Merge pull request #11225 from FernandoS27/no-laxatives-in-santas-cookies

Y.F.C: Rework the Query Cache.

src/common/settings.cpp

@@ -130,13 +130,17 @@ void LogSettings() {
     log_path("DataStorage_SDMCDir", Common::FS::GetYuzuPath(Common::FS::YuzuPath::SDMCDir));
 }
 
+void UpdateGPUAccuracy() {
+    values.current_gpu_accuracy = values.gpu_accuracy.GetValue();
+}
+
 bool IsGPULevelExtreme() {
-    return values.gpu_accuracy.GetValue() == GpuAccuracy::Extreme;
+    return values.current_gpu_accuracy == GpuAccuracy::Extreme;
 }
 
 bool IsGPULevelHigh() {
-    return values.gpu_accuracy.GetValue() == GpuAccuracy::Extreme ||
+    return values.current_gpu_accuracy == GpuAccuracy::Extreme ||
-           values.gpu_accuracy.GetValue() == GpuAccuracy::High;
+           values.current_gpu_accuracy == GpuAccuracy::High;
 }
 
 bool IsFastmemEnabled() {

src/common/settings.h

@@ -307,6 +307,7 @@ struct Values {
                                                       Specialization::Default,
                                                       true,
                                                       true};
+    GpuAccuracy current_gpu_accuracy{GpuAccuracy::High};
     SwitchableSetting<AnisotropyMode, true> max_anisotropy{
         linkage,          AnisotropyMode::Automatic, AnisotropyMode::Automatic, AnisotropyMode::X16,
         "max_anisotropy", Category::RendererAdvanced};
@@ -522,6 +523,7 @@ struct Values {
 
 extern Values values;
 
+void UpdateGPUAccuracy();
 bool IsGPULevelExtreme();
 bool IsGPULevelHigh();
 

src/video_core/CMakeLists.txt

@@ -95,6 +95,12 @@ add_library(video_core STATIC
     memory_manager.h
     precompiled_headers.h
     pte_kind.h
+    query_cache/bank_base.h
+    query_cache/query_base.h
+    query_cache/query_cache_base.h
+    query_cache/query_cache.h
+    query_cache/query_stream.h
+    query_cache/types.h
     query_cache.h
     rasterizer_accelerated.cpp
     rasterizer_accelerated.h

src/video_core/buffer_cache/buffer_cache.h

@@ -272,13 +272,19 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_ad
     if (!cpu_addr) {
         return {&slot_buffers[NULL_BUFFER_ID], 0};
     }
-    const BufferId buffer_id = FindBuffer(*cpu_addr, size);
+    return ObtainCPUBuffer(*cpu_addr, size, sync_info, post_op);
+}
+
+template <class P>
+std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(
+    VAddr cpu_addr, u32 size, ObtainBufferSynchronize sync_info, ObtainBufferOperation post_op) {
+    const BufferId buffer_id = FindBuffer(cpu_addr, size);
     Buffer& buffer = slot_buffers[buffer_id];
 
     // synchronize op
     switch (sync_info) {
     case ObtainBufferSynchronize::FullSynchronize:
-        SynchronizeBuffer(buffer, *cpu_addr, size);
+        SynchronizeBuffer(buffer, cpu_addr, size);
         break;
     default:
         break;
@@ -286,11 +292,11 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_ad
 
     switch (post_op) {
     case ObtainBufferOperation::MarkAsWritten:
-        MarkWrittenBuffer(buffer_id, *cpu_addr, size);
+        MarkWrittenBuffer(buffer_id, cpu_addr, size);
         break;
     case ObtainBufferOperation::DiscardWrite: {
-        VAddr cpu_addr_start = Common::AlignDown(*cpu_addr, 64);
+        VAddr cpu_addr_start = Common::AlignDown(cpu_addr, 64);
-        VAddr cpu_addr_end = Common::AlignUp(*cpu_addr + size, 64);
+        VAddr cpu_addr_end = Common::AlignUp(cpu_addr + size, 64);
         IntervalType interval{cpu_addr_start, cpu_addr_end};
         ClearDownload(interval);
         common_ranges.subtract(interval);
@@ -300,7 +306,7 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_ad
         break;
     }
 
-    return {&buffer, buffer.Offset(*cpu_addr)};
+    return {&buffer, buffer.Offset(cpu_addr)};
 }
 
 template <class P>

src/video_core/buffer_cache/buffer_cache_base.h

@@ -295,6 +295,10 @@ public:
     [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(GPUVAddr gpu_addr, u32 size,
                                                        ObtainBufferSynchronize sync_info,
                                                        ObtainBufferOperation post_op);
+
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainCPUBuffer(VAddr gpu_addr, u32 size,
+                                                          ObtainBufferSynchronize sync_info,
+                                                          ObtainBufferOperation post_op);
     void FlushCachedWrites();
 
     /// Return true when there are uncommitted buffers to be downloaded
@@ -335,6 +339,14 @@ public:
 
     [[nodiscard]] std::pair<Buffer*, u32> GetDrawIndirectBuffer();
 
+    template <typename Func>
+    void BufferOperations(Func&& func) {
+        do {
+            channel_state->has_deleted_buffers = false;
+            func();
+        } while (channel_state->has_deleted_buffers);
+    }
+
     std::recursive_mutex mutex;
     Runtime& runtime;
 

src/video_core/control/channel_state_cache.h

@@ -51,7 +51,7 @@ public:
     virtual void CreateChannel(Tegra::Control::ChannelState& channel);
 
     /// Bind a channel for execution.
-    void BindToChannel(s32 id);
+    virtual void BindToChannel(s32 id);
 
     /// Erase channel's state.
     void EraseChannel(s32 id);

src/video_core/engines/draw_manager.h

@@ -46,6 +46,7 @@ public:
     };
 
     struct IndirectParams {
+        bool is_byte_count;
         bool is_indexed;
         bool include_count;
         GPUVAddr count_start_address;

src/video_core/engines/maxwell_3d.cpp

@@ -20,8 +20,6 @@
 
 namespace Tegra::Engines {
 
-using VideoCore::QueryType;
-
 /// First register id that is actually a Macro call.
 constexpr u32 MacroRegistersStart = 0xE00;
 
@@ -500,27 +498,21 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
 }
 
 void Maxwell3D::ProcessQueryGet() {
+    VideoCommon::QueryPropertiesFlags flags{};
+    if (regs.report_semaphore.query.short_query == 0) {
+        flags |= VideoCommon::QueryPropertiesFlags::HasTimeout;
+    }
+    const GPUVAddr sequence_address{regs.report_semaphore.Address()};
+    const VideoCommon::QueryType query_type =
+        static_cast<VideoCommon::QueryType>(regs.report_semaphore.query.report.Value());
+    const u32 payload = regs.report_semaphore.payload;
+    const u32 subreport = regs.report_semaphore.query.sub_report;
     switch (regs.report_semaphore.query.operation) {
     case Regs::ReportSemaphore::Operation::Release:
         if (regs.report_semaphore.query.short_query != 0) {
-            const GPUVAddr sequence_address{regs.report_semaphore.Address()};
+            flags |= VideoCommon::QueryPropertiesFlags::IsAFence;
-            const u32 payload = regs.report_semaphore.payload;
-            std::function<void()> operation([this, sequence_address, payload] {
-                memory_manager.Write<u32>(sequence_address, payload);
-            });
-            rasterizer->SignalFence(std::move(operation));
-        } else {
-            struct LongQueryResult {
-                u64_le value;
-                u64_le timestamp;
-            };
-            const GPUVAddr sequence_address{regs.report_semaphore.Address()};
-            const u32 payload = regs.report_semaphore.payload;
-            [this, sequence_address, payload] {
-                memory_manager.Write<u64>(sequence_address + sizeof(u64), system.GPU().GetTicks());
-                memory_manager.Write<u64>(sequence_address, payload);
-            }();
         }
+        rasterizer->Query(sequence_address, query_type, flags, payload, subreport);
         break;
     case Regs::ReportSemaphore::Operation::Acquire:
         // TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
@@ -528,11 +520,7 @@ void Maxwell3D::ProcessQueryGet() {
         UNIMPLEMENTED_MSG("Unimplemented query operation ACQUIRE");
         break;
     case Regs::ReportSemaphore::Operation::ReportOnly:
-        if (const std::optional<u64> result = GetQueryResult()) {
+        rasterizer->Query(sequence_address, query_type, flags, payload, subreport);
-            // If the query returns an empty optional it means it's cached and deferred.
-            // In this case we have a non-empty result, so we stamp it immediately.
-            StampQueryResult(*result, regs.report_semaphore.query.short_query == 0);
-        }
         break;
     case Regs::ReportSemaphore::Operation::Trap:
         UNIMPLEMENTED_MSG("Unimplemented query operation TRAP");
@@ -544,6 +532,10 @@ void Maxwell3D::ProcessQueryGet() {
 }
 
