"Merge Tagged PR 11385"

src/CMakeLists.txt

@@ -85,6 +85,7 @@ if (MSVC)
         /wd4100 # 'identifier': unreferenced formal parameter
         /wd4324 # 'struct_name': structure was padded due to __declspec(align())
         /wd4201 # nonstandard extension used : nameless struct/union
+        /wd4702 # unreachable code (when used with LTO)
     )
 
     if (USE_CCACHE OR YUZU_USE_PRECOMPILED_HEADERS)

src/audio_core/adsp/apps/audio_renderer/audio_renderer.cpp

@@ -88,8 +88,13 @@ MailboxMessage AudioRenderer::Receive(Direction dir, bool block) {
     return mailbox.Receive(dir, block);
 }
 
-void AudioRenderer::SetCommandBuffer(s32 session_id, CommandBuffer& buffer) noexcept {
-    command_buffers[session_id] = buffer;
+void AudioRenderer::SetCommandBuffer(s32 session_id, CpuAddr buffer, u64 size, u64 time_limit,
+                                     u64 applet_resource_user_id, bool reset) noexcept {
+    command_buffers[session_id].buffer = buffer;
+    command_buffers[session_id].size = size;
+    command_buffers[session_id].time_limit = time_limit;
+    command_buffers[session_id].applet_resource_user_id = applet_resource_user_id;
+    command_buffers[session_id].reset_buffer = reset;
 }
 
 u32 AudioRenderer::GetRemainCommandCount(s32 session_id) const noexcept {

src/audio_core/adsp/apps/audio_renderer/audio_renderer.h

@@ -75,7 +75,8 @@ public:
     void Send(Direction dir, MailboxMessage message);
     MailboxMessage Receive(Direction dir, bool block = true);
 
-    void SetCommandBuffer(s32 session_id, CommandBuffer& buffer) noexcept;
+    void SetCommandBuffer(s32 session_id, CpuAddr buffer, u64 size, u64 time_limit,
+                          u64 applet_resource_user_id, bool reset) noexcept;
     u32 GetRemainCommandCount(s32 session_id) const noexcept;
     void ClearRemainCommandCount(s32 session_id) noexcept;
     u64 GetRenderingStartTick(s32 session_id) const noexcept;

src/audio_core/adsp/apps/audio_renderer/command_list_processor.cpp

@@ -37,11 +37,6 @@ u32 CommandListProcessor::GetRemainingCommandCount() const {
     return command_count - processed_command_count;
 }
 
-void CommandListProcessor::SetBuffer(const CpuAddr buffer, const u64 size) {
-    commands = reinterpret_cast<u8*>(buffer + sizeof(Renderer::CommandListHeader));
-    commands_buffer_size = size;
-}
-
 Sink::SinkStream* CommandListProcessor::GetOutputSinkStream() const {
     return stream;
 }

src/audio_core/adsp/apps/audio_renderer/command_list_processor.h

@@ -56,14 +56,6 @@ public:
      */
     u32 GetRemainingCommandCount() const;
 
-    /**
-     * Set the command buffer.
-     *
-     * @param buffer - The buffer to use.
-     * @param size   - The size of the buffer.
-     */
-    void SetBuffer(CpuAddr buffer, u64 size);
-
     /**
      * Get the stream for this command list.
      *

src/audio_core/renderer/command/data_source/adpcm.cpp

@@ -20,6 +20,12 @@ void AdpcmDataSourceVersion1Command::Process(const AudioRenderer::CommandListPro
     auto out_buffer{processor.mix_buffers.subspan(output_index * processor.sample_count,
                                                   processor.sample_count)};
 
+    for (auto& wave_buffer : wave_buffers) {
+        wave_buffer.loop_start_offset = wave_buffer.start_offset;
+        wave_buffer.loop_end_offset = wave_buffer.end_offset;
+        wave_buffer.loop_count = wave_buffer.loop ? -1 : 0;
+    }
+
     DecodeFromWaveBuffersArgs args{
         .sample_format{SampleFormat::Adpcm},
         .output{out_buffer},

src/audio_core/renderer/command/data_source/decode.cpp

@@ -123,11 +123,13 @@ static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
         return 0;
     }
 
