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);
}
// 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);
}
}
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;
}
}
}
f64 dotProduct(const Size2D &_size,
const s8 * src0Base, ptrdiff_t src0Stride,
const s8 * 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 131071 schar multiplication results in sint32 without overflow
// We process 16 elements and accumulate two new elements per step. So we could handle 131071/2*16 elements
#define DOT_INT_BLOCKSIZE 131070*8
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const s8 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const s8 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
int64x2_t ws = vmovq_n_s64(0);
while(i + 16 <= size.width)
{
size_t lim = std::min(i + DOT_UINT_BLOCKSIZE, size.width) - 16;
int32x4_t s1 = vmovq_n_s32(0);
int32x4_t s2 = vmovq_n_s32(0);
for (; i <= lim; i += 16)
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
int8x16_t vs1 = vld1q_s8(src0 + i);
int8x16_t vs2 = vld1q_s8(src1 + i);
int16x8_t vdot1 = vmull_s8(vget_low_s8(vs1), vget_low_s8(vs2));
int16x8_t vdot2 = vmull_s8(vget_high_s8(vs1), vget_high_s8(vs2));
s1 = vpadalq_s16(s1, vdot1);
s2 = vpadalq_s16(s2, vdot2);
}
ws = vpadalq_s32(ws, s1);
ws = vpadalq_s32(ws, s2);
}
if(i + 8 <= size.width)
{
int8x8_t vs1 = vld1_s8(src0 + i);
int8x8_t vs2 = vld1_s8(src1 + i);
ws = vpadalq_s32(ws, vpaddlq_s16(vmull_s8(vs1, vs2)));
i += 8;
}
result += (double)vget_lane_s64(vadd_s64(vget_low_s64(ws), vget_high_s64(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
}
