/* Routine optimized for shuffling a buffer for a type size of 8 bytes. */
shuffle8_neon(uint8_t
* const dest,
const uint8_t* const src,
const size_t vectorizable_elements,
const size_t total_elements
)
{
size_t i, j, k, l;
static const size_t bytesoftype = 8;
uint8x8x2_t r0[4];
uint16x4x2_t r1[4];
uint32x2x2_t r2[4];
for(
i = 0, k = 0;
i<vectorizable_elements*bytesoftype;
i += 64, k++) {
/* Load and interleave groups of 8 bytes (64 bytes) to the structure r0 */
for(
j = 0;
j < 4; j++) {
r0[j] =
vzip_u8(vld1_u8(src + i + (2 * j) * 8), vld1_u8(src + i + (2 * j + 1) * 8)
);
}
/* Interleave 16 bytes */
for(
j = 0;
j < 2; j++) {
for(
l = 0;
l < 2; l++) {
r1[j*2+l] =
vzip_u16(vreinterpret_u16_u8(r0[j * 2].val[l]), vreinterpret_u16_u8(r0[j * 2 + 1].val[l])
);
}
}
/* Interleave 32 bytes */
for(
j = 0;
j < 2; j++) {
for(
l = 0;
l < 2; l++) {
r2[j*2+l] =
vzip_u32(vreinterpret_u32_u16(r1[j].val[l]), vreinterpret_u32_u16(r1[j + 2].val[l])
);
}
}
/* Store the results in the destination vector */
for(
j = 0;
j < 4; j++) {
for(
l = 0;
l < 2; l++) {
vst1_u8(dest
+ k*8 + (j*2+l)*total_elements,
vreinterpret_u8_u32(r2[j]
.val[l]));
}
}
}
}
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;
}
/* Routine optimized for unshuffling a buffer for a type size of 16 bytes. */
unshuffle16_neon(uint8_t
* const dest,
const uint8_t* const src,
const size_t vectorizable_elements,
const size_t total_elements
)
{
size_t i, j, k, l, m;
static const size_t bytesoftype = 16;
uint8x8x2_t r0[8];
uint16x4x2_t r1[8];
uint32x2x2_t r2[8];
for(
i = 0, k = 0;
i<vectorizable_elements*bytesoftype;
i += 128, k++) {
/* Load and interleave groups of 16 bytes (128 bytes) to the structure r0*/
for(
j = 0;
j < 8; j++) {
r0[j] =
vzip_u8(vld1_u8(src + (2 * j) * total_elements + k * 8), vld1_u8(src + (2 * j + 1) * total_elements + k * 8)
);
}
/* Interleave 16 bytes */
for(
j = 0;
j < 4; j++) {
for(
l = 0;
l < 2; l++) {
r1[2*j+l] =
vzip_u16(vreinterpret_u16_u8(r0[2 * j].val[l]), vreinterpret_u16_u8(r0[2 * j + 1].val[l])
);
}
}
/* Interleave 32 bytes */
for(
j = 0;
j < 2; j++) {
for(
l = 0;
l < 2; l++) {
for(
m = 0;
m < 2; m++) {
r2[j*2+l+4*m] =
vzip_u32(vreinterpret_u32_u16(r1[j + 4 * m].val[l]), vreinterpret_u32_u16(r1[j + 2 + 4 * m].val[l])
);
}
}
}
/* Store the results in the destination vector */
for(
j = 0;
j < 4; j++) {
for(
l = 0;
l < 2; l++) {
for(
m = 0;
m < 2; m++) {
vst1_u8(dest
+ i + (4*j+m+2*l)*8,
vreinterpret_u8_u32(r2[j + 4 * m]
.val[l]));
}
}
}
}
}
