Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Rendervulkan.cpp: Use less conservative (memory and execution) barriers #1108

Closed
wants to merge 15 commits into from
Closed

Rendervulkan.cpp: Use less conservative (memory and execution) barriers #1108

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
54eb009
rendervulkan: Use somewhat less conservative exec & memory barriers
sharkautarch Jan 20, 2024
eb000f1
Merge branch 'master' into rendervulkan_relaxed_barriers
sharkautarch Jan 20, 2024
2d77a07
fixes
sharkautarch Jan 21, 2024
7a43566
minor cleanup
sharkautarch Jan 21, 2024
f487532
fix Write-After-Write (WAW) sync hazard
sharkautarch Jan 21, 2024
b332388
comment out vulkan string helper include
sharkautarch Jan 22, 2024
10bd1f4
remove useless *_READ on .srcAccessMask
sharkautarch Jan 22, 2024
3ad1c95
fix another sync hazard
sharkautarch Jan 25, 2024
b1994b4
slight refactor
sharkautarch Jan 26, 2024
9f56960
fix another sync hazard
sharkautarch Jan 27, 2024
4bd9bf9
handle insertBarrier() in reshade_effect_manager.cpp
sharkautarch Jan 27, 2024
a1d27df
fix unused variable warning and add missing break
sharkautarch Jan 27, 2024
79d9236
insertBarrier() use VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT instea…
sharkautarch Jan 27, 2024
fb98ddc
fixup reshade insertBarrier()/fix reshade sync hazard
sharkautarch Jan 27, 2024
a268a52
fix vulkan validation warning about layout
sharkautarch Jan 27, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
181 changes: 160 additions & 21 deletions src/rendervulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@
#include <bitset>
#include <thread>
#include <vulkan/vulkan_core.h>
//#include <vulkan/vk_enum_string_helper.h>

#if defined(__linux__)
#include <sys/sysmacros.h>
Expand Down Expand Up @@ -1322,7 +1323,12 @@ void CVulkanCmdBuffer::begin()

void CVulkanCmdBuffer::end()
{
insertBarrier(true);
const barrier_info_t barrier_info = {
.task_type = pipeline_task::end
};

insertBarrier(&barrier_info);

vk_check( m_device->vk.EndCommandBuffer(m_cmdBuffer) );
}

Expand Down Expand Up @@ -1393,7 +1399,7 @@ void CVulkanCmdBuffer::bindPipeline(VkPipeline pipeline)
m_device->vk.CmdBindPipeline(m_cmdBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
}

void CVulkanCmdBuffer::dispatch(uint32_t x, uint32_t y, uint32_t z)
void CVulkanCmdBuffer::dispatch(uint32_t x, uint32_t y, uint32_t z, unsigned int total_dispatches, unsigned int curr_dispatch_no)
{
for (auto src : m_boundTextures)
{
Expand All @@ -1402,7 +1408,16 @@ void CVulkanCmdBuffer::dispatch(uint32_t x, uint32_t y, uint32_t z)
}
assert(m_target != nullptr);
prepareDestImage(m_target);
insertBarrier();

const barrier_info_t barrier_info = {
.task_type = pipeline_task::shader,
.shader_sync_info = {
.curr_sync_point = curr_dispatch_no,
.total_sync_points = total_dispatches
}
};

insertBarrier(&barrier_info);

VkDescriptorSet descriptorSet = m_device->descriptorSet();

Expand Down Expand Up @@ -1555,7 +1570,12 @@ void CVulkanCmdBuffer::copyImage(std::shared_ptr<CVulkanTexture> src, std::share
m_textureRefs.emplace(dst.get(), dst);
prepareSrcImage(src.get());
prepareDestImage(dst.get());
insertBarrier();

const barrier_info_t barrier_info = {
.task_type = pipeline_task::copy,
};

insertBarrier(&barrier_info);

VkImageCopy region = {
.srcSubresource = {
Expand All @@ -1582,7 +1602,13 @@ void CVulkanCmdBuffer::copyBufferToImage(VkBuffer buffer, VkDeviceSize offset, u
{
m_textureRefs.emplace(dst.get(), dst);
prepareDestImage(dst.get());
insertBarrier();

const barrier_info_t barrier_info = {
.task_type = pipeline_task::copy,
};

insertBarrier(&barrier_info);

