include/gpudefs.h

#ifndef GPUDEFS_H
#define GPUDEFS_H

//== INCLUDES =================================================================

#include <cuda_runtime.h>

#include <defs.h>
#include <extension.h>

//== DEFINITIONS ===============================================================

#define HWS 16 // Half Warp Size
#define DW 8 // Default number of warps (computational block height)
#define CHW 7 // Carry-heavy number of warps (computational block height for some kernels)
#define OW 6 // Optimized number of warps (computational block height for some kernels)
#define DNB 6 // Default number of blocks per SM (minimum blocks per SM launch bounds)
#define ONB 5 // Optimized number of blocks per SM (minimum blocks per SM for some kernels)
#define MTS 192 // Maximum number of threads per block with 8 blocks per SM
#define MBO 8 // Maximum number of blocks per SM using optimize or maximum warps
#define CHB 7 // Carry-heavy number of blocks per SM using default number of warps
#define MW 6 // Maximum number of warps per block with 8 blocks per SM (with all warps computing)
#define SOW 5 // Dual-scheduler optimized number of warps per block (with 8 blocks per SM and to use the dual scheduler with 1 computing warp)
#define MBH 3 // Maximum number of blocks per SM using half-warp size

//== NAMESPACES ===============================================================

namespace gpufilter {

//== CLASS DEFINITION =========================================================

typedef struct _alg_setup {
    int width, 
        height, 
        m_size, 
        n_size, 
        last_m, 
        last_n, 
        border, 
        carry_height, 
        carry_width; 
    float inv_width, 
        inv_height; 
} alg_setup; 

//== PROTOTYPES ===============================================================

extern
void calc_alg_setup( alg_setup& algs,
                     const int& w,
                     const int& h );

extern
void calc_alg_setup( alg_setup& algs,
                     const int& w,
                     const int& h,
                     const int& extb );

extern
void up_alg_setup( const alg_setup& algs );

extern
void up_constants_coefficients1( const float& b0,
                                 const float& a1 );

extern
void up_constants_coefficients2( const float& b0,
                                 const float& a1,
                                 const float& a2 );

//=============================================================================
} // namespace gpufilter
//=============================================================================
#endif // GPUDEFS_H
//=============================================================================