void
foo ()
{
_mm256_zeroupper ();
x = y;
_mm256_zeroupper ();
_mm256_zeroupper ();
_mm256_zeroupper ();
}
void
foo ()
{
x = y;
_mm256_zeroall ();
_mm256_zeroupper ();
_mm256_zeroupper ();
_mm256_zeroupper ();
}
int main(int argc, char **argv) {
_mm256_zeroupper();
float a = 3;
float b = atoi(argv[1]);
if(argc>2) _mm256_zeroupper();
for(int k = 0; k < 5000000; k++) {
b = a + b * b;
a = b - a * a;
}
return (b == 0)? 1 : 0;
}
__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;
}
int SymmColumnVec_32f_Unsymm_AVX(const float** src, const float* ky, float* dst, float delta, int width, int ksize2)
{
int i = 0, k;
const float *S, *S2;
const __m128 d4 = _mm_set1_ps(delta);
const __m256 d8 = _mm256_set1_ps(delta);
for (; i <= width - 16; i += 16)
{
__m256 f, s0 = d8, s1 = d8;
__m256 x0;
S = src[0] + i;
for (k = 1; k <= ksize2; k++)
{
S = src[k] + i;
S2 = src[-k] + i;
f = _mm256_set1_ps(ky[k]);
x0 = _mm256_sub_ps(_mm256_loadu_ps(S), _mm256_loadu_ps(S2));
#if CV_FMA3
s0 = _mm256_fmadd_ps(x0, f, s0);
#else
s0 = _mm256_add_ps(s0, _mm256_mul_ps(x0, f));
#endif
x0 = _mm256_sub_ps(_mm256_loadu_ps(S + 8), _mm256_loadu_ps(S2 + 8));
#if CV_FMA3
s1 = _mm256_fmadd_ps(x0, f, s1);
#else
s1 = _mm256_add_ps(s1, _mm256_mul_ps(x0, f));
#endif
}
_mm256_storeu_ps(dst + i, s0);
_mm256_storeu_ps(dst + i + 8, s1);
}
for (; i <= width - 4; i += 4)
{
__m128 f, x0, s0 = d4;
for (k = 1; k <= ksize2; k++)
{
f = _mm_set1_ps(ky[k]);
x0 = _mm_sub_ps(_mm_load_ps(src[k] + i), _mm_load_ps(src[-k] + i));
s0 = _mm_add_ps(s0, _mm_mul_ps(x0, f));
}
_mm_storeu_ps(dst + i, s0);
}
_mm256_zeroupper();
return i;
}
void _Run(OutputPixelType aaOutput[ciHeight][ciWidth], InputPixelType_1 aaInput1[ciHeight][ciWidth], InputPixelType_2 aaInput2[ciHeight][ciWidth])
{
for (int iY = 0; iY < ciHeight; ++iY)
{
_mm256_zeroall();
OutputPixelType *pOutput = aaOutput[iY];
InputPixelType_1 *pInput1 = aaInput1[iY];
InputPixelType_2 *pInput2 = aaInput2[iY];
for (int iX = 0; iX < ciWidth; iX += VectorWidth)
{
__m256 mmIn1 = _mm256_loadu_ps( pInput1 + iX );
__m256 mmIn2 = _mm256_loadu_ps( pInput2 + iX );
_mm256_storeu_ps( pOutput + iX, _mm256_add_ps(mmIn1, mmIn2) );
}
_mm256_zeroupper();
}
}
int RowVec_32f_AVX(const float* src0, const float* _kx, float* dst, int width, int cn, int _ksize)
{
int i = 0, k;
for (; i <= width - 8; i += 8)
{
const float* src = src0 + i;
__m256 f, x0;
__m256 s0 = _mm256_set1_ps(0.0f);
for (k = 0; k < _ksize; k++, src += cn)
{
f = _mm256_set1_ps(_kx[k]);
x0 = _mm256_loadu_ps(src);
#if CV_FMA3
s0 = _mm256_fmadd_ps(x0, f, s0);
#else
s0 = _mm256_add_ps(s0, _mm256_mul_ps(x0, f));
#endif
}
_mm256_storeu_ps(dst + i, s0);
}
_mm256_zeroupper();
return i;
}
int warpAffineBlockline(int *adelta, int *bdelta, short* xy, short* alpha, int X0, int Y0, int bw)
{
const int AB_BITS = MAX(10, (int)INTER_BITS);
int x1 = 0;
__m256i fxy_mask = _mm256_set1_epi32(INTER_TAB_SIZE - 1);
__m256i XX = _mm256_set1_epi32(X0), YY = _mm256_set1_epi32(Y0);
for (; x1 <= bw - 16; x1 += 16)
{
__m256i tx0, tx1, ty0, ty1;
tx0 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(adelta + x1)), XX);
ty0 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(bdelta + x1)), YY);
tx1 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(adelta + x1 + 8)), XX);
ty1 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(bdelta + x1 + 8)), YY);
tx0 = _mm256_srai_epi32(tx0, AB_BITS - INTER_BITS);
ty0 = _mm256_srai_epi32(ty0, AB_BITS - INTER_BITS);
tx1 = _mm256_srai_epi32(tx1, AB_BITS - INTER_BITS);
ty1 = _mm256_srai_epi32(ty1, AB_BITS - INTER_BITS);
__m256i fx_ = _mm256_packs_epi32(_mm256_and_si256(tx0, fxy_mask),
_mm256_and_si256(tx1, fxy_mask));
__m256i fy_ = _mm256_packs_epi32(_mm256_and_si256(ty0, fxy_mask),
_mm256_and_si256(ty1, fxy_mask));