VkBufferImageCopy region = {
.bufferOffset = offset,
.bufferRowLength = stride,
Expand Down Expand Up @@ -1641,7 +1667,7 @@ void CVulkanCmdBuffer::markDirty(CVulkanTexture *image)
result->second.dirty = true;
}

void CVulkanCmdBuffer::insertBarrier(bool flush)
void CVulkanCmdBuffer::insertBarrier(const barrier_info_t * const barrier_info)
{
std::vector<VkImageMemoryBarrier> barriers;

Expand All @@ -1654,6 +1680,89 @@ void CVulkanCmdBuffer::insertBarrier(bool flush)
.layerCount = 1
};


VkFlags srcStageMask = m_previousCopy ? VK_PIPELINE_STAGE_TRANSFER_BIT : 0;
VkAccessFlags src_write_bits = m_previousCopy ? VK_ACCESS_TRANSFER_WRITE_BIT : 0;

VkFlags dstStageMask = 0;
VkAccessFlags dst_write_bits = 0;
VkAccessFlags dst_read_bits = 0;

bool flush = false;
assert( barrier_info != nullptr);

switch (barrier_info->task_type) {
case (pipeline_task::reshade): {
if (barrier_info->reshade_target == reshade_target::init) {
srcStageMask = dstStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
dst_write_bits = src_write_bits = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
dst_read_bits = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT;
} else {
dstStageMask |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
dst_read_bits |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT;
}
break;
}
case (pipeline_task::shader): {

const bool isFirst = (barrier_info->shader_sync_info.curr_sync_point == 1u);
//const bool isLast = (barrier_info->shader_sync_info.curr_sync_point == barrier_info->shader_sync_info.total_sync_points);
const bool multipleShaders = barrier_info->shader_sync_info.total_sync_points > 1u;

#ifdef DEBUG_BARRIER
printf("\n pipeline_task::shader\n");
printf("\n isFirst = %s, isLast = %s\ncurr_sync_point = %u, total_sync_points = %u\n", isFirst ? "true" : "false", isLast ? "true" : "false", barrier_info->shader_sync_info.curr_sync_point, barrier_info->shader_sync_info.total_sync_points);
#endif

src_write_bits |= (!isFirst ? VK_ACCESS_SHADER_WRITE_BIT : 0);

dst_read_bits = multipleShaders ? VK_ACCESS_SHADER_READ_BIT : 0;
/* ^ TODO: if we ever move to syncronization2, could change dst_read_bits to
* shader sampler read, when multipleShaders == true && isLast == true
*/

dst_write_bits = VK_ACCESS_SHADER_WRITE_BIT;

srcStageMask |= (!isFirst ? VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT : 0);

dstStageMask = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;

break;

} case (pipeline_task::copy): {
#ifdef DEBUG_BARRIER
printf("\n pipeline_task::copy\n");
#endif

dst_read_bits = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
dst_write_bits = VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_SHADER_WRITE_BIT;

dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
break;

} case (pipeline_task::end): {
flush = true;

#ifdef DEBUG_BARRIER
printf("\n pipeline_task::end\n");
#endif
dst_read_bits = dst_write_bits = 0;
srcStageMask |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
break;
}
default:
{
__builtin_unreachable();
break;
}
}

if (srcStageMask == 0)
srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
if (dstStageMask == 0)
dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;


for (auto& pair : m_textureState)
{
CVulkanTexture *image = pair.first;
Expand All @@ -1668,25 +1777,40 @@ void CVulkanCmdBuffer::insertBarrier(bool flush)
if (!state.discarded && !state.dirty && !state.needsImport && !isExport && !isPresent)
continue;

const VkAccessFlags write_bits = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
const VkAccessFlags read_bits = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT;

if (image->queueFamily == VK_QUEUE_FAMILY_IGNORED)
image->queueFamily = m_queueFamily;


const VkAccessFlags src_read_bits = 0u; //*_READ on .srcAccessMask for CmdPipelineBarrier is always the same as a no-op
//https://github.com/KhronosGroup/Vulkan-Docs/issues/131
VkImageMemoryBarrier memoryBarrier =
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.srcAccessMask = state.dirty ? write_bits : 0u,
.dstAccessMask = flush ? 0u : read_bits | write_bits,
.oldLayout = (state.discarded || state.needsImport) ? VK_IMAGE_LAYOUT_UNDEFINED : VK_IMAGE_LAYOUT_GENERAL,
.srcAccessMask = ( state.dirty ? (src_write_bits | src_read_bits) : 0u)
| ( (isPresent && state.dirty) ? VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT : 0u),

