inline avx_m256_t newsin_ps(avx_m256_t x) {
avx_m256_t sign_bit = _mm256_and_ps(x, _ps_sign_mask);
x = _mm256_and_ps(x, _ps_inv_sign_mask);
avx_m256_t y = _mm256_mul_ps(x, _ps_cephes_FOPI);
avx_m256i_t emm2 = _mm256_cvttps_epi32(y);
emm2 = _mm256_add_epi32(emm2, _pi32_1);
emm2 = _mm256_and_si256(emm2, _pi32_inv1);
y = _mm256_cvtepi32_ps(emm2);
avx_m256i_t emm0 = _mm256_and_si256(emm2, _pi32_4);
emm0 = _mm256_slli_epi32(emm0, 29);
emm2 = _mm256_and_si256(emm2, _pi32_2);
emm2 = _mm256_cmpeq_epi32(emm2, _mm256_setzero_si256());
avx_m256_t swap_sign_bit = _mm256_castsi256_ps(emm0);
avx_m256_t poly_mask = _mm256_castsi256_ps(emm2);
sign_bit = _mm256_xor_ps(sign_bit, swap_sign_bit);
avx_m256_t temp = _ps_minus_cephes_DP123;
temp = _mm256_mul_ps(y, temp);
x = _mm256_add_ps(x, temp);
avx_m256_t x2 = _mm256_mul_ps(x, x);
avx_m256_t x3 = _mm256_mul_ps(x2, x);
avx_m256_t x4 = _mm256_mul_ps(x2, x2);
y = _ps_coscof_p0;
avx_m256_t y2 = _ps_sincof_p0;
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p1);
y2 = _mm256_add_ps(y2, _ps_sincof_p1);
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p2);
y2 = _mm256_add_ps(y2, _ps_sincof_p2);
y = _mm256_mul_ps(y, x4);
y2 = _mm256_mul_ps(y2, x3);
temp = _mm256_mul_ps(x2, _ps_0p5);
temp = _mm256_sub_ps(temp, _ps_1);
y = _mm256_sub_ps(y, temp);
y2 = _mm256_add_ps(y2, x);
y = _mm256_andnot_ps(poly_mask, y);
y2 = _mm256_and_ps(poly_mask, y2);
y = _mm256_add_ps(y, y2);
y = _mm256_xor_ps(y, sign_bit);
return y;
} // newsin_ps()
// Compare rank with all values currently in the queue. Returns -1 if the value already exists
// or is larger than all values.
// Otherwise, returns the index of the register in which the value should be inserted.
// Mask is replicated to both lanes, so it can be used for both value and rank lane.
int PriorityQueue_AVX2::compare(__m256i mrank, int &field, __m256i >mask)
{
static const __m256i eq4mask = _mm256_set_epi32(0, 0, 0, 0, -1, -1, -1, -1);
__m256i eq, eq4;
int reg, mask;
// Because items are sorted in ascending order within each (double) register, the mask after GT
// comparison must be of the form 000...1111, which is one less than a power of two.
{
__m256i r0_7 = _mm256_permute2x128_si256(_rv[1], _rv[0], 0x20); // [0 .. 7]
gtmask = _mm256_cmpgt_epi32(r0_7, mrank);
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask));
eq = _mm256_cmpeq_epi32(r0_7, mrank);
_ASSERTE(((mask + 1) & mask) == 0);
reg = 1;
}
if (!mask) {
__m256i r8_15 = _mm256_permute2x128_si256(_rv[3], _rv[2], 0x20); // [8 .. 15]
gtmask = _mm256_cmpgt_epi32(r8_15, mrank);
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask));
eq = _mm256_or_si256(eq, _mm256_cmpeq_epi32(r8_15, mrank));
_ASSERTE(((mask + 1) & mask) == 0);
reg = 3;
}
if (!mask) {
gtmask = _mm256_cmpgt_epi32(_rv[4], mrank); // [16 .. 19]; don't care about value
eq4 = _mm256_and_si256(eq4mask, _mm256_cmpeq_epi32(mrank, _rv[4])); // .. ditto
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask)) & 0xF; // ignore comparison with values
eq = _mm256_or_si256(eq, eq4);
_ASSERTE(((mask + 1) & mask) == 0);
reg = 4;
}
if (_mm256_movemask_ps(_mm256_castsi256_ps(eq)) != 0)
mask = 0;
if (!mask)
return -1;
// Adjust register according to mask (higher 128-bits i double register: one register lower)
// There is no "previous" register to test against for equality if we need to insert in the
// very first register. Also duplicate the same mask to both lanes.
if (mask > 0xF) {
mask >>= 4;
--reg;
gtmask = _mm256_permute2x128_si256(gtmask, gtmask, 0x11); // replicate high lane to both
}
__m256 mm256_exp_ps(__m256 x) {
__m256 tmp = _mm256_setzero_ps(), fx;
__m256i emm0;
__m256 one = *(__m256*)m256_ps_1;
x = _mm256_min_ps(x, *(__m256*)m256_ps_exp_hi);
x = _mm256_max_ps(x, *(__m256*)m256_ps_exp_lo);
/* express exp(x) as exp(g + n*log(2)) */
fx = _mm256_mul_ps(x, *(__m256*)m256_ps_cephes_LOG2EF);
fx = _mm256_add_ps(fx, *(__m256*)m256_ps_0p5);
/* how to perform a floorf with SSE: just below */
/* step 1 : cast to int */
emm0 = _mm256_cvttps_epi32(fx);
/* step 2 : cast back to float */
tmp = _mm256_cvtepi32_ps(emm0);
/* if greater, substract 1 */
__m256 mask = _mm256_cmp_ps( tmp, fx, _CMP_GT_OS );
mask = _mm256_and_ps(mask, one);
fx = _mm256_sub_ps(tmp, mask);
tmp = _mm256_mul_ps(fx, *(__m256*)m256_ps_cephes_exp_C1);
__m256 z = _mm256_mul_ps(fx, *(__m256*)m256_ps_cephes_exp_C2);
x = _mm256_sub_ps(x, tmp);
x = _mm256_sub_ps(x, z);
z = _mm256_mul_ps(x,x);
__m256 y = *(__m256*)m256_ps_cephes_exp_p0;
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(__m256*)m256_ps_cephes_exp_p1);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(__m256*)m256_ps_cephes_exp_p2);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(__m256*)m256_ps_cephes_exp_p3);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(__m256*)m256_ps_cephes_exp_p4);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(__m256*)m256_ps_cephes_exp_p5);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, x);
y = _mm256_add_ps(y, one);
/* build 2^n */
emm0 = _mm256_cvttps_epi32(fx);
emm0 = _mm256_add_epi32(emm0, *(__m256i*)m256_pi32_0x7f);
emm0 = _mm256_slli_epi32(emm0, 23);
__m256 pow2n = _mm256_castsi256_ps(emm0);
y = _mm256_mul_ps(y, pow2n);
_mm256_zeroupper();
return y;
}
v8sf exp256_ps(v8sf x) {
v8sf tmp = _mm256_setzero_ps(), fx;
v8si imm0;
v8sf one = *(v8sf*)_ps256_1;
x = _mm256_min_ps(x, *(v8sf*)_ps256_exp_hi);
x = _mm256_max_ps(x, *(v8sf*)_ps256_exp_lo);
/* express exp(x) as exp(g + n*log(2)) */
fx = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_LOG2EF);
fx = _mm256_add_ps(fx, *(v8sf*)_ps256_0p5);
/* how to perform a floorf with SSE: just below */
//imm0 = _mm256_cvttps_epi32(fx);
//tmp = _mm256_cvtepi32_ps(imm0);
tmp = _mm256_floor_ps(fx);
/* if greater, substract 1 */
//v8sf mask = _mm256_cmpgt_ps(tmp, fx);
v8sf mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS);
mask = _mm256_and_ps(mask, one);
fx = _mm256_sub_ps(tmp, mask);
tmp = _mm256_mul_ps(fx, *(v8sf*)_ps256_cephes_exp_C1);
v8sf z = _mm256_mul_ps(fx, *(v8sf*)_ps256_cephes_exp_C2);
x = _mm256_sub_ps(x, tmp);
x = _mm256_sub_ps(x, z);
z = _mm256_mul_ps(x,x);
v8sf y = *(v8sf*)_ps256_cephes_exp_p0;
