/**
* @brief Computes an array of the exponentials in the transport equation,
* \f$ exp(-\frac{\Sigma_t * l}{sin(\theta)}) \f$, for each energy group
* and polar angle for a given Track segment.
* @param curr_segment pointer to the Track segment of interest
* @param exponentials the array to store the exponential values
*/
void VectorizedSolver::computeExponentials(segment* curr_segment,
FP_PRECISION* exponentials) {
FP_PRECISION length = curr_segment->_length;
FP_PRECISION* sigma_t = curr_segment->_material->getSigmaT();
/* Evaluate the exponentials using the linear interpolation table */
if (_interpolate_exponential) {
FP_PRECISION tau;
int index;
for (int e=0; e < _num_groups; e++) {
for (int p=0; p < _num_polar; p++) {
tau = sigma_t[e] * length;
index = round_to_int(tau * _inverse_exp_table_spacing);
index *= _two_times_num_polar;
exponentials(p,e) = (1. - (_exp_table[index+2 * p] * tau +
_exp_table[index + 2 * p +1]));
}
}
}
/* Evalute the exponentials using the intrinsic exp(...) function */
else {
int tid = omp_get_thread_num();
FP_PRECISION* sinthetas = _quad->getSinThetas();
FP_PRECISION* taus = &_thread_taus[tid*_polar_times_groups];
/* Initialize the tau argument for the exponentials */
for (int p=0; p < _num_polar; p++) {
for (int v=0; v < _num_vector_lengths; v++) {
#pragma simd vectorlength(VEC_LENGTH)
for (int e=v*VEC_LENGTH; e < (v+1)*VEC_LENGTH; e++)
taus(p,e) = -sigma_t[e] * length / sinthetas[p];
}
}
/* Evaluate the negative of the exponentials using Intel's MKL */
#ifdef SINGLE
vsExp(_polar_times_groups, taus, exponentials);
#else
vdExp(_polar_times_groups, taus, exponentials);
#endif
/* Compute one minus the exponentials */
for (int p=0; p < _num_polar; p++) {
for (int v=0; v < _num_vector_lengths; v++) {
#pragma simd vectorlength(VEC_LENGTH)
for (int e=v*VEC_LENGTH; e < (v+1)*VEC_LENGTH; e++)
exponentials(p,e) = 1.0 - exponentials(p,e);
}
}
}
}
/**
* @brief Computes an array of the exponentials in the transport equation,
* \f$ exp(-\frac{\Sigma_t * l}{sin(\theta)}) \f$, for each energy group
* and polar angle for a given Track segment.
* @param curr_segment pointer to the Track segment of interest
* @param exponentials the array to store the exponential values
*/
void VectorizedSolver::computeExponentials(segment* curr_segment,
FP_PRECISION* exponentials) {
FP_PRECISION length = curr_segment->_length;
FP_PRECISION* sigma_t = curr_segment->_material->getSigmaT();
/* Evaluate the exponentials using the linear interpolation table */
if (_exp_evaluator->isUsingInterpolation()) {
FP_PRECISION tau;
for (int e=0; e < _num_groups; e++) {
tau = length * sigma_t[e];
for (int p=0; p < _num_polar; p++)
exponentials(p,e) = _exp_evaluator->computeExponential(tau, p);
}
}
/* Evalute the exponentials using the intrinsic exp(...) function */
else {
int tid = omp_get_thread_num();
FP_PRECISION* sin_thetas = _polar_quad->getSinThetas();
FP_PRECISION* taus = &_thread_taus[tid*_polar_times_groups];
/* Initialize the tau argument for the exponentials */
for (int p=0; p < _num_polar; p++) {
for (int v=0; v < _num_vector_lengths; v++) {
#pragma simd vectorlength(VEC_LENGTH)
for (int e=v*VEC_LENGTH; e < (v+1)*VEC_LENGTH; e++)
taus(p,e) = -sigma_t[e] * length;
#pragma simd vectorlength(VEC_LENGTH)
for (int e=v*VEC_LENGTH; e < (v+1)*VEC_LENGTH; e++)
taus(p,e) /= sin_thetas[p];
}
}
/* Evaluate the negative of the exponentials using Intel's MKL */
#ifdef SINGLE
vsExp(_polar_times_groups, taus, exponentials);
#else
vdExp(_polar_times_groups, taus, exponentials);
#endif
/* Compute one minus the exponentials */
for (int p=0; p < _num_polar; p++) {
for (int v=0; v < _num_vector_lengths; v++) {
#pragma simd vectorlength(VEC_LENGTH)
for (int e=v*VEC_LENGTH; e < (v+1)*VEC_LENGTH; e++)
exponentials(p,e) = 1.0 - exponentials(p,e);
}
}
}
}
/*
* Class: com_intel_analytics_bigdl_mkl_MKL
* Method: vsExp
* Signature: (I[FI[FI)V
*/
JNIEXPORT void JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_vsExp
(JNIEnv * env, jclass cls, jint n, jfloatArray a, jint aOffset, jfloatArray y,
jint yOffset) {
jfloat * jni_a = (*env)->GetPrimitiveArrayCritical(env, a, JNI_FALSE);
jfloat * jni_y = (*env)->GetPrimitiveArrayCritical(env, y, JNI_FALSE);
vsExp( n, jni_a + aOffset, jni_y + yOffset);
(*env)->ReleasePrimitiveArrayCritical(env, y, jni_y, 0);
(*env)->ReleasePrimitiveArrayCritical(env, a, jni_a, 0);
}
// To force linking in all needed Intel routines
// See compileCX for details (linking against .a)
void Dummy(void)
{
vsAdd();
vsSub();
vsDiv();
vsSqr();
vsMul();
vsAbs();
vsInv();
vsSin();
vsCos();
vsSinCos();
vsTan();
vsAsin();
vsAcos();
vsAtan();
vsAtan2();
vsSinh();
vsCosh();
vsTanh();
vsAsinh();
vsAcosh();
vsAtanh();
vsPow();
vsPowx();
vsSqrt();
vsCbrt();
vsInvSqrt();
vsInvCbrt();
vsHypot();
vsFloor();
vsCeil();
vsRound();
vsTrunc();
vsRint();
vsNearbyInt();
vsModf();
vsExp();
vsLn();
vsLog10();
vsErf();
vsErfc();
vsErfInv();
}
void caffe_exp<float>(const int n, const float* a, float* y) {
vsExp(n, a, y);
}
