template<> void
xSpMdV<float>::xSpMdV_Function( bool flush )
{
clsparseStatus status = clsparseScsrmv( &a, &csrMtx, &x, &b, &y, control );
if( flush )
clFinish( queue );
}
void test_csrmv()
{
clsparseStatus status;
cl_int cl_status;
clsparseEnableExtendedPrecision(CLSE::control, extended_precision);
if (typeid(T) == typeid(cl_float) )
{
status = clsparseScsrmv(&gAlpha, &CSRE::csrSMatrix, &gX,
&gBeta, &gY, CLSE::control);
ASSERT_EQ(clsparseSuccess, status);
float* vals = (float*)&CSRE::ublasSCsr.value_data()[0];
int* rows = &CSRE::ublasSCsr.index1_data()[0];
int* cols = &CSRE::ublasSCsr.index2_data()[0];
for (int row = 0; row < CSRE::n_rows; row++)
{
// Summation done at a higher precision to decrease
// summation errors from rounding.
hY[row] *= hBeta;
int row_end = rows[row+1];
double temp_sum;
temp_sum = hY[row];
for (int i = rows[row]; i < rows[row+1]; i++)
{
// Perform: hY[row] += hAlpha * vals[i] * hX[cols[i]];
temp_sum += hAlpha * vals[i] * hX[cols[i]];
}
hY[row] = temp_sum;
}
T* host_result = (T*) ::clEnqueueMapBuffer(CLSE::queue, gY.values,
CL_TRUE, CL_MAP_READ,
0, gY.num_values * sizeof(T),
0, nullptr, nullptr, &cl_status);
ASSERT_EQ(CL_SUCCESS, cl_status);
uint64_t max_ulps = 0;
uint64_t min_ulps = UINT64_MAX;
uint64_t total_ulps = 0;
for (int i = 0; i < hY.size(); i++)
{
long long int intDiff = (long long int)boost::math::float_distance(hY[i], host_result[i]);
intDiff = llabs(intDiff);
total_ulps += intDiff;
if (max_ulps < intDiff)
max_ulps = intDiff;
if (min_ulps > intDiff)
min_ulps = intDiff;
// Debug printouts.
//std::cout << "Row " << i << " Float Ulps: " << intDiff << std::endl;
//std::cout.precision(9);
//std::cout << "\tFloat hY[" << i << "] = " << std::scientific << hY[i] << " (0x" << std::hex << *(uint32_t *)&hY[i] << "), " << std::dec;
//std::cout << "host_result[" << i << "] = " << std::scientific << host_result[i] << " (0x" << std::hex << *(uint32_t *)&host_result[i] << ")" << std::dec << std::endl;
}
#ifndef NDEBUG
if (extended_precision)
{
std::cout << "Float Min ulps: " << min_ulps << std::endl;
std::cout << "Float Max ulps: " << max_ulps << std::endl;
std::cout << "Float Total ulps: " << total_ulps << std::endl;
std::cout << "Float Average ulps: " << (double)total_ulps/(double)hY.size() << " (Size: " << hY.size() << ")" << std::endl;
}
#endif
for (int i = 0; i < hY.size(); i++)
{
double compare_val = 0.;
if (extended_precision)
{
// The limit here is somewhat weak because some GPUs don't
// support correctly rounded denorms in SPFP mode.
if (boost::math::isnormal(hY[i]))
compare_val = fabs(hY[i]*1e-3);
}
else
{
if (boost::math::isnormal(hY[i]))
compare_val = fabs(hY[i]*0.1);
}
if (compare_val < 10*FLT_EPSILON)
compare_val = 10*FLT_EPSILON;
ASSERT_NEAR(hY[i], host_result[i], compare_val);
}
cl_status = ::clEnqueueUnmapMemObject(CLSE::queue, gY.values,
host_result, 0, nullptr, nullptr);
ASSERT_EQ(CL_SUCCESS, cl_status);
}
if (typeid(T) == typeid(cl_double) )
{
status = clsparseDcsrmv(&gAlpha, &CSRE::csrDMatrix, &gX,
&gBeta, &gY, CLSE::control);
ASSERT_EQ(clsparseSuccess, status);
double* vals = (double*)&CSRE::ublasDCsr.value_data()[0];
int* rows = &CSRE::ublasDCsr.index1_data()[0];
int* cols = &CSRE::ublasDCsr.index2_data()[0];
for (int row = 0; row < CSRE::n_rows; row++)
{
// Summation done using a compensated summation to decrease
// summation errors from rounding. This allows us to get
// smaller errors without requiring quad precision support.
