void DecoderBinaural::process(const double* inputs, double* outputs)
{
m_decoder->process(inputs, m_channels_vector_double);
--m_index;
m_channels_inputs_left[0][m_index] = m_channels_vector_double[0];
outputs[1] = outputs[0] = cblas_sdot(m_impulses_size, m_channels_inputs_left[0]+m_index, 1, m_impulses_vector[0], 1);
for(int i = 1; i < m_number_of_virtual_channels; i++)
{
m_channels_inputs_left[i][m_index] = m_channels_vector_double[i];
m_channels_inputs_right[i][m_index] = m_channels_vector_double[i];
outputs[0] += cblas_sdot(m_impulses_size, m_channels_inputs_left[i]+m_index, 1, m_impulses_vector[m_number_of_virtual_channels - i], 1);
outputs[1] += cblas_sdot(m_impulses_size, m_channels_inputs_right[i]+m_index, 1, m_impulses_vector[i], 1);
}
if(m_index <= 0)
{
m_index = m_impulses_size;
cblas_scopy(m_impulses_size, m_channels_inputs_left[0], 1, m_channels_inputs_left[0]+m_impulses_size, 1);
for(int i = 1; i < m_number_of_virtual_channels; i++)
{
cblas_scopy(m_impulses_size, m_channels_inputs_left[i], 1, m_channels_inputs_left[i]+m_impulses_size, 1);
cblas_scopy(m_impulses_size, m_channels_inputs_right[i], 1, m_channels_inputs_right[i]+m_impulses_size, 1);
}
}
}
void preprocess1(int panelSz, int D, float *XX, float*X, float*Y, float lamda, float*Z, float*B){
// panelSz: the number of points to process in each round
int i,j;
// step 1: compute all X[i]'*X[j] (i>j)
for (i=panelSz-1;i>0;i--)
for (j=i-1;j>=0;j--){
XX[i*panelSz+j] = cblas_sdot(D, &(X[i*D]), 1, &(X[j*D]), 1);
// printf("XX[%d]=%8.4f, X[%d]=%8.4f, X[%d]=%8.4f\n", i*panelSz+j, XX[i*panelSz+j], i*D, X[i*D], j*D, X[j*D]);
}
// step 2: compute all Z vectors
// Z0=lamda*X[0], B=lamda*X[0]*Y[0]
cblas_scopy(D, X, 1, Z, 1);
cblas_sscal(D, lamda, Z, 1);
float alpha=lamda*Y[0];
cblas_scopy(D, X, 1, B, 1);
cblas_sscal(D, alpha, B, 1);
for (i=1; i<panelSz;i++){
cblas_scopy(D, &(X[i*D]), 1, &(Z[i*D]),1);
// Z[i] = lamda*(X[i] - sum_{j<i} XX[i,j]*Z[j]);
for (j=0;j<i;j++){
cblas_saxpy(D, -1*XX[i*panelSz+j], &(Z[j*D]), 1, &(Z[i*D]), 1);
}
cblas_sscal(D, lamda, &(Z[i*D]), 1);
// B = lamda*(Y[i] - X[i]*B) X[i] + B;
float alpha = lamda*(Y[i]-cblas_sdot(D, &(X[i*D]), 1, B, 1));
cblas_saxpy(D, alpha, &(X[i*D]), 1, B, 1);
}
}
void replace_nans_avgs(int vec_blocksize, type_precision* vec,
int rows , int cols, list<long int>* ar_nan_idxs)
{
type_precision* ones = new type_precision[rows];
for (int i = 0; i < rows; i++)
{
ones[i] = 1.0;
}
for (int i = 0; i < vec_blocksize * cols; i++)
{
type_precision sum = cblas_sdot(rows, ones, 1, &vec[i * rows], 1);
type_precision avg = sum / (rows - ar_nan_idxs[i].size());
for (list<long int>::iterator it = ar_nan_idxs[i].begin();
it != ar_nan_idxs[i].end();
it++)
{
int idx = i * rows + (*it);
vec[idx] = avg;
}
}
delete []ones;
}
real THBlas_(dot)(long n, real *x, long incx, real *y, long incy)
{
if(n == 1)
{
incx = 1;
incy = 1;
}
#if defined(USE_BLAS) && (defined(TH_REAL_IS_DOUBLE) || defined(TH_REAL_IS_FLOAT))
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if defined(TH_REAL_IS_DOUBLE)
return cblas_ddot(i_n, x, i_incx, y, i_incy);
#else
return cblas_sdot(i_n, x, i_incx, y, i_incy);
#endif
}
#endif
{
long i;
