__m128 modulus_128(
const __m128& x,
const __m128& y) {
__m128 z = x;
__m128 n = _mm_round_ps((-z) / y, _MM_FROUND_TO_ZERO) + _mm_set1_ps(1.f);
__m128 mask = _mm_cmplt_ps(z, _mm_set1_ps(0.f));
z = applyMask_ps(mask, z + n * y, z);
n = _mm_round_ps(z / y, _MM_FROUND_TO_ZERO);
return z - n * y;
}
__m128 exp_128(
const __m128& x) {
//! Clip the value
__m128 y = _mm_max_ps(_mm_min_ps(x, _mm_set1_ps(88.3762626647949f)),
_mm_set1_ps(-88.3762626647949f));
//! Express exp(x) as exp(g + n * log(2))
__m128 fx = y * _mm_set1_ps(1.44269504088896341) + _mm_set1_ps(0.5f);
//! Floor
const __m128 tmp = _mm_round_ps(fx, _MM_FROUND_TO_ZERO);
//! If greater, substract 1
const __m128 mask = _mm_and_ps(_mm_cmpgt_ps(tmp, fx), _mm_set1_ps(1.f));
fx = tmp - mask;
y -= fx * _mm_set1_ps(0.693359375 - 2.12194440e-4);
const __m128 z = y * y;
const __m128 t = (((((_mm_set1_ps(1.9875691500E-4) * y +
_mm_set1_ps(1.3981999507E-3)) * y +
_mm_set1_ps(8.3334519073E-3)) * y +
_mm_set1_ps(4.1665795894E-2)) * y +
_mm_set1_ps(1.6666665459E-1)) * y +
_mm_set1_ps(5.0000001201E-1)) * z + y + _mm_set1_ps(1.f);
//! Build 2^n
const __m128i emm0 = _mm_add_epi32(_mm_cvttps_epi32(fx), _mm_set1_epi32(0x7f));
//! Return the result
return t * _mm_castsi128_ps(_mm_slli_epi32(emm0, 23));
}
__m128 ori_to_bin_128(
const __m128& ori,
const int nbins) {
//! For convenience
const __m128 x2PI = _mm_set1_ps(2 * M_PI);
const __m128 xbins = _mm_set1_ps(nbins);
//! Get it positive
const __m128 mask = _mm_cmplt_ps(ori, _mm_setzero_ps());
//! Get the value
const __m128 val = _mm_round_ps(applyMask_ps(mask, ori + x2PI, ori)
/ x2PI * xbins + _mm_set1_ps(0.5f), _MM_FROUND_TO_ZERO);
//! Return the modulo of it
return val - xbins * _mm_round_ps(val / xbins, _MM_FROUND_TO_ZERO);
}
void
test4bit (void)
{
d1 = _mm_round_pd (d2, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
d1 = _mm_round_sd (d2, d3, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
a1 = _mm_round_ps (a2, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
a1 = _mm_round_ss (a2, a2, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
a1 = _mm_blend_ps (a2, a3, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
e1 = _mm256_blend_pd (e2, e3, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
e1 = _mm256_round_pd (e2, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
b1 = _mm256_round_ps (b2, k4); /* { dg-error "the last argument must be a 4-bit immediate" } */
}
int dist_mic(const float* xyz, const int* pairs, const float* box_matrix,
float* distance_out, float* displacement_out,
const int n_frames, const int n_atoms, const int n_pairs) {
/* Compute the distance/displacement between pairs of atoms in every frame
of xyz following the minimum image convention in periodic boundary
conditions.
The computation follows scheme B.9 in Tukerman, M. "Statistical
Mechanics: Theory and Molecular Simulation", 2010.
Parameters
----------
xyz : array, shape=(n_frames, n_atoms, 3)
Cartesian coordinates of the atoms in every frame, in contiguous C order.
pairs : array, shape=(n_pairs, 2)
The specific pairs of atoms whose distance you want to compute. A 2d
array of pairs, in C order.
box_matrix : array, shape=(3,3)
The box matrix for a single frame. All of the frames are assumed to
use this box vector.
distance_out : array, shape=(n_frames, n_pairs)
Array where the distances between pairs will be stored, in contiguous
C order.
displacement_out : array, shaoe=(n_frames, n_pairs, 3), optional
An optional return value: if you'd also like to save the displacement
vectors between the pairs, you can pass a pointer here. If
displacement_out is NULL, then this variable will not be saved back
to memory.
All of the arrays are assumed to be contiguous. This code will
segfault if they're not.
*/
#ifndef __SSE4_1__
_MM_SET_ROUNDING_MODE(_MM_ROUND_NEAREST);
int rounding_mode = _MM_GET_ROUNDING_MODE();
#endif
int i, j;
int store_displacement = displacement_out == NULL ? 0 : 1;
int store_distance = distance_out == NULL ? 0 : 1;
__m128 r1, r2, s12, r12, s, r12_2;
__m128 hinv[3];
__m128 h[3];
for (i = 0; i < n_frames; i++) {
// Store the columns of the box matrix in three float4s. This format
// is fast for matrix * vector product. See, for example, this S.O. question:
// http://stackoverflow.com/questions/14967969/efficient-4x4-matrix-vector-multiplication-with-sse-horizontal-add-and-dot-prod
h[0] = _mm_setr_ps(box_matrix[0], box_matrix[3], box_matrix[6], 0.0f);
h[1] = _mm_setr_ps(box_matrix[1], box_matrix[4], box_matrix[7], 0.0f);
h[2] = _mm_setr_ps(box_matrix[2], box_matrix[5], box_matrix[8], 0.0f);
// Calculate the inverse of the box matrix, and also store it in the same
// format.
inverse33(box_matrix, hinv+0, hinv+1, hinv+2);
for (j = 0; j < n_pairs; j++) {
// Load the two vectors whos distance we want to compute
r1 = load_float3(xyz + 3*pairs[2*j + 0]);
r2 = load_float3(xyz + 3*pairs[2*j + 1]);
r12 = _mm_sub_ps(r2, r1);
// s12 = INVERSE(H) * r12
s12 = _mm_add_ps(_mm_add_ps(
_mm_mul_ps(hinv[0], _mm_shuffle_ps(r12, r12, _MM_SHUFFLE(0,0,0,0))),
_mm_mul_ps(hinv[1], _mm_shuffle_ps(r12, r12, _MM_SHUFFLE(1,1,1,1)))),
_mm_mul_ps(hinv[2], _mm_shuffle_ps(r12, r12, _MM_SHUFFLE(2,2,2,2))));
// s12 = s12 - NEAREST_INTEGER(s12)
#ifdef __SSE4_1__
s12 = _mm_sub_ps(s12, _mm_round_ps(s12, _MM_FROUND_TO_NEAREST_INT));
#else
s12 = _mm_sub_ps(s12, _mm_cvtepi32_ps(_mm_cvtps_epi32(s12)));
#endif
r12 = _mm_add_ps(_mm_add_ps(
_mm_mul_ps(h[0], _mm_shuffle_ps(s12, s12, _MM_SHUFFLE(0,0,0,0))),
_mm_mul_ps(h[1], _mm_shuffle_ps(s12, s12, _MM_SHUFFLE(1,1,1,1)))),
_mm_mul_ps(h[2], _mm_shuffle_ps(s12, s12, _MM_SHUFFLE(2,2,2,2))));
if (store_displacement) {
// store the two lower entries (x,y) in memory
_mm_storel_pi((__m64*)(displacement_out), r12);
displacement_out += 2;
// swap high-low and then store the z entry in the memory
_mm_store_ss(displacement_out++, _mm_movehl_ps(r12, r12));
}
if (store_distance) {
// out = sqrt(sum(r12**2))
r12_2 = _mm_mul_ps(r12, r12);
s = _mm_hadd_ps(r12_2, r12_2);
s = _mm_hadd_ps(s, s);
s = _mm_sqrt_ps(s);
_mm_store_ss(distance_out++, s);
}
}
// advance to the next frame
xyz += n_atoms*3;
box_matrix += 9;
}
#ifndef __SSE4_1__
_MM_SET_ROUNDING_MODE(rounding_mode);
#endif
return 1;
}
void Permutohedral::init ( const float* feature, int feature_size, int N )
{
// Compute the lattice coordinates for each feature [there is going to be a lot of magic here
N_ = N;
d_ = feature_size;
HashTable hash_table( d_, N_/**(d_+1)*/ );
const int blocksize = sizeof(__m128) / sizeof(float);
const __m128 invdplus1 = _mm_set1_ps( 1.0f / (d_+1) );
const __m128 dplus1 = _mm_set1_ps( d_+1 );
const __m128 Zero = _mm_set1_ps( 0 );
const __m128 One = _mm_set1_ps( 1 );
// Allocate the class memory
if (offset_) delete [] offset_;
offset_ = new int[ (d_+1)*(N_+16) ];
memset( offset_, 0, (d_+1)*(N_+16)*sizeof(int) );
if (barycentric_) delete [] barycentric_;
barycentric_ = new float[ (d_+1)*(N_+16) ];
memset( barycentric_, 0, (d_+1)*(N_+16)*sizeof(float) );
// Allocate the local memory
__m128 * scale_factor = (__m128*) _mm_malloc( (d_ )*sizeof(__m128) , 16 );
__m128 * f = (__m128*) _mm_malloc( (d_ )*sizeof(__m128) , 16 );
__m128 * elevated = (__m128*) _mm_malloc( (d_+1)*sizeof(__m128) , 16 );
__m128 * rem0 = (__m128*) _mm_malloc( (d_+1)*sizeof(__m128) , 16 );
__m128 * rank = (__m128*) _mm_malloc( (d_+1)*sizeof(__m128), 16 );
float * barycentric = new float[(d_+2)*blocksize];
short * canonical = new short[(d_+1)*(d_+1)];
short * key = new short[d_+1];
// Compute the canonical simplex
for( int i=0; i<=d_; i++ ){
for( int j=0; j<=d_-i; j++ )
canonical[i*(d_+1)+j] = i;
for( int j=d_-i+1; j<=d_; j++ )
canonical[i*(d_+1)+j] = i - (d_+1);
}
// Expected standard deviation of our filter (p.6 in [Adams etal 2010])
float inv_std_dev = sqrt(2.0 / 3.0)*(d_+1);
// Compute the diagonal part of E (p.5 in [Adams etal 2010])
for( int i=0; i<d_; i++ )
scale_factor[i] = _mm_set1_ps( 1.0 / sqrt( float((i+2)*(i+1) ) * inv_std_dev) );
// Setup the SSE rounding
#ifndef __SSE4_1__
const unsigned int old_rounding = _mm_getcsr();
_mm_setcsr( (old_rounding&~_MM_ROUND_MASK) | _MM_ROUND_NEAREST );
#endif
// Compute the simplex each feature lies in
for( int k=0; k<N_; k+=blocksize ){
// Load the feature from memory
float * ff = (float*)f;
for( int j=0; j<d_; j++ )
for( int i=0; i<blocksize; i++ )
ff[ j*blocksize + i ] = k+i < N_ ? feature[ (k+i)*d_+j ] : 0.0;
// Elevate the feature ( y = Ep, see p.5 in [Adams etal 2010])
// sm contains the sum of 1..n of our faeture vector
__m128 sm = Zero;
for( int j=d_; j>0; j-- ){
__m128 cf = f[j-1]*scale_factor[j-1];
elevated[j] = sm - _mm_set1_ps(j)*cf;
sm += cf;
}
elevated[0] = sm;
// Find the closest 0-colored simplex through rounding
__m128 sum = Zero;
for( int i=0; i<=d_; i++ ){
__m128 v = invdplus1 * elevated[i];
#ifdef __SSE4_1__
v = _mm_round_ps( v, _MM_FROUND_TO_NEAREST_INT );
#else
v = _mm_cvtepi32_ps( _mm_cvtps_epi32( v ) );
#endif
rem0[i] = v*dplus1;
sum += v;
}
// Find the simplex we are in and store it in rank (where rank describes what position coorinate i has in the sorted order of the features values)
for( int i=0; i<=d_; i++ )
rank[i] = Zero;
for( int i=0; i<d_; i++ ){
__m128 di = elevated[i] - rem0[i];
for( int j=i+1; j<=d_; j++ ){
__m128 dj = elevated[j] - rem0[j];
__m128 c = _mm_and_ps( One, _mm_cmplt_ps( di, dj ) );
rank[i] += c;
rank[j] += One-c;
}
}
// If the point doesn't lie on the plane (sum != 0) bring it back
for( int i=0; i<=d_; i++ ){
rank[i] += sum;
__m128 add = _mm_and_ps( dplus1, _mm_cmplt_ps( rank[i], Zero ) );
__m128 sub = _mm_and_ps( dplus1, _mm_cmpge_ps( rank[i], dplus1 ) );
rank[i] += add-sub;
rem0[i] += add-sub;
}
// Compute the barycentric coordinates (p.10 in [Adams etal 2010])
for( int i=0; i<(d_+2)*blocksize; i++ )
barycentric[ i ] = 0;
for( int i=0; i<=d_; i++ ){
__m128 v = (elevated[i] - rem0[i])*invdplus1;
// Didn't figure out how to SSE this
float * fv = (float*)&v;
float * frank = (float*)&rank[i];
for( int j=0; j<blocksize; j++ ){
int p = d_-frank[j];
barycentric[j*(d_+2)+p ] += fv[j];
barycentric[j*(d_+2)+p+1] -= fv[j];
}
}
// The rest is not SSE'd
for( int j=0; j<blocksize; j++ ){
// Wrap around
barycentric[j*(d_+2)+0]+= 1 + barycentric[j*(d_+2)+d_+1];
float * frank = (float*)rank;
float * frem0 = (float*)rem0;
// Compute all vertices and their offset
for( int remainder=0; remainder<=d_; remainder++ ){
for( int i=0; i<d_; i++ ){
key[i] = frem0[i*blocksize+j] + canonical[ remainder*(d_+1) + (int)frank[i*blocksize+j] ];
}
offset_[ (j+k)*(d_+1)+remainder ] = hash_table.find( key, true );
barycentric_[ (j+k)*(d_+1)+remainder ] = barycentric[ j*(d_+2)+remainder ];
}
}
}
_mm_free( scale_factor );
_mm_free( f );
_mm_free( elevated );
_mm_free( rem0 );
_mm_free( rank );
delete [] barycentric;
delete [] canonical;
delete [] key;
// Reset the SSE rounding
#ifndef __SSE4_1__
_mm_setcsr( old_rounding );
#endif
// This is normally fast enough so no SSE needed here
// Find the Neighbors of each lattice point
// Get the number of vertices in the lattice
M_ = hash_table.size();
// Create the neighborhood structure
if(blur_neighbors_) delete[] blur_neighbors_;
blur_neighbors_ = new Neighbors[ (d_+1)*M_ ];
short * n1 = new short[d_+1];
short * n2 = new short[d_+1];
// For each of d+1 axes,
for( int j = 0; j <= d_; j++ ){
for( int i=0; i<M_; i++ ){
const short * key = hash_table.getKey( i );
for( int k=0; k<d_; k++ ){
n1[k] = key[k] - 1;
n2[k] = key[k] + 1;
}
n1[j] = key[j] + d_;
n2[j] = key[j] - d_;
blur_neighbors_[j*M_+i].n1 = hash_table.find( n1 );
blur_neighbors_[j*M_+i].n2 = hash_table.find( n2 );
}
}
delete[] n1;
delete[] n2;
}
static __m128i cielabv (union hvrgbpix rgb)
{
