[Texture_cache] Better memory handling for devices with lower memory allocations (#233)

Means games like Minecraft Dungeons, Sea of Stars, Luigi Mansion 2, Astroneer, Alan Wake, etc are now playable.
It also cleans up the recent abi.cpp and bindless texture commits a bit.
Everything is in #ifdef ANDROID - The biggest change is CACHING_PAGEBITS = 12.
Without that the way the buffercache grows and joins buffers can cause Android to run out of memory (as you end up with just one big buffer that needs to be copied every time it grows)
Also patches up ffmpeg issues.

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/233
Co-authored-by: JPikachu <jpikachu.eden@gmail.com>
Co-committed-by: JPikachu <jpikachu.eden@gmail.com>

externals/dynarmic/src/dynarmic/backend/x64/abi.cpp

@@ -116,18 +116,13 @@ void ABI_PopCallerSaveRegistersAndAdjustStack(BlockOfCode& code, const std::size
     ABI_PopRegistersAndAdjustStack(code, frame_size, ABI_ALL_CALLER_SAVE);
 }
 
-static consteval size_t ABI_AllCallerSaveSize() noexcept {
+static std::vector<HostLoc> ABI_AllCallerSaveExcept(const std::size_t except) noexcept {
-    return ABI_ALL_CALLER_SAVE.max_size();
+    std::vector<HostLoc> arr;
-}
+    std::remove_copy(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end(), std::back_inserter(arr), static_cast<HostLoc>(except));
-static consteval std::array<HostLoc, ABI_AllCallerSaveSize() - 1> ABI_AllCallerSaveExcept(const std::size_t except) noexcept {
-    std::array<HostLoc, ABI_AllCallerSaveSize() - 1> arr;
-    for(std::size_t i = 0; i < arr.size(); ++i) {
-        arr[i] = static_cast<HostLoc>(i + (i >= except ? 1 : 0));
-    }
     return arr;
 }
 
-alignas(64) static constinit std::array<HostLoc, ABI_AllCallerSaveSize() - 1> ABI_CALLER_SAVED_EXCEPT_TABLE[32] = {
+alignas(64) static std::vector<HostLoc> ABI_CALLER_SAVED_EXCEPT_TABLE[32] = {
     ABI_AllCallerSaveExcept(0),
     ABI_AllCallerSaveExcept(1),
     ABI_AllCallerSaveExcept(2),
@@ -163,24 +158,12 @@ alignas(64) static constinit std::array<HostLoc, ABI_AllCallerSaveSize() - 1> AB
 };
 
 void ABI_PushCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) {
-#ifdef _MSC_VER
-    std::vector<HostLoc> regs;
-    std::remove_copy(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end(), std::back_inserter(regs), exception);
-    ABI_PushRegistersAndAdjustStack(code, 0, regs);
-#else
     ASSUME(size_t(exception) < 32);
     ABI_PushRegistersAndAdjustStack(code, 0, ABI_CALLER_SAVED_EXCEPT_TABLE[size_t(exception)]);
-#endif
 }
 
 void ABI_PopCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) {
-#ifdef _MSC_VER
-    std::vector<HostLoc> regs;
-    std::remove_copy(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end(), std::back_inserter(regs), exception);
-    ABI_PopRegistersAndAdjustStack(code, 0, regs);
-#else
     ASSUME(size_t(exception) < 32);
     ABI_PopRegistersAndAdjustStack(code, 0, ABI_CALLER_SAVED_EXCEPT_TABLE[size_t(exception)]);
-#endif
 }
 }  // namespace Dynarmic::Backend::X64

src/core/arm/nce/arm_nce.cpp

@@ -224,7 +224,7 @@ HaltReason ArmNce::RunThread(Kernel::KThread* thread) {
     if (auto it = post_handlers.find(m_guest_ctx.pc); it != post_handlers.end()) {
         hr = ReturnToRunCodeByTrampoline(thread_params, &m_guest_ctx, it->second);
     } else {
-        hr = ReturnToRunCodeByExceptionLevelChange(m_thread_id, thread_params);
+        hr = ReturnToRunCodeByExceptionLevelChange(m_thread_id, thread_params);  // Android: Use "process handle SIGUSR2 -n true -p true -s false" (and SIGURG) in LLDB when debugging
     }
 
     // Critical section for thread cleanup

src/core/hle/service/sockets/sfdnsres.cpp

@@ -151,7 +151,12 @@ static std::pair<u32, GetAddrInfoError> GetHostByNameRequestImpl(HLERequestConte
     // For now, ignore options, which are in input buffer 1 for GetHostByNameRequestWithOptions.
 
