static void
TEST (void)
{
union
{
__m128i x[NUM / 2];
long long ll[NUM];
} dst;
union
{
__m128i x[NUM / 2];
int i[NUM * 2];
} src1, src2;
int i, sign = 1;
long long value;
for (i = 0; i < NUM * 2; i += 2)
{
src1.i[i] = i * i * sign;
src2.i[i] = (i + 20) * sign;
sign = -sign;
}
for (i = 0; i < NUM; i += 2)
dst.x[i / 2] = _mm_mul_epi32 (src1.x[i / 2], src2.x[i / 2]);
for (i = 0; i < NUM; i++)
{
value = (long long) src1.i[i * 2] * (long long) src2.i[i * 2];
if (value != dst.ll[i])
abort ();
}
}
int oneThread(int threadId)
{
int *aa;
int *bb;
int k;
int itr;
aa = (int *)_mm_malloc(sizeof(int)*ARRAY_SIZE, 16);
bb = (int *)_mm_malloc(sizeof(int)*ARRAY_SIZE, 16);
memset(&aa[0], 1, ARRAY_SIZE*4);
memset(&bb[0], 2, ARRAY_SIZE*4);
__m128i a0,a1,a2,a3,b0,b1,b2,b3;
__m128i a4,a5,a6,a7,b4,b5,b6,b7;
__m128i c0,c1,c2,c3;
__m128i c4,c5,c6,c7;
__m128i cc;
cc = _mm_set_epi32 (0, 0, 0, 0);
for (k = 0; k < REPS; k++)
{
for (itr = 0; itr<ARRAY_SIZE; itr+=32)
{
a0 = _mm_load_si128((__m128i*)&aa[itr]);
a1 = _mm_load_si128((__m128i*)&aa[itr+4]);
a2 = _mm_load_si128((__m128i*)&aa[itr+8]);
a3 = _mm_load_si128((__m128i*)&aa[itr+12]);
a4 = _mm_load_si128((__m128i*)&aa[itr+16]);
a5 = _mm_load_si128((__m128i*)&aa[itr+20]);
a6 = _mm_load_si128((__m128i*)&aa[itr+24]);
a7 = _mm_load_si128((__m128i*)&aa[itr+28]);
b0 = _mm_load_si128((__m128i*)&bb[itr]);
b1 = _mm_load_si128((__m128i*)&bb[itr+4]);
b2 = _mm_load_si128((__m128i*)&bb[itr+8]);
b3 = _mm_load_si128((__m128i*)&bb[itr+12]);
b4 = _mm_load_si128((__m128i*)&bb[itr+16]);
b5 = _mm_load_si128((__m128i*)&bb[itr+20]);
b6 = _mm_load_si128((__m128i*)&bb[itr+24]);
b7 = _mm_load_si128((__m128i*)&bb[itr+28]);
c0 = _mm_mul_epi32(a0, b0);
c1 = _mm_mul_epi32(a1, b1);
c2 = _mm_mul_epi32(a2, b2);
c3 = _mm_mul_epi32(a3, b3);
c4 = _mm_mul_epi32(a4, b4);
c5 = _mm_mul_epi32(a5, b5);
c6 = _mm_mul_epi32(a6, b6);
c7 = _mm_mul_epi32(a7, b7);
c0 = _mm_add_epi32(c0,c1);
c1 = _mm_add_epi32(c2,c3);
c2 = _mm_add_epi32(c4,c5);
c3 = _mm_add_epi32(c6,c7);
c0 = _mm_add_epi32(c0,c1);
c1 = _mm_add_epi32(c2,c3);
c0 = _mm_add_epi32(c0,c1);
cc = _mm_add_epi32(cc,c0);
}
}
cc = _mm_hadd_epi32(cc,cc);
cc = _mm_hadd_epi32(cc,cc);
int count =0;
count = _mm_cvtsi128_si32(cc) ;
free(aa);
free(bb);
return count;
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified_sse4_1(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D, /* I Order */
int arch /* I Run-time architecture */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 xcorr[ SILK_MAX_ORDER_LPC ];
__m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210;
__m128i CONST1 = _mm_set1_epi32(1);
silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE);
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length);
if(rshifts > MAX_RSHIFTS) {
C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS);
silk_assert(C0 > 0);
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32(C0) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if(rshifts_extra > 0) {
rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts);
C0 = silk_RSHIFT32(C0, rshifts_extra);
} else {
rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts);
C0 = silk_LSHIFT32(C0, -rshifts_extra);
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts);
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
int i;
opus_int32 d;
x_ptr = x + s * subfr_length;
celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch);
for(n = 1; n < D + 1; n++) {
for (i = n + subfr_length - D, d = 0; i < subfr_length; i++)
d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]);
xcorr[ n - 1 ] += d;
}
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts);
}
}
}
silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for(n = 0; n < D; n++) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if(rshifts > -2) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */
for(k = 0; k < n; k++) {
C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */
for(k = 0; k <= n; k++) {
CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */
X1_3210 = _mm_set1_epi32(x1);
X2_3210 = _mm_set1_epi32(x2);
TMP1_3210 = _mm_setzero_si128();
TMP2_3210 = _mm_setzero_si128();
for(k = 0; k < n - 3; k += 4) {
PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]);
SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]);
FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]);
PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3));
LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]);
ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]);
T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210);
T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7);
ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1);
FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210);
LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210);
PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210);
SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210);
_mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210);
_mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210);
TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210);
TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210);
}
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210));
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E));
tmp1 += _mm_cvtsi128_si32(TMP1_3210);
tmp2 += _mm_cvtsi128_si32(TMP2_3210);
for(; k < n; k++) {
C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */
tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */
tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
{
__m128i xmm_tmp1, xmm_tmp2;
__m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1;
__m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1;
xmm_tmp1 = _mm_set1_epi32(tmp1);
xmm_tmp2 = _mm_set1_epi32(tmp2);
for(k = 0; k <= n - 3; k += 4) {
xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]);
xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]);
xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3));
xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1);
xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1);
/* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/
xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1);
xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1);
xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2);
xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2);
xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16);
xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16);
xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16);
xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16);
xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC);
xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC);
X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]);
PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]);
X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0);
PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0);
_mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210);
_mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210);
}
for(; k <= n; k++) {
CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1,
silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2,
silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */
}
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q(-rshifts) */
tmp2 = C_last_row[ n ]; /* Q(-rshifts) */
num = 0; /* Q(-rshifts) */
nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */
for(k = 0; k < n; k++) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1;
lz = silk_min(32 - QA, lz);
Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */
tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]),
Atmp1), 32 - QA - lz); /* Q(1-rshifts) */
}
CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */
CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */
num = silk_ADD32(num, tmp2); /* Q(-rshifts) */
num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */
/* Calculate the next order reflection (parcor) coefficient */
if(silk_abs(num) < nrg) {
rc_Q31 = silk_DIV32_varQ(num, nrg, 31);
} else {
rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31);
tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2);
if(tmp1 <= minInvGain_Q30) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */
rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */
rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */
if(num < 0) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for(k = 0; k < (n + 1) >> 1; k++) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */
if(reached_max_gain) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for(k = n + 1; k < D; k++) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for(k = 0; k <= n + 1; k++) {
tmp1 = CAf[ k ]; /* Q(-rshifts) */
tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */
CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */
}
}
if(reached_max_gain) {
for(k = 0; k < D; k++) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);
}
/* Subtract energy of preceding samples from C0 */
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts);
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts);
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2);
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q(-rshifts) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for(k = 0; k < D; k++) {
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */
nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */
tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */
*res_nrg_Q = -rshifts;
}
}
__m128i test_mm_mul_epi32(__m128i x, __m128i y) {
// CHECK-LABEL: test_mm_mul_epi32
// CHECK: call <2 x i64> @llvm.x86.sse41.pmuldq
// CHECK-ASM: pmuldq %xmm{{.*}}, %xmm{{.*}}
return _mm_mul_epi32(x, y);
}
BOOST_FORCEINLINE __m128i __vectorcall operator * ( __m128i const left, __m128i const right ) {
return _mm_mul_epi32( left, right );
}
__m128i test_mm_mul_epi32(__m128i x, __m128i y) {
// CHECK-LABEL: test_mm_mul_epi32
// CHECK: call <2 x i64> @llvm.x86.sse41.pmuldq(<4 x i32> %{{.*}}, <4 x i32> %{{.*}})
return _mm_mul_epi32(x, y);
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC_sse4_1(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int16 *in_Q14, /* I input vector to be quantized */
const opus_int32 *W_Q18, /* I weighting matrix */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int16 diff_Q14[ 5 ];
opus_int32 sum1_Q14, sum2_Q16;
__m128i C_tmp1, C_tmp2, C_tmp3, C_tmp4, C_tmp5;
/* Loop over codebook */
*rate_dist_Q14 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
for( k = 0; k < L; k++ ) {
gain_tmp_Q7 = cb_gain_Q7[k];
diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 );
C_tmp1 = OP_CVTEPI16_EPI32_M64( &in_Q14[ 1 ] );
C_tmp2 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] );
C_tmp2 = _mm_slli_epi32( C_tmp2, 7 );
C_tmp1 = _mm_sub_epi32( C_tmp1, C_tmp2 );
diff_Q14[ 1 ] = _mm_extract_epi16( C_tmp1, 0 );
diff_Q14[ 2 ] = _mm_extract_epi16( C_tmp1, 2 );
diff_Q14[ 3 ] = _mm_extract_epi16( C_tmp1, 4 );
diff_Q14[ 4 ] = _mm_extract_epi16( C_tmp1, 6 );
/* Weighted rate */
sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] );
/* Penalty for too large gain */
sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 );
silk_assert( sum1_Q14 >= 0 );
/* first row of W_Q18 */
C_tmp3 = _mm_loadu_si128( (__m128i *)(&W_Q18[ 1 ] ) );
C_tmp4 = _mm_mul_epi32( C_tmp3, C_tmp1 );
C_tmp4 = _mm_srli_si128( C_tmp4, 2 );
C_tmp1 = _mm_shuffle_epi32( C_tmp1, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
C_tmp3 = _mm_shuffle_epi32( C_tmp3, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
C_tmp5 = _mm_mul_epi32( C_tmp3, C_tmp1 );
C_tmp5 = _mm_srli_si128( C_tmp5, 2 );
C_tmp5 = _mm_add_epi32( C_tmp4, C_tmp5 );
C_tmp5 = _mm_slli_epi32( C_tmp5, 1 );
C_tmp5 = _mm_add_epi32( C_tmp5, _mm_shuffle_epi32( C_tmp5, _MM_SHUFFLE( 0, 0, 0, 2 ) ) );
sum2_Q16 = _mm_cvtsi128_si32( C_tmp5 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] );
/* second row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] );
/* third row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] );
/* fourth row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] );
/* last row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] );
silk_assert( sum1_Q14 >= 0 );
/* find best */
if( sum1_Q14 < *rate_dist_Q14 ) {
*rate_dist_Q14 = sum1_Q14;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}