TheRock/rocm-systems/projects/hip-tests/catch/unit/deviceLib/hipBfloat16.cc

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

#include <hip_test_common.hh>
#include <hip/hip_bfloat16.h>
#include <type_traits>
#include <random>
#include <climits>

#define SIZE 100

static std::random_device dev;
static std::mt19937 rng(dev());

inline float getRandomFloat(int16_t min = 10, int64_t max = LONG_MAX) {
  std::uniform_real_distribution<float> gen(min, max);
  return gen(rng);
}
__host__ __device__ bool testRelativeAccuracy(float a, hip_bfloat16 b) {
  float c = static_cast<float>(b);
  // float relative error should be less than 1/(2^7) since bfloat16
  // has 7 bits mantissa.
  if (fabs(c - a) / a <= 1.0 / 128) {
    return true;
  }
  return false;
}
__host__ __device__ bool testOperations(const float& fa, const float& fb) {
  bool testPass = true;
  hip_bfloat16 bf_a(fa);
  hip_bfloat16 bf_b(fb);
  float fc = static_cast<float>(bf_a);
  float fd = static_cast<float>(bf_b);

  testPass &= testRelativeAccuracy(fa, bf_a);
  testPass &= testRelativeAccuracy(fb, bf_b);

  testPass &= testRelativeAccuracy(fc + fd, bf_a + bf_b);
  // when checked as above for add, operation sub fails on GPU
  if (hip_bfloat16(fc - fd) == (bf_a - bf_b)) {
    testPass &= true;
  }
  testPass &= testRelativeAccuracy(fc * fd, bf_a * bf_b);
  testPass &= testRelativeAccuracy(fc / fd, bf_a / bf_b);

  hip_bfloat16 bf_x;
  bf_x = bf_a;
  bf_x++;
  bf_x--;
  ++bf_x;
  --bf_x;
  // hip_bfloat16 is converted to float and then inc/decremented,
  // hence check with reduced precision
  testPass &= testRelativeAccuracy(bf_x, bf_a);

  bf_x = bf_a;
  bf_x += bf_b;
  bf_x = bf_a;
  bf_x -= bf_b;
  bf_x = bf_a;
  bf_x *= bf_b;
  bf_x = bf_a;
  bf_x /= bf_b;

  hip_bfloat16 bf_rounded = hip_bfloat16::round_to_bfloat16(fa);
  if (std::isnan(bf_rounded)) {
    if (std::isnan(bf_rounded) || std::isinf(bf_rounded)) {
      testPass &= true;
    }
  }
  return testPass;
}
__global__ void testOperationsGPU(float* d_a, float* d_b, bool* testPass) {
  int id = threadIdx.x;
  if (id > SIZE) return;
  float& a = d_a[id];
  float& b = d_b[id];
  *testPass = testOperations(a, b);
}
HIP_TEST_CASE(Unit_hipBfloat16) {
  float *h_fa, *h_fb;
  float *d_fa, *d_fb;
  bool *d_fc, h_fc = false;

  h_fa = new float[SIZE];
  h_fb = new float[SIZE];

  bool result = false;
  for (int i = 0; i < SIZE; i++) {
    h_fa[i] = getRandomFloat();
    h_fb[i] = getRandomFloat();
    result = testOperations(h_fa[i], h_fb[i]);
    REQUIRE(result == true);
  }

  HIP_CHECK(hipMalloc(&d_fa, sizeof(float) * SIZE));
  HIP_CHECK(hipMalloc(&d_fb, sizeof(float) * SIZE));
  HIP_CHECK(hipMalloc(&d_fc, sizeof(bool)));

  HIP_CHECK(hipMemcpy(d_fa, h_fa, sizeof(float) * SIZE, hipMemcpyHostToDevice));
  HIP_CHECK(hipMemcpy(d_fb, h_fb, sizeof(float) * SIZE, hipMemcpyHostToDevice));
  HIP_CHECK(hipMemcpy(d_fc, &h_fc, sizeof(bool), hipMemcpyHostToDevice));

  hipLaunchKernelGGL(testOperationsGPU, 1, SIZE, 0, 0, d_fa, d_fb, d_fc);
  HIP_CHECK(hipDeviceSynchronize());
  HIP_CHECK(hipMemcpy(&h_fc, d_fc, sizeof(bool), hipMemcpyDeviceToHost));

  REQUIRE(h_fc == true);

  delete[] h_fa;
  delete[] h_fb;
  HIP_CHECK(hipFree(d_fa));
  HIP_CHECK(hipFree(d_fb));
  HIP_CHECK(hipFree(d_fc));
}