tx0 = _mm256_packs_epi32(_mm256_srai_epi32(tx0, INTER_BITS),
_mm256_srai_epi32(tx1, INTER_BITS));
ty0 = _mm256_packs_epi32(_mm256_srai_epi32(ty0, INTER_BITS),
_mm256_srai_epi32(ty1, INTER_BITS));
fx_ = _mm256_adds_epi16(fx_, _mm256_slli_epi16(fy_, INTER_BITS));
fx_ = _mm256_permute4x64_epi64(fx_, (3 << 6) + (1 << 4) + (2 << 2) + 0);
_mm256_storeu_si256((__m256i*)(xy + x1 * 2), _mm256_unpacklo_epi16(tx0, ty0));
_mm256_storeu_si256((__m256i*)(xy + x1 * 2 + 16), _mm256_unpackhi_epi16(tx0, ty0));
_mm256_storeu_si256((__m256i*)(alpha + x1), fx_);
}
_mm256_zeroupper();
return x1;
}
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps)
{
const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
uint32_t partitions = 1u << max_partition_order;
FLAC__ASSERT(default_partition_samples > predictor_order);
/* first do max_partition_order */
{
const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples);
uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order);
if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) {
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-7; residual_sample+=8) {
__m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample)));
sum256 = _mm256_add_epi32(sum256, res256);
}
sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
sum128 = _mm_add_epi32(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi32(sum128, res128);
}
sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_SHUFFLE(1,0,3,2)));
sum128 = _mm_add_epi32(sum128, _mm_shufflelo_epi16(sum128, _MM_SHUFFLE(1,0,3,2)));
abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128);
/* workaround for MSVC bugs (at least versions 2015 and 2017 are affected) */
#if (defined _MSC_VER) && (defined FLAC__CPU_X86_64)
abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/
#endif
}
}
else { /* have to pessimistically use 64 bits for accumulator */
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
__m256i res256 = _mm256_cvtepu32_epi64(res128);
sum256 = _mm256_add_epi64(sum256, res256);
}
sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-1; residual_sample+=2) {
__m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample)));
res128 = _mm_cvtepu32_epi64(res128);
sum128 = _mm_add_epi64(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi64(sum128, res128);
}
sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8));
_mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128);
}
}
}
/* now merge partitions for lower orders */
{
uint32_t from_partition = 0, to_partition = partitions;
int partition_order;
for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
uint32_t i;
partitions >>= 1;
for(i = 0; i < partitions; i++) {
abs_residual_partition_sums[to_partition++] =
abs_residual_partition_sums[from_partition ] +
abs_residual_partition_sums[from_partition+1];
from_partition += 2;
}
}
}
_mm256_zeroupper();
}
void AVX2FMA3DNoise(Vector3d& result, const Vector3d& EPoint)
{
#if CHECK_FUNCTIONAL
Vector3d param(EPoint);
#endif
AVX2TABLETYPE *mp;
// TODO FIXME - global statistics reference
// Stats[Calls_To_DNoise]++;
const __m256d ONE_PD = _mm256_set1_pd(1.0);
const __m128i short_si128 = _mm_set1_epi32(0xffff);
const __m256d xyzn = _mm256_setr_pd(EPoint[X], EPoint[Y], EPoint[Z], 0);
const __m256d epsy = _mm256_set1_pd(1.0 - EPSILON);
const __m256d xyzn_e = _mm256_sub_pd(xyzn, epsy);
const __m128i tmp_xyzn = _mm256_cvttpd_epi32(_mm256_blendv_pd(xyzn, xyzn_e, xyzn));
const __m128i noise_min_xyzn = _mm_setr_epi32(NOISE_MINX, NOISE_MINY, NOISE_MINZ, 0);
const __m256d xyz_ixyzn = _mm256_sub_pd(xyzn, _mm256_cvtepi32_pd(tmp_xyzn));
const __m256d xyz_jxyzn = _mm256_sub_pd(xyz_ixyzn, ONE_PD);