 void Maxwell3D::ProcessQueryCondition() {
+    if (rasterizer->AccelerateConditionalRendering()) {
+        execute_on = true;
+        return;
+    }
     const GPUVAddr condition_address{regs.render_enable.Address()};
     switch (regs.render_enable_override) {
     case Regs::RenderEnable::Override::AlwaysRender:
@@ -553,10 +545,6 @@ void Maxwell3D::ProcessQueryCondition() {
         execute_on = false;
         break;
     case Regs::RenderEnable::Override::UseRenderEnable: {
-        if (rasterizer->AccelerateConditionalRendering()) {
-            execute_on = true;
-            return;
-        }
         switch (regs.render_enable.mode) {
         case Regs::RenderEnable::Mode::True: {
             execute_on = true;
@@ -598,15 +586,9 @@ void Maxwell3D::ProcessQueryCondition() {
 }
 
 void Maxwell3D::ProcessCounterReset() {
-#if ANDROID
-    if (!Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        return;
-    }
-#endif
     switch (regs.clear_report_value) {
     case Regs::ClearReport::ZPassPixelCount:
-        rasterizer->ResetCounter(QueryType::SamplesPassed);
+        rasterizer->ResetCounter(VideoCommon::QueryType::ZPassPixelCount64);
         break;
     default:
         LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}", regs.clear_report_value);
@@ -620,28 +602,6 @@ void Maxwell3D::ProcessSyncPoint() {
     rasterizer->SignalSyncPoint(sync_point);
 }
 
-std::optional<u64> Maxwell3D::GetQueryResult() {
-    switch (regs.report_semaphore.query.report) {
-    case Regs::ReportSemaphore::Report::Payload:
-        return regs.report_semaphore.payload;
-    case Regs::ReportSemaphore::Report::ZPassPixelCount64:
-#if ANDROID
-        if (!Settings::IsGPULevelHigh()) {
-            // This is problematic on Android, disable on GPU Normal.
-            return 120;
-        }
-#endif
-        // Deferred.
-        rasterizer->Query(regs.report_semaphore.Address(), QueryType::SamplesPassed,
-                          system.GPU().GetTicks());
-        return std::nullopt;
-    default:
-        LOG_DEBUG(HW_GPU, "Unimplemented query report type {}",
-                  regs.report_semaphore.query.report.Value());
-        return 1;
-    }
-}
-
 void Maxwell3D::ProcessCBBind(size_t stage_index) {
     // Bind the buffer currently in CB_ADDRESS to the specified index in the desired shader
     // stage.

src/video_core/engines/maxwell_3d.h

@@ -3182,9 +3182,6 @@ private:
     /// Handles writes to syncing register.
     void ProcessSyncPoint();
 
-    /// Returns a query's value or an empty object if the value will be deferred through a cache.
-    std::optional<u64> GetQueryResult();
-
     void RefreshParametersImpl();
 
     bool IsMethodExecutable(u32 method);

src/video_core/engines/maxwell_dma.cpp

@@ -362,21 +362,17 @@ void MaxwellDMA::ReleaseSemaphore() {
     const auto type = regs.launch_dma.semaphore_type;
     const GPUVAddr address = regs.semaphore.address;
     const u32 payload = regs.semaphore.payload;
+    VideoCommon::QueryPropertiesFlags flags{VideoCommon::QueryPropertiesFlags::IsAFence};
     switch (type) {
     case LaunchDMA::SemaphoreType::NONE:
         break;
     case LaunchDMA::SemaphoreType::RELEASE_ONE_WORD_SEMAPHORE: {
-        std::function<void()> operation(
+        rasterizer->Query(address, VideoCommon::QueryType::Payload, flags, payload, 0);
-            [this, address, payload] { memory_manager.Write<u32>(address, payload); });
-        rasterizer->SignalFence(std::move(operation));
         break;
     }
     case LaunchDMA::SemaphoreType::RELEASE_FOUR_WORD_SEMAPHORE: {
-        std::function<void()> operation([this, address, payload] {
+        rasterizer->Query(address, VideoCommon::QueryType::Payload,
-            memory_manager.Write<u64>(address + sizeof(u64), system.GPU().GetTicks());
+                          flags | VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
-            memory_manager.Write<u64>(address, payload);
-        });
-        rasterizer->SignalFence(std::move(operation));
         break;
     }
     default:

src/video_core/engines/puller.cpp

@@ -82,10 +82,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
     if (op == GpuSemaphoreOperation::WriteLong) {
         const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
         const u32 payload = regs.semaphore_sequence;
-        [this, sequence_address, payload] {
+        rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
-            memory_manager.Write<u64>(sequence_address + sizeof(u64), gpu.GetTicks());
+                          VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
-            memory_manager.Write<u64>(sequence_address, payload);
-        }();
     } else {
         do {
             const u32 word{memory_manager.Read<u32>(regs.semaphore_address.SemaphoreAddress())};
@@ -120,10 +118,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
 void Puller::ProcessSemaphoreRelease() {
     const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
     const u32 payload = regs.semaphore_release;
-    std::function<void()> operation([this, sequence_address, payload] {
+    rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
-        memory_manager.Write<u32>(sequence_address, payload);
+                      VideoCommon::QueryPropertiesFlags::IsAFence, payload, 0);
-    });
-    rasterizer->SignalFence(std::move(operation));
 }
 
 void Puller::ProcessSemaphoreAcquire() {
@@ -132,7 +128,6 @@ void Puller::ProcessSemaphoreAcquire() {
     while (word != value) {
         regs.acquire_active = true;
         regs.acquire_value = value;
-        std::this_thread::sleep_for(std::chrono::milliseconds(1));
         rasterizer->ReleaseFences();
         word = memory_manager.Read<u32>(regs.semaphore_address.SemaphoreAddress());
         // TODO(kemathe73) figure out how to do the acquire_timeout

src/video_core/fence_manager.h

@@ -55,6 +55,9 @@ public:
 
     // Unlike other fences, this one doesn't
     void SignalOrdering() {
+        if constexpr (!can_async_check) {
+            TryReleasePendingFences<false>();
+        }
         std::scoped_lock lock{buffer_cache.mutex};
         buffer_cache.AccumulateFlushes();
     }
@@ -104,9 +107,25 @@ public:
         SignalFence(std::move(func));
     }
 
-    void WaitPendingFences() {
+    void WaitPendingFences([[maybe_unused]] bool force) {
         if constexpr (!can_async_check) {
             TryReleasePendingFences<true>();
+        } else {
+            if (!force) {
+                return;
+            }
+            std::mutex wait_mutex;
+            std::condition_variable wait_cv;
+            std::atomic<bool> wait_finished{};
+            std::function<void()> func([&] {
+                std::scoped_lock lk(wait_mutex);
+                wait_finished.store(true, std::memory_order_relaxed);
+                wait_cv.notify_all();
+            });
+            SignalFence(std::move(func));
+            std::unique_lock lk(wait_mutex);
+            wait_cv.wait(
+                lk, [&wait_finished] { return wait_finished.load(std::memory_order_relaxed); });
         }
     }
 

src/video_core/gpu.cpp

@@ -102,7 +102,8 @@ struct GPU::Impl {
 
     /// Signal the ending of command list.
     void OnCommandListEnd() {
-        rasterizer->ReleaseFences();
+        rasterizer->ReleaseFences(false);
+        Settings::UpdateGPUAccuracy();
     }
 
     /// Request a host GPU memory flush from the CPU.
@@ -220,6 +221,7 @@ struct GPU::Impl {
     /// This can be used to launch any necessary threads and register any necessary
     /// core timing events.
     void Start() {
+        Settings::UpdateGPUAccuracy();
         gpu_thread.StartThread(*renderer, renderer->Context(), *scheduler);
     }
 

src/video_core/host_shaders/CMakeLists.txt

@@ -41,6 +41,9 @@ set(SHADER_FILES
     pitch_unswizzle.comp
     present_bicubic.frag
     present_gaussian.frag
+    queries_prefix_scan_sum.comp
+    queries_prefix_scan_sum_nosubgroups.comp
+    resolve_conditional_render.comp
     smaa_edge_detection.vert
     smaa_edge_detection.frag
     smaa_blending_weight_calculation.vert
@@ -70,6 +73,7 @@ if ("${GLSLANGVALIDATOR}" STREQUAL "GLSLANGVALIDATOR-NOTFOUND")
 endif()
 
 set(GLSL_FLAGS "")
+set(SPIR_V_VERSION "spirv1.3")
 set(QUIET_FLAG "--quiet")
 
 set(SHADER_INCLUDE ${CMAKE_CURRENT_BINARY_DIR}/include)
@@ -123,7 +127,7 @@ foreach(FILENAME IN ITEMS ${SHADER_FILES})
             OUTPUT
                 ${SPIRV_HEADER_FILE}
             COMMAND
-                ${GLSLANGVALIDATOR} -V ${QUIET_FLAG} -I"${FIDELITYFX_INCLUDE_DIR}" ${GLSL_FLAGS} --variable-name ${SPIRV_VARIABLE_NAME} -o ${SPIRV_HEADER_FILE} ${SOURCE_FILE}
+                ${GLSLANGVALIDATOR} -V ${QUIET_FLAG} -I"${FIDELITYFX_INCLUDE_DIR}" ${GLSL_FLAGS} --variable-name ${SPIRV_VARIABLE_NAME} -o ${SPIRV_HEADER_FILE} ${SOURCE_FILE} --target-env ${SPIR_V_VERSION}
             MAIN_DEPENDENCY
                 ${SOURCE_FILE}
         )

src/video_core/host_shaders/queries_prefix_scan_sum.comp

@@ -0,0 +1,173 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#version 460 core
+
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_shuffle : require
+#extension GL_KHR_shader_subgroup_shuffle_relative : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+#ifdef VULKAN
+
+#define HAS_EXTENDED_TYPES 1
+#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
+#define END_PUSH_CONSTANTS };
+#define UNIFORM(n)
+#define BINDING_INPUT_BUFFER 0
+#define BINDING_OUTPUT_IMAGE 1
+
+#else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
+
+#extension GL_NV_gpu_shader5 : enable
+#ifdef GL_NV_gpu_shader5
+#define HAS_EXTENDED_TYPES 1
+#else
+#define HAS_EXTENDED_TYPES 0
+#endif
+#define BEGIN_PUSH_CONSTANTS
+#define END_PUSH_CONSTANTS
+#define UNIFORM(n) layout(location = n) uniform
+#define BINDING_INPUT_BUFFER 0
+#define BINDING_OUTPUT_IMAGE 0
+
+#endif
+
+BEGIN_PUSH_CONSTANTS
+UNIFORM(0) uint min_accumulation_base;
+UNIFORM(1) uint max_accumulation_base;
+UNIFORM(2) uint accumulation_limit;
+UNIFORM(3) uint buffer_offset;
+END_PUSH_CONSTANTS
+
+#define LOCAL_RESULTS 8
+#define QUERIES_PER_INVOC 2048
+
+layout(local_size_x = QUERIES_PER_INVOC / LOCAL_RESULTS) in;
+
+layout(std430, binding = 0) readonly buffer block1 {
+    uvec2 input_data[];
+};
+
+layout(std430, binding = 1) coherent buffer block2 {
+    uvec2 output_data[];
+};
+
+layout(std430, binding = 2) coherent buffer block3 {
+    uvec2 accumulated_data;
+};
+
+shared uvec2 shared_data[128];
+
+// Simple Uint64 add that uses 2 uint variables for GPUs that don't support uint64
+uvec2 AddUint64(uvec2 value_1, uvec2 value_2) {
+    uint carry = 0;
+    uvec2 result;
+    result.x = uaddCarry(value_1.x, value_2.x, carry);
+    result.y = value_1.y + value_2.y + carry;
+    return result;
+}
+
+// do subgroup Prefix Sum using Hillis and Steele's algorithm
+uvec2 subgroupInclusiveAddUint64(uvec2 value) {
+    uvec2 result = value;
+    for (uint i = 1; i < gl_SubgroupSize; i *= 2) {
+        uvec2 other = subgroupShuffleUp(result, i); // get value from subgroup_inv_id - i;
+        if (i <= gl_SubgroupInvocationID) {
+            result = AddUint64(result, other);
+        }
+    }
+    return result;
+}
+
+// Writes down the results to the output buffer and to the accumulation buffer
+void WriteResults(uvec2 results[LOCAL_RESULTS]) {
+    const uint current_id = gl_LocalInvocationID.x;
+    const uvec2 accum = accumulated_data;
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        uvec2 base_data = current_id * LOCAL_RESULTS + i < min_accumulation_base ? accum : uvec2(0, 0);
+        AddUint64(results[i], base_data);
+    }
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        output_data[buffer_offset + current_id * LOCAL_RESULTS + i] = results[i];
+    }
+    uint index = accumulation_limit % LOCAL_RESULTS;
+    uint base_id = accumulation_limit / LOCAL_RESULTS;
+    if (min_accumulation_base >= accumulation_limit + 1) {
+        if (current_id == base_id) {
+            accumulated_data = results[index];
+        }
+        return;
+    }
+    // We have that ugly case in which the accumulation data is reset in the middle somewhere.
+    barrier();
+    groupMemoryBarrier();
+
+    if (current_id == base_id) {
+        uvec2 reset_value = output_data[max_accumulation_base - 1];
+        // Calculate two complement / negate manually
+        reset_value = AddUint64(uvec2(1,0), ~reset_value);
+        accumulated_data = AddUint64(results[index], reset_value);
+    }
+}
+
+void main() {
+    const uint subgroup_inv_id = gl_SubgroupInvocationID;
+    const uint subgroup_id = gl_SubgroupID + gl_WorkGroupID.x * gl_NumSubgroups;
+    const uint last_subgroup_id = subgroupMax(subgroup_inv_id);
+    const uint current_id = gl_LocalInvocationID.x;
+    const uint total_work = accumulation_limit;
+    const uint last_result_id = LOCAL_RESULTS - 1;
+    uvec2 data[LOCAL_RESULTS];
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        data[i] = input_data[buffer_offset + current_id * LOCAL_RESULTS + i];
+    }
+    uvec2 results[LOCAL_RESULTS];
+    results[0] = data[0];
+    for (uint i = 1; i < LOCAL_RESULTS; i++) {
+        results[i] = AddUint64(data[i], results[i - 1]);
+    }
+    // make sure all input data has been loaded
+    subgroupBarrier();
+    subgroupMemoryBarrier();
+
+    // on the last local result, do a subgroup inclusive scan sum
+    results[last_result_id] = subgroupInclusiveAddUint64(results[last_result_id]);
+    // get the last local result from the subgroup behind the current
+    uvec2 result_behind = subgroupShuffleUp(results[last_result_id], 1);
+    if (subgroup_inv_id != 0) {
+        for (uint i = 1; i < LOCAL_RESULTS; i++) {
+            results[i - 1] = AddUint64(results[i - 1], result_behind);
+        }
+    }
+
+    // if we had less queries than our subgroup, just write down the results.
+    if (total_work <= gl_SubgroupSize * LOCAL_RESULTS) { // This condition is constant per dispatch.
+        WriteResults(results);
+        return;
+    }
+
+    // We now have more, so lets write the last result into shared memory.
+    // Only pick the last subgroup.
+    if (subgroup_inv_id == last_subgroup_id) {
+        shared_data[subgroup_id] = results[last_result_id];
+    }
+    // wait until everyone loaded their stuffs
+    barrier();
+    memoryBarrierShared();
+
+    // only if it's not the first subgroup
+    if (subgroup_id != 0) {
+        // get the results from some previous invocation
+        uvec2 tmp = shared_data[subgroup_inv_id];
+        subgroupBarrier();
+        subgroupMemoryBarrierShared();
+        tmp = subgroupInclusiveAddUint64(tmp);
+        // obtain the result that would be equivalent to the previous result
+        uvec2 shuffled_result = subgroupShuffle(tmp, subgroup_id - 1);
+        for (uint i = 0; i < LOCAL_RESULTS; i++) {
+            results[i] = AddUint64(results[i], shuffled_result);
+        }
+    }
+    WriteResults(results);
+}

src/video_core/host_shaders/queries_prefix_scan_sum_nosubgroups.comp

@@ -0,0 +1,138 @@
+// SPDX-FileCopyrightText: Copyright 2015 Graham Sellers, Richard Wright Jr. and Nicholas Haemel
+// SPDX-License-Identifier: MIT
+
+// Code obtained from OpenGL SuperBible, Seventh Edition by Graham Sellers, Richard Wright Jr. and
+// Nicholas Haemel. Modified to suit needs.
+
+#version 460 core
+
+#ifdef VULKAN
+
+#define HAS_EXTENDED_TYPES 1
+#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
+#define END_PUSH_CONSTANTS };
+#define UNIFORM(n)
+#define BINDING_INPUT_BUFFER 0
+#define BINDING_OUTPUT_IMAGE 1
+
+#else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
+
+#extension GL_NV_gpu_shader5 : enable
+#ifdef GL_NV_gpu_shader5
+#define HAS_EXTENDED_TYPES 1
+#else
+#define HAS_EXTENDED_TYPES 0
+#endif
+#define BEGIN_PUSH_CONSTANTS
+#define END_PUSH_CONSTANTS
+#define UNIFORM(n) layout(location = n) uniform
+#define BINDING_INPUT_BUFFER 0
+#define BINDING_OUTPUT_IMAGE 0
+
+#endif
+
+BEGIN_PUSH_CONSTANTS
+UNIFORM(0) uint min_accumulation_base;
+UNIFORM(1) uint max_accumulation_base;
+UNIFORM(2) uint accumulation_limit;
+UNIFORM(3) uint buffer_offset;
+END_PUSH_CONSTANTS
+
+#define LOCAL_RESULTS 4
+#define QUERIES_PER_INVOC 2048
+
+layout(local_size_x = QUERIES_PER_INVOC / LOCAL_RESULTS) in;
+
+layout(std430, binding = 0) readonly buffer block1 {
+    uvec2 input_data[gl_WorkGroupSize.x * LOCAL_RESULTS];
+};
+
+layout(std430, binding = 1) writeonly coherent buffer block2 {
+    uvec2 output_data[gl_WorkGroupSize.x * LOCAL_RESULTS];
+};
+
+layout(std430, binding = 2) coherent buffer block3 {
+    uvec2 accumulated_data;
+};
+
+shared uvec2 shared_data[gl_WorkGroupSize.x * LOCAL_RESULTS];
+
+uvec2 AddUint64(uvec2 value_1, uvec2 value_2) {
+    uint carry = 0;
+    uvec2 result;
+    result.x = uaddCarry(value_1.x, value_2.x, carry);
+    result.y = value_1.y + value_2.y + carry;
+    return result;
+}
+
+void main(void) {
+    uint id = gl_LocalInvocationID.x;
+    uvec2 base_value[LOCAL_RESULTS];
+    const uvec2 accum = accumulated_data;
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        base_value[i] = (buffer_offset + id * LOCAL_RESULTS + i) < min_accumulation_base
+                            ? accumulated_data
+                            : uvec2(0);
+    }
+    uint work_size = gl_WorkGroupSize.x;
+    uint rd_id;
+    uint wr_id;
+    uint mask;
+    uvec2 inputs[LOCAL_RESULTS];
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        inputs[i] = input_data[buffer_offset + id * LOCAL_RESULTS + i];
+    }
+    // The number of steps is the log base 2 of the
+    // work group size, which should be a power of 2
+    const uint steps = uint(log2(work_size)) + uint(log2(LOCAL_RESULTS));
+    uint step = 0;
+
+    // Each invocation is responsible for the content of
+    // two elements of the output array
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        shared_data[id * LOCAL_RESULTS + i] = inputs[i];
+    }
+    // Synchronize to make sure that everyone has initialized
+    // their elements of shared_data[] with data loaded from
+    // the input arrays
+    barrier();
+    memoryBarrierShared();
+    // For each step...
+    for (step = 0; step < steps; step++) {
+        // Calculate the read and write index in the
+        // shared array
+        mask = (1 << step) - 1;
+        rd_id = ((id >> step) << (step + 1)) + mask;
+        wr_id = rd_id + 1 + (id & mask);
+        // Accumulate the read data into our element
+
+        shared_data[wr_id] = AddUint64(shared_data[rd_id], shared_data[wr_id]);
+        // Synchronize again to make sure that everyone
+        // has caught up with us
+        barrier();
+        memoryBarrierShared();
+    }
+    // Add the accumulation
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        shared_data[id * LOCAL_RESULTS + i] =
+            AddUint64(shared_data[id * LOCAL_RESULTS + i], base_value[i]);
+    }
+    barrier();
+    memoryBarrierShared();
+
+    // Finally write our data back to the output buffer
+    for (uint i = 0; i < LOCAL_RESULTS; i++) {
+        output_data[buffer_offset + id * LOCAL_RESULTS + i] = shared_data[id * LOCAL_RESULTS + i];
+    }
+    if (id == 0) {
+        if (min_accumulation_base >= accumulation_limit + 1) {
+            accumulated_data = shared_data[accumulation_limit];
+            return;
+        }
+        uvec2 reset_value = shared_data[max_accumulation_base - 1];
+        uvec2 final_value = shared_data[accumulation_limit];
+        // Two complements
+        reset_value = AddUint64(uvec2(1, 0), ~reset_value);
+        accumulated_data = AddUint64(final_value, reset_value);
+    }
+}

src/video_core/host_shaders/resolve_conditional_render.comp

@@ -0,0 +1,20 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#version 450
+
+layout(local_size_x = 1) in;
+
+layout(std430, binding = 0) buffer Query {
+    uvec2 initial;
+    uvec2 unknown;
+    uvec2 current;
+};
+
+layout(std430, binding = 1) buffer Result {
+    uint result;
+};
+
+void main() {
+    result = all(equal(initial, current)) ? 1 : 0;
+}

src/video_core/macro/macro_hle.cpp

@@ -67,6 +67,7 @@ public:
         }
 
         auto& params = maxwell3d.draw_manager->GetIndirectParams();
+        params.is_byte_count = false;
         params.is_indexed = false;
         params.include_count = false;
         params.count_start_address = 0;
@@ -161,6 +162,7 @@ public:
                 0, 0x644, Maxwell3D::HLEReplacementAttributeType::BaseInstance);
         }
         auto& params = maxwell3d.draw_manager->GetIndirectParams();
+        params.is_byte_count = false;
         params.is_indexed = true;
         params.include_count = false;
         params.count_start_address = 0;
@@ -256,6 +258,7 @@ public:
         const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
         auto& params = maxwell3d.draw_manager->GetIndirectParams();
+        params.is_byte_count = false;
         params.is_indexed = true;
         params.include_count = true;
         params.count_start_address = maxwell3d.GetMacroAddress(4);
@@ -319,6 +322,47 @@ private:
     }
 };
 
+class HLE_DrawIndirectByteCount final : public HLEMacroImpl {
+public:
+    explicit HLE_DrawIndirectByteCount(Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
+
+    void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
+        auto topology = static_cast<Maxwell3D::Regs::PrimitiveTopology>(parameters[0] & 0xFFFFU);
+        if (!maxwell3d.AnyParametersDirty() || !IsTopologySafe(topology)) {
+            Fallback(parameters);
+            return;
+        }
+
+        auto& params = maxwell3d.draw_manager->GetIndirectParams();
+        params.is_byte_count = true;
+        params.is_indexed = false;
+        params.include_count = false;
+        params.count_start_address = 0;
+        params.indirect_start_address = maxwell3d.GetMacroAddress(2);
+        params.buffer_size = 4;
+        params.max_draw_counts = 1;
+        params.stride = parameters[1];
+        maxwell3d.regs.draw.begin = parameters[0];
+        maxwell3d.regs.draw_auto_stride = parameters[1];
+        maxwell3d.regs.draw_auto_byte_count = parameters[2];
+
+        maxwell3d.draw_manager->DrawArrayIndirect(topology);
+    }
+
+private:
+    void Fallback(const std::vector<u32>& parameters) {
+        maxwell3d.RefreshParameters();
+
+        maxwell3d.regs.draw.begin = parameters[0];
+        maxwell3d.regs.draw_auto_stride = parameters[1];
+        maxwell3d.regs.draw_auto_byte_count = parameters[2];
+
+        maxwell3d.draw_manager->DrawArray(
+            maxwell3d.regs.draw.topology, 0,
+            maxwell3d.regs.draw_auto_byte_count / maxwell3d.regs.draw_auto_stride, 0, 1);
+    }
+};
+
 class HLE_C713C83D8F63CCF3 final : public HLEMacroImpl {
 public:
     explicit HLE_C713C83D8F63CCF3(Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
@@ -536,6 +580,11 @@ HLEMacro::HLEMacro(Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {
                          [](Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
                              return std::make_unique<HLE_TransformFeedbackSetup>(maxwell3d__);
                          }));
+    builders.emplace(0xB5F74EDB717278ECULL,
+                     std::function<std::unique_ptr<CachedMacro>(Maxwell3D&)>(
+                         [](Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
+                             return std::make_unique<HLE_DrawIndirectByteCount>(maxwell3d__);
+                         }));
 }
 
 HLEMacro::~HLEMacro() = default;

src/video_core/query_cache.h

@@ -25,6 +25,13 @@
 #include "video_core/rasterizer_interface.h"
 #include "video_core/texture_cache/slot_vector.h"
 
+namespace VideoCore {
+enum class QueryType {
+    SamplesPassed,
+};
+constexpr std::size_t NumQueryTypes = 1;
+} // namespace VideoCore
+
 namespace VideoCommon {
 
 using AsyncJobId = SlotId;
@@ -98,9 +105,9 @@ private:
 };
 
 template <class QueryCache, class CachedQuery, class CounterStream, class HostCounter>
-class QueryCacheBase : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo> {
+class QueryCacheLegacy : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo> {
 public:
-    explicit QueryCacheBase(VideoCore::RasterizerInterface& rasterizer_,
+    explicit QueryCacheLegacy(VideoCore::RasterizerInterface& rasterizer_,
                               Core::Memory::Memory& cpu_memory_)
         : rasterizer{rasterizer_},
           // Use reinterpret_cast instead of static_cast as workaround for

src/video_core/query_cache/bank_base.h

@@ -0,0 +1,104 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <atomic>
+#include <deque>
+#include <utility>
+
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+class BankBase {
+protected:
+    const size_t base_bank_size{};
+    size_t bank_size{};
+    std::atomic<size_t> references{};
+    size_t current_slot{};
+
+public:
+    explicit BankBase(size_t bank_size_) : base_bank_size{bank_size_}, bank_size(bank_size_) {}
+
+    virtual ~BankBase() = default;
+
+    virtual std::pair<bool, size_t> Reserve() {
+        if (IsClosed()) {
+            return {false, bank_size};
+        }
+        const size_t result = current_slot++;
+        return {true, result};
+    }
+
+    virtual void Reset() {
+        current_slot = 0;
+        references = 0;
+        bank_size = base_bank_size;
+    }
+
+    size_t Size() const {
+        return bank_size;
+    }
+
+    void AddReference(size_t how_many = 1) {
+        references.fetch_add(how_many, std::memory_order_relaxed);
+    }
+
+    void CloseReference(size_t how_many = 1) {
+        if (how_many > references.load(std::memory_order_relaxed)) {
+            UNREACHABLE();
+        }
+        references.fetch_sub(how_many, std::memory_order_relaxed);
+    }
+
+    void Close() {
+        bank_size = current_slot;
+    }
+
+    bool IsClosed() const {
+        return current_slot >= bank_size;
+    }
+
+    bool IsDead() const {
+        return IsClosed() && references == 0;
+    }
+};
+
+template <typename BankType>
+class BankPool {
+private:
+    std::deque<BankType> bank_pool;
+    std::deque<size_t> bank_indices;
+
+public:
+    BankPool() = default;
+    ~BankPool() = default;
+
+    // Reserve a bank from the pool and return its index
+    template <typename Func>
+    size_t ReserveBank(Func&& builder) {
+        if (!bank_indices.empty() && bank_pool[bank_indices.front()].IsDead()) {
+            size_t new_index = bank_indices.front();
+            bank_indices.pop_front();
+            bank_pool[new_index].Reset();
+            return new_index;
+        }
+        size_t new_index = bank_pool.size();
+        builder(bank_pool, new_index);
+        bank_indices.push_back(new_index);
+        return new_index;
+    }
+
+    // Get a reference to a bank using its index
+    BankType& GetBank(size_t index) {
+        return bank_pool[index];
+    }
+
+    // Get the total number of banks in the pool
+    size_t BankCount() const {
+        return bank_pool.size();
+    }
+};
+
+} // namespace VideoCommon

src/video_core/query_cache/query_base.h

@@ -0,0 +1,70 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/common_funcs.h"
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+enum class QueryFlagBits : u32 {
+    HasTimestamp = 1 << 0,       ///< Indicates if this query has a timestamp.
+    IsFinalValueSynced = 1 << 1, ///< Indicates if the query has been synced in the host
+    IsHostSynced = 1 << 2,       ///< Indicates if the query has been synced in the host
+    IsGuestSynced = 1 << 3,      ///< Indicates if the query has been synced with the guest.
+    IsHostManaged = 1 << 4,      ///< Indicates if this query points to a host query
+    IsRewritten = 1 << 5,        ///< Indicates if this query was rewritten by another query
+    IsInvalidated = 1 << 6,      ///< Indicates the value of th query has been nullified.
+    IsOrphan = 1 << 7,           ///< Indicates the query has not been set by a guest query.
+    IsFence = 1 << 8,            ///< Indicates the query is a fence.
+};
+DECLARE_ENUM_FLAG_OPERATORS(QueryFlagBits)
+
+class QueryBase {
+public:
+    VAddr guest_address{};
+    QueryFlagBits flags{};
+    u64 value{};
+
+protected:
+    // Default constructor
+    QueryBase() = default;
+
+    // Parameterized constructor
+    QueryBase(VAddr address, QueryFlagBits flags_, u64 value_)
+        : guest_address(address), flags(flags_), value{value_} {}
+};
+
+class GuestQuery : public QueryBase {
+public:
+    // Parameterized constructor
+    GuestQuery(bool isLong, VAddr address, u64 queryValue)
+        : QueryBase(address, QueryFlagBits::IsFinalValueSynced, queryValue) {
+        if (isLong) {
+            flags |= QueryFlagBits::HasTimestamp;
+        }
+    }
+};
+
+class HostQueryBase : public QueryBase {
+public:
+    // Default constructor
+    HostQueryBase() : QueryBase(0, QueryFlagBits::IsHostManaged | QueryFlagBits::IsOrphan, 0) {}
+
+    // Parameterized constructor
+    HostQueryBase(bool has_timestamp, VAddr address)
+        : QueryBase(address, QueryFlagBits::IsHostManaged, 0), start_bank_id{}, size_banks{},
+          start_slot{}, size_slots{} {
+        if (has_timestamp) {
+            flags |= QueryFlagBits::HasTimestamp;
+        }
+    }
+
+    u32 start_bank_id{};
+    u32 size_banks{};
+    size_t start_slot{};
+    size_t size_slots{};
+};
+
+} // namespace VideoCommon

