TheRock/rocm-systems/projects/hip-tests/catch/unit/math/single_precision_intrinsics.cc

/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

#include <hip_test_common.hh>

#include "unary_common.hh"
#include "binary_common.hh"
#include "ternary_common.hh"

/********** Unary Functions **********/

#define MATH_UNARY_SP_KERNEL_DEF(func_name)                                                        \
  __global__ void func_name##_kernel(float* const ys, const size_t num_xs, float* const xs) {      \
    const auto tid = cg::this_grid().thread_rank();                                                \
    const auto stride = cg::this_grid().size();                                                    \
                                                                                                   \
    for (size_t i = tid; i < num_xs; i += stride) {                                                \
      ys[i] = func_name(xs[i]);                                                                    \
    }                                                                                              \
  }

#define MATH_UNARY_SP_TEST_DEF_IMPL(func_name, ref_func, validator_builder)                        \
  HIP_TEST_CASE(Unit_Device_##func_name##_Accuracy_Positive) {                                     \
    UnarySinglePrecisionTest(func_name##_kernel, ref_func, validator_builder);                     \
  }

#define MATH_UNARY_SP_TEST_DEF(func_name, ref_func)                                                \
  MATH_UNARY_SP_TEST_DEF_IMPL(func_name, ref_func, func_name##_validator_builder)

#define MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(func_name)                                             \
  static std::unique_ptr<MatcherBase<float>> func_name##_validator_builder(float target, float x)


static float __frcp_rn_ref(float x) { return 1.0f / x; }

MATH_UNARY_SP_KERNEL_DEF(__frcp_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__frcp_rn(x)` for all possible inputs. The error bounds are
 * IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF_IMPL(__frcp_rn, __frcp_rn_ref, EqValidatorBuilderFactory<float>());


MATH_UNARY_SP_KERNEL_DEF(__fsqrt_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fsqrt_rn(x)` for all possible inputs. The results are
 * compared against reference function `float std::sqrt(float)`. The error bounds are
 * IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF_IMPL(__fsqrt_rn, static_cast<float (*)(float)>(std::sqrt),
                            EqValidatorBuilderFactory<float>());


static float __frsqrt_rn_ref(float x) { return 1.0f / std::sqrt(x); }

MATH_UNARY_SP_KERNEL_DEF(__frsqrt_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__frsqrt_rn(x)` for all possible inputs. The results are
 * compared against reference function `float std::sqrt(float)`. The error bounds are
 * IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF_IMPL(__frsqrt_rn, __frsqrt_rn_ref, EqValidatorBuilderFactory<float>());


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__expf) {
  const int64_t ulp_err = 2 + static_cast<int64_t>(std::floor(std::abs(1.16f * x)));
  return ULPValidatorBuilderFactory<float>(ulp_err)(target);
}

MATH_UNARY_SP_KERNEL_DEF(__expf);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__expf(x)` for all possible inputs. The results are
 * compared against reference function `double std::exp(double)`. The maximum ulp error is `2 +
 * floor(abs(1.16 * x))`.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__expf, static_cast<double (*)(double)>(std::exp));


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__exp10f) {
  const int64_t ulp_err = 2 + static_cast<int64_t>(std::floor(std::abs(2.95f * x)));
  return ULPValidatorBuilderFactory<float>(ulp_err)(target);
}

MATH_UNARY_SP_KERNEL_DEF(__exp10f);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__exp10f(x)` for all possible inputs. The results are
 * compared against reference function `double exp10(double)`. The maximum ulp error is `2 +
 * floor(abs(2.95 * x))`.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__exp10f, static_cast<double (*)(double)>(exp10));


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__logf) {
  if (0.5f <= x && x <= 2.0f) {
    const auto abs_err = std::pow(2.0, -21.41);
    return AbsValidatorBuilderFactory<float>(abs_err)(target);
  } else {
    return ULPValidatorBuilderFactory<float>(3)(target);
  }
}

MATH_UNARY_SP_KERNEL_DEF(__logf);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__logf(x)` for all possible inputs. The results are
 * compared against reference function `double std::log(double)`. For `x` in [0.5, 2], the maximum
 * absolute error is 2^-21.41, otherwise, the maximum ulp error is 3.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__logf, static_cast<double (*)(double)>(std::log));


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__log2f) {
  if (0.5f <= x && x <= 2.0f) {
    const auto abs_err = std::pow(2.0, -22.0);
    return AbsValidatorBuilderFactory<float>(abs_err)(target);
  } else {
    return ULPValidatorBuilderFactory<float>(2)(target);
  }
}

