inline void Multiply(const PVector4df &v, PVector4df &out) {
#ifdef __SSE_AVAIL__
__m128 v1 = _mm_load_ps(v._Vec);
__m128 m0 = _mm_load_ps(Row1);
__m128 m1 = _mm_load_ps(Row2);
__m128 m2 = _mm_load_ps(Row3);
__m128 m3 = _mm_load_ps(Row4);
m0 = _mm_mul_ps(m0, v1); //(e11 * v.X) , (e21 * v.Y), (e31 * v.Z), (e41 * v.W)
m1 = _mm_mul_ps(m1, v1); //(e12 * v.X) , (e22 * v.Y), (e32 * v.Z), (e42 * v.W)
m2 = _mm_mul_ps(m2, v1); //(e13 * v.X) , (e23 * v.Y), (e33 * v.Z), (e43 * v.W)
m3 = _mm_mul_ps(m3, v1); //(e14 * v.X) , (e24 * v.Y), (e34 * v.Z), (e44 * v.W)
m0 = _mm_hadd_ps(m0, m1);
m0 = _mm_hadd_ps(m0, m0);
m2 = _mm_hadd_ps(m2, m3);
m2 = _mm_hadd_ps(m2, m2);
_mm_store_ps(out._Vec, _mm_movehl_ps(m2, m0));
#else
out.X = v.X * e11 + v.Y * e21 + v.Z * e31 + v.W * e41;
out.Y = v.X * e12 + v.Y * e22 + v.Z * e32 + v.W * e42;
out.Z = v.X * e13 + v.Y * e23 + v.Z * e33 + v.W * e43;
out.W = v.X * e14 + v.Y * e24 + v.Z * e34 + v.W * e44;
#endif
}
void shz::math::matrix<shz::math::f32, 4, 4>::mul(const shz::math::f32* left, const shz::math::f32*right, shz::math::f32* target) {
shz::math::f32 _ALIGNED(16) left_transposed[shz::math::matrix<shz::math::f32, 4, 4>::size];
matrix<shz::math::f32, 4, 4>::transpose(left, left_transposed);
shz::math::f32 _ALIGNED(16) temp[4];
for(size_t m=0; m<4; ++m){
__m128 right_operand = _mm_load_ps(right);
shz::math::f32* transposed = left_transposed;
// This probably can be further optimized. Needs benchmarking
// Idea: unroll loop and exploit _mm_hadd_ps to sum partials from 2 rows at once
// AVX: with AVX instruction set this loop is trivial
for(size_t n=0; n<4; ++n){
__m128 left_operand = _mm_load_ps(transposed);
__m128 mul_result = _mm_mul_ps(left_operand, right_operand);
__m128 added = _mm_hadd_ps(mul_result, mul_result);
added = _mm_hadd_ps(added, added);
_mm_store_ps(temp, added);
*target = temp[0];
target++;
transposed += 4;
}
right += 4;
}
}
inline PVector4df operator*(const PVector4df &v) {
#ifdef __SSE_AVAIL__
__m128 v1 = _mm_load_ps(v._Vec);
__m128 m0 = _mm_load_ps(Row1);
__m128 m1 = _mm_load_ps(Row2);
__m128 m2 = _mm_load_ps(Row3);
__m128 m3 = _mm_load_ps(Row4);
m0 = _mm_mul_ps(m0, v1); //(e11 * v.X) , (e21 * v.Y), (e31 * v.Z), (e41 * v.W)
m1 = _mm_mul_ps(m1, v1); //(e12 * v.X) , (e22 * v.Y), (e32 * v.Z), (e42 * v.W)
m2 = _mm_mul_ps(m2, v1); //(e13 * v.X) , (e23 * v.Y), (e33 * v.Z), (e43 * v.W)
m3 = _mm_mul_ps(m3, v1); //(e14 * v.X) , (e24 * v.Y), (e34 * v.Z), (e44 * v.W)
m0 = _mm_hadd_ps(m0, m1);
m0 = _mm_hadd_ps(m0, m0);
m2 = _mm_hadd_ps(m2, m3);
m2 = _mm_hadd_ps(m2, m2);
m0 = _mm_movehl_ps(m2, m0);
PVector4df val;
_mm_store_ps(val._Vec, m0);
return val;
#else
return PVector4df(v.X * e11 + v.Y * e21 + v.Z * e31 + v.W * e41,
v.X * e12 + v.Y * e22 + v.Z * e32 + v.W * e42,
v.X * e13 + v.Y * e23 + v.Z * e33 + v.W * e43,
