     m_oclbase->ocl_setKernelArg(13, sizeof(double), &N0_m);
     m_oclbase->ocl_setKernelArg(14, sizeof(double), &bkg_m);
-    m_oclbase->ocl_setKernelArg(15, sizeof(double)*numpar, NULL);
-    m_oclbase->ocl_setKernelArg(16, sizeof(double)*numfunc, NULL);
-    m_oclbase->ocl_setKernelArg(17, sizeof(int)*nummap, NULL);
+    num = numpar; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(15, sizeof(double)*num, NULL);
+    num = numfunc; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(16, sizeof(double)*num, NULL);
+    num = nummap; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(17, sizeof(int)*num, NULL);
 
     if (ierr != DKS_SUCCESS)
       return ierr;
   } else if (fitType == FITTYPE_ASYMMETRY) {
     //create kernel
     ierr = m_oclbase->ocl_createKernel("kernelChiSquareAsymmetry");
-
     if (ierr != DKS_SUCCESS)
       return ierr;
 
     //set kernel args
+    size_t num=1;
     m_oclbase->ocl_setKernelArg(0, sizeof(cl_mem), &cl_mem_data);
     m_oclbase->ocl_setKernelArg(1, sizeof(cl_mem), &cl_mem_err);
     m_oclbase->ocl_setKernelArg(2, sizeof(cl_mem), &cl_param);
@@ -200,9 +203,12 @@ int OpenCLChiSquareRuntime::launchChiSquare(int fitType,
     m_oclbase->ocl_setKernelArg(11, sizeof(double), &timeStep);
     m_oclbase->ocl_setKernelArg(12, sizeof(double), &alpha_m);
     m_oclbase->ocl_setKernelArg(13, sizeof(double), &beta_m);
-    m_oclbase->ocl_setKernelArg(14, sizeof(double)*numpar, NULL);
-    m_oclbase->ocl_setKernelArg(15, sizeof(double)*numfunc, NULL);
-    m_oclbase->ocl_setKernelArg(16, sizeof(int)*nummap, NULL);
+    num = numpar; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(14, sizeof(double)*num, NULL);
+    num = numfunc; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(15, sizeof(double)*num, NULL);
+    num = nummap; if (num == 0) num = 1;
+    m_oclbase->ocl_setKernelArg(16, sizeof(int)*num, NULL);
 
     if (ierr != DKS_SUCCESS)
       return ierr;
@@ -321,4 +327,3 @@ int OpenCLChiSquareRuntime::checkChiSquareKernels(int fitType, int &threadsPerBl
   return ierr;
 
 }
-

src/OpenCL/OpenCLKernels/OpenCLChiSquareRuntime.cl

@@ -106,6 +106,56 @@ double ifld(double t, double alpha, double phi, double nu, double lambdaT, doubl
   return alpha*cos(wt+ph)*exp(-lambdaT*t) + (1.0-alpha)*exp(-lambdaL*t);
 }
 
+double ifgk(double t, double alpha, double nu, double sigma, double lambda, double beta) {
+  double wt = TWO_PI*nu*t;
+  double rate2 = sigma*sigma*t*t;
+  double rateL = 0.0;
+  double result = 0.0;
+
+  // make sure lambda > 0
+  if (lambda < 0.0)
+    return 0.0;
+
+  if (beta < 0.001) {
+    rateL = 1.0;
+  } else {
+    rateL = pow(lambda*t, beta);
+  }
+
+  if (nu < 0.01) {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(1.0-rate2)*exp(-0.5*rate2);
+  } else {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(cos(wt)-sigma*sigma*t*t/(wt)*sin(wt))*exp(-0.5*rate2);
+  }
+
+  return result;
+}
+
+double ifll(double t, double alpha, double nu, double a, double lambda, double beta) {
+  double wt = TWO_PI*nu*t;
+  double at = a*t;
+  double rateL = 0.0;
+  double result = 0.0;
+
+  // make sure lambda > 0
+  if (lambda < 0.0)
+    return 0.0;
+
+  if (beta < 0.001) {
+    rateL = 1.0;
+  } else {
+    rateL = pow(lambda*t, beta);
+  }
+
+  if (nu < 0.01) {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(1.0-at)*exp(-at);
+  } else {
+    result = (1.0-alpha)*exp(-rateL) + alpha*(cos(wt)-a/(TWO_PI*nu)*sin(wt))*exp(-at);
+  }
+
+  return result;
+}
+
 double b(double t, double phi, double nu) {
   return bessj0(TWO_PI*nu*t + DEG_TO_RAD*phi);
 }