MATH_UNARY_SP_KERNEL_DEF(__log2f);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__log2f(x)` for all possible inputs. The results are
 * compared against reference function `double std::log2(double)`. For `x` in [0.5, 2], the maximum
 * absolute error is 2^-22, otherwise, the maximum ulp error is 2.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__log2f, static_cast<double (*)(double)>(std::log2));


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__log10f) {
  if (0.5f <= x && x <= 2.0f) {
    const auto abs_err = std::pow(2.0, -24.0);
    return AbsValidatorBuilderFactory<float>(abs_err)(target);
  } else {
    return ULPValidatorBuilderFactory<float>(3)(target);
  }
}

MATH_UNARY_SP_KERNEL_DEF(__log10f);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__log10f(x)` for all possible inputs. The results are
 * compared against reference function `double std::log10(double)`. For `x` in [0.5, 2], the maximum
 * absolute error is 2^-24, otherwise, the maximum ulp error is 3.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__log10f, static_cast<double (*)(double)>(std::log10));


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__sinf) {
  if (-M_PI <= x && x <= M_PI) {
    const auto abs_err = std::pow(2.0, -21.41);
    return AbsValidatorBuilderFactory<float>(abs_err)(target);
  } else {
    return NopValidatorBuilderFactory<float>()();
  }
}

MATH_UNARY_SP_KERNEL_DEF(__sinf);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__sinf(x)` for all possible inputs. The results are
 * compared against reference function `double std::sin(double)`. For `x` in [-PI, PI], the maximum
 * absolute error is 2^-21.41, and larger otherwise.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__sinf, static_cast<double (*)(double)>(std::sin));


__device__ float __sincosf_sin(float x) {
  float sin, cos;
  __sincosf(x, &sin, &cos);
  return sin;
}

MATH_UNARY_SP_KERNEL_DEF(__sincosf_sin);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__sincosf(x, sptr, cptr)` for all possible inputs. The
 * results in `sptr` are compared against reference function `double std::sin(double)`. For `x` in
 * [-PI, PI], the maximum absolute error is 2^-21.41, and larger otherwise.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF_IMPL(__sincosf_sin, static_cast<double (*)(double)>(std::sin),
                            __sinf_validator_builder);


MATH_UNARY_SP_VALIDATOR_BUILDER_DEF(__cosf) {
  if (-M_PI <= x && x <= M_PI) {
    const auto abs_err = std::pow(2.0, -21.19);
    return AbsValidatorBuilderFactory<float>(abs_err)(target);
  } else {
    return NopValidatorBuilderFactory<float>()();
  }
}

MATH_UNARY_SP_KERNEL_DEF(__cosf);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__cosf(x)` for all possible inputs. The results are
 * compared against reference function `double std::cos(double)`. For `x` in [-PI, PI], the maximum
 * absolute error is 2^-21.19, and larger otherwise.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF(__cosf, static_cast<double (*)(double)>(std::cos));


__device__ float __sincosf_cos(float x) {
  float sin, cos;
  __sincosf(x, &sin, &cos);
  return cos;
}

MATH_UNARY_SP_KERNEL_DEF(__sincosf_cos);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__sincosf(x, sptr, cptr)` for all possible inputs. The
 * results in `cptr` are compared against reference function `double std::cos(double)`. For `x` in
 * [-PI, PI], the maximum absolute error is 2^-21.19, and larger otherwise.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_UNARY_SP_TEST_DEF_IMPL(__sincosf_cos, static_cast<double (*)(double)>(std::cos),
                            __cosf_validator_builder);


/********** Binary Functions **********/

#define MATH_BINARY_SP_KERNEL_DEF(func_name)                                                       \
  __global__ void func_name##_kernel(float* const ys, const size_t num_xs, float* const x1s,       \
                                     float* const x2s) {                                           \
    const auto tid = cg::this_grid().thread_rank();                                                \
    const auto stride = cg::this_grid().size();                                                    \
                                                                                                   \
    for (size_t i = tid; i < num_xs; i += stride) {                                                \
      ys[i] = func_name(x1s[i], x2s[i]);                                                           \
    }                                                                                              \
  }

#define MATH_BINARY_SP_TEST_DEF_IMPL(func_name, ref_func, validator_builder)                       \
  HIP_TEST_CASE(Unit_Device_##func_name##_Accuracy_Positive) {                                     \
    BinaryFloatingPointTest(func_name##_kernel, ref_func, validator_builder);                      \
  }

#define MATH_BINARY_SP_TEST_DEF(func_name, ref_func)                                               \
  MATH_BINARY_SP_TEST_DEF_IMPL(func_name, ref_func, func_name##_validator_builder)

#define MATH_BINARY_SP_VALIDATOR_BUILDER_DEF(func_name)                                            \
  static std::unique_ptr<MatcherBase<float>> func_name##_validator_builder(float target, float x1, \
                                                                           float x2)

static float __fadd_rn_ref(float x1, float x2) { return x1 + x2; }

MATH_BINARY_SP_KERNEL_DEF(__fadd_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fadd_rn(x,y)` against a table of difficult values,
 * followed by a large number of randomly generated values. The error bounds are IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_BINARY_SP_TEST_DEF_IMPL(__fadd_rn, __fadd_rn_ref, EqValidatorBuilderFactory<float>());


static float __fsub_rn_ref(float x1, float x2) { return x1 - x2; }

MATH_BINARY_SP_KERNEL_DEF(__fsub_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fsub_rn(x,y)` against a table of difficult values,
 * followed by a large number of randomly generated values. The error bounds are IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_BINARY_SP_TEST_DEF_IMPL(__fsub_rn, __fsub_rn_ref, EqValidatorBuilderFactory<float>());


static float __fmul_rn_ref(float x1, float x2) { return x1 * x2; }

MATH_BINARY_SP_KERNEL_DEF(__fmul_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fmul_rn(x,y)` against a table of difficult values,
 * followed by a large number of randomly generated values. The error bounds are IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_BINARY_SP_TEST_DEF_IMPL(__fmul_rn, __fmul_rn_ref, EqValidatorBuilderFactory<float>());


static float __fdiv_rn_ref(float x1, float x2) { return x1 / x2; }

MATH_BINARY_SP_KERNEL_DEF(__fdiv_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fdiv_rn(x,y)` against a table of difficult values,
 * followed by a large number of randomly generated values. The error bounds are IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_BINARY_SP_TEST_DEF_IMPL(__fdiv_rn, __fdiv_rn_ref, EqValidatorBuilderFactory<float>());


MATH_BINARY_SP_VALIDATOR_BUILDER_DEF(__fdividef) {
  x1 = 2.0f;
  const auto abs_x2 = std::abs(x2);
  if (std::pow(x1, -126.0f) <= abs_x2 && abs_x2 <= std::pow(x1, 126.0f)) {
    return ULPValidatorBuilderFactory<float>(2)(target);
  } else {
    return NopValidatorBuilderFactory<float>()();
  }
}

MATH_BINARY_SP_KERNEL_DEF(__fdividef);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fdividef(x,y)` against a table of difficult values,
 * followed by a large number of randomly generated values. For `|y|` in [2^-126, 2^126], the
 * maximum ulp error is 2.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_BINARY_SP_TEST_DEF(__fdividef, __fdiv_rn_ref);


/********** Ternary Functions **********/

#define MATH_TERNARY_SP_KERNEL_DEF(func_name)                                                      \
  __global__ void func_name##_kernel(float* const ys, const size_t num_xs, float* const x1s,       \
                                     float* const x2s, float* const x3s) {                         \
    const auto tid = cg::this_grid().thread_rank();                                                \
    const auto stride = cg::this_grid().size();                                                    \
                                                                                                   \
    for (size_t i = tid; i < num_xs; i += stride) {                                                \
      ys[i] = func_name(x1s[i], x2s[i], x3s[i]);                                                   \
    }                                                                                              \
  }

#define MATH_TERNARY_SP_TEST_DEF_IMPL(func_name, ref_func, validator_builder)                      \
  HIP_TEST_CASE(Unit_Device_##func_name##_Accuracy_Positive) {                                     \
    TernaryFloatingPointTest(func_name##_kernel, ref_func, validator_builder);                     \
  }

#define MATH_TERNARY_SP_TEST_DEF(func_name, ref_func, validator_builder)                           \
  MATH_TERNARY_SP_TEST_DEF_IMPL(func_name, ref_func, func_name##_validator_builder)

#define MATH_TERNARY_SP_VALIDATOR_BUILDER_DEF(func_name)                                           \
  static std::unique_ptr<MatcherBase<float>> func_name##_validator_builder(float target, float x1, \
                                                                           float x2, float x3)


MATH_TERNARY_SP_KERNEL_DEF(__fmaf_rn);

/**
 * Test Description
 * ------------------------
 *    - Tests the numerical accuracy of `__fmaf(x,y,z)` against a table of difficult values,
 * followed by a large number of randomly generated values. The error bounds are IEEE-compliant.
 *
 * Test source
 * ------------------------
 *    - unit/math/single_precision_intrinsics.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
MATH_TERNARY_SP_TEST_DEF_IMPL(__fmaf_rn, static_cast<float (*)(float, float, float)>(std::fma),
                              EqValidatorBuilderFactory<float>());