const __m128i i_xyzn = _mm_and_si128(_mm_sub_epi32(tmp_xyzn, noise_min_xyzn),
_mm_set1_epi32(0xfff));
const __m256d s_xyzn = _mm256_mul_pd(xyz_ixyzn,
_mm256_mul_pd(xyz_ixyzn,
_mm256_sub_pd(_mm256_set1_pd(3.0),
_mm256_add_pd(xyz_ixyzn, xyz_ixyzn))));
const __m256d t_xyzn = _mm256_sub_pd(ONE_PD, s_xyzn);
const __m256d txtysxsy = _mm256_permute2f128_pd(t_xyzn, s_xyzn, 0x20);
const __m256d txsxtxsx = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(2, 0, 2, 0));
const __m256d tytysysy = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(3, 3, 1, 1));
const __m256d txtysxtytxsysxsy = _mm256_mul_pd(txsxtxsx, tytysysy);
const __m256d incrsump_s1 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(t_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
const __m256d incrsump_s2 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(s_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
int ints[4];
_mm_storeu_si128((__m128i*)(ints), i_xyzn);
const int ixiy_hash = Hash2d(ints[0], ints[1]);
const int jxiy_hash = Hash2d(ints[0] + 1, ints[1]);
const int ixjy_hash = Hash2d(ints[0], ints[1] + 1);
const int jxjy_hash = Hash2d(ints[0] + 1, ints[1] + 1);
const int iz = ints[2];
const __m256d iii = _mm256_blend_pd(PERMUTE4x64(xyz_ixyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
const __m256d jjj = _mm256_blend_pd(PERMUTE4x64(xyz_jxyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
__m256d ss;
__m256d blend;
__m256d x = _mm256_setzero_pd(), y = _mm256_setzero_pd(), z = _mm256_setzero_pd();
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(0, 0, 0, 0));
// blend = _mm256_blend_pd(iii, jjj, 0);
INCSUMAVX_VECTOR(mp, ss, iii);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(1, 1, 1, 1));
blend = _mm256_blend_pd(iii, jjj, 2);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(3, 3, 3, 3));
blend = _mm256_blend_pd(iii, jjj, 6);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(2, 2, 2, 2));
blend = _mm256_blend_pd(iii, jjj, 4);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(2, 2, 2, 2));
blend = _mm256_blend_pd(iii, jjj, 12);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(3, 3, 3, 3));
// blend = _mm256_blend_pd(iii, jjj, 14);
INCSUMAVX_VECTOR(mp, ss, jjj);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(1, 1, 1, 1));
blend = _mm256_blend_pd(iii, jjj, 10);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(0, 0, 0, 0));
blend = _mm256_blend_pd(iii, jjj, 8);
INCSUMAVX_VECTOR(mp, ss, blend);
__m256d xy = _mm256_hadd_pd(x,y);
__m128d xy_up = _mm256_extractf128_pd(xy,1);
xy_up = _mm_add_pd(_mm256_castpd256_pd128(xy),xy_up);
_mm_storeu_pd(&result[X],xy_up);
__m128d z_up = _mm256_extractf128_pd(z,1);
z_up = _mm_add_pd(_mm256_castpd256_pd128(z),z_up);
z_up = _mm_hadd_pd(z_up,z_up);
result[Z] = _mm_cvtsd_f64(z_up);
#if CHECK_FUNCTIONAL
{
Vector3d portable_res;
PortableDNoise(portable_res , param);
if (fabs(portable_res[X] - result[X]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise X error");
}
if (fabs(portable_res[Y] - result[Y]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise Y error");
}
if (fabs(portable_res[Z] - result[Z]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise Z error");
}
}
#endif
_mm256_zeroupper();
return;
}
DBL AVX2FMA3Noise(const Vector3d& EPoint, int noise_generator)
{
AVX2TABLETYPE *mp;
DBL sum = 0.0;
// TODO FIXME - global statistics reference
// Stats[Calls_To_Noise]++;
if (noise_generator == kNoiseGen_Perlin)
{
// The 1.59 and 0.985 are to correct for some biasing problems with
// the random # generator used to create the noise tables. Final
// range of values is about 5.0e-4 below 0.0 and above 1.0. Mean
// value is 0.49 (ideally it would be 0.5).