     // Prevent resolution of Nintendo servers
-    if (host.find("srv.nintendo.net") != std::string::npos) {
+    if (host.find("srv.nintendo.net") != std::string::npos ||
+        host.find("battle.net") != std::string::npos ||
+        host.find("microsoft.com") != std::string::npos ||
+        host.find("mojang.com") != std::string::npos ||
+        host.find("xboxlive.com") != std::string::npos ||
+        host.find("minecraftservices.com") != std::string::npos) {
         LOG_WARNING(Network, "Resolution of hostname {} requested, returning EAI_AGAIN", host);
         return {0, GetAddrInfoError::AGAIN};
     }
@@ -268,7 +273,12 @@ static std::pair<u32, GetAddrInfoError> GetAddrInfoRequestImpl(HLERequestContext
     const std::string host = Common::StringFromBuffer(host_buffer);
 
     // Prevent resolution of Nintendo servers
-    if (host.find("srv.nintendo.net") != std::string::npos) {
+    if (host.find("srv.nintendo.net") != std::string::npos ||
+        host.find("battle.net") != std::string::npos ||
+        host.find("microsoft.com") != std::string::npos ||
+        host.find("mojang.com") != std::string::npos ||
+        host.find("xboxlive.com") != std::string::npos ||
+        host.find("minecraftservices.com") != std::string::npos) {
         LOG_WARNING(Network, "Resolution of hostname {} requested, returning EAI_AGAIN", host);
         return {0, GetAddrInfoError::AGAIN};
     }

src/shader_recompiler/frontend/maxwell/translate/impl/warp_shuffle.cpp

@@ -4,6 +4,7 @@
 #include "common/bit_field.h"
 #include "common/common_types.h"
 #include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+#include <fstream>
 
 namespace Shader::Maxwell {
 namespace {
@@ -33,6 +34,17 @@ enum class ShuffleMode : u64 {
     }
 }
 
+bool IsKONA() {
+    std::ifstream machineFile("/sys/devices/soc0/machine");
+    if (machineFile.is_open()) {
+        std::string line;
+        std::getline(machineFile, line);
+        if (line == "KONA")
+            return true;
+    }
+    return false;
+}
+
 void Shuffle(TranslatorVisitor& v, u64 insn, const IR::U32& index, const IR::U32& mask) {
     union {
         u64 insn;
@@ -44,7 +56,10 @@ void Shuffle(TranslatorVisitor& v, u64 insn, const IR::U32& index, const IR::U32
 
     const IR::U32 result{ShuffleOperation(v.ir, v.X(shfl.src_reg), index, mask, shfl.mode)};
     v.ir.SetPred(shfl.pred, v.ir.GetInBoundsFromOp(result));
-    v.X(shfl.dest_reg, result);
+    if (IsKONA())
+        v.X(shfl.dest_reg, v.ir.Imm32(0xffffffff)); // This fixes the freeze for Retroid / Snapdragon SD865
+    else
+        v.X(shfl.dest_reg, result);
 }
 } // Anonymous namespace
 

src/shader_recompiler/ir_opt/texture_pass.cpp

@@ -327,30 +327,14 @@ std::optional<ConstBufferAddr> TryGetConstBuffer(const IR::Inst* inst, Environme
     };
 }
 
-// TODO:xbzk: shall be dropped when Track method cover all bindless stuff
-static ConstBufferAddr last_valid_addr = ConstBufferAddr{
-    .index = 0,
-    .offset = 0,
-    .shift_left = 0,
-    .secondary_index = 0,
-    .secondary_offset = 0,
-    .secondary_shift_left = 0,
-    .dynamic_offset = {},
-    .count = 1,
-    .has_secondary = false,
-};
-
 TextureInst MakeInst(Environment& env, IR::Block* block, IR::Inst& inst) {
-    ConstBufferAddr addr;
+    ConstBufferAddr addr{};
     if (IsBindless(inst)) {
         const std::optional<ConstBufferAddr> track_addr{Track(inst.Arg(0), env)};
-
         if (!track_addr) {
-            //throw NotImplementedException("Failed to track bindless texture constant buffer");
+            LOG_WARNING(Shader, "Failed to track bindless texture constant buffer");
-            addr = last_valid_addr; // TODO:xbzk: shall be dropped when Track method cover all bindless stuff
         } else {
             addr = *track_addr;
-            last_valid_addr = addr; // TODO:xbzk: shall be dropped when Track method cover all bindless stuff
         }
     } else {
         addr = ConstBufferAddr{

src/video_core/buffer_cache/buffer_cache.h

@@ -50,7 +50,7 @@ BufferCache<P>::~BufferCache() = default;
 template <class P>
 void BufferCache<P>::RunGarbageCollector() {
     const bool aggressive_gc = total_used_memory >= critical_memory;
-    const u64 ticks_to_destroy = aggressive_gc ? 60 : 120;
+    const u64 ticks_to_destroy = aggressive_gc ? 60 : 150;
     int num_iterations = aggressive_gc ? 64 : 32;
     const auto clean_up = [this, &num_iterations](BufferId buffer_id) {
         if (num_iterations == 0) {
@@ -1380,6 +1380,9 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
     });
     new_buffer.MarkUsage(copies[0].dst_offset, copies[0].size);
     runtime.CopyBuffer(new_buffer, overlap, copies, true);
+#ifdef ANDROID
+    runtime.Finish();
+#endif
     DeleteBuffer(overlap_id, true);
 }
 
@@ -1671,7 +1674,12 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
     }
 
     Unregister(buffer_id);
+    
+#ifdef ANDROID
+    if (!do_not_mark)
+#endif
     delayed_destruction_ring.Push(std::move(slot_buffers[buffer_id]));
+    
     slot_buffers.erase(buffer_id);
 
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {

src/video_core/buffer_cache/buffer_cache_base.h

@@ -156,7 +156,11 @@ template <class P>
 class BufferCache : public VideoCommon::ChannelSetupCaches<BufferCacheChannelInfo> {
     // Page size for caching purposes.
     // This is unrelated to the CPU page size and it can be changed as it seems optimal.
+#ifdef ANDROID
+    static constexpr u32 CACHING_PAGEBITS = 12;
+#else
     static constexpr u32 CACHING_PAGEBITS = 16;
+#endif
     static constexpr u64 CACHING_PAGESIZE = u64{1} << CACHING_PAGEBITS;
 
     static constexpr bool IS_OPENGL = P::IS_OPENGL;
@@ -170,9 +174,15 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<BufferCacheChannelInf
     static constexpr bool SEPARATE_IMAGE_BUFFERS_BINDINGS = P::SEPARATE_IMAGE_BUFFER_BINDINGS;
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = P::USE_MEMORY_MAPS_FOR_UPLOADS;
 
+#ifdef ANDROID
+    static constexpr s64 DEFAULT_EXPECTED_MEMORY = 512_MiB;
+    static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB;
+    static constexpr s64 TARGET_THRESHOLD = 3_GiB;
+#else
     static constexpr s64 DEFAULT_EXPECTED_MEMORY = 512_MiB;
     static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB;
     static constexpr s64 TARGET_THRESHOLD = 4_GiB;
+#endif
 
     // Debug Flags.
 
@@ -448,7 +458,12 @@ private:
     Tegra::MaxwellDeviceMemoryManager& device_memory;
 
     Common::SlotVector<Buffer> slot_buffers;
-    DelayedDestructionRing<Buffer, 8> delayed_destruction_ring;
+#ifdef ANDROID
+    static constexpr size_t TICKS_TO_DESTROY = 2;
+#else
+    static constexpr size_t TICKS_TO_DESTROY = 8;
+#endif
+    DelayedDestructionRing<Buffer, TICKS_TO_DESTROY> delayed_destruction_ring;
 
     const Tegra::Engines::DrawManager::IndirectParams* current_draw_indirect{};
 

src/video_core/cdma_pusher.cpp

@@ -20,6 +20,9 @@ CDmaPusher::CDmaPusher(Host1x::Host1x& host1x_, s32 id)
     : host1x{host1x_}, memory_manager{host1x.GMMU()},
       host_processor{std::make_unique<Host1x::Control>(host1x_)}, current_class{
                                                                       static_cast<ChClassId>(id)} {
+#ifdef ANDROID
+    std::this_thread::sleep_for(std::chrono::milliseconds{500}); // HACK: Fix for Astroneer - doesn't always start without this delay. Happens on Windows too (but rarer)
+#endif
     thread = std::jthread([this](std::stop_token stop_token) { ProcessEntries(stop_token); });
 }
 

src/video_core/delayed_destruction_ring.h

@@ -22,7 +22,6 @@ public:
     void Push(T&& object) {
         elements[index].push_back(std::move(object));
     }
-
 private:
     size_t index = 0;
     std::array<std::vector<T>, TICKS_TO_DESTROY> elements;

src/video_core/host1x/ffmpeg/ffmpeg.cpp

@@ -237,16 +237,23 @@ std::shared_ptr<Frame> DecoderContext::ReceiveFrame() {
     if (m_codec_context->hw_device_ctx) {
         // If we have a hardware context, make a separate frame here to receive the
         // hardware result before sending it to the output.
-        Frame intermediate_frame;
+        std::shared_ptr<Frame> intermediate_frame = std::make_shared<Frame>();
 
-        if (!receive(intermediate_frame.GetFrame())) {
+        if (!receive(intermediate_frame->GetFrame())) {
             return {};
 		}
 		