.dstAccessMask = dst_read_bits | dst_write_bits,
.oldLayout = ( (state.discarded || state.needsImport) ) ? VK_IMAGE_LAYOUT_UNDEFINED : VK_IMAGE_LAYOUT_GENERAL,
.newLayout = isPresent ? presentLayout : VK_IMAGE_LAYOUT_GENERAL,
.srcQueueFamilyIndex = isExport ? image->queueFamily : state.needsImport ? externalQueue : image->queueFamily,
.dstQueueFamilyIndex = isExport ? externalQueue : state.needsImport ? m_queueFamily : m_queueFamily,
.image = image->vkImage(),
.subresourceRange = subResRange
};

/*#ifdef DEBUG_BARRIER
char buf[256] = ".oldLayout = ";
strcat(buf, string_VkImageLayout(memoryBarrier.oldLayout));
const char * next = "\n.newLayout = ";
strcat(buf, next);
strcat(buf, string_VkImageLayout(memoryBarrier.newLayout));
const char * next2 = "\n";
strcat(buf, next2);
printf(buf);
#endif*/



barriers.push_back(memoryBarrier);

state.discarded = false;
Expand All @@ -1695,8 +1819,11 @@ void CVulkanCmdBuffer::insertBarrier(bool flush)
}

// TODO replace VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
m_device->vk.CmdPipelineBarrier(m_cmdBuffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
// ^ Done ^_^
m_device->vk.CmdPipelineBarrier(m_cmdBuffer, srcStageMask, dstStageMask,
0, 0, nullptr, 0, nullptr, barriers.size(), barriers.data());

m_previousCopy = (barrier_info->task_type == pipeline_task::copy);
}

static CVulkanDevice g_device;
Expand Down Expand Up @@ -3725,6 +3852,18 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh
for (uint32_t i = 0; i < EOTF_Count; i++)
cmdBuffer->bindColorMgmtLuts(i, frameInfo->shaperLut[i], frameInfo->lut3D[i]);



unsigned int total_dispatches = 1;
if ( frameInfo->useFSRLayer0 || frameInfo->useNISLayer0 || frameInfo->blurLayer0 )
total_dispatches = 2;

if ( pPipewireTexture != nullptr )
total_dispatches++;

unsigned int curr_dispatch_no = 1;


if ( frameInfo->useFSRLayer0 )
{
uint32_t inputX = frameInfo->layers[0].tex->width();
Expand All @@ -3745,7 +3884,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh

int pixelsPerGroup = 16;

cmdBuffer->dispatch(div_roundup(tempX, pixelsPerGroup), div_roundup(tempY, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(tempX, pixelsPerGroup), div_roundup(tempY, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);

cmdBuffer->bindPipeline(g_device.pipeline(SHADER_TYPE_RCAS, frameInfo->layerCount, frameInfo->ycbcrMask() & ~1, 0u, frameInfo->colorspaceMask(), outputTF ));
bind_all_layers(cmdBuffer.get(), frameInfo);
Expand All @@ -3756,7 +3895,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh
cmdBuffer->bindTarget(compositeImage);
cmdBuffer->uploadConstants<RcasPushData_t>(frameInfo, g_upscaleFilterSharpness / 10.0f);

cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);
}
else if ( frameInfo->useNISLayer0 )
{
Expand Down Expand Up @@ -3787,7 +3926,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh
int pixelsPerGroupX = 32;
int pixelsPerGroupY = 24;