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p1);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p2);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p3);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p4);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p5);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, x);
y = _mm256_add_ps(y, one);
/* build 2^n */
imm0 = _mm256_cvttps_epi32(fx);
// another two AVX2 instructions
imm0 = _mm256_add_epi32(imm0, *(v8si*)_pi32_256_0x7f);
imm0 = _mm256_slli_epi32(imm0, 23);
v8sf pow2n = _mm256_castsi256_ps(imm0);
y = _mm256_mul_ps(y, pow2n);
return y;
}
__m256 _inner_mm256_exp_ps1(__m256 arg)
{
arg = _mm256_mul_ps(arg, _mm256_set1_ps(1.4426950408889634073599246810018921374266459541529859f));
__m256i e = _mm256_add_epi32(
_mm256_castps_si256(_mm256_cmp_ps(arg, _mm256_set1_ps(0.0f), _CMP_LT_OQ)),
_mm256_cvttps_epi32(arg));
arg = _mm256_sub_ps(arg, _mm256_cvtepi32_ps(e));
__m256 intermediate_result;
intermediate_result = _mm256_fmadd_ps(_mm256_set1_ps(0.0136779459179717f), arg, _mm256_set1_ps(0.0517692205767896f));
intermediate_result = _mm256_fmadd_ps(intermediate_result, arg, _mm256_set1_ps(0.241554388295527f));
intermediate_result = _mm256_fmadd_ps(intermediate_result, arg, _mm256_set1_ps(0.692998430056128f));
intermediate_result = _mm256_fmadd_ps(intermediate_result, arg, _mm256_set1_ps(0.999999804292074f));
arg = intermediate_result;
__m256 res = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_add_epi32(e, _mm256_set1_epi32(127)), 23));
res = _mm256_mul_ps(res, arg);
return res;
}
inline void newsincos_ps_dual(avx_m256_t x1, avx_m256_t x2, avx_m256_t *s1, avx_m256_t *s2,
avx_m256_t *c1, avx_m256_t *c2) {
avx_m256_t tempa = _ps_sign_mask;
avx_m256_t tempb = _ps_inv_sign_mask;
avx_m256_t sign_bit1 = _mm256_and_ps(x1, tempa);
avx_m256_t sign_bit2 = _mm256_and_ps(x2, tempa);
x1 = _mm256_and_ps(x1, tempb);
x2 = _mm256_and_ps(x2, tempb);
tempa = _ps_cephes_FOPI;
avx_m256_t y1 = _mm256_mul_ps(x1, tempa);
avx_m256_t y2 = _mm256_mul_ps(x2, tempa);
//avx_m256i_t emm21 = _mm256_cvttps_epi32(y1);
//avx_m256i_t emm22 = _mm256_cvttps_epi32(y2);
//emm21 = _mm256_add_epi32(emm21, _pi32_1);
//emm22 = _mm256_add_epi32(emm22, _pi32_1);
avx_m256i_t emm21 = _mm256_cvttps_epi32(_mm256_add_ps(y1, _ps_1));
avx_m256i_t emm22 = _mm256_cvttps_epi32(_mm256_add_ps(y2, _ps_1));
//emm21 = _mm256_and_si256(emm21, _pi32_inv1);
//emm22 = _mm256_and_si256(emm22, _pi32_inv1);
emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21), _mm256_castsi256_ps(_pi32_inv1)));
emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22), _mm256_castsi256_ps(_pi32_inv1)));
y1 = _mm256_cvtepi32_ps(emm21);
y2 = _mm256_cvtepi32_ps(emm22);
//avx_m256i_t tempia = _pi32_2;
//avx_m256i_t cos_emm21 = _mm256_sub_epi32(emm21, tempia);
//avx_m256i_t cos_emm22 = _mm256_sub_epi32(emm22, tempia);
avx_m256i_t cos_emm21 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm21), _ps_2));
avx_m256i_t cos_emm22 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm22), _ps_2));
//avx_m256i_t tempib = _pi32_4;
//avx_m256i_t emm01 = _mm256_and_si256(emm21, tempib);
//avx_m256i_t emm02 = _mm256_and_si256(emm22, tempib);
avx_m256i_t emm01 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21),
_mm256_castsi256_ps(_pi32_4)));
avx_m256i_t emm02 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22),
_mm256_castsi256_ps(_pi32_4)));
//avx_m256i_t cos_emm01 = _mm256_andnot_si256(cos_emm21, tempib);
//avx_m256i_t cos_emm02 = _mm256_andnot_si256(cos_emm22, tempib);
avx_m256i_t cos_emm01 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm21),
_mm256_castsi256_ps(_pi32_4)));
avx_m256i_t cos_emm02 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm22),
_mm256_castsi256_ps(_pi32_4)));
//emm01 = _mm256_slli_epi32(emm01, 29);
__m128i emm0hi1 = _mm256_extractf128_si256(emm01, 0);
__m128i emm0lo1 = _mm256_extractf128_si256(emm01, 1);
emm0hi1 = _mm_slli_epi32(emm0hi1, 29);
emm0lo1 = _mm_slli_epi32(emm0lo1, 29);
emm01 = _mm256_insertf128_si256(emm01, emm0hi1, 0);
emm01 = _mm256_insertf128_si256(emm01, emm0lo1, 1);
//emm02 = _mm256_slli_epi32(emm02, 29);
__m128i emm0hi2 = _mm256_extractf128_si256(emm02, 0);
__m128i emm0lo2 = _mm256_extractf128_si256(emm02, 1);
emm0hi2 = _mm_slli_epi32(emm0hi2, 29);
emm0lo2 = _mm_slli_epi32(emm0lo2, 29);
emm02 = _mm256_insertf128_si256(emm02, emm0hi1, 0);
emm02 = _mm256_insertf128_si256(emm02, emm0lo1, 1);
//cos_emm01 = _mm256_slli_epi32(cos_emm01, 29);
__m128i cos_emm0hi1 = _mm256_extractf128_si256(cos_emm01, 0);
__m128i cos_emm0lo1 = _mm256_extractf128_si256(cos_emm01, 1);
cos_emm0hi1 = _mm_slli_epi32(cos_emm0hi1, 29);
cos_emm0lo1 = _mm_slli_epi32(cos_emm0lo1, 29);
cos_emm01 = _mm256_insertf128_si256(cos_emm01, cos_emm0hi1, 0);
cos_emm01 = _mm256_insertf128_si256(cos_emm01, cos_emm0lo1, 1);
//cos_emm02 = _mm256_slli_epi32(cos_emm02, 29);
__m128i cos_emm0hi2 = _mm256_extractf128_si256(cos_emm02, 0);
__m128i cos_emm0lo2 = _mm256_extractf128_si256(cos_emm02, 1);
cos_emm0hi2 = _mm_slli_epi32(cos_emm0hi2, 29);
cos_emm0lo2 = _mm_slli_epi32(cos_emm0lo2, 29);
cos_emm02 = _mm256_insertf128_si256(cos_emm02, cos_emm0hi2, 0);
cos_emm02 = _mm256_insertf128_si256(cos_emm02, cos_emm0lo2, 1);
//tempia = _pi32_2;
//tempib = _mm256_setzero_si256();
//emm21 = _mm256_and_si256(emm21, tempia);
//emm22 = _mm256_and_si256(emm22, tempia);
emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21),
_mm256_castsi256_ps(_pi32_2)));
emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22),
_mm256_castsi256_ps(_pi32_2)));
//cos_emm21 = _mm256_and_si256(cos_emm21, tempia);
//cos_emm22 = _mm256_and_si256(cos_emm22, tempia);
cos_emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm21),
_mm256_castsi256_ps(_pi32_2)));
cos_emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm22),
_mm256_castsi256_ps(_pi32_2)));
//emm21 = _mm256_cmpeq_epi32(emm21, tempib);
//emm22 = _mm256_cmpeq_epi32(emm22, tempib);
emm21 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm21), _mm256_setzero_ps(), _CMP_EQ_UQ));
emm22 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm22), _mm256_setzero_ps(), _CMP_EQ_UQ));
//cos_emm21 = _mm256_cmpeq_epi32(cos_emm21, tempib);
//cos_emm22 = _mm256_cmpeq_epi32(cos_emm22, tempib);