real sum = 0;
for(i = 0; i < n; i++)
sum += x[i*incx]*y[i*incy];
return sum;
}
}
/* Trains a network by presenting an example and
* adjusts the weights by stochastic gradient
* descent to reduce a squared hinge loss
*/
void train(nnet_t* n, sparse_t* v, int target){
int i;
/* Forward pass */
cblas_scopy(n->hidden,n->b1,1,n->a1,1);
for(i=0; i<v->nz; i++){
cblas_saxpy(n->hidden, v->x[i], n->W1[v->idx[i]], 1, n->a1, 1);
}
activation(n->a1,n->x1,n->g1,n->hidden);
n->a2 = n->b2 + cblas_sdot(n->hidden, n->W2, 1, n->x1, 1);
activation(&n->a2,&n->x2,&n->g2,1);
if(target*n->x2 > 1)
/* Hinge loss, no error -> no need to backpropagate */
return;
/* Backward pass */
n->d2 = (target-n->x2)*n->g2;
cblas_scopy(n->hidden,n->W2,1,n->d1,1);
for(i=0; i<n->hidden; i++)
n->d1[i] *= n->d2*n->g1[i];
n->b2 += n->eta*n->d2;
cblas_saxpy(n->hidden, n->eta*n->d2, n->x1, 1, n->W2, 1);
cblas_saxpy(n->hidden, n->eta, n->d1, 1, n->b1, 1);
/* Sparse inputs imply sparse gradients.
* This update saves a lot of computation
* compared to general purpose neural net
* implementations.
*/
for(i=0; i<v->nz; i++){
cblas_saxpy(n->hidden, n->eta*v->x[i], n->d1, 1, n->W1[v->idx[i]], 1);
}
}
real THBlas_dot(long size, real *x, long xStride, real *y, long yStride)
{
if(size == 1)
{
xStride = 1;
yStride = 1;
}
#if USE_CBLAS
if( (size < INT_MAX) && (xStride < INT_MAX) && (yStride < INT_MAX) )
{
#ifdef USE_DOUBLE
return cblas_ddot(size, x, xStride, y, yStride);
#else
return cblas_sdot(size, x, xStride, y, yStride);
#endif
}
#endif
{
long i;
real sum = 0;
for(i = 0; i < size; i++)
sum += x[i*xStride]*y[i*yStride];
return sum;
}
}
value_t<A> dot(const A& a, const B& b) {
if constexpr (all_dma<A, B>) {
a.ensure_cpu_up_to_date();
b.ensure_cpu_up_to_date();
if constexpr (all_single_precision<A, B>) {
return cblas_sdot(etl::size(a), a.memory_start(), 1, b.memory_start(), 1);
} else {
return cblas_ddot(etl::size(a), a.memory_start(), 1, b.memory_start(), 1);
JNIEXPORT jfloat JNICALL Java_uncomplicate_neanderthal_CBLAS_sdot
(JNIEnv *env, jclass clazz, jint N,
jobject X, jint offsetX, jint incX,
jobject Y, jint offsetY, jint incY) {
float *cX = (float *) (*env)->GetDirectBufferAddress(env, X);
float *cY = (float *) (*env)->GetDirectBufferAddress(env, Y);
return cblas_sdot(N, cX + offsetX, incX, cY + offsetY, incY);
};
void Vector::processVelocity(const float* inputs, float* outputs)
{
float veclocitySum = 0.f, velocityAbscissa = 0.f, velocityOrdinate = 0.f;
veclocitySum = 0.;
for(int i = 0; i < m_number_of_channels; i++)
veclocitySum += inputs[i];
velocityAbscissa = cblas_sdot(m_number_of_channels, inputs, 1, m_channels_abscissa_float, 1);
velocityOrdinate = cblas_sdot(m_number_of_channels, inputs, 1, m_channels_ordinate_float, 1);
if(veclocitySum)
{
outputs[0] = velocityAbscissa / veclocitySum;
outputs[1] = velocityOrdinate / veclocitySum;
}
else
{
outputs[0] = 0.;
outputs[1] = 0.;
}
}
float HostVector<float>::Dot(const BaseVector<float> &x) const {
assert(&x != NULL);
const HostVector<float> *cast_x = dynamic_cast<const HostVector<float>*> (&x);
assert(cast_x != NULL);
assert(this->size_ == cast_x->size_);
return cblas_sdot(this->size_, this->vec_, 1, cast_x->vec_, 1);