// This method is like performing summation at quad precision and
// casting down to double in the end.
hY[row] *= hBeta;
int row_end = rows[row+1];
double temp_sum;
temp_sum = hY[row];
T sumk_err = 0.;
for (int i = rows[row]; i < rows[row+1]; i++)
{
// Perform: hY[row] += hAlpha * vals[i] * hX[cols[i]];
temp_sum = two_sum(temp_sum, hAlpha*vals[i]*hX[cols[i]], &sumk_err);
}
hY[row] = temp_sum + sumk_err;
}
T* host_result = (T*) ::clEnqueueMapBuffer(CLSE::queue, gY.values,
CL_TRUE, CL_MAP_READ,
0, gY.num_values * sizeof(T),
0, nullptr, nullptr, &cl_status);
ASSERT_EQ(CL_SUCCESS, cl_status);
uint64_t max_ulps = 0;
uint64_t min_ulps = ULLONG_MAX;
uint64_t total_ulps = 0;
for (int i = 0; i < hY.size(); i++)
{
long long int intDiff = (long long int)boost::math::float_distance(hY[i], host_result[i]);
intDiff = llabs(intDiff);
total_ulps += intDiff;
if (max_ulps < intDiff)
max_ulps = intDiff;
if (min_ulps > intDiff)
min_ulps = intDiff;
// Debug printouts.
//std::cout << "Row " << i << " Double Ulps: " << intDiff << std::endl;
//std::cout.precision(17);
//std::cout << "\tDouble hY[" << i << "] = " << std::scientific << hY[i] << " (0x" << std::hex << *(uint64_t *)&hY[i] << "), " << std::dec;
//std::cout << "host_result[" << i << "] = " << std::scientific << host_result[i] << " (0x" << std::hex << *(uint64_t *)&host_result[i] << ")" << std::dec << std::endl;
}
if (extended_precision)
{
#ifndef NDEBUG
std::cout << "Double Min ulps: " << min_ulps << std::endl;
std::cout << "Double Max ulps: " << max_ulps << std::endl;
std::cout << "Double Total ulps: " << total_ulps << std::endl;
std::cout << "Double Average ulps: " << (double)total_ulps/(double)hY.size() << " (Size: " << hY.size() << ")" << std::endl;
#endif
for (int i = 0; i < hY.size(); i++)
{
double compare_val = fabs(hY[i]*1e-14);
if (compare_val < 10*DBL_EPSILON)
compare_val = 10*DBL_EPSILON;
ASSERT_NEAR(hY[i], host_result[i], compare_val);
}
}
else
{
for (int i = 0; i < hY.size(); i++)
{
double compare_val = 0.;
if (boost::math::isnormal(hY[i]))
compare_val = fabs(hY[i]*0.1);
if (compare_val < 10*DBL_EPSILON)
compare_val = 10*DBL_EPSILON;
ASSERT_NEAR(hY[i], host_result[i], compare_val);
}
}
cl_status = ::clEnqueueUnmapMemObject(CLSE::queue, gY.values,
host_result, 0, nullptr, nullptr);
ASSERT_EQ(CL_SUCCESS, cl_status);
}
// Reset output buffer for next test.
::clReleaseMemObject(gY.values);
clsparseInitVector(&gY);
gY.values = clCreateBuffer(CLSE::context,
CL_MEM_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
hY.size() * sizeof(T), hY.data().begin(),
&cl_status);
gY.num_values = hY.size();
ASSERT_EQ(CL_SUCCESS, cl_status);
}
