int main ()
{
*vecfloat++ = vec_andc(vecint[0], vecfloat[1]);
*vecfloat++ = vec_andc(vecfloat[0], vecint[1]);
*vecfloat++ = vec_vxor(vecint[0], vecfloat[1]);
*vecfloat++ = vec_vxor(vecfloat[0], vecint[1]);
*varpixel++ = vec_packpx(vecuint[0], vecuint[1]);
*varpixel++ = vec_vpkpx(vecuint[0], vecuint[1]);
*vecshort++ = vec_vmulosb(vecchar[0], vecchar[1]);
*vecint++ = vec_ld(var_int[0], longp[1]);
*vecint++ = vec_lde(var_int[0], longp[1]);
*vecint++ = vec_ldl(var_int[0], longp[1]);
*vecint++ = vec_lvewx(var_int[0], longp[1]);
*vecint++ = vec_unpackh(vecshort[0]);
*vecint++ = vec_unpackl(vecshort[0]);
*vecushort++ = vec_andc(vecshort[0], vecushort[1]);
*vecushort++ = vec_andc(vecushort[0], vecshort[1]);
*vecushort++ = vec_vxor(vecshort[0], vecushort[1]);
*vecushort++ = vec_vxor(vecushort[0], vecshort[1]);
*vecuint++ = vec_ld(var_int[0], ulongp[1]);
*vecuint++ = vec_lvx(var_int[0], ulongp[1]);
*vecuint++ = vec_vmsumubm(vecuchar[0], vecuchar[1], vecuint[2]);
*vecuchar++ = vec_xor(vecuchar[0], vecchar[1]);
return 0;
}
void
b()
{
z = vec_add (x, y);
/* Make sure the predicates accept correct argument types. */
int1 = vec_all_in (f, g);
int1 = vec_all_ge (f, g);
int1 = vec_all_eq (c, d);
int1 = vec_all_ne (s, t);
int1 = vec_any_eq (i, j);
int1 = vec_any_ge (f, g);
int1 = vec_all_ngt (f, g);
int1 = vec_any_ge (c, d);
int1 = vec_any_ge (s, t);
int1 = vec_any_ge (i, j);
int1 = vec_any_ge (c, d);
int1 = vec_any_ge (s, t);
int1 = vec_any_ge (i, j);
vec_mtvscr (i);
vec_dssall ();
s = (vector signed short) vec_mfvscr ();
vec_dss (3);
vec_dst (pi, int1 + int2, 3);
vec_dstst (pi, int1 + int2, 3);
vec_dststt (pi, int1 + int2, 3);
vec_dstt (pi, int1 + int2, 3);
uc = (vector unsigned char) vec_lvsl (int1 + 69, (signed int *) pi);
uc = (vector unsigned char) vec_lvsr (int1 + 69, (signed int *) pi);
c = vec_lde (int1, (signed char *) pi);
s = vec_lde (int1, (signed short *) pi);
i = vec_lde (int1, (signed int *) pi);
i = vec_ldl (int1, pi);
i = vec_ld (int1, pi);
vec_st (i, int2, pi);
vec_ste (c, int2, (signed char *) pi);
vec_ste (s, int2, (signed short *) pi);
vec_ste (i, int2, (signed int *) pi);
vec_stl (i, int2, pi);
}
void f22() {
*var_vec_u32++ = vec_ld(var_int[0], var_unsigned_long_ptr[1]);
*var_vec_u32++ = vec_lde(var_int[0], var_unsigned_long_ptr[1]);
*var_vec_u32++ = vec_ldl(var_int[0], var_unsigned_long_ptr[1]);
*var_vec_u32++ = vec_lvewx(var_int[0], var_unsigned_long_ptr[1]);
*var_vec_u32++ = vec_lvx(var_int[0], var_unsigned_long_ptr[1]);
*var_vec_u32++ = vec_lvxl(var_int[0], var_unsigned_long_ptr[1]);
}
void f13() {
*var_vec_s32++ = vec_ld(var_int[0], var_long_ptr[1]);
*var_vec_s32++ = vec_lde(var_int[0], var_long_ptr[1]);
*var_vec_s32++ = vec_ldl(var_int[0], var_long_ptr[1]);
*var_vec_s32++ = vec_lvewx(var_int[0], var_long_ptr[1]);
*var_vec_s32++ = vec_lvx(var_int[0], var_long_ptr[1]);
*var_vec_s32++ = vec_lvxl(var_int[0], var_long_ptr[1]);
}
static void test ()
{
vector unsigned char vuc;
vector signed char vsc;
vector unsigned short vus;
vector signed short vss;
vector unsigned int vui;
vector signed int vsi;
vector float vf;
init ();
vuc = vec_lde (9*1, (unsigned char *)svuc);
vsc = vec_lde (14*1, (signed char *)svsc);
vus = vec_lde (7*2, (unsigned short *)svus);
vss = vec_lde (1*2, (signed short *)svss);
vui = vec_lde (3*4, (unsigned int *)svui);
vsi = vec_lde (2*4, (signed int *)svsi);
vf = vec_lde (0*4, (float *)svf);
check (vec_extract (vuc, 9) == 9, "vuc");
check (vec_extract (vsc, 14) == 6, "vsc");
check (vec_extract (vus, 7) == 7, "vus");
check (vec_extract (vss, 1) == -3, "vss");
check (vec_extract (vui, 3) == 3, "vui");
check (vec_extract (vsi, 2) == 0, "vsi");
check (vec_extract (vf, 0) == 0.0, "vf");
}
int main (int argc, const char * argv[])
{
int i;
const float cf = 1.0;
vector float v;
const vector float cv = (vector float){1.0, 2.0, 3.0, 4.0};
vec_dst(&cv, i, 0);
v = vec_ld(0, &cv);
v = vec_lde(0, &cf);
vec_lvsl(0, &cf);
return 0;
}
static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
{
vec_s16 dc16;
vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
vec_s32 v_dc32;
LOAD_ZERO;
DECLARE_ALIGNED(16, int, dc);
int i;
dc = (block[0] + 32) >> 6;
block[0] = 0;
v_dc32 = vec_lde(0, &dc);
dc16 = VEC_SPLAT16((vec_s16)v_dc32, 1);
if (size == 4)
dc16 = VEC_SLD16(dc16, zero_s16v, 8);
dcplus = vec_packsu(dc16, zero_s16v);
dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
aligner = vec_lvsr(0, dst);
#if !HAVE_BIGENDIAN
aligner = vec_perm(aligner, zero_u8v, vcswapc());
#endif
dcplus = vec_perm(dcplus, dcplus, aligner);
dcminus = vec_perm(dcminus, dcminus, aligner);
for (i = 0; i < size; i += 4) {
v0 = vec_ld(0, dst+0*stride);
v1 = vec_ld(0, dst+1*stride);
v2 = vec_ld(0, dst+2*stride);
v3 = vec_ld(0, dst+3*stride);
v0 = vec_adds(v0, dcplus);
v1 = vec_adds(v1, dcplus);
v2 = vec_adds(v2, dcplus);
v3 = vec_adds(v3, dcplus);
v0 = vec_subs(v0, dcminus);
v1 = vec_subs(v1, dcminus);
v2 = vec_subs(v2, dcminus);
v3 = vec_subs(v3, dcminus);
vec_st(v0, 0, dst+0*stride);
vec_st(v1, 0, dst+1*stride);
vec_st(v2, 0, dst+2*stride);
vec_st(v3, 0, dst+3*stride);
dst += 4*stride;
}
}
static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
{
vec_s16 dc16;
vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
LOAD_ZERO;
DECLARE_ALIGNED(16, int, dc);
int i;
dc = (block[0] + 32) >> 6;
block[0] = 0;
dc16 = vec_splat((vec_s16) vec_lde(0, &dc), 1);
if (size == 4)
dc16 = vec_sld(dc16, zero_s16v, 8);
dcplus = vec_packsu(dc16, zero_s16v);
dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
aligner = vec_lvsr(0, dst);
dcplus = vec_perm(dcplus, dcplus, aligner);
dcminus = vec_perm(dcminus, dcminus, aligner);
for (i = 0; i < size; i += 4) {
v0 = vec_ld(0, dst+0*stride);
v1 = vec_ld(0, dst+1*stride);
v2 = vec_ld(0, dst+2*stride);
v3 = vec_ld(0, dst+3*stride);
v0 = vec_adds(v0, dcplus);
v1 = vec_adds(v1, dcplus);
v2 = vec_adds(v2, dcplus);
v3 = vec_adds(v3, dcplus);
v0 = vec_subs(v0, dcminus);
v1 = vec_subs(v1, dcminus);
v2 = vec_subs(v2, dcminus);
v3 = vec_subs(v3, dcminus);
vec_st(v0, 0, dst+0*stride);
vec_st(v1, 0, dst+1*stride);
vec_st(v2, 0, dst+2*stride);
vec_st(v3, 0, dst+3*stride);
dst += 4*stride;
}
}
/* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
* to preserve proper dst alignment. */
void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */,
int stride, int h, int x16, int y16, int rounder)
{
int i;
const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = {
(16 - x16) * (16 - y16), /* A */
(x16) * (16 - y16), /* B */
(16 - x16) * (y16), /* C */
(x16) * (y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero =
(const vector unsigned char) vec_splat_u8(0);