-    auto samples_to_process{
-        std::min(req.end_offset - req.start_offset - req.offset, req.samples_to_read)};
+    auto start_pos{req.start_offset + req.offset};
+    auto samples_to_process{std::min(req.end_offset - start_pos, req.samples_to_read)};
+    if (samples_to_process == 0) {
+        return 0;
+    }
 
     auto samples_to_read{samples_to_process};
-    auto start_pos{req.start_offset + req.offset};
     auto samples_remaining_in_frame{start_pos % SamplesPerFrame};
     auto position_in_frame{(start_pos / SamplesPerFrame) * NibblesPerFrame +
                            samples_remaining_in_frame};
@@ -225,13 +227,24 @@ static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
  * @param args   - The wavebuffer data, and information for how to decode it.
  */
 void DecodeFromWaveBuffers(Core::Memory::Memory& memory, const DecodeFromWaveBuffersArgs& args) {
+    static constexpr auto EndWaveBuffer = [](auto& voice_state, auto& wavebuffer, auto& index,
+                                             auto& played_samples, auto& consumed) -> void {
+        voice_state.wave_buffer_valid[index] = false;
+        voice_state.loop_count = 0;
+
+        if (wavebuffer.stream_ended) {
+            played_samples = 0;
+        }
+
+        index = (index + 1) % MaxWaveBuffers;
+        consumed++;
+    };
     auto& voice_state{*args.voice_state};
     auto remaining_sample_count{args.sample_count};
     auto fraction{voice_state.fraction};
 
-    const auto sample_rate_ratio{
-        (Common::FixedPoint<49, 15>(args.source_sample_rate) / args.target_sample_rate) *
-        args.pitch};
+    const auto sample_rate_ratio{Common::FixedPoint<49, 15>(
+        (f32)args.source_sample_rate / (f32)args.target_sample_rate * (f32)args.pitch)};
     const auto size_required{fraction + remaining_sample_count * sample_rate_ratio};
 
     if (size_required < 0) {
@@ -298,22 +311,23 @@ void DecodeFromWaveBuffers(Core::Memory::Memory& memory, const DecodeFromWaveBuf
             auto end_offset{wavebuffer.end_offset};
 
             if (wavebuffer.loop && voice_state.loop_count > 0 &&
-                wavebuffer.loop_start_offset != 0 && wavebuffer.loop_end_offset != 0 &&
                 wavebuffer.loop_start_offset <= wavebuffer.loop_end_offset) {
                 start_offset = wavebuffer.loop_start_offset;
                 end_offset = wavebuffer.loop_end_offset;
             }
 
-            DecodeArg decode_arg{.buffer{wavebuffer.buffer},
-                                 .buffer_size{wavebuffer.buffer_size},
-                                 .start_offset{start_offset},
-                                 .end_offset{end_offset},
-                                 .channel_count{args.channel_count},
-                                 .coefficients{},
-                                 .adpcm_context{nullptr},
-                                 .target_channel{args.channel},
-                                 .offset{offset},
-                                 .samples_to_read{samples_to_read - samples_read}};
+            DecodeArg decode_arg{
+                .buffer{wavebuffer.buffer},
+                .buffer_size{wavebuffer.buffer_size},
+                .start_offset{start_offset},
+                .end_offset{end_offset},
+                .channel_count{args.channel_count},
+                .coefficients{},
+                .adpcm_context{nullptr},
+                .target_channel{args.channel},
+                .offset{offset},
+                .samples_to_read{samples_to_read - samples_read},
+            };
 
             s32 samples_decoded{0};
 
@@ -350,42 +364,30 @@ void DecodeFromWaveBuffers(Core::Memory::Memory& memory, const DecodeFromWaveBuf
             temp_buffer_pos += samples_decoded;
             offset += samples_decoded;
 
-            if (samples_decoded == 0 || offset >= end_offset - start_offset) {
-                offset = 0;
-                if (!wavebuffer.loop) {
-                    voice_state.wave_buffer_valid[wavebuffer_index] = false;
-                    voice_state.loop_count = 0;
+            if (samples_decoded && offset < end_offset - start_offset) {
+                continue;
+            }
 