}
/*
* Class: com_intel_analytics_bigdl_mkl_MKL
* Method: vsdot
* Signature: (I[FII[FII)V
*/
JNIEXPORT float JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_vsdot
(JNIEnv * env, jclass cls, jint n, jfloatArray x, jint xOffset, jint incx,
jfloatArray y, jint yOffset, jint incy) {
jfloat * jni_x = (*env)->GetPrimitiveArrayCritical(env, x, JNI_FALSE);
jfloat * jni_y = (*env)->GetPrimitiveArrayCritical(env, y, JNI_FALSE);
float res = cblas_sdot(n, jni_x + xOffset, incx, jni_y + yOffset, incy);
(*env)->ReleasePrimitiveArrayCritical(env, y, jni_y, 0);
(*env)->ReleasePrimitiveArrayCritical(env, x, jni_x, 0);
return res;
}
static void dotprod(const RView& X, const RView& Y,
T& result)
{
#ifdef MKL
const float * x = X.data();
const float * y = Y.data();
result = cblas_sdot(DIM_N, x, 1, y, 1);
#else
for (int i=0; i<DIM_N; ++i)
result += X[i] * Y[i];
#endif
}
/* Given an input vector v, compute the output of the network. */
float value(nnet_t* n, sparse_t* v){
int i;
cblas_scopy(n->hidden,n->b1,1,n->a1,1);
for(i=0; i<v->nz; i++){
cblas_saxpy(n->hidden, v->x[i], n->W1[v->idx[i]], 1, n->a1, 1);
}
activation(n->a1,n->x1,n->g1,n->hidden);
n->a2 = n->b2;
n->a2 += cblas_sdot(n->hidden, n->W2, 1, n->x1, 1);
activation(&n->a2,&n->x2,&n->g2,1);
return n->x2;
}
void bli_sdot( conj_t conj, int n, float* x, int incx, float* y, int incy, float* rho )
{
#ifdef BLIS_ENABLE_CBLAS_INTERFACES
*rho = cblas_sdot( n,
x, incx,
y, incy );
#else
*rho = F77_sdot( &n,
x, &incx,
y, &incy );
#endif
}
JNIEXPORT jfloat JNICALL Java_edu_berkeley_bid_CBLAS_sdotxx
(JNIEnv * env, jobject calling_obj, jint N, jfloatArray jX, jint startX, jfloatArray jY, jint startY){
jfloat * X = (*env)->GetPrimitiveArrayCritical(env, jX, JNI_FALSE);
jfloat * Y = (*env)->GetPrimitiveArrayCritical(env, jY, JNI_FALSE);
jfloat returnValue;
returnValue = cblas_sdot(N, X+startX, 1, Y+startY, 1);
(*env)->ReleasePrimitiveArrayCritical(env, jY, Y, 0);
(*env)->ReleasePrimitiveArrayCritical(env, jX, X, 0);
return returnValue;
}
value_t<A> sum(const A& a) {
if constexpr (is_dma<A>) {
etl::dyn_vector<value_t<A>> ones(etl::size(a));
ones = value_t<A>(1);
a.ensure_cpu_up_to_date();
[[maybe_unused]] const auto* m_a = a.memory_start();
[[maybe_unused]] const auto* m_b = ones.memory_start();
if constexpr (is_single_precision<A>) {
return cblas_sdot(etl::size(a), m_a, 1, m_b, 1);
} else if constexpr (is_double_precision<A>) {
void Vector::processEnergy(const float* inputs, float* outputs)
{
float energySum = 0.f, energyAbscissa = 0.f, energyOrdinate = 0.f;
cblas_scopy(m_number_of_channels, inputs, 1, m_channels_float, 1);
for(int i = 0; i < m_number_of_channels; i++)
m_channels_float[i] *= m_channels_float[i];
energySum = cblas_sasum(m_number_of_channels, m_channels_float, 1);
energyAbscissa = cblas_sdot(m_number_of_channels, m_channels_float, 1, m_channels_abscissa_float, 1);
energyOrdinate = cblas_sdot(m_number_of_channels, m_channels_float, 1, m_channels_ordinate_float, 1);
if(energySum)
{
outputs[0] = energyAbscissa / energySum;
outputs[1] = energyOrdinate / energySum;
}
else
{
outputs[0] = 0.;
outputs[1] = 0.;
}
}
inline float
