void GGLAssembler::component_sat(const component_t& v)
{
const int one = ((1<<v.size())-1)<<v.l;
CMP(AL, v.reg, imm( 1<<v.h ));
if (isValidImmediate(one)) {
MOV(HS, 0, v.reg, imm( one ));
} else if (isValidImmediate(~one)) {
MVN(HS, 0, v.reg, imm( ~one ));
} else {
MOV(HS, 0, v.reg, imm( 1<<v.h ));
SUB(HS, 0, v.reg, v.reg, imm( 1<<v.l ));
}
}
void GGLAssembler::extract(integer_t& d, int s, int h, int l, int bits)
{
const int maskLen = h-l;
#ifdef __mips__
assert(maskLen<=11);
#else
assert(maskLen<=8);
#endif
assert(h);
#if __ARM_ARCH__ >= 7
const int mask = (1<<maskLen)-1;
if ((h == bits) && !l && (s != d.reg)) {
MOV(AL, 0, d.reg, s); // component = packed;
} else if ((h == bits) && l) {
MOV(AL, 0, d.reg, reg_imm(s, LSR, l)); // component = packed >> l;
} else if (!l && isValidImmediate(mask)) {
AND(AL, 0, d.reg, s, imm(mask)); // component = packed & mask;
} else if (!l && isValidImmediate(~mask)) {
BIC(AL, 0, d.reg, s, imm(~mask)); // component = packed & mask;
} else {
UBFX(AL, d.reg, s, l, maskLen); // component = (packed & mask) >> l;
}
#else
if (h != bits) {
const int mask = ((1<<maskLen)-1) << l;
if (isValidImmediate(mask)) {
AND(AL, 0, d.reg, s, imm(mask)); // component = packed & mask;
} else if (isValidImmediate(~mask)) {
BIC(AL, 0, d.reg, s, imm(~mask)); // component = packed & mask;
} else {
MOV(AL, 0, d.reg, reg_imm(s, LSL, 32-h));
l += 32-h;
h = 32;
}
s = d.reg;
}
if (l) {
MOV(AL, 0, d.reg, reg_imm(s, LSR, l)); // component = packed >> l;
s = d.reg;
}
if (s != d.reg) {
MOV(AL, 0, d.reg, s);
}
#endif
d.s = maskLen;
}
void GGLAssembler::downshift(
pixel_t& d, int component, component_t s, const reg_t& dither)
{
const needs_t& needs = mBuilderContext.needs;
Scratch scratches(registerFile());
int sh = s.h;
int sl = s.l;
int maskHiBits = (sh!=32) ? ((s.flags & CLEAR_HI)?1:0) : 0;
int maskLoBits = (sl!=0) ? ((s.flags & CLEAR_LO)?1:0) : 0;
int sbits = sh - sl;
int dh = d.format.c[component].h;
int dl = d.format.c[component].l;
int dbits = dh - dl;
int dithering = 0;
ALOGE_IF(sbits<dbits, "sbits (%d) < dbits (%d) in downshift", sbits, dbits);
if (sbits>dbits) {
// see if we need to dither
dithering = mDithering;
}
int ireg = d.reg;
if (!(d.flags & FIRST)) {
if (s.flags & CORRUPTIBLE) {
ireg = s.reg;
} else {
ireg = scratches.obtain();
}
}
d.flags &= ~FIRST;
if (maskHiBits) {
// we need to mask the high bits (and possibly the lowbits too)
// and we might be able to use immediate mask.
