void add_pixels_clamped_mvi(const int16_t *block, uint8_t *pixels,
ptrdiff_t line_size)
{
int h = 8;
/* Keep this function a leaf function by generating the constants
manually (mainly for the hack value ;-). */
uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
uint64_t signmask = zap(-1, 0x33);
signmask ^= signmask >> 1; /* 0x8000800080008000 */
do {
uint64_t shorts0, pix0, signs0;
uint64_t shorts1, pix1, signs1;
shorts0 = ldq(block);
shorts1 = ldq(block + 4);
pix0 = unpkbw(ldl(pixels));
/* Signed subword add (MMX paddw). */
signs0 = shorts0 & signmask;
shorts0 &= ~signmask;
shorts0 += pix0;
shorts0 ^= signs0;
/* Clamp. */
shorts0 = maxsw4(shorts0, 0);
shorts0 = minsw4(shorts0, clampmask);
/* Next 4. */
pix1 = unpkbw(ldl(pixels + 4));
signs1 = shorts1 & signmask;
shorts1 &= ~signmask;
shorts1 += pix1;
shorts1 ^= signs1;
shorts1 = maxsw4(shorts1, 0);
shorts1 = minsw4(shorts1, clampmask);
stl(pkwb(shorts0), pixels);
stl(pkwb(shorts1), pixels + 4);
pixels += line_size;
block += 8;
} while (--h);
}
static void dct_unquantize_h263_inter_axp(MpegEncContext *s, DCTELEM *block,
int n, int qscale)
{
int i, n_coeffs;
uint64_t qmul, qadd;
uint64_t correction;
qadd = WORD_VEC((qscale - 1) | 1);
qmul = qscale << 1;
/* This mask kills spill from negative subwords to the next subword. */
correction = WORD_VEC((qmul - 1) + 1); /* multiplication / addition */
n_coeffs = s->intra_scantable.raster_end[s->block_last_index[n]];
for(i = 0; i <= n_coeffs; block += 4, i += 4) {
uint64_t levels, negmask, zeros, add;
levels = ldq(block);
if (levels == 0)
continue;
#ifdef __alpha_max__
/* I don't think the speed difference justifies runtime
detection. */
negmask = maxsw4(levels, -1); /* negative -> ffff (-1) */
negmask = minsw4(negmask, 0); /* positive -> 0000 (0) */
#else
negmask = cmpbge(WORD_VEC(0x7fff), levels);
negmask &= (negmask >> 1) | (1 << 7);
negmask = zap(-1, negmask);
#endif
zeros = cmpbge(0, levels);
zeros &= zeros >> 1;
/* zeros |= zeros << 1 is not needed since qadd <= 255, so
zapping the lower byte suffices. */
levels *= qmul;
levels -= correction & (negmask << 16);
/* Negate qadd for negative levels. */
add = qadd ^ negmask;
add += WORD_VEC(0x0001) & negmask;
/* Set qadd to 0 for levels == 0. */
add = zap(add, zeros);
levels += add;
stq(levels, block);
}
}
/* These functions were the base for the optimized assembler routines,
and remain here for documentation purposes. */
static void put_pixels_clamped_mvi(const int16_t *block, uint8_t *pixels,
ptrdiff_t line_size)
{
int i = 8;
uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
do {
uint64_t shorts0, shorts1;
shorts0 = ldq(block);
shorts0 = maxsw4(shorts0, 0);
shorts0 = minsw4(shorts0, clampmask);
stl(pkwb(shorts0), pixels);
shorts1 = ldq(block + 4);
shorts1 = maxsw4(shorts1, 0);
shorts1 = minsw4(shorts1, clampmask);
stl(pkwb(shorts1), pixels + 4);
pixels += line_size;
block += 8;
} while (--i);
}
static void dct_unquantize_h263_axp(DCTELEM *block, int n_coeffs,
uint64_t qscale, uint64_t qadd)
{
uint64_t qmul = qscale << 1;
uint64_t correction = WORD_VEC(qmul * 255 >> 8);
int i;
qadd = WORD_VEC(qadd);
for(i = 0; i <= n_coeffs; block += 4, i += 4)
{
uint64_t levels, negmask, zeros, add, sub;
levels = ldq(block);
if (levels == 0)
continue;
#ifdef __alpha_max__
/* I don't think the speed difference justifies runtime
detection. */
negmask = maxsw4(levels, -1); /* negative -> ffff (-1) */
negmask = minsw4(negmask, 0); /* positive -> 0000 (0) */
#else
negmask = cmpbge(WORD_VEC(0x7fff), levels);
negmask &= (negmask >> 1) | (1 << 7);
negmask = zap(-1, negmask);
#endif
zeros = cmpbge(0, levels);
zeros &= zeros >> 1;
/* zeros |= zeros << 1 is not needed since qadd <= 255, so
zapping the lower byte suffices. */
levels *= qmul;
levels -= correction & (negmask << 16);
add = qadd & ~negmask;
sub = qadd & negmask;
/* Set qadd to 0 for levels == 0. */
add = zap(add, zeros);
levels += add;
levels -= sub;
stq(levels, block);
}
}
static void dct_unquantize_h263_intra_axp(MpegEncContext *s, DCTELEM *block,
int n, int qscale)
{
int i, n_coeffs;
uint64_t qmul, qadd;
uint64_t correction;
DCTELEM *orig_block = block;
DCTELEM block0; /* might not be used uninitialized */
qadd = WORD_VEC((qscale - 1) | 1);
qmul = qscale << 1;
/* This mask kills spill from negative subwords to the next subword. */
correction = WORD_VEC((qmul - 1) + 1); /* multiplication / addition */
if (!s->h263_aic) {
if (n < 4)
block0 = block[0] * s->y_dc_scale;
else
block0 = block[0] * s->c_dc_scale;
} else {
qadd = 0;
}
n_coeffs = 63; // does not always use zigzag table
for(i = 0; i <= n_coeffs; block += 4, i += 4) {
uint64_t levels, negmask, zeros, add;
levels = ldq(block);
if (levels == 0)
continue;
#ifdef __alpha_max__
/* I don't think the speed difference justifies runtime
detection. */
negmask = maxsw4(levels, -1); /* negative -> ffff (-1) */
negmask = minsw4(negmask, 0); /* positive -> 0000 (0) */
#else
negmask = cmpbge(WORD_VEC(0x7fff), levels);
negmask &= (negmask >> 1) | (1 << 7);
negmask = zap(-1, negmask);
#endif
zeros = cmpbge(0, levels);
zeros &= zeros >> 1;
/* zeros |= zeros << 1 is not needed since qadd <= 255, so
zapping the lower byte suffices. */
levels *= qmul;
levels -= correction & (negmask << 16);
/* Negate qadd for negative levels. */
add = qadd ^ negmask;
add += WORD_VEC(0x0001) & negmask;
/* Set qadd to 0 for levels == 0. */
add = zap(add, zeros);
levels += add;
stq(levels, block);
}
if (s->mb_intra && !s->h263_aic)
orig_block[0] = block0;
}