int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
YV12_BUFFER_CONFIG *running_avg,
MACROBLOCK *signal, unsigned int motion_magnitude,
int y_offset, int uv_offset) {
/* 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.
*/
const uint8x16_t v_level1_adjustment = vdupq_n_u8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 : 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 = vdupq_n_u8(4);
const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
/* Local variables for array pointers and strides. */
unsigned char *sig = signal->thismb;
int sig_stride = 16;
unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
int mc_running_avg_y_stride = mc_running_avg->y_stride;
unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
int running_avg_y_stride = running_avg->y_stride;
/* Go over lines. */
int i;
int sum_diff = 0;
for (i = 0; i < 16; ++i) {
int8x16_t v_sum_diff = vdupq_n_s8(0);
uint8x16_t v_running_avg_y;
/* 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);
v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
v_sum_diff = vqaddq_s8(v_sum_diff,
vreinterpretq_s8_u8(v_pos_adjustment));
v_sum_diff = vqsubq_s8(v_sum_diff,
vreinterpretq_s8_u8(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.
*/
{
int s0 = vgetq_lane_s8(v_sum_diff, 0) +
vgetq_lane_s8(v_sum_diff, 1) +
vgetq_lane_s8(v_sum_diff, 2) +
vgetq_lane_s8(v_sum_diff, 3);
int s1 = vgetq_lane_s8(v_sum_diff, 4) +
vgetq_lane_s8(v_sum_diff, 5) +
vgetq_lane_s8(v_sum_diff, 6) +
vgetq_lane_s8(v_sum_diff, 7);
int s2 = vgetq_lane_s8(v_sum_diff, 8) +
vgetq_lane_s8(v_sum_diff, 9) +
vgetq_lane_s8(v_sum_diff, 10) +
vgetq_lane_s8(v_sum_diff, 11);
int s3 = vgetq_lane_s8(v_sum_diff, 12) +
vgetq_lane_s8(v_sum_diff, 13) +
vgetq_lane_s8(v_sum_diff, 14) +
vgetq_lane_s8(v_sum_diff, 15);
sum_diff += s0 + s1+ s2 + s3;
}
/* 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. */
if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
return COPY_BLOCK;
/* Tell above level that block was filtered. */
vp8_copy_mem16x16(running_avg->y_buffer + y_offset, running_avg_y_stride,
signal->thismb, sig_stride);
return FILTER_BLOCK;
}
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);
}
static INLINE void vp8_loop_filter_simple_horizontal_edge_neon(
unsigned char *s, int p, const unsigned char *blimit) {
uint8_t *sp;
uint8x16_t qblimit, q0u8;
uint8x16_t q5u8, q6u8, q7u8, q8u8, q9u8, q10u8, q14u8, q15u8;
int16x8_t q2s16, q3s16, q13s16;
int8x8_t d8s8, d9s8;
int8x16_t q2s8, q3s8, q4s8, q10s8, q11s8, q14s8;
qblimit = vdupq_n_u8(*blimit);
sp = s - (p << 1);
q5u8 = vld1q_u8(sp);
sp += p;
q6u8 = vld1q_u8(sp);
sp += p;
q7u8 = vld1q_u8(sp);
sp += p;
q8u8 = vld1q_u8(sp);
q15u8 = vabdq_u8(q6u8, q7u8);
q14u8 = vabdq_u8(q5u8, q8u8);
q15u8 = vqaddq_u8(q15u8, q15u8);
q14u8 = vshrq_n_u8(q14u8, 1);
q0u8 = vdupq_n_u8(0x80);
q13s16 = vdupq_n_s16(3);
q15u8 = vqaddq_u8(q15u8, q14u8);
q5u8 = veorq_u8(q5u8, q0u8);
q6u8 = veorq_u8(q6u8, q0u8);
q7u8 = veorq_u8(q7u8, q0u8);
q8u8 = veorq_u8(q8u8, q0u8);
q15u8 = vcgeq_u8(qblimit, q15u8);
q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7u8)),
vget_low_s8(vreinterpretq_s8_u8(q6u8)));
q3s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7u8)),
vget_high_s8(vreinterpretq_s8_u8(q6u8)));
q4s8 = vqsubq_s8(vreinterpretq_s8_u8(q5u8), vreinterpretq_s8_u8(q8u8));
q2s16 = vmulq_s16(q2s16, q13s16);
q3s16 = vmulq_s16(q3s16, q13s16);
q10u8 = vdupq_n_u8(3);
q9u8 = vdupq_n_u8(4);
q2s16 = vaddw_s8(q2s16, vget_low_s8(q4s8));
q3s16 = vaddw_s8(q3s16, vget_high_s8(q4s8));
d8s8 = vqmovn_s16(q2s16);
d9s8 = vqmovn_s16(q3s16);
q4s8 = vcombine_s8(d8s8, d9s8);
q14s8 = vandq_s8(q4s8, vreinterpretq_s8_u8(q15u8));
q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q10u8));
