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_vget_highs64 (void)
{
int64x1_t out_int64x1_t;
int64x2_t arg0_int64x2_t;
out_int64x1_t = vget_high_s64 (arg0_int64x2_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);
}
static inline void DotProductWithScaleNeon(int32_t* cross_correlation,
const int16_t* vector1,
const int16_t* vector2,
size_t length,
int scaling) {
size_t i = 0;
size_t len1 = length >> 3;
size_t len2 = length & 7;
int64x2_t sum0 = vdupq_n_s64(0);
int64x2_t sum1 = vdupq_n_s64(0);
for (i = len1; i > 0; i -= 1) {
int16x8_t seq1_16x8 = vld1q_s16(vector1);
int16x8_t seq2_16x8 = vld1q_s16(vector2);
#if defined(WEBRTC_ARCH_ARM64)
int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8),
vget_low_s16(seq2_16x8));
int32x4_t tmp1 = vmull_high_s16(seq1_16x8, seq2_16x8);
#else
int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8),
vget_low_s16(seq2_16x8));
int32x4_t tmp1 = vmull_s16(vget_high_s16(seq1_16x8),
vget_high_s16(seq2_16x8));
#endif
sum0 = vpadalq_s32(sum0, tmp0);
sum1 = vpadalq_s32(sum1, tmp1);
vector1 += 8;
vector2 += 8;
}
// Calculate the rest of the samples.
int64_t sum_res = 0;
for (i = len2; i > 0; i -= 1) {
sum_res += WEBRTC_SPL_MUL_16_16(*vector1, *vector2);
vector1++;
vector2++;
}
sum0 = vaddq_s64(sum0, sum1);
#if defined(WEBRTC_ARCH_ARM64)
int64_t sum2 = vaddvq_s64(sum0);
*cross_correlation = (int32_t)((sum2 + sum_res) >> scaling);
#else
int64x1_t shift = vdup_n_s64(-scaling);
int64x1_t sum2 = vadd_s64(vget_low_s64(sum0), vget_high_s64(sum0));
sum2 = vadd_s64(sum2, vdup_n_s64(sum_res));
sum2 = vshl_s64(sum2, shift);
vst1_lane_s32(cross_correlation, vreinterpret_s32_s64(sum2), 0);
#endif
}
// 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;
}
}
int64x1_t test_vget_high_s64(int64x2_t a) {
// CHECK-COMMON-LABEL: test_vget_high_s64:
return vget_high_s64(a);
// CHECK-AARCH64: dup d0, {{v[0-9]+}}.d[1]
// CHECK-ARM64: ext v0.16b, v0.16b, v0.16b, #8
}
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);
}
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
}
void mdrc5b_apply_limiter(MDRC5B_LOCAL_STRUCT_T *HeapPtr)
{
unsigned int LaIdx;
unsigned int NumMainCh;
unsigned int Samples;
unsigned int ch, k, n;
MMlong *Ptr;
MMlong *Ptr2;
MMlong *MemOutPtr;
MMshort PeakdB;
MMlong PeakMax;
int RmsMeasure;
MMshort LimiterAtCoef;
MMshort LimiterReCoef;
MMshort LimiterGainMant[MDRC5B_BLOCK_SIZE + 1];
MMshort LimiterGainExp;
MMshort LimiterTargetGaindB;
unsigned int LimiterHoldRem;
unsigned int LimiterHtSamp;
MMshort Exp, TargetGain;
MMshort MaxShiftBits;
unsigned int lookahead_len = (unsigned int) HeapPtr->LimiterLALen;
unsigned int cpt1, cpt2;
uint32x2x2_t Temp_u32x2x2;
uint32x2_t Ldbits_u32x2, Ldbits2_u32x2;
uint32x2_t bsl_u32x2;
int32x2_t LimGainMant_32x2;
int64x2_t TempX_64x2, MemOut_64x2;
int64x2_t Tmp_64x2;
