What approach to take for SIMD optimizations
- by goldenmean
Hi,
I am trying to optimize below code for SIMD operations (8way/4way/2way SIMD whiechever possible and if it gives gains in performance) I am tryin to analyze it first on paper to understand the algorithm used. How can i optimize it for SIMD:-
void idct(uint8_t *dst, int stride, int16_t *input, int type)
{
int16_t *ip = input;
uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
int A, B, C, D, Ad, Bd, Cd, Dd, E, F, G, H;
int Ed, Gd, Add, Bdd, Fd, Hd;
int i;
/* Inverse DCT on the rows now */
for (i = 0; i < 8; i++) {
/* Check for non-zero values */
if ( ip[0] | ip[1] | ip[2] | ip[3] | ip[4] | ip[5] | ip[6] | ip[7] ) {
A = M(xC1S7, ip[1]) + M(xC7S1, ip[7]);
B = M(xC7S1, ip[1]) - M(xC1S7, ip[7]);
C = M(xC3S5, ip[3]) + M(xC5S3, ip[5]);
D = M(xC3S5, ip[5]) - M(xC5S3, ip[3]);
Ad = M(xC4S4, (A - C));
Bd = M(xC4S4, (B - D));
Cd = A + C;
Dd = B + D;
E = M(xC4S4, (ip[0] + ip[4]));
F = M(xC4S4, (ip[0] - ip[4]));
G = M(xC2S6, ip[2]) + M(xC6S2, ip[6]);
H = M(xC6S2, ip[2]) - M(xC2S6, ip[6]);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
/* Final sequence of operations over-write original inputs. */
ip[0] = (int16_t)(Gd + Cd) ;
ip[7] = (int16_t)(Gd - Cd );
ip[1] = (int16_t)(Add + Hd);
ip[2] = (int16_t)(Add - Hd);
ip[3] = (int16_t)(Ed + Dd) ;
ip[4] = (int16_t)(Ed - Dd );
ip[5] = (int16_t)(Fd + Bdd);
ip[6] = (int16_t)(Fd - Bdd);
}
ip += 8; /* next row */
}
ip = input;
for ( i = 0; i < 8; i++) {
/* Check for non-zero values (bitwise or faster than ||) */
if ( ip[1 * 8] | ip[2 * 8] | ip[3 * 8] |
ip[4 * 8] | ip[5 * 8] | ip[6 * 8] | ip[7 * 8] ) {
A = M(xC1S7, ip[1*8]) + M(xC7S1, ip[7*8]);
B = M(xC7S1, ip[1*8]) - M(xC1S7, ip[7*8]);
C = M(xC3S5, ip[3*8]) + M(xC5S3, ip[5*8]);
D = M(xC3S5, ip[5*8]) - M(xC5S3, ip[3*8]);
Ad = M(xC4S4, (A - C));
Bd = M(xC4S4, (B - D));
Cd = A + C;
Dd = B + D;
E = M(xC4S4, (ip[0*8] + ip[4*8])) + 8;
F = M(xC4S4, (ip[0*8] - ip[4*8])) + 8;
if(type==1){ //HACK
E += 16*128;
F += 16*128;
}
G = M(xC2S6, ip[2*8]) + M(xC6S2, ip[6*8]);
H = M(xC6S2, ip[2*8]) - M(xC2S6, ip[6*8]);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
/* Final sequence of operations over-write original inputs. */
if(type==0){
ip[0*8] = (int16_t)((Gd + Cd ) >> 4);
ip[7*8] = (int16_t)((Gd - Cd ) >> 4);
ip[1*8] = (int16_t)((Add + Hd ) >> 4);
ip[2*8] = (int16_t)((Add - Hd ) >> 4);
ip[3*8] = (int16_t)((Ed + Dd ) >> 4);
ip[4*8] = (int16_t)((Ed - Dd ) >> 4);
ip[5*8] = (int16_t)((Fd + Bdd ) >> 4);
ip[6*8] = (int16_t)((Fd - Bdd ) >> 4);
}else if(type==1){
dst[0*stride] = cm[(Gd + Cd ) >> 4];
dst[7*stride] = cm[(Gd - Cd ) >> 4];
dst[1*stride] = cm[(Add + Hd ) >> 4];
dst[2*stride] = cm[(Add - Hd ) >> 4];
dst[3*stride] = cm[(Ed + Dd ) >> 4];
dst[4*stride] = cm[(Ed - Dd ) >> 4];
dst[5*stride] = cm[(Fd + Bdd ) >> 4];
dst[6*stride] = cm[(Fd - Bdd ) >> 4];
}else{
dst[0*stride] = cm[dst[0*stride] + ((Gd + Cd ) >> 4)];
dst[7*stride] = cm[dst[7*stride] + ((Gd - Cd ) >> 4)];
dst[1*stride] = cm[dst[1*stride] + ((Add + Hd ) >> 4)];
dst[2*stride] = cm[dst[2*stride] + ((Add - Hd ) >> 4)];
dst[3*stride] = cm[dst[3*stride] + ((Ed + Dd ) >> 4)];
dst[4*stride] = cm[dst[4*stride] + ((Ed - Dd ) >> 4)];
dst[5*stride] = cm[dst[5*stride] + ((Fd + Bdd ) >> 4)];
dst[6*stride] = cm[dst[6*stride] + ((Fd - Bdd ) >> 4)];
}
} else {
if(type==0){
ip[0*8] =
ip[1*8] =
ip[2*8] =
ip[3*8] =
ip[4*8] =
ip[5*8] =
ip[6*8] =
ip[7*8] = ((xC4S4 * ip[0*8] + (IdctAdjustBeforeShift<<16))>>20);
}else if(type==1){
dst[0*stride]=
dst[1*stride]=
dst[2*stride]=
dst[3*stride]=
dst[4*stride]=
dst[5*stride]=
dst[6*stride]=
dst[7*stride]= cm[128 + ((xC4S4 * ip[0*8] + (IdctAdjustBeforeShift<<16))>>20)];
}else{
if(ip[0*8]){
int v= ((xC4S4 * ip[0*8] + (IdctAdjustBeforeShift<<16))>>20);
dst[0*stride] = cm[dst[0*stride] + v];
dst[1*stride] = cm[dst[1*stride] + v];
dst[2*stride] = cm[dst[2*stride] + v];
dst[3*stride] = cm[dst[3*stride] + v];
dst[4*stride] = cm[dst[4*stride] + v];
dst[5*stride] = cm[dst[5*stride] + v];
dst[6*stride] = cm[dst[6*stride] + v];
dst[7*stride] = cm[dst[7*stride] + v];
}
}
}
ip++; /* next column */
dst++;
}
}
