Merge pull request #10457 from Kelebek1/optimise

Remove memory allocations in some hot paths

src/audio_core/device/audio_buffers.h

@@ -7,6 +7,7 @@
 #include <mutex>
 #include <span>
 #include <vector>
+#include <boost/container/static_vector.hpp>
 
 #include "audio_buffer.h"
 #include "audio_core/device/device_session.h"
@@ -48,7 +49,7 @@ public:
      *
      * @param out_buffers - The buffers which were registered.
      */
-    void RegisterBuffers(std::vector<AudioBuffer>& out_buffers) {
+    void RegisterBuffers(boost::container::static_vector<AudioBuffer, N>& out_buffers) {
         std::scoped_lock l{lock};
         const s32 to_register{std::min(std::min(appended_count, BufferAppendLimit),
                                        BufferAppendLimit - registered_count)};
@@ -162,7 +163,8 @@ public:
      * @param max_buffers     - Maximum number of buffers to released.
      * @return The number of buffers released.
      */
-    u32 GetRegisteredAppendedBuffers(std::vector<AudioBuffer>& buffers_flushed, u32 max_buffers) {
+    u32 GetRegisteredAppendedBuffers(
+        boost::container::static_vector<AudioBuffer, N>& buffers_flushed, u32 max_buffers) {
         std::scoped_lock l{lock};
         if (registered_count + appended_count == 0) {
             return 0;
@@ -270,7 +272,7 @@ public:
      */
     bool FlushBuffers(u32& buffers_released) {
         std::scoped_lock l{lock};
-        std::vector<AudioBuffer> buffers_flushed{};
+        boost::container::static_vector<AudioBuffer, N> buffers_flushed{};
 
         buffers_released = GetRegisteredAppendedBuffers(buffers_flushed, append_limit);
 

src/audio_core/device/device_session.cpp

@@ -79,7 +79,7 @@ void DeviceSession::ClearBuffers() {
     }
 }
 
-void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) const {
+void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) {
     for (const auto& buffer : buffers) {
         Sink::SinkBuffer new_buffer{
             .frames = buffer.size / (channel_count * sizeof(s16)),
@@ -88,13 +88,13 @@ void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) const {
             .consumed = false,
         };
 
+        tmp_samples.resize_destructive(buffer.size / sizeof(s16));
         if (type == Sink::StreamType::In) {
-            std::vector<s16> samples{};
-            stream->AppendBuffer(new_buffer, samples);
+            stream->AppendBuffer(new_buffer, tmp_samples);
         } else {
-            std::vector<s16> samples(buffer.size / sizeof(s16));
-            system.ApplicationMemory().ReadBlockUnsafe(buffer.samples, samples.data(), buffer.size);
-            stream->AppendBuffer(new_buffer, samples);
+            system.ApplicationMemory().ReadBlockUnsafe(buffer.samples, tmp_samples.data(),
+                                                       buffer.size);
+            stream->AppendBuffer(new_buffer, tmp_samples);
         }
     }
 }

src/audio_core/device/device_session.h

@@ -10,6 +10,7 @@
 
 #include "audio_core/common/common.h"
 #include "audio_core/sink/sink.h"
+#include "common/scratch_buffer.h"
 #include "core/hle/service/audio/errors.h"
 
 namespace Core {
@@ -62,7 +63,7 @@ public:
      *
      * @param buffers - The buffers to play.
      */
-    void AppendBuffers(std::span<const AudioBuffer> buffers) const;
+    void AppendBuffers(std::span<const AudioBuffer> buffers);
 
     /**
      * (Audio In only) Pop samples from the backend, and write them back to this buffer's address.
@@ -146,8 +147,8 @@ private:
     std::shared_ptr<Core::Timing::EventType> thread_event;
     /// Is this session initialised?
     bool initialized{};
-    /// Buffer queue
-    std::vector<AudioBuffer> buffer_queue{};
+    /// Temporary sample buffer
+    Common::ScratchBuffer<s16> tmp_samples{};
 };
 
 } // namespace AudioCore

src/audio_core/in/audio_in_system.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <mutex>
+
 #include "audio_core/audio_event.h"
 #include "audio_core/audio_manager.h"
 #include "audio_core/in/audio_in_system.h"
@@ -89,7 +90,7 @@ Result System::Start() {
     session->Start();
     state = State::Started;
 
-    std::vector<AudioBuffer> buffers_to_flush{};
+    boost::container::static_vector<AudioBuffer, BufferCount> buffers_to_flush{};
     buffers.RegisterBuffers(buffers_to_flush);
     session->AppendBuffers(buffers_to_flush);
     session->SetRingSize(static_cast<u32>(buffers_to_flush.size()));
@@ -134,7 +135,7 @@ bool System::AppendBuffer(const AudioInBuffer& buffer, const u64 tag) {
 
 void System::RegisterBuffers() {
     if (state == State::Started) {
-        std::vector<AudioBuffer> registered_buffers{};
+        boost::container::static_vector<AudioBuffer, BufferCount> registered_buffers{};
         buffers.RegisterBuffers(registered_buffers);
         session->AppendBuffers(registered_buffers);
     }

src/audio_core/out/audio_out_system.cpp

@@ -89,7 +89,7 @@ Result System::Start() {
     session->Start();
     state = State::Started;
 
-    std::vector<AudioBuffer> buffers_to_flush{};
+    boost::container::static_vector<AudioBuffer, BufferCount> buffers_to_flush{};
     buffers.RegisterBuffers(buffers_to_flush);
     session->AppendBuffers(buffers_to_flush);
     session->SetRingSize(static_cast<u32>(buffers_to_flush.size()));
@@ -134,7 +134,7 @@ bool System::AppendBuffer(const AudioOutBuffer& buffer, u64 tag) {
 
 void System::RegisterBuffers() {
     if (state == State::Started) {
-        std::vector<AudioBuffer> registered_buffers{};
+        boost::container::static_vector<AudioBuffer, BufferCount> registered_buffers{};
         buffers.RegisterBuffers(registered_buffers);
         session->AppendBuffers(registered_buffers);
     }

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

@@ -8,6 +8,7 @@
 #include "audio_core/renderer/command/resample/resample.h"
 #include "common/fixed_point.h"
 #include "common/logging/log.h"
+#include "common/scratch_buffer.h"
 #include "core/memory.h"
 
 namespace AudioCore::AudioRenderer {
@@ -27,6 +28,7 @@ constexpr std::array<u8, 3> PitchBySrcQuality = {4, 8, 4};
 template <typename T>
 static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
                      const DecodeArg& req) {
+    std::array<T, TempBufferSize> tmp_samples{};
     constexpr s32 min{std::numeric_limits<s16>::min()};
     constexpr s32 max{std::numeric_limits<s16>::max()};
 
@@ -49,18 +51,17 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
         const u64 size{channel_count * samples_to_decode};
         const u64 size_bytes{size * sizeof(T)};
 
-        std::vector<T> samples(size);
-        memory.ReadBlockUnsafe(source, samples.data(), size_bytes);
+        memory.ReadBlockUnsafe(source, tmp_samples.data(), size_bytes);
 
         if constexpr (std::is_floating_point_v<T>) {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                auto sample{static_cast<s32>(samples[i * channel_count + req.target_channel] *
+                auto sample{static_cast<s32>(tmp_samples[i * channel_count + req.target_channel] *
                                              std::numeric_limits<s16>::max())};
                 out_buffer[i] = static_cast<s16>(std::clamp(sample, min, max));
             }
         } else {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                out_buffer[i] = samples[i * channel_count + req.target_channel];
+                out_buffer[i] = tmp_samples[i * channel_count + req.target_channel];
             }
         }
     } break;
@@ -73,17 +74,16 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
         }
 
         const VAddr source{req.buffer + ((req.start_offset + req.offset) * sizeof(T))};
-        std::vector<T> samples(samples_to_decode);
-        memory.ReadBlockUnsafe(source, samples.data(), samples_to_decode * sizeof(T));
+        memory.ReadBlockUnsafe(source, tmp_samples.data(), samples_to_decode * sizeof(T));
 
         if constexpr (std::is_floating_point_v<T>) {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                auto sample{static_cast<s32>(samples[i * channel_count + req.target_channel] *
+                auto sample{static_cast<s32>(tmp_samples[i * channel_count + req.target_channel] *
                                              std::numeric_limits<s16>::max())};
                 out_buffer[i] = static_cast<s16>(std::clamp(sample, min, max));
             }
         } else {
-            std::memcpy(out_buffer.data(), samples.data(), samples_to_decode * sizeof(s16));
+            std::memcpy(out_buffer.data(), tmp_samples.data(), samples_to_decode * sizeof(s16));
         }
         break;
     }
@@ -101,6 +101,7 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
  */
 static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
                        const DecodeArg& req) {
+    std::array<u8, TempBufferSize> wavebuffer{};
     constexpr u32 SamplesPerFrame{14};
     constexpr u32 NibblesPerFrame{16};
 
@@ -138,9 +139,7 @@ static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
     }
 
     const auto size{std::max((samples_to_process / 8U) * SamplesPerFrame, 8U)};
-    std::vector<u8> wavebuffer(size);
-    memory.ReadBlockUnsafe(req.buffer + position_in_frame / 2, wavebuffer.data(),
-                           wavebuffer.size());
+    memory.ReadBlockUnsafe(req.buffer + position_in_frame / 2, wavebuffer.data(), size);
 
     auto context{req.adpcm_context};
     auto header{context->header};
@@ -258,7 +257,7 @@ void DecodeFromWaveBuffers(Core::Memory::Memory& memory, const DecodeFromWaveBuf
     u32 offset{voice_state.offset};
 
     auto output_buffer{args.output};
-    std::vector<s16> temp_buffer(TempBufferSize, 0);
+    std::array<s16, TempBufferSize> temp_buffer{};
 
     while (remaining_sample_count > 0) {
         const auto samples_to_write{std::min(remaining_sample_count, max_remaining_sample_count)};

src/audio_core/renderer/command/effect/compressor.cpp

@@ -44,8 +44,8 @@ static void InitializeCompressorEffect(const CompressorInfo::ParameterVersion2&
 
 static void ApplyCompressorEffect(const CompressorInfo::ParameterVersion2& params,
                                   CompressorInfo::State& state, bool enabled,
-                                  std::vector<std::span<const s32>> input_buffers,
-                                  std::vector<std::span<s32>> output_buffers, u32 sample_count) {
+                                  std::span<std::span<const s32>> input_buffers,
+                                  std::span<std::span<s32>> output_buffers, u32 sample_count) {
     if (enabled) {
         auto state_00{state.unk_00};
         auto state_04{state.unk_04};
@@ -124,8 +124,8 @@ void CompressorCommand::Dump([[maybe_unused]] const ADSP::CommandListProcessor&
 }
 
 void CompressorCommand::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (s16 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,

src/audio_core/renderer/command/effect/delay.cpp

@@ -51,7 +51,7 @@ static void InitializeDelayEffect(const DelayInfo::ParameterVersion1& params,
         state.delay_lines[channel].sample_count_max = sample_count_max.to_int_floor();
         state.delay_lines[channel].sample_count = sample_count.to_int_floor();
         state.delay_lines[channel].buffer.resize(state.delay_lines[channel].sample_count, 0);
-        if (state.delay_lines[channel].buffer.size() == 0) {
+        if (state.delay_lines[channel].sample_count == 0) {
             state.delay_lines[channel].buffer.push_back(0);
         }
         state.delay_lines[channel].buffer_pos = 0;
@@ -74,8 +74,8 @@ static void InitializeDelayEffect(const DelayInfo::ParameterVersion1& params,
  */
 template <size_t NumChannels>
 static void ApplyDelay(const DelayInfo::ParameterVersion1& params, DelayInfo::State& state,
-                       std::vector<std::span<const s32>>& inputs,
-                       std::vector<std::span<s32>>& outputs, const u32 sample_count) {
+                       std::span<std::span<const s32>> inputs, std::span<std::span<s32>> outputs,
+                       const u32 sample_count) {
     for (u32 sample_index = 0; sample_index < sample_count; sample_index++) {
         std::array<Common::FixedPoint<50, 14>, NumChannels> input_samples{};
         for (u32 channel = 0; channel < NumChannels; channel++) {
@@ -153,8 +153,8 @@ static void ApplyDelay(const DelayInfo::ParameterVersion1& params, DelayInfo::St
  * @param sample_count - Number of samples to process.
  */
 static void ApplyDelayEffect(const DelayInfo::ParameterVersion1& params, DelayInfo::State& state,
-                             const bool enabled, std::vector<std::span<const s32>>& inputs,
-                             std::vector<std::span<s32>>& outputs, const u32 sample_count) {
+                             const bool enabled, std::span<std::span<const s32>> inputs,
+                             std::span<std::span<s32>> outputs, const u32 sample_count) {
 
     if (!IsChannelCountValid(params.channel_count)) {
         LOG_ERROR(Service_Audio, "Invalid delay channels {}", params.channel_count);
@@ -208,8 +208,8 @@ void DelayCommand::Dump([[maybe_unused]] const ADSP::CommandListProcessor& proce
 }
 
 void DelayCommand::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (s16 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,

src/audio_core/renderer/command/effect/i3dl2_reverb.cpp

@@ -408,8 +408,8 @@ void I3dl2ReverbCommand::Dump([[maybe_unused]] const ADSP::CommandListProcessor&
 }
 
 void I3dl2ReverbCommand::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (u32 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,

src/audio_core/renderer/command/effect/light_limiter.cpp

@@ -47,8 +47,8 @@ static void InitializeLightLimiterEffect(const LightLimiterInfo::ParameterVersio
  */
 static void ApplyLightLimiterEffect(const LightLimiterInfo::ParameterVersion2& params,
                                     LightLimiterInfo::State& state, const bool enabled,
-                                    std::vector<std::span<const s32>>& inputs,
-                                    std::vector<std::span<s32>>& outputs, const u32 sample_count,
+                                    std::span<std::span<const s32>> inputs,
+                                    std::span<std::span<s32>> outputs, const u32 sample_count,
                                     LightLimiterInfo::StatisticsInternal* statistics) {
     constexpr s64 min{std::numeric_limits<s32>::min()};
     constexpr s64 max{std::numeric_limits<s32>::max()};
@@ -147,8 +147,8 @@ void LightLimiterVersion1Command::Dump([[maybe_unused]] const ADSP::CommandListP
 }
 
 void LightLimiterVersion1Command::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (u32 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,
@@ -190,8 +190,8 @@ void LightLimiterVersion2Command::Dump([[maybe_unused]] const ADSP::CommandListP
 }
 
 void LightLimiterVersion2Command::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (u32 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,

src/audio_core/renderer/command/effect/reverb.cpp

@@ -250,8 +250,8 @@ static Common::FixedPoint<50, 14> Axfx2AllPassTick(ReverbInfo::ReverbDelayLine&
  */
 template <size_t NumChannels>
 static void ApplyReverbEffect(const ReverbInfo::ParameterVersion2& params, ReverbInfo::State& state,
-                              std::vector<std::span<const s32>>& inputs,
-                              std::vector<std::span<s32>>& outputs, const u32 sample_count) {
+                              std::span<std::span<const s32>> inputs,
+                              std::span<std::span<s32>> outputs, const u32 sample_count) {
     static constexpr std::array<u8, ReverbInfo::MaxDelayTaps> OutTapIndexes1Ch{
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     };
@@ -369,8 +369,8 @@ static void ApplyReverbEffect(const ReverbInfo::ParameterVersion2& params, Rever
  * @param sample_count - Number of samples to process.
  */
 static void ApplyReverbEffect(const ReverbInfo::ParameterVersion2& params, ReverbInfo::State& state,
-                              const bool enabled, std::vector<std::span<const s32>>& inputs,
-                              std::vector<std::span<s32>>& outputs, const u32 sample_count) {
+                              const bool enabled, std::span<std::span<const s32>> inputs,
+                              std::span<std::span<s32>> outputs, const u32 sample_count) {
     if (enabled) {
         switch (params.channel_count) {
         case 0:
@@ -412,8 +412,8 @@ void ReverbCommand::Dump([[maybe_unused]] const ADSP::CommandListProcessor& proc
 }
 
 void ReverbCommand::Process(const ADSP::CommandListProcessor& processor) {
-    std::vector<std::span<const s32>> input_buffers(parameter.channel_count);
-    std::vector<std::span<s32>> output_buffers(parameter.channel_count);
+    std::array<std::span<const s32>, MaxChannels> input_buffers{};
+    std::array<std::span<s32>, MaxChannels> output_buffers{};
 
     for (u32 i = 0; i < parameter.channel_count; i++) {
         input_buffers[i] = processor.mix_buffers.subspan(inputs[i] * processor.sample_count,

src/audio_core/renderer/command/sink/circular_buffer.cpp

@@ -24,7 +24,7 @@ void CircularBufferSinkCommand::Process(const ADSP::CommandListProcessor& proces
     constexpr s32 min{std::numeric_limits<s16>::min()};
     constexpr s32 max{std::numeric_limits<s16>::max()};
 