cos_emm21 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm21), _mm256_setzero_ps(), _CMP_EQ_UQ));
cos_emm22 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm22), _mm256_setzero_ps(), _CMP_EQ_UQ));
avx_m256_t emm0f1 = _mm256_castsi256_ps(emm01);
avx_m256_t emm0f2 = _mm256_castsi256_ps(emm02);
avx_m256_t emm2f1 = _mm256_castsi256_ps(emm21);
avx_m256_t emm2f2 = _mm256_castsi256_ps(emm22);
avx_m256_t cos_emm0f1 = _mm256_castsi256_ps(cos_emm01);
avx_m256_t cos_emm0f2 = _mm256_castsi256_ps(cos_emm02);
avx_m256_t cos_emm2f1 = _mm256_castsi256_ps(cos_emm21);
avx_m256_t cos_emm2f2 = _mm256_castsi256_ps(cos_emm22);
sign_bit1 = _mm256_xor_ps(sign_bit1, emm0f1);
sign_bit2 = _mm256_xor_ps(sign_bit2, emm0f2);
tempa = _ps_minus_cephes_DP123;
tempb = _mm256_mul_ps(y2, tempa);
tempa = _mm256_mul_ps(y1, tempa);
x2 = _mm256_add_ps(x2, tempb);
x1 = _mm256_add_ps(x1, tempa);
avx_m256_t x21 = _mm256_mul_ps(x1, x1);
avx_m256_t x22 = _mm256_mul_ps(x2, x2);
avx_m256_t x31 = _mm256_mul_ps(x21, x1);
avx_m256_t x32 = _mm256_mul_ps(x22, x2);
avx_m256_t x41 = _mm256_mul_ps(x21, x21);
avx_m256_t x42 = _mm256_mul_ps(x22, x22);
tempa = _ps_coscof_p0;
tempb = _ps_sincof_p0;
y1 = _mm256_mul_ps(x21, tempa);
y2 = _mm256_mul_ps(x22, tempa);
avx_m256_t y21 = _mm256_mul_ps(x21, tempb);
avx_m256_t y22 = _mm256_mul_ps(x22, tempb);
tempa = _ps_coscof_p1;
tempb = _ps_sincof_p1;
y1 = _mm256_add_ps(y1, tempa);
y2 = _mm256_add_ps(y2, tempa);
y21 = _mm256_add_ps(y21, tempb);
y22 = _mm256_add_ps(y22, tempb);
y1 = _mm256_mul_ps(y1, x21);
y2 = _mm256_mul_ps(y2, x22);
y21 = _mm256_mul_ps(y21, x21);
y22 = _mm256_mul_ps(y22, x22);
tempa = _ps_coscof_p2;
tempb = _ps_sincof_p2;
y1 = _mm256_add_ps(y1, tempa);
y2 = _mm256_add_ps(y2, tempa);
y21 = _mm256_add_ps(y21, tempb);
y22 = _mm256_add_ps(y22, tempb);
y1 = _mm256_mul_ps(y1, x41);
y2 = _mm256_mul_ps(y2, x42);
y21 = _mm256_mul_ps(y21, x31);
y22 = _mm256_mul_ps(y22, x32);
tempa = _ps_0p5;
tempb = _ps_1;
avx_m256_t temp_21 = _mm256_mul_ps(x21, tempa);
avx_m256_t temp_22 = _mm256_mul_ps(x22, tempa);
y21 = _mm256_add_ps(y21, x1);
y22 = _mm256_add_ps(y22, x2);
temp_21 = _mm256_sub_ps(temp_21, tempb);
temp_22 = _mm256_sub_ps(temp_22, tempb);
y1 = _mm256_sub_ps(y1, temp_21);
y2 = _mm256_sub_ps(y2, temp_22);
avx_m256_t cos_y1 = y1;
avx_m256_t cos_y2 = y2;
avx_m256_t cos_y21 = y21;
avx_m256_t cos_y22 = y22;
y1 = _mm256_andnot_ps(emm2f1, y1);
y2 = _mm256_andnot_ps(emm2f2, y2);
cos_y1 = _mm256_andnot_ps(cos_emm2f1, cos_y1);
cos_y2 = _mm256_andnot_ps(cos_emm2f2, cos_y2);
y21 = _mm256_and_ps(emm2f1, y21);
y22 = _mm256_and_ps(emm2f2, y22);
cos_y21 = _mm256_and_ps(cos_emm2f1, cos_y21);
cos_y22 = _mm256_and_ps(cos_emm2f2, cos_y22);
y1 = _mm256_add_ps(y1, y21);
y2 = _mm256_add_ps(y2, y22);
cos_y1 = _mm256_add_ps(cos_y1, cos_y21);
cos_y2 = _mm256_add_ps(cos_y2, cos_y22);
*s1 = _mm256_xor_ps(y1, sign_bit1);
*s2 = _mm256_xor_ps(y2, sign_bit2);
*c1 = _mm256_xor_ps(cos_y1, cos_emm0f1);
*c2 = _mm256_xor_ps(cos_y2, cos_emm0f2);
} // newsincos_ps_dual()
/* since sin256_ps and cos256_ps are almost identical, sincos256_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos256_ps(v8sf x, v8sf *s, v8sf *c) {
v8sf xmm1, xmm2, xmm3 = _mm256_setzero_ps(), sign_bit_sin, y;
v8si imm0, imm2, imm4;
#ifndef __AVX2__
v4si imm0_1, imm0_2;
v4si imm2_1, imm2_2;
v4si imm4_1, imm4_2;
#endif
sign_bit_sin = x;
/* take the absolute value */
x = _mm256_and_ps(x, *(v8sf*)_ps256_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm256_and_ps(sign_bit_sin, *(v8sf*)_ps256_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_FOPI);
#ifdef __AVX2__
/* store the integer part of y in imm2 */
imm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
imm2 = _mm256_add_epi32(imm2, *(v8si*)_pi32_256_1);
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_inv1);
y = _mm256_cvtepi32_ps(imm2);
imm4 = imm2;
/* get the swap sign flag for the sine */
imm0 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_4);
imm0 = _mm256_slli_epi32(imm0, 29);
//v8sf swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
/* get the polynom selection mask for the sine*/
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_2);
imm2 = _mm256_cmpeq_epi32(imm2, *(v8si*)_pi32_256_0);
//v8sf poly_mask = _mm256_castsi256_ps(imm2);
#else
/* we use SSE2 routines to perform the integer ops */
COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y),imm2_1,imm2_2);
imm2_1 = _mm_add_epi32(imm2_1, *(v4si*)_pi32avx_1);
imm2_2 = _mm_add_epi32(imm2_2, *(v4si*)_pi32avx_1);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_inv1);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_inv1);
COPY_XMM_TO_IMM(imm2_1,imm2_2,imm2);
y = _mm256_cvtepi32_ps(imm2);
imm4_1 = imm2_1;
imm4_2 = imm2_2;
imm0_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_4);
imm0_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_4);
imm0_1 = _mm_slli_epi32(imm0_1, 29);
imm0_2 = _mm_slli_epi32(imm0_2, 29);
COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_2);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_2);
imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
#endif
v8sf swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
v8sf poly_mask = _mm256_castsi256_ps(imm2);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(v8sf*)_ps256_minus_cephes_DP1;
xmm2 = *(v8sf*)_ps256_minus_cephes_DP2;
xmm3 = *(v8sf*)_ps256_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
#ifdef __AVX2__
imm4 = _mm256_sub_epi32(imm4, *(v8si*)_pi32_256_2);
imm4 = _mm256_andnot_si128(imm4, *(v8si*)_pi32_256_4);
imm4 = _mm256_slli_epi32(imm4, 29);
#else
imm4_1 = _mm_sub_epi32(imm4_1, *(v4si*)_pi32avx_2);
imm4_2 = _mm_sub_epi32(imm4_2, *(v4si*)_pi32avx_2);
imm4_1 = _mm_andnot_si128(imm4_1, *(v4si*)_pi32avx_4);
imm4_2 = _mm_andnot_si128(imm4_2, *(v4si*)_pi32avx_4);
imm4_1 = _mm_slli_epi32(imm4_1, 29);
imm4_2 = _mm_slli_epi32(imm4_2, 29);
COPY_XMM_TO_IMM(imm4_1, imm4_2, imm4);
#endif
v8sf sign_bit_cos = _mm256_castsi256_ps(imm4);
sign_bit_sin = _mm256_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
v8sf z = _mm256_mul_ps(x,x);
y = *(v8sf*)_ps256_coscof_p0;
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