sum = 0.5 * (1.59 * SolidNoise(EPoint) + 0.985);
// Clamp final value to 0-1 range
if (sum < 0.0) sum = 0.0;
if (sum > 1.0) sum = 1.0;
return sum;
}
const __m256d ONE_PD = _mm256_set1_pd(1);
const __m128i short_si128 = _mm_set1_epi32(0xffff);
const __m256d xyzn = _mm256_setr_pd(EPoint[X], EPoint[Y], EPoint[Z], 0);
const __m256d epsy = _mm256_set1_pd(1.0 - EPSILON);
const __m256d xyzn_e = _mm256_sub_pd(xyzn, epsy);
const __m128i tmp_xyzn = _mm256_cvttpd_epi32(_mm256_blendv_pd(xyzn, xyzn_e, xyzn));
const __m128i noise_min_xyzn = _mm_setr_epi32(NOISE_MINX, NOISE_MINY, NOISE_MINZ, 0);
const __m256d xyz_ixyzn = _mm256_sub_pd(xyzn, _mm256_cvtepi32_pd(tmp_xyzn));
const __m256d xyz_jxyzn = _mm256_sub_pd(xyz_ixyzn, ONE_PD);
const __m128i i_xyzn = _mm_and_si128(_mm_sub_epi32(tmp_xyzn, noise_min_xyzn),
_mm_set1_epi32(0xfff));
const __m256d s_xyzn = _mm256_mul_pd(xyz_ixyzn,
_mm256_mul_pd(xyz_ixyzn,
_mm256_sub_pd(_mm256_set1_pd(3.0),
_mm256_add_pd(xyz_ixyzn, xyz_ixyzn))));
const __m256d t_xyzn = _mm256_sub_pd(ONE_PD, s_xyzn);
const __m256d txtysxsy = _mm256_permute2f128_pd(t_xyzn, s_xyzn, 0x20);
const __m256d txsxtxsx = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(2, 0, 2, 0));
const __m256d tytysysy = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(3, 3, 1, 1));
const __m256d txtysxtytxsysxsy = _mm256_mul_pd(txsxtxsx, tytysysy);
const __m256d incrsump_s1 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(t_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
const __m256d incrsump_s2 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(s_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
int ints[4];
_mm_storeu_si128((__m128i*)(ints), i_xyzn);
const int ixiy_hash = Hash2d(ints[0], ints[1]);
const int jxiy_hash = Hash2d(ints[0] + 1, ints[1]);
const int ixjy_hash = Hash2d(ints[0], ints[1] + 1);
const int jxjy_hash = Hash2d(ints[0] + 1, ints[1] + 1);
const int iz = ints[2];
const __m256d iii = _mm256_blend_pd(PERMUTE4x64(xyz_ixyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
const __m256d jjj = _mm256_blend_pd(PERMUTE4x64(xyz_jxyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
__m256d sumr = _mm256_setzero_pd();
__m256d sumr1 = _mm256_setzero_pd();
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz)];
INCSUMAVX_NOBLEND(sumr, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(0, 0, 0, 0)), iii);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(1, 1, 1, 1)), iii, jjj, 2);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(2, 2, 2, 2)), iii, jjj, 4);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(3, 3, 3, 3)), iii, jjj, 6);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz + 1)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(0, 0, 0, 0)), iii, jjj, 8);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz + 1)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(1, 1, 1, 1)), iii, jjj, 10);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz + 1)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(2, 2, 2, 2)), iii, jjj, 12);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz + 1)];
INCSUMAVX_NOBLEND(sumr1, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(3, 3, 3, 3)), jjj);
{
sumr = _mm256_add_pd(sumr, sumr1);
__m128d sumr_up = _mm256_extractf128_pd(sumr,1);
sumr_up = _mm_add_pd(_mm256_castpd256_pd128(sumr),sumr_up);
sumr_up = _mm_hadd_pd(sumr_up,sumr_up);
sum = _mm_cvtsd_f64(sumr_up);
}