-                    if (wavebuffer.stream_ended) {
-                        played_sample_count = 0;
-                    }
-
-                    wavebuffer_index = (wavebuffer_index + 1) % MaxWaveBuffers;
-                    wavebuffers_consumed++;
-                } else {
-                    voice_state.loop_count++;
-                    if (wavebuffer.loop_count >= 0 &&
-                        (voice_state.loop_count > wavebuffer.loop_count || samples_decoded == 0)) {
-                        voice_state.wave_buffer_valid[wavebuffer_index] = false;
-                        voice_state.loop_count = 0;
-
-                        if (wavebuffer.stream_ended) {
-                            played_sample_count = 0;
-                        }
-
-                        wavebuffer_index = (wavebuffer_index + 1) % MaxWaveBuffers;
-                        wavebuffers_consumed++;
-                    }
-
-                    if (samples_decoded == 0) {
-                        is_buffer_starved = true;
-                        break;
-                    }
-
-                    if (args.IsVoicePlayedSampleCountResetAtLoopPointSupported) {
-                        played_sample_count = 0;
-                    }
+            offset = 0;
+            if (wavebuffer.loop) {
+                voice_state.loop_count++;
+                if (wavebuffer.loop_count >= 0 &&
+                    (voice_state.loop_count > wavebuffer.loop_count || samples_decoded == 0)) {
+                    EndWaveBuffer(voice_state, wavebuffer, wavebuffer_index, played_sample_count,
+                                  wavebuffers_consumed);
                 }
+
+                if (samples_decoded == 0) {
+                    is_buffer_starved = true;
+                    break;
+                }
+
+                if (args.IsVoicePlayedSampleCountResetAtLoopPointSupported) {
+                    played_sample_count = 0;
+                }
+            } else {
+                EndWaveBuffer(voice_state, wavebuffer, wavebuffer_index, played_sample_count,
+                              wavebuffers_consumed);
             }
         }
 

src/audio_core/renderer/command/data_source/pcm_float.cpp

@@ -21,6 +21,12 @@ void PcmFloatDataSourceVersion1Command::Process(
     auto out_buffer = processor.mix_buffers.subspan(output_index * processor.sample_count,
                                                     processor.sample_count);
 
+    for (auto& wave_buffer : wave_buffers) {
+        wave_buffer.loop_start_offset = wave_buffer.start_offset;
+        wave_buffer.loop_end_offset = wave_buffer.end_offset;
+        wave_buffer.loop_count = wave_buffer.loop ? -1 : 0;
+    }
+
     DecodeFromWaveBuffersArgs args{
         .sample_format{SampleFormat::PcmFloat},
         .output{out_buffer},

src/audio_core/renderer/command/data_source/pcm_int16.cpp

@@ -23,6 +23,12 @@ void PcmInt16DataSourceVersion1Command::Process(
     auto out_buffer = processor.mix_buffers.subspan(output_index * processor.sample_count,
                                                     processor.sample_count);
 
+    for (auto& wave_buffer : wave_buffers) {
+        wave_buffer.loop_start_offset = wave_buffer.start_offset;
+        wave_buffer.loop_end_offset = wave_buffer.end_offset;
+        wave_buffer.loop_count = wave_buffer.loop ? -1 : 0;
+    }
+
     DecodeFromWaveBuffersArgs args{
         .sample_format{SampleFormat::PcmInt16},
         .output{out_buffer},

src/audio_core/renderer/system.cpp

@@ -609,17 +609,11 @@ void System::SendCommandToDsp() {
                 time_limit_percent = 70.0f;
             }
 
-            AudioRenderer::CommandBuffer command_buffer{
-                .buffer{translated_addr},
-                .size{command_size},
-                .time_limit{
-                    static_cast<u64>((time_limit_percent / 100) * 2'880'000.0 *
-                                     (static_cast<f32>(render_time_limit_percent) / 100.0f))},
-                .applet_resource_user_id{applet_resource_user_id},
-                .reset_buffer{reset_command_buffers},
-            };
-
-            audio_renderer.SetCommandBuffer(session_id, command_buffer);
+            auto time_limit{
+                static_cast<u64>((time_limit_percent / 100) * 2'880'000.0 *
+                                 (static_cast<f32>(render_time_limit_percent) / 100.0f))};
+            audio_renderer.SetCommandBuffer(session_id, translated_addr, command_size, time_limit,
+                                            applet_resource_user_id, reset_command_buffers);
             reset_command_buffers = false;
             command_buffer_size = command_size;
             if (remaining_command_count == 0) {

src/core/core.cpp

@@ -406,6 +406,7 @@ struct System::Impl {
             gpu_core->NotifyShutdown();
         }
 