v8sf tmp = _mm256_mul_ps(z, *(v8sf*)_ps256_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(v8sf*)_ps256_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v8sf y2 = *(v8sf*)_ps256_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
v8sf ysin2 = _mm256_and_ps(xmm3, y2);
v8sf ysin1 = _mm256_andnot_ps(xmm3, y);
y2 = _mm256_sub_ps(y2,ysin2);
y = _mm256_sub_ps(y, ysin1);
xmm1 = _mm256_add_ps(ysin1,ysin2);
xmm2 = _mm256_add_ps(y,y2);
/* update the sign */
*s = _mm256_xor_ps(xmm1, sign_bit_sin);
*c = _mm256_xor_ps(xmm2, sign_bit_cos);
}
inline void newsincos_ps(avx_m256_t x, avx_m256_t *s, avx_m256_t *c) {
avx_m256_t sign_bit = _mm256_and_ps(x, _ps_sign_mask);
x = _mm256_and_ps(x, _ps_inv_sign_mask);
avx_m256_t y = _mm256_mul_ps(x, _ps_cephes_FOPI);
//avx_m256i_t emm2 = _mm256_cvttps_epi32(y);
//emm2 = _mm256_add_epi32(emm2, _pi32_1);
avx_m256i_t emm2 = _mm256_cvttps_epi32(_mm256_add_ps(y, _ps_1));
//emm2 = _mm256_and_si256(emm2, _pi32_inv1);
emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2), _mm256_castsi256_ps(_pi32_inv1)));
y = _mm256_cvtepi32_ps(emm2);
//avx_m256i_t cos_emm2 = _mm256_sub_epi32(emm2, _pi32_2);
avx_m256i_t cos_emm2 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm2), _ps_2));
//avx_m256i_t emm0 = _mm256_and_si256(emm2, _pi32_4);
avx_m256i_t emm0 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2),
_mm256_castsi256_ps(_pi32_4)));
//avx_m256i_t cos_emm0 = _mm256_andnot_si256(cos_emm2, _pi32_4);
avx_m256i_t cos_emm0 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm2),
_mm256_castsi256_ps(_pi32_4)));
//emm0 = _mm256_slli_epi32(emm0, 29);
__m128i emm0hi = _mm256_extractf128_si256(emm0, 0);
__m128i emm0lo = _mm256_extractf128_si256(emm0, 1);
emm0hi = _mm_slli_epi32(emm0hi, 29);
emm0lo = _mm_slli_epi32(emm0lo, 29);
emm0 = _mm256_insertf128_si256(emm0, emm0hi, 0);
emm0 = _mm256_insertf128_si256(emm0, emm0lo, 1);
//cos_emm0 = _mm256_slli_epi32(cos_emm0, 29);
__m128i cos_emm0hi = _mm256_extractf128_si256(cos_emm0, 0);
__m128i cos_emm0lo = _mm256_extractf128_si256(cos_emm0, 1);
cos_emm0hi = _mm_slli_epi32(cos_emm0hi, 29);
cos_emm0lo = _mm_slli_epi32(cos_emm0lo, 29);
cos_emm0 = _mm256_insertf128_si256(cos_emm0, cos_emm0hi, 0);
cos_emm0 = _mm256_insertf128_si256(cos_emm0, cos_emm0lo, 1);
//emm2 = _mm256_and_si256(emm2, _pi32_2);
emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2),
_mm256_castsi256_ps(_pi32_2)));
//cos_emm2 = _mm256_and_si256(cos_emm2, _pi32_2);
cos_emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm2),
_mm256_castsi256_ps(_pi32_2)));
//emm2 = _mm256_cmpeq_epi32(emm2, _mm256_setzero_si256());
emm2 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm2), _mm256_setzero_ps(), _CMP_EQ_UQ));
//cos_emm2 = _mm256_cmpeq_epi32(cos_emm2, _mm256_setzero_si256());
cos_emm2 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm2), _mm256_setzero_ps(), _CMP_EQ_UQ));
avx_m256_t emm0f = _mm256_castsi256_ps(emm0);
avx_m256_t emm2f = _mm256_castsi256_ps(emm2);
avx_m256_t cos_emm0f = _mm256_castsi256_ps(cos_emm0);
avx_m256_t cos_emm2f = _mm256_castsi256_ps(cos_emm2);
sign_bit = _mm256_xor_ps(sign_bit, emm0f);
avx_m256_t temp_2 = _ps_minus_cephes_DP123;
temp_2 = _mm256_mul_ps(y, temp_2);
x = _mm256_add_ps(x, temp_2);
avx_m256_t x2 = _mm256_mul_ps(x, x);
avx_m256_t x3 = _mm256_mul_ps(x2, x);
avx_m256_t x4 = _mm256_mul_ps(x2, x2);
y = _ps_coscof_p0;
avx_m256_t y2 = _ps_sincof_p0;
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p1);
y2 = _mm256_add_ps(y2, _ps_sincof_p1);
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p2);
y2 = _mm256_add_ps(y2, _ps_sincof_p2);
y = _mm256_mul_ps(y, x4);
y2 = _mm256_mul_ps(y2, x3);
temp_2 = _mm256_mul_ps(x2, _ps_0p5);
y2 = _mm256_add_ps(y2, x);
temp_2 = _mm256_sub_ps(temp_2, _ps_1);
y = _mm256_sub_ps(y, temp_2);
avx_m256_t cos_y = y;
avx_m256_t cos_y2 = y2;
y = _mm256_andnot_ps(emm2f, y);
cos_y = _mm256_andnot_ps(cos_emm2f, cos_y);
y2 = _mm256_and_ps(emm2f, y2);
cos_y2 = _mm256_and_ps(cos_emm2f, cos_y2);
y = _mm256_add_ps(y, y2);
cos_y = _mm256_add_ps(cos_y, cos_y2);
*s = _mm256_xor_ps(y, sign_bit);
*c = _mm256_xor_ps(cos_y, cos_emm0f);
} // newsincos_ps()
INLINE const avxi operator ^( const avxi& a, const avxi& b ) { return _mm256_castps_si256(_mm256_xor_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b))); }
inline void newexp_ps_dual(avx_m256_t x1, avx_m256_t x2, avx_m256_t* exp1, avx_m256_t* exp2) {
avx_m256_t one = _ps_1;
avx_m256_t zero = _ps_0;
x1 = _mm256_min_ps(x1, _ps_exp_hi);
x2 = _mm256_min_ps(x2, _ps_exp_hi);
x1 = _mm256_max_ps(x1, _ps_exp_lo);
x2 = _mm256_max_ps(x2, _ps_exp_lo);
avx_m256_t temp_21 = _mm256_mul_ps(x1, _ps_cephes_LOG2EF);
avx_m256_t temp_22 = _mm256_mul_ps(x2, _ps_cephes_LOG2EF);
temp_21 = _mm256_add_ps(temp_21, _ps_0p5);
temp_22 = _mm256_add_ps(temp_22, _ps_0p5);
avx_m256i_t emm01 = _mm256_cvttps_epi32(temp_21);
avx_m256i_t emm02 = _mm256_cvttps_epi32(temp_22);
avx_m256_t temp_11 = _mm256_cvtepi32_ps(emm01);
avx_m256_t temp_12 = _mm256_cvtepi32_ps(emm02);
avx_m256_t temp_31 = _mm256_sub_ps(temp_11, temp_21);
avx_m256_t temp_32 = _mm256_sub_ps(temp_12, temp_22);
avx_m256_t mask1 = _mm256_cmp_ps(temp_31, zero, _CMP_GT_OQ);
avx_m256_t mask2 = _mm256_cmp_ps(temp_32, zero, _CMP_GT_OQ);
mask1 = _mm256_and_ps(mask1, one);
mask2 = _mm256_and_ps(mask2, one);
temp_21 = _mm256_sub_ps(temp_11, mask1);
temp_22 = _mm256_sub_ps(temp_12, mask2);
emm01 = _mm256_cvttps_epi32(temp_21);
emm02 = _mm256_cvttps_epi32(temp_22);
temp_11 = _mm256_mul_ps(temp_21, _ps_cephes_exp_C12);
temp_12 = _mm256_mul_ps(temp_22, _ps_cephes_exp_C12);
x1 = _mm256_sub_ps(x1, temp_11);
x2 = _mm256_sub_ps(x2, temp_12);
avx_m256_t x21 = _mm256_mul_ps(x1, x1);
avx_m256_t x22 = _mm256_mul_ps(x2, x2);
avx_m256_t x31 = _mm256_mul_ps(x21, x1);
avx_m256_t x32 = _mm256_mul_ps(x22, x2);
avx_m256_t x41 = _mm256_mul_ps(x21, x21);
avx_m256_t x42 = _mm256_mul_ps(x22, x22);
temp_11 = _mm256_add_ps(x1, one);
temp_12 = _mm256_add_ps(x2, one);
temp_21 = _mm256_mul_ps(x21, _ps_cephes_exp_p5);
temp_22 = _mm256_mul_ps(x22, _ps_cephes_exp_p5);
temp_31 = _mm256_mul_ps(x31, _ps_cephes_exp_p4);
temp_32 = _mm256_mul_ps(x32, _ps_cephes_exp_p4);
temp_11 = _mm256_add_ps(temp_11, temp_21);
temp_12 = _mm256_add_ps(temp_12, temp_22);