register const vector unsigned short vcsr8 =
(const vector unsigned short) vec_splat_u16(8);
register vector unsigned char dstv, dstv2, srcvB, srcvC, srcvD;
register vector unsigned short tempB, tempC, tempD;
unsigned long dst_odd = (unsigned long) dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long) src & 0x0000000F;
register vector unsigned short tempA =
vec_ld(0, (const unsigned short *) ABCD);
register vector unsigned short Av = vec_splat(tempA, 0);
register vector unsigned short Bv = vec_splat(tempA, 1);
register vector unsigned short Cv = vec_splat(tempA, 2);
register vector unsigned short Dv = vec_splat(tempA, 3);
register vector unsigned short rounderV =
vec_splat((vec_u16) vec_lde(0, &rounder_a), 0);
/* we'll be able to pick-up our 9 char elements at src from those
* 32 bytes we load the first batch here, as inside the loop we can
* reuse 'src + stride' from one iteration as the 'src' of the next. */
register vector unsigned char src_0 = vec_ld(0, src);
register vector unsigned char src_1 = vec_ld(16, src);
register vector unsigned char srcvA = vec_perm(src_0, src_1,
vec_lvsl(0, src));
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
else
srcvB = src_1;
srcvA = vec_mergeh(vczero, srcvA);
srcvB = vec_mergeh(vczero, srcvB);
for (i = 0; i < h; i++) {
dst_odd = (unsigned long) dst & 0x0000000F;
src_really_odd = (((unsigned long) src) + stride) & 0x0000000F;
dstv = vec_ld(0, dst);
/* We'll be able to pick-up our 9 char elements at src + stride from
* those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD
* as the next srcvA and srcvB. */
src_0 = vec_ld(stride + 0, src);
src_1 = vec_ld(stride + 16, src);
srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
else
srcvD = src_1;
srcvC = vec_mergeh(vczero, srcvC);
srcvD = vec_mergeh(vczero, srcvD);
/* OK, now we (finally) do the math :-)
* Those four instructions replace 32 int muls & 32 int adds.
* Isn't AltiVec nice? */
tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC);
srcvA = srcvC;
srcvB = srcvD;
tempD = vec_sr(tempD, vcsr8);
dstv2 = vec_pack(tempD, (vector unsigned short) vczero);
if (dst_odd)
dstv2 = vec_perm(dstv, dstv2, vcprm(0, 1, s0, s1));
else
dstv2 = vec_perm(dstv, dstv2, vcprm(s0, s1, 2, 3));
vec_st(dstv2, 0, dst);
dst += stride;
src += stride;
}
}
/* AltiVec version of dct_unquantize_h263
this code assumes `block' is 16 bytes-aligned */
static void dct_unquantize_h263_altivec(MpegEncContext *s,
DCTELEM *block, int n, int qscale)
{
int i, level, qmul, qadd;
int nCoeffs;
assert(s->block_last_index[n]>=0);
qadd = (qscale - 1) | 1;
qmul = qscale << 1;
if (s->mb_intra) {
if (!s->h263_aic) {
if (n < 4)
block[0] = block[0] * s->y_dc_scale;
else
block[0] = block[0] * s->c_dc_scale;
}else
qadd = 0;
i = 1;
nCoeffs= 63; //does not always use zigzag table
} else {
i = 0;
nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
}
{
register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
DECLARE_ALIGNED(16, short, qmul8) = qmul;
DECLARE_ALIGNED(16, short, qadd8) = qadd;
register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
register vector bool short blockv_null, blockv_neg;
register short backup_0 = block[0];
register int j = 0;
qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
nqaddv = vec_sub(vczero, qaddv);
#if 0 // block *is* 16 bytes-aligned, it seems.