dot_call( dot_params< float >& p )
{
//std::cout << "calling blas sdot (single precision) " << std::endl;
float vvi = cblas_sdot(
p.n,
p.x,
p.inc_x,
p.y,
p.inc_y
);
return vvi;
}
static void
FLOAT_dot(void *a, intp stridea, void *b, intp strideb, void *res,
intp n, void *tmp)
{
register int na = stridea / sizeof(float);
register int nb = strideb / sizeof(float);
if ((sizeof(float) * na == stridea) &&
(sizeof(float) * nb == strideb) &&
(na >= 0) && (nb >= 0))
*((float *)res) = cblas_sdot((int)n, (float *)a, na, (float *)b, nb);
else
oldFunctions[PyArray_FLOAT](a, stridea, b, strideb, res, n, tmp);
}
JNIEXPORT void JNICALL Java_edu_berkeley_bid_CBLAS_sdotm
(JNIEnv * env, jobject calling_obj, jint nrows, jint ncols, jfloatArray jX, jint ldx, jfloatArray jY, jint ldy, jfloatArray jZ){
jfloat * X = (*env)->GetPrimitiveArrayCritical(env, jX, JNI_FALSE);
jfloat * Y = (*env)->GetPrimitiveArrayCritical(env, jY, JNI_FALSE);
jfloat * Z = (*env)->GetPrimitiveArrayCritical(env, jZ, JNI_FALSE);
int i, j;
for (i = 0; i < ncols; i++) {
Z[i] = cblas_sdot(nrows, X+i*ldx, 1, Y+i*ldy, 1);
}
(*env)->ReleasePrimitiveArrayCritical(env, jZ, Z, 0);
(*env)->ReleasePrimitiveArrayCritical(env, jY, Y, 0);
(*env)->ReleasePrimitiveArrayCritical(env, jX, X, 0);
}
bool dot(REAL &inner, const RealVector &x, const RealVector &y) {
bool flag = true;
UINT N; INT incX, incY;
N = x.size;
incX = 1;
incY = 1;
if (NULL == &x || NULL == &y) {
flag = false;
goto end;
}
if (x.size != y.size) {
flag = false;
goto end;
}
inner = cblas_sdot(N, x.M, incX, y.M, incY);
end:
return flag;
}
inline static void
magma_slarfxsym_v2(magma_int_t n,
float *A, magma_int_t lda,
float *V, float *TAU,
float *work)
{
/*
WORK (workspace) float real array, dimension N
*/
magma_int_t ione = 1;
float dtmp;
float c_zero = MAGMA_S_ZERO;
float c_neg_one= MAGMA_S_NEG_ONE;
float c_half = MAGMA_S_HALF;
/* X = AVtau */
blasf77_ssymv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione);
/* compute dtmp= X'*V */
#if defined(PRECISION_z) || defined(PRECISION_c)
dtmp = c_zero;
for (magma_int_t j = 0; j < n ; j++)
dtmp = dtmp + MAGMA_S_CNJG(work[j]) * V[j];
//cblas_sdot_sub(n, work, ione, V, ione, &dtmp);
#else
dtmp = cblas_sdot(n, work, ione, V, ione);
#endif
/* compute 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * c_half * (*TAU);
/* compute W=X-1/2VX'Vt = X - dtmp*V */
blasf77_saxpy(&n, &dtmp, V, &ione, work, &ione);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_ssyr2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda);
}
typename TypeTraits<typename Vector::ScalarType>::magnitude_type
dot(const Vector& x,
const Vector& y)
{
int n = x.coefs.size();
#ifdef MINIFE_DEBUG
if (y.local_size < n) {
std::cerr << "miniFE::dot ERROR, y must be at least as long as x."<<std::endl;
n = y.local_size;
}
#endif
typedef typename Vector::ScalarType Scalar;
typedef typename TypeTraits<typename Vector::ScalarType>::magnitude_type magnitude;
const Scalar* xcoefs = &x.coefs[0];
const Scalar* ycoefs = &y.coefs[0];
#if defined(MINIFE_MKL_DOUBLE)
magnitude result = cblas_ddot(
#elif defined(MINIFE_MKL_FLOAT)
magnitude result = cblas_sdot(
#else
#error Unknown MINIFE_SCALAR type.
#endif
(MKL_INT) n,
xcoefs,
(MKL_INT) 1,
ycoefs,
(MKL_INT) 1);
#ifdef HAVE_MPI
magnitude local_dot = result, global_dot = 0;
MPI_Datatype mpi_dtype = TypeTraits<magnitude>::mpi_type();
MPI_Allreduce(&local_dot, &global_dot, 1, mpi_dtype, MPI_SUM, MPI_COMM_WORLD);
return global_dot;
#else
return result;
#endif
}
//#define NUMERIC_PartialDeriv_CALC
void BuildMatrix(BkgFitMatrixPacker *fit,bool accum, bool debug)
{
#if 1
(void) debug;
fit->BuildMatrix(accum);
#else
mat_assembly_instruction *pinst = fit->instList;
int lineInc=0;
// build JTJ and RHS matricies
for (int i=0; i < fit->nInstr; i++)
{
double sum=0.0;
for (int j=0; j < pinst->cnt; j++)
sum += cblas_sdot(pinst->si[j].len,pinst->si[j].src1,1,pinst->si[j].src2,1);
if (accum)
* (pinst->dst) += sum;
else
* (pinst->dst) = sum;
if (debug)
{
char *src1Name = findName(pinst->si[0].src1);
char *src2Name = findName(pinst->si[0].src2);
printf("%d(%s--%s)(%lf) ",i,src1Name,src2Name,* (pinst->dst));
if (lineInc++ > 6)
{
printf("\n ");
lineInc = 0;
}
}
pinst++;
}
#endif
}
void
test_dot (void) {
const double flteps = 1e-4, dbleps = 1e-6;
{
int N = 1;
float alpha = 0.0f;
float X[] = { 0.733f };
float Y[] = { 0.825f };
int incX = 1;
int incY = -1;
float expected = 0.604725f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 1)");
};
{
int N = 1;
float alpha = 0.1f;
float X[] = { 0.733f };
float Y[] = { 0.825f };
int incX = 1;
int incY = -1;
float expected = 0.704725f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 2)");
};
{
int N = 1;
float alpha = 1.0f;
float X[] = { 0.733f };
float Y[] = { 0.825f };
int incX = 1;
int incY = -1;
float expected = 1.604725f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 3)");
};
{
int N = 1;
float alpha = 0.0f;
float X[] = { -0.812f };
float Y[] = { -0.667f };
int incX = -1;
int incY = 1;
float expected = 0.541604f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 4)");
};
{
int N = 1;
float alpha = 0.1f;
float X[] = { -0.812f };
float Y[] = { -0.667f };
int incX = -1;
int incY = 1;
float expected = 0.641604f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 5)");
};
{
int N = 1;
float alpha = 1.0f;
float X[] = { -0.812f };
float Y[] = { -0.667f };
int incX = -1;
int incY = 1;
float expected = 1.541604f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 6)");
};
{
int N = 1;
float alpha = 0.0f;
float X[] = { 0.481f };
float Y[] = { 0.523f };
int incX = -1;
int incY = -1;
float expected = 0.251563f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 7)");
};
{
int N = 1;
float alpha = 0.1f;
float X[] = { 0.481f };
float Y[] = { 0.523f };
int incX = -1;
int incY = -1;
float expected = 0.351563f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 8)");
};
{
int N = 1;
float alpha = 1.0f;
float X[] = { 0.481f };
float Y[] = { 0.523f };
int incX = -1;
int incY = -1;
float expected = 1.251563f;
float f;
f = cblas_sdsdot (N, alpha, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdsdot(case 9)");
};
{
int N = 1;
float X[] = { 0.785f };
float Y[] = { -0.7f };
int incX = 1;
int incY = -1;
float expected = -0.5495f;
float f;
f = cblas_sdot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdot(case 10)");
};
{
int N = 1;
double X[] = { 0.79 };
double Y[] = { -0.679 };
int incX = 1;
int incY = -1;
double expected = -0.53641;
double f;