if (noise_generator == kNoiseGen_RangeCorrected)
{
/* details of range here:
Min, max: -1.05242, 0.988997
Mean: -0.0191481, Median: -0.535493, Std Dev: 0.256828
We want to change it to as close to [0,1] as possible.
*/
sum += 1.05242;
sum *= 0.48985582;
/*sum *= 0.5;
sum += 0.5;*/
if (sum < 0.0)
sum = 0.0;
if (sum > 1.0)
sum = 1.0;
}
else
{
sum = sum + 0.5; /* range at this point -0.5 - 0.5... */
if (sum < 0.0)
sum = 0.0;
if (sum > 1.0)
sum = 1.0;
}
#if CHECK_FUNCTIONAL
{
DBL orig_sum = PortableNoise(EPoint, noise_generator);
if (fabs(orig_sum - sum) >= EPSILON)
{
throw POV_EXCEPTION_STRING("Noise error");
}
}
#endif
_mm256_zeroupper();
return (sum);
}
void
foo ()
{
bar2 (y);
_mm256_zeroupper ();
}
void TransLut::process_plane_flt_any_avx2 (uint8_t *dst_ptr, const uint8_t *src_ptr, int stride_dst, int stride_src, int w, int h)
{
assert (dst_ptr != 0);
assert (src_ptr != 0);
assert (stride_dst != 0 || h == 1);
assert (stride_src != 0 || h == 1);
assert (w > 0);
assert (h > 0);
for (int y = 0; y < h; ++y)
{
const FloatIntMix * s_ptr =
reinterpret_cast <const FloatIntMix *> (src_ptr);
TD * d_ptr =
reinterpret_cast < TD *> (dst_ptr);
for (int x = 0; x < w; x += 8)
{
union
{
__m256i _vect;
uint32_t _scal [8];
} index;
__m256 lerp;
TransLut_FindIndexAvx2 <M>::find_index (s_ptr + x, index._vect, lerp);
#if 1 // Looks as fast as _mm256_set_ps
// G++ complains about sizeof() as argument
__m256 val = _mm256_i32gather_ps (
&_lut.use <float> (0), index._vect, 4 // 4 == sizeof (float)
);
const __m256 va2 = _mm256_i32gather_ps (
&_lut.use <float> (1), index._vect, 4 // 4 == sizeof (float)
);
#else
__m256 val = _mm256_set_ps (
_lut.use <float> (index._scal [7] ),
_lut.use <float> (index._scal [6] ),
_lut.use <float> (index._scal [5] ),
_lut.use <float> (index._scal [4] ),
_lut.use <float> (index._scal [3] ),
_lut.use <float> (index._scal [2] ),
_lut.use <float> (index._scal [1] ),
_lut.use <float> (index._scal [0] )
);
const __m256 va2 = _mm256_set_ps (
_lut.use <float> (index._scal [7] + 1),
_lut.use <float> (index._scal [6] + 1),
_lut.use <float> (index._scal [5] + 1),
_lut.use <float> (index._scal [4] + 1),
_lut.use <float> (index._scal [3] + 1),
_lut.use <float> (index._scal [2] + 1),
_lut.use <float> (index._scal [1] + 1),
_lut.use <float> (index._scal [0] + 1)
);
#endif
const __m256 dif = _mm256_sub_ps (va2, val);
val = _mm256_add_ps (val, _mm256_mul_ps (dif, lerp));
TransLut_store_avx2 (&d_ptr [x], val);
}
src_ptr += stride_src;
dst_ptr += stride_dst;
}
_mm256_zeroupper (); // Back to SSE state
}
void aom_sad64x64x4d_avx2(const uint8_t *src, int src_stride,
const uint8_t *const ref[4], int ref_stride,
uint32_t res[4]) {
__m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg;
__m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg;
__m256i ref3_reg, ref3next_reg;
__m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
__m256i sum_mlow, sum_mhigh;
int i;
const uint8_t *ref0, *ref1, *ref2, *ref3;
ref0 = ref[0];
ref1 = ref[1];
ref2 = ref[2];
ref3 = ref[3];
sum_ref0 = _mm256_set1_epi16(0);
sum_ref1 = _mm256_set1_epi16(0);
sum_ref2 = _mm256_set1_epi16(0);
sum_ref3 = _mm256_set1_epi16(0);
for (i = 0; i < 64; i++) {
// load 64 bytes from src and all refs
src_reg = _mm256_loadu_si256((const __m256i *)src);
srcnext_reg = _mm256_loadu_si256((const __m256i *)(src + 32));
ref0_reg = _mm256_loadu_si256((const __m256i *)ref0);
ref0next_reg = _mm256_loadu_si256((const __m256i *)(ref0 + 32));
