// 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
}
INLINE uint movemask(const avxb& a) {return _mm256_movemask_ps(a);}
INLINE bool all(const avxb& a) {return _mm256_movemask_ps(a) == (uint)0xff;}
INLINE bool reduce_and(const avxb& a) {return _mm256_movemask_ps(a) == (uint)0xff;}
// reduction operations
INLINE size_t popcnt(const avxb& a) {return __popcnt(_mm256_movemask_ps(a));}
CPLErr
GDALGridInverseDistanceToAPower2NoSmoothingNoSearchAVX(
const void *poOptions,
GUInt32 nPoints,
CPL_UNUSED const double *unused_padfX,
CPL_UNUSED const double *unused_padfY,
CPL_UNUSED const double *unused_padfZ,
double dfXPoint, double dfYPoint,
double *pdfValue,
void* hExtraParamsIn )
{
size_t i = 0;
GDALGridExtraParameters* psExtraParams = (GDALGridExtraParameters*) hExtraParamsIn;
const float* pafX = psExtraParams->pafX;
const float* pafY = psExtraParams->pafY;
const float* pafZ = psExtraParams->pafZ;
const float fEpsilon = 0.0000000000001f;
const float fXPoint = (float)dfXPoint;
const float fYPoint = (float)dfYPoint;
const __m256 ymm_small = GDAL_mm256_load1_ps(fEpsilon);
const __m256 ymm_x = GDAL_mm256_load1_ps(fXPoint);
const __m256 ymm_y = GDAL_mm256_load1_ps(fYPoint);
__m256 ymm_nominator = _mm256_setzero_ps();
__m256 ymm_denominator = _mm256_setzero_ps();
int mask = 0;
#undef LOOP_SIZE
#if defined(__x86_64) || defined(_M_X64)
/* This would also work in 32bit mode, but there are only 8 XMM registers */
/* whereas we have 16 for 64bit */
#define LOOP_SIZE 16
size_t nPointsRound = (nPoints / LOOP_SIZE) * LOOP_SIZE;
for ( i = 0; i < nPointsRound; i += LOOP_SIZE )
{
__m256 ymm_rx = _mm256_sub_ps(_mm256_load_ps(pafX + i), ymm_x); /* rx = pafX[i] - fXPoint */
__m256 ymm_rx_8 = _mm256_sub_ps(_mm256_load_ps(pafX + i + 8), ymm_x);
__m256 ymm_ry = _mm256_sub_ps(_mm256_load_ps(pafY + i), ymm_y); /* ry = pafY[i] - fYPoint */
__m256 ymm_ry_8 = _mm256_sub_ps(_mm256_load_ps(pafY + i + 8), ymm_y);
__m256 ymm_r2 = _mm256_add_ps(_mm256_mul_ps(ymm_rx, ymm_rx), /* r2 = rx * rx + ry * ry */
_mm256_mul_ps(ymm_ry, ymm_ry));
__m256 ymm_r2_8 = _mm256_add_ps(_mm256_mul_ps(ymm_rx_8, ymm_rx_8),
_mm256_mul_ps(ymm_ry_8, ymm_ry_8));
__m256 ymm_invr2 = _mm256_rcp_ps(ymm_r2); /* invr2 = 1.0f / r2 */
__m256 ymm_invr2_8 = _mm256_rcp_ps(ymm_r2_8);
ymm_nominator = _mm256_add_ps(ymm_nominator, /* nominator += invr2 * pafZ[i] */
_mm256_mul_ps(ymm_invr2, _mm256_load_ps(pafZ + i)));
ymm_nominator = _mm256_add_ps(ymm_nominator,
_mm256_mul_ps(ymm_invr2_8, _mm256_load_ps(pafZ + i + 8)));
ymm_denominator = _mm256_add_ps(ymm_denominator, ymm_invr2); /* denominator += invr2 */
ymm_denominator = _mm256_add_ps(ymm_denominator, ymm_invr2_8);
mask = _mm256_movemask_ps(_mm256_cmp_ps(ymm_r2, ymm_small, _CMP_LT_OS)) | /* if( r2 < fEpsilon) */
(_mm256_movemask_ps(_mm256_cmp_ps(ymm_r2_8, ymm_small, _CMP_LT_OS)) << 8);
if( mask )
break;
}
#else
#define LOOP_SIZE 8
size_t nPointsRound = (nPoints / LOOP_SIZE) * LOOP_SIZE;
for ( i = 0; i < nPointsRound; i += LOOP_SIZE )
{
__m256 ymm_rx = _mm256_sub_ps(_mm256_load_ps((float*)pafX + i), ymm_x); /* rx = pafX[i] - fXPoint */
__m256 ymm_ry = _mm256_sub_ps(_mm256_load_ps((float*)pafY + i), ymm_y); /* ry = pafY[i] - fYPoint */
__m256 ymm_r2 = _mm256_add_ps(_mm256_mul_ps(ymm_rx, ymm_rx), /* r2 = rx * rx + ry * ry */
_mm256_mul_ps(ymm_ry, ymm_ry));
__m256 ymm_invr2 = _mm256_rcp_ps(ymm_r2); /* invr2 = 1.0f / r2 */
ymm_nominator = _mm256_add_ps(ymm_nominator, /* nominator += invr2 * pafZ[i] */
_mm256_mul_ps(ymm_invr2, _mm256_load_ps((float*)pafZ + i)));
ymm_denominator = _mm256_add_ps(ymm_denominator, ymm_invr2); /* denominator += invr2 */
mask = _mm256_movemask_ps(_mm256_cmp_ps(ymm_r2, ymm_small, _CMP_LT_OS)); /* if( r2 < fEpsilon) */
if( mask )
break;
}
#endif
/* Find which i triggered r2 < fEpsilon */
if( mask )
{
for(int j = 0; j < LOOP_SIZE; j++ )
{
if( mask & (1 << j) )
{
(*pdfValue) = (pafZ)[i + j];
// GCC and MSVC need explicit zeroing
#if !defined(__clang__)
_mm256_zeroupper();
#endif
return CE_None;
}
}
}
#undef LOOP_SIZE
/* Get back nominator and denominator values for YMM registers */
float afNominator[8], afDenominator[8];
_mm256_storeu_ps(afNominator, ymm_nominator);
_mm256_storeu_ps(afDenominator, ymm_denominator);
// MSVC doesn't emit AVX afterwards but may use SSE, so clear upper bits
// Other compilers will continue using AVX for the below floating points operations
#if defined(_MSC_FULL_VER)
_mm256_zeroupper();
#endif
float fNominator = afNominator[0] + afNominator[1] +
afNominator[2] + afNominator[3] +
afNominator[4] + afNominator[5] +
afNominator[6] + afNominator[7];
float fDenominator = afDenominator[0] + afDenominator[1] +
afDenominator[2] + afDenominator[3] +
afDenominator[4] + afDenominator[5] +
afDenominator[6] + afDenominator[7];
/* Do the few remaining loop iterations */
for ( ; i < nPoints; i++ )
{
const float fRX = pafX[i] - fXPoint;
const float fRY = pafY[i] - fYPoint;
const float fR2 =
fRX * fRX + fRY * fRY;
// If the test point is close to the grid node, use the point
// value directly as a node value to avoid singularity.
if ( fR2 < 0.0000000000001 )
{
break;
}
else
{
const float fInvR2 = 1.0f / fR2;
fNominator += fInvR2 * pafZ[i];
fDenominator += fInvR2;
}
}
if( i != nPoints )
{
(*pdfValue) = pafZ[i];
}
else
if ( fDenominator == 0.0 )
{
(*pdfValue) =
((GDALGridInverseDistanceToAPowerOptions*)poOptions)->dfNoDataValue;
}
else
(*pdfValue) = fNominator / fDenominator;
// GCC needs explicit zeroing
#if defined(__GNUC__) && !defined(__clang__)
_mm256_zeroupper();
#endif
return CE_None;
}
template<class Extension,class Info>
struct call<nbtrue_,tag::simd_(tag::arithmetic_,Extension),Info>
{
typedef int32_t result_type;
NT2_FUNCTOR_CALL_DISPATCH(
1,
typename nt2::meta::scalar_of<A0>::type,
(3, (float,double,arithmetic_))
)
NT2_FUNCTOR_CALL_EVAL_IF(1, float)
{
typedef typename meta::as_real<A0>::type type;
int32_t r = _mm256_movemask_ps(isnez(a0));
return (r&1)+((r>>1)&1)+((r>>2)&1)+(r>>3&1)+((r>>4)&1)+((r>>5)&1)+(r>>6&1)+(r>>7);
// return __builtin_popcount(_mm_movemask_ps(isnez(cast<type>(a0))));
}
NT2_FUNCTOR_CALL_EVAL_IF(1, double)
{
int32_t r = _mm256_movemask_pd(isnez(a0));
return (r&1)+(r>>1&1)+((r>>2)&1)+(r>>3);
}
NT2_FUNCTOR_CALL_EVAL_IF(1, arithmetic_)
{
typedef typename meta::scalar_of<A0>::type sctype;
typedef typename simd::native<sctype, tag::sse_ > svtype;
svtype a00 = { _mm256_extractf128_si256(a0, 0)};
svtype a01 = { _mm256_extractf128_si256(a0, 1)};
return nbtrue(a00)+nbtrue(a01);
void
plot(u32 w, u32 h, float x1, float y1, float x2, float y2, float dx,
float dy, u32 max_iter = 4096)
{
assert(w % 8 == 0);
// AVX Constants
float const constants[] {
x1,
y1,
dx,
dy,
1.0f,
4.0f
};
__m256 const vx1 = _mm256_broadcast_ss(constants);
__m256 const vy1 = _mm256_broadcast_ss(constants + 1);
__m256 const vdx = _mm256_broadcast_ss(constants + 2);
__m256 const vdy = _mm256_broadcast_ss(constants + 3);
__m256 const v1 = _mm256_broadcast_ss(constants + 4);
__m256 const v4 = _mm256_broadcast_ss(constants + 5);
// Start timing
std::chrono::time_point<std::chrono::high_resolution_clock> t1, t2;
std::chrono::duration<double> dt;
t1 = std::chrono::high_resolution_clock::now();
// Zero line counter
__m256 vj = _mm256_xor_ps(v1, v1);
for (u32 j = 0; j < h; j++) {
for (u32 i = 0; i < w; i += 8) {
// Fill column counter
float const vi_[8] { i+0.f, i+1.f, i+2.f, i+3.f, i+4.f, i+5.f, i+6.f, i+7.f };
__m256 vi = _mm256_load_ps(vi_);
// Compute start point
__m256 vx0 = _mm256_mul_ps(vi, vdx);
vx0 = _mm256_add_ps(vx0, vx1);
__m256 vy0 = _mm256_mul_ps(vj, vdy);
vy0 = _mm256_add_ps(vy0, vy1);
__m256 vx = vx0;
__m256 vy = vy0;
__m256 vcount = _mm256_xor_ps(v1, v1); // Zero iteration counter
u32 iter = 0;
u8 no_overflow = 0;
do {
// Compute products
__m256 vxx = _mm256_mul_ps(vx, vx);
__m256 vyy = _mm256_mul_ps(vy, vy);
// Check termination condition
__m256 vtmp = _mm256_add_ps(vxx, vyy);
vtmp = _mm256_cmp_ps(vtmp, v4, _CMP_LT_OQ);
no_overflow = _mm256_movemask_ps(vtmp) & 0xff;
// Accumulate iteration counter
vtmp = _mm256_and_ps(vtmp, v1);
vcount = _mm256_add_ps(vcount, vtmp);
// Step
vtmp = _mm256_mul_ps(vx, vy);
vtmp = _mm256_add_ps(vtmp, vtmp);
vy = _mm256_add_ps(vtmp, vy0);
vtmp = _mm256_sub_ps(vxx, vyy);
vx = _mm256_add_ps(vtmp, vx0);
++iter;
} while (no_overflow && (iter < max_iter));
for (u32 k = 0; k < 8; k++) {
u32 n = ((float *) &vcount)[k] + 0.5f;
if (n == max_iter) n = 0;
char c = ' ';
if (n > 0) {
static char const charset[] = ".,c8M@jawrpogOQEPGJ";
c = charset[n % (sizeof(charset) - 1)];
}
attron(COLOR_PAIR((n % 7) + 1));
addch(c);
attroff(COLOR_PAIR((n % 7) + 1));
if (i + k + 1 == w) addch('\n');
}
}
// Increment line counter
vj = _mm256_add_ps(vj, v1);
}
// End timing
t2 = std::chrono::high_resolution_clock::now();
dt = t2 - t1;
std::string info = std::to_string(dt.count() * 1000.0) + "ms";
attron(COLOR_PAIR(1));
printw(info.c_str());
attroff(COLOR_PAIR(1));
}
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
}
void molec_quadrant_neighbor_interaction_fma(molec_Quadrant_t q, molec_Quadrant_t q_n, float* Epot_)
{
#ifdef __AVX2__
const __m256 sigLJ = _mm256_set1_ps(molec_parameter->sigLJ);
