liverpool: Use pending downloads API for rewind

Prevents creating huge buffers
This commit is contained in:
IndecisiveTurtle 2025-06-23 01:10:22 +03:00
parent 35c6e57469
commit e580073430
6 changed files with 15 additions and 15 deletions

View File

@ -708,9 +708,7 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
}
case PM4ItOpcode::Rewind: {
if (rasterizer) {
const VAddr flush_addr = VAddr(reinterpret_cast<const u32*>(header));
const u32 flush_size = dcb.size_bytes();
rasterizer->ReadMemory(flush_addr, flush_size);
rasterizer->CommitPendingDownloads(true);
}
const PM4CmdRewind* rewind = reinterpret_cast<const PM4CmdRewind*>(header);
while (!rewind->Valid()) {
@ -915,9 +913,7 @@ Liverpool::Task Liverpool::ProcessCompute(const u32* acb, u32 acb_dwords, u32 vq
}
case PM4ItOpcode::Rewind: {
if (rasterizer) {
const VAddr flush_addr = VAddr(reinterpret_cast<const u32*>(header));
const u32 flush_size = acb_dwords * sizeof(u32);
rasterizer->ReadMemory(flush_addr, flush_size);
rasterizer->CommitPendingDownloads(true);
}
const PM4CmdRewind* rewind = reinterpret_cast<const PM4CmdRewind*>(header);
while (!rewind->Valid()) {

View File

@ -27,7 +27,7 @@ static constexpr size_t UboStreamBufferSize = 128_MB;
static constexpr size_t DownloadBufferSize = 128_MB;
static constexpr size_t DeviceBufferSize = 128_MB;
static constexpr size_t MaxPageFaults = 1024;
static constexpr size_t DownloadSizeThreshold = 2_MB;
static constexpr size_t DownloadSizeThreshold = 1_MB;
BufferCache::BufferCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
AmdGpu::Liverpool* liverpool_, TextureCache& texture_cache_,
@ -188,7 +188,7 @@ void BufferCache::DownloadBufferMemory(const Buffer& buffer, VAddr device_addr,
}
}
bool BufferCache::CommitPendingDownloads() {
bool BufferCache::CommitPendingDownloads(bool wait_done) {
if (pending_download_ranges.Empty()) {
return false;
}
@ -254,7 +254,11 @@ bool BufferCache::CommitPendingDownloads() {
}
}
});
scheduler.Flush();
if (wait_done) {
scheduler.Finish();
} else {
scheduler.Flush();
}
return true;
}

View File

@ -128,7 +128,7 @@ public:
void CopyBuffer(VAddr dst, VAddr src, u32 num_bytes, bool dst_gds, bool src_gds);
/// Schedules pending GPU modified ranges since last commit to be copied back the host memory.
bool CommitPendingDownloads();
bool CommitPendingDownloads(bool wait_done);
/// Obtains a buffer for the specified region.
[[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(VAddr gpu_addr, u32 size, bool is_written,

View File

@ -60,9 +60,9 @@ void Rasterizer::CpSync() {
vk::DependencyFlagBits::eByRegion, ib_barrier, {}, {});
}
bool Rasterizer::CommitPendingDownloads() {
bool Rasterizer::CommitPendingDownloads(bool wait_done) {
scheduler.PopPendingOperations();
return buffer_cache.CommitPendingDownloads();
return buffer_cache.CommitPendingDownloads(wait_done);
}
bool Rasterizer::FilterDraw() {

View File

@ -65,7 +65,7 @@ public:
void UnmapMemory(VAddr addr, u64 size);
void CpSync();
bool CommitPendingDownloads();
bool CommitPendingDownloads(bool wait_done = false);
u64 Flush();
void Finish();
void ProcessFaults();

View File

@ -357,8 +357,8 @@ public:
}
/// Defers an operation until the gpu has reached the current cpu tick.
void DeferOperation(Common::UniqueFunction<void>&& func, bool prev_tick = false) {
pending_ops.emplace(std::move(func), prev_tick ? CurrentTick() - 1 : CurrentTick());
void DeferOperation(Common::UniqueFunction<void>&& func) {
pending_ops.emplace(std::move(func), CurrentTick());
}
static std::mutex submit_mutex;