Use quantised gradients in gpu_hist histograms (#8246)

tests/cpp/common/test_device_helpers.cu

@@ -196,72 +196,4 @@ TEST(DeviceHelpers, ArgSort) {
                                 thrust::greater<size_t>{}));
 }
 
-namespace {
-// Atomic add as type cast for test.
-XGBOOST_DEV_INLINE int64_t atomicAdd(int64_t *dst, int64_t src) {  // NOLINT
-  uint64_t* u_dst = reinterpret_cast<uint64_t*>(dst);
-  uint64_t u_src = *reinterpret_cast<uint64_t*>(&src);
-  uint64_t ret = ::atomicAdd(u_dst, u_src);
-  return *reinterpret_cast<int64_t*>(&ret);
-}
-}
-
-void TestAtomicAdd() {
-  size_t n_elements = 1024;
-  dh::device_vector<int64_t> result_a(1, 0);
-  auto d_result_a = result_a.data().get();
-
-  dh::device_vector<int64_t> result_b(1, 0);
-  auto d_result_b = result_b.data().get();
-
-  /**
-   * Test for simple inputs
-   */
-  std::vector<int64_t> h_inputs(n_elements);
-  for (size_t i = 0; i < h_inputs.size(); ++i) {
-    h_inputs[i] = (i % 2 == 0) ? i : -i;
-  }
-  dh::device_vector<int64_t> inputs(h_inputs);
-  auto d_inputs = inputs.data().get();
-
-  dh::LaunchN(n_elements, [=] __device__(size_t i) {
-    dh::AtomicAdd64As32(d_result_a, d_inputs[i]);
-    atomicAdd(d_result_b, d_inputs[i]);
-  });
-  ASSERT_EQ(result_a[0], result_b[0]);
-
-  /**
-   * Test for positive values that don't fit into 32 bit integer.
-   */
-  thrust::fill(inputs.begin(), inputs.end(),
-               (std::numeric_limits<uint32_t>::max() / 2));
-  thrust::fill(result_a.begin(), result_a.end(), 0);
-  thrust::fill(result_b.begin(), result_b.end(), 0);
-  dh::LaunchN(n_elements, [=] __device__(size_t i) {
-    dh::AtomicAdd64As32(d_result_a, d_inputs[i]);
-    atomicAdd(d_result_b, d_inputs[i]);
-  });
-  ASSERT_EQ(result_a[0], result_b[0]);
-  ASSERT_GT(result_a[0], std::numeric_limits<uint32_t>::max());
-  CHECK_EQ(thrust::reduce(inputs.begin(), inputs.end(), int64_t(0)), result_a[0]);
-
-  /**
-   * Test for negative values that don't fit into 32 bit integer.
-   */
-  thrust::fill(inputs.begin(), inputs.end(),
-               (std::numeric_limits<int32_t>::min() / 2));
-  thrust::fill(result_a.begin(), result_a.end(), 0);
-  thrust::fill(result_b.begin(), result_b.end(), 0);
-  dh::LaunchN(n_elements, [=] __device__(size_t i) {
-    dh::AtomicAdd64As32(d_result_a, d_inputs[i]);
-    atomicAdd(d_result_b, d_inputs[i]);
-  });
-  ASSERT_EQ(result_a[0], result_b[0]);
-  ASSERT_LT(result_a[0], std::numeric_limits<int32_t>::min());
-  CHECK_EQ(thrust::reduce(inputs.begin(), inputs.end(), int64_t(0)), result_a[0]);
-}
-
-TEST(AtomicAdd, Int64) {
-  TestAtomicAdd();
-}
 }  // namespace xgboost