-    std::vector<s16> output(processor.sample_count);
+    std::array<s16, TargetSampleCount * MaxChannels> output{};
     for (u32 channel = 0; channel < input_count; channel++) {
         auto input{processor.mix_buffers.subspan(inputs[channel] * processor.sample_count,
                                                  processor.sample_count)};
@@ -33,7 +33,7 @@ void CircularBufferSinkCommand::Process(const ADSP::CommandListProcessor& proces
         }
 
         processor.memory->WriteBlockUnsafe(address + pos, output.data(),
-                                           output.size() * sizeof(s16));
+                                           processor.sample_count * sizeof(s16));
         pos += static_cast<u32>(processor.sample_count * sizeof(s16));
         if (pos >= size) {
             pos = 0;

src/audio_core/renderer/command/sink/device.cpp

@@ -33,8 +33,7 @@ void DeviceSinkCommand::Process(const ADSP::CommandListProcessor& processor) {
         .consumed{false},
     };
 
-    std::vector<s16> samples(out_buffer.frames * input_count);
-
+    std::array<s16, TargetSampleCount * MaxChannels> samples{};
     for (u32 channel = 0; channel < input_count; channel++) {
         const auto offset{inputs[channel] * out_buffer.frames};
 
@@ -45,7 +44,7 @@ void DeviceSinkCommand::Process(const ADSP::CommandListProcessor& processor) {
     }
 
     out_buffer.tag = reinterpret_cast<u64>(samples.data());
-    stream->AppendBuffer(out_buffer, samples);
+    stream->AppendBuffer(out_buffer, {samples.data(), out_buffer.frames * input_count});
 
     if (stream->IsPaused()) {
         stream->Start();

src/audio_core/renderer/mix/mix_context.cpp

@@ -125,10 +125,10 @@ bool MixContext::TSortInfo(const SplitterContext& splitter_context) {
         return false;
     }
 
-    std::vector<s32> sorted_results{node_states.GetSortedResuls()};
-    const auto result_size{std::min(count, static_cast<s32>(sorted_results.size()))};
+    auto sorted_results{node_states.GetSortedResuls()};
+    const auto result_size{std::min(count, static_cast<s32>(sorted_results.second))};
     for (s32 i = 0; i < result_size; i++) {
-        sorted_mix_infos[i] = &mix_infos[sorted_results[i]];
+        sorted_mix_infos[i] = &mix_infos[sorted_results.first[i]];
     }
 
     CalcMixBufferOffset();

src/audio_core/renderer/nodes/node_states.cpp

@@ -134,8 +134,8 @@ u32 NodeStates::GetNodeCount() const {
     return node_count;
 }
 
-std::vector<s32> NodeStates::GetSortedResuls() const {
-    return {results.rbegin(), results.rbegin() + result_pos};
+std::pair<std::span<u32>::reverse_iterator, size_t> NodeStates::GetSortedResuls() const {
+    return {results.rbegin(), result_pos};
 }
 
 } // namespace AudioCore::AudioRenderer

src/audio_core/renderer/nodes/node_states.h

@@ -175,7 +175,7 @@ public:
      *
      * @return Vector of nodes in reverse order.
      */
-    std::vector<s32> GetSortedResuls() const;
+    std::pair<std::span<u32>::reverse_iterator, size_t> GetSortedResuls() const;
 
 private:
     /// Number of nodes in the graph

src/audio_core/renderer/system.cpp

@@ -444,6 +444,7 @@ Result System::Update(std::span<const u8> input, std::span<u8> performance, std:
     std::scoped_lock l{lock};
 
     const auto start_time{core.CoreTiming().GetClockTicks()};
+    std::memset(output.data(), 0, output.size());
 
     InfoUpdater info_updater(input, output, process_handle, behavior);
 

src/audio_core/sink/null_sink.h

@@ -20,7 +20,7 @@ public:
     explicit NullSinkStreamImpl(Core::System& system_, StreamType type_)
         : SinkStream{system_, type_} {}
     ~NullSinkStreamImpl() override {}
-    void AppendBuffer(SinkBuffer&, std::vector<s16>&) override {}
+    void AppendBuffer(SinkBuffer&, std::span<s16>) override {}
     std::vector<s16> ReleaseBuffer(u64) override {
         return {};
     }

src/audio_core/sink/sink_stream.cpp

@@ -18,7 +18,7 @@
 
 namespace AudioCore::Sink {
 
-void SinkStream::AppendBuffer(SinkBuffer& buffer, std::vector<s16>& samples) {
+void SinkStream::AppendBuffer(SinkBuffer& buffer, std::span<s16> samples) {
     if (type == StreamType::In) {
         queue.enqueue(buffer);
         queued_buffers++;
@@ -66,15 +66,16 @@ void SinkStream::AppendBuffer(SinkBuffer& buffer, std::vector<s16>& samples) {
                 static_cast<s16>(std::clamp(right_sample, min, max));
         }
 
-        samples.resize(samples.size() / system_channels * device_channels);
+        samples = samples.subspan(0, samples.size() / system_channels * device_channels);
 
     } else if (system_channels == 2 && device_channels == 6) {
         // We need moar samples! Not all games will provide 6 channel audio.
         // TODO: Implement some upmixing here. Currently just passthrough, with other
         // channels left as silence.
-        std::vector<s16> new_samples(samples.size() / system_channels * device_channels, 0);
+        auto new_size = samples.size() / system_channels * device_channels;
+        tmp_samples.resize_destructive(new_size);
 
-        for (u32 read_index = 0, write_index = 0; read_index < samples.size();
+        for (u32 read_index = 0, write_index = 0; read_index < new_size;
              read_index += system_channels, write_index += device_channels) {
             const auto left_sample{static_cast<s16>(std::clamp(
                 static_cast<s32>(
@@ -82,7 +83,7 @@ void SinkStream::AppendBuffer(SinkBuffer& buffer, std::vector<s16>& samples) {
                     volume),
                 min, max))};
 
-            new_samples[write_index + static_cast<u32>(Channels::FrontLeft)] = left_sample;
+            tmp_samples[write_index + static_cast<u32>(Channels::FrontLeft)] = left_sample;
 
             const auto right_sample{static_cast<s16>(std::clamp(
                 static_cast<s32>(
@@ -90,9 +91,9 @@ void SinkStream::AppendBuffer(SinkBuffer& buffer, std::vector<s16>& samples) {
                     volume),
                 min, max))};
 
-            new_samples[write_index + static_cast<u32>(Channels::FrontRight)] = right_sample;
+            tmp_samples[write_index + static_cast<u32>(Channels::FrontRight)] = right_sample;
         }
-        samples = std::move(new_samples);
+        samples = std::span<s16>(tmp_samples);
 
     } else if (volume != 1.0f) {
         for (u32 i = 0; i < samples.size(); i++) {

src/audio_core/sink/sink_stream.h

@@ -16,6 +16,7 @@
 #include "common/polyfill_thread.h"
 #include "common/reader_writer_queue.h"
 #include "common/ring_buffer.h"
+#include "common/scratch_buffer.h"
 #include "common/thread.h"
 
 namespace Core {
@@ -170,7 +171,7 @@ public:
      * @param buffer  - Audio buffer information to be queued.
      * @param samples - The s16 samples to be queue for playback.
      */
-    virtual void AppendBuffer(SinkBuffer& buffer, std::vector<s16>& samples);
+    virtual void AppendBuffer(SinkBuffer& buffer, std::span<s16> samples);
 
     /**
      * Release a buffer. Audio In only, will fill a buffer with recorded samples.
@@ -255,6 +256,8 @@ private:
     /// Signalled when ring buffer entries are consumed
     std::condition_variable_any release_cv;
     std::mutex release_mutex;
+    /// Temporary buffer for appending samples when upmixing
+    Common::ScratchBuffer<s16> tmp_samples{};
 };
 
 using SinkStreamPtr = std::unique_ptr<SinkStream>;

src/common/ring_buffer.h

@@ -9,6 +9,7 @@
 #include <cstddef>
 #include <cstring>
 #include <new>
+#include <span>
 #include <type_traits>
 #include <vector>
 
@@ -53,7 +54,7 @@ public:
         return push_count;
     }
 
-    std::size_t Push(const std::vector<T>& input) {
+    std::size_t Push(const std::span<T> input) {
         return Push(input.data(), input.size());
     }
 

src/common/scratch_buffer.h

@@ -3,6 +3,9 @@
 
 #pragma once
 
+#include <iterator>
+
+#include "common/concepts.h"
 #include "common/make_unique_for_overwrite.h"
 
 namespace Common {
@@ -16,6 +19,12 @@ namespace Common {
 template <typename T>
 class ScratchBuffer {
 public:
+    using iterator = T*;
+    using const_iterator = const T*;
+    using value_type = T;
+    using element_type = T;
+    using iterator_category = std::contiguous_iterator_tag;
+
     ScratchBuffer() = default;
 
     explicit ScratchBuffer(size_t initial_capacity)

src/core/hle/kernel/k_synchronization_object.cpp

@@ -3,6 +3,7 @@
 
 #include "common/assert.h"
 #include "common/common_types.h"
+#include "common/scratch_buffer.h"
 #include "core/hle/kernel/k_scheduler.h"
 #include "core/hle/kernel/k_scoped_scheduler_lock_and_sleep.h"
 #include "core/hle/kernel/k_synchronization_object.h"
@@ -75,7 +76,7 @@ Result KSynchronizationObject::Wait(KernelCore& kernel, s32* out_index,
                                     KSynchronizationObject** objects, const s32 num_objects,
                                     s64 timeout) {
     // Allocate space on stack for thread nodes.
-    std::vector<ThreadListNode> thread_nodes(num_objects);
+    std::array<ThreadListNode, Svc::ArgumentHandleCountMax> thread_nodes;
 
     // Prepare for wait.
     KThread* thread = GetCurrentThreadPointer(kernel);

src/core/hle/kernel/k_thread.cpp

@@ -909,7 +909,7 @@ Result KThread::SetActivity(Svc::ThreadActivity activity) {
     R_SUCCEED();
 }
 
-Result KThread::GetThreadContext3(std::vector<u8>& out) {
+Result KThread::GetThreadContext3(Common::ScratchBuffer<u8>& out) {
     // Lock ourselves.
     KScopedLightLock lk{m_activity_pause_lock};
 
@@ -927,15 +927,13 @@ Result KThread::GetThreadContext3(std::vector<u8>& out) {
                 // Mask away mode bits, interrupt bits, IL bit, and other reserved bits.
                 auto context = GetContext64();
                 context.pstate &= 0xFF0FFE20;
-
-                out.resize(sizeof(context));
+                out.resize_destructive(sizeof(context));
                 std::memcpy(out.data(), std::addressof(context), sizeof(context));
             } else {
                 // Mask away mode bits, interrupt bits, IL bit, and other reserved bits.
                 auto context = GetContext32();
                 context.cpsr &= 0xFF0FFE20;
-
-                out.resize(sizeof(context));
+                out.resize_destructive(sizeof(context));
                 std::memcpy(out.data(), std::addressof(context), sizeof(context));
             }
         }

src/core/hle/kernel/k_thread.h

@@ -15,6 +15,7 @@
 #include "common/intrusive_list.h"
 
 #include "common/intrusive_red_black_tree.h"
+#include "common/scratch_buffer.h"
 #include "common/spin_lock.h"
 #include "core/arm/arm_interface.h"
 #include "core/hle/kernel/k_affinity_mask.h"
@@ -567,7 +568,7 @@ public:
 
     void RemoveWaiter(KThread* thread);
 
-    Result GetThreadContext3(std::vector<u8>& out);
+    Result GetThreadContext3(Common::ScratchBuffer<u8>& out);
 
     KThread* RemoveUserWaiterByKey(bool* out_has_waiters, KProcessAddress key) {
         return this->RemoveWaiterByKey(out_has_waiters, key, false);

src/core/hle/kernel/svc/svc_ipc.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include "common/scope_exit.h"
+#include "common/scratch_buffer.h"
 #include "core/core.h"
 #include "core/hle/kernel/k_client_session.h"
 #include "core/hle/kernel/k_process.h"
@@ -45,11 +46,11 @@ Result ReplyAndReceive(Core::System& system, s32* out_index, uint64_t handles_ad
                  handles_addr, static_cast<u64>(sizeof(Handle) * num_handles)),
              ResultInvalidPointer);
 
-    std::vector<Handle> handles(num_handles);
+    std::array<Handle, Svc::ArgumentHandleCountMax> handles;
     GetCurrentMemory(kernel).ReadBlock(handles_addr, handles.data(), sizeof(Handle) * num_handles);
 
     // Convert handle list to object table.
-    std::vector<KSynchronizationObject*> objs(num_handles);
+    std::array<KSynchronizationObject*, Svc::ArgumentHandleCountMax> objs;
     R_UNLESS(handle_table.GetMultipleObjects<KSynchronizationObject>(objs.data(), handles.data(),
                                                                      num_handles),
              ResultInvalidHandle);
@@ -80,7 +81,7 @@ Result ReplyAndReceive(Core::System& system, s32* out_index, uint64_t handles_ad
         // Wait for an object.
         s32 index;
         Result result = KSynchronizationObject::Wait(kernel, std::addressof(index), objs.data(),
-                                                     static_cast<s32>(objs.size()), timeout_ns);
+                                                     num_handles, timeout_ns);
         if (result == ResultTimedOut) {
             R_RETURN(result);
         }

src/core/hle/kernel/svc/svc_synchronization.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include "common/scope_exit.h"
+#include "common/scratch_buffer.h"
 #include "core/core.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/kernel/k_readable_event.h"
@@ -54,7 +55,7 @@ static Result WaitSynchronization(Core::System& system, int32_t* out_index, cons
     // Get the synchronization context.
     auto& kernel = system.Kernel();
     auto& handle_table = GetCurrentProcess(kernel).GetHandleTable();
-    std::vector<KSynchronizationObject*> objs(num_handles);
+    std::array<KSynchronizationObject*, Svc::ArgumentHandleCountMax> objs;
 
     // Copy user handles.
     if (num_handles > 0) {
@@ -72,8 +73,8 @@ static Result WaitSynchronization(Core::System& system, int32_t* out_index, cons
     });
 
     // Wait on the objects.
-    Result res = KSynchronizationObject::Wait(kernel, out_index, objs.data(),
-                                              static_cast<s32>(objs.size()), timeout_ns);
+    Result res =
+        KSynchronizationObject::Wait(kernel, out_index, objs.data(), num_handles, timeout_ns);
 
     R_SUCCEED_IF(res == ResultSessionClosed);
     R_RETURN(res);
@@ -87,8 +88,7 @@ Result WaitSynchronization(Core::System& system, int32_t* out_index, u64 user_ha
 
