gl_staging_buffers: Optimization to reduce fence waiting

This commit is contained in:
ameerj 2023-05-27 22:04:16 -04:00
parent 642c14f0c7
commit cb0a410907
2 changed files with 22 additions and 4 deletions

View File

@ -9,8 +9,12 @@
#include "common/alignment.h" #include "common/alignment.h"
#include "common/assert.h" #include "common/assert.h"
#include "common/bit_util.h"
#include "common/microprofile.h"
#include "video_core/renderer_opengl/gl_staging_buffer_pool.h" #include "video_core/renderer_opengl/gl_staging_buffer_pool.h"
MICROPROFILE_DEFINE(OpenGL_BufferRequest, "OpenGL", "BufferRequest", MP_RGB(128, 128, 192));
namespace OpenGL { namespace OpenGL {
StagingBufferMap::~StagingBufferMap() { StagingBufferMap::~StagingBufferMap() {
@ -25,8 +29,11 @@ StagingBuffers::StagingBuffers(GLenum storage_flags_, GLenum map_flags_)
StagingBuffers::~StagingBuffers() = default; StagingBuffers::~StagingBuffers() = default;
StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence) { StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence) {
MICROPROFILE_SCOPE(OpenGL_BufferRequest);
const size_t index = RequestBuffer(requested_size); const size_t index = RequestBuffer(requested_size);
OGLSync* const sync = insert_fence ? &syncs[index] : nullptr; OGLSync* const sync = insert_fence ? &syncs[index] : nullptr;
sync_indices[index] = insert_fence ? ++current_sync_index : 0;
return StagingBufferMap{ return StagingBufferMap{
.mapped_span = std::span(maps[index], requested_size), .mapped_span = std::span(maps[index], requested_size),
.sync = sync, .sync = sync,
@ -41,13 +48,14 @@ size_t StagingBuffers::RequestBuffer(size_t requested_size) {
OGLBuffer& buffer = buffers.emplace_back(); OGLBuffer& buffer = buffers.emplace_back();
buffer.Create(); buffer.Create();
glNamedBufferStorage(buffer.handle, requested_size, nullptr, const auto next_pow2_size = Common::NextPow2(requested_size);
glNamedBufferStorage(buffer.handle, next_pow2_size, nullptr,
storage_flags | GL_MAP_PERSISTENT_BIT); storage_flags | GL_MAP_PERSISTENT_BIT);
maps.push_back(static_cast<u8*>(glMapNamedBufferRange(buffer.handle, 0, requested_size, maps.push_back(static_cast<u8*>(glMapNamedBufferRange(buffer.handle, 0, next_pow2_size,
map_flags | GL_MAP_PERSISTENT_BIT))); map_flags | GL_MAP_PERSISTENT_BIT)));
syncs.emplace_back(); syncs.emplace_back();
sizes.push_back(requested_size); sync_indices.emplace_back();
sizes.push_back(next_pow2_size);
ASSERT(syncs.size() == buffers.size() && buffers.size() == maps.size() && ASSERT(syncs.size() == buffers.size() && buffers.size() == maps.size() &&
maps.size() == sizes.size()); maps.size() == sizes.size());
@ -56,6 +64,7 @@ size_t StagingBuffers::RequestBuffer(size_t requested_size) {
} }
std::optional<size_t> StagingBuffers::FindBuffer(size_t requested_size) { std::optional<size_t> StagingBuffers::FindBuffer(size_t requested_size) {
size_t known_unsignaled_index = current_sync_index + 1;
size_t smallest_buffer = std::numeric_limits<size_t>::max(); size_t smallest_buffer = std::numeric_limits<size_t>::max();
std::optional<size_t> found; std::optional<size_t> found;
const size_t num_buffers = sizes.size(); const size_t num_buffers = sizes.size();
@ -65,7 +74,14 @@ std::optional<size_t> StagingBuffers::FindBuffer(size_t requested_size) {
continue; continue;
} }
if (syncs[index].handle != 0) { if (syncs[index].handle != 0) {
if (sync_indices[index] >= known_unsignaled_index) {
// This fence is later than a fence that is known to not be signaled
continue;
}
if (!syncs[index].IsSignaled()) { if (!syncs[index].IsSignaled()) {
// Since this fence hasn't been signaled, it's safe to assume all later
// fences haven't been signaled either
known_unsignaled_index = std::min(known_unsignaled_index, sync_indices[index]);
continue; continue;
} }
syncs[index].Release(); syncs[index].Release();

View File

@ -42,8 +42,10 @@ struct StagingBuffers {
std::vector<OGLBuffer> buffers; std::vector<OGLBuffer> buffers;
std::vector<u8*> maps; std::vector<u8*> maps;
std::vector<size_t> sizes; std::vector<size_t> sizes;
std::vector<size_t> sync_indices;
GLenum storage_flags; GLenum storage_flags;
GLenum map_flags; GLenum map_flags;
size_t current_sync_index = 0;
}; };
class StreamBuffer { class StreamBuffer {