Skip to content

Commit

Permalink
mxp: GPU .view() + complex tests
Browse files Browse the repository at this point in the history
  • Loading branch information
@cmpfeil
cmpfeil committed Jan 28, 2025
1 parent 02f4099 commit be61ef5
Showing 1 changed file with 300 additions and 0 deletions.
300 changes: 300 additions & 0 deletions tests/test_mxp.cxx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1784,4 +1784,304 @@ TEST(mxp, device_view_view_axaxaxpy)
EXPECT_EQ(y, (gt::gtensor_device<float, 1>{y_init, y_init, y_init + x_init,
y_init, y_init}));
}

TEST(mxp, device_view_complex_axaxaxpy)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int nx{3};
const int ny{5};
const complex32_t x_init{exp2f(-23), -exp2f(-24)};
const complex32_t y_init{1.f, 1.f};
const float a_init{1.f / 3.f};

EXPECT_NE(y_init.real(), y_init.real() + x_init.real());
EXPECT_NE(y_init.imag(), y_init.imag() + x_init.imag());

const gt::gtensor_device<float, 1> a(nx, a_init);
const gt::gtensor_device<complex32_t, 1> x(nx, x_init);
/* */ gt::gtensor_device<complex32_t, 1> y(ny, y_init);

const auto gt_a =
gt::adapt_device<1>(gt::raw_pointer_cast(a.data()), a.size());
const auto gt_x =
gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto gt_y =
gt::adapt_device<1>(gt::raw_pointer_cast(y.data()), y.size());

using gt::placeholders::_all;
using gt::placeholders::_s;

gt_y.view(_s(1, -1)) =
gt_y.view(_s(1, -1)) + gt_a.view(_all) * gt_x.view(_all) +
gt_a.view(_all) * gt_x.view(_all) + gt_a.view(_all) * gt_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>(ny, y_init)));

const auto mxp_a =
mxp::adapt_device<1, double>(gt::raw_pointer_cast(a.data()), a.size());
const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto mxp_y =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(y.data()), y.size());

mxp_y.view(_s(1, -1)) =
mxp_y.view(_s(1, -1)) + mxp_a.view(_all) * mxp_x.view(_all) +
mxp_a.view(_all) * mxp_x.view(_all) + mxp_a.view(_all) * mxp_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>{y_init, y_init + x_init,
y_init + x_init,
y_init + x_init, y_init}));
}

TEST(mxp, device_view_complex_op_plus)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int nx{3};
const int ny{5};
const complex32_t x_init{exp2f(-23) / 3.f, -exp2f(-24) / 3.f};
const complex32_t y_init{1.f, 1.f};

const complex32_t mxp_ref{y_init + 3.f * x_init};

EXPECT_NE(y_init.real(), mxp_ref.real());
EXPECT_NE(y_init.imag(), mxp_ref.imag());

const gt::gtensor_device<complex32_t, 1> x(nx, x_init);
/* */ gt::gtensor_device<complex32_t, 1> y(ny, y_init);

const auto gt_x =
gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto gt_y =
gt::adapt_device<1>(gt::raw_pointer_cast(y.data()), y.size());

using gt::placeholders::_all;
using gt::placeholders::_s;

gt_y.view(_s(1, -1)) =
gt_y.view(_s(1, -1)) + gt_x.view(_all) + gt_x.view(_all) + gt_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>(ny, y_init)));

const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto mxp_y =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(y.data()), y.size());

mxp_y.view(_s(1, -1)) = mxp_y.view(_s(1, -1)) + mxp_x.view(_all) +
mxp_x.view(_all) + mxp_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>{y_init, mxp_ref, mxp_ref,
mxp_ref, y_init}));
}

TEST(mxp, device_view_complex_op_minus)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int nx{3};
const int ny{5};
const complex32_t x_init{-exp2f(-23) / 3.f, exp2f(-24) / 3.f};
const complex32_t y_init{1.f, 1.f};

const complex32_t mxp_ref{y_init - 3.f * x_init};