v.X * e14 + v.Y * e24 + v.Z * e34 + v.W * e44);
#endif
}
HW_FORCE_INLINE Vec<N> spreadDot(const Vec<N>& a, const Vec<N>& b) {
__m128 x = _mm_mul_ps(a.xmm, b.xmm);
x = _mm_hadd_ps(x, x);
x = _mm_hadd_ps(x, x);
return Vec<N>(x);
}
static void
mexsoftmax(float* y, float* shift, mwSize m, mwSize n) {
__m128 i1, i2;
__m128 o1, o2;
while (m>0)
{
mwSize curn = n;
float sum = 0.0f;
declconst128(zero, 0.0f);
while (curn>0 && ((unsigned long)(y+curn) & 15) != 0)
{
--curn;
y[curn]=fastexp(y[curn]-*shift);
sum += y[curn];
}
__m128 s1 = _mm_load1_ps (shift);
__m128 sum1 = zero;
while (curn>7) {
i1 = _mm_load_ps (y+curn-4);
i2 = _mm_load_ps (y+curn-8);
i1 = _mm_sub_ps (i1, s1);
i2 = _mm_sub_ps (i2, s1);
o1 = vfastexp(i1);
o2 = vfastexp(i2);
_mm_store_ps (y+curn-4, o1);
sum1 = _mm_add_ps (sum1, o1);
_mm_store_ps (y+curn-8, o2);
sum1 = _mm_add_ps (sum1, o2);
curn-=8;
}
sum1 = _mm_hadd_ps (sum1, sum1);
sum1 = _mm_hadd_ps (sum1, sum1);
sum += _mm_cvtss_f32 (sum1);
while(curn>0) {
--curn;
y[curn]=fastexp(y[curn]-*shift);
sum += y[curn];
}
sum = 1.0f / sum;
ptrdiff_t n_pdt = n;
ptrdiff_t one_pdt = 1;
sscal (&n_pdt, &sum, y, &one_pdt);
++shift;
y+=n;
--m;
}
}
HW_FORCE_INLINE float dot(const Vec<N>& a, const Vec<N>& b) {
__m128 x = _mm_mul_ps(a.xmm, b.xmm);
x = _mm_hadd_ps(x, x);
x = _mm_hadd_ps(x, x);
float tmp;
_mm_store_ss(&tmp, x);
return tmp;
}
static inline __m128 horizontal_add(const __m128 a)
{
#if 0 //!! needs SSE3
const __m128 ftemp = _mm_hadd_ps(a, a);
return _mm_hadd_ps(ftemp, ftemp);
#else
const __m128 ftemp = _mm_add_ps(a, _mm_movehl_ps(a, a)); //a0+a2,a1+a3
return _mm_add_ss(ftemp, _mm_shuffle_ps(ftemp, ftemp, _MM_SHUFFLE(1, 1, 1, 1))); //(a0+a2)+(a1+a3)
#endif
}
HW_FORCE_INLINE Vec<N> normalized(const Vec<N>& a) {
__m128 x = _mm_mul_ps(a.xmm, a.xmm);
x = _mm_hadd_ps(x, x);
x = _mm_hadd_ps(x, x);
x = _mm_rsqrt_ps(x);
x = _mm_mul_ps(a.xmm, x);
return Vec<N>(x);
}
_XOINL float QuaternionSquareSum(const Quaternion& q)
{
#if defined(XO_SSE)
__m128 square = _mm_mul_ps(q.xmm, q.xmm);
square = _mm_hadd_ps(square, square);
square = _mm_hadd_ps(square, square);
return _mm_cvtss_f32(square);
#else
return q.x * q.x + q.y * q.y + q.z * q.z + q.w * q.w;
#endif
}
float length2() const {
Vec3 a = *this;
a.w = 0.0f;
__m128 &D = a.m128;
D = _mm_mul_ps(D, D);
D = _mm_hadd_ps(D, D);
D = _mm_hadd_ps(D, D);
return a.x;
}
inline float hadd(const vector4f& rhs)
{
#if SSE_INSTR_SET >= 3 // SSE3
__m128 tmp0 = _mm_hadd_ps(rhs, rhs);
__m128 tmp1 = _mm_hadd_ps(tmp0, tmp0);
#else
__m128 tmp0 = _mm_add_ps(rhs, _mm_movehl_ps(rhs, rhs));
__m128 tmp1 = _mm_add_ss(tmp0, _mm_shuffle_ps(tmp0, tmp0, 1));
#endif
return _mm_cvtss_f32(tmp1);
}
// use MMX/SSE extensions
void dotprod_rrrf_execute_mmx(dotprod_rrrf _q,
float * _x,
float * _y)
{
// first cut: ...