ref1_reg = _mm256_loadu_si256((const __m256i *)ref1);
ref1next_reg = _mm256_loadu_si256((const __m256i *)(ref1 + 32));
ref2_reg = _mm256_loadu_si256((const __m256i *)ref2);
ref2next_reg = _mm256_loadu_si256((const __m256i *)(ref2 + 32));
ref3_reg = _mm256_loadu_si256((const __m256i *)ref3);
ref3next_reg = _mm256_loadu_si256((const __m256i *)(ref3 + 32));
// sum of the absolute differences between every ref-i to src
ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg);
ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg);
ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg);
ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg);
// sum every ref-i
sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg);
sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg);
sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg);
sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg);
src += src_stride;
ref0 += ref_stride;
ref1 += ref_stride;
ref2 += ref_stride;
ref3 += ref_stride;
}
{
__m128i sum;
// in sum_ref-i the result is saved in the first 4 bytes
// the other 4 bytes are zeroed.
// sum_ref1 and sum_ref3 are shifted left by 4 bytes
sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
// merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
// merge every 64 bit from each sum_ref-i
sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
// add the low 64 bit to the high 64 bit
sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
// add the low 128 bit to the high 128 bit
sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
_mm256_extractf128_si256(sum_mlow, 1));
_mm_storeu_si128((__m128i *)(res), sum);
}
_mm256_zeroupper();
}
void fastGEMM( const float* aptr, size_t astep, const float* bptr,
size_t bstep, float* cptr, size_t cstep,
int ma, int na, int nb )
{
int n = 0;
for( ; n <= nb - 16; n += 16 )
{
for( int m = 0; m < ma; m += 4 )
{
const float* aptr0 = aptr + astep*m;
const float* aptr1 = aptr + astep*std::min(m+1, ma-1);
const float* aptr2 = aptr + astep*std::min(m+2, ma-1);
const float* aptr3 = aptr + astep*std::min(m+3, ma-1);
float* cptr0 = cptr + cstep*m;
float* cptr1 = cptr + cstep*std::min(m+1, ma-1);
float* cptr2 = cptr + cstep*std::min(m+2, ma-1);
float* cptr3 = cptr + cstep*std::min(m+3, ma-1);
__m256 d00 = _mm256_setzero_ps(), d01 = _mm256_setzero_ps();
__m256 d10 = _mm256_setzero_ps(), d11 = _mm256_setzero_ps();
__m256 d20 = _mm256_setzero_ps(), d21 = _mm256_setzero_ps();
__m256 d30 = _mm256_setzero_ps(), d31 = _mm256_setzero_ps();
for( int k = 0; k < na; k++ )
{
__m256 a0 = _mm256_set1_ps(aptr0[k]);
__m256 a1 = _mm256_set1_ps(aptr1[k]);
__m256 a2 = _mm256_set1_ps(aptr2[k]);
__m256 a3 = _mm256_set1_ps(aptr3[k]);
__m256 b0 = _mm256_loadu_ps(bptr + k*bstep + n);
__m256 b1 = _mm256_loadu_ps(bptr + k*bstep + n + 8);
d00 = _mm256_fmadd_ps(a0, b0, d00);
d01 = _mm256_fmadd_ps(a0, b1, d01);
d10 = _mm256_fmadd_ps(a1, b0, d10);
d11 = _mm256_fmadd_ps(a1, b1, d11);
d20 = _mm256_fmadd_ps(a2, b0, d20);
d21 = _mm256_fmadd_ps(a2, b1, d21);
d30 = _mm256_fmadd_ps(a3, b0, d30);
d31 = _mm256_fmadd_ps(a3, b1, d31);
}
_mm256_storeu_ps(cptr0 + n, d00);
_mm256_storeu_ps(cptr0 + n + 8, d01);
_mm256_storeu_ps(cptr1 + n, d10);
_mm256_storeu_ps(cptr1 + n + 8, d11);
_mm256_storeu_ps(cptr2 + n, d20);
_mm256_storeu_ps(cptr2 + n + 8, d21);
_mm256_storeu_ps(cptr3 + n, d30);
_mm256_storeu_ps(cptr3 + n + 8, d31);
}
}
for( ; n < nb; n++ )
{
for( int m = 0; m < ma; m++ )