EXPECT_NE(y_init.real(), mxp_ref.real());
EXPECT_NE(y_init.imag(), mxp_ref.imag());

const gt::gtensor_device<complex32_t, 1> x(nx, x_init);
/* */ gt::gtensor_device<complex32_t, 1> y(ny, y_init);

const auto gt_x =
gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto gt_y =
gt::adapt_device<1>(gt::raw_pointer_cast(y.data()), y.size());

using gt::placeholders::_all;
using gt::placeholders::_s;

gt_y.view(_s(1, -1)) =
gt_y.view(_s(1, -1)) - gt_x.view(_all) - gt_x.view(_all) - gt_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>(ny, y_init)));

const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto mxp_y =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(y.data()), y.size());

mxp_y.view(_s(1, -1)) = mxp_y.view(_s(1, -1)) - mxp_x.view(_all) -
mxp_x.view(_all) - mxp_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>{y_init, mxp_ref, mxp_ref,
mxp_ref, y_init}));
}

TEST(mxp, device_view_complex_op_multiply)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int nx{3};
const int ny{5};
const complex32_t x_init{1.f + exp2f(-12), 0.f};
const complex32_t y_init{-654.321f};

const complex32_t gt_ref{1.f + exp2f(-11) + exp2f(-12) + exp2f(-23), 0.f};
const complex32_t mxp_ref{1.f + exp2f(-11) + exp2f(-12) + exp2f(-22), 0.f};

EXPECT_NE(gt_ref.real(), mxp_ref.real());

const gt::gtensor_device<complex32_t, 1> x(nx, x_init);
/* */ gt::gtensor_device<complex32_t, 1> y(ny, y_init);

const auto gt_x =
gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto gt_y =
gt::adapt_device<1>(gt::raw_pointer_cast(y.data()), y.size());

using gt::placeholders::_all;
using gt::placeholders::_s;

gt_y.view(_s(1, -1)) = gt_x.view(_all) * gt_x.view(_all) * gt_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>{y_init, gt_ref, gt_ref,
gt_ref, y_init}));

const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto mxp_y =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(y.data()), y.size());

mxp_y.view(_s(1, -1)) =
mxp_x.view(_all) * mxp_x.view(_all) * mxp_x.view(_all);

EXPECT_EQ(y, (gt::gtensor_device<complex32_t, 1>{y_init, mxp_ref, mxp_ref,
mxp_ref, y_init}));
}

TEST(mxp, device_view_complex_op_divide)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int nx{3};
const int ny{5};
double val = 1.5 + exp2f(-8) + exp2f(-15) + exp2f(-23);
double invval = 1. / val;
double ref = val / invval / invval;

#ifdef GTENSOR_DEVICE_CUDA
const double lb_err_expect_gt = 1.98e-07;
#else
const double lb_err_expect_gt = 2.7e-07;
#endif
const double ub_err_expect_mxp = 4.0e-08;

const complex32_t x_init{float(invval), 0.f};
const complex32_t y_init{float(val), 0.f};

const gt::gtensor_device<complex32_t, 1> x(nx, x_init);
/* */ gt::gtensor_device<complex32_t, 1> y_a(ny, y_init);
/* */ gt::gtensor_device<complex32_t, 1> y_b(ny, y_init);

const auto gt_x =
gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto gt_y =
gt::adapt_device<1>(gt::raw_pointer_cast(y_a.data()), y_a.size());

using gt::placeholders::_all;
using gt::placeholders::_s;

gt_y.view(_s(1, -1)) =
gt_y.view(_s(1, -1)) / gt_x.view(_all) / gt_x.view(_all);

/* */ gt::gtensor<complex32_t, 1> h_y(ny);
gt::copy(y_a, h_y);
double gt_err = gt::abs(h_y(1).real() - ref);

EXPECT_EQ(h_y(0), y_init);
EXPECT_GT(gt_err, lb_err_expect_gt);
EXPECT_EQ(h_y(2), h_y(1));
EXPECT_EQ(h_y(3), h_y(1));
EXPECT_EQ(h_y(4), y_init);

const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), x.size());
/* */ auto mxp_y = mxp::adapt_device<1, complex64_t>(
gt::raw_pointer_cast(y_b.data()), y_b.size());

mxp_y.view(_s(1, -1)) =
mxp_y.view(_s(1, -1)) / mxp_x.view(_all) / mxp_x.view(_all);

gt::copy(y_b, h_y);
double mxp_err = std::abs(h_y(1).real() - ref);

EXPECT_EQ(h_y(0), y_init);
EXPECT_LT(mxp_err, ub_err_expect_mxp);
EXPECT_EQ(h_y(2), h_y(1));
EXPECT_EQ(h_y(3), h_y(1));
EXPECT_EQ(h_y(4), y_init);
}

TEST(mxp, device_view_placeholders_complex_aXaXaXpY_2D)
{
using complex32_t = gt::complex<float>;
using complex64_t = gt::complex<double>;

const int mn[2]{4, 3};
const int lk = 1, uk = 2;

const complex32_t x_init{exp2f(-23), -exp2f(-24)};
const complex32_t y_init{1.f, 1.f};
const float a_init{1.f / 3.f};

EXPECT_NE(y_init.real(), y_init.real() + x_init.real());
EXPECT_NE(y_init.imag(), y_init.imag() + x_init.imag());

const gt::gtensor_device<float, 1> a(mn[0], a_init);
const gt::gtensor_device<complex32_t, 1> x(mn[0], x_init);
/* */ gt::gtensor_device<complex32_t, 2> Y(gt::shape(mn[0], mn[1]), y_init);

const auto gt_a = gt::adapt_device<1>(gt::raw_pointer_cast(a.data()), mn[0]);
const auto gt_x = gt::adapt_device<1>(gt::raw_pointer_cast(x.data()), mn[0]);
/* */ auto gt_Y = gt::adapt_device<2>(gt::raw_pointer_cast(Y.data()), mn);

using gt::placeholders::_all;
using gt::placeholders::_newaxis;
using gt::placeholders::_s;

gt_Y.view(_all, _s(lk, uk)) =
gt_Y.view(_all, _s(lk, uk)) +
gt_a.view(_all, _newaxis) * gt_x.view(_all, _newaxis) +
gt_a.view(_all, _newaxis) * gt_x.view(_all, _newaxis) +
gt_a.view(_all, _newaxis) * gt_x.view(_all, _newaxis);

EXPECT_EQ(
Y, (gt::gtensor_device<complex32_t, 2>(gt::shape(mn[0], mn[1]), y_init)));

const auto mxp_a =
mxp::adapt_device<1, double>(gt::raw_pointer_cast(a.data()), mn[0]);
const auto mxp_x =
mxp::adapt_device<1, complex64_t>(gt::raw_pointer_cast(x.data()), mn[0]);
/* */ auto mxp_Y =
mxp::adapt_device<2, complex64_t>(gt::raw_pointer_cast(Y.data()), mn);

mxp_Y.view(_all, _s(lk, uk)) =
mxp_Y.view(_all, _s(lk, uk)) +
mxp_a.view(_all, _newaxis) * mxp_x.view(_all, _newaxis) +
mxp_a.view(_all, _newaxis) * mxp_x.view(_all, _newaxis) +
mxp_a.view(_all, _newaxis) * mxp_x.view(_all, _newaxis);

EXPECT_EQ(
Y, (gt::gtensor_device<complex32_t, 2>{
{y_init, y_init, y_init, y_init},
{y_init + x_init, y_init + x_init, y_init + x_init, y_init + x_init},
{y_init, y_init, y_init, y_init}}));
}

#endif

0 comments on commit be61ef5

Please sign in to comment.