__m128 v; // input vector
__m128 h; // coefficients vector
__m128 s; // dot product
__m128 sum = _mm_setzero_ps(); // load zeros into sum register
// t = 4*(floor(_n/4))
unsigned int t = (_q->n >> 2) << 2;
//
unsigned int i;
for (i=0; i<t; i+=4) {
// load inputs into register (unaligned)
v = _mm_loadu_ps(&_x[i]);
// load coefficients into register (aligned)
h = _mm_load_ps(&_q->h[i]);
// compute multiplication
s = _mm_mul_ps(v, h);
// parallel addition
sum = _mm_add_ps( sum, s );
}
// aligned output array
float w[4] __attribute__((aligned(16)));
#if HAVE_PMMINTRIN_H
// fold down into single value
__m128 z = _mm_setzero_ps();
sum = _mm_hadd_ps(sum, z);
sum = _mm_hadd_ps(sum, z);
// unload single (lower value)
_mm_store_ss(w, sum);
float total = w[0];
#else
// unload packed array
_mm_store_ps(w, sum);
float total = w[0] + w[1] + w[2] + w[3];
#endif
// cleanup
for (; i<_q->n; i++)
total += _x[i] * _q->h[i];
// set return value
*_y = total;
}
HW_FORCE_INLINE float invLength(const Vec<N>& a) {
__m128 x = _mm_mul_ps(a.xmm, a.xmm);
x = _mm_hadd_ps(x, x);
x = _mm_hadd_ps(x, x);
x = _mm_rsqrt_ss(x);
float tmp;
_mm_store_ss(&tmp, x);
return tmp;
}
inline float dot_product(__m128 a, __m128 b)
{
#if defined(SSE4)
__m128 m = _mm_dp_ps(a, b, 0xff);
return m.m128_f32[0];
#elif defined(SSE3)
__m128 m = _mm_mul_ps(a, b);
m = _mm_hadd_ps(m, m);
m = _mm_hadd_ps(m, m);
return m.m128_f32[0];
#else
__m128 m = _mm_mul_ps(a, b);
return m.m128_f32[0] + m.m128_f32[1] + m.m128_f32[2] + m.m128_f32[3];
#endif
}
// ~~~~~~~~~~~~~~~ Task2
void mulVectorSse(MATRIX_TYPE** matrix, MATRIX_TYPE* vector, MATRIX_TYPE* result, size_t size) {
for (size_t i = 0; i < size; i++) {
__m128 localSum = _mm_setzero_ps();
for (size_t j = 0; j < size; j += 4) {
__m128 tempMatix = _mm_load_ps(&matrix[i][j]);
__m128 tempVector = _mm_load_ps(&vector[j]);
localSum = _mm_add_ps(localSum, _mm_mul_ps(tempMatix, tempVector));
}
localSum = _mm_hadd_ps(localSum, localSum);
localSum = _mm_hadd_ps(localSum, localSum);
