// Copyright 2024 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$TESTNAME = f"F32Simd{ARCH.upper()}Test"

$if ARCH_MACRO:
  // This header needs to go first for the arch test macros.
  #include "xnnpack/common.h"

  #if ${ARCH_MACRO}

#include <algorithm>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <random>
#include <vector>

#include <gmock/gmock.h>
#include <gtest/gtest.h>
#include "xnnpack/isa-checks.h"
#include "xnnpack/simd/f32-${ARCH}.h"
#include "replicable_random_device.h"

namespace xnnpack {

class ${TESTNAME} : public ::testing::Test {
 protected:
  void SetUp() override {
    $if TEST_REQUIRES:
      ${TEST_REQUIRES};
    inputs_.resize(3 * xnn_simd_size_f32);
    output_.resize(xnn_simd_size_f32);
    std::uniform_real_distribution<float> f32dist(-10.0f, 10.0f);
    std::generate(inputs_.begin(), inputs_.end(),
                  [&]() { return f32dist(rng_); });
  }

  xnnpack::ReplicableRandomDevice rng_;
  std::vector<float> inputs_;
  std::vector<float> output_;
};

TEST_F(${TESTNAME}, SetZero) {
  xnn_storeu_f32(output_.data(), xnn_zero_f32());
  EXPECT_THAT(output_, testing::Each(testing::Eq(0.0f)));
}

TEST_F(${TESTNAME}, Add) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_add_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], inputs_[k] + inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, Mul) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_mul_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], inputs_[k] * inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, Fmadd) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t c =
      xnn_loadu_f32(inputs_.data() + 2 * xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_fmadd_f32(a, b, c);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
#if XNN_SIMD_HAS_NATIVE_FMA
    // If an arch claims to support FMA, it better also round things correctly.
    ASSERT_EQ(output_[k],
              static_cast<float>(
                  static_cast<double>(inputs_[k]) *
                      static_cast<double>(inputs_[k + xnn_simd_size_f32]) +
                  static_cast<double>(inputs_[k + 2 * xnn_simd_size_f32])));
#else
    ASSERT_EQ(output_[k],
              inputs_[k] * inputs_[k + xnn_simd_size_f32] +
                  inputs_[k + 2 * xnn_simd_size_f32]);
#endif  // XNN_SIMD_HAS_NATIVE_FMA
  }
}

TEST_F(${TESTNAME}, Fmsub) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t c =
      xnn_loadu_f32(inputs_.data() + 2 * xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_fmsub_f32(a, b, c);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
#if XNN_SIMD_HAS_NATIVE_FMA
    // If an arch claims to support FMA, it better also round things correctly.
    ASSERT_EQ(output_[k],
              static_cast<float>(
                  static_cast<double>(inputs_[k]) *
                      static_cast<double>(inputs_[k + xnn_simd_size_f32]) -
                  static_cast<double>(inputs_[k + 2 * xnn_simd_size_f32])));
#else
    ASSERT_EQ(output_[k],
              inputs_[k] * inputs_[k + xnn_simd_size_f32] -
                  inputs_[k + 2 * xnn_simd_size_f32]);
#endif  // XNN_SIMD_HAS_NATIVE_FMA
  }
}

TEST_F(${TESTNAME}, Fnmadd) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t c =
      xnn_loadu_f32(inputs_.data() + 2 * xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_fnmadd_f32(a, b, c);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
#if XNN_SIMD_HAS_NATIVE_FMA
    // If an arch claims to support FMA, it better also round things correctly.
    ASSERT_EQ(output_[k],
              static_cast<float>(
                  static_cast<double>(-inputs_[k]) *
                      static_cast<double>(inputs_[k + xnn_simd_size_f32]) +
                  static_cast<double>(inputs_[k + 2 * xnn_simd_size_f32])));
#else
    ASSERT_EQ(output_[k],
              -inputs_[k] * inputs_[k + xnn_simd_size_f32] +
                  inputs_[k + 2 * xnn_simd_size_f32]);
#endif  // XNN_SIMD_HAS_NATIVE_FMA
  }
}

TEST_F(${TESTNAME}, Sub) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_sub_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], inputs_[k] - inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, Div) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_div_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_NEAR(output_[k], inputs_[k] / inputs_[k + xnn_simd_size_f32],
    2 * std::numeric_limits<float>::epsilon() * std::abs(output_[k]));
  }
}

TEST_F(${TESTNAME}, Max) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_max_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], std::max(inputs_[k], inputs_[k + xnn_simd_size_f32]));
  }
}

TEST_F(${TESTNAME}, Min) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_min_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], std::min(inputs_[k], inputs_[k + xnn_simd_size_f32]));
  }
}

TEST_F(${TESTNAME}, Abs) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t res = xnn_abs_f32(a);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], std::abs(inputs_[k]));
  }
}

TEST_F(${TESTNAME}, Neg) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t res = xnn_neg_f32(a);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], -inputs_[k]);
  }
}

TEST_F(${TESTNAME}, And) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_and_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(*(uint32_t *)&output_[k],
              *(uint32_t *)&inputs_[k] &
                  *(uint32_t *)&inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, Or) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_or_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(*(uint32_t *)&output_[k],
              *(uint32_t *)&inputs_[k] |
                  *(uint32_t *)&inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, Xor) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_xor_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(*(uint32_t *)&output_[k],
              *(uint32_t *)&inputs_[k] ^
                  *(uint32_t *)&inputs_[k + xnn_simd_size_f32]);
  }
}

TEST_F(${TESTNAME}, ShiftLeft) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  // Not using a loop since the `bits` parameter needs to be a compile-time
  // constant, e.g. for `neon`.
  $for BITS in range(1, 32):
    {
      const xnn_simd_f32_t res = xnn_sll_f32(a, ${BITS});
      xnn_storeu_f32(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f32; k++) {
        ASSERT_EQ(*(uint32_t *)&output_[k], *(uint32_t *)&inputs_[k] << ${BITS});
      }
    }
}

TEST_F(${TESTNAME}, ShiftRight) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  // Not using a loop since the `bits` parameter needs to be a compile-time
  // constant, e.g. for `neon`.
  $for BITS in range(1, 32):
    {
      const xnn_simd_f32_t res = xnn_srl_f32(a, ${BITS});
      xnn_storeu_f32(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f32; k++) {
        ASSERT_EQ(*(uint32_t *)&output_[k], *(uint32_t *)&inputs_[k] >> ${BITS});
      }
    }
}

TEST_F(${TESTNAME}, ShiftRightSigned) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  // Not using a loop since the `bits` parameter needs to be a compile-time
  // constant, e.g. for `neon`.
  $for BITS in range(1, 32):
    {
      const xnn_simd_f32_t res = xnn_sra_f32(a, ${BITS});
      xnn_storeu_f32(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f32; k++) {
        ASSERT_EQ(*(int32_t *)&output_[k], *(int32_t *)&inputs_[k] >> ${BITS});
      }
    }
}

TEST_F(${TESTNAME}, CmpEq) {
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    if (rng_() & 1) {
      inputs_[k + xnn_simd_size_f32] = inputs_[k];
    }
  }
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t b = xnn_loadu_f32(inputs_.data() + xnn_simd_size_f32);
  const xnn_simd_f32_t res = xnn_cmpeq_f32(a, b);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(*(uint32_t *)&output_[k],
              inputs_[k] == inputs_[k + xnn_simd_size_f32] ? 0xFFFFFFFF : 0);
  }
}

TEST_F(${TESTNAME}, GetExp) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  const xnn_simd_f32_t res = xnn_getexp_f32(a);
  xnn_storeu_f32(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f32; k++) {
    ASSERT_EQ(output_[k], std::logb(inputs_[k]));
  }
}

TEST_F(${TESTNAME}, StoreTail) {
  const xnn_simd_f32_t a = xnn_loadu_f32(inputs_.data());
  for (size_t num_elements = 1; num_elements < xnn_simd_size_f32;
      num_elements++) {
    xnn_store_tail_f32(output_.data(), a, num_elements);
    for (size_t k = 0; k < num_elements; k++) {
      ASSERT_EQ(output_[k], inputs_[k]);
    }
    for (size_t k = num_elements; k < xnn_simd_size_f32; k++) {
      ASSERT_EQ(output_[k], 0.0f);
    }
  }
}

}  // namespace xnnpack

$if ARCH_MACRO:
  #endif  // ${ARCH_MACRO}