{
const float* aptr0 = aptr + astep*m;
float* cptr0 = cptr + cstep*m;
float d0 = 0.f;
for( int k = 0; k < na; k++ )
d0 += aptr0[k]*bptr[k*bstep + n];
cptr0[n] = d0;
}
}
_mm256_zeroupper();
}
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps)
{
const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
uint32_t partitions = 1u << max_partition_order;
FLAC__ASSERT(default_partition_samples > predictor_order);
/* first do max_partition_order */
{
const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples);
uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order);
if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) {
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-7; residual_sample+=8) {
__m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample)));
sum256 = _mm256_add_epi32(sum256, res256);
}
sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
sum128 = _mm_add_epi32(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi32(sum128, res128);
}
sum128 = _mm_hadd_epi32(sum128, sum128);
sum128 = _mm_hadd_epi32(sum128, sum128);
abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128);
/* workaround for a bug in MSVC2015U2 - see https://connect.microsoft.com/VisualStudio/feedback/details/2659191/incorrect-code-generation-for-x86-64 */
#if (defined _MSC_VER) && (_MSC_FULL_VER == 190023918) && (defined FLAC__CPU_X86_64)
abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/
#endif
}
}
else { /* have to pessimistically use 64 bits for accumulator */
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
__m256i res256 = _mm256_cvtepu32_epi64(res128);
sum256 = _mm256_add_epi64(sum256, res256);
}
sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-1; residual_sample+=2) {
__m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample)));
res128 = _mm_cvtepu32_epi64(res128);
sum128 = _mm_add_epi64(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi64(sum128, res128);
}
sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8));
_mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128);
}
}
}
/* now merge partitions for lower orders */
{
uint32_t from_partition = 0, to_partition = partitions;
int partition_order;
for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
uint32_t i;
partitions >>= 1;
for(i = 0; i < partitions; i++) {
abs_residual_partition_sums[to_partition++] =
abs_residual_partition_sums[from_partition ] +
abs_residual_partition_sums[from_partition+1];
from_partition += 2;
}
}
}
_mm256_zeroupper();
}
extern void
baseInputs(const TanBoard anBoard, float arInput[])
{
int i = 3;
const unsigned int *pB = &anBoard[0][0];
float *pInput = &arInput[0];
register __m128 vec0;
register __m128 vec1;
register __m128 vec2;
register __m128 vec3;
register __m128 vec4;
register __m128 vec5;
register __m128 vec6;
register __m128 vec7;
while (i--) {
vec0 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput, vec0);
vec1 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec1);
vec2 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec2);
vec3 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec3);
vec4 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec4);
vec5 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec5);
vec6 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec6);
vec7 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec7);
pInput += 4;
}
/* bar */
vec0 = _mm_load_ps(inpvecb[*pB++]);