q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q9u8));
q2s8 = vshrq_n_s8(q2s8, 3);
q3s8 = vshrq_n_s8(q3s8, 3);
q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6u8), q2s8);
q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q7u8), q3s8);
q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
vst1q_u8(s, q7u8);
s -= p;
vst1q_u8(s, q6u8);
return;
}
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;
}
}
}
static INLINE void loop_filter_neon_16(uint8x16_t qblimit, // blimit
uint8x16_t qlimit, // limit
uint8x16_t qthresh, // thresh
uint8x16_t q3, // p3
uint8x16_t q4, // p2
uint8x16_t q5, // p1
uint8x16_t q6, // p0
uint8x16_t q7, // q0
uint8x16_t q8, // q1
uint8x16_t q9, // q2
uint8x16_t q10, // q3
uint8x16_t *q5r, // p1
uint8x16_t *q6r, // p0
uint8x16_t *q7r, // q0
uint8x16_t *q8r) { // q1
uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
int16x8_t q2s16, q11s16;
uint16x8_t q4u16;
int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
int8x8_t d2s8, d3s8;
q11u8 = vabdq_u8(q3, q4);
q12u8 = vabdq_u8(q4, q5);
q13u8 = vabdq_u8(q5, q6);
q14u8 = vabdq_u8(q8, q7);
q3 = vabdq_u8(q9, q8);
q4 = vabdq_u8(q10, q9);
q11u8 = vmaxq_u8(q11u8, q12u8);
q12u8 = vmaxq_u8(q13u8, q14u8);
q3 = vmaxq_u8(q3, q4);
q15u8 = vmaxq_u8(q11u8, q12u8);
q9 = vabdq_u8(q6, q7);
// vp8_hevmask
q13u8 = vcgtq_u8(q13u8, qthresh);
q14u8 = vcgtq_u8(q14u8, qthresh);
q15u8 = vmaxq_u8(q15u8, q3);
q2u8 = vabdq_u8(q5, q8);
q9 = vqaddq_u8(q9, q9);
q15u8 = vcgeq_u8(qlimit, q15u8);
// vp8_filter() function
// convert to signed
q10 = vdupq_n_u8(0x80);
q8 = veorq_u8(q8, q10);
q7 = veorq_u8(q7, q10);
q6 = veorq_u8(q6, q10);
q5 = veorq_u8(q5, q10);
q2u8 = vshrq_n_u8(q2u8, 1);
q9 = vqaddq_u8(q9, q2u8);
q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
vget_low_s8(vreinterpretq_s8_u8(q6)));
q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
vget_high_s8(vreinterpretq_s8_u8(q6)));
q9 = vcgeq_u8(qblimit, q9);
q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5), vreinterpretq_s8_u8(q8));
q14u8 = vorrq_u8(q13u8, q14u8);
q4u16 = vdupq_n_u16(3);
q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
q15u8 = vandq_u8(q15u8, q9);
q1s8 = vreinterpretq_s8_u8(q1u8);
q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
q4 = vdupq_n_u8(3);
q9 = vdupq_n_u8(4);
// vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
d2s8 = vqmovn_s16(q2s16);
d3s8 = vqmovn_s16(q11s16);
q1s8 = vcombine_s8(d2s8, d3s8);
q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
q1s8 = vreinterpretq_s8_u8(q1u8);
q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
q2s8 = vshrq_n_s8(q2s8, 3);
q1s8 = vshrq_n_s8(q1s8, 3);
q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
q1s8 = vrshrq_n_s8(q1s8, 1);
q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
*q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
*q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
*q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
*q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
return;
}
operator int8x16_t () const { return vreinterpretq_s8_u8 (val); }
static INLINE void vp8_mbloop_filter_neon(
uint8x16_t qblimit, // mblimit
uint8x16_t qlimit, // limit
uint8x16_t qthresh, // thresh
uint8x16_t q3, // p2
uint8x16_t q4, // p2
uint8x16_t q5, // p1
uint8x16_t q6, // p0
uint8x16_t q7, // q0
uint8x16_t q8, // q1
uint8x16_t q9, // q2
uint8x16_t q10, // q3
uint8x16_t *q4r, // p1
uint8x16_t *q5r, // p1
uint8x16_t *q6r, // p0
uint8x16_t *q7r, // q0
uint8x16_t *q8r, // q1
uint8x16_t *q9r) { // q1
uint8x16_t q0u8, q1u8, q11u8, q12u8, q13u8, q14u8, q15u8;
int16x8_t q0s16, q2s16, q11s16, q12s16, q13s16, q14s16, q15s16;
int8x16_t q1s8, q6s8, q7s8, q2s8, q11s8, q13s8;