-		m_temp_frame->SetFormat(PreferredGpuFormat);
+        const auto fmt = intermediate_frame->GetPixelFormat();
-        if (int ret = av_hwframe_transfer_data(m_temp_frame->GetFrame(), intermediate_frame.GetFrame(), 0); ret < 0) {
+        const auto desc = av_pix_fmt_desc_get(fmt);
-            LOG_ERROR(HW_GPU, "av_hwframe_transfer_data error: {}", AVError(ret));
+        if (desc && (desc->flags & AV_PIX_FMT_FLAG_HWACCEL)) {
-            return {};
+		    m_temp_frame->SetFormat(PreferredGpuFormat);
+            if (int ret = av_hwframe_transfer_data(m_temp_frame->GetFrame(), intermediate_frame->GetFrame(), 0); ret < 0) {
+                LOG_ERROR(HW_GPU, "av_hwframe_transfer_data error: {}", AVError(ret));
+                return {};
+            }
+        }
+        else {
+            m_temp_frame = std::move(intermediate_frame);
         }
     } else {
         // Otherwise, decode the frame as normal.

src/video_core/host1x/ffmpeg/ffmpeg.h

@@ -21,6 +21,7 @@ extern "C" {
 
 #include <libavcodec/avcodec.h>
 #include <libavutil/opt.h>
+#include <libavutil/pixdesc.h>
 #include <libavcodec/codec_internal.h>
 
 #if defined(__GNUC__) || defined(__clang__)

src/video_core/renderer_vulkan/present/layer.cpp

@@ -137,6 +137,7 @@ void Layer::CreateStagingBuffer(const Tegra::FramebufferConfig& framebuffer) {
         .pQueueFamilyIndices = nullptr,
     };
 
+    buffer.reset();
     buffer = memory_allocator.CreateBuffer(ci, MemoryUsage::Upload);
 }
 

src/video_core/texture_cache/texture_cache.h

@@ -78,6 +78,9 @@ void TextureCache<P>::RunGarbageCollector() {
     size_t num_iterations = 0;
 
     const auto Configure = [&](bool allow_aggressive) {
+#ifdef ANDROID
+        high_priority_mode = true;
+#endif
         high_priority_mode = total_used_memory >= expected_memory;
         aggressive_mode = allow_aggressive && total_used_memory >= critical_memory;
         ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 50ULL;

src/video_core/texture_cache/texture_cache_base.h

@@ -107,10 +107,17 @@ class TextureCache : public VideoCommon::ChannelSetupCaches<TextureCacheChannelI
 
     static constexpr size_t UNSET_CHANNEL{std::numeric_limits<size_t>::max()};
 
+    #ifdef ANDROID
+    static constexpr s64 TARGET_THRESHOLD = 3_GiB;
+    static constexpr s64 DEFAULT_EXPECTED_MEMORY = 1_GiB + 125_MiB;
+    static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB + 625_MiB;
+    static constexpr size_t GC_EMERGENCY_COUNTS = 2;
+    #else
     static constexpr s64 TARGET_THRESHOLD = 4_GiB;
     static constexpr s64 DEFAULT_EXPECTED_MEMORY = 1_GiB + 125_MiB;
     static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB + 625_MiB;
     static constexpr size_t GC_EMERGENCY_COUNTS = 2;
+    #endif
 
     using Runtime = typename P::Runtime;
     using Image = typename P::Image;
@@ -476,7 +483,11 @@ private:
     };
     Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
 
+ #ifdef ANDROID
+    static constexpr size_t TICKS_TO_DESTROY = 2;
+ #else
     static constexpr size_t TICKS_TO_DESTROY = 8;
+#endif
     DelayedDestructionRing<Image, TICKS_TO_DESTROY> sentenced_images;
     DelayedDestructionRing<ImageView, TICKS_TO_DESTROY> sentenced_image_view;
     DelayedDestructionRing<Framebuffer, TICKS_TO_DESTROY> sentenced_framebuffers;

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -268,7 +268,10 @@ vk::Buffer MemoryAllocator::CreateBuffer(const VkBufferCreateInfo& ci, MemoryUsa
     VmaAllocation allocation{};
     VkMemoryPropertyFlags property_flags{};
 
-    vk::Check(vmaCreateBuffer(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info));
+    VkResult result = vmaCreateBuffer(allocator, &ci, &alloc_ci, &handle, &allocation, &alloc_info);
+    if (result == VK_ERROR_OUT_OF_DEVICE_MEMORY) {
+        LOG_ERROR(Render_Vulkan, "Out of memory creating buffer (size: {})", ci.size);
+    }
     vmaGetAllocationMemoryProperties(allocator, allocation, &property_flags);
 
     u8* data = reinterpret_cast<u8*>(alloc_info.pMappedData);