Skip to content

Commit

Permalink
gsdx ogl: replace eight 8MB PBO with a single fat 64MB PBO
Browse files Browse the repository at this point in the history 
It ought to be the same in performance but code will be easier this way

v2: print the sync status
v3: use a performance print so it doesn't spam the console
  • Loading branch information
@gregory38
gregory38 committed Jun 18, 2016
1 parent 0857902 commit 113c78b
Show file tree
Hide file tree
Showing 2 changed files with 64 additions and 63 deletions.
112 changes: 57 additions & 55 deletions plugins/GSdx/GSTextureOGL.cpp
View file
Expand Up @@ -35,13 +35,14 @@ extern uint64 g_real_texture_upload_byte;
// FIXME OGL4: investigate, only 1 unpack buffer always bound
namespace PboPool {

GLuint m_pool[PBO_POOL_SIZE];
uptr m_offset[PBO_POOL_SIZE];
char* m_map[PBO_POOL_SIZE];
uint32 m_current_pbo = 0;
const uint32 m_pbo_size = 64*1024*1024;
const uint32 m_seg_size = 16*1024*1024;

GLuint m_buffer;
uptr m_offset;
char* m_map;
uint32 m_size;
GLsync m_fence[PBO_POOL_SIZE];
const uint32 m_pbo_size = 8*1024*1024;
GLsync m_fence[m_pbo_size/m_seg_size];

// Option for buffer storage
// XXX: actually does I really need coherent and barrier???
Expand All @@ -52,20 +53,20 @@ namespace PboPool {
const GLbitfield create_flags = common_flags | GL_CLIENT_STORAGE_BIT;

void Init() {
glGenBuffers(countof(m_pool), m_pool);
glGenBuffers(1, &m_buffer);

for (size_t i = 0; i < countof(m_pool); i++) {
BindPbo();
BindPbo();

string pretty_name = "PBO" + to_string(i);
glObjectLabel(GL_BUFFER, m_pool[i], pretty_name.size(), pretty_name.c_str());
glObjectLabel(GL_BUFFER, m_buffer, -1, "PBO");

glBufferStorage(GL_PIXEL_UNPACK_BUFFER, m_pbo_size, NULL, create_flags);
m_map[m_current_pbo] = (char*)glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, m_pbo_size, map_flags);
m_fence[m_current_pbo] = 0;
glBufferStorage(GL_PIXEL_UNPACK_BUFFER, m_pbo_size, NULL, create_flags);
m_map = (char*)glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, m_pbo_size, map_flags);
m_offset = 0;

NextPbo();
for (size_t i = 0; i < countof(m_fence); i++) {
m_fence[i] = 0;
}

UnbindPbo();
}

Expand All @@ -77,69 +78,70 @@ namespace PboPool {
fprintf(stderr, "BUG: PBO too small %d but need %d\n", m_pbo_size, m_size);
}

if (m_offset[m_current_pbo] + m_size >= m_pbo_size) {
//NextPbo(); // For test purpose
NextPboWithSync();
}

// Note: texsubimage will access currently bound buffer
// Pbo ready let's get a pointer
BindPbo();

map = m_map[m_current_pbo] + m_offset[m_current_pbo];
Sync();

map = m_map + m_offset;

return map;
}

void Unmap() {
glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, m_offset[m_current_pbo], m_size);
glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, m_offset, m_size);
}

uptr Offset() {
return m_offset[m_current_pbo];
return m_offset;
}

void Destroy() {
for (size_t i = 0; i < countof(m_pool); i++) {
m_map[i] = NULL;
m_offset[i] = 0;
glDeleteSync(m_fence[i]);
m_map = NULL;
m_offset = 0;

// Don't know if we must do it
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, m_pool[i]);
glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
}
// Don't know if we must do it
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, m_buffer);
glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
glDeleteBuffers(countof(m_pool), m_pool);
}

void BindPbo() {
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, m_pool[m_current_pbo]);
for (size_t i = 0; i < countof(m_fence); i++) {
glDeleteSync(m_fence[i]);
}

glDeleteBuffers(1, &m_buffer);
}

void NextPbo() {
m_current_pbo = (m_current_pbo + 1) & (countof(m_pool)-1);
// Mark new PBO as free
m_offset[m_current_pbo] = 0;
void BindPbo() {
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, m_buffer);
}

void NextPboWithSync() {
m_fence[m_current_pbo] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
NextPbo();
if (m_fence[m_current_pbo]) {
#ifdef ENABLE_OGL_DEBUG_FENCE
GLenum status = glClientWaitSync(m_fence[m_current_pbo], GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
#else
glClientWaitSync(m_fence[m_current_pbo], GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
#endif
glDeleteSync(m_fence[m_current_pbo]);
m_fence[m_current_pbo] = 0;
void Sync() {
uint32 segment_current = m_offset / m_seg_size;
uint32 segment_next = (m_offset + m_size) / m_seg_size;

#ifdef ENABLE_OGL_DEBUG_FENCE
if (status != GL_ALREADY_SIGNALED) {
fprintf(stderr, "GL_PIXEL_UNPACK_BUFFER: Sync Sync! Buffer too small\n");
if (segment_current != segment_next) {
if (segment_next >= countof(m_fence)) {
segment_next = 0;
}
// Align current transfer on the start of the segment
m_offset = m_seg_size * segment_next;

// protect the left segment
m_fence[segment_current] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);

// Check next segment is free
if (m_fence[segment_next]) {
GLenum status = glClientWaitSync(m_fence[segment_next], GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
// Potentially it doesn't work on AMD driver which might always return GL_CONDITION_SATISFIED
if (status != GL_ALREADY_SIGNALED) {
GL_PERF("GL_PIXEL_UNPACK_BUFFER: Sync Sync (%x)! Buffer too small ?", status);
}

glDeleteSync(m_fence[segment_next]);
m_fence[segment_next] = 0;
}
#endif
}
}

Expand All @@ -149,7 +151,7 @@ namespace PboPool {

void EndTransfer() {
// Note: keep offset aligned for SSE/AVX
m_offset[m_current_pbo] = (m_offset[m_current_pbo] + m_size + 63) & ~0x3F;
m_offset += (m_size + 63) & ~0x3F;
}
}

Expand Down
15 changes: 7 additions & 8 deletions plugins/GSdx/GSTextureOGL.h
View file
Expand Up @@ -24,15 +24,14 @@
#include "GSTexture.h"

namespace PboPool {
void BindPbo();
void UnbindPbo();
void NextPbo();
void NextPboWithSync();
inline void BindPbo();
inline void UnbindPbo();
inline void Sync();

char* Map(uint32 size);
void Unmap();
uptr Offset();
void EndTransfer();
inline char* Map(uint32 size);
inline void Unmap();
inline uptr Offset();
inline void EndTransfer();

void Init();
void Destroy();
Expand Down

0 comments on commit 113c78b

Please sign in to comment.