/normxcorr/trunk

#ifndef NORMXCORR_HW_H

#define NORMXCORR_HW_H

#include <cuda.h>

#include <cuda_runtime.h>

#include <cublas.h>

#include <cufft.h>

#include <cudpp.h>

#include "dict_image.h"

#define TRANSPOSE_SIZE 64	// should equal to CPU cache line size

#define BLOCK_SIZE_1D 		64

#define BLOCK_SIZE_2D 		16

#define CP_BLOCK 256		// should be divisable by CP_BLOCK_SIZE, BLOCK_SIZE_xD

#define CP_BLOCK_SIZE BLOCK_SIZE_2D

#define SIDE_BLOCK_SIZE BLOCK_SIZE_2D

#define USE_UNDOCUMENTED

//#define USE_SSE

struct STProcessingState {

    int stored;				// flag indicating if we already have coordinates in coords stored

    unsigned char *banlist;		// control points banned from computations for various reasons

    float *points;			// various information on control points

					// 	base_x, base_y, data_x, data_y

					//	base_frac_x, base_frac_y

					//	move_x, move_y

    float *cuda_points;			// Various information on control points:

					// 0: data_x

					// 1: data_y

					// 2: sum

					// 3: denom

    unsigned char *input_buffer;	// Input Image buffer / Host

    unsigned char *cuda_input_buffer;	// Input Image buffer / Device

    cufftReal *cuda_base_buffer;	// Temporary buffer for FFT inputs, pre-zeroed

    cufftReal *cuda_data_buffer;	// Temporary buffer for FFT inputs, pre-zeroed

    void *cuda_temp_buffer;		// Main computational buffer, temporary

    cufftComplex *cuda_fft_cache;	// Stored FFT's of the template image

    float *cuda_lsum_cache;		// Cache of local sums

    float *cuda_denom_cache;		// Cahce of denoms

    float *cuda_lsum_temp;		// Temporary buffer for local sum comp, first two pre-zeroed

    float *res_x;			// External points buffer

    float *res_y;			// External points buffer

    int fft_initialized;		// Flag indicating if CUFFT plan is initialized

    cufftHandle cufft_plan;

    cufftHandle cufft_r2c_plan;

    cufftHandle cufft_c2r_plan;

    int cudpp_initialized;		// Flag indicating if CUDPP plan is initialized

    CUDPPHandle cudpp_plan;

    int matlab_mode;		// 2D stored transposed (i.e. by columns)

    int mode;			// 1 - image mode, 0 - fragment mode

    int base_mode;		// 1 - image mode, 0 - fragment mode

    float minx,miny,maxx,maxy;	// Coordinates of actualy used rectangle of image

    int ncp;			// Number of control points

    int ncp_alloc_size;

    int width;			// Images width

    int height;			// Images height

    int corr_size;		// CORR_SIZE 

    int precision;		// PRECISION

    int side_alloc_size;	// allocation size for 1 side of fft

    int fft_size;		// Matrix Size for FFT (base_size + input_size - 1)

    int fft_real_size;		// Rounded to next power of 2

    int fft_alloc_size;		// cuda optimized size2

    int subimage_size;		// Size of neighborhood (4*corr_size + 1)

    int lsum_size;		// Dimmensions of local sums (2*corr_size + 1)

    int lsum_temp_size;		// Matrix Size for computing local sums

    int lsum_alloc_size;	// Size of allocated line to store lsum row

    int lsum_aligned_size;	// CUDA optimized lsum_temp_size

    int lsum_short_aligned_size;// CUDA optimized lsum_temp_size - lsum_size - 1

    int side_blocks_power;	// Indicates if amount of side blocks is power of 2

    int base_blocks_power;	// Indicates if amount of base blocks is power of 2

    DICTImageType image_type;

    void *image_buf;		// Temporary buffer for images (specific-format)

    unsigned char *image;	// Reduced format

#ifdef DICT_HW_MEASURE_TIMINGS

    int time[24];		// Timing counters (SHOULD BE LAST in struct!!!): 

				//  0- 3: hardware, init, set_base_points, set_cur_points

				//  4-11: comp_base, comp_image, (copy_fragment, load_fragment, preprocess, process, postprocess, reserve)

				// 12-15: load_base_tiff, base_image_reduction, load_tiff, image_reduction

				// 16   :get_results

#endif /* DICT_HW_MEASURE_TIMINGS */

};

typedef struct STProcessingState TProcessingState;

#endif /* NORMXCORR_HW_H */