version 2.0, interleaved streams

define.h

@@ -1,3 +1,4 @@
+
 #ifndef M_PI
 #define _USE_MATH_DEFINES  //for Windows
 #endif
@@ -25,7 +26,7 @@
 #endif
 
 
-#define VERSION 1.721
+#define VERSION 2.0
 
 
 #define def_T0 296.0 		//Reference Temperature in K
@@ -41,14 +42,24 @@
 
 #define def_TOL 1.43e-17		//Tolerance in the Voigt function 3.58e-9 2.48e-12 1.43e-17 3.25e-27 1.69e-33
 #define def_TOLF 2.48e-12f		//Tolerance in the Voigt function
-#define def_nthmax 32768   		//Maximum number of threads in 2.0 Cards
+#define def_nthmax 1048576   		//Maximum number of threads for line and sort kernels
 #define def_nlmax 32768			//Maximum number of lines per kernel launch, to prevent from time out on Desktop machines
-//#define def_maxlines 30000000ll		//maximum number of lines stored on the GPU, Should not be less than maximum in HITEMP lines
-#define def_maxlines 8000000ll		//maximum number of lines stored on the GPU, Should not be less than maximum in HITEMP lines, must be long long int type
+#define def_maxlines 1000000ll		//maximum number of lines stored on the GPU, Should not be less than maximum in HITEMP lines, must be long long int type
 #define def_maxfiles 500		//maximum number of files per molecule
+#define def_doTuning 1			//use the self-tuning routines
 
 #define def_NmaxSample 100		//Maximum Number of resample coefficients for K(y)
 
+
+#define def_nlA 1024			//split lines into blocks of this size, A Line kernels
+#define def_nlB 1024			//split lines into blocks of this size, A Line kernels
+#define def_nlC 1024			//split lines into blocks of this size, A Line kernels
+
+#define def_rBs 256			//number of streams in readBuffer copy
+
+#define def_KSn 32
+
+
 //default values of parameters
 #define def_qALPHA_L 0.5	//q value in the Lorentz half width q = Pself / P
 #define def_gammaF 1.0		//scaling factor for Lorentzian halfwidth
@@ -138,29 +149,55 @@ struct Param{
 	int units;
 	int useIndividualX;
 	int replaceFiles;
-	int RLOW;
 	int profile;
 	double gammaF;
+	int doTuning;
 };
 
 struct Line{
-	double *nu_h, *nu_d;            //Wavenumber
-	double *S_h, *S_d;        //Intensity
+	double *nu_h, *nu_d;		//Wavenumber
+	double *S_h, *S_d;		//Intensity
 	float *Sf_d;
-	double *S1_d;     //modified Intensity
+	double *S1_d;			//modified Intensity
 	float *S1f_d;
-	double *A_h, *A_d;              //Einstein A coefficient
-	double *delta_h, *delta_d;      //pressure induced line shift
-	double *EL_h, *EL_d;  		//Energy of lower state
-	double *vy_h, *vy_d; 	//Lorentz Halfwidth / Doppler Halfwidth
+	double *A_h, *A_d;		//Einstein A coefficient
+	double *delta_h, *delta_d;	//pressure induced line shift
+	double *EL_h, *EL_d;		//Energy of lower state
+	double *vy_h, *vy_d;		//Lorentz Halfwidth / Doppler Halfwidth
 	float *vyf_d;
-	float *va_d;	 	//(numin - nu) * ialphaD
-	float *vb_d;	 	//dnu * ialphaD
-	float *vcut2_d; 	//(cut * ialphaD)^2
-	double *ialphaD_h, *ialphaD_d;  //Doppler Halfwidth
-	double *n_h, *n_d;              //temperature dependent exponent
-	double *Q_h, *Q_d;              //partition function
-	int *ID_h, *ID_d;		//line id used for sorting
-
-
+	float *va_d;			//(numin - nu) * ialphaD
+	float *vb_d;			//dnu * ialphaD
+	float *vcut2_d;			//(cut * ialphaD)^2
+	double *ialphaD_h, *ialphaD_d;	//Doppler Halfwidth
+	double *n_h, *n_d;		//temperature dependent exponent
+	double *Sort_d;			//helper array used to sort the other arrays
+	int *ID_d;			//line id used for sorting
+	double *nuLimitsA0_d;	//limits for Line blocks, min
+	double *nuLimitsA1_d;	//limits for Line blocks, max
+	double *nuLimitsAL0_d;	//limits for Line blocks, min
+	double *nuLimitsAL1_d;	//limits for Line blocks, max
+	double *nuLimitsAR0_d;	//limits for Line blocks, min
+	double *nuLimitsAR1_d;	//limits for Line blocks, max
+	double *nuLimitsB0_d;	//limits for Line blocks, min
+	double *nuLimitsB1_d;	//limits for Line blocks, max
+	double *nuLimitsC0_d;	//limits for Line blocks, min
+	double *nuLimitsC1_d;	//limits for Line blocks, max
+
+	long long int *iiLimitsA0_h, *iiLimitsA0_d;	//limits for Line blocks, min
+	long long int *iiLimitsA1_h, *iiLimitsA1_d;	//limits for Line blocks, max
+	long long int *iiLimitsAL0_h, *iiLimitsAL0_d;	//limits for Line blocks, min
+	long long int *iiLimitsAL1_h, *iiLimitsAL1_d;	//limits for Line blocks, max
+	long long int *iiLimitsAR0_h, *iiLimitsAR0_d;	//limits for Line blocks, min
+	long long int *iiLimitsAR1_h, *iiLimitsAR1_d;	//limits for Line blocks, max
+	long long int *iiLimitsB0_h, *iiLimitsB0_d;	//limits for Line blocks, min
+	long long int *iiLimitsB1_h, *iiLimitsB1_d;	//limits for Line blocks, max
+	long long int *iiLimitsC0_h, *iiLimitsC0_d;	//limits for Line blocks, min
+	long long int *iiLimitsC1_h, *iiLimitsC1_d;	//limits for Line blocks, max
+
+	long long int *iiLimitsAT_m, *iiLimitsAT_d;	//limits for all blocks, mapped memory
+	long long int *iiLimitsALT_m, *iiLimitsALT_d;	//limits for all blocks, mapped memory
+	long long int *iiLimitsART_m, *iiLimitsART_d;	//limits for all blocks, mapped memory
+	long long int *iiLimitsBT_m, *iiLimitsBT_d;	//limits for all blocks, mapped memory
+	long long int *iiLimitsCT_m, *iiLimitsCT_d;	//limits for all blocks, mapped memory
+
 };