TheRock/rocm-systems/projects/hip-tests/catch/unit/warp/warp_shfl.cc

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

#include "warp_shfl_common.hh"

#include <bitset>

/**
 * @addtogroup shfl shfl
 * @{
 * @ingroup DeviceLanguageTest
 * `T __shfl(T var, int src_lane, int width = warpSize)` -
 * Contains unit test for warp shfl function
 */

namespace cg = cooperative_groups;

template <typename T> __global__ void shfl(T* const out, const T* const in,
                                           const uint64_t* const active_masks,
                                           const uint8_t* const src_lanes, const int width) {
  if (deactivate_thread(active_masks)) {
    return;
  }
  const auto grid = cg::this_grid();
  const auto block = cg::this_thread_block();
  T var = in[grid.thread_rank()];
  out[grid.thread_rank()] = __shfl(var, src_lanes[block.thread_rank() % width], width);
}

template <typename T> class WarpShfl : public WarpShflTest<WarpShfl<T>, T> {
 public:
  void launch_kernel(T* const arr_dev, T* const input_dev, const uint64_t* const active_masks) {
    width_ = generate_width(this->warp_size_);
    INFO("Width: " << width_);
    const auto alloc_size = width_ * sizeof(uint8_t);
    LinearAllocGuard<uint8_t> src_lanes_dev(LinearAllocs::hipMalloc, alloc_size);
    src_lanes_.resize(width_);
    std::generate(src_lanes_.begin(), src_lanes_.end(),
                  [this] { return GenRandomInteger(0, static_cast<int>(2 * width_)); });

    HIP_CHECK(hipMemcpy(src_lanes_dev.ptr(), src_lanes_.data(), alloc_size, hipMemcpyHostToDevice));
    shfl<<<this->grid_.grid_dim_, this->grid_.block_dim_>>>(arr_dev, input_dev, active_masks,
                                                            src_lanes_dev.ptr(), width_);
  }

  void validate(const T* const arr, const T* const input) {
    ArrayAllOf(arr, this->grid_.thread_count_, [this, &input](unsigned int i) -> std::optional<T> {
      const auto rank_in_block = this->grid_.thread_rank_in_block(i).value();
      const auto rank_in_warp = rank_in_block % this->warp_size_;
      const auto rank_in_partition = rank_in_block % width_;
      const int src_lane = src_lanes_[rank_in_partition] % width_;
      const int src_offset = src_lane - rank_in_partition;

      const auto mask_idx = this->warps_in_block_ * (i / this->grid_.threads_in_block_count_) +
                            rank_in_block / this->warp_size_;
      const std::bitset<sizeof(uint64_t) * 8> active_mask(this->active_masks_[mask_idx]);

      if (!active_mask.test(rank_in_warp) || (!active_mask.test((rank_in_warp + src_offset))) ||
          (rank_in_block + src_offset >= this->grid_.threads_in_block_count_)) {
        return std::nullopt;
      }

      return input[i + src_offset];
    });
  };

 private:
  std::vector<uint8_t> src_lanes_;
  int width_;
};

/**
 * Test Description
 * ------------------------
 *  - Validates the warp shuffle behavior for all valid width sizes {2, 4, 8, 16, 32,
 * 64(if supported)} for generated shuffle target lanes. The threads are deactivated based on the
 * passed active mask. The test is run for all overloads of shfl.
 * Test source
 * ------------------------
 *  - unit/warp/warp_shfl.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 5.2
 *  - Device supports warp shuffle
 */
HIP_TEMPLATE_TEST_CASE(Unit_Warp_Shfl_Positive_Basic, int, unsigned int, long, unsigned long,
                   long long, unsigned long long, float, double, __half, __half2) {
  int device;
  hipDeviceProp_t device_properties;
  HIP_CHECK(hipGetDevice(&device));
  HIP_CHECK(hipGetDeviceProperties(&device_properties, device));

  if (!device_properties.arch.hasWarpShuffle) {
    HIP_SKIP_TEST(HipTest::SkipReason::kWarpShuffleUnsupported);
  }

  SECTION("Shfl with specified active mask and input values") { WarpShfl<TestType>().run(false); }

  SECTION("Shfl with random active mask and input values") { WarpShfl<TestType>().run(true); }
}

/**
 * End doxygen group DeviceLanguageTest.
 * @}
 */