__m128 xvxyz[2] = {_mm_set1_ps(0.5),_mm_set1_ps(0.5) }; //,0.5,0.5,0.5);
__m128 vcam0 = _mm_setr_ps(cielab_xyz_cam[0][0],cielab_xyz_cam[1][0],cielab_xyz_cam[2][0],0);
__m128 vcam1 = _mm_setr_ps(cielab_xyz_cam[0][1],cielab_xyz_cam[1][1],cielab_xyz_cam[2][1],0);
__m128 vcam2 = _mm_setr_ps(cielab_xyz_cam[0][2],cielab_xyz_cam[1][2],cielab_xyz_cam[2][2],0);
__m128 vrgb0h = _mm_set1_ps(rgb.h.c[0]);
__m128 vrgb1h = _mm_set1_ps(rgb.h.c[1]);
__m128 vrgb2h = _mm_set1_ps(rgb.h.c[2]);
__m128 vrgb0v = _mm_set1_ps(rgb.v.c[0]);
__m128 vrgb1v = _mm_set1_ps(rgb.v.c[1]);
__m128 vrgb2v = _mm_set1_ps(rgb.v.c[2]);
xvxyz[0] = _mm_add_ps(xvxyz[0], _mm_mul_ps(vcam0,vrgb0h));
xvxyz[0] = _mm_add_ps(xvxyz[0], _mm_mul_ps(vcam1,vrgb1h));
xvxyz[0] = _mm_add_ps(xvxyz[0], _mm_mul_ps(vcam2,vrgb2h));
xvxyz[1] = _mm_add_ps(xvxyz[1], _mm_mul_ps(vcam0,vrgb0v));
xvxyz[1] = _mm_add_ps(xvxyz[1], _mm_mul_ps(vcam1,vrgb1v));
xvxyz[1] = _mm_add_ps(xvxyz[1], _mm_mul_ps(vcam2,vrgb2v));
xvxyz[0] = _mm_max_ps(_mm_set1_ps(0),
_mm_min_ps(_mm_set1_ps(0xffff),
_mm_round_ps(xvxyz[0], _MM_FROUND_TO_ZERO)));
xvxyz[1] = _mm_max_ps(_mm_set1_ps(0),
_mm_min_ps(_mm_set1_ps(0xffff),
_mm_round_ps(xvxyz[1], _MM_FROUND_TO_ZERO)));
__m128i loadaddrh = _mm_cvttps_epi32(xvxyz[0]);
__m128i loadaddrv = _mm_cvttps_epi32(xvxyz[1]);
#ifdef __AVX__
__m256 vlab,
vxyz = { cielab_cbrt[_mm_extract_epi32(loadaddrh,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,0)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,1)],
0,
cielab_cbrt[_mm_extract_epi32(loadaddrv,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,0)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,1)],
0},
vxyz2 = {0,
cielab_cbrt[_mm_extract_epi32(loadaddrh,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,2)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,0)],
0,
cielab_cbrt[_mm_extract_epi32(loadaddrv,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,2)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,0)]};
vlab = _mm256_sub_ps(vxyz,vxyz2);
vlab = _mm256_mul_ps(vlab, _mm256_setr_ps(116,500,200,0,116,500,200,0));
vlab = _mm256_sub_ps(vlab, _mm256_setr_ps(16,0,0,0,16,0,0,0));
vlab = _mm256_mul_ps(vlab,_mm256_set1_ps(64));
vlab = _mm256_round_ps(vlab, _MM_FROUND_TO_ZERO);
__m256i vlabi = _mm256_cvtps_epi32(vlab);
return _mm_packs_epi32(_mm256_castsi256_si128(vlabi), ((__m128i*)&vlabi)[1]);
#else
__m128 vlabh, vxyzh = {cielab_cbrt[_mm_extract_epi32(loadaddrh,0)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrh,2)],
0};
__m128 vlabv, vxyzv = {cielab_cbrt[_mm_extract_epi32(loadaddrv,0)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,1)],
cielab_cbrt[_mm_extract_epi32(loadaddrv,2)],
0};
vlabh = _mm_sub_ps(_mm_shuffle_ps(vxyzh,vxyzh,_MM_SHUFFLE(0,1,0,1)),
_mm_shuffle_ps(vxyzh,vxyzh,_MM_SHUFFLE(0,2,1,3)));
vlabh = _mm_mul_ps(vlabh,_mm_setr_ps(116,500,200,0));
vlabh = _mm_sub_ps(vlabh,_mm_setr_ps(16,0,0,0));
vlabh = _mm_mul_ps(vlabh,_mm_set_ps1(64));
vlabh = _mm_round_ps(vlabh, _MM_FROUND_TO_ZERO);
vlabv = _mm_sub_ps(_mm_shuffle_ps(vxyzv,vxyzv,_MM_SHUFFLE(0,1,0,1)),
_mm_shuffle_ps(vxyzv,vxyzv,_MM_SHUFFLE(0,2,1,3)));
vlabv = _mm_mul_ps(vlabv,_mm_setr_ps(116,500,200,0));
vlabv = _mm_sub_ps(vlabv,_mm_setr_ps(16,0,0,0));
vlabv = _mm_mul_ps(vlabv,_mm_set_ps1(64));
vlabv = _mm_round_ps(vlabv, _MM_FROUND_TO_ZERO);
return _mm_set_epi64(_mm_cvtps_pi16(vlabv),_mm_cvtps_pi16(vlabh));
#endif
}
void convert_to_rgb_fast()
{
unsigned i,j,c;
int row, col, k;
ushort *img;
float out_cam[3][4];
double num, inverse[3][3];
static const double xyzd50_srgb[3][3] =
{ { 0.436083, 0.385083, 0.143055 },
{ 0.222507, 0.716888, 0.060608 },
{ 0.013930, 0.097097, 0.714022 } };
static const double rgb_rgb[3][3] =
{ { 1,0,0 }, { 0,1,0 }, { 0,0,1 } };
static const double adobe_rgb[3][3] =
{ { 0.715146, 0.284856, 0.000000 },
{ 0.000000, 1.000000, 0.000000 },
{ 0.000000, 0.041166, 0.958839 } };
static const double wide_rgb[3][3] =
{ { 0.593087, 0.404710, 0.002206 },
{ 0.095413, 0.843149, 0.061439 },
{ 0.011621, 0.069091, 0.919288 } };
static const double prophoto_rgb[3][3] =
{ { 0.529317, 0.330092, 0.140588 },
{ 0.098368, 0.873465, 0.028169 },
{ 0.016879, 0.117663, 0.865457 } };
static const double (*out_rgb[])[3] =
{ rgb_rgb, adobe_rgb, wide_rgb, prophoto_rgb, xyz_rgb };
static const char *name[] =
{ "sRGB", "Adobe RGB (1998)", "WideGamut D65", "ProPhoto D65", "XYZ" };
static const unsigned phead[] =
{ 1024, 0, 0x2100000, 0x6d6e7472, 0x52474220, 0x58595a20, 0, 0, 0,
0x61637370, 0, 0, 0x6e6f6e65, 0, 0, 0, 0, 0xf6d6, 0x10000, 0xd32d };
unsigned pbody[] =
{ 10, 0x63707274, 0, 36, /* cprt */
0x64657363, 0, 40, /* desc */
0x77747074, 0, 20, /* wtpt */
0x626b7074, 0, 20, /* bkpt */
0x72545243, 0, 14, /* rTRC */
0x67545243, 0, 14, /* gTRC */
0x62545243, 0, 14, /* bTRC */
0x7258595a, 0, 20, /* rXYZ */
0x6758595a, 0, 20, /* gXYZ */
0x6258595a, 0, 20 }; /* bXYZ */
static const unsigned pwhite[] = { 0xf351, 0x10000, 0x116cc };
unsigned pcurve[] = { 0x63757276, 0, 1, 0x1000000 };
gamma_curve (gamm[0], gamm[1], 0, 0);
memcpy (out_cam, rgb_cam, sizeof out_cam);
raw_color |= colors == 1 || document_mode ||
output_color < 1 || output_color > 5;
if (!raw_color) {
oprof = (unsigned *) calloc (phead[0], 1);
merror (oprof, "convert_to_rgb()");
memcpy (oprof, phead, sizeof phead);
if (output_color == 5) oprof[4] = oprof[5];
oprof[0] = 132 + 12*pbody[0];
for (i=0; i < pbody[0]; i++) {
oprof[oprof[0]/4] = i ? (i > 1 ? 0x58595a20 : 0x64657363) : 0x74657874;
pbody[i*3+2] = oprof[0];
oprof[0] += (pbody[i*3+3] + 3) & -4;
}
memcpy (oprof+32, pbody, sizeof pbody);
oprof[pbody[5]/4+2] = strlen(name[output_color-1]) + 1;