int64x2_t LimiterGainExp_64x2, sample_64x2;
int64x1_t TempX_64x1, sample_64x1;
int32_t *LimiterGainMant_ptr;
int32x2_t Tmp_32x2, Ldbits_32x2, n_32x2;
int32x2_t TempX_low_32x2, TempX_high_32x2;
int32x2x2_t Tmp_32x2x2;
int64x1_t Peak_64x1, PeakMax_64x1, Tmp_64x1, diffX_64x1;
int64x1_t Peak_scale_pow_64x1, Peak_scale_64x1, Zero_s64x1;
int64x1_t MaxShiftBits_neg_64x1, MaxShiftBits_hd_64x1;
int64x2_t diffX_64x2;
uint64x1_t bsl_u64x1;
int32x2_t LimiterPeakCoef_32x2, diffX_low_32x2, diffX_high_32x2;
int32x2_t TargetGain_32x2;
uint32x2x2_t Peak_u32x2x2;
uint32x2_t Peak_exp_u32x2, Peak_exp2_u32x2, Peak_mant_u32x2;
int32x2_t x_32x2, xn_32x2, PeakdB_32x2, Peak_exp_32x2;
int32x2_t LimiterTargetGaindB_32x2, Exp_32x2, LimiterCoef_32x2;
int32x4_t Tmp_32x4;
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_APPLY_LIMITER)
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
Samples = (unsigned int) HeapPtr->BlockSize;
NumMainCh = (unsigned int) HeapPtr->NumMainCh;
TempX_64x2 = vdupq_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
// compute the number of bits needs to be shifted to avoid overflow
for(k = (Samples >> 1); k > 0; k--)
{
sample_64x2 = vld1q_s64(Ptr);
Ptr +=2;
sample_64x2 = veorq_s64(sample_64x2, vshrq_n_s64(sample_64x2, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, sample_64x2);
}
if(Samples & 1)
{
sample_64x1 = vld1_s64(Ptr);
sample_64x1 = veor_s64(sample_64x1, vshr_n_s64(sample_64x1, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, vcombine_s64(sample_64x1, sample_64x1));
}
}
TempX_64x1 = vorr_s64(vget_low_s64(TempX_64x2), vget_high_s64(TempX_64x2));
Temp_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(TempX_64x1), vreinterpret_u32_s64(TempX_64x1));
bsl_u32x2 = vceq_u32(Temp_u32x2x2.val[1], vdup_n_u32(0)); // MSB == 0 ?
// use clz instead of cls because we are sure that input value is positive
// and because cls(LSB) could be wrong (if MSB is equal to 0 and bit 31 of LSL is 1)
// thus clz result will be 1 more than cls result (that's why you may see (Ldbits - 1)
// instead of Ldbits below)
Ldbits_u32x2 = vadd_u32(vclz_u32(Temp_u32x2x2.val[0]), vdup_n_u32(32)); // clz(LSB)+32
Ldbits2_u32x2 = vclz_u32(Temp_u32x2x2.val[1]); // clz(MSB)
Ldbits_u32x2 = vbsl_u32(bsl_u32x2, Ldbits_u32x2, Ldbits2_u32x2); // MSB == 0 ? clz(LSB)+32 : clz(MSB)
bsl_u32x2 = vceq_u32(Ldbits_u32x2, vdup_n_u32(64)); // Ldbits == 64 ? (i.e. TempX == 0 ?)
// the aim of MaxShiftBits is that sample will be shifted so that it occupies
// 24 significant bits for 24 bits samples or 32 significant bits for 32 bits samples
// but we are in 64 bits architecture on CA9/NEON
// so we must right shift of ((64 - 24) - (Ldbits - 1)) bits for 24 bits samples
// or of ((64 - 32) - (Ldbits - 1)) bits for 32 bits samples
// and we add 1 because it was done this way on MMDSP (I don't know why !)