+        Network::CancelPendingSocketOperations();
         kernel.SuspendApplication(true);
         if (services) {
             services->KillNVNFlinger();
@@ -427,6 +428,7 @@ struct System::Impl {
         debugger.reset();
         kernel.Shutdown();
         memory.Reset();
+        Network::RestartSocketOperations();
 
         if (auto room_member = room_network.GetRoomMember().lock()) {
             Network::GameInfo game_info{};

src/core/internal_network/network.cpp

@@ -48,15 +48,32 @@ enum class CallType {
 
 using socklen_t = int;
 
+SOCKET interrupt_socket = static_cast<SOCKET>(-1);
+
+void InterruptSocketOperations() {
+    closesocket(interrupt_socket);
+}
+
+void AcknowledgeInterrupt() {
+    interrupt_socket = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP);
+}
+
 void Initialize() {
     WSADATA wsa_data;
     (void)WSAStartup(MAKEWORD(2, 2), &wsa_data);
+
+    AcknowledgeInterrupt();
 }
 
 void Finalize() {
+    InterruptSocketOperations();
     WSACleanup();
 }
 
+SOCKET GetInterruptSocket() {
+    return interrupt_socket;
+}
+
 sockaddr TranslateFromSockAddrIn(SockAddrIn input) {
     sockaddr_in result;
 
@@ -157,9 +174,42 @@ constexpr int SD_RECEIVE = SHUT_RD;
 constexpr int SD_SEND = SHUT_WR;
 constexpr int SD_BOTH = SHUT_RDWR;
 
-void Initialize() {}
+int interrupt_pipe_fd[2] = {-1, -1};
 
-void Finalize() {}
+void Initialize() {
+    if (pipe(interrupt_pipe_fd) != 0) {
+        LOG_ERROR(Network, "Failed to create interrupt pipe!");
+    }
+    int flags = fcntl(interrupt_pipe_fd[0], F_GETFL);
+    ASSERT_MSG(fcntl(interrupt_pipe_fd[0], F_SETFL, flags | O_NONBLOCK) == 0,
+               "Failed to set nonblocking state for interrupt pipe");
+}
+
+void Finalize() {
+    if (interrupt_pipe_fd[0] >= 0) {
+        close(interrupt_pipe_fd[0]);
+    }
+    if (interrupt_pipe_fd[1] >= 0) {
+        close(interrupt_pipe_fd[1]);
+    }
+}
+
+void InterruptSocketOperations() {
+    u8 value = 0;
+    ASSERT(write(interrupt_pipe_fd[1], &value, sizeof(value)) == 1);
+}
+
+void AcknowledgeInterrupt() {
+    u8 value = 0;
+    ssize_t ret = read(interrupt_pipe_fd[0], &value, sizeof(value));
+    if (ret != 1 && errno != EAGAIN && errno != EWOULDBLOCK) {
+        LOG_ERROR(Network, "Failed to acknowledge interrupt on shutdown");
+    }
+}
+
+SOCKET GetInterruptSocket() {
+    return interrupt_pipe_fd[0];
+}
 
 sockaddr TranslateFromSockAddrIn(SockAddrIn input) {
     sockaddr_in result;
@@ -490,6 +540,14 @@ NetworkInstance::~NetworkInstance() {
     Finalize();
 }
 
+void CancelPendingSocketOperations() {
+    InterruptSocketOperations();
+}
+
+void RestartSocketOperations() {
+    AcknowledgeInterrupt();
+}
+
 std::optional<IPv4Address> GetHostIPv4Address() {
     const auto network_interface = Network::GetSelectedNetworkInterface();
     if (!network_interface.has_value()) {
@@ -560,7 +618,14 @@ std::pair<s32, Errno> Poll(std::vector<PollFD>& pollfds, s32 timeout) {
         return result;
     });
 
