Updating miniapp (076c10d1) · Commits · Berrill, Mark A / RayTrace-miniapp

build/configure

+1 −0

Original line number	Diff line number	Diff line
		@@ -6,6 +6,7 @@ cmake \
		-D CMAKE_CXX_COMPILER=g++ \
		-D CXX_STD=11 \
		-D USE_OPENACC=0 \
		-D USE_OPENMP=1 \
		-D USE_KOKKOS=0 \
		-D KOKKOS_DIRECTORY=/usr/local/kokkos \
		-D KOKKOS_WRAPPER=/usr/local/kokkos/config/nvcc_wrapper \

src/CMakeLists.txt

+19 −2

Original line number	Diff line number	Diff line
		@@ -43,6 +43,23 @@ INCLUDE( Find_TIMER.cmake )
		CONFIGURE_TIMER( 0 "" )


		# Enable OpenMP
		IF ( USE_OPENMP )
		ADD_DEFINITIONS( -DUSE_OPENMP )
		IF ( USING_GCC )
		SET(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fopenmp")
		SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
		ELSEIF ( USING_CRAY )
		ELSEIF ( USING_PGCC )
		SET(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -openmp")
		SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -openmp")
		ELSE()
		MESSAGE(FATAL_ERROR "Compiling with OpenMP is not yet set for this compiler")
		ENDIF()
		MESSAGE( "Using OpenMP" )
		ENDIF()


		# Enable OpenACC
		IF ( USE_OPENACC )
		ADD_DEFINITIONS( -DUSE_OPENACC )
		@@ -96,7 +113,7 @@ IF ( USE_CUDA )
		ENDIF()
		IF(NOT CUDA_NVCC_FLAGS)
		# Set minimum requirements
		SET( CUDA_NVCC_FLAGS "-arch=sm_20 ${CXX_STD_FLAG}" )
		SET( CUDA_NVCC_FLAGS "-arch=sm_30 ${CXX_STD_FLAG}" )
		ENDIF()
		IF ( ${CMAKE_BUILD_TYPE} STREQUAL "Debug" )
		SET( CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -lineinfo" )
		@@ -138,7 +155,7 @@ ADD_DISTCLEAN( libRayTrace.* null_timer CreateImage* )
		# Create the library
		INCLUDE_DIRECTORIES( ${RAYTRACE_SOURCE_DIR} )
		ADD_DEFINITIONS( -DDISABLE_WRITE_FAILED_RAYS )
		SET( SOURCES RayTrace RayTrace.cpp RayTraceStructures.cpp utilities/RayUtilities.cpp interp.cpp RayTraceImageCPU.cpp )
		SET( SOURCES RayTrace RayTraceImage.cpp RayTraceStructures.cpp utilities/RayUtilities.cpp interp.cpp RayTraceImageCPU.cpp )
		IF ( USE_OPENACC )
		SET( SOURCES ${SOURCES} RayTraceImageOpenACC.cpp )
		ENDIF()

src/CreateImage.cpp

+202 −160

Original line number	Diff line number	Diff line
		@@ -2,28 +2,39 @@
		// This miniapp mimics the behavior of create_image

		#include "RayTrace.h"
		#include <cstring>
		#include <iostream>
		#include <math.h>
		#include <sstream>
		#include <stdint.h>
		#include <string>
		#include <cstring>
		#include <sstream>
		#include <math.h>

		#if defined( WIN32 ) \|\| defined( _WIN32 ) \|\| defined( WIN64 ) \|\| defined( _WIN64 )
		#include <windows.h>
		#define get_time( x ) QueryPerformanceCounter( x )
		#define get_frequency( f ) QueryPerformanceFrequency( f )
		#define get_diff(start,end,f) \
		static_cast<double>(end.QuadPart-start.QuadPart)/static_cast<double>(f.QuadPart)
		#define TIME_TYPE LARGE_INTEGER
		#define sleep_ms Sleep
		inline double get_diff( TIME_TYPE start, TIME_TYPE end, TIME_TYPE f )
		{
		return ( ( (double) ( end.QuadPart - start.QuadPart ) ) / ( (double) f.QuadPart ) );
		}
		#else
		#include <sys/time.h>
		#define get_time( x ) gettimeofday( x, NULL );
		#define get_frequency( f ) ( *f = timeval() )
		#define get_diff(start,end,f) 1e-6*static_cast<double>( \
		0xF4240*(static_cast<int64_t>(end.tv_sec)-static_cast<int64_t>(start.tv_sec)) + \
		(static_cast<int64_t>(end.tv_usec)-static_cast<int64_t>(start.tv_usec)) )
		#define TIME_TYPE timeval
		#define sleep_ms( X ) \
		do { \
		struct timespec ts; \
		ts.tv_sec = X / 1000; \
		ts.tv_nsec = ( X % 1000 ) * 1000000; \
		nanosleep( &ts, NULL ); \
		} while ( 0 )
		inline double get_diff( TIME_TYPE start, TIME_TYPE end, TIME_TYPE )
		{
		return ( (double) end.tv_sec - start.tv_sec ) + 1e-6 * ( (double) end.tv_usec - start.tv_usec );
		}
		#endif

		#ifdef USE_KOKKOS
		@@ -42,8 +53,8 @@ inline void fread2( void ptr, size_t size, size_t count, FILE fid )


		// Check the answer
		inline int check_ans( const double image0, const double I_ang0,
		const RayTrace::create_image_struct& data )
		inline int check_ans(
		const double image0, const double I_ang0, const RayTrace::create_image_struct &data )
		{
		size_t N_image = data.euv_beam->nx * data.euv_beam->ny * data.euv_beam->nv;
		size_t N_ang = data.euv_beam->na * data.euv_beam->nb;
		@@ -66,9 +77,11 @@ inline int check_ans( const double image0, const double I_ang0,
		norm1[1] = sqrt( norm1[1] );
		error[0] = sqrt( error[0] ) / norm0[0];
		error[1] = sqrt( error[1] ) / norm0[1];
		const double tol = 5e-6; // RayTrace uses single precision for some calculations (may need to adjust to 1e-5)
		const double tol =
		5e-6; // RayTrace uses single precision for some calculations (may need to adjust to 1e-5)
		// bool pass = error[0]<=tol && error[1]<=tol;
		bool pass = (norm0[0]-norm1[0])/norm0[0]<=tol && (norm0[1]-norm1[1])/norm0[1]<=tol;
		bool pass =
		( norm0[0] - norm1[0] ) / norm0[0] <= tol && ( norm0[1] - norm1[1] ) / norm0[1] <= tol;
		if ( !pass ) {
		std::cerr << " Answers do not match:" << std::endl;
		std::cerr << " image: " << error[0] << " " << norm0[0] << " " << norm1[0] << std::endl;
		@@ -79,7 +92,8 @@ inline int check_ans( const double image0, const double I_ang0,


