WindowServer: Optimize msync calls to synchronize less bytes
This *probably* makes framebuffer much faster :D Window server does not have to do page aligned and page sized syncs which used way too many pixels.
This commit is contained in:
parent
f206e72447
commit
36d07065fb
|
|
@ -80,3 +80,32 @@ struct Circle
|
|||
}
|
||||
|
||||
};
|
||||
|
||||
struct Range
|
||||
{
|
||||
uint32_t start { 0 };
|
||||
uint32_t count { 0 };
|
||||
|
||||
bool is_continuous_with(const Range& range) const
|
||||
{
|
||||
return start <= range.start + range.count && range.start <= start + count;
|
||||
}
|
||||
|
||||
uint32_t distance_between(const Range& range) const
|
||||
{
|
||||
if (is_continuous_with(range))
|
||||
return 0;
|
||||
if (start < range.start)
|
||||
return range.start - (start + count);
|
||||
return start - (range.start + range.count);
|
||||
}
|
||||
|
||||
void merge_with(const Range& range)
|
||||
{
|
||||
const uint32_t new_start = BAN::Math::min(start, range.start);
|
||||
const uint32_t new_end = BAN::Math::max(start + count, range.start + range.count);
|
||||
|
||||
start = new_start;
|
||||
count = new_end - new_start;
|
||||
}
|
||||
};
|
||||
|
|
|
|||
|
|
@ -24,6 +24,8 @@ WindowServer::WindowServer(Framebuffer& framebuffer, int32_t corner_radius)
|
|||
MUST(bitmap.resize(m_framebuffer.width * m_framebuffer.height, { 0x10, 0x10, 0x10, 0xFF }));
|
||||
m_background_image = MUST(BAN::UniqPtr<LibImage::Image>::create(m_framebuffer.width, m_framebuffer.height, BAN::move(bitmap)));
|
||||
|
||||
MUST(m_pending_syncs.resize(m_framebuffer.height));
|
||||
|
||||
invalidate(m_framebuffer.area());
|
||||
}
|
||||
|
||||
|
|
@ -910,24 +912,71 @@ void WindowServer::invalidate(Rectangle area)
|
|||
mark_pending_sync(area);
|
||||
}
|
||||
|
||||
void WindowServer::mark_pending_sync(Rectangle area)
|
||||
void WindowServer::RangeList::add_range(const Range& range)
|
||||
{
|
||||
// FIXME: this marks too many pages
|
||||
|
||||
const uintptr_t mmap_start = reinterpret_cast<uintptr_t>(m_framebuffer.mmap) + area.y * m_framebuffer.width * 4;
|
||||
const uintptr_t mmap_end = mmap_start + (area.height + 1) * m_framebuffer.width * 4;
|
||||
|
||||
uintptr_t mmap_addr = mmap_start & ~(uintptr_t)0xFFF;
|
||||
while (mmap_addr < mmap_end)
|
||||
if (range_count == 0)
|
||||
{
|
||||
size_t index = (mmap_addr - reinterpret_cast<uintptr_t>(m_framebuffer.mmap)) / 4096;
|
||||
size_t byte = index / 8;
|
||||
size_t bit = index % 8;
|
||||
//dprintln("{}/{}", byte, m_pages_to_sync_bitmap.size());
|
||||
if (byte < m_pages_to_sync_bitmap.size())
|
||||
m_pages_to_sync_bitmap[byte] |= 1 << bit;
|
||||
mmap_addr += 4096;
|
||||
ranges[0] = range;
|
||||
range_count++;
|
||||
return;
|
||||
}
|
||||
|
||||
size_t min_distance_value = SIZE_MAX;
|
||||
size_t min_distance_index = 0;
|
||||
for (size_t i = 0; i < range_count; i++)
|
||||
{
|
||||
if (ranges[i].is_continuous_with(range))
|
||||
{
|
||||
ranges[i].merge_with(range);
|
||||
|
||||
size_t last_continuous = i;
|
||||
for (size_t j = i + 1; j < range_count; j++)
|
||||
{
|
||||
if (!ranges[i].is_continuous_with(ranges[j]))
|
||||
break;
|
||||
last_continuous = j;
|
||||
}
|
||||
|
||||
if (last_continuous != i)
|
||||
{
|
||||
ranges[i].merge_with(ranges[last_continuous]);
|
||||
for (size_t j = 1; last_continuous + j < range_count; j++)
|
||||
ranges[i + j] = ranges[last_continuous + j];
|
||||
range_count -= last_continuous - i;
|
||||
}
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
const auto distance = ranges[i].distance_between(range);
|
||||
if (distance < min_distance_value)
|
||||
{
|
||||
min_distance_value = distance;
|
||||
min_distance_index = i;
|
||||
}
|
||||
}
|
||||
|
||||
if (range_count >= ranges.size())