src/video_core/query_cache/query_cache.h

@@ -0,0 +1,580 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+#include <deque>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+#include <utility>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "common/logging/log.h"
+#include "common/scope_exit.h"
+#include "common/settings.h"
+#include "core/memory.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
+#include "video_core/query_cache/bank_base.h"
+#include "video_core/query_cache/query_base.h"
+#include "video_core/query_cache/query_cache_base.h"
+#include "video_core/query_cache/query_stream.h"
+#include "video_core/query_cache/types.h"
+
+namespace VideoCommon {
+
+using Maxwell = Tegra::Engines::Maxwell3D;
+
+struct SyncValuesStruct {
+    VAddr address;
+    u64 value;
+    u64 size;
+
+    static constexpr bool GeneratesBaseBuffer = true;
+};
+
+template <typename Traits>
+class GuestStreamer : public SimpleStreamer<GuestQuery> {
+public:
+    using RuntimeType = typename Traits::RuntimeType;
+
+    GuestStreamer(size_t id_, RuntimeType& runtime_)
+        : SimpleStreamer<GuestQuery>(id_), runtime{runtime_} {}
+
+    virtual ~GuestStreamer() = default;
+
+    size_t WriteCounter(VAddr address, bool has_timestamp, u32 value,
+                        std::optional<u32> subreport = std::nullopt) override {
+        auto new_id = BuildQuery(has_timestamp, address, static_cast<u64>(value));
+        pending_sync.push_back(new_id);
+        return new_id;
+    }
+
+    bool HasPendingSync() const override {
+        return !pending_sync.empty();
+    }
+
+    void SyncWrites() override {
+        if (pending_sync.empty()) {
+            return;
+        }
+        std::vector<SyncValuesStruct> sync_values;
+        sync_values.reserve(pending_sync.size());
+        for (size_t pending_id : pending_sync) {
+            auto& query = slot_queries[pending_id];
+            if (True(query.flags & QueryFlagBits::IsRewritten) ||
+                True(query.flags & QueryFlagBits::IsInvalidated)) {
+                continue;
+            }
+            query.flags |= QueryFlagBits::IsHostSynced;
+            sync_values.emplace_back(SyncValuesStruct{
+                .address = query.guest_address,
+                .value = query.value,
+                .size = static_cast<u64>(True(query.flags & QueryFlagBits::HasTimestamp) ? 8 : 4)});
+        }
+        pending_sync.clear();
+        if (sync_values.size() > 0) {
+            runtime.template SyncValues<SyncValuesStruct>(sync_values);
+        }
+    }
+
+private:
+    RuntimeType& runtime;
+    std::deque<size_t> pending_sync;
+};
+
+template <typename Traits>
+class StubStreamer : public GuestStreamer<Traits> {
+public:
+    using RuntimeType = typename Traits::RuntimeType;
+
+    StubStreamer(size_t id_, RuntimeType& runtime_, u32 stub_value_)
+        : GuestStreamer<Traits>(id_, runtime_), stub_value{stub_value_} {}
+
+    ~StubStreamer() override = default;
+
+    size_t WriteCounter(VAddr address, bool has_timestamp, [[maybe_unused]] u32 value,
+                        std::optional<u32> subreport = std::nullopt) override {
+        size_t new_id =
+            GuestStreamer<Traits>::WriteCounter(address, has_timestamp, stub_value, subreport);
+        return new_id;
+    }
+
+private:
+    u32 stub_value;
+};
+
+template <typename Traits>
+struct QueryCacheBase<Traits>::QueryCacheBaseImpl {
+    using RuntimeType = typename Traits::RuntimeType;
+
+    QueryCacheBaseImpl(QueryCacheBase<Traits>* owner_, VideoCore::RasterizerInterface& rasterizer_,
+                       Core::Memory::Memory& cpu_memory_, RuntimeType& runtime_, Tegra::GPU& gpu_)
+        : owner{owner_}, rasterizer{rasterizer_},
+          cpu_memory{cpu_memory_}, runtime{runtime_}, gpu{gpu_} {
+        streamer_mask = 0;
+        for (size_t i = 0; i < static_cast<size_t>(QueryType::MaxQueryTypes); i++) {
+            streamers[i] = runtime.GetStreamerInterface(static_cast<QueryType>(i));
+            if (streamers[i]) {
+                streamer_mask |= 1ULL << streamers[i]->GetId();
+            }
+        }
+    }
+
+    template <typename Func>
+    void ForEachStreamerIn(u64 mask, Func&& func) {
+        static constexpr bool RETURNS_BOOL =
+            std::is_same_v<std::invoke_result<Func, StreamerInterface*>, bool>;
+        while (mask != 0) {
+            size_t position = std::countr_zero(mask);
+            mask &= ~(1ULL << position);
+            if constexpr (RETURNS_BOOL) {
+                if (func(streamers[position])) {
+                    return;
+                }
+            } else {
+                func(streamers[position]);
+            }
+        }
+    }
+
+    template <typename Func>
+    void ForEachStreamer(Func&& func) {
+        ForEachStreamerIn(streamer_mask, func);
+    }
+
+    QueryBase* ObtainQuery(QueryCacheBase<Traits>::QueryLocation location) {
+        size_t which_stream = location.stream_id.Value();
+        auto* streamer = streamers[which_stream];
+        if (!streamer) {
+            return nullptr;
+        }
+        return streamer->GetQuery(location.query_id.Value());
+    }
+
+    QueryCacheBase<Traits>* owner;
+    VideoCore::RasterizerInterface& rasterizer;
+    Core::Memory::Memory& cpu_memory;
+    RuntimeType& runtime;
+    Tegra::GPU& gpu;
+    std::array<StreamerInterface*, static_cast<size_t>(QueryType::MaxQueryTypes)> streamers;
+    u64 streamer_mask;
+    std::mutex flush_guard;
+    std::deque<u64> flushes_pending;
+    std::vector<QueryCacheBase<Traits>::QueryLocation> pending_unregister;
+};
+
+template <typename Traits>
+QueryCacheBase<Traits>::QueryCacheBase(Tegra::GPU& gpu_,
+                                       VideoCore::RasterizerInterface& rasterizer_,
+                                       Core::Memory::Memory& cpu_memory_, RuntimeType& runtime_)
+    : cached_queries{} {
+    impl = std::make_unique<QueryCacheBase<Traits>::QueryCacheBaseImpl>(
+        this, rasterizer_, cpu_memory_, runtime_, gpu_);
+}
+
+template <typename Traits>
+QueryCacheBase<Traits>::~QueryCacheBase() = default;
+
+template <typename Traits>
+void QueryCacheBase<Traits>::CounterEnable(QueryType counter_type, bool is_enabled) {
+    size_t index = static_cast<size_t>(counter_type);
+    StreamerInterface* streamer = impl->streamers[index];
+    if (!streamer) [[unlikely]] {
+        UNREACHABLE();
+        return;
+    }
+    if (is_enabled) {
+        streamer->StartCounter();
+    } else {
+        streamer->PauseCounter();
+    }
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::CounterClose(QueryType counter_type) {
+    size_t index = static_cast<size_t>(counter_type);
+    StreamerInterface* streamer = impl->streamers[index];
+    if (!streamer) [[unlikely]] {
+        UNREACHABLE();
+        return;
+    }
+    streamer->CloseCounter();
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::CounterReset(QueryType counter_type) {
+    size_t index = static_cast<size_t>(counter_type);
+    StreamerInterface* streamer = impl->streamers[index];
+    if (!streamer) [[unlikely]] {
+        UNIMPLEMENTED();
+        return;
+    }
+    streamer->ResetCounter();
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::BindToChannel(s32 id) {
+    VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo>::BindToChannel(id);
+    impl->runtime.Bind3DEngine(maxwell3d);
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type,
+                                           QueryPropertiesFlags flags, u32 payload, u32 subreport) {
+    const bool has_timestamp = True(flags & QueryPropertiesFlags::HasTimeout);
+    const bool is_fence = True(flags & QueryPropertiesFlags::IsAFence);
+    size_t streamer_id = static_cast<size_t>(counter_type);
+    auto* streamer = impl->streamers[streamer_id];
+    if (streamer == nullptr) [[unlikely]] {
+        counter_type = QueryType::Payload;
+        payload = 1U;
+        streamer_id = static_cast<size_t>(counter_type);
+        streamer = impl->streamers[streamer_id];
+    }
+    auto cpu_addr_opt = gpu_memory->GpuToCpuAddress(addr);
+    if (!cpu_addr_opt) [[unlikely]] {
+        return;
+    }
+    VAddr cpu_addr = *cpu_addr_opt;
+    const size_t new_query_id = streamer->WriteCounter(cpu_addr, has_timestamp, payload, subreport);
+    auto* query = streamer->GetQuery(new_query_id);
+    if (is_fence) {
+        query->flags |= QueryFlagBits::IsFence;
+    }
+    QueryLocation query_location{};
+    query_location.stream_id.Assign(static_cast<u32>(streamer_id));
+    query_location.query_id.Assign(static_cast<u32>(new_query_id));
+    const auto gen_caching_indexing = [](VAddr cur_addr) {
+        return std::make_pair<u64, u32>(cur_addr >> Core::Memory::YUZU_PAGEBITS,
+                                        static_cast<u32>(cur_addr & Core::Memory::YUZU_PAGEMASK));
+    };
+    u8* pointer = impl->cpu_memory.GetPointer(cpu_addr);
+    u8* pointer_timestamp = impl->cpu_memory.GetPointer(cpu_addr + 8);
+    bool is_synced = !Settings::IsGPULevelHigh() && is_fence;
+
+    std::function<void()> operation([this, is_synced, streamer, query_base = query, query_location,
+                                     pointer, pointer_timestamp] {
+        if (True(query_base->flags & QueryFlagBits::IsInvalidated)) {
+            if (!is_synced) [[likely]] {
+                impl->pending_unregister.push_back(query_location);
+            }
+            return;
+        }
+        if (False(query_base->flags & QueryFlagBits::IsFinalValueSynced)) [[unlikely]] {
+            UNREACHABLE();
+            return;
+        }
+        query_base->value += streamer->GetAmmendValue();
+        streamer->SetAccumulationValue(query_base->value);
+        if (True(query_base->flags & QueryFlagBits::HasTimestamp)) {
+            u64 timestamp = impl->gpu.GetTicks();
+            std::memcpy(pointer_timestamp, &timestamp, sizeof(timestamp));
+            std::memcpy(pointer, &query_base->value, sizeof(query_base->value));
+        } else {
+            u32 value = static_cast<u32>(query_base->value);
+            std::memcpy(pointer, &value, sizeof(value));
+        }
+        if (!is_synced) [[likely]] {
+            impl->pending_unregister.push_back(query_location);
+        }
+    });
+    if (is_fence) {
+        impl->rasterizer.SignalFence(std::move(operation));
+    } else {
+        if (!Settings::IsGPULevelHigh() && counter_type == QueryType::Payload) {
+            if (has_timestamp) {
+                u64 timestamp = impl->gpu.GetTicks();
+                u64 value = static_cast<u64>(payload);