_mm_store_ss(&result[i], localSum);
}
}
inline float32x4_t dot(const float32x4_t xmm1, const float32x4_t xmm2)
{
#if TWIST_ARCH & TWIST_ARCH_SSE3_BIT
float32x4_t mul0 = _mm_mul_ps(xmm1, xmm2);
float32x4_t hadd0 = _mm_hadd_ps(mul0, mul0);
float32x4_t hadd1 = _mm_hadd_ps(hadd0, hadd0);
return hadd1;
#else // SSE3
float32x4_t mul0 = _mm_mul_ps(xmm1, xmm2);
float32x4_t swap0 = _mm_shuffle_ps(mul0, mul0, _MM_SHUFFLE(2, 3, 0, 1));
float32x4_t add0 = _mm_add_ps(mul0, swap0);
float32x4_t swap1 = _mm_shuffle_ps(add0, add0, _MM_SHUFFLE(0, 1, 2, 3));
float32x4_t add1 = _mm_add_ps(add0, swap1);
return add1;
#endif // SSE
}
inline vector4f haddp(const vector4f* row)
{
#if SSE_INSTR_SET >= 3 // SSE3
return _mm_hadd_ps(_mm_hadd_ps(row[0], row[1]),
_mm_hadd_ps(row[2], row[3]));
#else
__m128 tmp0 = _mm_unpacklo_ps(row[0], row[1]);
__m128 tmp1 = _mm_unpackhi_ps(row[0], row[1]);
__m128 tmp2 = _mm_unpackhi_ps(row[2], row[3]);
tmp0 = _mm_add_ps(tmp0, tmp1);
tmp1 = _mm_unpacklo_ps(row[2], row[3]);
tmp1 = _mm_add_ps(tmp1, tmp2);
tmp2 = _mm_movehl_ps(tmp1, tmp0);
tmp0 = _mm_movelh_ps(tmp0, tmp1);
return _mm_add_ps(tmp0, tmp2);
#endif
}
float reduction_sum_sse(float *v, int n)
{
int i;
float sum;
__m128 *v4 = (__m128 *)v;
__m128 vsum = _mm_set1_ps(0.0f);
for (i = 0; i < n / 4; i++)
vsum = _mm_add_ps(vsum, v4[i]);
vsum = _mm_hadd_ps(vsum, vsum);
vsum = _mm_hadd_ps(vsum, vsum);
_mm_store_ss(&sum, vsum);
return sum;
}
core::F32_t core::Vector3::dot( core::Vector3 &a, core::Vector3 &b )
{
ALIGNED_16 core::F32_t aVector[] = {a.x, a.y, a.z, 0};
ALIGNED_16 core::F32_t bVector[] = {b.x, b.y, b.z, 0};
__m128 ma;
__m128 mb;
//__m128 mr;
ma = _mm_load_ps(aVector);
mb = _mm_load_ps(bVector);
ALIGNED_16 core::F32_t res[4];
ma = _mm_mul_ps(ma, mb);
ma = _mm_hadd_ps(ma, ma);
ma = _mm_hadd_ps(ma, ma);
_mm_store_ps(&res[0], ma);
return res[0];
}
const Vec3 &normalize() {
w = 0.f;
__m128 D = m128;
D = _mm_mul_ps(D, D);
D = _mm_hadd_ps(D, D);
D = _mm_hadd_ps(D, D);
// 1 iteration of Newton-raphson -- Idea from Intel's Embree.
__m128 r = _mm_rsqrt_ps(D);
r = _mm_add_ps(
_mm_mul_ps(_mm_set_ps(1.5f, 1.5f, 1.5f, 1.5f), r),
_mm_mul_ps(_mm_mul_ps(_mm_mul_ps(D, _mm_set_ps(-0.5f, -0.5f, -0.5f, -0.5f)), r), _mm_mul_ps(r, r)));
m128 = _mm_mul_ps(m128, r);
return *this;
}
float DotProductSIMD(const float* a, const float* b, std::size_t n) {
std::size_t i = 0;
__m128 sum = _mm_setzero_ps();
for(; i < ROUND_DOWN(n, 4); i += 4) {
__m128 x = _mm_loadu_ps(a + i);
__m128 y = _mm_loadu_ps(a + i);
x = _mm_mul_ps(x, y);
sum = _mm_add_ps(x, sum);
}
sum = _mm_hadd_ps(sum, sum);
sum = _mm_hadd_ps(sum, sum);
float product = _mm_cvtss_f32(sum);
for(; i < n; i++) {
product += a[i] * b[i];
}
return product;
}
int
distance_scan_to_map(
map_t * map,
scan_t * scan,
position_t position)
{
int npoints = 0; /* number of points where scan matches map */
int64_t sum = 0; /* sum of map values at those points */
/* Pre-compute sine and cosine of angle for rotation */
double position_theta_radians = radians(position.theta_degrees);
double costheta = cos(position_theta_radians) * map->scale_pixels_per_mm;
double sintheta = sin(position_theta_radians) * map->scale_pixels_per_mm;
/* Pre-compute pixel offset for translation */
double pos_x_pix = position.x_mm * map->scale_pixels_per_mm;
double pos_y_pix = position.y_mm * map->scale_pixels_per_mm;
__m128 sincos128 = _mm_set_ps (costheta, -sintheta, sintheta, costheta);
__m128 posxy128 = _mm_set_ps (pos_x_pix, pos_y_pix, pos_x_pix, pos_y_pix);
int i = 0;
for (i=0; i<scan->npoints; i++)
{
/* Consider only scan points representing obstacles */
if (scan->value[i] == OBSTACLE)
{
/* Compute coordinate pair using SSE */
__m128 xy128 = _mm_set_ps (scan->x_mm[i], scan->y_mm[i], scan->x_mm[i], scan->y_mm[i]);
xy128 = _mm_mul_ps(sincos128, xy128);
xy128 = _mm_hadd_ps(xy128, xy128);
xy128 = _mm_add_ps(xy128, posxy128);
cs_pos_mmx_t pos;
pos.mmx = _mm_cvtps_pi32(xy128);
/* Extract coordinates */
int x = pos.pos.x;
int y = pos.pos.y;
/* Empty the multimedia state to avoid floating-point errors later */
_mm_empty();
/* Add point if in map bounds */
if (x >= 0 && x < map->size_pixels && y >= 0 && y < map->size_pixels)
{
sum += map->pixels[y * map->size_pixels + x];
npoints++;
}
}
}
/* Return sum scaled by number of points, or -1 if none */
return npoints ? (int)(sum * 1024 / npoints) : -1;
}
v4f step_t::operator () (float t) const
{
// Evaluate the polynomial f by Estrin's method. Return
// (0 0 0 0) if t < t0,
// (f f f f) if t0 <= t < t1,
// (1 1 1 1) if t > t1.
v4f c4 = load4f (c);
v4f one = { 1.0f, 1.0f, 1.0f, 1.0f };
v4f tttt = _mm_set1_ps (t); // t t t t
v4f tt = _mm_unpacklo_ps (one, tttt); // 1 t 1 t
v4f f0 = c4 * tt; // c0 c1*t c2 c3*t
v4f ha = _mm_hadd_ps (f0, f0) * tt * tt;
v4f f = _mm_hadd_ps (ha, ha); // f f f f
v4f f1 = _mm_unpacklo_ps (f, one); // f 1 f 1
v4f tx = load4f (T); // t0 t1 t1 inf
v4f lo = _mm_movelh_ps (tx, tx); // t0 t1 t0 t1
v4f hi = _mm_movehl_ps (tx, tx); // t1 inf t1 inf
v4f sel = _mm_and_ps (_mm_cmpge_ps (tttt, lo), _mm_cmplt_ps (tttt, hi));
v4f val = _mm_and_ps (sel, f1); // f? 1? f? 1?