|
||||
{
|
||||
ranges[min_distance_index].merge_with(range);
|
||||
return;
|
||||
}
|
||||
|
||||
size_t insert_idx = 0;
|
||||
for (; insert_idx < range_count; insert_idx++)
|
||||
if (range.start < ranges[insert_idx].start)
|
||||
break;
|
||||
for (size_t i = range_count; i > insert_idx; i--)
|
||||
ranges[i] = ranges[i - 1];
|
||||
ranges[insert_idx] = range;
|
||||
range_count++;
|
||||
}
|
||||
|
||||
void WindowServer::mark_pending_sync(Rectangle to_sync)
|
||||
{
|
||||
ASSERT(to_sync == to_sync.get_overlap(m_framebuffer.area()).value());
|
||||
for (int32_t y_off = 0; y_off < to_sync.height; y_off++)
|
||||
m_pending_syncs[to_sync.y + y_off].add_range({ static_cast<uint32_t>(to_sync.x), static_cast<uint32_t>(to_sync.width) });
|
||||
}
|
||||
|
||||
void WindowServer::sync()
|
||||
|
|
@ -951,33 +1000,44 @@ void WindowServer::sync()
|
|||
dir_y = -dir_y;
|
||||
}
|
||||
|
||||
for (size_t i = 0; i < m_pages_to_sync_bitmap.size() * 8; i++)
|
||||
size_t range_start = 0;
|
||||
size_t range_count = 0;
|
||||
for (int32_t y = 0; y < m_framebuffer.height; y++)
|
||||
{
|
||||
size_t byte = i / 8;
|
||||
size_t bit = i % 8;
|
||||
if (!(m_pages_to_sync_bitmap[byte] & (1 << bit)))
|
||||
continue;
|
||||
auto& range_list = m_pending_syncs[y];
|
||||
|
||||
size_t len = 1;
|
||||
while (i + len < m_pages_to_sync_bitmap.size() * 8)
|
||||
for (size_t i = 0; i < range_list.range_count; i++)
|
||||
{
|
||||
size_t byte = (i + len) / 8;
|
||||
size_t bit = (i + len) % 8;
|
||||
if (!(m_pages_to_sync_bitmap[byte] & (1 << bit)))
|
||||
break;
|
||||
len++;
|
||||
const size_t cur_start = y * m_framebuffer.width + range_list.ranges[i].start;
|
||||
const size_t cur_count = range_list.ranges[i].count;
|
||||
|
||||
if (range_count == 0)
|
||||
{
|
||||
range_start = cur_start;
|
||||
range_count = cur_count;
|
||||
}
|
||||
else
|
||||
{
|
||||
const size_t distance = cur_start - (range_start + range_count);
|
||||
|
||||
// combine nearby ranges to reduce msync calls
|
||||
// NOTE: value of 128 is an arbitary constant that *just* felt nice
|
||||
if (distance <= 128)
|
||||
range_count = (cur_start + cur_count) - range_start;
|
||||
else
|
||||
{
|
||||
msync(m_framebuffer.mmap + range_start, range_count * 4, MS_SYNC);
|
||||
range_start = cur_start;
|
||||
range_count = cur_count;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
msync(
|
||||
reinterpret_cast<uint8_t*>(m_framebuffer.mmap) + i * 4096,
|
||||
len * 4096,
|
||||
MS_SYNC
|
||||
);
|
||||
|
||||
i += len;
|
||||
range_list.range_count = 0;
|
||||
}
|
||||
|
||||
memset(m_pages_to_sync_bitmap.data(), 0, m_pages_to_sync_bitmap.size());
|
||||
if (range_count)
|
||||
msync(m_framebuffer.mmap + range_start, range_count * 4, MS_SYNC);
|
||||
}
|
||||
|
||||
Rectangle WindowServer::cursor_area() const
|
||||
|
|
|
|||
|
|
@ -3,6 +3,7 @@
|
|||
#include "Framebuffer.h"
|
||||
#include "Window.h"
|
||||
|
||||
#include <BAN/Array.h>
|
||||
#include <BAN/Function.h>
|
||||
#include <BAN/Iteration.h>
|
||||
#include <BAN/Vector.h>
|
||||
|
|
@ -59,6 +60,14 @@ public:
|
|||
private:
|
||||
void mark_pending_sync(Rectangle area);
|
||||
|
||||
private:
|
||||
struct RangeList
|
||||
{
|
||||
size_t range_count { 0 };
|
||||
BAN::Array<Range, 32> ranges;
|
||||
void add_range(const Range& range);
|
||||
};
|
||||
|
||||
private:
|
||||
Framebuffer& m_framebuffer;
|
||||
BAN::Vector<BAN::RefPtr<Window>> m_client_windows;
|
||||
|
|
@ -67,7 +76,7 @@ private:
|
|||
|
||||
const int32_t m_corner_radius;
|
||||
|
||||
BAN::Vector<uint8_t> m_pages_to_sync_bitmap;
|
||||
BAN::Vector<RangeList> m_pending_syncs;
|
||||
|
||||
BAN::UniqPtr<LibImage::Image> m_background_image;
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue