inline void mul_code(backend::source_generator &o, bool invert) {
o.begin_function<T2>("mul");
o.begin_function_parameters();
o.template parameter<T2>("a");
o.template parameter<T2>("b");
o.end_function_parameters();
if(invert) { // conjugate b
o.new_line() << type_name<T2>() << " r = {"
"a.x * b.x + a.y * b.y, "
"a.y * b.x - a.x * b.y};";
} else {
o.new_line() << type_name<T2>() << " r = {"
"a.x * b.x - a.y * b.y, "
"a.y * b.x + a.x * b.y};";
}
o.new_line() << "return r;";
o.end_function();
}
inline void twiddle_code(backend::source_generator &o) {
o.begin_function<T2>("twiddle");
o.begin_function_parameters();
o.template parameter<T>("alpha");
o.end_function_parameters();
if(std::is_same<T, cl_double>::value) {
// use sincos with double since we probably want higher precision
#if defined(VEXCL_BACKEND_OPENCL) || defined(VEXCL_BACKEND_COMPUTE)
o.new_line() << type_name<T>() << " cs, sn = sincos(alpha, &cs);";
#else
o.new_line() << type_name<T>() << " sn, cs;";
o.new_line() << "sincos(alpha, &sn, &cs);";
#endif
o.new_line() << type_name<T2>() << " r = {cs, sn};";
} else {
// use native with float since we probably want higher performance
#if defined(VEXCL_BACKEND_OPENCL) || defined(VEXCL_BACKEND_COMPUTE)
o.new_line() << type_name<T2>() << " r = {"
"native_cos(alpha), native_sin(alpha)};";
#elif defined(VEXCL_BACKEND_CUDA)
o.new_line() << type_name<T>() << " sn, cs;";
o.new_line() << "__sincosf(alpha, &sn, &cs);";
o.new_line() << type_name<T2>() << " r = {cs, sn};";
#elif defined(VEXCL_BACKEND_JIT)
o.new_line() << type_name<T>() << " sn, cs;";
o.new_line() << "sincosf(alpha, &sn, &cs);";
o.new_line() << type_name<T2>() << " r = {cs, sn};";
#else
# error Unsupported backend!
#endif
}
o.new_line() << "return r;";
o.end_function();
}
static void define(backend::source_generator &src, const std::string &fname)
{
const size_t N = cl_vector_length<T>::value;
typedef typename std::conditional<
sizeof(T) < 32, cl_uint, cl_ulong
>::type ctr_t;
const size_t ctr_n = sizeof(T) <= 8 ? 2 : 4;
typedef typename Generator::template function<ctr_t, ctr_n> generator;
const size_t key_n = generator::K;
generator::define(src);
src.begin_function<T>(fname);
src.begin_function_parameters();
src.template parameter<cl_ulong>("prm1");
src.template parameter<cl_ulong>("prm2");
src.end_function_parameters();
src.new_line() << "union ";
src.open("{");
src.new_line() << type_name<ctr_t>() << " ctr[" << ctr_n << "];";
if (std::is_same<Ts, cl_float>::value) {
src.new_line()
<< type_name<cl_uint>() << " res_i[" << N << "];";
src.new_line()
<< type_name<cl_float>() << " res_f[" << N << "];";
} else if (std::is_same<Ts, cl_double>::value) {
src.new_line()
<< type_name<cl_ulong>() << " res_i[" << N << "];";
src.new_line()
<< type_name<cl_double>() << " res_f[" << N << "];";
}
src.new_line() << type_name<T>() << " res;";
src.close("} ctr;");
src.new_line() << type_name<ctr_t>() << " key[" << key_n << "];";
for(size_t i = 0; i < ctr_n; i += 2)
src.new_line()
<< "ctr.ctr[" << i << "] = prm1; "