cmdBuffer->dispatch(div_roundup(tempX, pixelsPerGroupX), div_roundup(tempY, pixelsPerGroupY));
cmdBuffer->dispatch(div_roundup(tempX, pixelsPerGroupX), div_roundup(tempY, pixelsPerGroupY), 1, total_dispatches, curr_dispatch_no++);

struct FrameInfo_t nisFrameInfo = *frameInfo;
nisFrameInfo.layers[0].tex = g_output.tmpOutput;
Expand All @@ -3801,7 +3940,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh

int pixelsPerGroup = 8;

cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);
}
else if ( frameInfo->blurLayer0 )
{
Expand All @@ -3827,7 +3966,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh

int pixelsPerGroup = 8;

cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);

bool useSrgbView = frameInfo->layers[0].colorspace == GAMESCOPE_APP_TEXTURE_COLORSPACE_LINEAR;

Expand All @@ -3840,7 +3979,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh
cmdBuffer->setSamplerUnnormalized(VKR_BLUR_EXTRA_SLOT, true);
cmdBuffer->setSamplerNearest(VKR_BLUR_EXTRA_SLOT, false);

cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);
}
else
{
Expand All @@ -3851,7 +3990,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh

const int pixelsPerGroup = 8;

cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup));
cmdBuffer->dispatch(div_roundup(currentOutputWidth, pixelsPerGroup), div_roundup(currentOutputHeight, pixelsPerGroup), 1, total_dispatches, curr_dispatch_no++);
}

if ( pPipewireTexture != nullptr )
Expand Down Expand Up @@ -3897,7 +4036,7 @@ std::optional<uint64_t> vulkan_composite( struct FrameInfo_t *frameInfo, std::sh
// For ycbcr, we operate on 2 pixels at a time, so use the half-extent.
const int dispatchSize = ycbcr ? pixelsPerGroup * 2 : pixelsPerGroup;

cmdBuffer->dispatch(div_roundup(pPipewireTexture->width(), dispatchSize), div_roundup(pPipewireTexture->height(), dispatchSize));
cmdBuffer->dispatch(div_roundup(pPipewireTexture->width(), dispatchSize), div_roundup(pPipewireTexture->height(), dispatchSize), 1, total_dispatches, curr_dispatch_no++);
}
}

Expand Down
37 changes: 35 additions & 2 deletions src/rendervulkan.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -846,6 +846,37 @@ struct TextureState
}
};

typedef enum class pipeline_task {
reshade,
shader,
copy,
end
} pipeline_task_t;

typedef struct {
unsigned int curr_sync_point;
unsigned int total_sync_points;
} shader_sync_info_t;

typedef enum class reshade_target {
init,
runtime
} reshade_target_t;


typedef struct {
pipeline_task_t task_type;

union {
shader_sync_info_t shader_sync_info;
reshade_target_t reshade_target;
};

} barrier_info_t;

//#define DEBUG_BARRIER 1


class CVulkanCmdBuffer
{
public:
Expand All @@ -871,7 +902,7 @@ class CVulkanCmdBuffer
template<class PushData, class... Args>
void uploadConstants(Args&&... args);
void bindPipeline(VkPipeline pipeline);
void dispatch(uint32_t x, uint32_t y = 1, uint32_t z = 1);
void dispatch(uint32_t x, uint32_t y = 1, uint32_t z = 1, unsigned int total_dispatches = 1, unsigned int curr_dispatch_no = 1);
void copyImage(std::shared_ptr<CVulkanTexture> src, std::shared_ptr<CVulkanTexture> dst);
void copyBufferToImage(VkBuffer buffer, VkDeviceSize offset, uint32_t stride, std::shared_ptr<CVulkanTexture> dst);

Expand All @@ -880,12 +911,14 @@ class CVulkanCmdBuffer
void prepareDestImage(CVulkanTexture *image);
void discardImage(CVulkanTexture *image);
void markDirty(CVulkanTexture *image);
void insertBarrier(bool flush = false);
void insertBarrier(const barrier_info_t * const barrier_info = nullptr);

VkQueue queue() { return m_queue; }
uint32_t queueFamily() { return m_queueFamily; }

private:
bool m_previousCopy = false;

VkCommandBuffer m_cmdBuffer;
CVulkanDevice *m_device;

Expand Down
Loading
Loading