static INLINE
uint8x8x4_t read_4x8(unsigned char *src, int pitch) {
uint8x8x4_t x;
const uint8x8_t a = vld1_u8(src);
const uint8x8_t b = vld1_u8(src + pitch * 1);
const uint8x8_t c = vld1_u8(src + pitch * 2);
const uint8x8_t d = vld1_u8(src + pitch * 3);
const uint8x8_t e = vld1_u8(src + pitch * 4);
const uint8x8_t f = vld1_u8(src + pitch * 5);
const uint8x8_t g = vld1_u8(src + pitch * 6);
const uint8x8_t h = vld1_u8(src + pitch * 7);
const uint32x2x2_t r04_u32 = vtrn_u32(vreinterpret_u32_u8(a),
vreinterpret_u32_u8(e));
const uint32x2x2_t r15_u32 = vtrn_u32(vreinterpret_u32_u8(b),
vreinterpret_u32_u8(f));
const uint32x2x2_t r26_u32 = vtrn_u32(vreinterpret_u32_u8(c),
vreinterpret_u32_u8(g));
const uint32x2x2_t r37_u32 = vtrn_u32(vreinterpret_u32_u8(d),
vreinterpret_u32_u8(h));
const uint16x4x2_t r02_u16 = vtrn_u16(vreinterpret_u16_u32(r04_u32.val[0]),
vreinterpret_u16_u32(r26_u32.val[0]));
const uint16x4x2_t r13_u16 = vtrn_u16(vreinterpret_u16_u32(r15_u32.val[0]),
vreinterpret_u16_u32(r37_u32.val[0]));
const uint8x8x2_t r01_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[0]),
vreinterpret_u8_u16(r13_u16.val[0]));
const uint8x8x2_t r23_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[1]),
vreinterpret_u8_u16(r13_u16.val[1]));
/*
* after vtrn_u32
00 01 02 03 | 40 41 42 43
10 11 12 13 | 50 51 52 53
20 21 22 23 | 60 61 62 63
30 31 32 33 | 70 71 72 73
---
* after vtrn_u16
00 01 20 21 | 40 41 60 61
02 03 22 23 | 42 43 62 63
10 11 30 31 | 50 51 70 71
12 13 32 33 | 52 52 72 73
00 01 20 21 | 40 41 60 61
10 11 30 31 | 50 51 70 71
02 03 22 23 | 42 43 62 63
12 13 32 33 | 52 52 72 73
---
* after vtrn_u8
00 10 20 30 | 40 50 60 70
01 11 21 31 | 41 51 61 71
02 12 22 32 | 42 52 62 72
03 13 23 33 | 43 53 63 73
*/
x.val[0] = r01_u8.val[0];
x.val[1] = r01_u8.val[1];
x.val[2] = r23_u8.val[0];
x.val[3] = r23_u8.val[1];
return x;
}
static INLINE void write_2x4(unsigned char *dst, int pitch,
const uint8x8x2_t result) {
/*
* uint8x8x2_t result
00 01 02 03 | 04 05 06 07
10 11 12 13 | 14 15 16 17
---
* after vtrn_u8
00 10 02 12 | 04 14 06 16
01 11 03 13 | 05 15 07 17
*/
const uint8x8x2_t r01_u8 = vtrn_u8(result.val[0],
result.val[1]);
const uint16x4_t x_0_4 = vreinterpret_u16_u8(r01_u8.val[0]);
const uint16x4_t x_1_5 = vreinterpret_u16_u8(r01_u8.val[1]);
vst1_lane_u16((uint16_t *)dst, x_0_4, 0);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 0);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 1);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 1);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 2);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 2);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 3);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 3);
}
void aom_lpf_vertical_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
const uint8_t *limit, const uint8_t *thresh) {
int i;
uint8_t *s;
uint8x8_t dblimit, dlimit, dthresh;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
uint8x8_t d16u8, d17u8, d18u8;
uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
uint8x8x4_t d4Result;
uint8x8x2_t d2Result;
dblimit = vld1_u8(blimit);
dlimit = vld1_u8(limit);
dthresh = vld1_u8(thresh);
for (i = 0; i < 1; i++) {
s = src + (i * (pitch << 3)) - 4;
d3u8 = vld1_u8(s);
s += pitch;
d4u8 = vld1_u8(s);
s += pitch;
d5u8 = vld1_u8(s);
s += pitch;
d6u8 = vld1_u8(s);
s += pitch;
d7u8 = vld1_u8(s);
s += pitch;
d16u8 = vld1_u8(s);
s += pitch;
d17u8 = vld1_u8(s);
s += pitch;
d18u8 = vld1_u8(s);
d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
vreinterpret_u16_u32(d2tmp2.val[0]));
d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
vreinterpret_u16_u32(d2tmp3.val[0]));
d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
vreinterpret_u16_u32(d2tmp2.val[1]));
d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
vreinterpret_u16_u32(d2tmp3.val[1]));
d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
vreinterpret_u8_u16(d2tmp5.val[0]));
d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
vreinterpret_u8_u16(d2tmp5.val[1]));
d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