return _mm_hadd_ps (val, val);
}
// ~~~~~~~~~~~~~~~ Task3
void mulMatrix_IJKAlgSse(MATRIX_TYPE** first, MATRIX_TYPE** second, MATRIX_TYPE** result, size_t size) {
transpose(second, size);
for (size_t i = 0; i < size; i++) {
for (size_t j = 0; j < size; j++) {
__m128 temp = _mm_setzero_ps();
for (size_t k = 0; k < size; k += 4) {
__m128 tempFirst = _mm_load_ps(&first[i][k]);
__m128 tempSecond = _mm_load_ps(&second[j][k]);
temp = _mm_add_ps(temp, _mm_mul_ps(tempFirst, tempSecond));
}
temp = _mm_hadd_ps(temp, temp);
temp = _mm_hadd_ps(temp, temp);
_mm_store_ss(&result[i][j], temp);
}
}
transpose(second, size);
}
SSH float* ssh_mtx_mtx(const float* m1, const float* m2)
{
static float flt[16];
__m128 _m[4];
_m[0] = _mm_set_ps(m2[0], m2[4], m2[8], m2[12]);
_m[1] = _mm_set_ps(m2[1], m2[5], m2[9], m2[13]);
_m[2] = _mm_set_ps(m2[2], m2[6], m2[10], m2[14]);
_m[3] = _mm_set_ps(m2[3], m2[7], m2[11], m2[15]);
for(ssh_u i = 0; i < 4; i++)
{
for(ssh_u j = 0; j < 4; j++)
{
__m128 _tmp(_mm_mul_ps(*(__m128*)&m1[i * 4], _m[j]));
_tmp = _mm_hadd_ps(_tmp, _tmp);
flt[i * 4 + j] = _mm_hadd_ps(_tmp, _tmp).m128_f32[0];
}
}
return flt;
}
/** transform vector by rigid transform */
inline Matrix<float, 4, 1> operator * (const RigidTransform<float>& mat, const Matrix<float, 4, 1>& vec)
{
#ifdef SIMPLE_GL_USE_SSE4
__m128 res;
__m128 dotProd;
res = _mm_dp_ps(mat[0].m128, vec.m128, 0xEE);\
dotProd = _mm_dp_ps(mat[1].m128, vec.m128, 0xEE);\
res = _mm_blend_ps( res, dotProd, _MM_SHUFFLE(0, 1, 1, 1) );\
dotProd = _mm_dp_ps(mat[2].m128, vec.m128, 0xEE);\
res = _mm_blend_ps( res, dotProd, _MM_SHUFFLE(0, 0, 1, 1) );\
dotProd = _mm_dp_ps(mat[3].m128, vec.m128, 0xEE);\
res = _mm_blend_ps( res, dotProd, _MM_SHUFFLE(0, 0, 0, 1) );
return Matrix<float, 4, 1>(res);
#elif defined(SIMPLE_GL_USE_SSE3)
__m128 res;
__m128 dotProd0 = _mm_mul_ps(mat[0].m128, vec.m128);
dotProd0 = _mm_hadd_ps(dotProd0, dotProd0);
dotProd0 = _mm_hadd_ps(dotProd0, dotProd0);
__m128 dotProd1 = _mm_mul_ps(mat[1].m128, vec.m128);
dotProd1 = _mm_hadd_ps(dotProd1, dotProd1);
dotProd1 = _mm_hadd_ps(dotProd1, dotProd1);
__m128 dotProd2 = _mm_mul_ps(mat[2].m128, vec.m128);
dotProd2 = _mm_hadd_ps(dotProd2, dotProd2);
dotProd2 = _mm_hadd_ps(dotProd2, dotProd2);
__m128 dotProd3 = _mm_mul_ps(mat[3].m128, vec.m128);
dotProd3 = _mm_hadd_ps(dotProd3, dotProd3);
dotProd3 = _mm_hadd_ps(dotProd3, dotProd3);
__m128 vec01 = _mm_unpacklo_ps(dotProd0, dotProd1);
__m128 vec23 = _mm_unpackhi_ps(dotProd2, dotProd3);
res = _mm_movelh_ps(vec01, vec23);
return Matrix<float, 4, 1>(res);
#else // SSE2
// TODO: Think about good sse optimization
Matrix<float, 4, 1> res;
res[0] = mat[0][0] * res[0] + mat[0][1] * res[1] + mat[0][2] * res[2] + mat[0][3] * res[3];