		// Scale the input problem
		template<class TYPE> void scale_beam( TYPE& beam, double scale )
		template <class TYPE>
		void scale_beam( TYPE &beam, double scale )
		{
		const double x[2] = { beam.x[0] - 0.5 * beam.dx, beam.x[beam.nx - 1] + 0.5 * beam.dx };
		const double y[2] = { beam.y[0] - 0.5 * beam.dy, beam.y[beam.ny - 1] + 0.5 * beam.dy };
		@@ -89,24 +103,41 @@ template<class TYPE> void scale_beam( TYPE& beam, double scale )
		int ny = static_cast<int>( beam.ny * scale );
		int na = static_cast<int>( beam.na * scale );
		int nb = static_cast<int>( beam.nb * scale );
		delete [] beam.x; beam.x = new double[nx];
		delete [] beam.y; beam.y = new double[ny];
		delete [] beam.a; beam.a = new double[na];
		delete [] beam.b; beam.b = new double[nb];
		beam.nx = nx; beam.dx = (x[1]-x[0])/nx;
		beam.ny = ny; beam.dy = (y[1]-y[0])/ny;
		beam.na = na; beam.da = (a[1]-a[0])/na;
		beam.nb = nb; beam.db = (b[1]-b[0])/nb;
		for (int i=0; i<nx; i++) { beam.x[i] = x[0] + (0.5+i)*beam.dx; }
		for (int i=0; i<ny; i++) { beam.y[i] = y[0] + (0.5+i)*beam.dy; }
		for (int i=0; i<na; i++) { beam.a[i] = a[0] + (0.5+i)*beam.da; }
		for (int i=0; i<nb; i++) { beam.b[i] = b[0] + (0.5+i)*beam.db; }
		delete[] beam.x;
		beam.x = new double[nx];
		delete[] beam.y;
		beam.y = new double[ny];
		delete[] beam.a;
		beam.a = new double[na];
		delete[] beam.b;
		beam.b = new double[nb];
		beam.nx = nx;
		beam.dx = ( x[1] - x[0] ) / nx;
		beam.ny = ny;
		beam.dy = ( y[1] - y[0] ) / ny;
		beam.na = na;
		beam.da = ( a[1] - a[0] ) / na;
		beam.nb = nb;
		beam.db = ( b[1] - b[0] ) / nb;
		for ( int i = 0; i < nx; i++ ) {
		beam.x[i] = x[0] + ( 0.5 + i ) * beam.dx;
		}
		for ( int i = 0; i < ny; i++ ) {
		beam.y[i] = y[0] + ( 0.5 + i ) * beam.dy;
		}
		for ( int i = 0; i < na; i++ ) {
		beam.a[i] = a[0] + ( 0.5 + i ) * beam.da;
		}
		for ( int i = 0; i < nb; i++ ) {
		beam.b[i] = b[0] + ( 0.5 + i ) * beam.db;
		}
		}
		void scale_problem( RayTrace::create_image_struct &info, double scale )
		{
		scale_beam( const_cast<RayTrace::EUV_beam_struct >( info.euv_beam ), pow( scale, 0.25 ) );
		if ( info.seed_beam != NULL )
		scale_beam(const_cast<RayTrace::seed_beam_struct>(info.seed_beam),pow(scale,0.25));
		scale_beam(
		const_cast<RayTrace::seed_beam_struct >( info.seed_beam ), pow( scale, 0.25 ) );
		}


		@@ -118,26 +149,33 @@ int main(int argc, char *argv[])

		// Initialize kokkos
		#ifdef USE_KOKKOS
		/*int argc2 = 1;
		#ifdef KOKKOS_HAVE_PTHREAD
		int argc2 = 1;
		const char *argv2[10] = { NULL };
		argv2[0] = argv[0];
		#ifdef KOKKOS_HAVE_PTHREAD
		argv2[argc2] = "--kokkos-threads=4";
		argv2[argc2] = "--kokkos-threads=16";
		argc2++;
		#endif
		Kokkos::initialize(argc2,(char*)argv2);/
		Kokkos::initialize( argc2, (char **) argv2 );
		#else
		Kokkos::initialize( argc, argv );
		#endif
		#endif

		// Check the input arguments
		const char *err_msg = "CreateImage called with the wrong number of arguments:\n"
		" CreateImage <args> file.dat\n"
		"Optional arguments:\n"
		" -methods=METHODS Comma seperated list of methods to test. Default is all availible methods\n"
		" cpu, threads, Cuda, OpenAcc, Kokkos-Serial, Kokkos-Thread, Kokkos-OpenMP, Kokkos-Cuda\n";
		" -iterations=N Number of iterations to run. Time returned will be the average time/iteration.\n"
		" -scale=factor Increate the size of the problem by ~ this factor. (2.0 - twice as expensive)\n"
		" Note: this will disable checking the answer.\n";
		" -methods=METHODS Comma seperated list of methods to test. Default "
		"is all availible methods\n"
		" cpu, threads, OpenMP, Cuda, OpenAcc, Kokkos-Serial, "
		"Kokkos-Thread, Kokkos-OpenMP, Kokkos-Cuda\n"
		" -iterations=N Number of iterations to run. Time returned will be "
		"the average time/iteration.\n"
		" -scale=factor Increate the size of the problem by ~ this factor. "
		"(2.0 - twice as expensive)\n"
		" Note: this will disable checking the answer.\n"
		" Note: the scale factor is only approximate.\n";
		if ( argc < 2 ) {
		std::cerr << err_msg;
		@@ -149,7 +187,6 @@ int main(int argc, char *argv[])
		double scale = 1.0;
		for ( int i = 1; i < argc - 1; i++ ) {
		if ( strncmp( argv[i], "-methods=", 9 ) == 0 ) {
		std::string tmp(&argv[i][9]);
		std::stringstream ss( &argv[i][9] );
		std::string token;
		while ( std::getline( ss, token, ',' ) )
		@@ -179,14 +216,17 @@ int main(int argc, char *argv[])
		// Create the image structure
		RayTrace::create_image_struct info;
		info.unpack( std::pair<char *, size_t>( data, N_bytes ) );
		delete [] data; data = NULL;
		delete[] data;
		data = NULL;
		const double *image0 = info.image;
		const double *I_ang0 = info.I_ang;
		info.image = NULL;
		info.I_ang = NULL;
		if ( scale != 1.0 ) {
		delete [] image0; image0 = NULL;
		delete [] I_ang0; I_ang0 = NULL;
		delete[] image0;
		image0 = NULL;
		delete[] I_ang0;
		I_ang0 = NULL;
		scale_problem( info, scale );
		}

