parallel-gx1/Source/Core/VideoBackends/Vulkan/VKGfx.cpp

// Copyright 2016 Dolphin Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later

#include "VideoBackends/Vulkan/VKGfx.h"

#include <cstddef>
#include <cstdio>
#include <utility>

#include "Common/CommonTypes.h"
#include "Common/Logging/Log.h"
#include "Common/MsgHandler.h"

#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/ObjectCache.h"
#include "VideoBackends/Vulkan/StateTracker.h"
#include "VideoBackends/Vulkan/VKPipeline.h"
#include "VideoBackends/Vulkan/VKShader.h"
#include "VideoBackends/Vulkan/VKTexture.h"
#include "VideoBackends/Vulkan/VKVertexFormat.h"

#include "VideoCommon/DriverDetails.h"
#include "VideoCommon/FrameOutput.h"
#include "VideoCommon/FramebufferManager.h"
#include "VideoCommon/Present.h"
#include "VideoCommon/RenderState.h"
#include "VideoCommon/VideoConfig.h"
#include "VideoCommon/VideoBackendKind.h"

namespace Vulkan
{
VKGfx::VKGfx(FrameOutput* output, float backbuffer_scale)
    : m_output(output), m_backbuffer_scale(backbuffer_scale)
{
  g_video_backend_kind = VIDEO_BACKEND_VULKAN;
  UpdateActiveConfig();
  for (SamplerState& sampler_state : m_sampler_states)
    sampler_state = RenderState::GetPointSamplerState();

  // Various initialization routines will have executed commands on the command buffer.
  // Execute what we have done before beginning the first frame.
  ExecuteCommandBuffer(true, false);
}

VKGfx::~VKGfx()
{
  FrameOutputDelete(m_output);
}

bool VKGfx::IsHeadless() const
{
  return m_output == nullptr;
}

AbstractTexture* VKGfx::CreateTexture(const TextureConfig& config,
                                                      std::string_view name)
{
  return VKTexture::Create(config, name);
}

AbstractStagingTexture* VKGfx::CreateStagingTexture(StagingTextureType type,
                                                                    const TextureConfig& config)
{
  return VKStagingTexture::Create(type, config);
}

AbstractShader*
VKGfx::CreateShaderFromSource(ShaderStage stage, ShaderCode* source,
                              std::string_view name)
{
  return VKShader::CreateFromSource(stage, source, name);
}

AbstractShader* VKGfx::CreateShaderFromBuild(ShaderStage stage, shader_build_fn fn,
                                             const void* arg, std::string_view name)
{
  return VKShader::CreateFromBuild(stage, fn, arg, name);
}

AbstractShader* VKGfx::CreateShaderFromBinary(ShaderStage stage, const void* data,
                                                              size_t length, std::string_view name)
{
  return VKShader::CreateFromBinary(stage, data, length, name);
}

NativeVertexFormat*
VKGfx::CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl)
{
  return new VertexFormat(vtx_decl);
}

AbstractPipeline* VKGfx::CreatePipeline(const AbstractPipelineConfig& config,
                                                        const void* cache_data,
                                                        size_t cache_data_length)
{
  return VKPipeline::Create(config);
}

AbstractFramebuffer*
VKGfx::CreateFramebuffer(AbstractTexture* color_attachment, AbstractTexture* depth_attachment)
{
  return VKFramebuffer::Create(static_cast<VKTexture*>(color_attachment),
                               static_cast<VKTexture*>(depth_attachment));
}

void VKGfx::SetPipeline(const AbstractPipeline* pipeline)
{
  StateTracker::GetInstance()->SetPipeline(static_cast<const VKPipeline*>(pipeline));
}