vreinterpret_u8_u16(d2tmp7.val[0]));
d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
vreinterpret_u8_u16(d2tmp7.val[1]));
d3u8 = d2tmp8.val[0];
d4u8 = d2tmp8.val[1];
d5u8 = d2tmp9.val[0];
d6u8 = d2tmp9.val[1];
d7u8 = d2tmp10.val[0];
d16u8 = d2tmp10.val[1];
d17u8 = d2tmp11.val[0];
d18u8 = d2tmp11.val[1];
mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
&d5u8);
d4Result.val[0] = d0u8;
d4Result.val[1] = d1u8;
d4Result.val[2] = d2u8;
d4Result.val[3] = d3u8;
d2Result.val[0] = d4u8;
d2Result.val[1] = d5u8;
s = src - 3;
vst4_lane_u8(s, d4Result, 0);
s += pitch;
vst4_lane_u8(s, d4Result, 1);
s += pitch;
vst4_lane_u8(s, d4Result, 2);
s += pitch;
vst4_lane_u8(s, d4Result, 3);
s += pitch;
vst4_lane_u8(s, d4Result, 4);
s += pitch;
vst4_lane_u8(s, d4Result, 5);
s += pitch;
vst4_lane_u8(s, d4Result, 6);
s += pitch;
vst4_lane_u8(s, d4Result, 7);
s = src + 1;
vst2_lane_u8(s, d2Result, 0);
s += pitch;
vst2_lane_u8(s, d2Result, 1);
s += pitch;
vst2_lane_u8(s, d2Result, 2);
s += pitch;
vst2_lane_u8(s, d2Result, 3);
s += pitch;
vst2_lane_u8(s, d2Result, 4);
s += pitch;
vst2_lane_u8(s, d2Result, 5);
s += pitch;
vst2_lane_u8(s, d2Result, 6);
s += pitch;
vst2_lane_u8(s, d2Result, 7);
}
return;
}
uint8x8x2_t test_vtrn_u8(uint8x8_t a, uint8x8_t b) {
// CHECK-LABEL: test_vtrn_u8
return vtrn_u8(a, b);
// CHECK: trn1 {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
// CHECK: trn2 {{v[0-9]+}}.8b, {{v[0-9]+}}.8b
}
void vp9_lpf_vertical_4_neon(
unsigned char *src,
int pitch,
unsigned char *blimit,
unsigned char *limit,
unsigned char *thresh,
int count) {
int i, pitch8;
uint8_t *s;
uint8x8_t dblimit, dlimit, dthresh;
uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
uint8x8x4_t d4Result;
if (count == 0) // end_vp9_lf_h_edge
return;
dblimit = vld1_u8(blimit);
dlimit = vld1_u8(limit);
dthresh = vld1_u8(thresh);
pitch8 = pitch * 8;
for (i = 0; i < count; i++, src += pitch8) {
s = src - (i + 1) * 4;
d3u8 = vld1_u8(s);
s += pitch;
d4u8 = vld1_u8(s);
s += pitch;
d5u8 = vld1_u8(s);
s += pitch;
d6u8 = vld1_u8(s);
s += pitch;
d7u8 = vld1_u8(s);
s += pitch;
d16u8 = vld1_u8(s);
s += pitch;
d17u8 = vld1_u8(s);
s += pitch;
d18u8 = vld1_u8(s);
d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
vreinterpret_u32_u8(d7u8));
d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
vreinterpret_u32_u8(d16u8));
d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
vreinterpret_u32_u8(d17u8));
d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
vreinterpret_u32_u8(d18u8));
d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
vreinterpret_u16_u32(d2tmp2.val[0]));
d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
vreinterpret_u16_u32(d2tmp3.val[0]));
d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
vreinterpret_u16_u32(d2tmp2.val[1]));
d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
vreinterpret_u16_u32(d2tmp3.val[1]));
d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
vreinterpret_u8_u16(d2tmp5.val[0]));
d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
vreinterpret_u8_u16(d2tmp5.val[1]));
d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
vreinterpret_u8_u16(d2tmp7.val[0]));
d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
vreinterpret_u8_u16(d2tmp7.val[1]));
d3u8 = d2tmp8.val[0];
d4u8 = d2tmp8.val[1];
d5u8 = d2tmp9.val[0];
d6u8 = d2tmp9.val[1];
d7u8 = d2tmp10.val[0];
d16u8 = d2tmp10.val[1];
d17u8 = d2tmp11.val[0];
d18u8 = d2tmp11.val[1];
vp9_loop_filter_neon(dblimit, dlimit, dthresh,
d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
&d4u8, &d5u8, &d6u8, &d7u8);
d4Result.val[0] = d4u8;
d4Result.val[1] = d5u8;
d4Result.val[2] = d6u8;
d4Result.val[3] = d7u8;
src -= 2;
vst4_lane_u8(src, d4Result, 0);
src += pitch;
vst4_lane_u8(src, d4Result, 1);
src += pitch;
vst4_lane_u8(src, d4Result, 2);
src += pitch;
vst4_lane_u8(src, d4Result, 3);
src += pitch;
vst4_lane_u8(src, d4Result, 4);
src += pitch;
vst4_lane_u8(src, d4Result, 5);
src += pitch;
vst4_lane_u8(src, d4Result, 6);
src += pitch;
vst4_lane_u8(src, d4Result, 7);
}
return;
}
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;
}