82
static int32_t scalarproduct_int16_altivec(const int16_t * v1, const int16_t * v2, int order, const int shift)
81
static int32_t scalarproduct_int16_altivec(const int16_t *v1, const int16_t *v2,
86
register vec_s16 vec1, *pv;
87
register vec_s16 vec1;
87
88
register vec_s32 res = vec_splat_s32(0), t;
88
register vec_u32 shifts;
92
if(shift & 0x10) shifts = vec_add(shifts, vec_sl(vec_splat_u32(0x08), vec_splat_u32(0x1)));
93
if(shift & 0x08) shifts = vec_add(shifts, vec_splat_u32(0x08));
94
if(shift & 0x04) shifts = vec_add(shifts, vec_splat_u32(0x04));
95
if(shift & 0x02) shifts = vec_add(shifts, vec_splat_u32(0x02));
96
if(shift & 0x01) shifts = vec_add(shifts, vec_splat_u32(0x01));
98
91
for(i = 0; i < order; i += 8){
92
pv = (const vec_s16*)v1;
100
93
vec1 = vec_perm(pv[0], pv[1], vec_lvsl(0, v1));
101
94
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
102
t = vec_sr(t, shifts);
103
95
res = vec_sums(t, res);
115
107
vec_s16 *pv1 = (vec_s16*)v1;
116
vec_s16 *pv2 = (vec_s16*)v2;
117
vec_s16 *pv3 = (vec_s16*)v3;
118
108
register vec_s16 muls = {mul,mul,mul,mul,mul,mul,mul,mul};
119
register vec_s16 t0, t1, i0, i1;
120
register vec_s16 i2 = pv2[0], i3 = pv3[0];
109
register vec_s16 t0, t1, i0, i1, i4;
110
register vec_s16 i2 = vec_ld(0, v2), i3 = vec_ld(0, v3);
121
111
register vec_s32 res = zero_s32v;
122
112
register vec_u8 align = vec_lvsl(0, v2);
126
t0 = vec_perm(i2, pv2[1], align);
128
t1 = vec_perm(pv2[1], i2, align);
117
t0 = vec_perm(i2, i1, align);
119
t1 = vec_perm(i1, i2, align);
131
122
res = vec_msum(t0, i0, res);
132
123
res = vec_msum(t1, i1, res);
133
t0 = vec_perm(i3, pv3[1], align);
135
t1 = vec_perm(pv3[1], i3, align);
125
t0 = vec_perm(i3, i4, align);
127
t1 = vec_perm(i4, i3, align);
136
128
pv1[0] = vec_mladd(t0, muls, i0);
137
129
pv1[1] = vec_mladd(t1, muls, i1);
141
133
} while(--order);
142
134
res = vec_splat(vec_sums(res, zero_s32v), 3);
143
135
vec_ste(res, 0, &ires);
147
void int_init_altivec(DSPContext* c, AVCodecContext *avctx)
139
void ff_int_init_altivec(DSPContext* c, AVCodecContext *avctx)
149
141
c->ssd_int8_vs_int16 = ssd_int8_vs_int16_altivec;
150
142
c->scalarproduct_int16 = scalarproduct_int16_altivec;