sglang_v0.5.2/pytorch_2.8.0/third_party/XNNPACK/test/f16-simd.cc.in

// 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"F16Simd{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 <random>
#include <vector>

#include <gmock/gmock.h>
#include <gtest/gtest.h>
#include "xnnpack/isa-checks.h"
#include "xnnpack/simd/f16-${ARCH}.h"
#include "xnnpack/fp16.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_f16);
    output_.resize(xnn_simd_size_f16);
    std::uniform_real_distribution<float> f32dist(-10.0f, 10.0f);
    std::generate(inputs_.begin(), inputs_.end(),
                  [&]() { return fp16_ieee_from_fp32_value(f32dist(rng_)); });
  }

  static std::vector<float> ToFloat32(const std::vector<uint16_t> &values) {
    std::vector<float> result;
    result.reserve(values.size());
    for (const uint16_t &value : values) {
      result.push_back(fp16_ieee_to_fp32_value(value));
    }
    return result;
  }

  static float TruncToF16(float value) {
    return fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(value));
  }

  xnnpack::ReplicableRandomDevice rng_;
  std::vector<uint16_t> inputs_;
  std::vector<uint16_t> output_;
};

TEST_F(${TESTNAME}, SetZero) {
  xnn_storeu_f16(output_.data(), xnn_zero_f16());
  EXPECT_THAT(ToFloat32(output_), testing::Each(testing::Eq(0.0f)));
}

TEST_F(${TESTNAME}, Add) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_add_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(inputs_f32[k] + inputs_f32[k + xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Mul) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_mul_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(inputs_f32[k] * inputs_f32[k + xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Fmadd) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t c =
      xnn_loadu_f16(inputs_.data() + 2 * xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_fmadd_f16(a, b, c);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(inputs_f32[k] * inputs_f32[k + xnn_simd_size_f16] +
                         inputs_f32[k + 2 * xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Fmsub) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t c =
      xnn_loadu_f16(inputs_.data() + 2 * xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_fmsub_f16(a, b, c);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(inputs_f32[k] * inputs_f32[k + xnn_simd_size_f16] -
                         inputs_f32[k + 2 * xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Fnmadd) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t c =
      xnn_loadu_f16(inputs_.data() + 2 * xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_fnmadd_f16(a, b, c);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(-inputs_f32[k] * inputs_f32[k + xnn_simd_size_f16] +
                         inputs_f32[k + 2 * xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Sub) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_sub_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              TruncToF16(inputs_f32[k] - inputs_f32[k + xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Div) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_div_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_NEAR(output_f32[k],
                inputs_f32[k] / inputs_f32[k + xnn_simd_size_f16],
                2 * 9.77e-04 * std::abs(output_f32[k]));
  }
}

TEST_F(${TESTNAME}, Max) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_max_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              std::max(inputs_f32[k], inputs_f32[k + xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Min) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_min_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k],
              std::min(inputs_f32[k], inputs_f32[k + xnn_simd_size_f16]));
  }
}

TEST_F(${TESTNAME}, Abs) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t res = xnn_abs_f16(a);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k], std::abs(inputs_f32[k]));
  }
}

TEST_F(${TESTNAME}, Neg) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t res = xnn_neg_f16(a);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k], -inputs_f32[k]);
  }
}

TEST_F(${TESTNAME}, And) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_and_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_[k], inputs_[k] & inputs_[k + xnn_simd_size_f16]);
  }
}

TEST_F(${TESTNAME}, Or) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_or_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_[k], inputs_[k] | inputs_[k + xnn_simd_size_f16]);
  }
}

TEST_F(${TESTNAME}, Xor) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_xor_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_[k], inputs_[k] ^ inputs_[k + xnn_simd_size_f16]);
  }
}

TEST_F(${TESTNAME}, ShiftLeft) {
  const xnn_simd_f16_t a = xnn_loadu_f16(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, 16):
    {
      const xnn_simd_f16_t res = xnn_sll_f16(a, ${BITS});
      xnn_storeu_f16(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f16; k++) {
        ASSERT_EQ(output_[k], (inputs_[k] << ${BITS}) & 0xFFFF);
      }
    }
}

TEST_F(${TESTNAME}, ShiftRight) {
  const xnn_simd_f16_t a = xnn_loadu_f16(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, 16):
    {
      const xnn_simd_f16_t res = xnn_srl_f16(a, ${BITS});
      xnn_storeu_f16(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f16; k++) {
        ASSERT_EQ(output_[k], inputs_[k] >> ${BITS});
      }
    }
}

TEST_F(${TESTNAME}, ShiftRightSigned) {
  const xnn_simd_f16_t a = xnn_loadu_f16(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, 16):
    {
      const xnn_simd_f16_t res = xnn_sra_f16(a, ${BITS});
      xnn_storeu_f16(output_.data(), res);
      for (size_t k = 0; k < xnn_simd_size_f16; k++) {
        ASSERT_EQ(output_[k], inputs_[k] >> ${BITS});
      }
    }
}

TEST_F(${TESTNAME}, CmpEq) {
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    if (rng_() & 1) {
      inputs_[k + xnn_simd_size_f16] = inputs_[k];
    }
  }
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t b = xnn_loadu_f16(inputs_.data() + xnn_simd_size_f16);
  const xnn_simd_f16_t res = xnn_cmpeq_f16(a, b);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_[k],
              inputs_[k] == inputs_[k + xnn_simd_size_f16] ? 0xFFFF : 0);
  }
}

TEST_F(${TESTNAME}, GetExp) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  const xnn_simd_f16_t res = xnn_getexp_f16(a);
  xnn_storeu_f16(output_.data(), res);
  std::vector<float> output_f32 = ToFloat32(output_);
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t k = 0; k < xnn_simd_size_f16; k++) {
    ASSERT_EQ(output_f32[k], std::logb(inputs_f32[k]));
  }
}

TEST_F(${TESTNAME}, StoreTail) {
  const xnn_simd_f16_t a = xnn_loadu_f16(inputs_.data());
  std::vector<float> inputs_f32 = ToFloat32(inputs_);
  for (size_t num_elements = 1; num_elements < xnn_simd_size_f16;
      num_elements++) {
    xnn_store_tail_f16(output_.data(), a, num_elements);
    std::vector<float> output_f32 = ToFloat32(output_);
    for (size_t k = 0; k < num_elements; k++) {
      ASSERT_EQ(output_f32[k], inputs_f32[k]);
    }
    for (size_t k = num_elements; k < xnn_simd_size_f16; k++) {
      ASSERT_EQ(output_f32[k], fp16_ieee_from_fp32_value(0.0f));
    }
  }
}

}  // namespace xnnpack

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