memcpy ((char *)oprof+pbody[8]+8, pwhite, sizeof pwhite);
pcurve[3] = (short)(256/gamm[5]+0.5) << 16;
for (i=4; i < 7; i++)
memcpy ((char *)oprof+pbody[i*3+2], pcurve, sizeof pcurve);
pseudoinverse ((double (*)[3])out_rgb[output_color-1], inverse, 3);
for (i=0; i < 3; i++)
for (j=0; j < 3; j++) {
for (num = k=0; k < 3; k++)
num += xyzd50_srgb[i][k] * inverse[j][k];
oprof[pbody[j*3+23]/4+i+2] = num * 0x10000 + 0.5;
}
for (i=0; i < phead[0]/4; i++)
oprof[i] = htonl(oprof[i]);
strcpy ((char *)oprof+pbody[2]+8, "auto-generated by dcraw");
strcpy ((char *)oprof+pbody[5]+12, name[output_color-1]);
for (i=0; i < 3; i++)
for (j=0; j < colors; j++)
for (out_cam[i][j] = k=0; k < 3; k++)
out_cam[i][j] += out_rgb[output_color-1][i][k] * rgb_cam[k][j];
}
if (verbose)
fprintf (stderr, raw_color ? _("Building histograms...\n") :
_("Converting to %s colorspace...\n"), name[output_color-1]);
memset (histogram, 0, sizeof histogram);
if(!raw_color) {
__m128 outcam0= {out_cam[0][0],out_cam[1][0],out_cam[2][0],0},
outcam1= {out_cam[0][1],out_cam[1][1],out_cam[2][1],0},
outcam2= {out_cam[0][2],out_cam[1][2],out_cam[2][2],0},
outcam3= {out_cam[0][3],out_cam[1][3],out_cam[2][3],0};
for (img=image[0]; img < image[width*height]; img+=4) {
__m128 out0;
__m128 vimg0 = {img[0],img[0],img[0],0},
vimg1 = {img[1],img[1],img[1],0},
vimg2 = {img[2],img[2],img[2],0},
vimg3 = {img[3],img[3],img[3],0};
// out[0] = out_cam[0][0] * img[0]
// +out_cam[0][1] * img[1]
// +out_cam[0][2] * img[2]
// +out_cam[0][3] * img[3];
// out[1] = out_cam[1][0] * img[0]
// +out_cam[1][1] * img[1]
// +out_cam[1][2] * img[2]
// +out_cam[1][3] * img[3];
// out[2] = out_cam[2][0] * img[0]
// +out_cam[2][1] * img[1]
// +out_cam[2][2] * img[2]
// +out_cam[2][3] * img[3];
out0 = _mm_add_ps(_mm_add_ps(
_mm_mul_ps(vimg0, outcam0),
_mm_mul_ps(vimg1, outcam1)
), _mm_add_ps(
_mm_mul_ps(vimg2, outcam2),
_mm_mul_ps(vimg3, outcam3)
));
//clip
out0 = _mm_max_ps(_mm_set1_ps(0), _mm_min_ps(_mm_set1_ps(0xffff), _mm_round_ps(out0, _MM_FROUND_TO_ZERO)));
__m128i o = _mm_cvtps_epi32(out0);
o = _mm_packus_epi32(o,_mm_setzero_si128());
memcpy(img, &o, sizeof(short)*3);
FORCC histogram[c][img[c] >> 3]++;
}
} else if (document_mode) {
__m128 test_mm_round_ps(__m128 x) {
// CHECK: define {{.*}} @test_mm_round_ps
// CHECK: @llvm.x86.sse41.round.ps
return _mm_round_ps(x, 2);
}
__m128 test_mm_round_ps(__m128 x) {
// CHECK-LABEL: test_mm_round_ps
// CHECK: call <4 x float> @llvm.x86.sse41.round.ps
// CHECK-ASM: roundps $2, %xmm{{.*}}, %xmm{{.*}}
return _mm_round_ps(x, 2);
}
__m128 test_mm_round_ps(__m128 x) {
// CHECK-LABEL: test_mm_round_ps
// CHECK: call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %{{.*}}, i32 4)
return _mm_round_ps(x, 4);
}