-    const int result = WSAPoll(host_pollfds.data(), static_cast<ULONG>(num), timeout);
+    host_pollfds.push_back(WSAPOLLFD{
+        .fd = GetInterruptSocket(),
+        .events = POLLIN,
+        .revents = 0,
+    });
+
+    const int result =
+        WSAPoll(host_pollfds.data(), static_cast<ULONG>(host_pollfds.size()), timeout);
     if (result == 0) {
         ASSERT(std::all_of(host_pollfds.begin(), host_pollfds.end(),
                            [](WSAPOLLFD fd) { return fd.revents == 0; }));
@@ -627,6 +692,24 @@ Errno Socket::Initialize(Domain domain, Type type, Protocol protocol) {
 std::pair<SocketBase::AcceptResult, Errno> Socket::Accept() {
     sockaddr_in addr;
     socklen_t addrlen = sizeof(addr);
+
+    std::vector<WSAPOLLFD> host_pollfds{
+        WSAPOLLFD{fd, POLLIN, 0},
+        WSAPOLLFD{GetInterruptSocket(), POLLIN, 0},
+    };
+
+    while (true) {
+        const int pollres =
+            WSAPoll(host_pollfds.data(), static_cast<ULONG>(host_pollfds.size()), -1);
+        if (host_pollfds[1].revents != 0) {
+            // Interrupt signaled before a client could be accepted, break
+            return {AcceptResult{}, Errno::AGAIN};
+        }
+        if (pollres > 0) {
+            break;
+        }
+    }
+
     const SOCKET new_socket = accept(fd, reinterpret_cast<sockaddr*>(&addr), &addrlen);
 
     if (new_socket == INVALID_SOCKET) {

src/core/internal_network/network.h

@@ -96,6 +96,9 @@ public:
     ~NetworkInstance();
 };
 
+void CancelPendingSocketOperations();
+void RestartSocketOperations();
+
 #ifdef _WIN32
 constexpr IPv4Address TranslateIPv4(in_addr addr) {
     auto& bytes = addr.S_un.S_un_b;

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -204,9 +204,7 @@ Id TextureImage(EmitContext& ctx, IR::TextureInstInfo info, const IR::Value& ind
         if (def.count > 1) {
             throw NotImplementedException("Indirect texture sample");
         }
-        const Id sampler_id{def.id};
-        const Id id{ctx.OpLoad(ctx.sampled_texture_buffer_type, sampler_id)};
-        return ctx.OpImage(ctx.image_buffer_type, id);
+        return ctx.OpLoad(ctx.image_buffer_type, def.id);
     } else {
         const TextureDefinition& def{ctx.textures.at(info.descriptor_index)};
         if (def.count > 1) {

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -74,6 +74,11 @@ spv::ImageFormat GetImageFormat(ImageFormat format) {
     throw InvalidArgument("Invalid image format {}", format);
 }
 
+spv::ImageFormat GetImageFormatForBuffer(ImageFormat format) {
+    const auto spv_format = GetImageFormat(format);
+    return spv_format == spv::ImageFormat::Unknown ? spv::ImageFormat::R32ui : spv_format;
+}
+
 Id ImageType(EmitContext& ctx, const ImageDescriptor& desc) {
     const spv::ImageFormat format{GetImageFormat(desc.format)};
     const Id type{ctx.U32[1]};
@@ -1242,9 +1247,8 @@ void EmitContext::DefineTextureBuffers(const Info& info, u32& binding) {
     }
     const spv::ImageFormat format{spv::ImageFormat::Unknown};
     image_buffer_type = TypeImage(F32[1], spv::Dim::Buffer, 0U, false, false, 1, format);
-    sampled_texture_buffer_type = TypeSampledImage(image_buffer_type);
 
-    const Id type{TypePointer(spv::StorageClass::UniformConstant, sampled_texture_buffer_type)};
+    const Id type{TypePointer(spv::StorageClass::UniformConstant, image_buffer_type)};
     texture_buffers.reserve(info.texture_buffer_descriptors.size());
     for (const TextureBufferDescriptor& desc : info.texture_buffer_descriptors) {
         if (desc.count != 1) {
@@ -1271,7 +1275,7 @@ void EmitContext::DefineImageBuffers(const Info& info, u32& binding) {
         if (desc.count != 1) {
             throw NotImplementedException("Array of image buffers");
         }
-        const spv::ImageFormat format{GetImageFormat(desc.format)};
+        const spv::ImageFormat format{GetImageFormatForBuffer(desc.format)};
         const Id image_type{TypeImage(U32[1], spv::Dim::Buffer, false, false, false, 2, format)};
         const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, image_type)};
         const Id id{AddGlobalVariable(pointer_type, spv::StorageClass::UniformConstant)};

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -206,7 +206,6 @@ public:
     Id output_u32{};
 
     Id image_buffer_type{};
-    Id sampled_texture_buffer_type{};
     Id image_u32{};
 
     std::array<UniformDefinitions, Info::MAX_CBUFS> cbufs{};