     // Ensure number of handles is valid.
     R_UNLESS(0 <= num_handles && num_handles <= Svc::ArgumentHandleCountMax, ResultOutOfRange);
-
-    std::vector<Handle> handles(num_handles);
+    std::array<Handle, Svc::ArgumentHandleCountMax> handles;
     if (num_handles > 0) {
         GetCurrentMemory(system.Kernel())
             .ReadBlock(user_handles, handles.data(), num_handles * sizeof(Handle));

src/core/hle/kernel/svc/svc_thread.cpp

@@ -174,7 +174,7 @@ Result GetThreadContext3(Core::System& system, u64 out_context, Handle thread_ha
         }
 
         // Get the thread context.
-        std::vector<u8> context;
+        static thread_local Common::ScratchBuffer<u8> context;
         R_TRY(thread->GetThreadContext3(context));
 
         // Copy the thread context to user space.

src/core/hle/service/audio/audin_u.cpp

@@ -5,6 +5,7 @@
 #include "audio_core/renderer/audio_device.h"
 #include "common/common_funcs.h"
 #include "common/logging/log.h"
+#include "common/settings.h"
 #include "common/string_util.h"
 #include "core/core.h"
 #include "core/hle/kernel/k_event.h"
@@ -123,19 +124,13 @@ private:
 
     void GetReleasedAudioInBuffer(HLERequestContext& ctx) {
         const auto write_buffer_size = ctx.GetWriteBufferNumElements<u64>();
-        std::vector<u64> released_buffers(write_buffer_size);
+        tmp_buffer.resize_destructive(write_buffer_size);
+        tmp_buffer[0] = 0;
 
-        const auto count = impl->GetReleasedBuffers(released_buffers);
+        const auto count = impl->GetReleasedBuffers(tmp_buffer);
 
-        [[maybe_unused]] std::string tags{};
-        for (u32 i = 0; i < count; i++) {
-            tags += fmt::format("{:08X}, ", released_buffers[i]);
-        }
-        [[maybe_unused]] auto sessionid{impl->GetSystem().GetSessionId()};
-        LOG_TRACE(Service_Audio, "called. Session {} released {} buffers: {}", sessionid, count,
-                  tags);
+        ctx.WriteBuffer(tmp_buffer);
 
-        ctx.WriteBuffer(released_buffers);
         IPC::ResponseBuilder rb{ctx, 3};
         rb.Push(ResultSuccess);
         rb.Push(count);
@@ -200,6 +195,7 @@ private:
     KernelHelpers::ServiceContext service_context;
     Kernel::KEvent* event;
     std::shared_ptr<AudioCore::AudioIn::In> impl;
+    Common::ScratchBuffer<u64> tmp_buffer;
 };
 
 AudInU::AudInU(Core::System& system_)

src/core/hle/service/audio/audout_u.cpp

@@ -123,19 +123,13 @@ private:
 
     void GetReleasedAudioOutBuffers(HLERequestContext& ctx) {
         const auto write_buffer_size = ctx.GetWriteBufferNumElements<u64>();
-        std::vector<u64> released_buffers(write_buffer_size);
+        tmp_buffer.resize_destructive(write_buffer_size);
+        tmp_buffer[0] = 0;
 
-        const auto count = impl->GetReleasedBuffers(released_buffers);
+        const auto count = impl->GetReleasedBuffers(tmp_buffer);
 
-        [[maybe_unused]] std::string tags{};
-        for (u32 i = 0; i < count; i++) {
-            tags += fmt::format("{:08X}, ", released_buffers[i]);
-        }
-        [[maybe_unused]] const auto sessionid{impl->GetSystem().GetSessionId()};
-        LOG_TRACE(Service_Audio, "called. Session {} released {} buffers: {}", sessionid, count,
-                  tags);
+        ctx.WriteBuffer(tmp_buffer);
 
-        ctx.WriteBuffer(released_buffers);
         IPC::ResponseBuilder rb{ctx, 3};
         rb.Push(ResultSuccess);
         rb.Push(count);
@@ -211,6 +205,7 @@ private:
     KernelHelpers::ServiceContext service_context;
     Kernel::KEvent* event;
     std::shared_ptr<AudioCore::AudioOut::Out> impl;
+    Common::ScratchBuffer<u64> tmp_buffer;
 };
 
 AudOutU::AudOutU(Core::System& system_)

src/core/hle/service/audio/audren_u.cpp

@@ -116,28 +116,26 @@ private:
         // These buffers are written manually to avoid an issue with WriteBuffer throwing errors for
         // checking size 0. Performance size is 0 for most games.
 
-        std::vector<u8> output{};
-        std::vector<u8> performance{};
         auto is_buffer_b{ctx.BufferDescriptorB()[0].Size() != 0};
         if (is_buffer_b) {
             const auto buffersB{ctx.BufferDescriptorB()};
-            output.resize(buffersB[0].Size(), 0);
-            performance.resize(buffersB[1].Size(), 0);
+            tmp_output.resize_destructive(buffersB[0].Size());
+            tmp_performance.resize_destructive(buffersB[1].Size());
         } else {
             const auto buffersC{ctx.BufferDescriptorC()};
-            output.resize(buffersC[0].Size(), 0);
-            performance.resize(buffersC[1].Size(), 0);
+            tmp_output.resize_destructive(buffersC[0].Size());
+            tmp_performance.resize_destructive(buffersC[1].Size());
         }
 
-        auto result = impl->RequestUpdate(input, performance, output);
+        auto result = impl->RequestUpdate(input, tmp_performance, tmp_output);
 
         if (result.IsSuccess()) {
             if (is_buffer_b) {
-                ctx.WriteBufferB(output.data(), output.size(), 0);
-                ctx.WriteBufferB(performance.data(), performance.size(), 1);
+                ctx.WriteBufferB(tmp_output.data(), tmp_output.size(), 0);
+                ctx.WriteBufferB(tmp_performance.data(), tmp_performance.size(), 1);
             } else {
-                ctx.WriteBufferC(output.data(), output.size(), 0);
-                ctx.WriteBufferC(performance.data(), performance.size(), 1);
+                ctx.WriteBufferC(tmp_output.data(), tmp_output.size(), 0);
+                ctx.WriteBufferC(tmp_performance.data(), tmp_performance.size(), 1);
             }
         } else {
             LOG_ERROR(Service_Audio, "RequestUpdate failed error 0x{:02X}!", result.description);
@@ -235,6 +233,8 @@ private:
     Kernel::KEvent* rendered_event;
     Manager& manager;
     std::unique_ptr<Renderer> impl;
+    Common::ScratchBuffer<u8> tmp_output;
+    Common::ScratchBuffer<u8> tmp_performance;
 };
 
 class IAudioDevice final : public ServiceFramework<IAudioDevice> {

src/core/hle/service/audio/audren_u.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "audio_core/audio_render_manager.h"
+#include "common/scratch_buffer.h"
 #include "core/hle/service/kernel_helpers.h"
 #include "core/hle/service/service.h"
 

src/core/hle/service/audio/hwopus.cpp

@@ -68,13 +68,13 @@ private:
                                  ExtraBehavior extra_behavior) {
         u32 consumed = 0;
         u32 sample_count = 0;
-        std::vector<opus_int16> samples(ctx.GetWriteBufferNumElements<opus_int16>());
+        tmp_samples.resize_destructive(ctx.GetWriteBufferNumElements<opus_int16>());
 
         if (extra_behavior == ExtraBehavior::ResetContext) {
             ResetDecoderContext();
         }
 
-        if (!DecodeOpusData(consumed, sample_count, ctx.ReadBuffer(), samples, performance)) {
+        if (!DecodeOpusData(consumed, sample_count, ctx.ReadBuffer(), tmp_samples, performance)) {
             LOG_ERROR(Audio, "Failed to decode opus data");
             IPC::ResponseBuilder rb{ctx, 2};
             // TODO(ogniK): Use correct error code
@@ -90,11 +90,11 @@ private:
         if (performance) {
             rb.Push<u64>(*performance);
         }
-        ctx.WriteBuffer(samples);
+        ctx.WriteBuffer(tmp_samples);
     }
 
     bool DecodeOpusData(u32& consumed, u32& sample_count, std::span<const u8> input,
-                        std::vector<opus_int16>& output, u64* out_performance_time) const {
+                        std::span<opus_int16> output, u64* out_performance_time) const {
         const auto start_time = std::chrono::steady_clock::now();
         const std::size_t raw_output_sz = output.size() * sizeof(opus_int16);
         if (sizeof(OpusPacketHeader) > input.size()) {
@@ -154,6 +154,7 @@ private:
     OpusDecoderPtr decoder;
     u32 sample_rate;
     u32 channel_count;
+    Common::ScratchBuffer<opus_int16> tmp_samples;
 };
 
 class IHardwareOpusDecoderManager final : public ServiceFramework<IHardwareOpusDecoderManager> {

src/core/hle/service/nvdrv/devices/nvdevice.h

@@ -34,7 +34,7 @@ public:
      * @returns The result code of the ioctl.
      */
     virtual NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output) = 0;
+                            std::span<u8> output) = 0;
 
     /**
      * Handles an ioctl2 request.
@@ -45,7 +45,7 @@ public:
      * @returns The result code of the ioctl.
      */
     virtual NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::span<const u8> inline_input, std::vector<u8>& output) = 0;
+                            std::span<const u8> inline_input, std::span<u8> output) = 0;
 
     /**
      * Handles an ioctl3 request.
@@ -56,7 +56,7 @@ public:
      * @returns The result code of the ioctl.
      */
     virtual NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output, std::vector<u8>& inline_output) = 0;
+                            std::span<u8> output, std::span<u8> inline_output) = 0;
 
     /**
      * Called once a device is opened

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -18,19 +18,19 @@ nvdisp_disp0::nvdisp_disp0(Core::System& system_, NvCore::Container& core)
 nvdisp_disp0::~nvdisp_disp0() = default;
 
 NvResult nvdisp_disp0::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output) {
+                              std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvdisp_disp0::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::span<const u8> inline_input, std::vector<u8>& output) {
+                              std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvdisp_disp0::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output, std::vector<u8>& inline_output) {
+                              std::span<u8> output, std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }

src/core/hle/service/nvdrv/devices/nvdisp_disp0.h

@@ -26,11 +26,11 @@ public:
     ~nvdisp_disp0() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;

src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp

@@ -28,7 +28,7 @@ nvhost_as_gpu::nvhost_as_gpu(Core::System& system_, Module& module_, NvCore::Con
 nvhost_as_gpu::~nvhost_as_gpu() = default;
 
 NvResult nvhost_as_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                               std::vector<u8>& output) {
+                               std::span<u8> output) {
     switch (command.group) {
     case 'A':
         switch (command.cmd) {
@@ -61,13 +61,13 @@ NvResult nvhost_as_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> i
 }
 
 NvResult nvhost_as_gpu::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                               std::span<const u8> inline_input, std::vector<u8>& output) {
+                               std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_as_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                               std::vector<u8>& output, std::vector<u8>& inline_output) {
+                               std::span<u8> output, std::span<u8> inline_output) {
     switch (command.group) {
     case 'A':
         switch (command.cmd) {
@@ -87,7 +87,7 @@ NvResult nvhost_as_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> i
 void nvhost_as_gpu::OnOpen(DeviceFD fd) {}
 void nvhost_as_gpu::OnClose(DeviceFD fd) {}
 
-NvResult nvhost_as_gpu::AllocAsEx(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::AllocAsEx(std::span<const u8> input, std::span<u8> output) {
     IoctlAllocAsEx params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -141,7 +141,7 @@ NvResult nvhost_as_gpu::AllocAsEx(std::span<const u8> input, std::vector<u8>& ou
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::AllocateSpace(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::AllocateSpace(std::span<const u8> input, std::span<u8> output) {
     IoctlAllocSpace params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -220,7 +220,7 @@ void nvhost_as_gpu::FreeMappingLocked(u64 offset) {
     mapping_map.erase(offset);
 }
 
-NvResult nvhost_as_gpu::FreeSpace(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::FreeSpace(std::span<const u8> input, std::span<u8> output) {
     IoctlFreeSpace params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -266,15 +266,14 @@ NvResult nvhost_as_gpu::FreeSpace(std::span<const u8> input, std::vector<u8>& ou
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::Remap(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::Remap(std::span<const u8> input, std::span<u8> output) {
     const auto num_entries = input.size() / sizeof(IoctlRemapEntry);
 
     LOG_DEBUG(Service_NVDRV, "called, num_entries=0x{:X}", num_entries);
 
-    std::vector<IoctlRemapEntry> entries(num_entries);
-    std::memcpy(entries.data(), input.data(), input.size());
-
     std::scoped_lock lock(mutex);
+    entries.resize_destructive(num_entries);
+    std::memcpy(entries.data(), input.data(), input.size());
 
     if (!vm.initialised) {
         return NvResult::BadValue;
@@ -320,7 +319,7 @@ NvResult nvhost_as_gpu::Remap(std::span<const u8> input, std::vector<u8>& output
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::MapBufferEx(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::MapBufferEx(std::span<const u8> input, std::span<u8> output) {
     IoctlMapBufferEx params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -424,7 +423,7 @@ NvResult nvhost_as_gpu::MapBufferEx(std::span<const u8> input, std::vector<u8>&
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::UnmapBuffer(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::UnmapBuffer(std::span<const u8> input, std::span<u8> output) {
     IoctlUnmapBuffer params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -463,7 +462,7 @@ NvResult nvhost_as_gpu::UnmapBuffer(std::span<const u8> input, std::vector<u8>&
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::BindChannel(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::BindChannel(std::span<const u8> input, std::span<u8> output) {
     IoctlBindChannel params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, fd={:X}", params.fd);
@@ -492,7 +491,7 @@ void nvhost_as_gpu::GetVARegionsImpl(IoctlGetVaRegions& params) {
     };
 }
 
-NvResult nvhost_as_gpu::GetVARegions(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_as_gpu::GetVARegions(std::span<const u8> input, std::span<u8> output) {
     IoctlGetVaRegions params{};
     std::memcpy(&params, input.data(), input.size());
 
@@ -511,8 +510,8 @@ NvResult nvhost_as_gpu::GetVARegions(std::span<const u8> input, std::vector<u8>&
     return NvResult::Success;
 }
 
-NvResult nvhost_as_gpu::GetVARegions(std::span<const u8> input, std::vector<u8>& output,
-                                     std::vector<u8>& inline_output) {
+NvResult nvhost_as_gpu::GetVARegions(std::span<const u8> input, std::span<u8> output,
+                                     std::span<u8> inline_output) {
     IoctlGetVaRegions params{};
     std::memcpy(&params, input.data(), input.size());
 

src/core/hle/service/nvdrv/devices/nvhost_as_gpu.h

@@ -15,6 +15,7 @@
 #include "common/address_space.h"
 #include "common/common_funcs.h"
 #include "common/common_types.h"
+#include "common/scratch_buffer.h"
 #include "common/swap.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
@@ -48,11 +49,11 @@ public:
     ~nvhost_as_gpu() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -138,18 +139,18 @@ private:
     static_assert(sizeof(IoctlGetVaRegions) == 16 + sizeof(VaRegion) * 2,
                   "IoctlGetVaRegions is incorrect size");
 