f = cblas_ddot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, dbleps, "ddot(case 11)");
};
{
int N = 1;
float X[] = { 0.474f, -0.27f };
float Y[] = { -0.144f, -0.392f };
int incX = 1;
int incY = -1;
float expected[2] = {-0.174096f, -0.146928f};
float f[2];
cblas_cdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotu(case 12) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotu(case 12) imag");
};
{
int N = 1;
float X[] = { 0.474f, -0.27f };
float Y[] = { -0.144f, -0.392f };
int incX = 1;
int incY = -1;
float expected[2] = {0.037584f, -0.224688f};
float f[2];
cblas_cdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotc(case 13) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotc(case 13) imag");
};
{
int N = 1;
double X[] = { -0.87, -0.631 };
double Y[] = { -0.7, -0.224 };
int incX = 1;
int incY = -1;
double expected[2] = {0.467656, 0.63658};
double f[2];
cblas_zdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotu(case 14) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotu(case 14) imag");
};
{
int N = 1;
double X[] = { -0.87, -0.631 };
double Y[] = { -0.7, -0.224 };
int incX = 1;
int incY = -1;
double expected[2] = {0.750344, -0.24682};
double f[2];
cblas_zdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotc(case 15) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotc(case 15) imag");
};
{
int N = 1;
float X[] = { -0.457f };
float Y[] = { 0.839f };
int incX = -1;
int incY = 1;
float expected = -0.383423f;
float f;
f = cblas_sdot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdot(case 16)");
};
{
int N = 1;
double X[] = { 0.949 };
double Y[] = { -0.873 };
int incX = -1;
int incY = 1;
double expected = -0.828477;
double f;
f = cblas_ddot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, dbleps, "ddot(case 17)");
};
{
int N = 1;
float X[] = { 0.852f, -0.045f };
float Y[] = { 0.626f, -0.164f };
int incX = -1;
int incY = 1;
float expected[2] = {0.525972f, -0.167898f};
float f[2];
cblas_cdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotu(case 18) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotu(case 18) imag");
};
{
int N = 1;
float X[] = { 0.852f, -0.045f };
float Y[] = { 0.626f, -0.164f };
int incX = -1;
int incY = 1;
float expected[2] = {0.540732f, -0.111558f};
float f[2];
cblas_cdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotc(case 19) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotc(case 19) imag");
};
{
int N = 1;
double X[] = { -0.786, -0.341 };
double Y[] = { -0.271, -0.896 };
int incX = -1;
int incY = 1;
double expected[2] = {-0.09253, 0.796667};
double f[2];
cblas_zdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotu(case 20) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotu(case 20) imag");
};
{
int N = 1;
double X[] = { -0.786, -0.341 };
double Y[] = { -0.271, -0.896 };
int incX = -1;
int incY = 1;
double expected[2] = {0.518542, 0.611845};
double f[2];
cblas_zdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotc(case 21) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotc(case 21) imag");
};
{
int N = 1;
float X[] = { -0.088f };
float Y[] = { -0.165f };
int incX = -1;
int incY = -1;
float expected = 0.01452f;
float f;
f = cblas_sdot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, flteps, "sdot(case 22)");
};
{
int N = 1;
double X[] = { -0.434 };
double Y[] = { -0.402 };
int incX = -1;
int incY = -1;
double expected = 0.174468;
double f;
f = cblas_ddot(N, X, incX, Y, incY);
gsl_test_rel(f, expected, dbleps, "ddot(case 23)");
};
{
int N = 1;
float X[] = { -0.347f, 0.899f };
float Y[] = { -0.113f, -0.858f };
int incX = -1;
int incY = -1;
float expected[2] = {0.810553f, 0.196139f};
float f[2];
cblas_cdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotu(case 24) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotu(case 24) imag");
};
{
int N = 1;
float X[] = { -0.347f, 0.899f };
float Y[] = { -0.113f, -0.858f };
int incX = -1;
int incY = -1;
float expected[2] = {-0.732131f, 0.399313f};
float f[2];
cblas_cdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], flteps, "cdotc(case 25) real");
gsl_test_rel(f[1], expected[1], flteps, "cdotc(case 25) imag");
};
{
int N = 1;
double X[] = { -0.897, -0.204 };
double Y[] = { -0.759, 0.557 };
int incX = -1;
int incY = -1;
double expected[2] = {0.794451, -0.344793};
double f[2];
cblas_zdotu_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotu(case 26) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotu(case 26) imag");
};
{
int N = 1;
double X[] = { -0.897, -0.204 };
double Y[] = { -0.759, 0.557 };
int incX = -1;
int incY = -1;
double expected[2] = {0.567195, -0.654465};
double f[2];
cblas_zdotc_sub(N, X, incX, Y, incY, &f);
gsl_test_rel(f[0], expected[0], dbleps, "zdotc(case 27) real");
gsl_test_rel(f[1], expected[1], dbleps, "zdotc(case 27) imag");
};
}
float caffe_cpu_strided_dot<float>(const int n, const float* x, const int incx,
const float* y, const int incy) {
return cblas_sdot(n, x, incx, y, incy);
}
static inline ffloat sdot_(const int *n, const float *x, const int *incx, const float *y, const int *incy)
{
return cblas_sdot(*n, x, *incx, y, *incy);
}
float wrapper_cblas_sdot(const int N, const float *X, const int incX, const float *Y, const int incY)
{
return cblas_sdot(N, X, incX, Y, incY);
}
static void master(int nslaves, char* parameterFile)
{
//VARIABLE DECLARATIONS
time_t startTime, computationStartTime, endTime;
int rank, i, j, accel, MAX_ITER, *slave_ldAs, total_ldA, rdA, numLambdas, error;
ISTAinstance_mpi* instance;
float *xvalue, *result, *b, *lambdas, lambdaStart, lambdaFinish, gamma, step, MIN_FUNCDIFF;
char regType, xfilename[MAX_FILENAME_SIZE], bfilename[MAX_FILENAME_SIZE], outfilename[MAX_FILENAME_SIZE];
//START TIMER
startTime = time(NULL);
//GET VALUES FROM PARAMETER FILE
getMasterParams(parameterFile, xfilename, bfilename, outfilename, &rdA,
&numLambdas, &lambdaStart, &lambdaFinish, &gamma, &step, ®Type, &accel,
&MAX_ITER, &MIN_FUNCDIFF);
//STORE EACH SLAVE'S INDIVIDUAL LDA AND CALCULATE TOTAL_LDA
slave_ldAs = (int*)malloc((nslaves+1)*sizeof(int));
int my_ldA = 0;
MPI_Gather(&my_ldA, 1, MPI_INT, slave_ldAs, 1, MPI_INT, 0, MPI_COMM_WORLD);
total_ldA = 0;
for(i=0; i<=nslaves; i++)
total_ldA += slave_ldAs[i];
fprintf(stdout, "TOTAL LDA IS %d\n", total_ldA);
//ALLOCATE MEMORY
xvalue = calloc(rdA+1,sizeof(float));
result = malloc((total_ldA+rdA)*sizeof(float));
b = malloc((total_ldA)*sizeof(float));
lambdas = malloc(numLambdas*sizeof(float));
if(xvalue==NULL || result==NULL || b==NULL || lambdas==NULL)
fprintf(stdout,"Unable to allocate memory!");
//ASSIGN VALUES TO XVALUE AND B