#ifdef SAMPLES_24_BITS
// MaxShiftBits = ((64 - 24) - (Ldbits - 1)) + 1
// = 42 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(42), vreinterpret_s32_u32(Ldbits_u32x2));
#else // SAMPLES_24_BITS
// MaxShiftBits = ((64 - 32) - (Ldbits - 1)) + 1
// = 34 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(34), vreinterpret_s32_u32(Ldbits_u32x2));
#endif // SAMPLES_24_BITS
Ldbits_32x2 = vmax_s32(vdup_n_s32(1), Ldbits_32x2);
Ldbits_32x2 = vbsl_s32(bsl_u32x2, vdup_n_s32(1), Ldbits_32x2); // if(TempX == 0) Ldbits = 1
MaxShiftBits = vget_lane_s32(Ldbits_32x2, 0);
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT\n");
sprintf(string, "MaxShiftBits=%d\n", MaxShiftBits);
debug_write_string(string);
}
#endif // DEBUG_LIMITER_OUTPUT
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
// insert the new subband samples into the lookahead buffers
RmsMeasure = HeapPtr->Limiter.RmsMeasure;
LaIdx = (unsigned int) HeapPtr->LimiterLaIdx;
if(LaIdx + Samples >= lookahead_len)
{
cpt1 = lookahead_len - LaIdx;
cpt2 = Samples - cpt1;
// update index
HeapPtr->LimiterLaIdx = (int) cpt2;
}
else
{
cpt1 = Samples;
cpt2 = 0;
// update index
HeapPtr->LimiterLaIdx = (int) (LaIdx + Samples);
}
LimiterPeakCoef_32x2 = vdup_n_s32(HeapPtr->LimiterPeakAtCoef); // LimiterPeakAtCoef, LimiterPeakAtCoef
LimiterPeakCoef_32x2 = vset_lane_s32(HeapPtr->LimiterPeakReCoef, LimiterPeakCoef_32x2, 1); // LimiterPeakAtCoef, LimiterPeakReCoef
Peak_scale_64x1 = vdup_n_s64(HeapPtr->PrevShiftBits - MaxShiftBits);
Peak_scale_pow_64x1 = vshl_n_s64(Peak_scale_64x1, 1);
MaxShiftBits_neg_64x1 = vdup_n_s64(-MaxShiftBits);
#ifdef SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(24 - MaxShiftBits);
#else // SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(32 - MaxShiftBits);
#endif // SAMPLES_24_BITS
PeakMax_64x1 = vdup_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
Ptr2 = HeapPtr->LimiterLABuf[ch] + LaIdx; // go to the first valid sample
Peak_64x1 = vdup_n_s64(HeapPtr->LimiterPeak[ch]);
if(RmsMeasure)
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_pow_64x1); // neg value => shift right rounding
// rms measure
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
else
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_64x1);
// amplitude measure
Zero_s64x1 = vdup_n_s64(0);
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
HeapPtr->LimiterPeak[ch] = vget_lane_s64(Peak_64x1, 0); // save history
} // for(ch = 0...)
