void test_vmovQ_nu8 (void)
{
uint8x16_t out_uint8x16_t;
uint8_t arg0_uint8_t;
out_uint8x16_t = vmovq_n_u8 (arg0_uint8_t);
}
static uint8x16_t xtime(uint8x16_t x)
{
uint8x16_t y = vshlq_n_u8(x,1);
x = vshrq_n_u8(x,7);
uint8x16_t n27 = vmovq_n_u8(0x1b);
x = vmulq_u8(x,n27);
x = veorq_u8(x,y);
return x;
}
void add3 (uint8x16_t *data) {
/* Set each sixteen values of the vector to 3.
*
* Remark: a 'q' suffix to intrinsics indicates
* the instruction run for 128 bits registers.
*/
uint8x16_t three = vmovq_n_u8 (3);
/* Add 3 to the value given in argument. */
*data = vaddq_u8 (*data, three);
}
uint8x16_t test_vmovq_n_u8(uint8_t v1) {
// CHECK: test_vmovq_n_u8
return vmovq_n_u8(v1);
// CHECK: dup {{v[0-9]+}}.16b, {{w[0-9]+}}
}
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;
}
}
}
/* u8x16 mv mul */
void mw_neon_mv_mul_u8x16(unsigned char * A, int Row, int T, unsigned char * B, unsigned char * C)
{
int i = 0;
int k = 0;
uint8x16_t neon_b, neon_c;
uint8x16_t neon_a0, neon_a1, neon_a2, neon_a3, neon_a4, neon_a5, neon_a6, neon_a7;
uint8x16_t neon_a8, neon_a9, neon_a10, neon_a11, neon_a12, neon_a13, neon_a14, neon_a15;
uint8x16_t neon_b0, neon_b1, neon_b2, neon_b3, neon_b4, neon_b5, neon_b6, neon_b7;
uint8x16_t neon_b8, neon_b9, neon_b10, neon_b11, neon_b12, neon_b13, neon_b14, neon_b15;
for (i = 0; i < Row; i+=16)
{
neon_c = vmovq_n_u8(0);
for (k = 0; k < T; k+=16)
{
int j = k * T + i;
neon_a0 = vld1q_u8(A + j);
j+=Row;
neon_a1 = vld1q_u8(A + j);
j+=Row;
neon_a2 = vld1q_u8(A + j);
j+=Row;
neon_a3 = vld1q_u8(A + j);
j+=Row;
neon_a4 = vld1q_u8(A + j);
j+=Row;
neon_a5 = vld1q_u8(A + j);
j+=Row;
neon_a6 = vld1q_u8(A + j);
j+=Row;
neon_a7 = vld1q_u8(A + j);
j+=Row;
neon_a8 = vld1q_u8(A + j);
j+=Row;
neon_a9 = vld1q_u8(A + j);
j+=Row;
neon_a10 = vld1q_u8(A + j);
j+=Row;
neon_a11 = vld1q_u8(A + j);
j+=Row;
neon_a12 = vld1q_u8(A + j);
j+=Row;
neon_a13 = vld1q_u8(A + j);
j+=Row;
neon_a14 = vld1q_u8(A + j);
j+=Row;
neon_a15 = vld1q_u8(A + j);
neon_b = vld1q_u8(B + k);
neon_b0 = vdupq_n_u8(vgetq_lane_u8(neon_b, 0));
neon_b1 = vdupq_n_u8(vgetq_lane_u8(neon_b, 1));
neon_b2 = vdupq_n_u8(vgetq_lane_u8(neon_b, 2));
neon_b3 = vdupq_n_u8(vgetq_lane_u8(neon_b, 3));
neon_b4 = vdupq_n_u8(vgetq_lane_u8(neon_b, 4));
neon_b5 = vdupq_n_u8(vgetq_lane_u8(neon_b, 5));
neon_b6 = vdupq_n_u8(vgetq_lane_u8(neon_b, 6));
neon_b7 = vdupq_n_u8(vgetq_lane_u8(neon_b, 7));
neon_b8 = vdupq_n_u8(vgetq_lane_u8(neon_b, 8));
neon_b9 = vdupq_n_u8(vgetq_lane_u8(neon_b, 9));
neon_b10 = vdupq_n_u8(vgetq_lane_u8(neon_b, 10));
neon_b11 = vdupq_n_u8(vgetq_lane_u8(neon_b, 11));
neon_b12 = vdupq_n_u8(vgetq_lane_u8(neon_b, 12));
neon_b13 = vdupq_n_u8(vgetq_lane_u8(neon_b, 13));
neon_b14 = vdupq_n_u8(vgetq_lane_u8(neon_b, 14));
neon_b15 = vdupq_n_u8(vgetq_lane_u8(neon_b, 15));
neon_c = vaddq_u8(vmulq_u8(neon_a0, neon_b0), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a1, neon_b1), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a2, neon_b2), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a3, neon_b3), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a4, neon_b4), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a5, neon_b5), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a6, neon_b6), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a7, neon_b7), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a8, neon_b8), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a9, neon_b9), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a10, neon_b10), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a11, neon_b11), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a12, neon_b12), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a13, neon_b13), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a14, neon_b14), neon_c);
neon_c = vaddq_u8(vmulq_u8(neon_a15, neon_b15), neon_c);
}
vst1q_u8(C + i, neon_c);
}
}
