void test_vtrnQu16 (void)
{
uint16x8x2_t out_uint16x8x2_t;
uint16x8_t arg0_uint16x8_t;
uint16x8_t arg1_uint16x8_t;
out_uint16x8x2_t = vtrnq_u16 (arg0_uint16x8_t, arg1_uint16x8_t);
}
uint16x8x2_t test_vtrnq_u16(uint16x8_t a, uint16x8_t b) {
// CHECK-LABEL: test_vtrnq_u16
return vtrnq_u16(a, b);
// CHECK: trn1 {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
// CHECK: trn2 {{v[0-9]+}}.8h, {{v[0-9]+}}.8h
}
static
void
neon_w32_split_4_32_multiply_region(gf_t *gf, uint32_t *src, uint32_t *dst,
uint32_t *d_end, uint8_t btable[8][4][16],
uint32_t val, int xor, int altmap)
{
int i, j;
#ifdef ARCH_AARCH64
uint8x16_t tables[8][4];
#else
uint8x8x2_t tables[8][4];
#endif
uint32x4_t v0, v1, v2, v3, s0, s1, s2, s3;
uint8x16_t p0, p1, p2, p3, si, mask1;
uint16x8x2_t r0, r1;
uint8x16x2_t q0, q1;
for (i = 0; i < 8; i++) {
for (j = 0; j < 4; j++) {
#ifdef ARCH_AARCH64
tables[i][j] = vld1q_u8(btable[i][j]);
#else
tables[i][j].val[0] = vld1_u8(btable[i][j]);
tables[i][j].val[1] = vld1_u8(btable[i][j] + 8);
#endif
}
}
mask1 = vdupq_n_u8(0xf);
while (dst < d_end) {
v0 = vld1q_u32(src); src += 4;
v1 = vld1q_u32(src); src += 4;
v2 = vld1q_u32(src); src += 4;
v3 = vld1q_u32(src); src += 4;
if (altmap) {
q0.val[0] = vreinterpretq_u8_u32(v0);
q0.val[1] = vreinterpretq_u8_u32(v1);
q1.val[0] = vreinterpretq_u8_u32(v2);
q1.val[1] = vreinterpretq_u8_u32(v3);
} else {
r0 = vtrnq_u16(vreinterpretq_u16_u32(v0), vreinterpretq_u16_u32(v2));
r1 = vtrnq_u16(vreinterpretq_u16_u32(v1), vreinterpretq_u16_u32(v3));
q0 = vtrnq_u8(vreinterpretq_u8_u16(r0.val[0]),
vreinterpretq_u8_u16(r1.val[0]));
q1 = vtrnq_u8(vreinterpretq_u8_u16(r0.val[1]),
vreinterpretq_u8_u16(r1.val[1]));
}
si = vandq_u8(q0.val[0], mask1);
p0 = vqtbl1q_u8(tables[0][0], si);
p1 = vqtbl1q_u8(tables[0][1], si);
p2 = vqtbl1q_u8(tables[0][2], si);
p3 = vqtbl1q_u8(tables[0][3], si);
si = vshrq_n_u8(q0.val[0], 4);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[1][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[1][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[1][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[1][3], si));
si = vandq_u8(q0.val[1], mask1);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[2][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[2][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[2][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[2][3], si));
si = vshrq_n_u8(q0.val[1], 4);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[3][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[3][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[3][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[3][3], si));
si = vandq_u8(q1.val[0], mask1);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[4][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[4][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[4][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[4][3], si));
si = vshrq_n_u8(q1.val[0], 4);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[5][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[5][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[5][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[5][3], si));
si = vandq_u8(q1.val[1], mask1);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[6][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[6][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[6][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[6][3], si));
si = vshrq_n_u8(q1.val[1], 4);
p0 = veorq_u8(p0, vqtbl1q_u8(tables[7][0], si));
p1 = veorq_u8(p1, vqtbl1q_u8(tables[7][1], si));
p2 = veorq_u8(p2, vqtbl1q_u8(tables[7][2], si));
p3 = veorq_u8(p3, vqtbl1q_u8(tables[7][3], si));
if (altmap) {
s0 = vreinterpretq_u32_u8(p0);
s1 = vreinterpretq_u32_u8(p1);
s2 = vreinterpretq_u32_u8(p2);
s3 = vreinterpretq_u32_u8(p3);
} else {
q0 = vtrnq_u8(p0, p1);
q1 = vtrnq_u8(p2, p3);
r0 = vtrnq_u16(vreinterpretq_u16_u8(q0.val[0]),
vreinterpretq_u16_u8(q1.val[0]));
r1 = vtrnq_u16(vreinterpretq_u16_u8(q0.val[1]),
vreinterpretq_u16_u8(q1.val[1]));
s0 = vreinterpretq_u32_u16(r0.val[0]);
s1 = vreinterpretq_u32_u16(r1.val[0]);
s2 = vreinterpretq_u32_u16(r0.val[1]);
s3 = vreinterpretq_u32_u16(r1.val[1]);
}
if (xor) {
v0 = vld1q_u32(dst);
v1 = vld1q_u32(dst + 4);
v2 = vld1q_u32(dst + 8);
v3 = vld1q_u32(dst + 12);
s0 = veorq_u32(s0, v0);
s1 = veorq_u32(s1, v1);
s2 = veorq_u32(s2, v2);
s3 = veorq_u32(s3, v3);
}
vst1q_u32(dst, s0);
vst1q_u32(dst + 4, s1);
vst1q_u32(dst + 8, s2);
vst1q_u32(dst + 12, s3);
dst += 16;
}
}
void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, // unused
int y_step_q4, // unused
int w, int h) {
int width;
const uint8_t *s, *psrc;
uint8_t *d, *pdst;
uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
uint8x16_t q12u8, q13u8, q14u8, q15u8;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16;
uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
int16x8_t q0s16;
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
uint16x8x2_t q0x2u16;
uint8x8x2_t d0x2u8, d1x2u8;
uint32x2x2_t d0x2u32;
uint16x4x2_t d0x2u16, d1x2u16;
uint32x4x2_t q0x2u32;
assert(x_step_q4 == 16);
(void)x_step_q4;
(void)y_step_q4;
(void)filter_y;
q0s16 = vld1q_s16(filter_x);
src -= 3; // adjust for taps
for (; h > 0; h -= 4, src += src_stride * 4,
dst += dst_stride * 4) { // loop_horiz_v
s = src;
d24u8 = vld1_u8(s);
s += src_stride;
d25u8 = vld1_u8(s);
s += src_stride;
d26u8 = vld1_u8(s);
s += src_stride;
d27u8 = vld1_u8(s);
q12u8 = vcombine_u8(d24u8, d25u8);
q13u8 = vcombine_u8(d26u8, d27u8);
q0x2u16 =
vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
d0x2u8 = vtrn_u8(d24u8, d25u8);
d1x2u8 = vtrn_u8(d26u8, d27u8);
__builtin_prefetch(src + src_stride * 4);
__builtin_prefetch(src + src_stride * 5);
__builtin_prefetch(src + src_stride * 6);
q8u16 = vmovl_u8(d0x2u8.val[0]);
q9u16 = vmovl_u8(d0x2u8.val[1]);
q10u16 = vmovl_u8(d1x2u8.val[0]);
q11u16 = vmovl_u8(d1x2u8.val[1]);
d16u16 = vget_low_u16(q8u16);
d17u16 = vget_high_u16(q8u16);
d18u16 = vget_low_u16(q9u16);
d19u16 = vget_high_u16(q9u16);
q8u16 = vcombine_u16(d16u16, d18u16); // vswp 17 18
q9u16 = vcombine_u16(d17u16, d19u16);
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16)); // vmov 23 21
for (width = w, psrc = src + 7, pdst = dst; width > 0;
width -= 4, psrc += 4, pdst += 4) { // loop_horiz
s = psrc;
d28u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d29u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d31u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d30u32 = vld1_dup_u32((const uint32_t *)s);
__builtin_prefetch(psrc + 64);
d0x2u16 =
vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
d1x2u16 =
vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]), // d28
vreinterpret_u8_u16(d1x2u16.val[0])); // d29
d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]), // d31
vreinterpret_u8_u16(d1x2u16.val[1])); // d30
__builtin_prefetch(psrc + 64 + src_stride);
q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
q0x2u32 =
vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
q12u16 = vmovl_u8(d28u8);
q13u16 = vmovl_u8(d29u8);
__builtin_prefetch(psrc + 64 + src_stride * 2);
d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
d23s16, d24s16, q0s16);
q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
d24s16, d26s16, q0s16);
q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
d26s16, d27s16, q0s16);
q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
d27s16, d25s16, q0s16);
__builtin_prefetch(psrc + 60 + src_stride * 3);
d2u16 = vqrshrun_n_s32(q1s32, 7);
d3u16 = vqrshrun_n_s32(q2s32, 7);
d4u16 = vqrshrun_n_s32(q14s32, 7);
d5u16 = vqrshrun_n_s32(q15s32, 7);
q1u16 = vcombine_u16(d2u16, d3u16);
q2u16 = vcombine_u16(d4u16, d5u16);
d2u8 = vqmovn_u16(q1u16);
d3u8 = vqmovn_u16(q2u16);
d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
vreinterpret_u32_u16(d0x2u16.val[1]));
d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
vreinterpret_u8_u32(d0x2u32.val[1]));
d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
d = pdst;
vst1_lane_u32((uint32_t *)d, d2u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d2u32, 1);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 1);
q8u16 = q9u16;
d20s16 = d23s16;
q11u16 = q12u16;
q9u16 = q13u16;
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
}
}
return;
}
void vp8_mbloop_filter_vertical_edge_uv_neon(
unsigned char *u,
int pitch,
unsigned char blimit,
unsigned char limit,
unsigned char thresh,
unsigned char *v) {
unsigned char *us, *ud;
unsigned char *vs, *vd;
uint8x16_t qblimit, qlimit, qthresh, q3, q4;
uint8x16_t q5, q6, q7, q8, q9, q10;
uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
uint8x8_t d15, d16, d17, d18, d19, d20, d21;
uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
qblimit = vdupq_n_u8(blimit);
qlimit = vdupq_n_u8(limit);
qthresh = vdupq_n_u8(thresh);
us = u - 4;
vs = v - 4;
d6 = vld1_u8(us);
us += pitch;
d7 = vld1_u8(vs);
vs += pitch;
d8 = vld1_u8(us);
us += pitch;
d9 = vld1_u8(vs);
vs += pitch;
d10 = vld1_u8(us);
us += pitch;
d11 = vld1_u8(vs);
vs += pitch;
d12 = vld1_u8(us);
us += pitch;
d13 = vld1_u8(vs);
vs += pitch;
d14 = vld1_u8(us);
us += pitch;
d15 = vld1_u8(vs);
vs += pitch;
d16 = vld1_u8(us);
us += pitch;
d17 = vld1_u8(vs);
vs += pitch;
d18 = vld1_u8(us);
us += pitch;
d19 = vld1_u8(vs);
vs += pitch;
d20 = vld1_u8(us);
d21 = vld1_u8(vs);
q3 = vcombine_u8(d6, d7);
q4 = vcombine_u8(d8, d9);
q5 = vcombine_u8(d10, d11);
q6 = vcombine_u8(d12, d13);
q7 = vcombine_u8(d14, d15);
q8 = vcombine_u8(d16, d17);
q9 = vcombine_u8(d18, d19);
q10 = vcombine_u8(d20, d21);
q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
vreinterpretq_u16_u32(q2tmp2.val[0]));
q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
vreinterpretq_u16_u32(q2tmp3.val[0]));
q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
vreinterpretq_u16_u32(q2tmp2.val[1]));
q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
vreinterpretq_u16_u32(q2tmp3.val[1]));
q2tmp8 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
vreinterpretq_u8_u16(q2tmp5.val[0]));
q2tmp9 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
vreinterpretq_u8_u16(q2tmp5.val[1]));
q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
vreinterpretq_u8_u16(q2tmp7.val[0]));
q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
vreinterpretq_u8_u16(q2tmp7.val[1]));
q3 = q2tmp8.val[0];
q4 = q2tmp8.val[1];
q5 = q2tmp9.val[0];
q6 = q2tmp9.val[1];
q7 = q2tmp10.val[0];
q8 = q2tmp10.val[1];
q9 = q2tmp11.val[0];
q10 = q2tmp11.val[1];
vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
q5, q6, q7, q8, q9, q10,
&q4, &q5, &q6, &q7, &q8, &q9);
q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
vreinterpretq_u16_u32(q2tmp2.val[0]));
q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
vreinterpretq_u16_u32(q2tmp3.val[0]));
q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
vreinterpretq_u16_u32(q2tmp2.val[1]));
q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
vreinterpretq_u16_u32(q2tmp3.val[1]));
q2tmp8 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
vreinterpretq_u8_u16(q2tmp5.val[0]));
q2tmp9 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
vreinterpretq_u8_u16(q2tmp5.val[1]));
q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
vreinterpretq_u8_u16(q2tmp7.val[0]));
q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
vreinterpretq_u8_u16(q2tmp7.val[1]));
q3 = q2tmp8.val[0];
q4 = q2tmp8.val[1];
q5 = q2tmp9.val[0];
q6 = q2tmp9.val[1];
q7 = q2tmp10.val[0];
q8 = q2tmp10.val[1];
q9 = q2tmp11.val[0];
q10 = q2tmp11.val[1];
ud = u - 4;
vst1_u8(ud, vget_low_u8(q3));
ud += pitch;
vst1_u8(ud, vget_low_u8(q4));
ud += pitch;
vst1_u8(ud, vget_low_u8(q5));
ud += pitch;
vst1_u8(ud, vget_low_u8(q6));
ud += pitch;
vst1_u8(ud, vget_low_u8(q7));
ud += pitch;
vst1_u8(ud, vget_low_u8(q8));
ud += pitch;
vst1_u8(ud, vget_low_u8(q9));
ud += pitch;
vst1_u8(ud, vget_low_u8(q10));
vd = v - 4;
vst1_u8(vd, vget_high_u8(q3));
vd += pitch;
vst1_u8(vd, vget_high_u8(q4));
vd += pitch;
vst1_u8(vd, vget_high_u8(q5));
vd += pitch;
vst1_u8(vd, vget_high_u8(q6));
vd += pitch;
vst1_u8(vd, vget_high_u8(q7));
vd += pitch;
vst1_u8(vd, vget_high_u8(q8));
vd += pitch;
vst1_u8(vd, vget_high_u8(q9));
vd += pitch;
vst1_u8(vd, vget_high_u8(q10));
return;
}