fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Proper(ish) VdSwap - fixes a bunch of things.

Browse Source 
Caching is working a bit better, now.
This commit is contained in:
Ben Vanik
2014-06-08 21:24:29 -07:00
parent 8337820500
commit 6e76c169d6
11 changed files with 101 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
src/xenia/gpu/resource_cache.cc
View File

@@ -9,6 +9,8 @@
#include <xenia/gpu/resource_cache.h>
#include <algorithm>
using namespace std;
using namespace xe;
@@ -110,6 +112,8 @@ uint64_t ResourceCache::HashRange(const MemoryRange& memory_range) {
}
void ResourceCache::SyncRange(uint32_t address, int length) {
SCOPE_profile_cpu_f("gpu");
// Scan the page table in sync with our resource list. This means
// we have O(n) complexity for updates, though we could definitely
// make this faster/cleaner.
@@ -118,15 +122,12 @@ void ResourceCache::SyncRange(uint32_t address, int length) {
// will not be changing, which allows us to do a foreach(res) and reload
// and then clear the table.
// DISABLED
return;
// total bytes = (512 * 1024 * 1024) / (16 * 1024) = 32768
// each byte = 1 page
// Walk as qwords so we can clear things up faster.
uint64_t* page_table = reinterpret_cast<uint64_t*>(
memory_->Translate(memory_->page_table()));
int page_size = 16 * 1024; // 16KB pages
uint32_t page_size = 16 * 1024; // 16KB pages
uint32_t lo_address = address % 0x20000000;
uint32_t hi_address = lo_address + length;
@@ -134,24 +135,38 @@ void ResourceCache::SyncRange(uint32_t address, int length) {
int start_page = lo_address / page_size;
int end_page = hi_address / page_size;
auto it = paged_resources_.upper_bound(lo_address);
auto end_it = paged_resources_.lower_bound(hi_address);
while (it != end_it) {
const auto& memory_range = it->second->memory_range();
int lo_page = (memory_range.guest_base % 0x20000000) / page_size;
int hi_page = lo_page + (memory_range.length / page_size);
for (int i = lo_page / 8; i <= hi_page / 8; ++i) {
uint64_t page_flags = page_table[i];
if (page_flags) {
// Dirty!
it->second->MarkDirty(i * 8 * page_size, (i * 8 + 7) * page_size);
{
SCOPE_profile_cpu_i("gpu", "SyncRange:mark");
auto it = lo_address > page_size ?
paged_resources_.upper_bound(lo_address - page_size) :
paged_resources_.begin();
auto end_it = paged_resources_.lower_bound(hi_address + page_size);
while (it != end_it) {
const auto& memory_range = it->second->memory_range();
int lo_page = (memory_range.guest_base % 0x20000000) / page_size;
int hi_page = lo_page + (memory_range.length / page_size);
lo_page = std::max(lo_page, start_page);
hi_page = std::min(hi_page, end_page);
if (lo_page > hi_page) {
++it;
continue;
}
for (int i = lo_page / 8; i <= hi_page / 8; ++i) {
uint64_t page_flags = page_table[i];
if (page_flags) {
// Dirty!
it->second->MarkDirty(i * 8 * page_size, (i * 8 + 7) * page_size);
}
}
++it;
}
++it;
}
// Reset page table.
for (auto i = start_page / 8; i <= end_page / 8; ++i) {
page_table[i] = 0;
{
SCOPE_profile_cpu_i("gpu", "SyncRange:reset");
for (auto i = start_page / 8; i <= end_page / 8; ++i) {
page_table[i] = 0;
}
}
}