-    NvResult AllocAsEx(std::span<const u8> input, std::vector<u8>& output);
-    NvResult AllocateSpace(std::span<const u8> input, std::vector<u8>& output);
-    NvResult Remap(std::span<const u8> input, std::vector<u8>& output);
-    NvResult MapBufferEx(std::span<const u8> input, std::vector<u8>& output);
-    NvResult UnmapBuffer(std::span<const u8> input, std::vector<u8>& output);
-    NvResult FreeSpace(std::span<const u8> input, std::vector<u8>& output);
-    NvResult BindChannel(std::span<const u8> input, std::vector<u8>& output);
+    NvResult AllocAsEx(std::span<const u8> input, std::span<u8> output);
+    NvResult AllocateSpace(std::span<const u8> input, std::span<u8> output);
+    NvResult Remap(std::span<const u8> input, std::span<u8> output);
+    NvResult MapBufferEx(std::span<const u8> input, std::span<u8> output);
+    NvResult UnmapBuffer(std::span<const u8> input, std::span<u8> output);
+    NvResult FreeSpace(std::span<const u8> input, std::span<u8> output);
+    NvResult BindChannel(std::span<const u8> input, std::span<u8> output);
 
     void GetVARegionsImpl(IoctlGetVaRegions& params);
-    NvResult GetVARegions(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetVARegions(std::span<const u8> input, std::vector<u8>& output,
-                          std::vector<u8>& inline_output);
+    NvResult GetVARegions(std::span<const u8> input, std::span<u8> output);
+    NvResult GetVARegions(std::span<const u8> input, std::span<u8> output,
+                          std::span<u8> inline_output);
 
     void FreeMappingLocked(u64 offset);
 
@@ -212,6 +213,7 @@ private:
         bool initialised{};
     } vm;
     std::shared_ptr<Tegra::MemoryManager> gmmu;
+    Common::ScratchBuffer<IoctlRemapEntry> entries;
 
     // s32 channel{};
     // u32 big_page_size{VM::DEFAULT_BIG_PAGE_SIZE};

src/core/hle/service/nvdrv/devices/nvhost_ctrl.cpp

@@ -35,7 +35,7 @@ nvhost_ctrl::~nvhost_ctrl() {
 }
 
 NvResult nvhost_ctrl::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                             std::vector<u8>& output) {
+                             std::span<u8> output) {
     switch (command.group) {
     case 0x0:
         switch (command.cmd) {
@@ -64,13 +64,13 @@ NvResult nvhost_ctrl::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> inp
 }
 
 NvResult nvhost_ctrl::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                             std::span<const u8> inline_input, std::vector<u8>& output) {
+                             std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_ctrl::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                             std::vector<u8>& output, std::vector<u8>& inline_outpu) {
+                             std::span<u8> output, std::span<u8> inline_outpu) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
@@ -79,7 +79,7 @@ void nvhost_ctrl::OnOpen(DeviceFD fd) {}
 
 void nvhost_ctrl::OnClose(DeviceFD fd) {}
 
-NvResult nvhost_ctrl::NvOsGetConfigU32(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl::NvOsGetConfigU32(std::span<const u8> input, std::span<u8> output) {
     IocGetConfigParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
     LOG_TRACE(Service_NVDRV, "called, setting={}!{}", params.domain_str.data(),
@@ -87,7 +87,7 @@ NvResult nvhost_ctrl::NvOsGetConfigU32(std::span<const u8> input, std::vector<u8
     return NvResult::ConfigVarNotFound; // Returns error on production mode
 }
 
-NvResult nvhost_ctrl::IocCtrlEventWait(std::span<const u8> input, std::vector<u8>& output,
+NvResult nvhost_ctrl::IocCtrlEventWait(std::span<const u8> input, std::span<u8> output,
                                        bool is_allocation) {
     IocCtrlEventWaitParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
@@ -231,7 +231,7 @@ NvResult nvhost_ctrl::FreeEvent(u32 slot) {
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl::IocCtrlEventRegister(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl::IocCtrlEventRegister(std::span<const u8> input, std::span<u8> output) {
     IocCtrlEventRegisterParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
     const u32 event_id = params.user_event_id;
@@ -252,7 +252,7 @@ NvResult nvhost_ctrl::IocCtrlEventRegister(std::span<const u8> input, std::vecto
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl::IocCtrlEventUnregister(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl::IocCtrlEventUnregister(std::span<const u8> input, std::span<u8> output) {
     IocCtrlEventUnregisterParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
     const u32 event_id = params.user_event_id & 0x00FF;
@@ -262,8 +262,7 @@ NvResult nvhost_ctrl::IocCtrlEventUnregister(std::span<const u8> input, std::vec
     return FreeEvent(event_id);
 }
 
-NvResult nvhost_ctrl::IocCtrlEventUnregisterBatch(std::span<const u8> input,
-                                                  std::vector<u8>& output) {
+NvResult nvhost_ctrl::IocCtrlEventUnregisterBatch(std::span<const u8> input, std::span<u8> output) {
     IocCtrlEventUnregisterBatchParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
     u64 event_mask = params.user_events;
@@ -281,7 +280,7 @@ NvResult nvhost_ctrl::IocCtrlEventUnregisterBatch(std::span<const u8> input,
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl::IocCtrlClearEventWait(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl::IocCtrlClearEventWait(std::span<const u8> input, std::span<u8> output) {
     IocCtrlEventClearParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
 

src/core/hle/service/nvdrv/devices/nvhost_ctrl.h

@@ -26,11 +26,11 @@ public:
     ~nvhost_ctrl() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -186,13 +186,12 @@ private:
     static_assert(sizeof(IocCtrlEventUnregisterBatchParams) == 8,
                   "IocCtrlEventKill is incorrect size");
 
-    NvResult NvOsGetConfigU32(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocCtrlEventWait(std::span<const u8> input, std::vector<u8>& output,
-                              bool is_allocation);
-    NvResult IocCtrlEventRegister(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocCtrlEventUnregister(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocCtrlEventUnregisterBatch(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocCtrlClearEventWait(std::span<const u8> input, std::vector<u8>& output);
+    NvResult NvOsGetConfigU32(std::span<const u8> input, std::span<u8> output);
+    NvResult IocCtrlEventWait(std::span<const u8> input, std::span<u8> output, bool is_allocation);
+    NvResult IocCtrlEventRegister(std::span<const u8> input, std::span<u8> output);
+    NvResult IocCtrlEventUnregister(std::span<const u8> input, std::span<u8> output);
+    NvResult IocCtrlEventUnregisterBatch(std::span<const u8> input, std::span<u8> output);
+    NvResult IocCtrlClearEventWait(std::span<const u8> input, std::span<u8> output);
 
     NvResult FreeEvent(u32 slot);
 

src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.cpp

@@ -22,7 +22,7 @@ nvhost_ctrl_gpu::~nvhost_ctrl_gpu() {
 }
 
 NvResult nvhost_ctrl_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                                 std::vector<u8>& output) {
+                                 std::span<u8> output) {
     switch (command.group) {
     case 'G':
         switch (command.cmd) {
@@ -54,13 +54,13 @@ NvResult nvhost_ctrl_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8>
 }
 
 NvResult nvhost_ctrl_gpu::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                                 std::span<const u8> inline_input, std::vector<u8>& output) {
+                                 std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_ctrl_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                                 std::vector<u8>& output, std::vector<u8>& inline_output) {
+                                 std::span<u8> output, std::span<u8> inline_output) {
     switch (command.group) {
     case 'G':
         switch (command.cmd) {
@@ -82,7 +82,7 @@ NvResult nvhost_ctrl_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8>
 void nvhost_ctrl_gpu::OnOpen(DeviceFD fd) {}
 void nvhost_ctrl_gpu::OnClose(DeviceFD fd) {}
 
-NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
     IoctlCharacteristics params{};
     std::memcpy(&params, input.data(), input.size());
@@ -127,8 +127,8 @@ NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::vec
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::vector<u8>& output,
-                                             std::vector<u8>& inline_output) {
+NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::span<u8> output,
+                                             std::span<u8> inline_output) {
     LOG_DEBUG(Service_NVDRV, "called");
     IoctlCharacteristics params{};
     std::memcpy(&params, input.data(), input.size());
@@ -175,7 +175,7 @@ NvResult nvhost_ctrl_gpu::GetCharacteristics(std::span<const u8> input, std::vec
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::span<u8> output) {
     IoctlGpuGetTpcMasksArgs params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, mask_buffer_size=0x{:X}", params.mask_buffer_size);
@@ -186,8 +186,8 @@ NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::vector<u8>
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::vector<u8>& output,
-                                      std::vector<u8>& inline_output) {
+NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::span<u8> output,
+                                      std::span<u8> inline_output) {
     IoctlGpuGetTpcMasksArgs params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, mask_buffer_size=0x{:X}", params.mask_buffer_size);
@@ -199,7 +199,7 @@ NvResult nvhost_ctrl_gpu::GetTPCMasks(std::span<const u8> input, std::vector<u8>
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::GetActiveSlotMask(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::GetActiveSlotMask(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlActiveSlotMask params{};
@@ -212,7 +212,7 @@ NvResult nvhost_ctrl_gpu::GetActiveSlotMask(std::span<const u8> input, std::vect
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::ZCullGetCtxSize(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::ZCullGetCtxSize(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlZcullGetCtxSize params{};
@@ -224,7 +224,7 @@ NvResult nvhost_ctrl_gpu::ZCullGetCtxSize(std::span<const u8> input, std::vector
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::ZCullGetInfo(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::ZCullGetInfo(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlNvgpuGpuZcullGetInfoArgs params{};
@@ -247,7 +247,7 @@ NvResult nvhost_ctrl_gpu::ZCullGetInfo(std::span<const u8> input, std::vector<u8
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::ZBCSetTable(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::ZBCSetTable(std::span<const u8> input, std::span<u8> output) {
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
 
     IoctlZbcSetTable params{};
@@ -263,7 +263,7 @@ NvResult nvhost_ctrl_gpu::ZBCSetTable(std::span<const u8> input, std::vector<u8>
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::ZBCQueryTable(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::ZBCQueryTable(std::span<const u8> input, std::span<u8> output) {
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
 
     IoctlZbcQueryTable params{};
@@ -273,7 +273,7 @@ NvResult nvhost_ctrl_gpu::ZBCQueryTable(std::span<const u8> input, std::vector<u
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::FlushL2(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::FlushL2(std::span<const u8> input, std::span<u8> output) {
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
 
     IoctlFlushL2 params{};
@@ -283,7 +283,7 @@ NvResult nvhost_ctrl_gpu::FlushL2(std::span<const u8> input, std::vector<u8>& ou
     return NvResult::Success;
 }
 
-NvResult nvhost_ctrl_gpu::GetGpuTime(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_ctrl_gpu::GetGpuTime(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlGetGpuTime params{};

src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.h

@@ -22,11 +22,11 @@ public:
     ~nvhost_ctrl_gpu() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -151,21 +151,21 @@ private:
     };
     static_assert(sizeof(IoctlGetGpuTime) == 0x10, "IoctlGetGpuTime is incorrect size");
 
-    NvResult GetCharacteristics(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetCharacteristics(std::span<const u8> input, std::vector<u8>& output,
-                                std::vector<u8>& inline_output);
+    NvResult GetCharacteristics(std::span<const u8> input, std::span<u8> output);
+    NvResult GetCharacteristics(std::span<const u8> input, std::span<u8> output,
+                                std::span<u8> inline_output);
 
-    NvResult GetTPCMasks(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetTPCMasks(std::span<const u8> input, std::vector<u8>& output,
-                         std::vector<u8>& inline_output);
+    NvResult GetTPCMasks(std::span<const u8> input, std::span<u8> output);
+    NvResult GetTPCMasks(std::span<const u8> input, std::span<u8> output,
+                         std::span<u8> inline_output);
 
-    NvResult GetActiveSlotMask(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ZCullGetCtxSize(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ZCullGetInfo(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ZBCSetTable(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ZBCQueryTable(std::span<const u8> input, std::vector<u8>& output);
-    NvResult FlushL2(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetGpuTime(std::span<const u8> input, std::vector<u8>& output);
+    NvResult GetActiveSlotMask(std::span<const u8> input, std::span<u8> output);
+    NvResult ZCullGetCtxSize(std::span<const u8> input, std::span<u8> output);
+    NvResult ZCullGetInfo(std::span<const u8> input, std::span<u8> output);
+    NvResult ZBCSetTable(std::span<const u8> input, std::span<u8> output);
+    NvResult ZBCQueryTable(std::span<const u8> input, std::span<u8> output);
+    NvResult FlushL2(std::span<const u8> input, std::span<u8> output);
+    NvResult GetGpuTime(std::span<const u8> input, std::span<u8> output);
 
     EventInterface& events_interface;
 

src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp

@@ -47,7 +47,7 @@ nvhost_gpu::~nvhost_gpu() {
 }
 
 NvResult nvhost_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output) {
+                            std::span<u8> output) {
     switch (command.group) {
     case 0x0:
         switch (command.cmd) {
@@ -99,7 +99,7 @@ NvResult nvhost_gpu::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> inpu
 };
 
 NvResult nvhost_gpu::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::span<const u8> inline_input, std::vector<u8>& output) {
+                            std::span<const u8> inline_input, std::span<u8> output) {
     switch (command.group) {
     case 'H':
         switch (command.cmd) {
@@ -113,7 +113,7 @@ NvResult nvhost_gpu::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> inpu
 }
 
 NvResult nvhost_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output, std::vector<u8>& inline_output) {
+                            std::span<u8> output, std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
@@ -121,7 +121,7 @@ NvResult nvhost_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> inpu
 void nvhost_gpu::OnOpen(DeviceFD fd) {}
 void nvhost_gpu::OnClose(DeviceFD fd) {}
 
-NvResult nvhost_gpu::SetNVMAPfd(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::SetNVMAPfd(std::span<const u8> input, std::span<u8> output) {
     IoctlSetNvmapFD params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, fd={}", params.nvmap_fd);
@@ -130,7 +130,7 @@ NvResult nvhost_gpu::SetNVMAPfd(std::span<const u8> input, std::vector<u8>& outp
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::SetClientData(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::SetClientData(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlClientData params{};
@@ -139,7 +139,7 @@ NvResult nvhost_gpu::SetClientData(std::span<const u8> input, std::vector<u8>& o
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::GetClientData(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::GetClientData(std::span<const u8> input, std::span<u8> output) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     IoctlClientData params{};
@@ -149,7 +149,7 @@ NvResult nvhost_gpu::GetClientData(std::span<const u8> input, std::vector<u8>& o
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::ZCullBind(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::ZCullBind(std::span<const u8> input, std::span<u8> output) {
     std::memcpy(&zcull_params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, gpu_va={:X}, mode={:X}", zcull_params.gpu_va,
               zcull_params.mode);
@@ -158,7 +158,7 @@ NvResult nvhost_gpu::ZCullBind(std::span<const u8> input, std::vector<u8>& outpu
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::SetErrorNotifier(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::SetErrorNotifier(std::span<const u8> input, std::span<u8> output) {
     IoctlSetErrorNotifier params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, offset={:X}, size={:X}, mem={:X}", params.offset,
@@ -168,14 +168,14 @@ NvResult nvhost_gpu::SetErrorNotifier(std::span<const u8> input, std::vector<u8>
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::SetChannelPriority(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::SetChannelPriority(std::span<const u8> input, std::span<u8> output) {
     std::memcpy(&channel_priority, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "(STUBBED) called, priority={:X}", channel_priority);
 
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::AllocGPFIFOEx2(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::AllocGPFIFOEx2(std::span<const u8> input, std::span<u8> output) {
     IoctlAllocGpfifoEx2 params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_WARNING(Service_NVDRV,
@@ -197,7 +197,7 @@ NvResult nvhost_gpu::AllocGPFIFOEx2(std::span<const u8> input, std::vector<u8>&
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::AllocateObjectContext(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::AllocateObjectContext(std::span<const u8> input, std::span<u8> output) {
     IoctlAllocObjCtx params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, class_num={:X}, flags={:X}", params.class_num,
@@ -208,7 +208,8 @@ NvResult nvhost_gpu::AllocateObjectContext(std::span<const u8> input, std::vecto
     return NvResult::Success;
 }
 
