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

#include <hip_test_common.hh>
#include <hip/hip_runtime.h>


static __global__ void kernel_syncthreads_or(int* syncTestD, int* allThreadsZeroD,
                                             int* allThreadsOneD, int* oneThreadOneD,
                                             int* allThreadsMinusOneD) {
  int blockSize = blockDim.x;
  int predicate = 10;
  // First block index starts with 0, and second block index starts
  // with blockSize
  int i = (blockIdx.x == 0) ? threadIdx.x : blockSize + threadIdx.x;

  // At very first, we need to ensure work-group level syncronization
  // properly happened, don't bother about predicate testing for now.
  // Thread 0 and thread 1 writes to shared memory. After call to api,
  // every thread reads shared memory, and store subtraction for verification
  __shared__ int sm[2];
  if (threadIdx.x == 0)
    sm[0] = 10;
  else if (threadIdx.x == 1)
    sm[1] = 20;
  __syncthreads_or(predicate);
  syncTestD[i] = sm[1] - sm[0];

  // All threads pass 0 as predicate value, result should be 0
  predicate = 0;
  allThreadsZeroD[i] = __syncthreads_or(predicate);

  // All threads pass 1 as predicate value, result should be 1
  predicate = 1;
  allThreadsOneD[i] = __syncthreads_or(predicate);

  // Thread 0 pass 1, and all other threads 0 as predicate value,
  // result should be 1
  predicate = (threadIdx.x == 0) ? 1 : 0;
  oneThreadOneD[i] = __syncthreads_or(predicate);

  // All threads pass -1 as predicate value, result should be 1
  predicate = -1;
  allThreadsMinusOneD[i] = __syncthreads_or(predicate);
}

static void test_syncthreads_or(int blockSize) {
  int nBytes = sizeof(int) * 2 * blockSize;
  int *syncTestD, *syncTestH;
  int *allThreadsZeroD, *allThreadsZeroH;
  int *allThreadsOneD, *allThreadsOneH;
  int *oneThreadOneD, *oneThreadOneH;
  int *allThreadsMinusOneD, *allThreadsMinusOneH;

  // Allocate device memory
  HIP_CHECK(hipMalloc((void**)&syncTestD, nBytes));
  HIP_CHECK(hipMalloc((void**)&allThreadsZeroD, nBytes));
  HIP_CHECK(hipMalloc((void**)&allThreadsOneD, nBytes));
  HIP_CHECK(hipMalloc((void**)&oneThreadOneD, nBytes));
  HIP_CHECK(hipMalloc((void**)&allThreadsMinusOneD, nBytes));

  // Allocate host memory
  HIP_CHECK(hipHostMalloc((void**)&syncTestH, nBytes));
  HIP_CHECK(hipHostMalloc((void**)&allThreadsZeroH, nBytes));
  HIP_CHECK(hipHostMalloc((void**)&allThreadsOneH, nBytes));
  HIP_CHECK(hipHostMalloc((void**)&oneThreadOneH, nBytes));
  HIP_CHECK(hipHostMalloc((void**)&allThreadsMinusOneH, nBytes));

  // Launch Kernel
  hipLaunchKernelGGL(kernel_syncthreads_or, 2, blockSize, 0, 0, syncTestD, allThreadsZeroD,
                     allThreadsOneD, oneThreadOneD, allThreadsMinusOneD);
  HIP_CHECK(hipGetLastError());
  // Copy result from device to host
  HIP_CHECK(hipMemcpy(syncTestH, syncTestD, nBytes, hipMemcpyDeviceToHost));
  HIP_CHECK(hipMemcpy(allThreadsZeroH, allThreadsZeroD, nBytes, hipMemcpyDeviceToHost));
  HIP_CHECK(hipMemcpy(allThreadsOneH, allThreadsOneD, nBytes, hipMemcpyDeviceToHost));
  HIP_CHECK(hipMemcpy(oneThreadOneH, oneThreadOneD, nBytes, hipMemcpyDeviceToHost));
  HIP_CHECK(hipMemcpy(allThreadsMinusOneH, allThreadsMinusOneD, nBytes, hipMemcpyDeviceToHost));

  // Validate results for both blocks together
  for (int i = 0; i < 2 * blockSize; ++i) {
    REQUIRE(syncTestH[i] == 10);
    REQUIRE(allThreadsZeroH[i] == 0);
    REQUIRE(allThreadsOneH[i] == 1);
    REQUIRE(oneThreadOneH[i] == 1);
    REQUIRE(allThreadsMinusOneH[i] == 1);
  }

  // Free device memory
  HIP_CHECK(hipFree(syncTestD));
  HIP_CHECK(hipFree(allThreadsZeroD));
  HIP_CHECK(hipFree(allThreadsOneD));
  HIP_CHECK(hipFree(oneThreadOneD));
  HIP_CHECK(hipFree(allThreadsMinusOneD));

  // Free host memory
  HIP_CHECK(hipHostFree(syncTestH));
  HIP_CHECK(hipHostFree(allThreadsZeroH));
  HIP_CHECK(hipHostFree(allThreadsOneH));
  HIP_CHECK(hipHostFree(oneThreadOneH));
  HIP_CHECK(hipHostFree(allThreadsMinusOneH));
}


HIP_TEST_CASE(Unit_syncthreads_or) {
  int blockSizes[] = {10, 40, 70, 130, 240, 723, 32, 64, 128, 256, 512, 1024};
  for (unsigned long i = 0; i < (sizeof(blockSizes) / sizeof(blockSizes[0])); ++i)
    test_syncthreads_or(blockSizes[i]);
}