static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
const int32x4_t a = vpaddlq_s16(v_16x8);
const int64x2_t b = vpaddlq_s32(a);
const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
vreinterpret_s32_s64(vget_high_s64(b)));
return vget_lane_s32(c, 0);
}
void test_vpaddlQs32 (void)
{
int64x2_t out_int64x2_t;
int32x4_t arg0_int32x4_t;
out_int64x2_t = vpaddlq_s32 (arg0_int32x4_t);
}
unsigned int vpx_get4x4sse_cs_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride) {
int16x4_t d22s16, d24s16, d26s16, d28s16;
int64x1_t d0s64;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
int32x4_t q7s32, q8s32, q9s32, q10s32;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int64x2_t q1s64;
d0u8 = vld1_u8(src_ptr);
src_ptr += source_stride;
d4u8 = vld1_u8(ref_ptr);
ref_ptr += recon_stride;
d1u8 = vld1_u8(src_ptr);
src_ptr += source_stride;
d5u8 = vld1_u8(ref_ptr);
ref_ptr += recon_stride;
d2u8 = vld1_u8(src_ptr);
src_ptr += source_stride;
d6u8 = vld1_u8(ref_ptr);
ref_ptr += recon_stride;
d3u8 = vld1_u8(src_ptr);
src_ptr += source_stride;
d7u8 = vld1_u8(ref_ptr);
ref_ptr += recon_stride;
q11u16 = vsubl_u8(d0u8, d4u8);
q12u16 = vsubl_u8(d1u8, d5u8);
q13u16 = vsubl_u8(d2u8, d6u8);
q14u16 = vsubl_u8(d3u8, d7u8);
d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));
q7s32 = vmull_s16(d22s16, d22s16);
q8s32 = vmull_s16(d24s16, d24s16);
q9s32 = vmull_s16(d26s16, d26s16);
q10s32 = vmull_s16(d28s16, d28s16);
q7s32 = vaddq_s32(q7s32, q8s32);
q9s32 = vaddq_s32(q9s32, q10s32);
q9s32 = vaddq_s32(q7s32, q9s32);
q1s64 = vpaddlq_s32(q9s32);
d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
}
// ref, src = [0, 510] - max diff = 16-bits
// bwl = {2, 3, 4}, width = {16, 32, 64}
int vp9_vector_var_neon(int16_t const *ref, int16_t const *src, const int bwl) {
int width = 4 << bwl;
int32x4_t sse = vdupq_n_s32(0);
int16x8_t total = vdupq_n_s16(0);
assert(width >= 8);
assert((width % 8) == 0);
do {
const int16x8_t r = vld1q_s16(ref);
const int16x8_t s = vld1q_s16(src);
const int16x8_t diff = vsubq_s16(r, s); // [-510, 510], 10 bits.
const int16x4_t diff_lo = vget_low_s16(diff);
const int16x4_t diff_hi = vget_high_s16(diff);
sse = vmlal_s16(sse, diff_lo, diff_lo); // dynamic range 26 bits.
sse = vmlal_s16(sse, diff_hi, diff_hi);
total = vaddq_s16(total, diff); // dynamic range 16 bits.
ref += 8;
src += 8;
width -= 8;
} while (width != 0);
{
// Note: 'total''s pairwise addition could be implemented similarly to
// horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
// with the summation of 'sse' performed better on a Cortex-A15.
const int32x4_t t0 = vpaddlq_s16(total); // cascading summation of 'total'
const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
const int32x2_t t2 = vpadd_s32(t1, t1);
const int t = vget_lane_s32(t2, 0);
const int64x2_t s0 = vpaddlq_s32(sse); // cascading summation of 'sse'.
const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
vreinterpret_s32_s64(vget_high_s64(s0)));
const int s = vget_lane_s32(s1, 0);
const int shift_factor = bwl + 2;
return s - ((t * t) >> shift_factor);
}
}
// coeff: 16 bits, dynamic range [-32640, 32640].
// length: value range {16, 64, 256, 1024}.
int aom_satd_neon(const int16_t *coeff, int length) {
const int16x4_t zero = vdup_n_s16(0);
int32x4_t accum = vdupq_n_s32(0);
do {
const int16x8_t src0 = vld1q_s16(coeff);
const int16x8_t src8 = vld1q_s16(coeff + 8);
accum = vabal_s16(accum, vget_low_s16(src0), zero);
accum = vabal_s16(accum, vget_high_s16(src0), zero);
accum = vabal_s16(accum, vget_low_s16(src8), zero);
accum = vabal_s16(accum, vget_high_s16(src8), zero);
length -= 16;
coeff += 16;
} while (length != 0);
{
// satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
const int64x2_t s0 = vpaddlq_s32(accum); // cascading summation of 'accum'.
const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
vreinterpret_s32_s64(vget_high_s64(s0)));
const int satd = vget_lane_s32(s1, 0);
return satd;
}
}
int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
int mc_running_avg_y_stride,
unsigned char *running_avg_y,
int running_avg_y_stride,
unsigned char *sig, int sig_stride,
unsigned int motion_magnitude,
int increase_denoising) {
/* If motion_magnitude is small, making the denoiser more aggressive by
* increasing the adjustment for each level, level1 adjustment is
* increased, the deltas stay the same.
*/
int shift_inc = (increase_denoising &&
motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
const uint8x16_t v_level1_adjustment = vmovq_n_u8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 + shift_inc : 3);
const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
int64x2_t v_sum_diff_total = vdupq_n_s64(0);
/* Go over lines. */
int r;
for (r = 0; r < 16; ++r) {
/* Load inputs. */
const uint8x16_t v_sig = vld1q_u8(sig);
const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
/* Calculate absolute difference and sign masks. */
const uint8x16_t v_abs_diff = vabdq_u8(v_sig, v_mc_running_avg_y);
const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
/* Figure out which level that put us in. */
const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold,
v_abs_diff);
const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold,
v_abs_diff);
const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold,
v_abs_diff);
/* Calculate absolute adjustments for level 1, 2 and 3. */
const uint8x16_t v_level2_adjustment = vandq_u8(v_level2_mask,
v_delta_level_1_and_2);
const uint8x16_t v_level3_adjustment = vandq_u8(v_level3_mask,
v_delta_level_2_and_3);
const uint8x16_t v_level1and2_adjustment = vaddq_u8(v_level1_adjustment,
v_level2_adjustment);
const uint8x16_t v_level1and2and3_adjustment = vaddq_u8(
v_level1and2_adjustment, v_level3_adjustment);
/* Figure adjustment absolute value by selecting between the absolute
* difference if in level0 or the value for level 1, 2 and 3.
*/
const uint8x16_t v_abs_adjustment = vbslq_u8(v_level1_mask,
v_level1and2and3_adjustment, v_abs_diff);
/* Calculate positive and negative adjustments. Apply them to the signal
* and accumulate them. Adjustments are less than eight and the maximum
* sum of them (7 * 16) can fit in a signed char.
*/
const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
v_abs_adjustment);
const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
v_abs_adjustment);
uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
/* Store results. */
vst1q_u8(running_avg_y, v_running_avg_y);
/* Sum all the accumulators to have the sum of all pixel differences
* for this macroblock.
*/
{
const int8x16_t v_sum_diff =
vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
vreinterpretq_s8_u8(v_neg_adjustment));
const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
const int32x4_t fedc_ba98_7654_3210 =
vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
const int64x2_t fedcba98_76543210 =
vpaddlq_s32(fedc_ba98_7654_3210);
v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
}
/* Update pointers for next iteration. */
sig += sig_stride;
mc_running_avg_y += mc_running_avg_y_stride;
running_avg_y += running_avg_y_stride;
}
/* Too much adjustments => copy block. */
{
int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
vget_low_s64(v_sum_diff_total));
int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
int sum_diff_thresh = SUM_DIFF_THRESHOLD;
if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
if (sum_diff > sum_diff_thresh) {
// Before returning to copy the block (i.e., apply no denoising),
// checK if we can still apply some (weaker) temporal filtering to
// this block, that would otherwise not be denoised at all. Simplest
// is to apply an additional adjustment to running_avg_y to bring it
// closer to sig. The adjustment is capped by a maximum delta, and
// chosen such that in most cases the resulting sum_diff will be
// within the accceptable range given by sum_diff_thresh.
// The delta is set by the excess of absolute pixel diff over the
// threshold.
int delta = ((sum_diff - sum_diff_thresh) >> 8) + 1;
// Only apply the adjustment for max delta up to 3.
if (delta < 4) {
const uint8x16_t k_delta = vmovq_n_u8(delta);
sig -= sig_stride * 16;
mc_running_avg_y -= mc_running_avg_y_stride * 16;
running_avg_y -= running_avg_y_stride * 16;
for (r = 0; r < 16; ++r) {
uint8x16_t v_running_avg_y = vld1q_u8(running_avg_y);
const uint8x16_t v_sig = vld1q_u8(sig);
const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
/* Calculate absolute difference and sign masks. */
const uint8x16_t v_abs_diff = vabdq_u8(v_sig,
v_mc_running_avg_y);
const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig,
v_mc_running_avg_y);
const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig,
v_mc_running_avg_y);
// Clamp absolute difference to delta to get the adjustment.
const uint8x16_t v_abs_adjustment =
vminq_u8(v_abs_diff, (k_delta));
const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
v_abs_adjustment);
const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
v_abs_adjustment);
v_running_avg_y = vqsubq_u8(v_running_avg_y, v_pos_adjustment);
v_running_avg_y = vqaddq_u8(v_running_avg_y, v_neg_adjustment);
/* Store results. */
vst1q_u8(running_avg_y, v_running_avg_y);
{
const int8x16_t v_sum_diff =
vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
vreinterpretq_s8_u8(v_pos_adjustment));
const int16x8_t fe_dc_ba_98_76_54_32_10 =
vpaddlq_s8(v_sum_diff);
const int32x4_t fedc_ba98_7654_3210 =
vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
const int64x2_t fedcba98_76543210 =
vpaddlq_s32(fedc_ba98_7654_3210);
v_sum_diff_total = vqaddq_s64(v_sum_diff_total,
fedcba98_76543210);
}
/* Update pointers for next iteration. */
sig += sig_stride;
mc_running_avg_y += mc_running_avg_y_stride;
running_avg_y += running_avg_y_stride;
}
{
// Update the sum of all pixel differences of this MB.
x = vqadd_s64(vget_high_s64(v_sum_diff_total),
vget_low_s64(v_sum_diff_total));
sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
if (sum_diff > sum_diff_thresh) {
return COPY_BLOCK;
}
}
} else {
return COPY_BLOCK;
}
}
}
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);
}
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);
}
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);
}