/* All of it in the init command buffer, which runs before the draws: no render pass ends for
 * it. The scratch texture is used there alone, and the destination is new or from the pool,
 * whose textures this command buffer's draws do not use, so the layouts each tracks hold. */
bool VKGfx::DecodeTextureAhead(const AbstractShader* shader, u32 groupsize_x, u32 groupsize_y,
                               u32 groups_x, u32 groups_y, AbstractTexture* scratch,
                               AbstractTexture* dst, const MathUtil::Rectangle<int>& rect,
                               u32 dst_level)
{
  const VkCommandBuffer cb = g_command_buffer_mgr->GetCurrentInitCommandBuffer();
  StateTracker* const state = StateTracker::GetInstance();
  VKTexture* const src_texture = static_cast<VKTexture*>(scratch);
  VKTexture* const dst_texture = static_cast<VKTexture*>(dst);
  const VkImageCopy image_copy = {
      {VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1},
      {rect.left, rect.top, 0},
      {VK_IMAGE_ASPECT_COLOR_BIT, dst_level, 0, 1},
      {rect.left, rect.top, 0},
      {static_cast<uint32_t>(rect.GetWidth()), static_cast<uint32_t>(rect.GetHeight()), 1}};

  (void)groupsize_x;
  (void)groupsize_y;
  state->SetImageTexture(0, src_texture->GetView());
  src_texture->TransitionToLayout(cb, VKTexture::ComputeImageLayout::WriteOnly);
  state->SetComputeShader(static_cast<const VKShader*>(shader));
  if (!state->BindComputeIn(cb))
    return false;
  vkCmdDispatch(cb, groups_x, groups_y, 1);
  src_texture->TransitionToLayout(cb, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
  dst_texture->TransitionToLayout(cb, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
  vkCmdCopyImage(cb, src_texture->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                 dst_texture->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &image_copy);
  dst_texture->TransitionToLayout(cb, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
  state->SetImageTexture(0, VK_NULL_HANDLE);
  return true;
}

void VKGfx::SetDrawState(const RasterizationState& rasterization, const DepthState& depth,
                         const BlendingState& blending)
{
  u32 state, blend;
  VKPipeline::GetDrawState(rasterization, depth, blending, &state, &blend);
  StateTracker::GetInstance()->SetDrawState(state, blend);
}

void VKGfx::ClearRegion(const MathUtil::Rectangle<int>& target_rc, bool color_enable,
                        bool alpha_enable, bool z_enable, u32 color, u32 z)
{
  VkRect2D target_vk_rc = {
      {target_rc.left, target_rc.top},
      {static_cast<uint32_t>(target_rc.GetWidth()), static_cast<uint32_t>(target_rc.GetHeight())}};

  // Convert RGBA8 -> floating-point values.
  VkClearValue clear_color_value = {};
  VkClearValue clear_depth_value = {};
  clear_color_value.color.float32[0] = static_cast<float>((color >> 16) & 0xFF) / 255.0f;
  clear_color_value.color.float32[1] = static_cast<float>((color >> 8) & 0xFF) / 255.0f;
  clear_color_value.color.float32[2] = static_cast<float>((color >> 0) & 0xFF) / 255.0f;
  clear_color_value.color.float32[3] = static_cast<float>((color >> 24) & 0xFF) / 255.0f;
  clear_depth_value.depthStencil.depth = static_cast<float>(z & 0xFFFFFF) / 16777216.0f;
  if (!g_backend_info.bSupportsReversedDepthRange)
    clear_depth_value.depthStencil.depth = 1.0f - clear_depth_value.depthStencil.depth;

  // If we're not in a render pass (start of the frame), we can use a clear render pass
  // to discard the data, rather than loading and then clearing.
  bool use_clear_attachments = (color_enable && alpha_enable) || z_enable;
  bool use_clear_render_pass =
      !StateTracker::GetInstance()->InRenderPass() && color_enable && alpha_enable && z_enable;

  // The NVIDIA Vulkan driver causes the GPU to lock up, or throw exceptions if MSAA is enabled,
  // a non-full clear rect is specified, and a clear loadop or vkCmdClearAttachments is used.
  if (g_ActiveConfig.iMultisamples > 1 &&
      DriverDetails::HasBug(DriverDetails::BUG_BROKEN_MSAA_CLEAR))
  {
    use_clear_render_pass = false;
    use_clear_attachments = false;
  }

  // This path cannot be used if the driver implementation doesn't guarantee pixels with no drawn
  // geometry in "this" renderpass won't be cleared
  if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_CLEAR_LOADOP_RENDERPASS))
    use_clear_render_pass = false;

  auto* vk_frame_buffer = static_cast<VKFramebuffer*>(m_current_framebuffer);

  // Fastest path: Use a render pass to clear the buffers.
  if (use_clear_render_pass)
  {
    vk_frame_buffer->SetAndClear(target_vk_rc, clear_color_value, clear_depth_value);
    return;
  }

  // Fast path: Use vkCmdClearAttachments to clear the buffers within a render path
  // We can't use this when preserving alpha but clearing color.
  if (use_clear_attachments)
  {
    /* colour, then depth */
    VkClearAttachment clear_attachments[2];
    u32 num_clear_attachments = 0;
    if (color_enable && alpha_enable)
    {
      VkClearAttachment clear_attachment;
      clear_attachment.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
      clear_attachment.colorAttachment = 0;
      clear_attachment.clearValue = clear_color_value;
      clear_attachments[num_clear_attachments++] = clear_attachment;
      color_enable = false;
      alpha_enable = false;
    }
    if (z_enable)
    {
      VkClearAttachment clear_attachment;
      clear_attachment.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
      clear_attachment.colorAttachment = 0;
      clear_attachment.clearValue = clear_depth_value;
      clear_attachments[num_clear_attachments++] = clear_attachment;
      z_enable = false;
    }
    if (num_clear_attachments)
    {
      VkClearRect vk_rect = {target_vk_rc, 0, g_framebuffer_manager->GetEFBLayers()};
      if (!StateTracker::GetInstance()->IsWithinRenderArea(
              target_vk_rc.offset.x, target_vk_rc.offset.y, target_vk_rc.extent.width,
              target_vk_rc.extent.height))
      {
        StateTracker::GetInstance()->EndClearRenderPass();
      }
      StateTracker::GetInstance()->BeginRenderPass();

      vkCmdClearAttachments(g_command_buffer_mgr->GetCurrentCommandBuffer(),
                            num_clear_attachments, clear_attachments, 1, &vk_rect);
    }
  }

  // Anything left over for the slow path?
  if (!color_enable && !alpha_enable && !z_enable)
    return;

  BaseClearRegion(target_rc, color_enable, alpha_enable, z_enable, color, z);
}

void VKGfx::Flush()
{
  ExecuteCommandBuffer(true, false);
}

void VKGfx::WaitForGPUIdle()
{
  ExecuteCommandBuffer(false, true);
}

bool VKGfx::BindBackbuffer(const ClearColor& clear_color)
{
  AbstractFramebuffer* fb;

  StateTracker::GetInstance()->EndRenderPass();
  g_command_buffer_mgr->WaitForWorkerThreadIdle();

  /* the output makes its image ready to draw into, outside a render pass */
  fb = m_output->Acquire();
  if (!fb)
    return false;
  SetAndClearFramebuffer(fb, clear_color);
  return true;
}

void VKGfx::PresentBackbuffer()
{
  /* the output leaves its image readable, submits and hands it on */
  StateTracker::GetInstance()->EndRenderPass();
  m_output->Present();

  // New cmdbuffer, so invalidate state.
  StateTracker::GetInstance()->InvalidateCachedState();
}

bool VKGfx::PresentTexture(const AbstractTexture* texture, u32 width, u32 height)
{
  bool presented;

  /* the output makes the texture readable outside a render pass, submits
   * on this thread and hands it on */
  StateTracker::GetInstance()->EndRenderPass();
  g_command_buffer_mgr->WaitForWorkerThreadIdle();
  presented = m_output->PresentTexture(texture, width, height);
  StateTracker::GetInstance()->InvalidateCachedState();
  return presented;
}

bool VKGfx::IsTextureHeld(const AbstractTexture* texture)
{
  return m_output->IsTextureHeld(texture);
}

void VKGfx::ExecuteCommandBuffer(bool submit_off_thread, bool wait_for_completion)
{
  StateTracker::GetInstance()->EndRenderPass();

  g_command_buffer_mgr->SubmitCommandBuffer(submit_off_thread, wait_for_completion);