if (!dithering) {
// we don't do this if we only have maskLoBits because we can
// do it more efficiently below (in the case where dl=0)
const int offset = sh - dbits;
if (dbits<=8 && offset >= 0) {
const uint32_t mask = ((1<<dbits)-1) << offset;
if (isValidImmediate(mask) || isValidImmediate(~mask)) {
build_and_immediate(ireg, s.reg, mask, 32);
sl = offset;
s.reg = ireg;
sbits = dbits;
maskLoBits = maskHiBits = 0;
}
}
} else {
// in the dithering case though, we need to preserve the lower bits
const uint32_t mask = ((1<<sbits)-1) << sl;
if (isValidImmediate(mask) || isValidImmediate(~mask)) {
build_and_immediate(ireg, s.reg, mask, 32);
s.reg = ireg;
maskLoBits = maskHiBits = 0;
}
}
}
// XXX: we could special case (maskHiBits & !maskLoBits)
// like we do for maskLoBits below, but it happens very rarely
// that we have maskHiBits only and the conditions necessary to lead
// to better code (like doing d |= s << 24)
if (maskHiBits) {
MOV(AL, 0, ireg, reg_imm(s.reg, LSL, 32-sh));
sl += 32-sh;
sh = 32;
s.reg = ireg;
maskHiBits = 0;
}
// Downsampling should be performed as follows:
// V * ((1<<dbits)-1) / ((1<<sbits)-1)
// V * [(1<<dbits)/((1<<sbits)-1) - 1/((1<<sbits)-1)]
// V * [1/((1<<sbits)-1)>>dbits - 1/((1<<sbits)-1)]
// V/((1<<(sbits-dbits))-(1>>dbits)) - (V>>sbits)/((1<<sbits)-1)>>sbits
// V/((1<<(sbits-dbits))-(1>>dbits)) - (V>>sbits)/(1-(1>>sbits))
//
// By approximating (1>>dbits) and (1>>sbits) to 0:
//
// V>>(sbits-dbits) - V>>sbits
//
// A good approximation is V>>(sbits-dbits),
// but better one (needed for dithering) is:
//
// (V>>(sbits-dbits)<<sbits - V)>>sbits
// (V<<dbits - V)>>sbits
// (V - V>>dbits)>>(sbits-dbits)
// Dithering is done here
if (dithering) {
comment("dithering");
if (sl) {
MOV(AL, 0, ireg, reg_imm(s.reg, LSR, sl));
sh -= sl;
sl = 0;
s.reg = ireg;
}
// scaling (V-V>>dbits)
SUB(AL, 0, ireg, s.reg, reg_imm(s.reg, LSR, dbits));
const int shift = (GGL_DITHER_BITS - (sbits-dbits));
if (shift>0) ADD(AL, 0, ireg, ireg, reg_imm(dither.reg, LSR, shift));
else if (shift<0) ADD(AL, 0, ireg, ireg, reg_imm(dither.reg, LSL,-shift));
else ADD(AL, 0, ireg, ireg, dither.reg);
s.reg = ireg;
}
if ((maskLoBits|dithering) && (sh > dbits)) {
int shift = sh-dbits;
if (dl) {
MOV(AL, 0, ireg, reg_imm(s.reg, LSR, shift));
if (ireg == d.reg) {
MOV(AL, 0, d.reg, reg_imm(ireg, LSL, dl));
} else {
ORR(AL, 0, d.reg, d.reg, reg_imm(ireg, LSL, dl));
}
} else {
if (ireg == d.reg) {
MOV(AL, 0, d.reg, reg_imm(s.reg, LSR, shift));
} else {
ORR(AL, 0, d.reg, d.reg, reg_imm(s.reg, LSR, shift));
}
}
} else {
int shift = sh-dh;
if (shift>0) {
if (ireg == d.reg) {
MOV(AL, 0, d.reg, reg_imm(s.reg, LSR, shift));
} else {
ORR(AL, 0, d.reg, d.reg, reg_imm(s.reg, LSR, shift));
}
} else if (shift<0) {
if (ireg == d.reg) {
MOV(AL, 0, d.reg, reg_imm(s.reg, LSL, -shift));
} else {
ORR(AL, 0, d.reg, d.reg, reg_imm(s.reg, LSL, -shift));
}
} else {
if (ireg == d.reg) {
if (s.reg != d.reg) {
MOV(AL, 0, d.reg, s.reg);
}
} else {
ORR(AL, 0, d.reg, d.reg, s.reg);
}
}
}
}