PeakMax = vget_lane_s64(PeakMax_64x1, 0);
HeapPtr->PrevShiftBits = MaxShiftBits;
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
if(PeakMax < MDRC5B_ALMOST_ZERO_THRESH)
{
PeakdB = (MDRC5B_POWER_DB_MINUS_INF << 16); // 8.16, [-128.0, 127.0] dB
}
else
{
Peak_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(PeakMax_64x1), vreinterpret_u32_s64(PeakMax_64x1));
bsl_u32x2 = vceq_u32(Peak_u32x2x2.val[1], vdup_n_u32(0));
Peak_exp_u32x2 = vadd_u32(vclz_u32(Peak_u32x2x2.val[0]), vdup_n_u32(32));
Peak_exp2_u32x2 = vclz_u32(Peak_u32x2x2.val[1]);
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, Peak_exp_u32x2, Peak_exp2_u32x2);
Peak_mant_u32x2 = vrshrn_n_u64(vshlq_u64(vreinterpretq_u64_s64(vcombine_s64(PeakMax_64x1, PeakMax_64x1)), vreinterpretq_s64_u64(vmovl_u32(Peak_exp_u32x2))), 32);
// if(Peak_mant >= sqrt(0.5))
// {
// Peak_exp--;
// Peak_mant >>= 1;
// }
bsl_u32x2 = vcge_u32(Peak_mant_u32x2, vdup_n_u32(0xB504F334));
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, vsub_u32(Peak_exp_u32x2, vdup_n_u32(1)), Peak_exp_u32x2);
Peak_mant_u32x2 = vbsl_u32(bsl_u32x2, vrshr_n_u32(Peak_mant_u32x2, 1), Peak_mant_u32x2);
Peak_exp_32x2 = vreinterpret_s32_u32(Peak_exp_u32x2);
#ifdef SAMPLES_24_BITS
// correction of 16 bits if input samples are 24 bits
Peak_exp_32x2 = vsub_s32(Peak_exp_32x2, vdup_n_s32(16));
#endif // SAMPLES_24_BITS
// at this point : sqrt(0.5)/2 <= Peak_mant < sqrt(0.5)
//
// ln(1+x) = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10 ... accuracy OK if |x| < 0.5
// sqrt(0.5)/2 <= Peak_mant < sqrt(0.5) => sqrt(0.5)-1 <= 2*Peak_mant-1 < 2*sqrt(0.5)-1
// => ln(Peak_mant) = ln(1+x)-ln(2) with x=2*Peak_mant-1, i.e. |x| < 0.414214...
// x=2*PeakMax_mant-1 in Q31
// => sqrt(0.5)-1 <= x < 2*sqrt(0.5)-1
x_32x2 = vreinterpret_s32_u32(vsub_u32(Peak_mant_u32x2, vdup_n_u32(0x80000000)));
PeakdB_32x2 = x_32x2; // PeakdB = x
xn_32x2 = vqrdmulh_s32(x_32x2, x_32x2); // xn = x^2
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 1)); // PeakdB = x - x^2/2
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^3
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x2AAAAAAB))); // PeakdB = x - x^2/2 + x^3/3
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^4
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 2)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^5
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x1999999A))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^6
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x15555555))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^7
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x12492492))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^8
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 3)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^9
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0E38E38E))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^10
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0CCCCCCD))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10
// at this point : PeakMaxdB contains ln(1+x) in Q31
if(RmsMeasure)
{
// dB(power) = 10*log10(power)
// PeakMaxdB = 10*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 10*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)-PeakMax_exp*10*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
//
// => RmsdB = 10/ln(10)*ln(1+x)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// => RmsdB (Q16) = 0x457CB*ln(1+x)+0x302A3*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// fractional mutiply 0x457CB*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x457CB));
// PeakdB_exp = 2*(HEADROOM+MaxShiftBits)-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(2 * (HEADROOM + MaxShiftBits)), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x302A3*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x302A3));
}
else
{
// dB(power) = 20*log10(abs)
// PeakMaxdB = 20*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 20*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)-PeakMax_exp*20*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
//
// => RmsdB = 20/ln(10)*ln(1+x)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
// => RmsdB (Q16) = 0x8AF96*ln(1+x)+0x60546*(HEADROOM+MaxShiftBits-PeakMax_exp)
// fractional mutiply 0x8AF96*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x8AF96));
// PeakdB_exp = HEADROOM+MaxShiftBits-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(HEADROOM + MaxShiftBits), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x60546*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x60546));
}
PeakdB = vget_lane_s32(PeakdB_32x2, 0);
