/******************************************************************************* Copyright(c) 2000 - 2006 Analog Devices. All Rights Reserved. By using this module you agree to the terms of the Analog Devices License Agreement for DSP Software. ******************************************************************************** Module Name: dilation.asm Label name : __dilation Description: For a detailed description, please refer to the file "readme.txt". which is included in the main project. *******************************************************************************/ .section L1_code; .global __dilation; .align 8; __dilation: [--SP] = (R7:6,P5:3); R7 = [SP+32]; // Address of structured element I1 = R7; // Initialize I register to structure element // address B1 = R7; // Initialize base register L1 = 18; // Initialize L register M0 = 8; // Initials to fetch h11 M1 = 6; // initialize to skip 3 element in structured // element array P5 = R0; // Address of input array M3 = R0; // duplicate the address P0 = R2; // Number of columns (N) P1 = R2; // Duplicate the number of columns R6 = R1-R1(NS)||I1+= M0;// Increment to fetch h11 element and clear r6 R7 = 1; P2 = R1; // Number of rows (M) P3 = R2; // Number of columns P0+=-4; // Columns -4 R3 = R2 << 1 ||R0 = B[P5++](Z)||R1.L = W[I1++]; // r3= 2*Col,fetch first input and h11 M0 = R3; r2 = R7 << 1; R3 = R3+R2(NS)||P4 = [SP+36]; // Address of output array M2 = R3; // 2*col +2 P3+= 2; // Col+2 P1+= -2; // Col-2 P2+= -2; // ROW-2; /////////////////////// FIRST ROW /////////////////////////////////// // First row and last row is computed separately. // In each row first and last elements are computed // separately. A0 = R0.L*R1.L(IS) || R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(0,1) and h12 R2 = W[P5++P1](Z) || R3.L = W[I1++]; // dummy fetches to increment P5 and I1 A0+= R0.L *R1.L(IS) ||R0 = B[P5++](Z) || R1.L = W[I1++]; // fetch x(1,0) and h21 A0+=R0.L*R1.L(IS) ||R0 = B[P5](Z) || R1.L = W[I1++]; // fetch x(1,1) and h22 R2 = (A0+=R0.L * R1.L )(IS)|| I1+= M1; // Increment I1 to fetch h10 CC = R2==0; // check if result is zero IF !CC R2 = R7; // if not R2==1 P5 = M3; // Initialize P5 to input array B[P4++] = R2; // store the result LSETUP(FIRST_ROW_ST,FIRST_ROW_END)LC0 = P1; // loop counter == COL-2 FIRST_ROW_ST: R0 = B[P5++](z) || R1.L = W[I1++]; // fetch X(0,0) and H10 A0 = R0.L *R1.L(IS) ||R0 = B[P5++](z) ||R1.L = W[I1++]; // fetch X(0,1) and H11 A0+=R0.L*R1.L(IS) ||R0 = B[P5] (z)||R1.L = W[I1++]; // fetch X(0,2) and H12 P5 = P5+P1; // increment P5 to fetch x(1,0) A0+= R0.L *R1.L (IS) ||R0=B[P5++](Z) || R1.L = W[I1++]; // fetch X(1,0) and H20 A0+= R0.L *R1.L (IS)||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch X(1,1) and H21 A0+= R0.L *R1.L(IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch X(1,2) and H22 R2 = (A0+= R0.L *R1.L)(IS) || I1+= M1; // Increment I1 to fetch h10 CC = R2 ==0; // check if result is 0 IF !CC R2 = R7; // if not r2 = = 1 P5-= P3; FIRST_ROW_END: B[P4++] = R2; // store the result R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(0,N-2) and H10 A0 = R0.L *R1.L (IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(0,n-1) and H11 R2 = W[P5++P1](Z) || R3.L = W[I1++]; // dummy fetch to increment P5 and I1 pointers A0+= R0.L * R1.L (IS)|| R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(1,N-2) and H20 A0+= R0.L * R1.L (IS)|| R0 = B[P5++](z)|| R1.L = W[I1++]; // fetch x(1,N-1) and H21 R2 = (A0+= R0.L *R1.L) (IS) ||I1+= 4; // Increment to fetch h01 CC = R2 == 0; IF !CC R2 = R7; P5 = M3; // Initialize P5 to starting of input array B[P4++] = R2; // store output of last element in first row //////////// MIDDLE LOOP ////////////////////////////////////////////// P3 = M0; // Initialized to 2*col LSETUP(ROW_ST,ROW_END)LC0 = P2; // loop counter = = ROW-2 (M-2) P2 = M2; // Initialized to 2*COL+2 ROW_ST: R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(0,0) and H01 A0 = R0.L *R1.L(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(0,1) and H02 R2 = W[P5++P1](Z) ||R3.L = W[I1++]; // dummy fetch to adjust P5 and I1 A0+= R0.L*R1.L(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(1,0) and H11 A0+= R0.L* R1.L(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(1,1) and H12 R2 = W[P5++P1](Z)|| R3.L = W[I1++]; // dummy fetch to adjust P5 and I1 A0 += R0.L*R1.L(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(2,0) and H21 A0 += R0.L* R1.L(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(2,1) and H22 P5 -= P2; // P5 modified to stating of the row R2 = (A0 += R0.L*R1.L )(IS)||R0 = B[P5++](z) || R1.L = W[I1++]; //fetch first element of the row and h00 R2 = MAX(R2,R6); R2 = MIN(R2,R7); B[P4++] = R2 || NOP;// store the output ///////////////// CORE LOOP ////////////////////////////////////// // Here data is processed to get two pixel output // at a time LSETUP(COL_ST,COL_END)LC1 = P1 >> 1; // loop counter = =(COL-2)/2 COL_ST: A0 = R0.L*R1.L(IS) ||R0 = B[P5++](z) || R1.H = W[I1++]; // fetch x(0,1) and h01 A1 = R0.L * R1.L ,A0+= R0.L*R1.H(IS) || R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(1,2) and h02 A1+= R0.L *R1.H ,A0+= R0.L *R1.L(IS) || R0 = B[P5++](z) || R1.H = W[I1++]; // fetch x(1,3) and h10 P5 = P5+P0; // move P5 to starting of next row A1+= R0.L *R1.L (IS) ||R0 = B[P5++](z) ||R1.L = W[I1++]; // fetch x(1,0) and h11 A0+= R0.L *R1.H(IS)|| R0 = B[P5++](z) ; // fetch x(1,1) A1+= R0.L * R1.H,A0+= R0.L *R1.L(IS) ||R0 = B[P5++](z) || R1.H = W[I1++]; // fetch x(1,2) and h12 A1+= R0.L *R1.L ,A0+= R0.L *R1.H(IS)|| R0 = B[P5++](z) ||R1.L = W[I1++]; // fetch x(1,3) and h20 P5 = P5+P0; // move P5 to starting of next row A1+= R0.L *R1.H(IS)||R0 = B[P5++](z)|| R1.H = W[I1++]; // fetch x(2,0) and h21 A0+= R0.L*R1.L(IS) ||R0 = B[P5++](z); // fetch x(2,1) A1+= R0.L * R1.L ,A0+= R0.L *R1.H(IS) ||R0 = B[P5++] (z) ||R1.L = W[I1++]; // fetch x(2,2) and h22 A1+= R0.L *R1.H,R2 = (A0+=R0.L *R1.L)(IS)||R0 = B[P5++](z); // fetch x(2,3) P5-=P2; // modify P5 to process next set of data R3 = (A1+=R0.L *R1.L)(IS)||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(0,2) and h00 CC = R2==0; IF !CC R2 = R7; CC = R3==0; IF !CC R3 = R7; MNOP||B[P4++] = R2; // store first output COL_END: B[P4++] = R3; // store next output A0 = R0.L *R1.L (IS)||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(0,N-2) and h01 R2 = W[P5++P1](Z) ||R3.L = W[I1++]; // dummy fetch to increment P5 and I1 A0+= R0.L *R1.L (IS)||R0=B[P5++](z) || R1.L = W[I1++]; // fetch X(1,N-2) and H10; A0+= R0.L*R1.L(IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch X(1,N-1) and H11; R2 = W[P5++P1](Z) || R3.L = W[I1++]; // dummy fetch to increment P5 and I1 A0+= R0.L *R1.L (IS)||R0 = B[P5++] (z)|| R1.L = W[I1++]; // fetch x(2,N-2) and h20 A0+= R0.L*R1.L(IS) ||R0 = B[P5++] (z)|| R1.L = W[I1++]; // fetch x(2,N-1) and h21; R2 = (A0+=R0.L *R1.L)(IS)|| I1+= 4; // modify I1 to fetch h01 CC = R2==0; IF !CC R2 = R7; P5-= P3; // modify P5 to starting of next row ROW_END: B[P4++] = R2; //////////////////// LAST ROW ////////////////////////////////////////////// P2=P1; P2+= 4; // P2==Col +2 R0 = B[P5++](z)||R1.L = W[I1++]; // fetch h01 and x(M-2,0) A0 = R0.L *R1.L(IS) || R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(M-2,1) and h02 R2 = W[P5++P1](Z)|| R3.L = W[I1++]; // dummy fetch to increment P5 and I1 A0+=R0.L*R1.L(IS)||R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(M-1,0) and h11 A0+=R0.L *R1.L (IS)||R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(M-1,2) and h12 P5-=P2; MNOP|| I1+=M1; // I1 is modified to fetch h00 R2 = (A0+=R0.L *R1.L) (IS)||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(M-2,0) and h00 CC = R2==0; IF !CC R2 = R7; B[P4++] = R2; // store the result LSETUP(LAST_ROW_ST,LAST_ROW_END)LC0 = P1; LAST_ROW_ST: A0 = R0.L*R1.L (IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(M-2,1) and h01 A0+=R0.L *R1.L(IS) ||R0 = B[P5](z) || R1.L = W[I1++]; // fetch x(M-2,2) and h02 P5 = P5+P1; // modify P5 to next row A0+=R0.L *R1.L(IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(M-1,0) and h10 A0+=R0.L *R1.L(IS) ||R0 = B[P5++](Z) || R1.L = W[I1++]; // fetch x(M-1,1) and h11 A0+=R0.L *R1.L(IS) ||R0 = B[P5++](z) || R1.L = W[I1++]; // fetch x(M-1,2) and h12 I1 += M1; // modify I1 to fetch h00 P5 -= P2; R2 = (A0+=R0.L *R1.L)(IS)||R0 = B[P5++] (z)|| R1.L = W[I1++]; // fetch h00 CC = R2==0; IF !CC R2 = R7; LAST_ROW_END: B[P4++] = R2; // store the result A0 = R0.L *R1.L(IS) || R0 = B[P5++](z)||R1.L = W[I1++]; // fetch x(m-2,N-1) and h01 R2 = W[P5++P1](Z) || R3.L = W[I1++]; // dummy fetch to increment P5 to next row and I1 // to fetch h10 A0+=R0.L*R1.L(IS)||R0 = B[P5++](z)||R1.L = W[I1++]; // fetch h10 and x(m-1,N-2) A0+= R0.L *R1.L (IS)||R0 = B[P5++](z)||R1.L = W[I1++]; // fetch h11 and x(m-1,N-1) R2 = (A0+=R0.L *R1.L )(IS); CC = R2 ==0; IF !CC R2 = R7; B[P4++] = R2 ; // store the final output pixel (R7:6,P5:3) = [SP++]; // POP r7 and P5-3 RTS; NOP; // To avoid one stall if LINK or UNLINK happens to // be the next instruction in the memory. __dilation.end: