static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
uint8_t *output_ptr,
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
unsigned int output_width,
const uint8_t *filter) {
const uint8x8_t f0 = vmov_n_u8(filter[0]);
const uint8x8_t f1 = vmov_n_u8(filter[1]);
unsigned int i, j;
for (i = 0; i < output_height; ++i) {
for (j = 0; j < output_width; j += 16) {
const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
}
// Next row...
src_ptr += src_pixels_per_line;
output_ptr += output_width;
}
}
inline unsigned int GetByteMask2(uint8x16_t a, uint8x16_t b)
{
uint8x16_t am = vandq_u8(a, compaction_mask);
uint8x16_t bm = vandq_u8(b, compaction_mask);
uint8x8_t a_sum = vpadd_u8(vget_high_u8(am), vget_low_u8(am));
uint8x8_t b_sum = vpadd_u8(vget_high_u8(bm), vget_low_u8(bm));
a_sum = vpadd_u8(b_sum, a_sum);
a_sum = vpadd_u8(a_sum, a_sum);
return vget_lane_u32(vreinterpret_u32_u8(a_sum), 0);
}
void test_vget_lowu8 (void)
{
register uint8x8_t out_uint8x8_t asm ("d18");
uint8x16_t arg0_uint8x16_t;
out_uint8x8_t = vget_low_u8 (arg0_uint8x16_t);
}
void test_vget_lowu8 (void)
{
uint8x8_t out_uint8x8_t;
uint8x16_t arg0_uint8x16_t;
out_uint8x8_t = vget_low_u8 (arg0_uint8x16_t);
}
SIMD_INLINE uint32x4_t SquaredDifferenceSumMasked(const uint8x16_t & a, const uint8x16_t & b, const uint8x16_t & mask)
{
uint8x16_t ad = vandq_u8(vabdq_u8(a, b), mask);
uint16x8_t lo = Square(vget_low_u8(ad));
uint16x8_t hi = Square(vget_high_u8(ad));
return vaddq_u32(vpaddlq_u16(lo), vpaddlq_u16(hi));
}
static v16
mulby(uint8_t x, v16 v)
{
#ifdef LIBRS_USE_NEON
#define uint8x16_to_8x8x2(v) ((uint8x8x2_t) { vget_low_u8(v), vget_high_u8(v) })
v16 lo, hi;
lo = v & VEC16(0x0f);
hi = vshrq_n_u8(v, 4);
lo = vcombine_u8(
vtbl2_u8(uint8x16_to_8x8x2(rs_nibmul[x].lo), vget_low_u8(lo)),
vtbl2_u8(uint8x16_to_8x8x2(rs_nibmul[x].lo), vget_high_u8(lo)));
hi = vcombine_u8(
vtbl2_u8(uint8x16_to_8x8x2(rs_nibmul[x].hi), vget_low_u8(hi)),
vtbl2_u8(uint8x16_to_8x8x2(rs_nibmul[x].hi), vget_high_u8(hi)));
return lo ^ hi;
#elif defined(LIBRS_USE_SSSE3)
v16 lo, hi;
lo = v & VEC16(0x0f);
hi = __builtin_ia32_psrawi128(v, 4);
hi &= VEC16(0x0f);
lo = __builtin_ia32_pshufb128(rs_nibmul[x].lo, lo);
hi = __builtin_ia32_pshufb128(rs_nibmul[x].hi, hi);
return lo ^ hi;
#else
v16 vv = VEC16(0);
while (x != 0) {
if (x & 1)
vv ^= v;
x >>= 1;
v = mul2(v);
}
return vv;
#endif
}
void vp9_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref,
const int ref_stride, const int height) {
int i;
uint16x8_t vec_sum_lo = vdupq_n_u16(0);
uint16x8_t vec_sum_hi = vdupq_n_u16(0);
const int shift_factor = ((height >> 5) + 3) * -1;
const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
for (i = 0; i < height; i += 8) {
const uint8x16_t vec_row1 = vld1q_u8(ref);
const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
ref += ref_stride * 8;
}
vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
hbuf += 8;
vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
}
static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
const uint32_t* const end = argb_data + (num_pixels & ~3);
const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
for (; argb_data < end; argb_data += 4) {
const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
const uint8x16_t greens =
vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
vtbl1_u8(vget_high_u8(argb), shuffle));
vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens));
}
// fallthrough and finish off with plain-C
VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3);
}
static INLINE void scaledconvolve_vert_w16(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
const int y0_q4, const int y_step_q4, const int w, const int h) {
int x, y;
int y_q4 = y0_q4;
src -= src_stride * (SUBPEL_TAPS / 2 - 1);
y = h;
do {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
if (y_q4 & SUBPEL_MASK) {
x = 0;
do {
const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
uint8x16_t ss[8];
uint8x8_t s[8], d[2];
load_u8_16x8(src_y, src_stride, &ss[0], &ss[1], &ss[2], &ss[3], &ss[4],
&ss[5], &ss[6], &ss[7]);
s[0] = vget_low_u8(ss[0]);
s[1] = vget_low_u8(ss[1]);
s[2] = vget_low_u8(ss[2]);
s[3] = vget_low_u8(ss[3]);
s[4] = vget_low_u8(ss[4]);
s[5] = vget_low_u8(ss[5]);
s[6] = vget_low_u8(ss[6]);
s[7] = vget_low_u8(ss[7]);
d[0] = scale_filter_8(s, filters);
s[0] = vget_high_u8(ss[0]);
s[1] = vget_high_u8(ss[1]);
s[2] = vget_high_u8(ss[2]);
s[3] = vget_high_u8(ss[3]);
s[4] = vget_high_u8(ss[4]);
s[5] = vget_high_u8(ss[5]);
s[6] = vget_high_u8(ss[6]);
s[7] = vget_high_u8(ss[7]);
d[1] = scale_filter_8(s, filters);
vst1q_u8(&dst[x], vcombine_u8(d[0], d[1]));
src_y += 16;
x += 16;
} while (x < w);
} else {
memcpy(dst, &src_y[3 * src_stride], w);
}
dst += dst_stride;
y_q4 += y_step_q4;
} while (--y);
}
int16_t vp9_int_pro_col_neon(uint8_t const *ref, const int width) {
int i;
uint16x8_t vec_sum = vdupq_n_u16(0);
for (i = 0; i < width; i += 16) {
const uint8x16_t vec_row = vld1q_u8(ref);
vec_sum = vaddw_u8(vec_sum, vget_low_u8(vec_row));
vec_sum = vaddw_u8(vec_sum, vget_high_u8(vec_row));
ref += 16;
}
return horizontal_add_u16x8(vec_sum);
}
void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint8x8_t d2u8 = vdup_n_u8(0);
uint8x16_t q0u8 = vdupq_n_u8(0);
uint8x16_t q1u8 = vdupq_n_u8(0);
(void)above;
for (k = 0; k < 2; k++, left += 16) {
q1u8 = vld1q_u8(left);
d2u8 = vget_low_u8(q1u8);
for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
q0u8 = vdupq_lane_u8(d2u8, 0);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 1);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 2);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 3);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 4);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 5);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 6);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 7);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
}
}
}
int normL1_(const uchar* a, const uchar* b, int n)
{
int j = 0, d = 0;
#if CV_SSE
__m128i d0 = _mm_setzero_si128();
for( ; j <= n - 16; j += 16 )
{
__m128i t0 = _mm_loadu_si128((const __m128i*)(a + j));
__m128i t1 = _mm_loadu_si128((const __m128i*)(b + j));
d0 = _mm_add_epi32(d0, _mm_sad_epu8(t0, t1));
}
for( ; j <= n - 4; j += 4 )
{
__m128i t0 = _mm_cvtsi32_si128(*(const int*)(a + j));
__m128i t1 = _mm_cvtsi32_si128(*(const int*)(b + j));
d0 = _mm_add_epi32(d0, _mm_sad_epu8(t0, t1));
}
d = _mm_cvtsi128_si32(_mm_add_epi32(d0, _mm_unpackhi_epi64(d0, d0)));
#elif CV_NEON
uint32x4_t v_sum = vdupq_n_u32(0.0f);
for ( ; j <= n - 16; j += 16)
{
uint8x16_t v_dst = vabdq_u8(vld1q_u8(a + j), vld1q_u8(b + j));
uint16x8_t v_low = vmovl_u8(vget_low_u8(v_dst)), v_high = vmovl_u8(vget_high_u8(v_dst));
v_sum = vaddq_u32(v_sum, vaddl_u16(vget_low_u16(v_low), vget_low_u16(v_high)));
v_sum = vaddq_u32(v_sum, vaddl_u16(vget_high_u16(v_low), vget_high_u16(v_high)));
}
uint CV_DECL_ALIGNED(16) buf[4];
vst1q_u32(buf, v_sum);
d = buf[0] + buf[1] + buf[2] + buf[3];
#endif
{
for( ; j <= n - 4; j += 4 )
{
d += std::abs(a[j] - b[j]) + std::abs(a[j+1] - b[j+1]) +
std::abs(a[j+2] - b[j+2]) + std::abs(a[j+3] - b[j+3]);
}
}
for( ; j < n; j++ )
d += std::abs(a[j] - b[j]);
return d;
}
void vp9_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
uint8x16_t q0u8, q1u8;
int16x8_t q0s16, q1s16, q8s16, q11s16;
uint16x4_t d20u16;
uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
for (k = 0; k < 2; k++, left += 8) {
d18u8 = vld1_u8(left);
q10u16 = vmovl_u8(d18u8);
d20u16 = vget_low_u16(q10u16);
for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
q0u16 = vdupq_lane_u16(d20u16, 0);
q8u16 = vdupq_lane_u16(d20u16, 1);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
q0u16 = vdupq_lane_u16(d20u16, 2);
q8u16 = vdupq_lane_u16(d20u16, 3);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
}
}
}
static INLINE void vp8_loop_filter_simple_vertical_edge_neon(
unsigned char *s,
int p,
const unsigned char *blimit) {
unsigned char *src1;
uint8x16_t qblimit, q0u8;
uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q11u8, q12u8, q14u8, q15u8;
int16x8_t q2s16, q13s16, q11s16;
int8x8_t d28s8, d29s8;
int8x16_t q2s8, q3s8, q10s8, q11s8, q14s8;
uint8x8x4_t d0u8x4; // d6, d7, d8, d9
uint8x8x4_t d1u8x4; // d10, d11, d12, d13
uint8x8x2_t d2u8x2; // d12, d13
uint8x8x2_t d3u8x2; // d14, d15
qblimit = vdupq_n_u8(*blimit);
src1 = s - 2;
d0u8x4 = read_4x8(src1, p);
src1 += p * 8;
d1u8x4 = read_4x8(src1, p);
q3u8 = vcombine_u8(d0u8x4.val[0], d1u8x4.val[0]); // d6 d10
q4u8 = vcombine_u8(d0u8x4.val[2], d1u8x4.val[2]); // d8 d12
q5u8 = vcombine_u8(d0u8x4.val[1], d1u8x4.val[1]); // d7 d11
q6u8 = vcombine_u8(d0u8x4.val[3], d1u8x4.val[3]); // d9 d13
q15u8 = vabdq_u8(q5u8, q4u8);
q14u8 = vabdq_u8(q3u8, q6u8);
q15u8 = vqaddq_u8(q15u8, q15u8);
q14u8 = vshrq_n_u8(q14u8, 1);
q0u8 = vdupq_n_u8(0x80);
q11s16 = vdupq_n_s16(3);
q15u8 = vqaddq_u8(q15u8, q14u8);
q3u8 = veorq_u8(q3u8, q0u8);
q4u8 = veorq_u8(q4u8, q0u8);
q5u8 = veorq_u8(q5u8, q0u8);
q6u8 = veorq_u8(q6u8, q0u8);
q15u8 = vcgeq_u8(qblimit, q15u8);
q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q4u8)),
vget_low_s8(vreinterpretq_s8_u8(q5u8)));
q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q4u8)),
vget_high_s8(vreinterpretq_s8_u8(q5u8)));
q14s8 = vqsubq_s8(vreinterpretq_s8_u8(q3u8),
vreinterpretq_s8_u8(q6u8));
q2s16 = vmulq_s16(q2s16, q11s16);
q13s16 = vmulq_s16(q13s16, q11s16);
q11u8 = vdupq_n_u8(3);
q12u8 = vdupq_n_u8(4);
q2s16 = vaddw_s8(q2s16, vget_low_s8(q14s8));
q13s16 = vaddw_s8(q13s16, vget_high_s8(q14s8));
d28s8 = vqmovn_s16(q2s16);
d29s8 = vqmovn_s16(q13s16);
q14s8 = vcombine_s8(d28s8, d29s8);
q14s8 = vandq_s8(q14s8, vreinterpretq_s8_u8(q15u8));
q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q11u8));
q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q12u8));
q2s8 = vshrq_n_s8(q2s8, 3);
q14s8 = vshrq_n_s8(q3s8, 3);
q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q5u8), q2s8);
q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q4u8), q14s8);
q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
d2u8x2.val[0] = vget_low_u8(q6u8); // d12
d2u8x2.val[1] = vget_low_u8(q7u8); // d14
d3u8x2.val[0] = vget_high_u8(q6u8); // d13
d3u8x2.val[1] = vget_high_u8(q7u8); // d15
src1 = s - 1;
write_2x8(src1, p, d2u8x2, d3u8x2);
}
inline bool isFound(uint8x16_t x)
{
uint8x8_t xx = vorr_u8(vget_low_u8(x), vget_high_u8(x));
return vget_lane_u64(vreinterpret_u64_u8(xx), 0);
}
void vp9_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
uint8x16_t q0u8, q1u8, q2u8;
int16x8_t q12s16, q13s16, q14s16, q15s16;
uint16x4_t d6u16;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u8 = vld1q_u8(above + 16);
q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
for (k = 0; k < 4; k++, left += 8) {
d26u8 = vld1_u8(left);
q3u16 = vmovl_u8(d26u8);
d6u16 = vget_low_u16(q3u16);
for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
q0u16 = vdupq_lane_u16(d6u16, 0);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 1);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 2);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 3);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
}
}
}
void qt_blend_argb32_on_argb32_neon(uchar *destPixels, int dbpl,
const uchar *srcPixels, int sbpl,
int w, int h,
int const_alpha)
{
const uint *src = (const uint *) srcPixels;
uint *dst = (uint *) destPixels;
int16x8_t half = vdupq_n_s16(0x80);
int16x8_t full = vdupq_n_s16(0xff);
if (const_alpha == 256) {
for (int y = 0; y < h; ++y) {
int x = 0;
for (; x < w-3; x += 4) {
int32x4_t src32 = vld1q_s32((int32_t *)&src[x]);
if ((src[x] & src[x+1] & src[x+2] & src[x+3]) >= 0xff000000) {
// all opaque
vst1q_s32((int32_t *)&dst[x], src32);
} else if (src[x] | src[x+1] | src[x+2] | src[x+3]) {
int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]);
const uint8x16_t src8 = vreinterpretq_u8_s32(src32);
const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32);
const uint8x8_t src8_low = vget_low_u8(src8);
const uint8x8_t dst8_low = vget_low_u8(dst8);
const uint8x8_t src8_high = vget_high_u8(src8);
const uint8x8_t dst8_high = vget_high_u8(dst8);
const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low));
const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low));
const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high));
const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high));
const int16x8_t result16_low = qvsource_over_s16(src16_low, dst16_low, half, full);
const int16x8_t result16_high = qvsource_over_s16(src16_high, dst16_high, half, full);
const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low));
const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high));
vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high));
}
}
for (; x<w; ++x) {
uint s = src[x];
if (s >= 0xff000000)
dst[x] = s;
else if (s != 0)
dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
}
dst = (quint32 *)(((uchar *) dst) + dbpl);
src = (const quint32 *)(((const uchar *) src) + sbpl);
}
} else if (const_alpha != 0) {
const_alpha = (const_alpha * 255) >> 8;
int16x8_t const_alpha16 = vdupq_n_s16(const_alpha);
for (int y = 0; y < h; ++y) {
int x = 0;
for (; x < w-3; x += 4) {
if (src[x] | src[x+1] | src[x+2] | src[x+3]) {
int32x4_t src32 = vld1q_s32((int32_t *)&src[x]);
int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]);
const uint8x16_t src8 = vreinterpretq_u8_s32(src32);
const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32);
const uint8x8_t src8_low = vget_low_u8(src8);
const uint8x8_t dst8_low = vget_low_u8(dst8);
const uint8x8_t src8_high = vget_high_u8(src8);
const uint8x8_t dst8_high = vget_high_u8(dst8);
const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low));
const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low));
const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high));
const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high));
const int16x8_t srcalpha16_low = qvbyte_mul_s16(src16_low, const_alpha16, half);
const int16x8_t srcalpha16_high = qvbyte_mul_s16(src16_high, const_alpha16, half);
const int16x8_t result16_low = qvsource_over_s16(srcalpha16_low, dst16_low, half, full);
const int16x8_t result16_high = qvsource_over_s16(srcalpha16_high, dst16_high, half, full);
const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low));
const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high));
vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high));
}
}
for (; x<w; ++x) {
uint s = src[x];
if (s != 0) {
s = BYTE_MUL(s, const_alpha);
dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
}
}
dst = (quint32 *)(((uchar *) dst) + dbpl);
src = (const quint32 *)(((const uchar *) src) + sbpl);
}
}
static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
const uint8x8_t shuffle) {
return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
vtbl1_u8(vget_high_u8(argb), shuffle));
}
inline uint8x8_t vget_low(const uint8x16_t & v) { return vget_low_u8 (v); }
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;
}
unsigned int vp8_variance_halfpixvar16x16_hv_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d10s16, d11s16, d12s16, d13s16;
int16x4_t d18s16, d19s16, d20s16, d21s16, d22s16, d23s16, d24s16, d25s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint16x8_t q0u16, q1u16, q5u16, q6u16, q9u16, q10u16, q11u16, q12u16;
int32x4_t q13s32, q14s32, q15s32;
int64x2_t q0s64, q1s64, q5s64;
q13s32 = vdupq_n_s32(0);
q14s32 = vdupq_n_s32(0);
q15s32 = vdupq_n_s32(0);
q0u8 = vld1q_u8(src_ptr);
q1u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q1u8 = vextq_u8(q0u8, q1u8, 1);
q0u8 = vrhaddq_u8(q0u8, q1u8);
for (i = 0; i < 4; i++) { // vp8_filt_fpo16x16s_4_0_loop_neon
q2u8 = vld1q_u8(src_ptr);
q3u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q4u8 = vld1q_u8(src_ptr);
q5u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q6u8 = vld1q_u8(src_ptr);
q7u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q8u8 = vld1q_u8(src_ptr);
q9u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q3u8 = vextq_u8(q2u8, q3u8, 1);
q5u8 = vextq_u8(q4u8, q5u8, 1);
q7u8 = vextq_u8(q6u8, q7u8, 1);
q9u8 = vextq_u8(q8u8, q9u8, 1);
q1u8 = vrhaddq_u8(q2u8, q3u8);
q2u8 = vrhaddq_u8(q4u8, q5u8);
q3u8 = vrhaddq_u8(q6u8, q7u8);
q4u8 = vrhaddq_u8(q8u8, q9u8);
q0u8 = vrhaddq_u8(q0u8, q1u8);
q1u8 = vrhaddq_u8(q1u8, q2u8);
q2u8 = vrhaddq_u8(q2u8, q3u8);
q3u8 = vrhaddq_u8(q3u8, q4u8);
q5u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q6u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q7u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q8u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
d4u8 = vget_low_u8(q2u8);
d5u8 = vget_high_u8(q2u8);
d6u8 = vget_low_u8(q3u8);
d7u8 = vget_high_u8(q3u8);
q9u16 = vsubl_u8(d0u8, vget_low_u8(q5u8));
q10u16 = vsubl_u8(d1u8, vget_high_u8(q5u8));
q11u16 = vsubl_u8(d2u8, vget_low_u8(q6u8));
q12u16 = vsubl_u8(d3u8, vget_high_u8(q6u8));
q0u16 = vsubl_u8(d4u8, vget_low_u8(q7u8));
q1u16 = vsubl_u8(d5u8, vget_high_u8(q7u8));
q5u16 = vsubl_u8(d6u8, vget_low_u8(q8u8));
q6u16 = vsubl_u8(d7u8, vget_high_u8(q8u8));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q9u16));
q14s32 = vmlal_s16(q14s32, d18s16, d18s16);
q15s32 = vmlal_s16(q15s32, d19s16, d19s16);
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q10u16));
q14s32 = vmlal_s16(q14s32, d20s16, d20s16);
q15s32 = vmlal_s16(q15s32, d21s16, d21s16);
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q11u16));
q14s32 = vmlal_s16(q14s32, d22s16, d22s16);
q15s32 = vmlal_s16(q15s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q12u16));
q14s32 = vmlal_s16(q14s32, d24s16, d24s16);
q15s32 = vmlal_s16(q15s32, d25s16, d25s16);
d0s16 = vreinterpret_s16_u16(vget_low_u16(q0u16));
d1s16 = vreinterpret_s16_u16(vget_high_u16(q0u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q0u16));
q14s32 = vmlal_s16(q14s32, d0s16, d0s16);
q15s32 = vmlal_s16(q15s32, d1s16, d1s16);
d2s16 = vreinterpret_s16_u16(vget_low_u16(q1u16));
d3s16 = vreinterpret_s16_u16(vget_high_u16(q1u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q1u16));
q14s32 = vmlal_s16(q14s32, d2s16, d2s16);
q15s32 = vmlal_s16(q15s32, d3s16, d3s16);
d10s16 = vreinterpret_s16_u16(vget_low_u16(q5u16));
d11s16 = vreinterpret_s16_u16(vget_high_u16(q5u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q5u16));
q14s32 = vmlal_s16(q14s32, d10s16, d10s16);
q15s32 = vmlal_s16(q15s32, d11s16, d11s16);
d12s16 = vreinterpret_s16_u16(vget_low_u16(q6u16));
d13s16 = vreinterpret_s16_u16(vget_high_u16(q6u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q6u16));
q14s32 = vmlal_s16(q14s32, d12s16, d12s16);
q15s32 = vmlal_s16(q15s32, d13s16, d13s16);
q0u8 = q4u8;
}
q15s32 = vaddq_s32(q14s32, q15s32);
q0s64 = vpaddlq_s32(q13s32);
q1s64 = vpaddlq_s32(q15s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
unsigned int vp8_sub_pixel_variance16x16_neon_func(
const unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
const unsigned char *dst_ptr,
int dst_pixels_per_line,
unsigned int *sse) {
int i;
DECLARE_ALIGNED_ARRAY(16, unsigned char, tmp, 528);
unsigned char *tmpp;
unsigned char *tmpp2;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d16u8, d17u8, d18u8;
uint8x8_t d19u8, d20u8, d21u8;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint8x16_t q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16, q6u16, q7u16, q8u16;
uint16x8_t q9u16, q10u16, q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
tmpp2 = tmp + 272;
tmpp = tmp;
if (xoffset == 0) { // secondpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
q11u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q13u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q14u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q15u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
} else if (yoffset == 0) { // firstpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
for (i = 4; i > 0 ; i--) {
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp2, q7u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q8u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q9u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q10u8);
tmpp2 += 16;
}
} else {
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
// First Pass: output_height lines x output_width columns (17x16)
for (i = 3; i > 0; i--) {
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q10u8);
tmpp += 16;
}
// First-pass filtering for rest 5 lines
d14u8 = vld1_u8(src_ptr);
d15u8 = vld1_u8(src_ptr + 8);
d16u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q9u16 = vmull_u8(d2u8, d0u8);
q10u16 = vmull_u8(d3u8, d0u8);
q11u16 = vmull_u8(d5u8, d0u8);
q12u16 = vmull_u8(d6u8, d0u8);
q13u16 = vmull_u8(d8u8, d0u8);
q14u16 = vmull_u8(d9u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
q9u16 = vmlal_u8(q9u16, d2u8, d1u8);
q11u16 = vmlal_u8(q11u16, d5u8, d1u8);
q13u16 = vmlal_u8(q13u16, d8u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
q10u16 = vmlal_u8(q10u16, d3u8, d1u8);
q12u16 = vmlal_u8(q12u16, d6u8, d1u8);
q14u16 = vmlal_u8(q14u16, d9u8, d1u8);
q1u16 = vmull_u8(d11u8, d0u8);
q2u16 = vmull_u8(d12u8, d0u8);
q3u16 = vmull_u8(d14u8, d0u8);
q4u16 = vmull_u8(d15u8, d0u8);
d11u8 = vext_u8(d11u8, d12u8, 1);
d14u8 = vext_u8(d14u8, d15u8, 1);
q1u16 = vmlal_u8(q1u16, d11u8, d1u8);
q3u16 = vmlal_u8(q3u16, d14u8, d1u8);
d12u8 = vext_u8(d12u8, d13u8, 1);
d15u8 = vext_u8(d15u8, d16u8, 1);
q2u16 = vmlal_u8(q2u16, d12u8, d1u8);
q4u16 = vmlal_u8(q4u16, d15u8, d1u8);
d10u8 = vqrshrn_n_u16(q9u16, 7);
d11u8 = vqrshrn_n_u16(q10u16, 7);
d12u8 = vqrshrn_n_u16(q11u16, 7);
d13u8 = vqrshrn_n_u16(q12u16, 7);
d14u8 = vqrshrn_n_u16(q13u16, 7);
d15u8 = vqrshrn_n_u16(q14u16, 7);
d16u8 = vqrshrn_n_u16(q1u16, 7);
d17u8 = vqrshrn_n_u16(q2u16, 7);
d18u8 = vqrshrn_n_u16(q3u16, 7);
d19u8 = vqrshrn_n_u16(q4u16, 7);
q5u8 = vcombine_u8(d10u8, d11u8);
q6u8 = vcombine_u8(d12u8, d13u8);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
vst1q_u8((uint8_t *)tmpp, q5u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q6u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
// secondpass_filter
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
tmpp = tmp;
tmpp2 = tmpp + 272;
q11u8 = vld1q_u8(tmpp);
tmpp += 16;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(tmpp);
tmpp += 16;
q13u8 = vld1q_u8(tmpp);
tmpp += 16;
q14u8 = vld1q_u8(tmpp);
tmpp += 16;
q15u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
}
// sub_pixel_variance16x16_neon
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
tmpp = tmp + 272;
for (i = 0; i < 8; i++) { // sub_pixel_variance16x16_neon_loop
q0u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u8 = vld1q_u8(tmpp);
tmpp += 16;
q2u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
q3u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
q11u16 = vsubl_u8(d0u8, vget_low_u8(q2u8));
q12u16 = vsubl_u8(d1u8, vget_high_u8(q2u8));
q13u16 = vsubl_u8(d2u8, vget_low_u8(q3u8));
q14u16 = vsubl_u8(d3u8, vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
void vpx_subtract_block_neon(int rows, int cols,
int16_t *diff, ptrdiff_t diff_stride,
const uint8_t *src, ptrdiff_t src_stride,
const uint8_t *pred, ptrdiff_t pred_stride) {
int r, c;
if (cols > 16) {
for (r = 0; r < rows; ++r) {
for (c = 0; c < cols; c += 32) {
const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
const uint8x16_t v_pred_00 = vld1q_u8(&pred[c + 0]);
const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
const uint16x8_t v_diff_lo_00 = vsubl_u8(vget_low_u8(v_src_00),
vget_low_u8(v_pred_00));
const uint16x8_t v_diff_hi_00 = vsubl_u8(vget_high_u8(v_src_00),
vget_high_u8(v_pred_00));
const uint16x8_t v_diff_lo_16 = vsubl_u8(vget_low_u8(v_src_16),
vget_low_u8(v_pred_16));
const uint16x8_t v_diff_hi_16 = vsubl_u8(vget_high_u8(v_src_16),
vget_high_u8(v_pred_16));
vst1q_s16(&diff[c + 0], vreinterpretq_s16_u16(v_diff_lo_00));
vst1q_s16(&diff[c + 8], vreinterpretq_s16_u16(v_diff_hi_00));
vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
}
diff += diff_stride;
pred += pred_stride;
src += src_stride;
}
} else if (cols > 8) {
for (r = 0; r < rows; ++r) {
const uint8x16_t v_src = vld1q_u8(&src[0]);
const uint8x16_t v_pred = vld1q_u8(&pred[0]);
const uint16x8_t v_diff_lo = vsubl_u8(vget_low_u8(v_src),
vget_low_u8(v_pred));
const uint16x8_t v_diff_hi = vsubl_u8(vget_high_u8(v_src),
vget_high_u8(v_pred));
vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
diff += diff_stride;
pred += pred_stride;
src += src_stride;
}
} else if (cols > 4) {
for (r = 0; r < rows; ++r) {
const uint8x8_t v_src = vld1_u8(&src[0]);
const uint8x8_t v_pred = vld1_u8(&pred[0]);
const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
diff += diff_stride;
pred += pred_stride;
src += src_stride;
}
} else {
for (r = 0; r < rows; ++r) {
for (c = 0; c < cols; ++c)
diff[c] = src[c] - pred[c];
diff += diff_stride;
pred += pred_stride;
src += src_stride;
}
}
}
void vp8_sixtap_predict8x8_neon(
unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
unsigned char *dst_ptr,
int dst_pitch) {
unsigned char *src, *tmpp;
unsigned char tmp[64];
int i;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d18u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8, d25u8;
uint8x8_t d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q9u8, q10u8, q11u8, q12u8;
if (xoffset == 0) { // secondpass_filter8x8_only
// load second_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// load src data
src = src_ptr - src_pixels_per_line * 2;
d18u8 = vld1_u8(src);
src += src_pixels_per_line;
d19u8 = vld1_u8(src);
src += src_pixels_per_line;
d20u8 = vld1_u8(src);
src += src_pixels_per_line;
d21u8 = vld1_u8(src);
src += src_pixels_per_line;
d22u8 = vld1_u8(src);
src += src_pixels_per_line;
d23u8 = vld1_u8(src);
src += src_pixels_per_line;
d24u8 = vld1_u8(src);
src += src_pixels_per_line;
d25u8 = vld1_u8(src);
src += src_pixels_per_line;
d26u8 = vld1_u8(src);
src += src_pixels_per_line;
d27u8 = vld1_u8(src);
src += src_pixels_per_line;
d28u8 = vld1_u8(src);
src += src_pixels_per_line;
d29u8 = vld1_u8(src);
src += src_pixels_per_line;
d30u8 = vld1_u8(src);
for (i = 2; i > 0; i--) {
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
d23u8 = d27u8;
d24u8 = d28u8;
d25u8 = d29u8;
d26u8 = d30u8;
vst1_u8(dst_ptr, d6u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d7u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d8u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d9u8);
dst_ptr += dst_pitch;
}
return;
}
// load first_pass filter
dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
// First pass: output_height lines x output_width columns (9x4)
if (yoffset == 0) // firstpass_filter4x4_only
src = src_ptr - 2;
else
src = src_ptr - 2 - (src_pixels_per_line * 2);
tmpp = tmp;
for (i = 2; i > 0; i--) {
q3u8 = vld1q_u8(src);
src += src_pixels_per_line;
q4u8 = vld1q_u8(src);
src += src_pixels_per_line;
q5u8 = vld1q_u8(src);
src += src_pixels_per_line;
q6u8 = vld1q_u8(src);
src += src_pixels_per_line;
__builtin_prefetch(src);
__builtin_prefetch(src + src_pixels_per_line);
__builtin_prefetch(src + src_pixels_per_line * 2);
q7u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
q8u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
q9u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
q7u16 = vmlsl_u8(q7u16, d28u8, d1u8);
q8u16 = vmlsl_u8(q8u16, d29u8, d1u8);
q9u16 = vmlsl_u8(q9u16, d30u8, d1u8);
q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
q7u16 = vmlsl_u8(q7u16, d28u8, d4u8);
q8u16 = vmlsl_u8(q8u16, d29u8, d4u8);
q9u16 = vmlsl_u8(q9u16, d30u8, d4u8);
q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
q7u16 = vmlal_u8(q7u16, d28u8, d2u8);
q8u16 = vmlal_u8(q8u16, d29u8, d2u8);
q9u16 = vmlal_u8(q9u16, d30u8, d2u8);
q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
q3u16 = vmull_u8(d28u8, d3u8);
q4u16 = vmull_u8(d29u8, d3u8);
q5u16 = vmull_u8(d30u8, d3u8);
q6u16 = vmull_u8(d31u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d22u8 = vqrshrun_n_s16(q7s16, 7);
d23u8 = vqrshrun_n_s16(q8s16, 7);
d24u8 = vqrshrun_n_s16(q9s16, 7);
d25u8 = vqrshrun_n_s16(q10s16, 7);
if (yoffset == 0) { // firstpass_filter8x4_only
vst1_u8(dst_ptr, d22u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d23u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d24u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d25u8);
dst_ptr += dst_pitch;
} else {
vst1_u8(tmpp, d22u8);
tmpp += 8;
vst1_u8(tmpp, d23u8);
tmpp += 8;
vst1_u8(tmpp, d24u8);
tmpp += 8;
vst1_u8(tmpp, d25u8);
tmpp += 8;
}
}
if (yoffset == 0)
return;
// First Pass on rest 5-line data
q3u8 = vld1q_u8(src);
src += src_pixels_per_line;
q4u8 = vld1q_u8(src);
src += src_pixels_per_line;
q5u8 = vld1q_u8(src);
src += src_pixels_per_line;
q6u8 = vld1q_u8(src);
src += src_pixels_per_line;
q7u8 = vld1q_u8(src);
q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
q8u16 = vmlsl_u8(q8u16, d27u8, d1u8);
q9u16 = vmlsl_u8(q9u16, d28u8, d1u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
q8u16 = vmlsl_u8(q8u16, d27u8, d4u8);
q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
q8u16 = vmlal_u8(q8u16, d27u8, d2u8);
q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
q8u16 = vmlal_u8(q8u16, d27u8, d5u8);
q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
q3u16 = vmull_u8(d27u8, d3u8);
q4u16 = vmull_u8(d28u8, d3u8);
q5u16 = vmull_u8(d29u8, d3u8);
q6u16 = vmull_u8(d30u8, d3u8);
q7u16 = vmull_u8(d31u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q11s16 = vreinterpretq_s16_u16(q11u16);
q12s16 = vreinterpretq_s16_u16(q12u16);
q8s16 = vqaddq_s16(q8s16, q3s16);
q9s16 = vqaddq_s16(q9s16, q4s16);
q10s16 = vqaddq_s16(q10s16, q5s16);
q11s16 = vqaddq_s16(q11s16, q6s16);
q12s16 = vqaddq_s16(q12s16, q7s16);
d26u8 = vqrshrun_n_s16(q8s16, 7);
d27u8 = vqrshrun_n_s16(q9s16, 7);
d28u8 = vqrshrun_n_s16(q10s16, 7);
d29u8 = vqrshrun_n_s16(q11s16, 7);
d30u8 = vqrshrun_n_s16(q12s16, 7);
// Second pass: 8x8
dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
d0s8 = vdup_lane_s8(dtmps8, 0);
d1s8 = vdup_lane_s8(dtmps8, 1);
d2s8 = vdup_lane_s8(dtmps8, 2);
d3s8 = vdup_lane_s8(dtmps8, 3);
d4s8 = vdup_lane_s8(dtmps8, 4);
d5s8 = vdup_lane_s8(dtmps8, 5);
d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
tmpp = tmp;
q9u8 = vld1q_u8(tmpp);
tmpp += 16;
q10u8 = vld1q_u8(tmpp);
tmpp += 16;
q11u8 = vld1q_u8(tmpp);
tmpp += 16;
q12u8 = vld1q_u8(tmpp);
d18u8 = vget_low_u8(q9u8);
d19u8 = vget_high_u8(q9u8);
d20u8 = vget_low_u8(q10u8);
d21u8 = vget_high_u8(q10u8);
d22u8 = vget_low_u8(q11u8);
d23u8 = vget_high_u8(q11u8);
d24u8 = vget_low_u8(q12u8);
d25u8 = vget_high_u8(q12u8);
for (i = 2; i > 0; i--) {
q3u16 = vmull_u8(d18u8, d0u8);
q4u16 = vmull_u8(d19u8, d0u8);
q5u16 = vmull_u8(d20u8, d0u8);
q6u16 = vmull_u8(d21u8, d0u8);
q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
q7u16 = vmull_u8(d21u8, d3u8);
q8u16 = vmull_u8(d22u8, d3u8);
q9u16 = vmull_u8(d23u8, d3u8);
q10u16 = vmull_u8(d24u8, d3u8);
q3s16 = vreinterpretq_s16_u16(q3u16);
q4s16 = vreinterpretq_s16_u16(q4u16);
q5s16 = vreinterpretq_s16_u16(q5u16);
q6s16 = vreinterpretq_s16_u16(q6u16);
q7s16 = vreinterpretq_s16_u16(q7u16);
q8s16 = vreinterpretq_s16_u16(q8u16);
q9s16 = vreinterpretq_s16_u16(q9u16);
q10s16 = vreinterpretq_s16_u16(q10u16);
q7s16 = vqaddq_s16(q7s16, q3s16);
q8s16 = vqaddq_s16(q8s16, q4s16);
q9s16 = vqaddq_s16(q9s16, q5s16);
q10s16 = vqaddq_s16(q10s16, q6s16);
d6u8 = vqrshrun_n_s16(q7s16, 7);
d7u8 = vqrshrun_n_s16(q8s16, 7);
d8u8 = vqrshrun_n_s16(q9s16, 7);
d9u8 = vqrshrun_n_s16(q10s16, 7);
d18u8 = d22u8;
d19u8 = d23u8;
d20u8 = d24u8;
d21u8 = d25u8;
d22u8 = d26u8;
d23u8 = d27u8;
d24u8 = d28u8;
d25u8 = d29u8;
d26u8 = d30u8;
vst1_u8(dst_ptr, d6u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d7u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d8u8);
dst_ptr += dst_pitch;
vst1_u8(dst_ptr, d9u8);
dst_ptr += dst_pitch;
}
return;
}
uint8x8_t test_vget_low_u8(uint8x16_t a) {
// CHECK-LABEL: test_vget_low_u8:
return vget_low_u8(a);
// CHECK-NEXT: ret
}
unsigned int vpx_mse16x16_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
int64x1_t d0s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8;
int32x4_t q7s32, q8s32, q9s32, q10s32;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int64x2_t q1s64;
q7s32 = vdupq_n_s32(0);
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
for (i = 0; i < 8; i++) { // mse16x16_neon_loop
q0u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q1u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q2u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q3u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q7s32 = vaddq_s32(q7s32, q8s32);
q9s32 = vaddq_s32(q9s32, q10s32);
q10s32 = vaddq_s32(q7s32, q9s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
}
f64 dotProduct(const Size2D &_size,
const u8 * src0Base, ptrdiff_t src0Stride,
const u8 * src1Base, ptrdiff_t src1Stride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
Size2D size(_size);
if (src0Stride == src1Stride &&
src0Stride == (ptrdiff_t)(size.width))
{
size.width *= size.height;
size.height = 1;
}
// It is possible to accumulate up to 66051 uchar multiplication results in uint32 without overflow
// We process 16 elements and accumulate two new elements per step. So we could handle 66051/2*16 elements
#define DOT_UINT_BLOCKSIZE 66050*8
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const u8 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const u8 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
uint64x2_t ws = vmovq_n_u64(0);
while(i + 16 <= size.width)
{
size_t lim = std::min(i + DOT_UINT_BLOCKSIZE, size.width) - 16;
uint32x4_t s1 = vmovq_n_u32(0);
uint32x4_t s2 = vmovq_n_u32(0);
for (; i <= lim; i += 16)
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
uint8x16_t vs1 = vld1q_u8(src0 + i);
uint8x16_t vs2 = vld1q_u8(src1 + i);
uint16x8_t vdot1 = vmull_u8(vget_low_u8(vs1), vget_low_u8(vs2));
uint16x8_t vdot2 = vmull_u8(vget_high_u8(vs1), vget_high_u8(vs2));
s1 = vpadalq_u16(s1, vdot1);
s2 = vpadalq_u16(s2, vdot2);
}
ws = vpadalq_u32(ws, s1);
ws = vpadalq_u32(ws, s2);
}
if(i + 8 <= size.width)
{
uint8x8_t vs1 = vld1_u8(src0 + i);
uint8x8_t vs2 = vld1_u8(src1 + i);
ws = vpadalq_u32(ws, vpaddlq_u16(vmull_u8(vs1, vs2)));
i += 8;
}
result += (double)vget_lane_u64(vadd_u64(vget_low_u64(ws), vget_high_u64(ws)), 0);
for (; i < size.width; ++i)
result += s32(src0[i]) * s32(src1[i]);
}
return result;
#else
(void)_size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
return 0;
#endif
}
// CHECK-LABEL: define <8 x i8> @test_vget_low_u8(<16 x i8> %a) #0 {
// CHECK: [[SHUFFLE_I:%.*]] = shufflevector <16 x i8> %a, <16 x i8> %a, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
// CHECK: ret <8 x i8> [[SHUFFLE_I]]
uint8x8_t test_vget_low_u8(uint8x16_t a) {
return vget_low_u8(a);
}
void vp8_mbloop_filter_horizontal_edge_uv_neon(
unsigned char *u,
int pitch,
unsigned char blimit,
unsigned char limit,
unsigned char thresh,
unsigned char *v) {
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;
qblimit = vdupq_n_u8(blimit);
qlimit = vdupq_n_u8(limit);
qthresh = vdupq_n_u8(thresh);
u -= (pitch << 2);
v -= (pitch << 2);
d6 = vld1_u8(u);
u += pitch;
d7 = vld1_u8(v);
v += pitch;
d8 = vld1_u8(u);
u += pitch;
d9 = vld1_u8(v);
v += pitch;
d10 = vld1_u8(u);
u += pitch;
d11 = vld1_u8(v);
v += pitch;
d12 = vld1_u8(u);
u += pitch;
d13 = vld1_u8(v);
v += pitch;
d14 = vld1_u8(u);
u += pitch;
d15 = vld1_u8(v);
v += pitch;
d16 = vld1_u8(u);
u += pitch;
d17 = vld1_u8(v);
v += pitch;
d18 = vld1_u8(u);
u += pitch;
d19 = vld1_u8(v);
v += pitch;
d20 = vld1_u8(u);
d21 = vld1_u8(v);
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);
vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
q5, q6, q7, q8, q9, q10,
&q4, &q5, &q6, &q7, &q8, &q9);
u -= (pitch * 6);
v -= (pitch * 6);
vst1_u8(u, vget_low_u8(q4));
u += pitch;
vst1_u8(v, vget_high_u8(q4));
v += pitch;
vst1_u8(u, vget_low_u8(q5));
u += pitch;
vst1_u8(v, vget_high_u8(q5));
v += pitch;
vst1_u8(u, vget_low_u8(q6));
u += pitch;
vst1_u8(v, vget_high_u8(q6));
v += pitch;
vst1_u8(u, vget_low_u8(q7));
u += pitch;
vst1_u8(v, vget_high_u8(q7));
v += pitch;
vst1_u8(u, vget_low_u8(q8));
u += pitch;
vst1_u8(v, vget_high_u8(q8));
v += pitch;
vst1_u8(u, vget_low_u8(q9));
vst1_u8(v, vget_high_u8(q9));
return;
}
unsigned int vp8_variance16x8_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
for (i = 0; i < 4; i++) { // variance16x8_neon_loop
q0u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q1u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
__builtin_prefetch(src_ptr);
q2u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q3u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
__builtin_prefetch(ref_ptr);
q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