}
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_PEAKMAX_PEAKDB\n");
sprintf(string, "PeakMax=0x%012llX, HEADROOM+MaxShiftBits=%d => PeakdB=0x%06X\n",
#ifdef SAMPLES_24_BITS
PeakMax & 0xFFFFFFFFFFFFLL,
#else // SAMPLES_24_BITS
(PeakMax >> 16) & 0xFFFFFFFFFFFFLL,
#endif // SAMPLES_24_BITS
HEADROOM + MaxShiftBits,
PeakdB & 0xFFFFFF);
debug_write_string(string);
}
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);
}
int64x1_t test_vget_high_s64(int64x2_t a) {
// CHECK-LABEL: test_vget_high_s64:
return vget_high_s64(a);
// CHECK: dup d0, {{v[0-9]+}}.d[1]
}
void silk_warped_autocorrelation_FIX_neon(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
)
{
if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
} else {
opus_int n, i, lsh;
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
opus_int64 corr_QC_orderT;
int64x2_t lsh_s64x2;
const opus_int orderT = ( order + 3 ) & ~3;
opus_int64 *corr_QCT;
opus_int32 *input_QS;
VARDECL( opus_int32, input_QST );
VARDECL( opus_int32, state );
SAVE_STACK;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
input_QS = input_QST;
/* input_QS has zero paddings in the beginning and end. */
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
/* Loop over samples */
for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
const int16x8_t t0_s16x4 = vld1q_s16( input + n );
vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
}
for( ; n < length; n++, input_QS++ ) {
input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
}
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
/* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */
/* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */
/* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
/* Keep the C code here to help understand the intrinsics optimization. */
/*
{
opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int32 *state_QST[ 3 ];
state_QST[ 0 ] = state_QS[ 0 ];
state_QST[ 1 ] = state_QS[ 1 ];
for( n = 0; n < length + order; n++, input_QS++ ) {
state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
for( i = 0; i < orderT; i++ ) {
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
}
state_QST[ 2 ] = state_QST[ 0 ];
state_QST[ 0 ] = state_QST[ 1 ];
state_QST[ 1 ] = state_QST[ 2 ];
}
}
*/
{
const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
const opus_int32 *in = input_QS + orderT;
opus_int o = orderT;
int32x4_t state_QS_s32x4[ 3 ][ 2 ];
ALLOC( state, length + orderT, opus_int32 );
state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
/* Calculate 8 taps of all inputs in each loop. */
do {
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
n = 0;
do {
calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
} while( ++n < ( length + order ) );
in = state;
o -= 8;
} while( o > 4 );
if( o ) {
/* Calculate the last 4 taps of all inputs. */
opus_int32 *stateT = state;
silk_assert( o == 4 );
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
n = length + order;
do {
calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
input_QS++;
stateT++;
} while( --n );
}
}
{
const opus_int16 *inputT = input;
int32x4_t t_s32x4;
int64x1_t t_s64x1;
int64x2_t t_s64x2 = vdupq_n_s64( 0 );
for( n = 0; n <= length - 8; n += 8 ) {
int16x8_t input_s16x8 = vld1q_s16( inputT );
t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
inputT += 8;
}
t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
for( ; n < length; n++ ) {
corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
}
corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
corr_QC[ orderT ] = corr_QC_orderT;
}
corr_QCT = corr_QC + orderT - order;
lsh = silk_CLZ64( corr_QC_orderT ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
lsh_s64x2 = vdupq_n_s64( lsh );
for( i = 0; i <= order - 3; i += 4 ) {
int32x4_t corr_s32x4;
int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
corr_s32x4 = vrev64q_s32( corr_s32x4 );
vst1q_s32( corr + order - i - 3, corr_s32x4 );
}
if( lsh >= 0 ) {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
}
} else {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
}
}
silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
RESTORE_STACK;
}
#ifdef OPUS_CHECK_ASM
{
opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int scale_c;
silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
silk_assert( scale_c == *scale );
}
#endif
}
// CHECK-LABEL: define <1 x i64> @test_vget_high_s64(<2 x i64> %a) #0 {
// CHECK: [[SHUFFLE_I:%.*]] = shufflevector <2 x i64> %a, <2 x i64> %a, <1 x i32> <i32 1>
// CHECK: ret <1 x i64> [[SHUFFLE_I]]
int64x1_t test_vget_high_s64(int64x2_t a) {
return vget_high_s64(a);
}