+                std::memcpy(pointer_timestamp, &timestamp, sizeof(timestamp));
+                std::memcpy(pointer, &value, sizeof(value));
+            } else {
+                std::memcpy(pointer, &payload, sizeof(payload));
+            }
+            streamer->Free(new_query_id);
+            return;
+        }
+        impl->rasterizer.SyncOperation(std::move(operation));
+    }
+    if (is_synced) {
+        streamer->Free(new_query_id);
+        return;
+    }
+    auto [cont_addr, base] = gen_caching_indexing(cpu_addr);
+    {
+        std::scoped_lock lock(cache_mutex);
+        auto it1 = cached_queries.try_emplace(cont_addr);
+        auto& sub_container = it1.first->second;
+        auto it_current = sub_container.find(base);
+        if (it_current == sub_container.end()) {
+            sub_container.insert_or_assign(base, query_location);
+            return;
+        }
+        auto* old_query = impl->ObtainQuery(it_current->second);
+        old_query->flags |= QueryFlagBits::IsRewritten;
+        sub_container.insert_or_assign(base, query_location);
+    }
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::UnregisterPending() {
+    const auto gen_caching_indexing = [](VAddr cur_addr) {
+        return std::make_pair<u64, u32>(cur_addr >> Core::Memory::YUZU_PAGEBITS,
+                                        static_cast<u32>(cur_addr & Core::Memory::YUZU_PAGEMASK));
+    };
+    std::scoped_lock lock(cache_mutex);
+    for (QueryLocation loc : impl->pending_unregister) {
+        const auto [streamer_id, query_id] = loc.unpack();
+        auto* streamer = impl->streamers[streamer_id];
+        if (!streamer) [[unlikely]] {
+            continue;
+        }
+        auto* query = streamer->GetQuery(query_id);
+        auto [cont_addr, base] = gen_caching_indexing(query->guest_address);
+        auto it1 = cached_queries.find(cont_addr);
+        if (it1 != cached_queries.end()) {
+            auto it2 = it1->second.find(base);
+            if (it2 != it1->second.end()) {
+                if (it2->second.raw == loc.raw) {
+                    it1->second.erase(it2);
+                }
+            }
+        }
+        streamer->Free(query_id);
+    }
+    impl->pending_unregister.clear();
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::NotifyWFI() {
+    bool should_sync = false;
+    impl->ForEachStreamer(
+        [&should_sync](StreamerInterface* streamer) { should_sync |= streamer->HasPendingSync(); });
+    if (!should_sync) {
+        return;
+    }
+
+    impl->ForEachStreamer([](StreamerInterface* streamer) { streamer->PresyncWrites(); });
+    impl->runtime.Barriers(true);
+    impl->ForEachStreamer([](StreamerInterface* streamer) { streamer->SyncWrites(); });
+    impl->runtime.Barriers(false);
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::NotifySegment(bool resume) {
+    if (resume) {
+        impl->runtime.ResumeHostConditionalRendering();
+    } else {
+        CounterClose(VideoCommon::QueryType::ZPassPixelCount64);
+        CounterClose(VideoCommon::QueryType::StreamingByteCount);
+        impl->runtime.PauseHostConditionalRendering();
+    }
+}
+
+template <typename Traits>
+bool QueryCacheBase<Traits>::AccelerateHostConditionalRendering() {
+    bool qc_dirty = false;
+    const auto gen_lookup = [this, &qc_dirty](GPUVAddr address) -> VideoCommon::LookupData {
+        auto cpu_addr_opt = gpu_memory->GpuToCpuAddress(address);
+        if (!cpu_addr_opt) [[unlikely]] {
+            return VideoCommon::LookupData{
+                .address = 0,
+                .found_query = nullptr,
+            };
+        }
+        VAddr cpu_addr = *cpu_addr_opt;
+        std::scoped_lock lock(cache_mutex);
+        auto it1 = cached_queries.find(cpu_addr >> Core::Memory::YUZU_PAGEBITS);
+        if (it1 == cached_queries.end()) {
+            return VideoCommon::LookupData{
+                .address = cpu_addr,
+                .found_query = nullptr,
+            };
+        }
+        auto& sub_container = it1->second;
+        auto it_current = sub_container.find(cpu_addr & Core::Memory::YUZU_PAGEMASK);
+
+        if (it_current == sub_container.end()) {
+            auto it_current_2 = sub_container.find((cpu_addr & Core::Memory::YUZU_PAGEMASK) + 4);
+            if (it_current_2 == sub_container.end()) {
+                return VideoCommon::LookupData{
+                    .address = cpu_addr,
+                    .found_query = nullptr,
+                };
+            }
+        }
+        auto* query = impl->ObtainQuery(it_current->second);
+        qc_dirty |= True(query->flags & QueryFlagBits::IsHostManaged) &&
+                    False(query->flags & QueryFlagBits::IsGuestSynced);
+        return VideoCommon::LookupData{
+            .address = cpu_addr,
+            .found_query = query,
+        };
+    };
+
+    auto& regs = maxwell3d->regs;
+    if (regs.render_enable_override != Maxwell::Regs::RenderEnable::Override::UseRenderEnable) {
+        impl->runtime.EndHostConditionalRendering();
+        return false;
+    }
+    const ComparisonMode mode = static_cast<ComparisonMode>(regs.render_enable.mode);
+    const GPUVAddr address = regs.render_enable.Address();
+    switch (mode) {
+    case ComparisonMode::True:
+        impl->runtime.EndHostConditionalRendering();
+        return false;
+    case ComparisonMode::False:
+        impl->runtime.EndHostConditionalRendering();
+        return false;
+    case ComparisonMode::Conditional: {
+        VideoCommon::LookupData object_1{gen_lookup(address)};
+        return impl->runtime.HostConditionalRenderingCompareValue(object_1, qc_dirty);
+    }
+    case ComparisonMode::IfEqual: {
+        VideoCommon::LookupData object_1{gen_lookup(address)};
+        VideoCommon::LookupData object_2{gen_lookup(address + 16)};
+        return impl->runtime.HostConditionalRenderingCompareValues(object_1, object_2, qc_dirty,
+                                                                   true);
+    }
+    case ComparisonMode::IfNotEqual: {
+        VideoCommon::LookupData object_1{gen_lookup(address)};
+        VideoCommon::LookupData object_2{gen_lookup(address + 16)};
+        return impl->runtime.HostConditionalRenderingCompareValues(object_1, object_2, qc_dirty,
+                                                                   false);
+    }
+    default:
+        return false;
+    }
+}
+
+// Async downloads
+template <typename Traits>
+void QueryCacheBase<Traits>::CommitAsyncFlushes() {
+    // Make sure to have the results synced in Host.
+    NotifyWFI();
+
+    u64 mask{};
+    {
+        std::scoped_lock lk(impl->flush_guard);
+        impl->ForEachStreamer([&mask](StreamerInterface* streamer) {
+            bool local_result = streamer->HasUnsyncedQueries();
+            if (local_result) {
+                mask |= 1ULL << streamer->GetId();
+            }
+        });
+        impl->flushes_pending.push_back(mask);
+    }
+    std::function<void()> func([this] { UnregisterPending(); });
+    impl->rasterizer.SyncOperation(std::move(func));
+    if (mask == 0) {
+        return;
+    }
+    u64 ran_mask = ~mask;
+    while (mask) {
+        impl->ForEachStreamerIn(mask, [&mask, &ran_mask](StreamerInterface* streamer) {
+            u64 dep_mask = streamer->GetDependentMask();
+            if ((dep_mask & ~ran_mask) != 0) {
+                return;
+            }
+            u64 index = streamer->GetId();
+            ran_mask |= (1ULL << index);
+            mask &= ~(1ULL << index);
+            streamer->PushUnsyncedQueries();
+        });
+    }
+}
+
+template <typename Traits>
+bool QueryCacheBase<Traits>::HasUncommittedFlushes() const {
+    bool result = false;
+    impl->ForEachStreamer([&result](StreamerInterface* streamer) {
+        result |= streamer->HasUnsyncedQueries();
+        return result;
+    });
+    return result;
+}
+
+template <typename Traits>
+bool QueryCacheBase<Traits>::ShouldWaitAsyncFlushes() {
+    std::scoped_lock lk(impl->flush_guard);
+    return !impl->flushes_pending.empty() && impl->flushes_pending.front() != 0ULL;
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::PopAsyncFlushes() {
+    u64 mask;
+    {
+        std::scoped_lock lk(impl->flush_guard);
+        mask = impl->flushes_pending.front();
+        impl->flushes_pending.pop_front();
+    }
+    if (mask == 0) {
+        return;
+    }
+    u64 ran_mask = ~mask;
+    while (mask) {
+        impl->ForEachStreamerIn(mask, [&mask, &ran_mask](StreamerInterface* streamer) {
+            u64 dep_mask = streamer->GetDependenceMask();
+            if ((dep_mask & ~ran_mask) != 0) {
+                return;
+            }
+            u64 index = streamer->GetId();
+            ran_mask |= (1ULL << index);
+            mask &= ~(1ULL << index);
+            streamer->PopUnsyncedQueries();
+        });
+    }
+}
+
+// Invalidation
+
+template <typename Traits>
+void QueryCacheBase<Traits>::InvalidateQuery(QueryCacheBase<Traits>::QueryLocation location) {
+    auto* query_base = impl->ObtainQuery(location);
+    if (!query_base) {
+        return;
+    }
+    query_base->flags |= QueryFlagBits::IsInvalidated;
+}
+
+template <typename Traits>
+bool QueryCacheBase<Traits>::IsQueryDirty(QueryCacheBase<Traits>::QueryLocation location) {
+    auto* query_base = impl->ObtainQuery(location);
+    if (!query_base) {
+        return false;
+    }
+    return True(query_base->flags & QueryFlagBits::IsHostManaged) &&
+           False(query_base->flags & QueryFlagBits::IsGuestSynced);
+}
+
+template <typename Traits>
+bool QueryCacheBase<Traits>::SemiFlushQueryDirty(QueryCacheBase<Traits>::QueryLocation location) {
+    auto* query_base = impl->ObtainQuery(location);
+    if (!query_base) {
+        return false;
+    }
+    if (True(query_base->flags & QueryFlagBits::IsFinalValueSynced) &&
+        False(query_base->flags & QueryFlagBits::IsGuestSynced)) {
+        auto* ptr = impl->cpu_memory.GetPointer(query_base->guest_address);
+        if (True(query_base->flags & QueryFlagBits::HasTimestamp)) {
+            std::memcpy(ptr, &query_base->value, sizeof(query_base->value));
+            return false;
+        }
+        u32 value_l = static_cast<u32>(query_base->value);
+        std::memcpy(ptr, &value_l, sizeof(value_l));
+        return false;
+    }
+    return True(query_base->flags & QueryFlagBits::IsHostManaged) &&
+           False(query_base->flags & QueryFlagBits::IsGuestSynced);
+}
+
+template <typename Traits>
+void QueryCacheBase<Traits>::RequestGuestHostSync() {
+    impl->rasterizer.ReleaseFences();
+}
+
+} // namespace VideoCommon

