//===- RocmRuntimeWrappers.cpp - MLIR ROCM runtime wrapper library --------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// Implements C wrappers around the ROCM library for easy linking in ORC jit.
// Also adds some debugging helpers that are helpful when writing MLIR code to
// run on GPUs.
//
//===----------------------------------------------------------------------===//

#include <cassert>
#include <cstdint>
#include <cstdio>

#include "mlir/ExecutionEngine/CRunnerUtils.h"

#include "hip/hip_runtime.h"

#ifdef _WIN32
#define MLIR_ROCM_WRAPPERS_EXPORT __declspec(dllexport)
#else
#define MLIR_ROCM_WRAPPERS_EXPORT __attribute__((visibility("default")))
#endif // _WIN32

#define HIP_REPORT_IF_ERROR(expr)                                              \
  [](hipError_t result) {                                                      \
    if (!result)                                                               \
      return;                                                                  \
    const char *name = hipGetErrorName(result);                                \
    if (!name)                                                                 \
      name = "<unknown>";                                                      \
    fprintf(stderr, "'%s' failed with '%s'\n", #expr, name);                   \
  }(expr)

thread_local static int32_t defaultDevice = 0;

extern "C" MLIR_ROCM_WRAPPERS_EXPORT hipModule_t
mgpuModuleLoad(void *data, size_t /*gpuBlobSize*/) {
  hipModule_t module = nullptr;
  HIP_REPORT_IF_ERROR(hipModuleLoadData(&module, data));
  return module;
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT hipModule_t
mgpuModuleLoadJIT(void *data, int optLevel, size_t /*assmeblySize*/) {
  assert(false && "This function is not available in HIP.");
  return nullptr;
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuModuleUnload(hipModule_t module) {
  HIP_REPORT_IF_ERROR(hipModuleUnload(module));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT hipFunction_t
mgpuModuleGetFunction(hipModule_t module, const char *name) {
  hipFunction_t function = nullptr;
  HIP_REPORT_IF_ERROR(hipModuleGetFunction(&function, module, name));
  return function;
}

// The wrapper uses intptr_t instead of ROCM's unsigned int to match
// the type of MLIR's index type. This avoids the need for casts in the
// generated MLIR code.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuLaunchKernel(hipFunction_t function, intptr_t gridX, intptr_t gridY,
                 intptr_t gridZ, intptr_t blockX, intptr_t blockY,
                 intptr_t blockZ, int32_t smem, hipStream_t stream,
                 void **params, void **extra, size_t /*paramsCount*/) {
  HIP_REPORT_IF_ERROR(hipModuleLaunchKernel(function, gridX, gridY, gridZ,
                                            blockX, blockY, blockZ, smem,
                                            stream, params, extra));
}

// Cooperative launch entry point. The cluster dimensions are accepted to
// match the CUDA wrapper signature, but HIP does not support thread block
// clusters; passing nonzero cluster dimensions is a usage error.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuLaunchKernelCooperative(
    hipFunction_t function, intptr_t gridX, intptr_t gridY, intptr_t gridZ,
    intptr_t clusterX, intptr_t clusterY, intptr_t clusterZ, intptr_t blockX,
    intptr_t blockY, intptr_t blockZ, int32_t smem, hipStream_t stream,
    void **params, void ** /*extra*/) {
  if (clusterX != 0 || clusterY != 0 || clusterZ != 0) {
    fprintf(stderr,
            "mgpuLaunchKernelCooperative: HIP does not support thread block "
            "clusters (got cluster=%ld,%ld,%ld)\n",
            clusterX, clusterY, clusterZ);
    abort();
  }
  HIP_REPORT_IF_ERROR(
      hipModuleLaunchCooperativeKernel(function, gridX, gridY, gridZ, blockX,
                                       blockY, blockZ, smem, stream, params));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT hipStream_t mgpuStreamCreate() {
  hipStream_t stream = nullptr;
  HIP_REPORT_IF_ERROR(hipStreamCreate(&stream));
  return stream;
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuStreamDestroy(hipStream_t stream) {
  HIP_REPORT_IF_ERROR(hipStreamDestroy(stream));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuStreamSynchronize(hipStream_t stream) {
  return HIP_REPORT_IF_ERROR(hipStreamSynchronize(stream));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuStreamWaitEvent(hipStream_t stream, hipEvent_t event) {
  HIP_REPORT_IF_ERROR(hipStreamWaitEvent(stream, event, /*flags=*/0));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT hipEvent_t mgpuEventCreate() {
  hipEvent_t event = nullptr;
  HIP_REPORT_IF_ERROR(hipEventCreateWithFlags(&event, hipEventDisableTiming));
  return event;
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuEventDestroy(hipEvent_t event) {
  HIP_REPORT_IF_ERROR(hipEventDestroy(event));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuEventSynchronize(hipEvent_t event) {
  HIP_REPORT_IF_ERROR(hipEventSynchronize(event));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuEventRecord(hipEvent_t event,
                                                          hipStream_t stream) {
  HIP_REPORT_IF_ERROR(hipEventRecord(event, stream));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void *mgpuMemAlloc(uint64_t sizeBytes,
                                                        hipStream_t /*stream*/,
                                                        bool /*isHostShared*/) {
  void *ptr;
  HIP_REPORT_IF_ERROR(hipMalloc(&ptr, sizeBytes));
  return ptr;
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuMemFree(void *ptr,
                                                      hipStream_t /*stream*/) {
  HIP_REPORT_IF_ERROR(hipFree(ptr));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemcpy(void *dst, void *src, size_t sizeBytes, hipStream_t stream) {
  HIP_REPORT_IF_ERROR(
      hipMemcpyAsync(dst, src, sizeBytes, hipMemcpyDefault, stream));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemset32(void *dst, int value, size_t count, hipStream_t stream) {
  HIP_REPORT_IF_ERROR(hipMemsetD32Async(reinterpret_cast<hipDeviceptr_t>(dst),
                                        value, count, stream));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemset16(void *dst, int short value, size_t count, hipStream_t stream) {
  HIP_REPORT_IF_ERROR(hipMemsetD16Async(reinterpret_cast<hipDeviceptr_t>(dst),
                                        value, count, stream));
}

/// Helper functions for writing mlir example code

// Allows to register byte array with the ROCM runtime. Helpful until we have
// transfer functions implemented.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemHostRegister(void *ptr, uint64_t sizeBytes) {
  HIP_REPORT_IF_ERROR(hipHostRegister(ptr, sizeBytes, /*flags=*/0));
}

// Allows to register a MemRef with the ROCm runtime. Helpful until we have
// transfer functions implemented.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemHostRegisterMemRef(int64_t rank, StridedMemRefType<char, 1> *descriptor,
                          int64_t elementSizeBytes) {
  int64_t *sizes = descriptor->sizes;
  [[maybe_unused]] int64_t *strides = &sizes[rank];
  int64_t runningStride = 1;
  // Only densely packed tensors are currently supported.
  for (int64_t i = rank - 1; i >= 0; --i) {
    assert(strides[i] == runningStride && "Mismatch in computed dense strides");
    runningStride *= sizes[i];
  }
  uint64_t sizeBytes = runningStride * elementSizeBytes;

  auto *ptr = descriptor->data + descriptor->offset * elementSizeBytes;
  mgpuMemHostRegister(ptr, sizeBytes);
}

// Allows to unregister byte array with the ROCM runtime. Helpful until we have
// transfer functions implemented.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuMemHostUnregister(void *ptr) {
  HIP_REPORT_IF_ERROR(hipHostUnregister(ptr));
}

// Allows to unregister a MemRef with the ROCm runtime. Helpful until we have
// transfer functions implemented.
extern "C" MLIR_ROCM_WRAPPERS_EXPORT void
mgpuMemHostUnregisterMemRef(int64_t rank,
                            StridedMemRefType<char, 1> *descriptor,
                            int64_t elementSizeBytes) {
  auto ptr = descriptor->data + descriptor->offset * elementSizeBytes;
  mgpuMemHostUnregister(ptr);
}

template <typename T>
void mgpuMemGetDevicePointer(T *hostPtr, T **devicePtr) {
  HIP_REPORT_IF_ERROR(hipSetDevice(0));
  HIP_REPORT_IF_ERROR(
      hipHostGetDevicePointer((void **)devicePtr, hostPtr, /*flags=*/0));
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT StridedMemRefType<float, 1>
mgpuMemGetDeviceMemRef1dFloat(float *allocated, float *aligned, int64_t offset,
                              int64_t size, int64_t stride) {
  float *devicePtr = nullptr;
  mgpuMemGetDevicePointer(aligned, &devicePtr);
  return {devicePtr, devicePtr, offset, {size}, {stride}};
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT StridedMemRefType<int32_t, 1>
mgpuMemGetDeviceMemRef1dInt32(int32_t *allocated, int32_t *aligned,
                              int64_t offset, int64_t size, int64_t stride) {
  int32_t *devicePtr = nullptr;
  mgpuMemGetDevicePointer(aligned, &devicePtr);
  return {devicePtr, devicePtr, offset, {size}, {stride}};
}

extern "C" MLIR_ROCM_WRAPPERS_EXPORT void mgpuSetDefaultDevice(int32_t device) {
  defaultDevice = device;
  HIP_REPORT_IF_ERROR(hipSetDevice(device));
}