// first make sure block[j] is 16 bytes-aligned
for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
level = block[j];
if (level) {
if (level < 0) {
level = level * qmul - qadd;
} else {
level = level * qmul + qadd;
}
block[j] = level;
}
}
#endif
// vectorize all the 16 bytes-aligned blocks
// of 8 elements
for(; (j + 7) <= nCoeffs ; j+=8) {
blockv = vec_ld(j << 1, block);
blockv_neg = vec_cmplt(blockv, vczero);
blockv_null = vec_cmpeq(blockv, vczero);
// choose between +qadd or -qadd as the third operand
temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
// multiply & add (block{i,i+7} * qmul [+-] qadd)
temp1 = vec_mladd(blockv, qmulv, temp1);
// put 0 where block[{i,i+7} used to have 0
blockv = vec_sel(temp1, blockv, blockv_null);
vec_st(blockv, j << 1, block);
}
// if nCoeffs isn't a multiple of 8, finish the job
// using good old scalar units.
// (we could do it using a truncated vector,
// but I'm not sure it's worth the hassle)
for(; j <= nCoeffs ; j++) {
level = block[j];
if (level) {
if (level < 0) {
level = level * qmul - qadd;
} else {
level = level * qmul + qadd;
}
block[j] = level;
}
}
if (i == 1) {
// cheat. this avoid special-casing the first iteration
block[0] = backup_0;
}
}
}
void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
{
POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
const DECLARE_ALIGNED_16(unsigned short, rounder_a) = rounder;
const DECLARE_ALIGNED_16(unsigned short, ABCD)[8] =
{
(16-x16)*(16-y16), /* A */
( x16)*(16-y16), /* B */
(16-x16)*( y16), /* C */
( x16)*( y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcsr8 = (const vector unsigned short)vec_splat_u16(8);
register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
int i;
unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
POWERPC_PERF_START_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
tempA = vec_ld(0, (unsigned short*)ABCD);
Av = vec_splat(tempA, 0);
Bv = vec_splat(tempA, 1);
Cv = vec_splat(tempA, 2);
Dv = vec_splat(tempA, 3);
rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);
// we'll be able to pick-up our 9 char elements
// at src from those 32 bytes
// we load the first batch here, as inside the loop
// we can re-use 'src+stride' from one iteration
// as the 'src' of the next.
src_0 = vec_ld(0, src);
src_1 = vec_ld(16, src);
srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));
if (src_really_odd != 0x0000000F) {
// if src & 0xF == 0xF, then (src+1) is properly aligned
// on the second vector.
srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
} else {
srcvB = src_1;
}
srcvA = vec_mergeh(vczero, srcvA);
srcvB = vec_mergeh(vczero, srcvB);
for(i=0; i<h; i++) {
dst_odd = (unsigned long)dst & 0x0000000F;
src_really_odd = (((unsigned long)src) + stride) & 0x0000000F;
dstv = vec_ld(0, dst);
// we we'll be able to pick-up our 9 char elements
// at src + stride from those 32 bytes
// then reuse the resulting 2 vectors srvcC and srcvD
// as the next srcvA and srcvB
src_0 = vec_ld(stride + 0, src);
src_1 = vec_ld(stride + 16, src);
srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
if (src_really_odd != 0x0000000F) {
// if src & 0xF == 0xF, then (src+1) is properly aligned
// on the second vector.
srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
} else {
srcvD = src_1;
}
srcvC = vec_mergeh(vczero, srcvC);
srcvD = vec_mergeh(vczero, srcvD);
// OK, now we (finally) do the math :-)
// those four instructions replaces 32 int muls & 32 int adds.
// isn't AltiVec nice ?
tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC);
srcvA = srcvC;
srcvB = srcvD;
tempD = vec_sr(tempD, vcsr8);
dstv2 = vec_pack(tempD, (vector unsigned short)vczero);
if (dst_odd) {
dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1));
} else {
dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3));
}
vec_st(dstv2, 0, dst);
dst += stride;
src += stride;
}
POWERPC_PERF_STOP_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
}
