/*******************************************************************************
Copyright(c) 2000 - 2002 Analog Devices. All Rights Reserved.
Developed by Joint Development Software Application Team, IPDC, Bangalore, India
for Blackfin DSPs  ( Micro Signal Architecture 1.0 specification).

By using this module you agree to the terms of the Analog Devices License
Agreement for DSP Software. 
********************************************************************************
Module Name     : mve_core.asm
Label Name      : __mve_core
Version         :   1.0
Change History  :

                Version     Date          Author            Comments
                1.0         07/02/2001    Vijay             Original 

Description     : The assembly function is used by the motion estimation routine
                  to compute the best matching macroblock for a given target 
                  block and to compute the motion vectors for that matching 
                  block. The subroutine fetches all the possible blocks within 
                  the search range in the reference frame and determines the 
                  best matching macro block for a give target block based on the
                  MAD. The MAD of a pair of 16x16 blocks denoted by TGT_BLK and 
                  REF_BLK is computed as follows :

                      MAD = Sum( ABS[TGT_BLK - REF_BLK ] )

                  The offset between the best matching reference block and the 
                  target block gives the motion vector. Half pixelation is not 
                  done inside this subroutine.

Prototype       : void _mve_core( srruct *input_parameters, struct *output);

                  The input to this routine is a set of two pointers. The first 
                  pointer points to a structure containing the following input 
                  parameters.

                  Input : Structure 1 :
                        struct {
                                    int hor_size;
                            -> Horizontal size of the video frame 
                                    int left_search_boundary;
                            -> (2^search_factor + 1) 
                                    unsigned char *tgt_blk_add;
                            -> Start address of the target block; 
                                    unsigned char *loc_tgt_blk_in_ref_frame;
                            -> Location of target block in the reference frame 
                                    unsigned char *search_begin_st_add;
                            -> Start address from where search should begin 
                                    int hor_search_range;
                            -> Search range in pixels in the horizontal 
                               direction 
                                    int ver_search_range;
                            -> Search range in pixels in the vertical direction 
                                }input_parameters;

                  The second structure acts as an output where the elements of 
                  the structure are modified by this routine

                  Output : Structure 2 :
                        struct {
                                    int least_mad;
                            -> The value of the least MAD after searching the 
                               search window fully 
                                    unsigned char *add_least_mad;
                            -> Address of the block having the least MAD 
                                    int vert_mv;
                            -> Vertical integer motion vector 
                                    int hor_mv;
                            -> Horizontal integer motion vector 
                                }output;

Registers used  : A0, A1, R0-R3, R5-R7, I0, I1, I3, M1-M3, L0, L1, L3, P0-P5, 
                  LC0 LC1.

Performance     :
                Code size : 388 Bytes.
*******************************************************************************/
.section L1_code;
.align 8;
.global __mve_core;
    
__mve_core:

    [--SP] = (R7:5, P5:3);
    L0 = 0;
    L1 = 0;
    L3 = 0;
    P5 = R0;
    I3 = R1;
    P3 = 16;                // Loop ctr(for 16 rows) initialized
    R7.L = -1;              // Initialize R7 to the positive maximum number
    R7.H = 0x7FFF;          // R7 contains the minimum MAD or MSE
    R0 = [P5];              // Horizontal size of the frame
    R1 = R0 << 4 || P1 = [P5 + 16];
                            // Starting address from where subroutine must start
                            // searching for matching blocks 
    M2 = R1;                // 16*Horizontal size
    R0 += -16;
    M1 = R0;                // Horizontal size - 16
    R0 += 4;
    M3 = R0;                // Horizontal size - 12
    MNOP || R1 = [P5 + 8];
    I0 = R1;                // Starting address of target block
    MNOP || R2 = [P5 + 20];
    P4 = R2;                // Horizontal search range
    R6 = [P5 + 24] || NOP;  // Vertical search range
    P2 = 0;
VERTICAL:
    P1 = P1 + P2;           // Update the reference block row address
    P2 = P1;                
    LSETUP(ST_SEARCH, END_SEARCH) LC1 = P4;
                            // Set horizontal span of the search range 
ST_SEARCH:
        I1 = P2;            // Fetch the start address of the reference block
/*********************** MEAN ABSOLUTE DIFFERENCE *****************************/
    
        A1=A0=0;            // Initialize accumulators for accumulation
        DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];
                            // Fetch the first data from the two blocks 
        LSETUP (MAD_START, MAD_END) LC0=P3;
MAD_START:  DISALGNEXCPT || R3 = [I1++];
            SAA (R1:0,R3:2) || R1 = [I0++]  || R2 = [I1++];
                            // Compute absolute difference and accumulate 
            SAA (R1:0,R3:2) (R) || R0 = [I0++] || R3 = [I1++];
                            //                  | 
            SAA (R1:0,R3:2) || R1 = [I0 ++ M3] || R2 = [I1++M1];
                            //  After fetch of 4th word of target blk, pointer 
                            // made to point next row 
MAD_END:    SAA (R1:0,R3:2) (R) || R0 = [I0++] || R2 = [I1++];

        R3=A1.L+A1.H,R2=A0.L+A0.H || I0 -= M2;    
        R0 = R2 + R3 (NS) || I0 -= 4;
                            // Add the accumulated values in both MACs 
    
    
/*************************** MINIMUM MAD COMPUTATION **************************/
        CC = R0 <= R7;      // Check if the latest MAD or MSE is less than the 
                            // previous ones
        IF !CC JUMP END_SEARCH (BP);
                            // If latest MAD is not lesser, then return 
        CC = R0 < R7;                           
        IF !CC JUMP EQUAL;  // If MAD is lesser jump to 'LESS'
        R7 = R0;            // If latest MAD is less, latest MAD or MSE is the 
                            // minimum
        P0 = P2;            // and corresponding block is better match
        JUMP END_SEARCH;
EQUAL:                      // Compute the MVs for the least, previous MAD
        R1 = P0;
        R0 = R0 - R0 (NS) || R2 = [P5 + 12];
        R3 = R1 - R2 (NS) || R5 = [P5];
                            // Find the difference between current block and 
                            // reference block 
        R1 = ABS R3 || R2 = [P5 + 4];
                            // Take the absolute value of the difference 
REPEAT1:
        CC = R1 < R5;
        IF CC JUMP FINISH_MOD1;                 
        R1 = R1 - R5;       // Divide the offset by COLUMNS
        R0 += 2;            // R0 has twice the quotient and R1 has the 
                            // remainder
        JUMP REPEAT1;
FINISH_MOD1:
        CC = R1 <= R2;
        IF CC JUMP NOTRANS1;
        R0 += 2;            // If remainder is greater than that value, 
                            // increment quotient by two
        R1 = R1 - R5;       // and remainder = remainder - COLUMNS
NOTRANS1:
        R1 <<= 1;           // Remainder*2 gives horizontal motion vector
        R2 = -R0;
        R5 = -R1;
        CC = R3 < 0;        // If R3 is negative (diff. negative), negate both 
                            // quotient and remainder
        IF CC R0 = R2;                          
        IF CC R1 = R5;                          
        A1 = R0.L*R0.L (IS);// Distance to previous best match from reference 
                            // block
        A1 += R1.L*R1.L (IS);
    
                            // Compute the MVs for the least, current MAD
        R1 = P2;
        R0 = R0 - R0 (NS) || R2 = [P5 + 12];
        R3 = R1 - R2 (NS) || R5 = [P5];
                            // Find the difference between current block and 
                            // reference block 
        R1 = ABS R3 || R2 = [P5 + 4];
                            // Take the absolute value of the difference
REPEAT2:
        CC = R1 < R5;
        IF CC JUMP FINISH_MOD2;                 
        R1 = R1 - R5;       // Divide the offset by COLUMNS
        R0 += 2;            // R0 has twice the quotient and R1 has the 
                            // remainder
        JUMP REPEAT2;
FINISH_MOD2:
        CC = R1 <= R2;
        IF CC JUMP NOTRANS2;
        R0 += 2;            // If remainder is greater than that value, 
                            // increment quotient by two
        R1 = R1 - R5;       // and remainder = remainder - COLUMNS
NOTRANS2:
        R1 <<= 1;           // Remainder*2 gives horizontal motion vector
        R2 = -R0;
        R5 = -R1;
        CC = R3 < 0;        // If R3 is negative (diff. negative), negate both 
                            // quotient and remainder
        IF CC R0 = R2;                          
        IF CC R1 = R5;                          
        A0 = R0.L*R0.L (IS);
        A0 += R1.L*R1.L (IS);
                            // Distance to current block from reference block 
    
        CC = A0 < A1;
        IF CC P0 = P2;      // Make current block as the best match till now
    
END_SEARCH:
        P2 += 1;            // Update the reference block column address
    P2 = [P5];              // Horizontal size of the video frame
    R6 += -1;
    CC = R6 <= 0;           // Check if the all the rows in the search area are 
                            // completed
    IF !CC JUMP  VERTICAL (BP);
                            // Repeat till the entire search area is matched 
    
                            // Compute the motion vectors
    R1 = P0;
    R0 = R0 - R0 (NS) || R2 = [P5 + 12];
    R3 = R1 - R2 (NS) || R5 = [P5];
                            // Find the difference between current block and 
                            // reference block 
    R1 = ABS R3 || R2 = [P5 + 4];
                            // Take the absolute value of the difference
REPEAT3:
    CC = R1 < R5;
    IF CC JUMP FINISH_MOD3;                 
    R1 = R1 - R5;           // Divide the offset by COLUMNS
    R0 += 2;                // R0 has twice the quotient and R1 has the 
                            // remainder
    JUMP REPEAT3;
FINISH_MOD3:
    CC = R1 <= R2;
    IF CC JUMP NOTRANS3;
    R0 += 2;                // If remainder is greater than that value, 
                            // increment quotient by two
    R1 = R1 - R5;           // and remainder = remainder - COLUMNS
NOTRANS3:
    R1 <<= 1;               // Remainder*2 gives horizontal motion vector
    R2 = -R0;
    R5 = -R1;
    CC = R3 < 0;            // If R3 is negative (diff. negative), negate both 
                            // quotient and remainder
    IF CC R0 = R2;                          
    IF CC R1 = R5;                          
    R3 = P0;
    [I3++] = R7;            // Store the least MAD in the output structure
    [I3++] = R3;            // Store the address of the least MAD in the output 
                            // structure
    [I3++] = R0;            // Store the integer vertical MV in the output 
                            // structure
    [I3++] = R1;            // Store the integer horizontal MV in the output 
                            // structure
    (R7:5, P5:3) = [SP++];          
    RTS;
    NOP;                    //to avoid one stall if LINK or UNLINK happens to be
                            //the next instruction after RTS in the memory.
__mve_core.end: