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

#include "top.hpp"
#include "os/os.hpp"
#include "device/device.hpp"
#include "rocsettings.hpp"
#include "device/rocm/rocglinterop.hpp"

namespace amd::roc {

// ================================================================================================
Settings::Settings() {
  // Initialize the HSA device default settings

  // Set this to true when we drop the flag
  doublePrecision_ = ::CL_KHR_FP64;

  enableLocalMemory_ = HSA_LOCAL_MEMORY_ENABLE;

  maxWorkGroupSize_ = 1024;
  preferredWorkGroupSize_ = 256;

  kernargPoolSize_ = HSA_KERNARG_POOL_SIZE;

  // Determine if user is requesting Non-Coherent mode
  // for system memory. By default system memory is
  // operates or is programmed to be in Coherent mode.
  // Users can turn it off for hardware that does not
  // support this feature naturally
  char* nonCoherentMode = getenv("OPENCL_USE_NC_MEMORY_POLICY");
  enableNCMode_ = (nonCoherentMode) ? true : false;

  // Disable image DMA by default (ROCM runtime doesn't support it)
  imageDMA_ = false;

  stagedXferSize_ = flagIsDefault(GPU_STAGING_BUFFER_SIZE) ? 1 * Mi : GPU_STAGING_BUFFER_SIZE * Mi;

  // Initialize transfer buffer size to 1MB by default
  xferBufSize_ = 1024 * Ki;

  pinnedXferSize_ = GPU_PINNED_XFER_SIZE * Mi;
  pinnedMinXferSize_ =
      flagIsDefault(GPU_PINNED_MIN_XFER_SIZE) ? 64 * Ki : GPU_PINNED_MIN_XFER_SIZE * Mi;

  sdmaCopyThreshold_ = GPU_FORCE_BLIT_COPY_SIZE * Ki;

  // Don't support Denormals for single precision by default
  singleFpDenorm_ = false;

  apuSystem_ = false;

  // Device enqueuing settings
  numDeviceEvents_ = 1024;
  numWaitEvents_ = 8;

  lcWavefrontSize64_ = true;
  imageBufferWar_ = false;

  sdma_p2p_threshold_ = ROC_P2P_SDMA_SIZE * Ki;
  hmmFlags_ = (!flagIsDefault(ROC_HMM_FLAGS)) ? ROC_HMM_FLAGS : 0;

  rocr_backend_ = true;

  cpu_wait_for_signal_ = ROC_CPU_WAIT_FOR_SIGNAL;
  system_scope_signal_ = ROC_SYSTEM_SCOPE_SIGNAL;

  // Use coarse grain system memory for kernel arguments by default (to keep GPU cache)
  fgs_kernel_arg_ = false;
  barrier_value_packet_ = false;
  ext_dispatch_packet_ = false;
  kernel_arg_impl_ = KernelArgImpl::HostKernelArgs;
  aql_device_ring_buf_ = false;
  gwsInitSupported_ = true;
  limit_blit_wg_ = 16;

  dynamic_queues_ = amd::IS_HIP ? DEBUG_HIP_DYNAMIC_QUEUES : 0;
  // note: OCL user events don't allow CPU blocking calls in DD mode
  blocking_blit_ = amd::IS_HIP;

  max_hw_queues_ = GPU_MAX_HW_QUEUES;

  queue_pipe_dist_ = false;
}

// ================================================================================================
bool Settings::create(bool fullProfile, const amd::Isa& isa, bool enableXNACK, bool coop_groups,
                      bool isXgmi) {
  uint32_t gfxipMajor = isa.versionMajor();
  uint32_t gfxipMinor = isa.versionMinor();
  uint32_t gfxStepping = isa.versionStepping();

  customHostAllocator_ = false;

  if (fullProfile) {
    pinnedXferSize_ = 0;
    stagedXferSize_ = 0;
    xferBufSize_ = 0;
    apuSystem_ = true;
  } else {
    pinnedXferSize_ = std::max(pinnedXferSize_, pinnedMinXferSize_);
  }
  enableXNACK_ = enableXNACK;

  // Enable extensions
  enableExtension(ClKhrByteAddressableStore);
  enableExtension(ClKhrGlobalInt32BaseAtomics);
  enableExtension(ClKhrGlobalInt32ExtendedAtomics);
  enableExtension(ClKhrLocalInt32BaseAtomics);
  enableExtension(ClKhrLocalInt32ExtendedAtomics);
  enableExtension(ClKhrInt64BaseAtomics);
  enableExtension(ClKhrInt64ExtendedAtomics);
  enableExtension(ClKhr3DImageWrites);
  enableExtension(ClAmdMediaOps);
  enableExtension(ClAmdMediaOps2);
  enableExtension(ClKhrImage2dFromBuffer);
  enableExtension(ClKhrGlSharing);

  // Enable platform extension
  enableExtension(ClAmdDeviceAttributeQuery);

  // Enable KHR double precision extension
  enableExtension(ClKhrFp64);
  enableExtension(ClKhrSubGroups);
  enableExtension(ClKhrDepthImages);
  enableExtension(ClAmdCopyBufferP2P);
  enableExtension(ClKhrFp16);
  supportDepthsRGB_ = true;

  enableExtension(ClAmdAssemblyProgram);
  // enable subnormals for gfx900 and later
  if (gfxipMajor >= 9) {
    singleFpDenorm_ = true;
    enableCoopGroups_ = GPU_ENABLE_COOP_GROUPS & coop_groups;
    enableCoopMultiDeviceGroups_ = GPU_ENABLE_COOP_GROUPS & coop_groups;
  }

  if ((gfxipMajor == 9 && gfxipMinor == 0 && gfxStepping == 10) ||
      ((gfxipMajor == 9 && gfxipMinor >= 4 &&
        (gfxStepping == 0 || gfxStepping == 1 || gfxStepping == 2)))) {
    // Enable Barrier Value packet is only for MI2XX/300
    barrier_value_packet_ = true;
    queue_pipe_dist_ = dynamic_queues_ >= 1;
  }

  if ((gfxipMajor == 9 && gfxipMinor >= 4) ||
      (gfxipMajor == 12 && gfxipMinor >= 5)) {
    sdma_swap_supported_ = true;
  }

  setKernelArgImpl(isa, isXgmi);

  if (gfxipMajor >= 10) {
     enableWave32Mode_ = true;
     // Disable wgp mode for gfx1250 and later
     if (gfxipMajor == 12 && gfxipMinor >= 5) {
        enableWgpMode_ = false;
     } else {
        enableWgpMode_ = GPU_ENABLE_WGP_MODE;
     }
     if (gfxipMajor == 10 && gfxipMinor == 1) {
       // GFX10.1 HW doesn't support custom pitch. Enable double copy workaround
       // TODO: This should be updated when ROCr support custom pitch
       imageBufferWar_ = GPU_IMAGE_BUFFER_WAR;
     }
  }

  if (!flagIsDefault(GPU_ENABLE_WAVE32_MODE)) {
    enableWave32Mode_ = GPU_ENABLE_WAVE32_MODE;
  }

  lcWavefrontSize64_ = !enableWave32Mode_;

  if (gfxipMajor > 10 || (gfxipMajor == 9 && gfxipMinor >= 4)) {
    gwsInitSupported_ = false;
  }

  if (GPU_MIPMAP) {
    enableExtension(ClKhrMipMapImage);
    enableExtension(ClKhrMipMapImageWrites);
  }

  if (gfxipMajor == 12 && gfxipMinor >= 5) {
    ext_dispatch_packet_ = true;
    groupMemCarveout_ = true;
  }

  // SDMA indirect copy uses the gfx1250 wait/signal-indirect SDMA
  // packet that dereferences a pointer-to-pointer slot before issuing the
  // copy.
  if (gfxipMajor == 12 && gfxipMinor == 5) {
    sdma_indirect_supported_ = true;
  }

#if defined(_WIN32)
  if (gfxipMajor >= 11) {
    // Due to driver limitation,
    // D3D10/11 sharing extensions are only supported on GFX11 or later,
    // and D3D9 sharing extension is not supported on any hardware as it is deprecated.
    enableExtension(ClKhrD3d10Sharing);
    enableExtension(ClKhrD3d11Sharing);
    enableExtension(ClAmdPlanarYuv);
  }
#endif

  // Override current device settings
  override();

  return true;
}

// ================================================================================================
void Settings::override() {
  // Limit reported workgroup size
  if (GPU_MAX_WORKGROUP_SIZE != 0) {
    preferredWorkGroupSize_ = GPU_MAX_WORKGROUP_SIZE;
  }

  if (!flagIsDefault(GPU_XFER_BUFFER_SIZE)) {
    xferBufSize_ = GPU_XFER_BUFFER_SIZE * Ki;
  }

  if (!flagIsDefault(AMD_GPU_FORCE_SINGLE_FP_DENORM)) {
    switch (AMD_GPU_FORCE_SINGLE_FP_DENORM) {
      case 0:
        singleFpDenorm_ = false;
        break;
      case 1:
        singleFpDenorm_ = true;
        break;
      default:
        break;
    }
  }
  if (!flagIsDefault(GPU_ENABLE_COOP_GROUPS)) {
    enableCoopGroups_ = GPU_ENABLE_COOP_GROUPS;
    enableCoopMultiDeviceGroups_ = GPU_ENABLE_COOP_GROUPS;
  }

  if (!flagIsDefault(ROC_USE_FGS_KERNARG)) {
    fgs_kernel_arg_ = ROC_USE_FGS_KERNARG;
  }

  if (!flagIsDefault(DEBUG_CLR_BLIT_KERNARG_OPT)) {
    kernel_arg_opt_ = DEBUG_CLR_BLIT_KERNARG_OPT;
  }
}

// ================================================================================================
void Settings::setKernelArgImpl(const amd::Isa& isa, bool isXgmi) {
  const uint32_t gfxipMajor = isa.versionMajor();
  const uint32_t gfxipMinor = isa.versionMinor();
  const uint32_t gfxStepping = isa.versionStepping();

  const bool isGfx94x = gfxipMajor == 9 && gfxipMinor >= 4 &&
                        (gfxStepping == 0 || gfxStepping == 1 || gfxStepping == 2);
  const bool isGfx90a = (gfxipMajor == 9 && gfxipMinor == 0 && gfxStepping == 10);
  const bool isGfx125x = (gfxipMajor == 12) && ((gfxipMinor >= 5));

  auto kernelArgImpl = KernelArgImpl::HostKernelArgs;

  if (isXgmi) {
    // The XGMI-connected path does not require the manual memory ordering
    // workarounds that the PCIe connected path requires
    kernelArgImpl = KernelArgImpl::DeviceKernelArgs;
  } else if (isGfx94x || isGfx90a || isGfx125x) {
    // Implement the kernel argument readback workaround
    // (write all args -> sfence -> write last byte -> mfence -> read last byte)
    kernelArgImpl = KernelArgImpl::DeviceKernelArgsReadback;
  }

  // Enable device kernel args by default on gfx94x/gfx125x. Device-memory AQL
  // ring buffers are controlled separately via DEBUG_CLR_AQL_DEV_QUEUE below.
  if (isGfx94x || isGfx125x) {
    kernel_arg_impl_ = kernelArgImpl;
    kernel_arg_opt_ = true;
  }

  if (!flagIsDefault(HIP_FORCE_DEV_KERNARG)) {
    kernel_arg_impl_ = kernelArgImpl & (HIP_FORCE_DEV_KERNARG ? 0xF : 0x0);
  }

  aql_device_ring_buf_ = (DEBUG_CLR_AQL_DEV_QUEUE > 0);
}
}  // namespace amd::roc