_mm_store_ps(pInput, vec0);
pInput += 4;
i = 3;
while (i--) {
vec0 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput, vec0);
vec1 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec1);
vec2 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec2);
vec3 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec3);
vec4 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec4);
vec5 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec5);
vec6 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec6);
vec7 = _mm_load_ps(inpvec[*pB++]);
_mm_store_ps(pInput += 4, vec7);
pInput += 4;
}
/* bar */
vec0 = _mm_load_ps(inpvecb[*pB]);
_mm_store_ps(pInput, vec0);
#if defined(USE_AVX)
_mm256_zeroupper();
#endif
return;
}
__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;
}
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;
}
void Scaler::process_plane_flt_avx2 (typename DST::Ptr::Type dst_ptr, typename SRC::PtrConst::Type src_ptr, int dst_stride, int src_stride, int width, int y_dst_beg, int y_dst_end) const
{
assert (DST::Ptr::check_ptr (dst_ptr, DST::ALIGN_W));
assert (SRC::PtrConst::check_ptr (src_ptr, SRC::ALIGN_R));
// When the destination is a buffer:
// mod4 is enough to guarantee alignment, but since we process pairs of
// vectors and write_partial() is not different from write() with float
// data (overwriting all the 64 bytes), we must take extra-care not to
// overflow from the output buffer.
// When the destination is not a buffer (output frame data), data may be
// unaligned anyway. (TO DO: check the algorithm and make sure this
// constraint is actual).
assert ((dst_stride & 15) == 0);
assert ((src_stride & 3) == 0);
assert (width > 0);
assert (y_dst_beg >= 0);
assert (y_dst_beg < y_dst_end);
assert (y_dst_end <= _dst_height);
const __m256i zero = _mm256_setzero_si256 ();
const __m256i mask_lsb = _mm256_set1_epi16 (0x00FF);
const __m256i sign_bit = _mm256_set1_epi16 (-0x8000);
const __m256 offset = _mm256_set1_ps (float (DST::OFFSET));
const __m256 add_cst = _mm256_set1_ps (float (_add_cst_flt));
const int w16 = width & -16;
const int w15 = width - w16;
for (int y = y_dst_beg; y < y_dst_end; ++y)
{
const KernelInfo& kernel_info = _kernel_info_arr [y];
const int kernel_size = kernel_info._kernel_size;
const float * coef_base_ptr = &_coef_flt_arr [kernel_info._coef_index];
const int ofs_y = kernel_info._start_line;
typename SRC::PtrConst::Type col_src_ptr = src_ptr;
SRC::PtrConst::jump (col_src_ptr, src_stride * ofs_y);
typename DST::Ptr::Type col_dst_ptr = dst_ptr;
typedef ScalerCopy <DST, 0, SRC, 0> ScCopy;
if (ScCopy::can_copy (kernel_info._copy_flt_flag))
{
ScCopy::copy (col_dst_ptr, col_src_ptr, width);
}
else
{
__m256 sum0;
__m256 sum1;
for (int x = 0; x < w16; x += 16)
{
typename SRC::PtrConst::Type pix_ptr = col_src_ptr;
process_vect_flt_avx2 <SRC, false> (
sum0, sum1, kernel_size, coef_base_ptr,
pix_ptr, zero, src_stride, add_cst, 0
);
DST::write_flt (
col_dst_ptr, sum0, sum1, mask_lsb, sign_bit, offset
);
DST::Ptr::jump (col_dst_ptr, 16);
SRC::PtrConst::jump (col_src_ptr, 16);
}
if (w15 > 0)
{
typename SRC::PtrConst::Type pix_ptr = col_src_ptr;
process_vect_flt_avx2 <SRC, true> (
sum0, sum1, kernel_size, coef_base_ptr,
pix_ptr, zero, src_stride, add_cst, w15
);
DST::write_flt_partial (
col_dst_ptr, sum0, sum1, mask_lsb, sign_bit, offset, w15
);
}
}
DST::Ptr::jump (dst_ptr, dst_stride);
}
_mm256_zeroupper (); // Back to SSE state
}