uint16x8_t q0u16, q11u16, q12u16, q13u16, q14u16, q15u16;
int8x16_t q0s8, q12s8, q14s8, q15s8;
int8x8_t d0, d1, d2, d3, d4, d5, d24, d25, d28, d29;
q11u8 = vabdq_u8(q3, q4);
q12u8 = vabdq_u8(q4, q5);
q13u8 = vabdq_u8(q5, q6);
q14u8 = vabdq_u8(q8, q7);
q1u8 = vabdq_u8(q9, q8);
q0u8 = vabdq_u8(q10, q9);
q11u8 = vmaxq_u8(q11u8, q12u8);
q12u8 = vmaxq_u8(q13u8, q14u8);
q1u8 = vmaxq_u8(q1u8, q0u8);
q15u8 = vmaxq_u8(q11u8, q12u8);
q12u8 = vabdq_u8(q6, q7);
// vp8_hevmask
q13u8 = vcgtq_u8(q13u8, qthresh);
q14u8 = vcgtq_u8(q14u8, qthresh);
q15u8 = vmaxq_u8(q15u8, q1u8);
q15u8 = vcgeq_u8(qlimit, q15u8);
q1u8 = vabdq_u8(q5, q8);
q12u8 = vqaddq_u8(q12u8, q12u8);
// vp8_filter() function
// convert to signed
q0u8 = vdupq_n_u8(0x80);
q9 = veorq_u8(q9, q0u8);
q8 = veorq_u8(q8, q0u8);
q7 = veorq_u8(q7, q0u8);
q6 = veorq_u8(q6, q0u8);
q5 = veorq_u8(q5, q0u8);
q4 = veorq_u8(q4, q0u8);
q1u8 = vshrq_n_u8(q1u8, 1);
q12u8 = vqaddq_u8(q12u8, q1u8);
q14u8 = vorrq_u8(q13u8, q14u8);
q12u8 = vcgeq_u8(qblimit, q12u8);
q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
vget_low_s8(vreinterpretq_s8_u8(q6)));
q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
vget_high_s8(vreinterpretq_s8_u8(q6)));
q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
vreinterpretq_s8_u8(q8));
q11s16 = vdupq_n_s16(3);
q2s16 = vmulq_s16(q2s16, q11s16);
q13s16 = vmulq_s16(q13s16, q11s16);
q15u8 = vandq_u8(q15u8, q12u8);
q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
q13s16 = vaddw_s8(q13s16, vget_high_s8(q1s8));
q12u8 = vdupq_n_u8(3);
q11u8 = vdupq_n_u8(4);
// vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
d2 = vqmovn_s16(q2s16);
d3 = vqmovn_s16(q13s16);
q1s8 = vcombine_s8(d2, d3);
q1s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q15u8));
q13s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
q2s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q11u8));
q13s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q12u8));
q2s8 = vshrq_n_s8(q2s8, 3);
q13s8 = vshrq_n_s8(q13s8, 3);
q7s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q2s8);
q6s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q13s8);
q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
q0u16 = q11u16 = q12u16 = q13u16 = q14u16 = q15u16 = vdupq_n_u16(63);
d5 = vdup_n_s8(9);
d4 = vdup_n_s8(18);
q0s16 = vmlal_s8(vreinterpretq_s16_u16(q0u16), vget_low_s8(q1s8), d5);
q11s16 = vmlal_s8(vreinterpretq_s16_u16(q11u16), vget_high_s8(q1s8), d5);
d5 = vdup_n_s8(27);
q12s16 = vmlal_s8(vreinterpretq_s16_u16(q12u16), vget_low_s8(q1s8), d4);
q13s16 = vmlal_s8(vreinterpretq_s16_u16(q13u16), vget_high_s8(q1s8), d4);
q14s16 = vmlal_s8(vreinterpretq_s16_u16(q14u16), vget_low_s8(q1s8), d5);
q15s16 = vmlal_s8(vreinterpretq_s16_u16(q15u16), vget_high_s8(q1s8), d5);
d0 = vqshrn_n_s16(q0s16 , 7);
d1 = vqshrn_n_s16(q11s16, 7);
d24 = vqshrn_n_s16(q12s16, 7);
d25 = vqshrn_n_s16(q13s16, 7);
d28 = vqshrn_n_s16(q14s16, 7);
d29 = vqshrn_n_s16(q15s16, 7);
q0s8 = vcombine_s8(d0, d1);
q12s8 = vcombine_s8(d24, d25);
q14s8 = vcombine_s8(d28, d29);
q11s8 = vqsubq_s8(vreinterpretq_s8_u8(q9), q0s8);
q0s8 = vqaddq_s8(vreinterpretq_s8_u8(q4), q0s8);
q13s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q12s8);
q12s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q12s8);
q15s8 = vqsubq_s8((q7s8), q14s8);
q14s8 = vqaddq_s8((q6s8), q14s8);
q1u8 = vdupq_n_u8(0x80);
*q9r = veorq_u8(vreinterpretq_u8_s8(q11s8), q1u8);
*q8r = veorq_u8(vreinterpretq_u8_s8(q13s8), q1u8);
*q7r = veorq_u8(vreinterpretq_u8_s8(q15s8), q1u8);
*q6r = veorq_u8(vreinterpretq_u8_s8(q14s8), q1u8);
*q5r = veorq_u8(vreinterpretq_u8_s8(q12s8), q1u8);
*q4r = veorq_u8(vreinterpretq_u8_s8(q0s8), q1u8);
return;
}