res[1] = mat[1][0] * res[0] + mat[1][1] * res[1] + mat[1][2] * res[2] + mat[1][3] * res[3];
res[2] = mat[2][0] * res[0] + mat[2][1] * res[1] + mat[2][2] * res[2] + mat[2][3] * res[3];
res[3] = mat[3][0] * res[0] + mat[3][1] * res[1] + mat[3][2] * res[2] + mat[3][3] * res[3];
return res;
#endif
}
float vsum(const float *a, int _n)
{
float sum;
int n = _n - _n%3;
__m128 vsum = _mm_set1_ps(0.0f);
assert((n & 3) == 0);
assert(((uintptr_t)a & 15) == 0);
for (int i = 0; i < n; i += 4)
{
__m128 v = _mm_load_ps(&a[i]);
vsum = _mm_add_ps(vsum, v);
}
vsum = _mm_hadd_ps(vsum, vsum);
vsum = _mm_hadd_ps(vsum, vsum);
_mm_store_ss(&sum, vsum);
for(int i=n;i<_n;i++){
sum += a[i];
}
return sum;
}
void lx_matmul_sse3_aligned(const float *in_A,
const float *in_x,
const float *out_y,
LXInteger n)
{
__m128 A0 = _mm_load_ps((const float *)(in_A + 0));
__m128 A1 = _mm_load_ps((const float *)(in_A + 4));
__m128 A2 = _mm_load_ps((const float *)(in_A + 8));
__m128 A3 = _mm_load_ps((const float *)(in_A + 12));
for (LXInteger i = 0; i < n; i++) {
__m128 x = _mm_load_ps((const float*)(in_x + i*4));
__m128 m0 = _mm_mul_ps(A0, x);
__m128 m1 = _mm_mul_ps(A1, x);
__m128 m2 = _mm_mul_ps(A2, x);
__m128 m3 = _mm_mul_ps(A3, x);
__m128 sum_01 = _mm_hadd_ps(m0, m1);
__m128 sum_23 = _mm_hadd_ps(m2, m3);
__m128 result = _mm_hadd_ps(sum_01, sum_23);
_mm_store_ps((float*)(out_y + i*4), result);
}
}
static void
matvec_sse()
{
/* Assume that the data size is an even multiple of the 128 bit
* SSE vectors (i.e. 4 floats) */
assert(!(SIZE & 0x3));
/* TASK: Implement your SSE version of the matrix-vector
* multiplication here.
*/
/* HINT: You might find at least the following instructions
* useful:
* - _mm_setzero_ps
* - _mm_load_ps
* - _mm_hadd_ps
* - _mm_cvtss_f32
*
* HINT: You can create the sum of all elements in a vector
* using two hadd instructions.
*/
__m128 dummy=_mm_setzero_ps();
for(int i=0;i<SIZE;++i){
__m128 temp=_mm_setzero_ps();
for(int j=0;j<SIZE;j+=4){
__m128 mm_vec_b=_mm_load_ps((__m128*)(vec_b+j));
__m128 mm_matr=_mm_load_ps((__m128*)(mat_a+MINDEX(i,j)));
__m128 out=_mm_mul_ps(mm_vec_b,mm_matr);
temp=_mm_add_ps(temp,out);
// vec_c[i]+=_mm_cvtss_f32(_mm_dp_ps(mm_matr,mm_vec_b,0xf1));
}
__m128 res=_mm_hadd_ps(_mm_hadd_ps(temp,dummy),dummy);
vec_c[i]=_mm_cvtss_f32(res);
}
}
int main(){
typedef union{
__m128 m128;
float flt[4];
} m128f;
__m128 x = {1.0,2.0,3.0,4.0};
__m128 y = {10.0,20.0,30.0,40.0};
m128f s,h;
s.m128=haddps(x,y);
h.m128=_mm_hadd_ps(x,y);
printf("Software hadd: %f %f %f %f\n",s.flt[0],s.flt[1],s.flt[2],s.flt[3]);
printf("Hardware hadd: %f %f %f %f\n",h.flt[0],h.flt[1],h.flt[2],h.flt[3]);
return;
}