const __m256 epsLJ = _mm256_set1_ps(molec_parameter->epsLJ);
const __m256 Rcut2 = _mm256_set1_ps(molec_parameter->Rcut2);
const int N = q.N;
const int N_n = q_n.N_pad;
__m256 Epot8 = _mm256_setzero_ps();
__m256 _1 = _mm256_set1_ps(1.f);
__m256 _2 = _mm256_set1_ps(2.f);
__m256 _24epsLJ = _mm256_mul_ps(_mm256_set1_ps(24.f), epsLJ);
for(int i = 0; i < N; ++i)
{
const __m256 xi = _mm256_set1_ps(q.x[i]);
const __m256 yi = _mm256_set1_ps(q.y[i]);
const __m256 zi = _mm256_set1_ps(q.z[i]);
__m256 f_xi = _mm256_setzero_ps();
__m256 f_yi = _mm256_setzero_ps();
__m256 f_zi = _mm256_setzero_ps();
for(int j = 0; j < N_n; j += 8)
{
// count number of interactions
if(MOLEC_CELLLIST_COUNT_INTERACTION)
++num_potential_interactions;
// load coordinates and fores into AVX vectors
const __m256 xj = _mm256_load_ps(&q_n.x[j]);
const __m256 yj = _mm256_load_ps(&q_n.y[j]);
const __m256 zj = _mm256_load_ps(&q_n.z[j]);
__m256 f_xj = _mm256_load_ps(&q_n.f_x[j]);
__m256 f_yj = _mm256_load_ps(&q_n.f_y[j]);
__m256 f_zj = _mm256_load_ps(&q_n.f_z[j]);
// distance computation
const __m256 xij = _mm256_sub_ps(xi, xj);
const __m256 yij = _mm256_sub_ps(yi, yj);
const __m256 zij = _mm256_sub_ps(zi, zj);
const __m256 zij2 = _mm256_mul_ps(zij, zij);
const __m256 r2 = _mm256_fmadd_ps(xij, xij, _mm256_fmadd_ps(yij, yij, zij2));
// r2 < Rcut2
const __m256 mask = _mm256_cmp_ps(r2, Rcut2, _CMP_LT_OQ);
// if( any(r2 < R2) )
if(_mm256_movemask_ps(mask))
{
const __m256 r2inv = _mm256_div_ps(_1, r2);
const __m256 s2 = _mm256_mul_ps(_mm256_mul_ps(sigLJ, sigLJ), r2inv);
const __m256 s6 = _mm256_mul_ps(_mm256_mul_ps(s2, s2), s2);
const __m256 s12 = _mm256_mul_ps(s6, s6);
const __m256 s12_minus_s6 = _mm256_sub_ps(s12, s6);
const __m256 two_s12_minus_s6 = _mm256_sub_ps(_mm256_mul_ps(_2, s12), s6);
Epot8 = _mm256_add_ps(Epot8, _mm256_and_ps(s12_minus_s6, mask));
const __m256 fr = _mm256_mul_ps(_mm256_mul_ps(_24epsLJ, r2inv), two_s12_minus_s6);
const __m256 fr_mask = _mm256_and_ps(fr, mask);
// update forces
f_xi = _mm256_fmadd_ps(fr_mask, xij,f_xi);
f_yi = _mm256_fmadd_ps(fr_mask, yij,f_yi);
f_zi = _mm256_fmadd_ps(fr_mask, zij,f_zi);
f_xj = _mm256_fnmadd_ps(fr_mask,xij,f_xj);
f_yj = _mm256_fnmadd_ps(fr_mask,yij,f_yj);
f_zj = _mm256_fnmadd_ps(fr_mask,zij,f_zj);
// store back j-forces
_mm256_store_ps(&q_n.f_x[j], f_xj);
_mm256_store_ps(&q_n.f_y[j], f_yj);
_mm256_store_ps(&q_n.f_z[j], f_zj);
}
}
// update i-forces
float MOLEC_ALIGNAS(32) f_array[8];
_mm256_store_ps(f_array, f_xi);
q.f_x[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
_mm256_store_ps(f_array, f_yi);
q.f_y[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
_mm256_store_ps(f_array, f_zi);
q.f_z[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
}
float MOLEC_ALIGNAS(32) E_pot_array[8];
_mm256_store_ps(E_pot_array, Epot8);
// perform reduction of potential energy
*Epot_ += 4
* molec_parameter->epsLJ*(E_pot_array[0] + E_pot_array[1] + E_pot_array[2]
+ E_pot_array[3] + E_pot_array[4] + E_pot_array[5]
+ E_pot_array[6] + E_pot_array[7]);
#endif
}