error=1;
if(strcmp(xfilename, "zeros")==0){
//do nothing - calloc already initialized xvalue to 0
}
else
error *= getVector(xvalue, rdA, xfilename);
error *= getVector(b, total_ldA, bfilename);
//CHECK FOR FILEOPEN ERRORS; IF ANY PRESENT END PROGRAM
for(i=1; i<=nslaves; i++)
if(slave_ldAs[i] == -1) error=0;
MPI_Bcast(&error, 1, MPI_INT, 0, MPI_COMM_WORLD);
if(error==0) {
free(result);
free(xvalue);
free(b);
free(lambdas);
return;
}
//PRINT INPUTS
/* fprintf(stdout, "Here's x:\n");
for(i=0; i < rdA; i++)
{
fprintf(stdout, "%f ", xvalue[i]);
}
fprintf(stdout, "\n and here's b:\n");
for(i=0; i < total_ldA; i++)
{
fprintf(stdout, "%f ", b[i]);
}
*/
//CREATE ISTA OBJECT
instance = ISTAinstance_mpi_new(slave_ldAs, total_ldA, rdA, b, lambdaStart, gamma,
accel, regType, xvalue, step,
nslaves, MPI_COMM_WORLD,
TAG_AX, TAG_ATX, TAG_ATAX, TAG_DIE);
//CENTER FEATURES
float* shifts = calloc(rdA, sizeof(float));
MPI_Reduce(shifts, instance->meanShifts, rdA, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
cblas_sscal(rdA, 1.0 / total_ldA, instance->meanShifts, 1);
MPI_Bcast(instance->meanShifts, rdA, MPI_FLOAT, 0, MPI_COMM_WORLD);
//SCALE FEATURES
float* norms = calloc(rdA, sizeof(float));
MPI_Reduce(norms, instance->scalingFactors, rdA, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
for(j=0; j<rdA; j++)
instance->scalingFactors[j] = pow(instance->scalingFactors[j], 0.5);
MPI_Bcast(instance->scalingFactors, rdA, MPI_FLOAT, 0, MPI_COMM_WORLD);
//CREATE LAMBDA PATH
calcLambdas(lambdas, numLambdas, lambdaStart, lambdaFinish, instance);
//DEBUGGING AREA
/*float* ones = calloc(rdA+1, sizeof(float));
for(i=0; i< rdA+1; i++) {
ones[i] = 1.0;
}
fprintf(stdout, "meanshifts: ");
for(i=0; i<5; i++) {
fprintf(stdout, "%f ", instance->meanShifts[i]);
}
fprintf(stdout, "\nscalingFactors: ");
for(i=0; i<5; i++) {
fprintf(stdout, "%f ", instance->scalingFactors[i]);
}
fprintf(stdout, "\nlambdas: ");
for(i=0; i<5; i++) {
fprintf(stdout, "%f ", lambdas[i]);
}
fprintf(stdout, "\nA * ones: ");
multiply_Ax(ones, result, instance);
for(i=0; i<5; i++) {
fprintf(stdout, "%f ", result[i]);
}
fprintf(stdout, "\n");
*/
//TIME UPDATE
computationStartTime = time(NULL);
//RUN ISTA
for(j=0; j < numLambdas; j++) {
instance->lambda = lambdas[j];
ISTAsolve_lite(instance, MAX_ITER, MIN_FUNCDIFF);
fprintf(stdout, "\n");
}
//UNDO RESCALING
for(i=0; i<(instance->rdA); i++) {
if(instance->scalingFactors[i] > 0.0001)
instance->xcurrent[i] = instance->xcurrent[i] / instance->scalingFactors[i];
}
instance->intercept = -1.0 * cblas_sdot(instance->rdA, instance->xcurrent, 1, instance->meanShifts, 1);
//WRITE RESULTS
writeResults(instance, outfilename, bfilename, lambdas[numLambdas-1]);
//STOP TIME
endTime = time(NULL);
fprintf(stdout,"Setup took %f seconds and computation took %f seconds\n",
difftime(computationStartTime, startTime), difftime(endTime, computationStartTime));
//CLOSE THE SLAVE PROCESSES AND FREE MEMORY
fprintf(stdout, "Closing the program\n");
for(rank=1; rank <= nslaves; rank++)
{
MPI_Send(0, 0, MPI_INT, rank, TAG_DIE, MPI_COMM_WORLD);
}
free(result);
ISTAinstance_mpi_free(instance);
free(shifts);
free(norms);
free(lambdas);
return;
}
float caffe_cpu_dot<float>(const int n, const float* x, const float* y) {
return cblas_sdot(n, x, 1, y, 1);
}