		@@ -196,6 +236,9 @@ int main(int argc, char *argv[])
		#if CXX_STD == 11 \|\| CXX_STD == 14
		methods.push_back( "threads" );
		#endif
		#ifdef USE_OPENMP
		methods.push_back( "OpenMP" );
		#endif
		#ifdef USE_CUDA
		methods.push_back( "Cuda" );
		#endif
		@@ -218,6 +261,7 @@ int main(int argc, char *argv[])

		// Call create_image for each method
		int N_errors = 0;
		sleep_ms( 50 );
		std::vector<double> time( methods.size() );
		for ( size_t i = 0; i < methods.size(); i++ ) {
		printf( "Running %s\n", methods[i].c_str() );
		@@ -256,5 +300,3 @@ int main(int argc, char *argv[])
		#endif
		return N_errors;
		}

src/RayTrace.cpp

deleted100644 → 0

+0 −427

File deleted.

Preview size limit exceeded, changes collapsed.

src/RayTrace.h

100755 → 100644

+78 −82

Original line number	Diff line number	Diff line
		@@ -4,9 +4,9 @@

		#include "RayTraceStructures.h"

		#include <ostream>
		#include <stdio.h>
		#include <stdlib.h>
		#include <ostream>


		/*! \namespace RayTrace
		@@ -34,9 +34,8 @@ namespace RayTrace {
		* @param[out] Iv The output intensity (1xK)
		* @param[out] ray2 The output ray properties (x,y,a,b)
		*/
		int calc_ray( const double ray[4], const int N, const double dz,
		const ray_gain_struct gain, const ray_seed_struct seed,
		int K, int method, double Iv, double ray2 );
		int calc_ray( const double ray[4], const int N, const double dz, const ray_gain_struct *gain,
		const ray_seed_struct seed, int K, int method, double Iv, double *ray2 );


		/*!
		@@ -68,9 +67,8 @@ namespace RayTrace {
		* @param[out] Ir The intensity vs length for the ray paths ( N x Nx x Ny x Na x Nb )
		*/
		int calc_ray_path( int Nx, int Ny, int Na, int Nb, const double x, const double y,
		const double a, const double b, const int N, const double dz,
		const ray_gain_struct gain, const ray_seed_struct seed,
		int K, const double *dv, int method, double c,
		const double a, const double b, const int N, const double dz, const ray_gain_struct *gain,
		const ray_seed_struct seed, int K, const double dv, int method, double c,
		std::vector<float> &xr, std::vector<float> &yr, std::vector<float> &Ir );


		@@ -91,13 +89,11 @@ namespace RayTrace {
		* @brief Function to create the EUV beam images
		* @details This function calculates the near and far field beam images. This function is
		* thread-safe provided unique arrays for mem, image, I_ang are provided in the info struct.
		* @param info Data structure containing the information needed to create the EUV beam images
		* @param info Data structure containing the information needed to create the EUV beam
		* images
		*/
		void create_image( create_image_struct *info, std::string method = "auto" );


		}


		#endif