temp_21 = _mm256_mul_ps(x31, _ps_cephes_exp_p0);
temp_22 = _mm256_mul_ps(x32, _ps_cephes_exp_p0);
temp_11 = _mm256_add_ps(temp_11, temp_31);
temp_12 = _mm256_add_ps(temp_12, temp_32);
avx_m256_t temp_41 = _mm256_mul_ps(x1, _ps_cephes_exp_p2);
avx_m256_t temp_42 = _mm256_mul_ps(x2, _ps_cephes_exp_p2);
temp_31 = _mm256_mul_ps(x21, _ps_cephes_exp_p1);
temp_32 = _mm256_mul_ps(x22, _ps_cephes_exp_p1);
//emm01 = _mm256_add_epi32(emm01, _pi32_0x7f);
//emm02 = _mm256_add_epi32(emm02, _pi32_0x7f);
emm01 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm01), _mm256_castsi256_ps(_pi32_0x7f)));
emm02 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm02), _mm256_castsi256_ps(_pi32_0x7f)));
temp_21 = _mm256_add_ps(temp_21, temp_31);
temp_22 = _mm256_add_ps(temp_22, temp_32);
temp_31 = _mm256_add_ps(temp_31, temp_41);
temp_32 = _mm256_add_ps(temp_32, temp_42);
//emm01 = _mm256_slli_epi32(emm01, 23);
__m128i emm0hi1 = _mm256_extractf128_si256(emm01, 0);
__m128i emm0lo1 = _mm256_extractf128_si256(emm01, 1);
emm0hi1 = _mm_slli_epi32(emm0hi1, 23);
emm0lo1 = _mm_slli_epi32(emm0lo1, 23);
emm01 = _mm256_insertf128_si256(emm01, emm0hi1, 0);
emm01 = _mm256_insertf128_si256(emm01, emm0lo1, 1);
//emm02 = _mm256_slli_epi32(emm02, 23);
__m128i emm0hi2 = _mm256_extractf128_si256(emm02, 0);
__m128i emm0lo2 = _mm256_extractf128_si256(emm02, 1);
emm0hi2 = _mm_slli_epi32(emm0hi2, 23);
emm0lo2 = _mm_slli_epi32(emm0lo2, 23);
emm02 = _mm256_insertf128_si256(emm02, emm0hi2, 0);
emm02 = _mm256_insertf128_si256(emm02, emm0lo2, 1);
avx_m256_t pow2n1 = _mm256_castsi256_ps(emm01);
avx_m256_t pow2n2 = _mm256_castsi256_ps(emm02);
temp_21 = _mm256_add_ps(temp_21, temp_31);
temp_22 = _mm256_add_ps(temp_22, temp_32);
temp_21 = _mm256_mul_ps(temp_21, x41);
temp_22 = _mm256_mul_ps(temp_22, x42);
avx_m256_t y1 = _mm256_add_ps(temp_11, temp_21);
avx_m256_t y2 = _mm256_add_ps(temp_12, temp_22);
*exp1 = _mm256_mul_ps(y1, pow2n1);
*exp2 = _mm256_mul_ps(y2, pow2n2);
} // newexp_ps_dual()
void TransLut_FindIndexAvx2 <TransLut::MapperLog>::find_index (const TransLut::FloatIntMix val_arr [8], __m256i &index, __m256 &frac)
{
assert (val_arr != 0);
// Constants
static const int mant_size = 23;
static const int exp_bias = 127;
static const uint32_t base = (exp_bias + LOGLUT_MIN_L2) << mant_size;
static const float val_min = 1.0f / (int64_t (1) << -LOGLUT_MIN_L2);
// static const float val_max = float (int64_t (1) << LOGLUT_MAX_L2);
static const int frac_size = mant_size - LOGLUT_RES_L2;
static const uint32_t frac_mask = (1 << frac_size) - 1;
const __m256 zero_f = _mm256_setzero_ps ();
const __m256 one_f = _mm256_set1_ps (1);
const __m256 frac_mul = _mm256_set1_ps (1.0f / (1 << frac_size));
const __m256 mul_eps = _mm256_set1_ps (1.0f / val_min);
const __m256 mask_abs_f = _mm256_load_ps (
reinterpret_cast <const float *> (fstb::ToolsAvx2::_mask_abs)
);
const __m256i zero_i = _mm256_setzero_si256 ();
const __m256i mask_abs_epi32 = _mm256_set1_epi32 (0x7FFFFFFF);
const __m256i one_epi32 = _mm256_set1_epi32 (1);
const __m256i base_epi32 = _mm256_set1_epi32 (int (base));
const __m256i frac_mask_epi32 = _mm256_set1_epi32 (frac_mask);
const __m256i val_min_epi32 =
_mm256_set1_epi32 ((LOGLUT_MIN_L2 + exp_bias) << mant_size);
const __m256i val_max_epi32 =
_mm256_set1_epi32 ((LOGLUT_MAX_L2 + exp_bias) << mant_size);
const __m256i index_max_epi32 =
_mm256_set1_epi32 ((LOGLUT_MAX_L2 - LOGLUT_MIN_L2) << LOGLUT_RES_L2);
const __m256i hsize_epi32 = _mm256_set1_epi32 (LOGLUT_HSIZE);
const __m256i mirror_epi32 = _mm256_set1_epi32 (LOGLUT_HSIZE - 1);
// It really starts here
const __m256 val_f = _mm256_load_ps (reinterpret_cast <const float *> (val_arr));
const __m256 val_a = _mm256_and_ps (val_f, mask_abs_f);
const __m256i val_i = _mm256_load_si256 (reinterpret_cast <const __m256i *> (val_arr));
const __m256i val_u = _mm256_and_si256 (val_i, mask_abs_epi32);
// Standard path
__m256i index_std = _mm256_sub_epi32 (val_u, base_epi32);
index_std = _mm256_srli_epi32 (index_std, frac_size);
index_std = _mm256_add_epi32 (index_std, one_epi32);
__m256i frac_stdi = _mm256_and_si256 (val_u, frac_mask_epi32);
__m256 frac_std = _mm256_cvtepi32_ps (frac_stdi);
frac_std = _mm256_mul_ps (frac_std, frac_mul);
// Epsilon path
__m256 frac_eps = _mm256_max_ps (val_a, zero_f);
frac_eps = _mm256_mul_ps (frac_eps, mul_eps);
// Range cases
const __m256i eps_flag_i = _mm256_cmpgt_epi32 (val_min_epi32, val_u);
const __m256i std_flag_i = _mm256_cmpgt_epi32 (val_max_epi32, val_u);
const __m256 eps_flag_f = _mm256_castsi256_ps (eps_flag_i);
const __m256 std_flag_f = _mm256_castsi256_ps (std_flag_i);
__m256i index_tmp =
fstb::ToolsAvx2::select (std_flag_i, index_std, index_max_epi32);
__m256 frac_tmp =
fstb::ToolsAvx2::select (std_flag_f, frac_std, one_f);
index_tmp = fstb::ToolsAvx2::select (eps_flag_i, zero_i, index_tmp);
frac_tmp = fstb::ToolsAvx2::select (eps_flag_f, frac_eps, frac_tmp);
// Sign cases
const __m256i neg_flag_i = _mm256_srai_epi32 (val_i, 31);
const __m256 neg_flag_f = _mm256_castsi256_ps (neg_flag_i);
const __m256i index_neg = _mm256_sub_epi32 (mirror_epi32, index_tmp);
const __m256i index_pos = _mm256_add_epi32 (hsize_epi32, index_tmp);
const __m256 frac_neg = _mm256_sub_ps (one_f, frac_tmp);
index = fstb::ToolsAvx2::select (neg_flag_i, index_neg, index_pos);
frac = fstb::ToolsAvx2::select (neg_flag_f, frac_neg, frac_tmp);
}
void Decoder::ADMMDecoder_deg_6_7_2_3_6()
{
int maxIter = maxIteration;
float mu = 5.5f;
float tableau[12] = { 0.0f };
if ((mBlocklength == 576) && (mNChecks == 288))
{
mu = 3.37309f;//penalty
tableau[2] = 0.00001f;
tableau[3] = 2.00928f;
tableau[6] = 4.69438f;
}
else if((mBlocklength == 2304) && (mNChecks == 1152) )
{
mu = 3.81398683f;//penalty
tableau[2] = 0.29669288f;
tableau[3] = 0.46964023f;
tableau[6] = 3.19548154f;
}
else
{
mu = 5.5;//penalty
tableau[2] = 0.8f;
tableau[3] = 0.8f;
tableau[6] = 0.8f;
}
const float rho = 1.9f; //over relaxation parameter;
const float un_m_rho = 1.0 - rho;
const auto _rho = _mm256_set1_ps( rho );
const auto _un_m_rho = _mm256_set1_ps( un_m_rho );
float tableaX[12];
//
// ON PRECALCULE LES CONSTANTES
//
#pragma unroll
for (int i = 0; i < 7; i++)
{
tableaX[i] = tableau[ i ] / mu;
}
const auto t_mu = _mm256_set1_ps ( mu );
const auto t2_amu = _mm256_set1_ps ( tableau[ 2 ] / mu );
const auto t3_amu = _mm256_set1_ps ( tableau[ 3 ] / mu );
const auto t6_amu = _mm256_set1_ps ( tableau[ 6 ] / mu );
const auto t2_2amu = _mm256_set1_ps ( 2.0f * tableau[ 2 ] / mu );
const auto t3_2amu = _mm256_set1_ps ( 2.0f * tableau[ 3 ] / mu );
const auto t6_2amu = _mm256_set1_ps ( 2.0f * tableau[ 6 ] / mu );
const auto t2_deg = _mm256_set1_ps ( 2.0f );
const auto t3_deg = _mm256_set1_ps ( 3.0f );
const auto t6_deg = _mm256_set1_ps ( 6.0f );
const auto zero = _mm256_set1_ps ( 0.0f );
const auto un = _mm256_set1_ps ( 1.0f );
const __m256 a = _mm256_set1_ps ( 0.0f );
const __m256 b = _mm256_set1_ps ( 0.5f );
//////////////////////////////////////////////////////////////////////////////////////
#pragma unroll
for( int j = 0; j < _mPCheckMapSize; j+=8 )
{
_mm256_store_ps(&Lambda [j], a);
_mm256_store_ps(&zReplica[j], b);
_mm256_store_ps(&latestProjVector[j], b);
}
//////////////////////////////////////////////////////////////////////////////////////
for(int i = 0; i < maxIter; i++)
{
int ptr = 0;
mIteration = i + 1;
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto start = timer();
#endif
//
// VN processing kernel
//
#pragma unroll
for (int j = 0; j < _mBlocklength; j++)
{
const int degVn = VariableDegree[j];
float M[8] __attribute__((aligned(64)));
if( degVn == 2 ){
#if 1
const int dVN = 2;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t2_amu), _mm256_sub_ps(t2_deg, t2_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 2;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 3 ){
#if 1
const int dVN = 3;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t3_amu), _mm256_sub_ps(t3_deg, t3_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 3;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 6 ){
#if 1
const int dVN = 6;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t6_amu), _mm256_sub_ps(t6_deg, t6_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 6;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}
}
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
t_vn += (timer() - start);
#endif
//
// CN processing kernel
//
int CumSumCheckDegree = 0; // cumulative position of currect edge in factor graph
int allVerified = 0;
float vector_before_proj[8] __attribute__((aligned(64)));
const auto zero = _mm256_set1_ps ( 0.0f );
const auto mask_6 = _mm256_set_epi32(0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto mask_7 = _mm256_set_epi32(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto dot5 = _mm256_set1_ps( 0.5f );
//
// MEASURE OF THE CN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto starT = timer();
#endif
const auto seuilProj = _mm256_set1_ps( 1e-5f );
for(int j = 0; j < _mNChecks; j++)
{
if( CheckDegree[j] == 6 ){
const int cDeg6 = 0x3F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_6), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
int test = (_mm256_movemask_ps( synd ) & cDeg6); // deg 6
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps (xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg6) == 0x00; // degree 6
if( skip == false )
{
const auto _ztemp = mp.projection_deg6( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda[CumSumCheckDegree], mask_6, mLambda);
_mm256_maskstore_ps(&zReplica[CumSumCheckDegree], mask_6, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_6, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 6;
}else if( CheckDegree[j] == 7 )
{
const int cDeg7 = 0x7F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_7), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
const int test = (_mm256_movemask_ps( synd ) & cDeg7); // deg 7
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps ( xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg7) == 0x00; // degree 7
if( skip == false )
{
const auto _ztemp = mp.projection_deg7( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda [CumSumCheckDegree], mask_7, mLambda);
_mm256_maskstore_ps(&zReplica [CumSumCheckDegree], mask_7, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_7, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 7;
}else{
exit( 0 );
}
}
//
// MEASURE OF THE CN LOOP EXECUTION TIME
//
#ifdef PROFILE_ON
t_cn += (timer() - starT);
#endif
#ifdef PROFILE_ON
t_ex += 1;
//FILE *ft=fopen("time.txt","a");
//fprintf(ft,"%d \n", t_cn/t_ex);
//fprintf(ft,"%d %d %d \n", t_cn, t_vn, t_pj);
//fclose(ft);
#endif
if(allVerified == 0)
{
mAlgorithmConverge = true;
mValidCodeword = true;
break;
}
}
//
// MEASURE OF THE NUMBER OF EXECUTION
//
// #ifdef PROFILE_ON
// t_ex += 1;
// #endif
}
const GSVector4 GSVector4::m_ps0123(0.0f, 1.0f, 2.0f, 3.0f);
const GSVector4 GSVector4::m_ps4567(4.0f, 5.0f, 6.0f, 7.0f);
const GSVector4 GSVector4::m_half(0.5f);
const GSVector4 GSVector4::m_one(1.0f);
const GSVector4 GSVector4::m_two(2.0f);
const GSVector4 GSVector4::m_four(4.0f);
const GSVector4 GSVector4::m_x4b000000(_mm_castsi128_ps(_mm_set1_epi32(0x4b000000)));
const GSVector4 GSVector4::m_x4f800000(_mm_castsi128_ps(_mm_set1_epi32(0x4f800000)));
const GSVector4 GSVector4::m_max(FLT_MAX);
const GSVector4 GSVector4::m_min(FLT_MIN);
#if _M_SSE >= 0x500
const GSVector8 GSVector8::m_half(0.5f);
const GSVector8 GSVector8::m_one(1.0f);
const GSVector8 GSVector8::m_x7fffffff(_mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff)));
const GSVector8 GSVector8::m_x80000000(_mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)));
const GSVector8 GSVector8::m_x4b000000(_mm256_castsi256_ps(_mm256_set1_epi32(0x4b000000)));
const GSVector8 GSVector8::m_x4f800000(_mm256_castsi256_ps(_mm256_set1_epi32(0x4f800000)));
const GSVector8 GSVector8::m_max(FLT_MAX);
const GSVector8 GSVector8::m_min(FLT_MIN);
#endif
#if _M_SSE >= 0x501
const GSVector8i GSVector8i::m_xff[33] =
{
GSVector8i(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000),
GSVector8i(0x000000ff, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000),
GSVector8i(0x0000ffff, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000),
/* evaluation of 8 sines at onces using AVX intrisics
The code is the exact rewriting of the cephes sinf function.
Precision is excellent as long as x < 8192 (I did not bother to
take into account the special handling they have for greater values
-- it does not return garbage for arguments over 8192, though, but
the extra precision is missing).
Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
surprising but correct result.
*/
v8sf sin256_ps(v8sf x) { // any x
v8sf xmm1, xmm2 = _mm256_setzero_ps(), xmm3, sign_bit, y;
v8si imm0, imm2;
#ifndef __AVX2__
v4si imm0_1, imm0_2;
v4si imm2_1, imm2_2;
#endif
sign_bit = x;
/* take the absolute value */
x = _mm256_and_ps(x, *(v8sf*)_ps256_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit = _mm256_and_ps(sign_bit, *(v8sf*)_ps256_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_FOPI);
/*
Here we start a series of integer operations, which are in the
realm of AVX2.
If we don't have AVX, let's perform them using SSE2 directives
*/
#ifdef __AVX2__
/* store the integer part of y in mm0 */
imm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
// another two AVX2 instruction
imm2 = _mm256_add_epi32(imm2, *(v8si*)_pi32_256_1);
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_inv1);
y = _mm256_cvtepi32_ps(imm2);
/* get the swap sign flag */
imm0 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_4);
imm0 = _mm256_slli_epi32(imm0, 29);
/* get the polynom selection mask
there is one polynom for 0 <= x <= Pi/4
and another one for Pi/4<x<=Pi/2
Both branches will be computed.
*/
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_2);
imm2 = _mm256_cmpeq_epi32(imm2,*(v8si*)_pi32_256_0);
#else
/* we use SSE2 routines to perform the integer ops */
COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y),imm2_1,imm2_2);
imm2_1 = _mm_add_epi32(imm2_1, *(v4si*)_pi32avx_1);
imm2_2 = _mm_add_epi32(imm2_2, *(v4si*)_pi32avx_1);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_inv1);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_inv1);
COPY_XMM_TO_IMM(imm2_1,imm2_2,imm2);
y = _mm256_cvtepi32_ps(imm2);
imm0_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_4);
imm0_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_4);
imm0_1 = _mm_slli_epi32(imm0_1, 29);
imm0_2 = _mm_slli_epi32(imm0_2, 29);
COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_2);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_2);
imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
#endif
v8sf swap_sign_bit = _mm256_castsi256_ps(imm0);
v8sf poly_mask = _mm256_castsi256_ps(imm2);
sign_bit = _mm256_xor_ps(sign_bit, swap_sign_bit);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(v8sf*)_ps256_minus_cephes_DP1;
xmm2 = *(v8sf*)_ps256_minus_cephes_DP2;
xmm3 = *(v8sf*)_ps256_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(v8sf*)_ps256_coscof_p0;
v8sf z = _mm256_mul_ps(x,x);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
v8sf tmp = _mm256_mul_ps(z, *(v8sf*)_ps256_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(v8sf*)_ps256_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v8sf y2 = *(v8sf*)_ps256_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
y = _mm256_andnot_ps(xmm3, y);
y = _mm256_add_ps(y,y2);
/* update the sign */
y = _mm256_xor_ps(y, sign_bit);
return y;
}
INLINE avxi unpackhi( const avxi& a, const avxi& b ) { return _mm256_castps_si256(_mm256_unpackhi_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b))); }
template<index_t index> INLINE const avxi shuffle( const avxi& a ) {
return _mm256_castps_si256(_mm256_permute_ps(_mm256_castsi256_ps(a), _MM_SHUFFLE(index, index, index, index)));
}
INLINE const avxi select( const avxb& mask, const avxi& t, const avxi& f ) { return _mm256_castps_si256(_mm256_blendv_ps(_mm256_castsi256_ps(f), _mm256_castsi256_ps(t), mask)); }
INLINE __m256 _mm256_maskload_ps (float const *ptr, __m256i mask) {
return _mm256_maskload_ps(ptr, _mm256_castsi256_ps(mask));
}
__m256 mm256_cos_ps(__m256 x) {
__m256 xmm1, xmm2 = _mm256_setzero_ps(), xmm3, y;
__m256i emm0, emm2;
/* take the absolute value */
x = _mm256_and_ps(x, *(__m256*)m256_ps_inv_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(__m256*)m256_ps_cephes_FOPI);
/* store the integer part of y in mm0 */
emm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm256_add_epi32(emm2, *(__m256i*)m256_pi32_1);
emm2 = _mm256_and_si256(emm2, *(__m256i*)m256_pi32_inv1);
y = _mm256_cvtepi32_ps(emm2);
emm2 = _mm256_sub_epi32(emm2, *(__m256i*)m256_pi32_2);
/* get the swap sign flag */
emm0 = _mm256_andnot_si256(emm2, *(__m256i*)m256_pi32_4);
emm0 = _mm256_slli_epi32(emm0, 29);
/* get the polynom selection mask */
emm2 = _mm256_and_si256(emm2, *(__m256i*)m256_pi32_2);
emm2 = _mm256_cmpeq_epi32(emm2, _mm256_setzero_si256());
__m256 sign_bit = _mm256_castsi256_ps(emm0);
__m256 poly_mask = _mm256_castsi256_ps(emm2);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m256*)m256_ps_minus_cephes_DP1;
xmm2 = *(__m256*)m256_ps_minus_cephes_DP2;
xmm3 = *(__m256*)m256_ps_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(__m256*)m256_ps_coscof_p0;
__m256 z = _mm256_mul_ps(x,x);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(__m256*)m256_ps_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(__m256*)m256_ps_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
__m256 tmp = _mm256_mul_ps(z, *(__m256*)m256_ps_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(__m256*)m256_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m256 y2 = *(__m256*)m256_ps_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(__m256*)m256_ps_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(__m256*)m256_ps_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
y = _mm256_andnot_ps(xmm3, y);
y = _mm256_add_ps(y,y2);
/* update the sign */
y = _mm256_xor_ps(y, sign_bit);
_mm256_zeroupper();
return y;
}
inline avx_m256_t newexp_ps(avx_m256_t x) {
avx_m256_t one = _ps_1;
avx_m256_t zero = _ps_0;
x = _mm256_min_ps(x, _ps_exp_hi);
x = _mm256_max_ps(x, _ps_exp_lo);
avx_m256_t temp_2 = _mm256_mul_ps(x, _ps_cephes_LOG2EF);
temp_2 = _mm256_add_ps(temp_2, _ps_0p5);
avx_m256i_t emm0 = _mm256_cvttps_epi32(temp_2);
avx_m256_t temp_1 = _mm256_cvtepi32_ps(emm0);
avx_m256_t temp_3 = _mm256_sub_ps(temp_1, temp_2);
avx_m256_t mask = _mm256_cmp_ps(temp_3, zero, _CMP_GT_OQ);
mask = _mm256_and_ps(mask, one);
temp_2 = _mm256_sub_ps(temp_1, mask);
emm0 = _mm256_cvttps_epi32(temp_2);
temp_1 = _mm256_mul_ps(temp_2, _ps_cephes_exp_C12);
x = _mm256_sub_ps(x, temp_1);
avx_m256_t x2 = _mm256_mul_ps(x, x);
avx_m256_t x3 = _mm256_mul_ps(x2, x);
avx_m256_t x4 = _mm256_mul_ps(x2, x2);
temp_1 = _mm256_add_ps(x, one);
temp_2 = _mm256_mul_ps(x2, _ps_cephes_exp_p5);
temp_3 = _mm256_mul_ps(x3, _ps_cephes_exp_p4);
temp_1 = _mm256_add_ps(temp_1, temp_2);