<< "ctr.ctr[" << i + 1 << "] = prm2;";
for(size_t i = 0; i < key_n; ++i)
src.new_line() << "key[" << i << "] = 0x12345678;";
src.new_line() << generator::name() << "(ctr.ctr, key);";
if(std::is_same<Ts, cl_float>::value) {
for(size_t i = 0; i < N; ++i)
src.new_line()
<< "ctr.res_f[" << i << "] = ctr.res_i[" << i
<< "] / " << std::numeric_limits<cl_uint>::max()
<< ".0f;";
} else if (std::is_same<Ts, cl_double>::value) {
for(size_t i = 0; i < N; ++i)
src.new_line()
<< "ctr.res_f[" << i << "] = ctr.res_i[" << i
<< "] / " << std::numeric_limits<cl_ulong>::max()
<< ".0;";
}
src.new_line() << "return ctr.res;";
src.end_function();
}
static void define(backend::source_generator &src, const std::string &fname)
{
const size_t N = cl_vector_length<T>::value;
const bool is_float = std::is_same<Ts, cl_float>::value;
const size_t ctr_n = is_float ? 2 : 4;
typedef typename Generator::template function<cl_uint, ctr_n> generator;
const size_t key_n = generator::K;
generator::define(src);
src.begin_function<T>(fname);
src.begin_function_parameters();
src.template parameter<cl_ulong>("prm1");
src.template parameter<cl_ulong>("prm2");
src.end_function_parameters();
#if defined(VEXCL_BACKEND_JIT)
src.new_line() << "#define cospi(x) cos(M_PI * (x))";
#endif
src.new_line() << "union ";
src.open("{");
src.new_line() << type_name<cl_uint>() << " ctr[" << ctr_n << "];";
if (is_float) {
src.new_line() << type_name<cl_uint>() << " res_i[2];";
} else {
src.new_line() << type_name<cl_ulong>() << " res_i[2];";
}
src.close("} ctr;");
src.new_line() << type_name<Ts>() << " u[2];";
src.new_line() << type_name<cl_uint>() << " key[" << key_n << "];";
for(size_t i = 0; i < ctr_n; i += 2)
src.new_line()
<< "ctr.ctr[" << i << "] = prm1; "
<< "ctr.ctr[" << i + 1 << "] = prm2;";
for(size_t i = 0; i < key_n; ++i)
src.new_line() << "key[" << i << "] = 0x12345678;";
if (N > 1) {
src.new_line() << "union ";
src.open("{");
src.new_line() << type_name<Ts>() << " z[" << N << "];";
src.new_line() << type_name<T>() << " v;";
src.close("} res;");
}
for(size_t i = 0 ; i < N ; i += 2) {
src.new_line() << generator::name() << "(ctr.ctr, key);";
if(is_float) {
for(size_t i = 0; i < 2; ++i)
src.new_line()
<< "u[" << i << "] = ctr.res_i[" << i
<< "] / " << std::numeric_limits<cl_uint>::max()
<< ".0f;";
} else {
for(size_t i = 0; i < 2; ++i)
src.new_line()
<< "u[" << i << "] = ctr.res_i[" << i
<< "] / " << std::numeric_limits<cl_ulong>::max()
<< ".0;";
}
if(N == 1) {
src.new_line()
<< "return sqrt(-2 * log(u[0])) * cospi(2 * u[1]);\n";
} else {
src.open("{");
src.new_line() << type_name<Ts>()
<< " l = sqrt(-2 * log(u[0])), cs, sn;";
#if defined(VEXCL_BACKEND_CUDA)
src.new_line() << "sincospi(2 * u[1], &sn, &cs);";
#else
src.new_line() << "sn = sincos("
<< std::setprecision(16)
<< boost::math::constants::two_pi<double>()
<< " * u[1], &cs);";
#endif
src.new_line() << "res.z[" << i << "] = l * cs;";
src.new_line() << "res.z[" << i + 1 << "] = l * sn;";
src.close("}");
}
}
if (N > 1)
src.new_line() << "return res.v;";
src.end_function();
}