-static std::vector<Tegra::CommandHeader> BuildWaitCommandList(NvFence fence) {
+static boost::container::small_vector<Tegra::CommandHeader, 512> BuildWaitCommandList(
+    NvFence fence) {
     return {
         Tegra::BuildCommandHeader(Tegra::BufferMethods::SyncpointPayload, 1,
                                   Tegra::SubmissionMode::Increasing),
@@ -219,35 +220,35 @@ static std::vector<Tegra::CommandHeader> BuildWaitCommandList(NvFence fence) {
     };
 }
 
-static std::vector<Tegra::CommandHeader> BuildIncrementCommandList(NvFence fence) {
-    std::vector<Tegra::CommandHeader> result{
+static boost::container::small_vector<Tegra::CommandHeader, 512> BuildIncrementCommandList(
+    NvFence fence) {
+    boost::container::small_vector<Tegra::CommandHeader, 512> result{
         Tegra::BuildCommandHeader(Tegra::BufferMethods::SyncpointPayload, 1,
                                   Tegra::SubmissionMode::Increasing),
         {}};
 
     for (u32 count = 0; count < 2; ++count) {
-        result.emplace_back(Tegra::BuildCommandHeader(Tegra::BufferMethods::SyncpointOperation, 1,
-                                                      Tegra::SubmissionMode::Increasing));
-        result.emplace_back(
+        result.push_back(Tegra::BuildCommandHeader(Tegra::BufferMethods::SyncpointOperation, 1,
+                                                   Tegra::SubmissionMode::Increasing));
+        result.push_back(
             BuildFenceAction(Tegra::Engines::Puller::FenceOperation::Increment, fence.id));
     }
 
     return result;
 }
 
-static std::vector<Tegra::CommandHeader> BuildIncrementWithWfiCommandList(NvFence fence) {
-    std::vector<Tegra::CommandHeader> result{
+static boost::container::small_vector<Tegra::CommandHeader, 512> BuildIncrementWithWfiCommandList(
+    NvFence fence) {
+    boost::container::small_vector<Tegra::CommandHeader, 512> result{
         Tegra::BuildCommandHeader(Tegra::BufferMethods::WaitForIdle, 1,
                                   Tegra::SubmissionMode::Increasing),
         {}};
-    const std::vector<Tegra::CommandHeader> increment{BuildIncrementCommandList(fence)};
-
-    result.insert(result.end(), increment.begin(), increment.end());
-
+    auto increment_list{BuildIncrementCommandList(fence)};
+    result.insert(result.end(), increment_list.begin(), increment_list.end());
     return result;
 }
 
-NvResult nvhost_gpu::SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::vector<u8>& output,
+NvResult nvhost_gpu::SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::span<u8> output,
                                       Tegra::CommandList&& entries) {
     LOG_TRACE(Service_NVDRV, "called, gpfifo={:X}, num_entries={:X}, flags={:X}", params.address,
               params.num_entries, params.flags.raw);
@@ -293,7 +294,7 @@ NvResult nvhost_gpu::SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::vector<u8>
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::SubmitGPFIFOBase(std::span<const u8> input, std::vector<u8>& output,
+NvResult nvhost_gpu::SubmitGPFIFOBase(std::span<const u8> input, std::span<u8> output,
                                       bool kickoff) {
     if (input.size() < sizeof(IoctlSubmitGpfifo)) {
         UNIMPLEMENTED();
@@ -315,7 +316,7 @@ NvResult nvhost_gpu::SubmitGPFIFOBase(std::span<const u8> input, std::vector<u8>
 }
 
 NvResult nvhost_gpu::SubmitGPFIFOBase(std::span<const u8> input, std::span<const u8> input_inline,
-                                      std::vector<u8>& output) {
+                                      std::span<u8> output) {
     if (input.size() < sizeof(IoctlSubmitGpfifo)) {
         UNIMPLEMENTED();
         return NvResult::InvalidSize;
@@ -327,7 +328,7 @@ NvResult nvhost_gpu::SubmitGPFIFOBase(std::span<const u8> input, std::span<const
     return SubmitGPFIFOImpl(params, output, std::move(entries));
 }
 
-NvResult nvhost_gpu::GetWaitbase(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::GetWaitbase(std::span<const u8> input, std::span<u8> output) {
     IoctlGetWaitbase params{};
     std::memcpy(&params, input.data(), sizeof(IoctlGetWaitbase));
     LOG_INFO(Service_NVDRV, "called, unknown=0x{:X}", params.unknown);
@@ -337,7 +338,7 @@ NvResult nvhost_gpu::GetWaitbase(std::span<const u8> input, std::vector<u8>& out
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::ChannelSetTimeout(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::ChannelSetTimeout(std::span<const u8> input, std::span<u8> output) {
     IoctlChannelSetTimeout params{};
     std::memcpy(&params, input.data(), sizeof(IoctlChannelSetTimeout));
     LOG_INFO(Service_NVDRV, "called, timeout=0x{:X}", params.timeout);
@@ -345,7 +346,7 @@ NvResult nvhost_gpu::ChannelSetTimeout(std::span<const u8> input, std::vector<u8
     return NvResult::Success;
 }
 
-NvResult nvhost_gpu::ChannelSetTimeslice(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_gpu::ChannelSetTimeslice(std::span<const u8> input, std::span<u8> output) {
     IoctlSetTimeslice params{};
     std::memcpy(&params, input.data(), sizeof(IoctlSetTimeslice));
     LOG_INFO(Service_NVDRV, "called, timeslice=0x{:X}", params.timeslice);

src/core/hle/service/nvdrv/devices/nvhost_gpu.h

@@ -41,11 +41,11 @@ public:
     ~nvhost_gpu() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -186,23 +186,23 @@ private:
     u32_le channel_priority{};
     u32_le channel_timeslice{};
 
-    NvResult SetNVMAPfd(std::span<const u8> input, std::vector<u8>& output);
-    NvResult SetClientData(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetClientData(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ZCullBind(std::span<const u8> input, std::vector<u8>& output);
-    NvResult SetErrorNotifier(std::span<const u8> input, std::vector<u8>& output);
-    NvResult SetChannelPriority(std::span<const u8> input, std::vector<u8>& output);
-    NvResult AllocGPFIFOEx2(std::span<const u8> input, std::vector<u8>& output);
-    NvResult AllocateObjectContext(std::span<const u8> input, std::vector<u8>& output);
-    NvResult SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::vector<u8>& output,
+    NvResult SetNVMAPfd(std::span<const u8> input, std::span<u8> output);
+    NvResult SetClientData(std::span<const u8> input, std::span<u8> output);
+    NvResult GetClientData(std::span<const u8> input, std::span<u8> output);
+    NvResult ZCullBind(std::span<const u8> input, std::span<u8> output);
+    NvResult SetErrorNotifier(std::span<const u8> input, std::span<u8> output);
+    NvResult SetChannelPriority(std::span<const u8> input, std::span<u8> output);
+    NvResult AllocGPFIFOEx2(std::span<const u8> input, std::span<u8> output);
+    NvResult AllocateObjectContext(std::span<const u8> input, std::span<u8> output);
+    NvResult SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::span<u8> output,
                               Tegra::CommandList&& entries);
-    NvResult SubmitGPFIFOBase(std::span<const u8> input, std::vector<u8>& output,
+    NvResult SubmitGPFIFOBase(std::span<const u8> input, std::span<u8> output,
                               bool kickoff = false);
     NvResult SubmitGPFIFOBase(std::span<const u8> input, std::span<const u8> input_inline,
-                              std::vector<u8>& output);
-    NvResult GetWaitbase(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ChannelSetTimeout(std::span<const u8> input, std::vector<u8>& output);
-    NvResult ChannelSetTimeslice(std::span<const u8> input, std::vector<u8>& output);
+                              std::span<u8> output);
+    NvResult GetWaitbase(std::span<const u8> input, std::span<u8> output);
+    NvResult ChannelSetTimeout(std::span<const u8> input, std::span<u8> output);
+    NvResult ChannelSetTimeslice(std::span<const u8> input, std::span<u8> output);
 
     EventInterface& events_interface;
     NvCore::Container& core;

src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp

@@ -16,7 +16,7 @@ nvhost_nvdec::nvhost_nvdec(Core::System& system_, NvCore::Container& core_)
 nvhost_nvdec::~nvhost_nvdec() = default;
 
 NvResult nvhost_nvdec::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output) {
+                              std::span<u8> output) {
     switch (command.group) {
     case 0x0:
         switch (command.cmd) {
@@ -56,13 +56,13 @@ NvResult nvhost_nvdec::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> in
 }
 
 NvResult nvhost_nvdec::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::span<const u8> inline_input, std::vector<u8>& output) {
+                              std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_nvdec::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output, std::vector<u8>& inline_output) {
+                              std::span<u8> output, std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }

src/core/hle/service/nvdrv/devices/nvhost_nvdec.h

@@ -14,11 +14,11 @@ public:
     ~nvhost_nvdec() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp

@@ -36,7 +36,7 @@ std::size_t SliceVectors(std::span<const u8> input, std::vector<T>& dst, std::si
 // Writes the data in src to an offset into the dst vector. The offset is specified in bytes
 // Returns the number of bytes written into dst.
 template <typename T>
-std::size_t WriteVectors(std::vector<u8>& dst, const std::vector<T>& src, std::size_t offset) {
+std::size_t WriteVectors(std::span<u8> dst, const std::vector<T>& src, std::size_t offset) {
     if (src.empty()) {
         return 0;
     }
@@ -72,8 +72,7 @@ NvResult nvhost_nvdec_common::SetNVMAPfd(std::span<const u8> input) {
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::Submit(DeviceFD fd, std::span<const u8> input,
-                                     std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::Submit(DeviceFD fd, std::span<const u8> input, std::span<u8> output) {
     IoctlSubmit params{};
     std::memcpy(&params, input.data(), sizeof(IoctlSubmit));
     LOG_DEBUG(Service_NVDRV, "called NVDEC Submit, cmd_buffer_count={}", params.cmd_buffer_count);
@@ -121,7 +120,7 @@ NvResult nvhost_nvdec_common::Submit(DeviceFD fd, std::span<const u8> input,
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::GetSyncpoint(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::GetSyncpoint(std::span<const u8> input, std::span<u8> output) {
     IoctlGetSyncpoint params{};
     std::memcpy(&params, input.data(), sizeof(IoctlGetSyncpoint));
     LOG_DEBUG(Service_NVDRV, "called GetSyncpoint, id={}", params.param);
@@ -133,7 +132,7 @@ NvResult nvhost_nvdec_common::GetSyncpoint(std::span<const u8> input, std::vecto
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::GetWaitbase(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::GetWaitbase(std::span<const u8> input, std::span<u8> output) {
     IoctlGetWaitbase params{};
     LOG_CRITICAL(Service_NVDRV, "called WAITBASE");
     std::memcpy(&params, input.data(), sizeof(IoctlGetWaitbase));
@@ -142,7 +141,7 @@ NvResult nvhost_nvdec_common::GetWaitbase(std::span<const u8> input, std::vector
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::MapBuffer(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::MapBuffer(std::span<const u8> input, std::span<u8> output) {
     IoctlMapBuffer params{};
     std::memcpy(&params, input.data(), sizeof(IoctlMapBuffer));
     std::vector<MapBufferEntry> cmd_buffer_handles(params.num_entries);
@@ -159,7 +158,7 @@ NvResult nvhost_nvdec_common::MapBuffer(std::span<const u8> input, std::vector<u
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::UnmapBuffer(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::UnmapBuffer(std::span<const u8> input, std::span<u8> output) {
     IoctlMapBuffer params{};
     std::memcpy(&params, input.data(), sizeof(IoctlMapBuffer));
     std::vector<MapBufferEntry> cmd_buffer_handles(params.num_entries);
@@ -173,7 +172,7 @@ NvResult nvhost_nvdec_common::UnmapBuffer(std::span<const u8> input, std::vector
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::SetSubmitTimeout(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvdec_common::SetSubmitTimeout(std::span<const u8> input, std::span<u8> output) {
     std::memcpy(&submit_timeout, input.data(), input.size());
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
     return NvResult::Success;

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h

@@ -108,12 +108,12 @@ protected:
 
     /// Ioctl command implementations
     NvResult SetNVMAPfd(std::span<const u8> input);
-    NvResult Submit(DeviceFD fd, std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetSyncpoint(std::span<const u8> input, std::vector<u8>& output);
-    NvResult GetWaitbase(std::span<const u8> input, std::vector<u8>& output);
-    NvResult MapBuffer(std::span<const u8> input, std::vector<u8>& output);
-    NvResult UnmapBuffer(std::span<const u8> input, std::vector<u8>& output);
-    NvResult SetSubmitTimeout(std::span<const u8> input, std::vector<u8>& output);
+    NvResult Submit(DeviceFD fd, std::span<const u8> input, std::span<u8> output);
+    NvResult GetSyncpoint(std::span<const u8> input, std::span<u8> output);
+    NvResult GetWaitbase(std::span<const u8> input, std::span<u8> output);
+    NvResult MapBuffer(std::span<const u8> input, std::span<u8> output);
+    NvResult UnmapBuffer(std::span<const u8> input, std::span<u8> output);
+    NvResult SetSubmitTimeout(std::span<const u8> input, std::span<u8> output);
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;
 

src/core/hle/service/nvdrv/devices/nvhost_nvjpg.cpp

@@ -13,7 +13,7 @@ nvhost_nvjpg::nvhost_nvjpg(Core::System& system_) : nvdevice{system_} {}
 nvhost_nvjpg::~nvhost_nvjpg() = default;
 
 NvResult nvhost_nvjpg::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output) {
+                              std::span<u8> output) {
     switch (command.group) {
     case 'H':
         switch (command.cmd) {
@@ -32,13 +32,13 @@ NvResult nvhost_nvjpg::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> in
 }
 
 NvResult nvhost_nvjpg::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::span<const u8> inline_input, std::vector<u8>& output) {
+                              std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_nvjpg::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                              std::vector<u8>& output, std::vector<u8>& inline_output) {
+                              std::span<u8> output, std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
@@ -46,7 +46,7 @@ NvResult nvhost_nvjpg::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> in
 void nvhost_nvjpg::OnOpen(DeviceFD fd) {}
 void nvhost_nvjpg::OnClose(DeviceFD fd) {}
 
-NvResult nvhost_nvjpg::SetNVMAPfd(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvhost_nvjpg::SetNVMAPfd(std::span<const u8> input, std::span<u8> output) {
     IoctlSetNvmapFD params{};
     std::memcpy(&params, input.data(), input.size());
     LOG_DEBUG(Service_NVDRV, "called, fd={}", params.nvmap_fd);

src/core/hle/service/nvdrv/devices/nvhost_nvjpg.h

@@ -16,11 +16,11 @@ public:
     ~nvhost_nvjpg() override;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -33,7 +33,7 @@ private:
 
     s32_le nvmap_fd{};
 
-    NvResult SetNVMAPfd(std::span<const u8> input, std::vector<u8>& output);
+    NvResult SetNVMAPfd(std::span<const u8> input, std::span<u8> output);
 };
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_vic.cpp

@@ -16,7 +16,7 @@ nvhost_vic::nvhost_vic(Core::System& system_, NvCore::Container& core_)
 nvhost_vic::~nvhost_vic() = default;
 
 NvResult nvhost_vic::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output) {
+                            std::span<u8> output) {
     switch (command.group) {
     case 0x0:
         switch (command.cmd) {
@@ -56,13 +56,13 @@ NvResult nvhost_vic::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> inpu
 }
 
 NvResult nvhost_vic::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::span<const u8> inline_input, std::vector<u8>& output) {
+                            std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
 NvResult nvhost_vic::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                            std::vector<u8>& output, std::vector<u8>& inline_output) {
+                            std::span<u8> output, std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }

src/core/hle/service/nvdrv/devices/nvhost_vic.h

@@ -13,11 +13,11 @@ public:
     ~nvhost_vic();
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;

src/core/hle/service/nvdrv/devices/nvmap.cpp

@@ -26,7 +26,7 @@ nvmap::nvmap(Core::System& system_, NvCore::Container& container_)
 nvmap::~nvmap() = default;
 
 NvResult nvmap::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                       std::vector<u8>& output) {
+                       std::span<u8> output) {
     switch (command.group) {
     case 0x1:
         switch (command.cmd) {
@@ -55,13 +55,13 @@ NvResult nvmap::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
 }
 
 NvResult nvmap::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                       std::span<const u8> inline_input, std::vector<u8>& output) {
+                       std::span<const u8> inline_input, std::span<u8> output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
 
-NvResult nvmap::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                       std::vector<u8>& output, std::vector<u8>& inline_output) {
+NvResult nvmap::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                       std::span<u8> inline_output) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl={:08X}", command.raw);
     return NvResult::NotImplemented;
 }
@@ -69,7 +69,7 @@ NvResult nvmap::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
 void nvmap::OnOpen(DeviceFD fd) {}
 void nvmap::OnClose(DeviceFD fd) {}
 
-NvResult nvmap::IocCreate(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocCreate(std::span<const u8> input, std::span<u8> output) {
     IocCreateParams params;
     std::memcpy(&params, input.data(), sizeof(params));
     LOG_DEBUG(Service_NVDRV, "called, size=0x{:08X}", params.size);
@@ -89,7 +89,7 @@ NvResult nvmap::IocCreate(std::span<const u8> input, std::vector<u8>& output) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocAlloc(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocAlloc(std::span<const u8> input, std::span<u8> output) {
     IocAllocParams params;
     std::memcpy(&params, input.data(), sizeof(params));
     LOG_DEBUG(Service_NVDRV, "called, addr={:X}", params.address);
@@ -137,7 +137,7 @@ NvResult nvmap::IocAlloc(std::span<const u8> input, std::vector<u8>& output) {
     return result;
 }
 
-NvResult nvmap::IocGetId(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocGetId(std::span<const u8> input, std::span<u8> output) {
     IocGetIdParams params;
     std::memcpy(&params, input.data(), sizeof(params));
 
@@ -161,7 +161,7 @@ NvResult nvmap::IocGetId(std::span<const u8> input, std::vector<u8>& output) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocFromId(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocFromId(std::span<const u8> input, std::span<u8> output) {
     IocFromIdParams params;
     std::memcpy(&params, input.data(), sizeof(params));
 
@@ -192,7 +192,7 @@ NvResult nvmap::IocFromId(std::span<const u8> input, std::vector<u8>& output) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocParam(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocParam(std::span<const u8> input, std::span<u8> output) {
     enum class ParamTypes { Size = 1, Alignment = 2, Base = 3, Heap = 4, Kind = 5, Compr = 6 };
 
     IocParamParams params;
@@ -241,7 +241,7 @@ NvResult nvmap::IocParam(std::span<const u8> input, std::vector<u8>& output) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocFree(std::span<const u8> input, std::vector<u8>& output) {
+NvResult nvmap::IocFree(std::span<const u8> input, std::span<u8> output) {
     IocFreeParams params;
     std::memcpy(&params, input.data(), sizeof(params));
 

src/core/hle/service/nvdrv/devices/nvmap.h

@@ -27,11 +27,11 @@ public:
     nvmap& operator=(const nvmap&) = delete;
 
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::vector<u8>& output) override;
+                    std::span<u8> output) override;
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output) override;
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output) override;
+                    std::span<const u8> inline_input, std::span<u8> output) override;
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output) override;
 
     void OnOpen(DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
@@ -106,12 +106,12 @@ private:
     };
     static_assert(sizeof(IocGetIdParams) == 8, "IocGetIdParams has wrong size");
 
-    NvResult IocCreate(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocAlloc(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocGetId(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocFromId(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocParam(std::span<const u8> input, std::vector<u8>& output);
-    NvResult IocFree(std::span<const u8> input, std::vector<u8>& output);
+    NvResult IocCreate(std::span<const u8> input, std::span<u8> output);
+    NvResult IocAlloc(std::span<const u8> input, std::span<u8> output);
+    NvResult IocGetId(std::span<const u8> input, std::span<u8> output);
+    NvResult IocFromId(std::span<const u8> input, std::span<u8> output);
+    NvResult IocParam(std::span<const u8> input, std::span<u8> output);
+    NvResult IocFree(std::span<const u8> input, std::span<u8> output);
 
     NvCore::Container& container;
     NvCore::NvMap& file;

src/core/hle/service/nvdrv/nvdrv.cpp

@@ -130,7 +130,7 @@ DeviceFD Module::Open(const std::string& device_name) {
 }
 
 NvResult Module::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                        std::vector<u8>& output) {
+                        std::span<u8> output) {
     if (fd < 0) {
         LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
         return NvResult::InvalidState;
@@ -147,7 +147,7 @@ NvResult Module::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
 }
 
 NvResult Module::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                        std::span<const u8> inline_input, std::vector<u8>& output) {
+                        std::span<const u8> inline_input, std::span<u8> output) {
     if (fd < 0) {
         LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
         return NvResult::InvalidState;
@@ -163,8 +163,8 @@ NvResult Module::Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
     return itr->second->Ioctl2(fd, command, input, inline_input, output);
 }
 
-NvResult Module::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                        std::vector<u8>& output, std::vector<u8>& inline_output) {
+NvResult Module::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                        std::span<u8> inline_output) {
     if (fd < 0) {
         LOG_ERROR(Service_NVDRV, "Invalid DeviceFD={}!", fd);
         return NvResult::InvalidState;

src/core/hle/service/nvdrv/nvdrv.h

@@ -80,13 +80,13 @@ public:
     DeviceFD Open(const std::string& device_name);
 
     /// Sends an ioctl command to the specified file descriptor.
-    NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output);
+    NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output);
 
     NvResult Ioctl2(DeviceFD fd, Ioctl command, std::span<const u8> input,
-                    std::span<const u8> inline_input, std::vector<u8>& output);
+                    std::span<const u8> inline_input, std::span<u8> output);
 
-    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::vector<u8>& output,
-                    std::vector<u8>& inline_output);
+    NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
+                    std::span<u8> inline_output);
 
     /// Closes a device file descriptor and returns operation success.
     NvResult Close(DeviceFD fd);

src/core/hle/service/nvdrv/nvdrv_interface.cpp

@@ -63,12 +63,12 @@ void NVDRV::Ioctl1(HLERequestContext& ctx) {
     }
 
     // Check device
-    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));
+    tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
     const auto input_buffer = ctx.ReadBuffer(0);
 
-    const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, output_buffer);
+    const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, tmp_output);
     if (command.is_out != 0) {
-        ctx.WriteBuffer(output_buffer);
+        ctx.WriteBuffer(tmp_output);
     }
 
     IPC::ResponseBuilder rb{ctx, 3};
@@ -90,12 +90,12 @@ void NVDRV::Ioctl2(HLERequestContext& ctx) {
 
     const auto input_buffer = ctx.ReadBuffer(0);
     const auto input_inlined_buffer = ctx.ReadBuffer(1);
-    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));
+    tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
 
     const auto nv_result =
-        nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, output_buffer);
+        nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, tmp_output);
     if (command.is_out != 0) {
-        ctx.WriteBuffer(output_buffer);
+        ctx.WriteBuffer(tmp_output);
     }
 
     IPC::ResponseBuilder rb{ctx, 3};
@@ -116,14 +116,12 @@ void NVDRV::Ioctl3(HLERequestContext& ctx) {
     }
 
     const auto input_buffer = ctx.ReadBuffer(0);
-    std::vector<u8> output_buffer(ctx.GetWriteBufferSize(0));
-    std::vector<u8> output_buffer_inline(ctx.GetWriteBufferSize(1));
-
-    const auto nv_result =
-        nvdrv->Ioctl3(fd, command, input_buffer, output_buffer, output_buffer_inline);
+    tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
+    tmp_output_inline.resize_destructive(ctx.GetWriteBufferSize(1));
+    const auto nv_result = nvdrv->Ioctl3(fd, command, input_buffer, tmp_output, tmp_output_inline);
     if (command.is_out != 0) {
-        ctx.WriteBuffer(output_buffer, 0);
-        ctx.WriteBuffer(output_buffer_inline, 1);
+        ctx.WriteBuffer(tmp_output, 0);
+        ctx.WriteBuffer(tmp_output_inline, 1);
     }
 
     IPC::ResponseBuilder rb{ctx, 3};

src/core/hle/service/nvdrv/nvdrv_interface.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include <memory>
+#include "common/scratch_buffer.h"
 #include "core/hle/service/nvdrv/nvdrv.h"
 #include "core/hle/service/service.h"
 
@@ -33,6 +34,8 @@ private:
 
     u64 pid{};
     bool is_initialized{};
+    Common::ScratchBuffer<u8> tmp_output;
+    Common::ScratchBuffer<u8> tmp_output_inline;
 };
 
 } // namespace Service::Nvidia

src/core/hle/service/nvnflinger/parcel.h

@@ -6,6 +6,7 @@
 #include <memory>
 #include <span>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 
 #include "common/alignment.h"
 #include "common/assert.h"
@@ -167,7 +168,7 @@ public:
 private:
     template <typename T>
         requires(std::is_trivially_copyable_v<T>)
-    void WriteImpl(const T& val, std::vector<u8>& buffer) {
+    void WriteImpl(const T& val, boost::container::small_vector<u8, 0x200>& buffer) {
         const size_t aligned_size = Common::AlignUp(sizeof(T), 4);
         const size_t old_size = buffer.size();
         buffer.resize(old_size + aligned_size);
@@ -176,8 +177,8 @@ private:
     }
 
 private:
-    std::vector<u8> m_data_buffer;
-    std::vector<u8> m_object_buffer;
+    boost::container::small_vector<u8, 0x200> m_data_buffer;
+    boost::container::small_vector<u8, 0x200> m_object_buffer;
 };
 
 } // namespace Service::android

src/shader_recompiler/backend/glsl/glsl_emit_context.cpp

@@ -479,7 +479,7 @@ void EmitContext::DefineGenericOutput(size_t index, u32 invocations) {
         const u32 remainder{4 - element};
         const TransformFeedbackVarying* xfb_varying{};
         const size_t xfb_varying_index{base_index + element};
-        if (xfb_varying_index < runtime_info.xfb_varyings.size()) {
+        if (xfb_varying_index < runtime_info.xfb_count) {
             xfb_varying = &runtime_info.xfb_varyings[xfb_varying_index];
             xfb_varying = xfb_varying->components > 0 ? xfb_varying : nullptr;
         }

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -387,7 +387,7 @@ void SetupSignedNanCapabilities(const Profile& profile, const IR::Program& progr
 }
 
 void SetupTransformFeedbackCapabilities(EmitContext& ctx, Id main_func) {
-    if (ctx.runtime_info.xfb_varyings.empty()) {
+    if (ctx.runtime_info.xfb_count == 0) {
         return;
     }
     ctx.AddCapability(spv::Capability::TransformFeedback);

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -160,7 +160,7 @@ void DefineGenericOutput(EmitContext& ctx, size_t index, std::optional<u32> invo
         const u32 remainder{4 - element};
         const TransformFeedbackVarying* xfb_varying{};
         const size_t xfb_varying_index{base_attr_index + element};
-        if (xfb_varying_index < ctx.runtime_info.xfb_varyings.size()) {
+        if (xfb_varying_index < ctx.runtime_info.xfb_count) {
             xfb_varying = &ctx.runtime_info.xfb_varyings[xfb_varying_index];
             xfb_varying = xfb_varying->components > 0 ? xfb_varying : nullptr;
         }

src/shader_recompiler/runtime_info.h

@@ -84,7 +84,8 @@ struct RuntimeInfo {
     bool glasm_use_storage_buffers{};
 
     /// Transform feedback state for each varying
-    std::vector<TransformFeedbackVarying> xfb_varyings;
+    std::array<TransformFeedbackVarying, 256> xfb_varyings{};
+    u32 xfb_count{0};
 };
 
 } // namespace Shader

src/video_core/buffer_cache/buffer_cache.h

@@ -207,7 +207,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
-    tmp_buffer.resize(amount);
+    tmp_buffer.resize_destructive(amount);
     cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
     cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
     return true;
@@ -1279,7 +1279,7 @@ template <class P>
 typename BufferCache<P>::OverlapResult BufferCache<P>::ResolveOverlaps(VAddr cpu_addr,
                                                                        u32 wanted_size) {
     static constexpr int STREAM_LEAP_THRESHOLD = 16;
-    std::vector<BufferId> overlap_ids;
+    boost::container::small_vector<BufferId, 16> overlap_ids;
     VAddr begin = cpu_addr;
     VAddr end = cpu_addr + wanted_size;
     int stream_score = 0;

src/video_core/buffer_cache/buffer_cache_base.h

@@ -229,7 +229,7 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<BufferCacheChannelInf
     using OverlapCounter = boost::icl::split_interval_map<VAddr, int>;
 
     struct OverlapResult {
-        std::vector<BufferId> ids;
+        boost::container::small_vector<BufferId, 16> ids;
         VAddr begin;
         VAddr end;
         bool has_stream_leap = false;
@@ -582,7 +582,7 @@ private:
     BufferId inline_buffer_id;
 
     std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
-    std::vector<u8> tmp_buffer;
+    Common::ScratchBuffer<u8> tmp_buffer;
 };
 
 } // namespace VideoCommon

src/video_core/cdma_pusher.h

@@ -63,7 +63,6 @@ struct ChCommand {
 };
 
 using ChCommandHeaderList = std::vector<ChCommandHeader>;
-using ChCommandList = std::vector<ChCommand>;
 
 struct ThiRegisters {
     u32_le increment_syncpt{};

src/video_core/dma_pusher.h

@@ -6,6 +6,7 @@
 #include <array>
 #include <span>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 #include <queue>
 
 #include "common/bit_field.h"
@@ -102,11 +103,12 @@ inline CommandHeader BuildCommandHeader(BufferMethods method, u32 arg_count, Sub
 struct CommandList final {
     CommandList() = default;
     explicit CommandList(std::size_t size) : command_lists(size) {}
-    explicit CommandList(std::vector<CommandHeader>&& prefetch_command_list_)
+    explicit CommandList(
+        boost::container::small_vector<CommandHeader, 512>&& prefetch_command_list_)
         : prefetch_command_list{std::move(prefetch_command_list_)} {}
 
-    std::vector<CommandListHeader> command_lists;
-    std::vector<CommandHeader> prefetch_command_list;
+    boost::container::small_vector<CommandListHeader, 512> command_lists;
+    boost::container::small_vector<CommandHeader, 512> prefetch_command_list;
 };
 
 /**

src/video_core/engines/maxwell_dma.cpp

@@ -108,9 +108,11 @@ void MaxwellDMA::Launch() {
         if (regs.launch_dma.remap_enable != 0 && is_const_a_dst) {
             ASSERT(regs.remap_const.component_size_minus_one == 3);
             accelerate.BufferClear(regs.offset_out, regs.line_length_in, regs.remap_consta_value);
-            std::vector<u32> tmp_buffer(regs.line_length_in, regs.remap_consta_value);
+            read_buffer.resize_destructive(regs.line_length_in * sizeof(u32));
+            std::span<u32> span(reinterpret_cast<u32*>(read_buffer.data()), regs.line_length_in);
+            std::ranges::fill(span, regs.remap_consta_value);
             memory_manager.WriteBlockUnsafe(regs.offset_out,
-                                            reinterpret_cast<u8*>(tmp_buffer.data()),
+                                            reinterpret_cast<u8*>(read_buffer.data()),
                                             regs.line_length_in * sizeof(u32));
         } else {
             memory_manager.FlushCaching();
@@ -126,32 +128,32 @@ void MaxwellDMA::Launch() {
                 UNIMPLEMENTED_IF(regs.line_length_in % 16 != 0);
                 UNIMPLEMENTED_IF(regs.offset_in % 16 != 0);
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
-                std::vector<u8> tmp_buffer(16);
+                read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
                     memory_manager.ReadBlockUnsafe(
                         convert_linear_2_blocklinear_addr(regs.offset_in + offset),
-                        tmp_buffer.data(), tmp_buffer.size());
-                    memory_manager.WriteBlockCached(regs.offset_out + offset, tmp_buffer.data(),
-                                                    tmp_buffer.size());
+                        read_buffer.data(), read_buffer.size());
+                    memory_manager.WriteBlockCached(regs.offset_out + offset, read_buffer.data(),
+                                                    read_buffer.size());
                 }
             } else if (is_src_pitch && !is_dst_pitch) {
                 UNIMPLEMENTED_IF(regs.line_length_in % 16 != 0);
                 UNIMPLEMENTED_IF(regs.offset_in % 16 != 0);
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
-                std::vector<u8> tmp_buffer(16);
+                read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
-                    memory_manager.ReadBlockUnsafe(regs.offset_in + offset, tmp_buffer.data(),
-                                                   tmp_buffer.size());
+                    memory_manager.ReadBlockUnsafe(regs.offset_in + offset, read_buffer.data(),
+                                                   read_buffer.size());
                     memory_manager.WriteBlockCached(
                         convert_linear_2_blocklinear_addr(regs.offset_out + offset),
-                        tmp_buffer.data(), tmp_buffer.size());
+                        read_buffer.data(), read_buffer.size());
                 }
             } else {
                 if (!accelerate.BufferCopy(regs.offset_in, regs.offset_out, regs.line_length_in)) {
-                    std::vector<u8> tmp_buffer(regs.line_length_in);
-                    memory_manager.ReadBlockUnsafe(regs.offset_in, tmp_buffer.data(),
+                    read_buffer.resize_destructive(regs.line_length_in);
+                    memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(),
                                                    regs.line_length_in);
-                    memory_manager.WriteBlockCached(regs.offset_out, tmp_buffer.data(),
+                    memory_manager.WriteBlockCached(regs.offset_out, read_buffer.data(),
                                                     regs.line_length_in);
                 }
             }
@@ -171,7 +173,8 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
     src_operand.address = regs.offset_in;
 
     DMA::BufferOperand dst_operand;
-    dst_operand.pitch = regs.pitch_out;
+    u32 abs_pitch_out = std::abs(static_cast<s32>(regs.pitch_out));
+    dst_operand.pitch = abs_pitch_out;
     dst_operand.width = regs.line_length_in;
     dst_operand.height = regs.line_count;
     dst_operand.address = regs.offset_out;
@@ -218,7 +221,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
     const size_t src_size =
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
 
-    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
+    const size_t dst_size = static_cast<size_t>(abs_pitch_out) * regs.line_count;
     read_buffer.resize_destructive(src_size);
     write_buffer.resize_destructive(dst_size);
 
@@ -227,7 +230,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
 
     UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
                      src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
-                     regs.pitch_out);
+                     abs_pitch_out);
 
     memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }

src/video_core/host1x/codecs/h264.cpp

@@ -4,6 +4,7 @@
 #include <array>
 #include <bit>
 
+#include "common/scratch_buffer.h"
 #include "common/settings.h"
 #include "video_core/host1x/codecs/h264.h"
 #include "video_core/host1x/host1x.h"
@@ -188,7 +189,8 @@ void H264BitWriter::WriteBit(bool state) {
 }
 
 void H264BitWriter::WriteScalingList(std::span<const u8> list, s32 start, s32 count) {
-    std::vector<u8> scan(count);
+    static Common::ScratchBuffer<u8> scan{};
+    scan.resize_destructive(count);
     if (count == 16) {
         std::memcpy(scan.data(), zig_zag_scan.data(), scan.size());
     } else {

src/video_core/memory_manager.cpp

@@ -587,7 +587,7 @@ void MemoryManager::InvalidateRegion(GPUVAddr gpu_addr, size_t size,
 
 void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size,
                               VideoCommon::CacheType which) {
-    std::vector<u8> tmp_buffer(size);
+    tmp_buffer.resize_destructive(size);
     ReadBlock(gpu_src_addr, tmp_buffer.data(), size, which);
 
     // The output block must be flushed in case it has data modified from the GPU.
@@ -670,9 +670,9 @@ bool MemoryManager::IsFullyMappedRange(GPUVAddr gpu_addr, std::size_t size) cons
     return result;
 }
 
-std::vector<std::pair<GPUVAddr, std::size_t>> MemoryManager::GetSubmappedRange(
-    GPUVAddr gpu_addr, std::size_t size) const {
-    std::vector<std::pair<GPUVAddr, std::size_t>> result{};
+boost::container::small_vector<std::pair<GPUVAddr, std::size_t>, 32>
+MemoryManager::GetSubmappedRange(GPUVAddr gpu_addr, std::size_t size) const {
+    boost::container::small_vector<std::pair<GPUVAddr, std::size_t>, 32> result{};
     GetSubmappedRangeImpl<true>(gpu_addr, size, result);
     return result;
 }
@@ -680,8 +680,9 @@ std::vector<std::pair<GPUVAddr, std::size_t>> MemoryManager::GetSubmappedRange(
 template <bool is_gpu_address>
 void MemoryManager::GetSubmappedRangeImpl(
     GPUVAddr gpu_addr, std::size_t size,
-    std::vector<std::pair<std::conditional_t<is_gpu_address, GPUVAddr, VAddr>, std::size_t>>&
-        result) const {
+    boost::container::small_vector<
+        std::pair<std::conditional_t<is_gpu_address, GPUVAddr, VAddr>, std::size_t>, 32>& result)
+    const {
     std::optional<std::pair<std::conditional_t<is_gpu_address, GPUVAddr, VAddr>, std::size_t>>
         last_segment{};
     std::optional<VAddr> old_page_addr{};

src/video_core/memory_manager.h

@@ -8,10 +8,12 @@
 #include <mutex>
 #include <optional>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 
 #include "common/common_types.h"
 #include "common/multi_level_page_table.h"
 #include "common/range_map.h"
+#include "common/scratch_buffer.h"
 #include "common/virtual_buffer.h"
 #include "video_core/cache_types.h"
 #include "video_core/pte_kind.h"
@@ -107,8 +109,8 @@ public:
      * if the region is continuous, a single pair will be returned. If it's unmapped, an empty
      * vector will be returned;
      */
-    std::vector<std::pair<GPUVAddr, std::size_t>> GetSubmappedRange(GPUVAddr gpu_addr,
-                                                                    std::size_t size) const;
+    boost::container::small_vector<std::pair<GPUVAddr, std::size_t>, 32> GetSubmappedRange(
+        GPUVAddr gpu_addr, std::size_t size) const;
 
     GPUVAddr Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size,
                  PTEKind kind = PTEKind::INVALID, bool is_big_pages = true);
@@ -165,7 +167,8 @@ private:
     template <bool is_gpu_address>
     void GetSubmappedRangeImpl(
         GPUVAddr gpu_addr, std::size_t size,
-        std::vector<std::pair<std::conditional_t<is_gpu_address, GPUVAddr, VAddr>, std::size_t>>&
+        boost::container::small_vector<
+            std::pair<std::conditional_t<is_gpu_address, GPUVAddr, VAddr>, std::size_t>, 32>&
             result) const;
 
     Core::System& system;
@@ -215,8 +218,8 @@ private:
     Common::VirtualBuffer<u32> big_page_table_cpu;
 
     std::vector<u64> big_page_continuous;
-    std::vector<std::pair<VAddr, std::size_t>> page_stash{};
-    std::vector<std::pair<VAddr, std::size_t>> page_stash2{};
+    boost::container::small_vector<std::pair<VAddr, std::size_t>, 32> page_stash{};
+    boost::container::small_vector<std::pair<VAddr, std::size_t>, 32> page_stash2{};
 
     mutable std::mutex guard;
 
@@ -226,6 +229,8 @@ private:
     std::unique_ptr<VideoCommon::InvalidationAccumulator> accumulator;
 
     static std::atomic<size_t> unique_identifier_generator;
+
+    Common::ScratchBuffer<u8> tmp_buffer;
 };
 
 } // namespace Tegra

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -85,7 +85,9 @@ Shader::RuntimeInfo MakeRuntimeInfo(const GraphicsPipelineKey& key,
     case Shader::Stage::VertexB:
     case Shader::Stage::Geometry:
         if (!use_assembly_shaders && key.xfb_enabled != 0) {
-            info.xfb_varyings = VideoCommon::MakeTransformFeedbackVaryings(key.xfb_state);
+            auto [varyings, count] = VideoCommon::MakeTransformFeedbackVaryings(key.xfb_state);
+            info.xfb_varyings = varyings;
+            info.xfb_count = count;
         }
         break;
     case Shader::Stage::TessellationEval:

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -361,7 +361,7 @@ void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
         .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
     };
     // Measuring a popular game, this number never exceeds the specified size once data is warmed up
-    boost::container::small_vector<VkBufferCopy, 3> vk_copies(copies.size());
+    boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
     std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([src_buffer, dst_buffer, vk_copies, barrier](vk::CommandBuffer cmdbuf) {

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -167,7 +167,10 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
                 info.fixed_state_point_size = point_size;
             }
             if (key.state.xfb_enabled) {
-                info.xfb_varyings = VideoCommon::MakeTransformFeedbackVaryings(key.state.xfb_state);
+                auto [varyings, count] =
+                    VideoCommon::MakeTransformFeedbackVaryings(key.state.xfb_state);
+                info.xfb_varyings = varyings;
+                info.xfb_count = count;
             }
             info.convert_depth_mode = gl_ndc;
         }
@@ -214,7 +217,10 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
             info.fixed_state_point_size = point_size;
         }
         if (key.state.xfb_enabled != 0) {
-            info.xfb_varyings = VideoCommon::MakeTransformFeedbackVaryings(key.state.xfb_state);
+            auto [varyings, count] =
+                VideoCommon::MakeTransformFeedbackVaryings(key.state.xfb_state);
+            info.xfb_varyings = varyings;
+            info.xfb_count = count;
         }
         info.convert_depth_mode = gl_ndc;
         break;

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -330,9 +330,9 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     };
 }
 
-[[maybe_unused]] [[nodiscard]] std::vector<VkBufferCopy> TransformBufferCopies(
-    std::span<const VideoCommon::BufferCopy> copies, size_t buffer_offset) {
-    std::vector<VkBufferCopy> result(copies.size());
+[[maybe_unused]] [[nodiscard]] boost::container::small_vector<VkBufferCopy, 16>
+TransformBufferCopies(std::span<const VideoCommon::BufferCopy> copies, size_t buffer_offset) {
+    boost::container::small_vector<VkBufferCopy, 16> result(copies.size());
     std::ranges::transform(
         copies, result.begin(), [buffer_offset](const VideoCommon::BufferCopy& copy) {
             return VkBufferCopy{
@@ -344,7 +344,7 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return result;
 }
 
-[[nodiscard]] std::vector<VkBufferImageCopy> TransformBufferImageCopies(
+[[nodiscard]] boost::container::small_vector<VkBufferImageCopy, 16> TransformBufferImageCopies(
     std::span<const BufferImageCopy> copies, size_t buffer_offset, VkImageAspectFlags aspect_mask) {
     struct Maker {
         VkBufferImageCopy operator()(const BufferImageCopy& copy) const {
@@ -377,14 +377,14 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
         VkImageAspectFlags aspect_mask;
     };
     if (aspect_mask == (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
-        std::vector<VkBufferImageCopy> result(copies.size() * 2);
+        boost::container::small_vector<VkBufferImageCopy, 16> result(copies.size() * 2);
         std::ranges::transform(copies, result.begin(),
                                Maker{buffer_offset, VK_IMAGE_ASPECT_DEPTH_BIT});
         std::ranges::transform(copies, result.begin() + copies.size(),
                                Maker{buffer_offset, VK_IMAGE_ASPECT_STENCIL_BIT});
         return result;
     } else {
-        std::vector<VkBufferImageCopy> result(copies.size());
+        boost::container::small_vector<VkBufferImageCopy, 16> result(copies.size());
         std::ranges::transform(copies, result.begin(), Maker{buffer_offset, aspect_mask});
         return result;
     }
@@ -867,8 +867,8 @@ void TextureCacheRuntime::BarrierFeedbackLoop() {
 
 void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,
                                            std::span<const VideoCommon::ImageCopy> copies) {
-    std::vector<VkBufferImageCopy> vk_in_copies(copies.size());
-    std::vector<VkBufferImageCopy> vk_out_copies(copies.size());
+    boost::container::small_vector<VkBufferImageCopy, 16> vk_in_copies(copies.size());
+    boost::container::small_vector<VkBufferImageCopy, 16> vk_out_copies(copies.size());
     const VkImageAspectFlags src_aspect_mask = src.AspectMask();
     const VkImageAspectFlags dst_aspect_mask = dst.AspectMask();
 
@@ -1157,7 +1157,7 @@ void TextureCacheRuntime::ConvertImage(Framebuffer* dst, ImageView& dst_view, Im
 
 void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
                                     std::span<const VideoCommon::ImageCopy> copies) {
-    std::vector<VkImageCopy> vk_copies(copies.size());
+    boost::container::small_vector<VkImageCopy, 16> vk_copies(copies.size());
     const VkImageAspectFlags aspect_mask = dst.AspectMask();
     ASSERT(aspect_mask == src.AspectMask());
 
@@ -1332,7 +1332,7 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
         ScaleDown(true);
     }
     scheduler->RequestOutsideRenderPassOperationContext();
-    std::vector vk_copies = TransformBufferImageCopies(copies, offset, aspect_mask);
+    auto vk_copies = TransformBufferImageCopies(copies, offset, aspect_mask);
     const VkBuffer src_buffer = buffer;
     const VkImage vk_image = *original_image;
     const VkImageAspectFlags vk_aspect_mask = aspect_mask;
@@ -1367,8 +1367,9 @@ void Image::DownloadMemory(std::span<VkBuffer> buffers_span, std::span<VkDeviceS
     if (is_rescaled) {
         ScaleDown();
     }
-    boost::container::small_vector<VkBuffer, 1> buffers_vector{};
-    boost::container::small_vector<std::vector<VkBufferImageCopy>, 1> vk_copies;
+    boost::container::small_vector<VkBuffer, 8> buffers_vector{};
+    boost::container::small_vector<boost::container::small_vector<VkBufferImageCopy, 16>, 8>
+        vk_copies;
     for (size_t index = 0; index < buffers_span.size(); index++) {
         buffers_vector.emplace_back(buffers_span[index]);
         vk_copies.emplace_back(
@@ -1858,7 +1859,7 @@ Framebuffer::~Framebuffer() = default;
 void Framebuffer::CreateFramebuffer(TextureCacheRuntime& runtime,
                                     std::span<ImageView*, NUM_RT> color_buffers,
                                     ImageView* depth_buffer, bool is_rescaled) {
-    std::vector<VkImageView> attachments;
+    boost::container::small_vector<VkImageView, NUM_RT + 1> attachments;
     RenderPassKey renderpass_key{};
     s32 num_layers = 1;
 

src/video_core/shader_cache.cpp

@@ -151,11 +151,9 @@ void ShaderCache::RemovePendingShaders() {
     marked_for_removal.erase(std::unique(marked_for_removal.begin(), marked_for_removal.end()),
                              marked_for_removal.end());
 
-    std::vector<ShaderInfo*> removed_shaders;
-    removed_shaders.reserve(marked_for_removal.size());
+    boost::container::small_vector<ShaderInfo*, 16> removed_shaders;
 
     std::scoped_lock lock{lookup_mutex};
-
     for (Entry* const entry : marked_for_removal) {
         removed_shaders.push_back(entry->data);
 

src/video_core/texture_cache/image_base.h

@@ -6,6 +6,7 @@
 #include <array>
 #include <optional>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 
 #include "common/common_funcs.h"
 #include "common/common_types.h"
@@ -108,8 +109,8 @@ struct ImageBase {
     std::vector<ImageViewInfo> image_view_infos;
     std::vector<ImageViewId> image_view_ids;
 
-    std::vector<u32> slice_offsets;
-    std::vector<SubresourceBase> slice_subresources;
+    boost::container::small_vector<u32, 16> slice_offsets;
+    boost::container::small_vector<SubresourceBase, 16> slice_subresources;
 
     std::vector<AliasedImage> aliased_images;
     std::vector<ImageId> overlapping_images;

src/video_core/texture_cache/texture_cache.h

@@ -526,7 +526,7 @@ void TextureCache<P>::WriteMemory(VAddr cpu_addr, size_t size) {
 
 template <class P>
 void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
-    std::vector<ImageId> images;
+    boost::container::small_vector<ImageId, 16> images;
     ForEachImageInRegion(cpu_addr, size, [&images](ImageId image_id, ImageBase& image) {
         if (!image.IsSafeDownload()) {
             return;
@@ -579,7 +579,7 @@ std::optional<VideoCore::RasterizerDownloadArea> TextureCache<P>::GetFlushArea(V
 
 template <class P>
 void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
-    std::vector<ImageId> deleted_images;
+    boost::container::small_vector<ImageId, 16> deleted_images;
     ForEachImageInRegion(cpu_addr, size, [&](ImageId id, Image&) { deleted_images.push_back(id); });
     for (const ImageId id : deleted_images) {
         Image& image = slot_images[id];
@@ -593,7 +593,7 @@ void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
 
 template <class P>
 void TextureCache<P>::UnmapGPUMemory(size_t as_id, GPUVAddr gpu_addr, size_t size) {
-    std::vector<ImageId> deleted_images;
+    boost::container::small_vector<ImageId, 16> deleted_images;
     ForEachImageInRegionGPU(as_id, gpu_addr, size,
                             [&](ImageId id, Image&) { deleted_images.push_back(id); });
     for (const ImageId id : deleted_images) {
@@ -1101,7 +1101,7 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
     const bool native_bgr = runtime.HasNativeBgr();
     const bool flexible_formats = True(options & RelaxedOptions::Format);
     ImageId image_id{};
-    boost::container::small_vector<ImageId, 1> image_ids;
+    boost::container::small_vector<ImageId, 8> image_ids;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
         if (True(existing_image.flags & ImageFlagBits::Remapped)) {
             return false;
@@ -1622,7 +1622,7 @@ ImageId TextureCache<P>::FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr)
         }
     }
     ImageId image_id{};
-    boost::container::small_vector<ImageId, 1> image_ids;
+    boost::container::small_vector<ImageId, 8> image_ids;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
         if (True(existing_image.flags & ImageFlagBits::Remapped)) {
             return false;
@@ -1942,7 +1942,7 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
         image.map_view_id = map_id;
         return;
     }
-    std::vector<ImageViewId> sparse_maps{};
+    boost::container::small_vector<ImageViewId, 16> sparse_maps;
     ForEachSparseSegment(
         image, [this, image_id, &sparse_maps](GPUVAddr gpu_addr, VAddr cpu_addr, size_t size) {
             auto map_id = slot_map_views.insert(gpu_addr, cpu_addr, size, image_id);
@@ -2217,7 +2217,7 @@ void TextureCache<P>::MarkModification(ImageBase& image) noexcept {
 
 template <class P>
 void TextureCache<P>::SynchronizeAliases(ImageId image_id) {
-    boost::container::small_vector<const AliasedImage*, 1> aliased_images;
+    boost::container::small_vector<const AliasedImage*, 8> aliased_images;
     Image& image = slot_images[image_id];
     bool any_rescaled = True(image.flags & ImageFlagBits::Rescaled);
     bool any_modified = True(image.flags & ImageFlagBits::GpuModified);

src/video_core/texture_cache/texture_cache_base.h

@@ -56,7 +56,7 @@ struct ImageViewInOut {
 struct AsyncDecodeContext {
     ImageId image_id;
     Common::ScratchBuffer<u8> decoded_data;
-    std::vector<BufferImageCopy> copies;
+    boost::container::small_vector<BufferImageCopy, 16> copies;
     std::mutex mutex;
     std::atomic_bool complete;
 };
@@ -429,7 +429,7 @@ private:
 
     std::unordered_map<u64, std::vector<ImageMapId>, Common::IdentityHash<u64>> page_table;
     std::unordered_map<u64, std::vector<ImageId>, Common::IdentityHash<u64>> sparse_page_table;
-    std::unordered_map<ImageId, std::vector<ImageViewId>> sparse_views;
+    std::unordered_map<ImageId, boost::container::small_vector<ImageViewId, 16>> sparse_views;
 
     VAddr virtual_invalid_space{};
 

src/video_core/texture_cache/util.cpp

@@ -329,13 +329,13 @@ template <u32 GOB_EXTENT>
 
 [[nodiscard]] std::optional<SubresourceExtent> ResolveOverlapRightAddress3D(
     const ImageInfo& new_info, GPUVAddr gpu_addr, const ImageBase& overlap, bool strict_size) {
-    const std::vector<u32> slice_offsets = CalculateSliceOffsets(new_info);
+    const auto slice_offsets = CalculateSliceOffsets(new_info);
     const u32 diff = static_cast<u32>(overlap.gpu_addr - gpu_addr);
     const auto it = std::ranges::find(slice_offsets, diff);
     if (it == slice_offsets.end()) {
         return std::nullopt;
     }
-    const std::vector subresources = CalculateSliceSubresources(new_info);
+    const auto subresources = CalculateSliceSubresources(new_info);
     const SubresourceBase base = subresources[std::distance(slice_offsets.begin(), it)];
     const ImageInfo& info = overlap.info;
     if (!IsBlockLinearSizeCompatible(new_info, info, base.level, 0, strict_size)) {
@@ -655,9 +655,9 @@ LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept {
     return sizes;
 }
 
-std::vector<u32> CalculateSliceOffsets(const ImageInfo& info) {
+boost::container::small_vector<u32, 16> CalculateSliceOffsets(const ImageInfo& info) {
     ASSERT(info.type == ImageType::e3D);
-    std::vector<u32> offsets;
+    boost::container::small_vector<u32, 16> offsets;
     offsets.reserve(NumSlices(info));
 
     const LevelInfo level_info = MakeLevelInfo(info);
@@ -679,9 +679,10 @@ std::vector<u32> CalculateSliceOffsets(const ImageInfo& info) {
     return offsets;
 }
 
-std::vector<SubresourceBase> CalculateSliceSubresources(const ImageInfo& info) {
+boost::container::small_vector<SubresourceBase, 16> CalculateSliceSubresources(
+    const ImageInfo& info) {
     ASSERT(info.type == ImageType::e3D);
-    std::vector<SubresourceBase> subresources;
+    boost::container::small_vector<SubresourceBase, 16> subresources;
     subresources.reserve(NumSlices(info));
     for (s32 level = 0; level < info.resources.levels; ++level) {
         const s32 depth = AdjustMipSize(info.size.depth, level);
@@ -723,8 +724,10 @@ ImageViewType RenderTargetImageViewType(const ImageInfo& info) noexcept {
     }
 }
 
-std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageInfo& src,
-                                             SubresourceBase base, u32 up_scale, u32 down_shift) {
+boost::container::small_vector<ImageCopy, 16> MakeShrinkImageCopies(const ImageInfo& dst,
+                                                                    const ImageInfo& src,
+                                                                    SubresourceBase base,
+                                                                    u32 up_scale, u32 down_shift) {
     ASSERT(dst.resources.levels >= src.resources.levels);
 
     const bool is_dst_3d = dst.type == ImageType::e3D;
@@ -733,7 +736,7 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
         ASSERT(src.resources.levels == 1);
     }
     const bool both_2d{src.type == ImageType::e2D && dst.type == ImageType::e2D};
-    std::vector<ImageCopy> copies;
+    boost::container::small_vector<ImageCopy, 16> copies;
     copies.reserve(src.resources.levels);
     for (s32 level = 0; level < src.resources.levels; ++level) {
         ImageCopy& copy = copies.emplace_back();
@@ -770,9 +773,10 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
     return copies;
 }
 
-std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src, u32 up_scale,
-                                                  u32 down_shift) {
-    std::vector<ImageCopy> copies;
+boost::container::small_vector<ImageCopy, 16> MakeReinterpretImageCopies(const ImageInfo& src,
+                                                                         u32 up_scale,
+                                                                         u32 down_shift) {
+    boost::container::small_vector<ImageCopy, 16> copies;
     copies.reserve(src.resources.levels);
     const bool is_3d = src.type == ImageType::e3D;
     for (s32 level = 0; level < src.resources.levels; ++level) {
@@ -824,9 +828,11 @@ bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config
     return gpu_memory.GpuToCpuAddress(address, guest_size_bytes).has_value();
 }
 
-std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                                            const ImageInfo& info, std::span<const u8> input,
-                                            std::span<u8> output) {
+boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
+                                                                   GPUVAddr gpu_addr,
+                                                                   const ImageInfo& info,
+                                                                   std::span<const u8> input,
+                                                                   std::span<u8> output) {
     const size_t guest_size_bytes = input.size_bytes();
     const u32 bpp_log2 = BytesPerBlockLog2(info.format);
     const Extent3D size = info.size;
@@ -861,7 +867,7 @@ std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GP
                                             info.tile_width_spacing);
     size_t guest_offset = 0;
     u32 host_offset = 0;
-    std::vector<BufferImageCopy> copies(num_levels);
+    boost::container::small_vector<BufferImageCopy, 16> copies(num_levels);
 
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
@@ -978,7 +984,7 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
     }
 }
 
-std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info) {
+boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(const ImageInfo& info) {
     const Extent3D size = info.size;
     const u32 bytes_per_block = BytesPerBlock(info.format);
     if (info.type == ImageType::Linear) {
@@ -1006,7 +1012,7 @@ std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info) {
 
     u32 host_offset = 0;
 
-    std::vector<BufferImageCopy> copies(num_levels);
+    boost::container::small_vector<BufferImageCopy, 16> copies(num_levels);
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
         const u32 num_blocks_per_layer = NumBlocks(level_size, tile_size);
@@ -1042,10 +1048,10 @@ Extent3D MipBlockSize(const ImageInfo& info, u32 level) {
     return AdjustMipBlockSize(num_tiles, level_info.block, level);
 }
 
-std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info) {
+boost::container::small_vector<SwizzleParameters, 16> FullUploadSwizzles(const ImageInfo& info) {
     const Extent2D tile_size = DefaultBlockSize(info.format);
     if (info.type == ImageType::Linear) {
-        return std::vector{SwizzleParameters{
+        return {SwizzleParameters{
             .num_tiles = AdjustTileSize(info.size, tile_size),
             .block = {},
             .buffer_offset = 0,
@@ -1057,7 +1063,7 @@ std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info) {
     const s32 num_levels = info.resources.levels;
 
     u32 guest_offset = 0;
-    std::vector<SwizzleParameters> params(num_levels);
+    boost::container::small_vector<SwizzleParameters, 16> params(num_levels);
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
         const Extent3D num_tiles = AdjustTileSize(level_size, tile_size);

src/video_core/texture_cache/util.h

@@ -5,6 +5,7 @@
 
 #include <optional>
 #include <span>
+#include <boost/container/small_vector.hpp>
 
 #include "common/common_types.h"
 #include "common/scratch_buffer.h"
@@ -40,9 +41,10 @@ struct OverlapResult {
 
 [[nodiscard]] LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept;
 
-[[nodiscard]] std::vector<u32> CalculateSliceOffsets(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<u32, 16> CalculateSliceOffsets(const ImageInfo& info);
 
-[[nodiscard]] std::vector<SubresourceBase> CalculateSliceSubresources(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<SubresourceBase, 16> CalculateSliceSubresources(
+    const ImageInfo& info);
 
 [[nodiscard]] u32 CalculateLevelStrideAlignment(const ImageInfo& info, u32 level);
 
@@ -51,21 +53,18 @@ struct OverlapResult {
 
 [[nodiscard]] ImageViewType RenderTargetImageViewType(const ImageInfo& info) noexcept;
 
-[[nodiscard]] std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst,
-                                                           const ImageInfo& src,
-                                                           SubresourceBase base, u32 up_scale = 1,
-                                                           u32 down_shift = 0);
+[[nodiscard]] boost::container::small_vector<ImageCopy, 16> MakeShrinkImageCopies(
+    const ImageInfo& dst, const ImageInfo& src, SubresourceBase base, u32 up_scale = 1,
+    u32 down_shift = 0);
 
-[[nodiscard]] std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src,
-                                                                u32 up_scale = 1,
-                                                                u32 down_shift = 0);
+[[nodiscard]] boost::container::small_vector<ImageCopy, 16> MakeReinterpretImageCopies(
+    const ImageInfo& src, u32 up_scale = 1, u32 down_shift = 0);
 
 [[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
 
-[[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
-                                                          GPUVAddr gpu_addr, const ImageInfo& info,
-                                                          std::span<const u8> input,
-                                                          std::span<u8> output);
+[[nodiscard]] boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(
+    Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
+    std::span<const u8> input, std::span<u8> output);
 
 [[nodiscard]] BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
                                           const ImageBase& image, std::span<u8> output);
@@ -73,13 +72,15 @@ struct OverlapResult {
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies);
 
-[[nodiscard]] std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(
+    const ImageInfo& info);
 
 [[nodiscard]] Extent3D MipSize(Extent3D size, u32 level);
 
 [[nodiscard]] Extent3D MipBlockSize(const ImageInfo& info, u32 level);
 
-[[nodiscard]] std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<SwizzleParameters, 16> FullUploadSwizzles(
+    const ImageInfo& info);
 
 void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
                   std::span<const BufferImageCopy> copies, std::span<const u8> memory,

src/video_core/transform_feedback.cpp

@@ -13,7 +13,7 @@
 
 namespace VideoCommon {
 
-std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
+std::pair<std::array<Shader::TransformFeedbackVarying, 256>, u32> MakeTransformFeedbackVaryings(
     const TransformFeedbackState& state) {
     static constexpr std::array VECTORS{
         28U,  // gl_Position
@@ -62,7 +62,8 @@ std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
         216U, // gl_TexCoord[6]
         220U, // gl_TexCoord[7]
     };
-    std::vector<Shader::TransformFeedbackVarying> xfb(256);
+    std::array<Shader::TransformFeedbackVarying, 256> xfb{};
+    u32 count{0};
     for (size_t buffer = 0; buffer < state.layouts.size(); ++buffer) {
         const auto& locations = state.varyings[buffer];
         const auto& layout = state.layouts[buffer];
@@ -103,11 +104,12 @@ std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
                 }
             }
             xfb[attribute] = varying;
+            count = std::max(count, attribute);
             highest = std::max(highest, (base_offset + varying.components) * 4);
         }
         UNIMPLEMENTED_IF(highest != layout.stride);
     }
-    return xfb;
+    return {xfb, count + 1};
 }
 
 } // namespace VideoCommon

src/video_core/transform_feedback.h

@@ -24,7 +24,7 @@ struct TransformFeedbackState {
         varyings;
 };
 
-std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
+std::pair<std::array<Shader::TransformFeedbackVarying, 256>, u32> MakeTransformFeedbackVaryings(
     const TransformFeedbackState& state);
 
 } // namespace VideoCommon

src/video_core/vulkan_common/vulkan_device.cpp

@@ -316,6 +316,7 @@ NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
 std::vector<const char*> ExtensionListForVulkan(
     const std::set<std::string, std::less<>>& extensions) {
     std::vector<const char*> output;
+    output.reserve(extensions.size());
     for (const auto& extension : extensions) {
         output.push_back(extension.c_str());
     }