temp_2 = _mm256_mul_ps(x3, _ps_cephes_exp_p0);
temp_1 = _mm256_add_ps(temp_1, temp_3);
avx_m256_t temp_4 = _mm256_mul_ps(x, _ps_cephes_exp_p2);
temp_3 = _mm256_mul_ps(x2, _ps_cephes_exp_p1);
emm0 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm0), _mm256_castsi256_ps(_pi32_0x7f)));
temp_2 = _mm256_add_ps(temp_2, temp_3);
temp_3 = _mm256_add_ps(temp_3, temp_4);
//emm0 = _mm256_slli_epi32(emm0, 23);
// convert emm0 into two 128-bit integer vectors
// perform shift on both vectors
// combine both vectors into 256-bit emm0
__m128i emm0hi = _mm256_extractf128_si256(emm0, 0);
__m128i emm0lo = _mm256_extractf128_si256(emm0, 1);
emm0hi = _mm_slli_epi32(emm0hi, 23);
emm0lo = _mm_slli_epi32(emm0lo, 23);
emm0 = _mm256_insertf128_si256(emm0, emm0hi, 0);
emm0 = _mm256_insertf128_si256(emm0, emm0lo, 1);
avx_m256_t pow2n = _mm256_castsi256_ps(emm0);
temp_2 = _mm256_add_ps(temp_2, temp_3);
temp_2 = _mm256_mul_ps(temp_2, x4);
avx_m256_t y = _mm256_add_ps(temp_1, temp_2);
y = _mm256_mul_ps(y, pow2n);
return y;
} // newexp_ps()
template<> INLINE const avxi shuffle<1, 1, 3, 3>( const avxi& b ) { return _mm256_castps_si256(_mm256_movehdup_ps(_mm256_castsi256_ps(b))); }
double bst_compute_129_m256_maskstore_root_aligned( void*_bst_obj, double* p, double* q, size_t nn ) {
segments_t* mem = (segments_t*) _bst_obj;
int n, i, r, l_end, j, l_end_pre;
double t, e_tmp;
double* e = mem->e, *w = mem->w;
int* root = mem->r;
__m256d v_tmp;
__m256d v00, v01, v02, v03;
__m256d v10, v11, v12, v13;
__m256d v20, v21, v22, v23;
__m256d v30, v31, v32, v33;
__m256i v_cur_roots;
__m256 v_rootmask0, v_rootmask1;
// initialization
// mem->n = nn;
n = nn; // subtractions with n potentially negative. say hello to all the bugs
int idx1, idx1_root;
int idx2;
int idx3, idx3_root;
int pad_root, pad, pad_r;
idx1 = ((int) mem->e_sz) - 1;
idx1_root = ((int) mem->r_sz);
// the conventio is that iteration i, idx1 points to the first element of line i+1
e[idx1++] = q[n];
// pad contains the padding for row i+1
// for row n it's always 3
pad = 3;
pad_root = 7;
for (i = n-1; i >= 0; --i) {
idx1 -= 2*(n-i)+1 + pad;
idx1_root -= 2*(n-i)+1 + pad_root;
idx2 = idx1 + 1;
e[idx1] = q[i];
w[idx1] = q[i];
for (j = i+1; j < n+1; ++j,++idx2) {
e[idx2] = INFINITY;
w[idx2] = w[idx2-1] + p[j-1] + q[j];
}
idx2 += pad; // padding of line i+1
// idx2 now points to the first element of the next line
idx3 = idx1;
idx3_root = idx1_root;
pad_r = pad;
for (r = i; r < n; ++r) {
pad_r = (pad_r+1)&3; // padding of line r+1
idx1 = idx3;
idx1_root = idx3_root;
l_end = idx2 + (n-r);
// l_end points to the first entry after the current row
e_tmp = e[idx1++];
idx1_root++;
// calculate until a multiple of 8 doubles is left
// 8 = 4 * 2 128-bit vectors
l_end_pre = idx2 + ((n-r)&15);
for( ; (idx2 < l_end_pre) && (idx2 < l_end); ++idx2 ) {
t = e_tmp + e[idx2] + w[idx1];
if (t < e[idx1]) {
e[idx1] = t;
root[idx1_root] = r;
}
idx1++;
idx1_root++;
}
v_tmp = _mm256_set_pd( e_tmp, e_tmp, e_tmp, e_tmp );
// execute the shit for 4 vectors of size 2
v_cur_roots = _mm256_set_epi32(r, r, r, r, r, r, r, r);
for( ; idx2 < l_end; idx2 += 16 ) {
v01 = _mm256_load_pd( &w[idx1 ] );
v11 = _mm256_load_pd( &w[idx1+ 4] );
v21 = _mm256_load_pd( &w[idx1+ 8] );
v31 = _mm256_load_pd( &w[idx1+12] );
v00 = _mm256_load_pd( &e[idx2 ] );
v01 = _mm256_add_pd( v01, v_tmp );
v10 = _mm256_load_pd( &e[idx2+ 4] );
v11 = _mm256_add_pd( v11, v_tmp );
v20 = _mm256_load_pd( &e[idx2+ 8] );
v21 = _mm256_add_pd( v21, v_tmp );
v30 = _mm256_load_pd( &e[idx2+12] );
v31 = _mm256_add_pd( v31, v_tmp );
v01 = _mm256_add_pd( v01, v00 );
v03 = _mm256_load_pd( &e[idx1 ] );
v11 = _mm256_add_pd( v11, v10 );
v13 = _mm256_load_pd( &e[idx1+ 4] );
v21 = _mm256_add_pd( v21, v20 );
v23 = _mm256_load_pd( &e[idx1+ 8] );
v31 = _mm256_add_pd( v31, v30 );
v33 = _mm256_load_pd( &e[idx1+12] );
v02 = _mm256_cmp_pd( v01, v03, _CMP_LT_OQ );
v12 = _mm256_cmp_pd( v11, v13, _CMP_LT_OQ );
v22 = _mm256_cmp_pd( v21, v23, _CMP_LT_OQ );
v32 = _mm256_cmp_pd( v31, v33, _CMP_LT_OQ );
_mm256_maskstore_pd( &e[idx1 ],
_mm256_castpd_si256( v02 ), v01 );
_mm256_maskstore_pd( &e[idx1+ 4],
_mm256_castpd_si256( v12 ), v11 );
v_rootmask0 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v02)),
_mm256_cvtpd_ps(v12) , 1
);
_mm256_maskstore_pd( &e[idx1+ 8],
_mm256_castpd_si256( v22 ), v21 );
_mm256_maskstore_pd( &e[idx1+12],
_mm256_castpd_si256( v32 ), v31 );
v_rootmask1 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v22)),
_mm256_cvtpd_ps(v32) , 1
);
_mm256_maskstore_ps( &root[idx1_root ],
_mm256_castps_si256( v_rootmask0 ),
_mm256_castsi256_ps( v_cur_roots ) );
_mm256_maskstore_ps( &root[idx1_root + 8],
_mm256_castps_si256( v_rootmask1 ),
_mm256_castsi256_ps( v_cur_roots ) );
idx1 += 16;
idx1_root += 16;
}
idx2 += pad_r;
idx3++;
idx3_root++;
}
pad = (pad -1)&3;
pad_root = (pad_root-1)&7;
}
// the index of the last item of the first row is ((n/4)+1)*4-1, due to the padding
// if n is even, the total number of entries in the first
// row of the table is odd, so we need padding
return e[ ((n/4)+1)*4 - 1 ];
}
template<> INLINE const avxi shuffle<0, 1, 0, 1>( const avxi& b ) { return _mm256_castps_si256(_mm256_castpd_ps(_mm256_movedup_pd(_mm256_castps_pd(_mm256_castsi256_ps(b))))); }
template<index_t index_0, index_t index_1, index_t index_2, index_t index_3> INLINE const avxi shuffle( const avxi& a, const avxi& b ) {
return _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), _MM_SHUFFLE(index_3, index_2, index_1, index_0)));
}
INLINE void _mm256_maskstore_ps (float *ptr, __m256i mask, __m256 data) {
_mm256_maskstore_ps(ptr, _mm256_castsi256_ps(mask), data);
}
real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
real scratch[4*DIM];
__m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
real * vdwioffsetptr0;
__m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
__m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
__m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
int nvdwtype;
__m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
int *vdwtype;
real *vdwparam;
__m256 one_sixth = _mm256_set1_ps(1.0/6.0);
__m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
__m256 dummy_mask,cutoff_mask;
__m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
__m256 one = _mm256_set1_ps(1.0);
__m256 two = _mm256_set1_ps(2.0);
x = xx[0];
f = ff[0];
nri = nlist->nri;
iinr = nlist->iinr;
jindex = nlist->jindex;
jjnr = nlist->jjnr;
shiftidx = nlist->shift;
gid = nlist->gid;
shiftvec = fr->shift_vec[0];
fshift = fr->fshift[0];
nvdwtype = fr->ntype;
vdwparam = fr->nbfp;
template<> INLINE const avxi shuffle<0, 0, 2, 2>( const avxi& b ) { return _mm256_castps_si256(_mm256_moveldup_ps(_mm256_castsi256_ps(b))); }
