Nyaaori/gamescope
1
0
Fork 0
Code Issues Pull requests Projects Releases 1 Packages Wiki Activity

Switch commit waiting over to polling the DMA-BUF fd.

Browse source 
Remove the implicit sync Vulkan fence hack, it took us 0.2ms of GPU time
to prepare the 1x1 copy to wait on.
This commit is contained in:
Pierre-Loup A. Griffais 2020-08-30 13:26:46 -07:00
parent 165946fad9
commit 6dd7f60540
3 changed files with 32 additions and 48 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

60
src/rendervulkan.cpp
View file

@ -346,6 +346,14 @@ bool CVulkanTexture::BInit( uint32_t width, uint32_t height, VkFormat format, cr
importMemoryInfo.pNext = allocInfo.pNext;
allocInfo.pNext = &importMemoryInfo;
// Take another copy to poll for implicit sync completion
m_FD = dup( pDMA->fd[0] );
if ( m_FD < 0 )
{
perror( "dup failed" );
return false;
}
}
if (vkAllocateMemory(device, &allocInfo, nullptr, &m_vkImageMemory) != VK_SUCCESS) {
@ -452,6 +460,12 @@ bool CVulkanTexture::BInit( uint32_t width, uint32_t height, VkFormat format, cr
CVulkanTexture::~CVulkanTexture( void )
{
if ( m_FD >= 0 )
{
close( m_FD );
m_FD = -1;
}
if ( m_pMappedData != nullptr )
{
vkUnmapMemory( device, g_output.pScreenshotImage->m_vkImageMemory );
@ -742,7 +756,7 @@ int init_device()
bufferCreateInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
bufferCreateInfo.pNext = nullptr;
bufferCreateInfo.size = 512 * 512 * 4;
bufferCreateInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT;
bufferCreateInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
result = vkCreateBuffer( device, &bufferCreateInfo, nullptr, &uploadBuffer );
@ -1263,7 +1277,6 @@ VulkanTexture_t vulkan_create_texture_from_dmabuf( struct wlr_dmabuf_attributes
CVulkanTexture::createFlags texCreateFlags;
texCreateFlags.bTextureable = true;
texCreateFlags.bTransferSrc = true;
if ( pTex->BInit( pDMA->width, pDMA->height, DRMFormatToVulkan( pDMA->format ), texCreateFlags, pDMA ) == false )
{
@ -1764,52 +1777,17 @@ uint32_t vulkan_texture_get_fbid( VulkanTexture_t vulkanTex )
return ret;
}
uint32_t vulkan_get_texture_fence( VulkanTexture_t vulkanTex )
int vulkan_get_texture_fence( VulkanTexture_t vulkanTex )
{
// Queue a trivial rendering operation referencing the texture, then wait for it
// This lets us simulate a "fence" for implicit-sync surfaces maybe
CVulkanTexture *pTex = g_mapVulkanTextures[ vulkanTex ];
if ( pTex == nullptr )
{
assert( 0 );
return -1;
}
VkCommandBuffer commandBuffer;
VkFence fence;
uint32_t handle = get_command_buffer( commandBuffer, &fence );
VkBufferImageCopy region = {};
region.imageSubresource = {
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.layerCount = 1
};
region.imageExtent = {
.width = 1,
.height = 1,
.depth = 1
};
vkCmdCopyImageToBuffer( commandBuffer, pTex->m_vkImage, VK_IMAGE_LAYOUT_GENERAL, uploadBuffer, 1, &region );
std::vector<CVulkanTexture *> refs;
refs.push_back( pTex );
submit_command_buffer( handle, refs );
return handle;
}
void vulkan_wait_for_fence( uint32_t handle )
{
assert( g_scratchCommandBuffers[ handle ].busy == true &&
g_scratchCommandBuffers[ handle ].haswaiter == true );
vkWaitForFences( device, 1, &g_scratchCommandBuffers[ handle ].fence, VK_TRUE, ~0 );
g_scratchCommandBuffers[ handle ].haswaiter = false;
return pTex->m_FD;
}
@ -1883,8 +1861,6 @@ void vulkan_wait_for_fence( uint32_t handle )

5
src/rendervulkan.hpp
View file

@ -87,6 +87,8 @@ public:
uint32_t m_unRowPitch = 0;
uint32_t m_FBID = 0;
int m_FD = -1;
int32_t nRefCount = 1;
@ -102,8 +104,7 @@ int vulkan_init(void);
VulkanTexture_t vulkan_create_texture_from_dmabuf( struct wlr_dmabuf_attributes *pDMA );
VulkanTexture_t vulkan_create_texture_from_bits( uint32_t width, uint32_t height, VkFormat format, void *bits );
uint32_t vulkan_get_texture_fence( VulkanTexture_t vulkanTex );
void vulkan_wait_for_fence( uint32_t );
int vulkan_get_texture_fence( VulkanTexture_t vulkanTex );
uint32_t vulkan_texture_get_fbid( VulkanTexture_t vulkanTex );

15
src/steamcompmgr.cpp
View file

@ -269,7 +269,7 @@ private:
sem waitListSem;
std::mutex waitListLock;
std::vector< std::pair< uint32_t, uint64_t > > waitList;
std::vector< std::pair< int, uint64_t > > waitList;
void imageWaitThreadRun( void )
{
@ -277,7 +277,7 @@ wait:
waitListSem.wait();
bool bFound = false;
uint32_t fence;
int fence;
uint64_t commitID;
retry:
@ -298,7 +298,8 @@ retry:
assert( bFound == true );
gpuvis_trace_printf( "wait fence begin_ctx=%lu\n", commitID );
vulkan_wait_for_fence( fence );
struct pollfd fd = { fence, POLLOUT, 0 };
poll( &fd, 1, -1 );
gpuvis_trace_printf( "wait fence end_ctx=%lu\n", commitID );
{
@ -2155,7 +2156,13 @@ void check_new_wayland_res( void )
w->commit_queue.push_back( newCommit );
}
uint32_t fence = vulkan_get_texture_fence( newCommit.vulkanTex );
int fence = vulkan_get_texture_fence( newCommit.vulkanTex );
if ( fence < 0 )
{
fprintf (stderr, "no fence for texture\n");
continue;
}
gpuvis_trace_printf( "pushing wait for commit %lu\n", newCommit.commitID );
{