  StateTracker::GetInstance()->InvalidateCachedState();
}

void VKGfx::OnConfigChanged(u32 bits)
{
  BaseOnConfigChanged(bits);

  if (bits & CONFIG_CHANGE_BIT_HOST_CONFIG)
    g_object_cache->ReloadPipelineCache();

  // Wipe sampler cache if force texture filtering or anisotropy changes.
  if (bits & (CONFIG_CHANGE_BIT_ANISOTROPY | CONFIG_CHANGE_BIT_FORCE_TEXTURE_FILTERING))
  {
    ExecuteCommandBuffer(false, true);
    ResetSamplerStates();
  }
}

void VKGfx::BindFramebuffer(VKFramebuffer* fb)
{
  StateTracker::GetInstance()->EndRenderPass();

  // Shouldn't be bound as a texture.
  fb->Unbind();

  fb->TransitionForRender();
  StateTracker::GetInstance()->SetFramebuffer(fb);
  m_current_framebuffer = fb;
}

void VKGfx::SetFramebuffer(AbstractFramebuffer* framebuffer)
{
  if (m_current_framebuffer == framebuffer)
    return;

  VKFramebuffer* vkfb = static_cast<VKFramebuffer*>(framebuffer);
  BindFramebuffer(vkfb);
}

void VKGfx::SetAndDiscardFramebuffer(AbstractFramebuffer* framebuffer)
{
  if (m_current_framebuffer == framebuffer)
    return;

  VKFramebuffer* vkfb = static_cast<VKFramebuffer*>(framebuffer);
  BindFramebuffer(vkfb);

  // If we're discarding, begin the discard pass, then switch to a load pass.
  // This way if the command buffer is flushed, we don't start another discard pass.
  StateTracker::GetInstance()->BeginDiscardRenderPass();
}

void VKGfx::SetAndClearFramebuffer(AbstractFramebuffer* framebuffer, const ClearColor& color_value,
                                   float depth_value)
{
  VKFramebuffer* vkfb = static_cast<VKFramebuffer*>(framebuffer);
  BindFramebuffer(vkfb);

  VkClearValue clear_color_value;
  std::memcpy(clear_color_value.color.float32, color_value.data(),
              sizeof(clear_color_value.color.float32));
  VkClearValue clear_depth_value;
  clear_depth_value.depthStencil.depth = depth_value;
  clear_depth_value.depthStencil.stencil = 0;
  vkfb->SetAndClear(vkfb->GetRect(), clear_color_value, clear_depth_value);
}

void VKGfx::SetTexture(u32 index, const AbstractTexture* texture)
{
  // Texture should always be in SHADER_READ_ONLY layout prior to use.
  // This is so we don't need to transition during render passes.
  const VKTexture* tex = static_cast<const VKTexture*>(texture);
  if (tex)
  {
    if (tex->GetLayout() != VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL)
    {
      if (StateTracker::GetInstance()->InRenderPass())
      {
        WARN_LOG_FMT(VIDEO, "Transitioning image in render pass in VKGfx::SetTexture()");
        StateTracker::GetInstance()->EndRenderPass();
      }

      tex->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
                              VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
    }

    StateTracker::GetInstance()->SetTexture(index, tex->GetView());
  }
  else
  {
    StateTracker::GetInstance()->SetTexture(0, VK_NULL_HANDLE);
  }
}

void VKGfx::SetSamplerState(u32 index, const SamplerState& state)
{
  // Skip lookup if the state hasn't changed.
  if (m_sampler_states[index] == state)
    return;

  // Look up new state and replace in state tracker.
  VkSampler sampler = g_object_cache->GetSampler(state);
  if (sampler == VK_NULL_HANDLE)
  {
    ERROR_LOG_FMT(VIDEO, "Failed to create sampler");
    sampler = g_object_cache->GetPointSampler();
  }

  StateTracker::GetInstance()->SetSampler(index, sampler);
  m_sampler_states[index] = state;
}

void VKGfx::SetComputeImageTexture(u32 index, AbstractTexture* texture, bool read, bool write)
{
  VKTexture* vk_texture = static_cast<VKTexture*>(texture);
  if (vk_texture)
  {
    StateTracker::GetInstance()->EndRenderPass();
    StateTracker::GetInstance()->SetImageTexture(index, vk_texture->GetView());
    vk_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
                                   read ? (write ? VKTexture::ComputeImageLayout::ReadWrite :
                                                   VKTexture::ComputeImageLayout::ReadOnly) :
                                          VKTexture::ComputeImageLayout::WriteOnly);
  }
  else
  {
    StateTracker::GetInstance()->SetImageTexture(index, VK_NULL_HANDLE);
  }
}

void VKGfx::UnbindTexture(const AbstractTexture* texture)
{
  StateTracker::GetInstance()->UnbindTexture(static_cast<const VKTexture*>(texture)->GetView());
}

void VKGfx::ResetSamplerStates()
{
  // Invalidate all sampler states, next draw will re-initialize them.
  for (u32 i = 0; i < m_sampler_states.size(); i++)
  {
    m_sampler_states[i] = RenderState::GetPointSamplerState();
    StateTracker::GetInstance()->SetSampler(i, g_object_cache->GetPointSampler());
  }

  // Invalidate all sampler objects (some will be unused now).
  g_object_cache->ClearSamplerCache();
}

void VKGfx::SetScissorRect(const MathUtil::Rectangle<int>& rc)
{
  VkRect2D scissor = {{rc.left, rc.top},
                      {static_cast<u32>(rc.GetWidth()), static_cast<u32>(rc.GetHeight())}};

  // See Vulkan spec for vkCmdSetScissor:
  // The x and y members of offset must be greater than or equal to 0.
  if (scissor.offset.x < 0)
  {
    scissor.extent.width -= -scissor.offset.x;
    scissor.offset.x = 0;
  }
  if (scissor.offset.y < 0)
  {
    scissor.extent.height -= -scissor.offset.y;
    scissor.offset.y = 0;
  }
  StateTracker::GetInstance()->SetScissor(scissor);
}

void VKGfx::SetViewport(float x, float y, float width, float height, float near_depth,
                        float far_depth)
{
  VkViewport viewport = {x, y, width, height, near_depth, far_depth};
  StateTracker::GetInstance()->SetViewport(viewport);
}

void VKGfx::Draw(u32 base_vertex, u32 num_vertices)
{
  if (!StateTracker::GetInstance()->Bind())
    return;

  vkCmdDraw(g_command_buffer_mgr->GetCurrentCommandBuffer(), num_vertices, 1, base_vertex, 0);
}

void VKGfx::DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex)
{
  if (!StateTracker::GetInstance()->Bind())
    return;

  vkCmdDrawIndexed(g_command_buffer_mgr->GetCurrentCommandBuffer(), num_indices, 1, base_index,
                   base_vertex, 0);
}

void VKGfx::DispatchComputeShader(const AbstractShader* shader, u32 groupsize_x, u32 groupsize_y,
                                  u32 groupsize_z, u32 groups_x, u32 groups_y, u32 groups_z)
{
  StateTracker::GetInstance()->SetComputeShader(static_cast<const VKShader*>(shader));
  if (StateTracker::GetInstance()->BindCompute())
    vkCmdDispatch(g_command_buffer_mgr->GetCurrentCommandBuffer(), groups_x, groups_y, groups_z);
}

SurfaceInfo VKGfx::GetSurfaceInfo() const
{
  return {m_output ? m_output->GetWidth() : 1u, m_output ? m_output->GetHeight() : 0u,
          m_backbuffer_scale,
          m_output ? m_output->GetFormat() : AbstractTextureFormat::Undefined};
}

}  // namespace Vulkan