src/video_core/query_cache/query_cache_base.h

@@ -0,0 +1,181 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <functional>
+#include <mutex>
+#include <optional>
+#include <span>
+#include <unordered_map>
+#include <utility>
+
+#include "common/assert.h"
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "core/memory.h"
+#include "video_core/control/channel_state_cache.h"
+#include "video_core/query_cache/query_base.h"
+#include "video_core/query_cache/types.h"
+
+namespace Core::Memory {
+class Memory;
+}
+
+namespace VideoCore {
+class RasterizerInterface;
+}
+
+namespace Tegra {
+class GPU;
+}
+
+namespace VideoCommon {
+
+struct LookupData {
+    VAddr address;
+    QueryBase* found_query;
+};
+
+template <typename Traits>
+class QueryCacheBase : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo> {
+    using RuntimeType = typename Traits::RuntimeType;
+
+public:
+    union QueryLocation {
+        BitField<27, 5, u32> stream_id;
+        BitField<0, 27, u32> query_id;
+        u32 raw;
+
+        std::pair<size_t, size_t> unpack() const {
+            return {static_cast<size_t>(stream_id.Value()), static_cast<size_t>(query_id.Value())};
+        }
+    };
+
+    explicit QueryCacheBase(Tegra::GPU& gpu, VideoCore::RasterizerInterface& rasterizer_,
+                            Core::Memory::Memory& cpu_memory_, RuntimeType& runtime_);
+
+    ~QueryCacheBase();
+
+    void InvalidateRegion(VAddr addr, std::size_t size) {
+        IterateCache<true>(addr, size,
+                           [this](QueryLocation location) { InvalidateQuery(location); });
+    }
+
+    void FlushRegion(VAddr addr, std::size_t size) {
+        bool result = false;
+        IterateCache<false>(addr, size, [this, &result](QueryLocation location) {
+            result |= SemiFlushQueryDirty(location);
+            return result;
+        });
+        if (result) {
+            RequestGuestHostSync();
+        }
+    }
+
+    static u64 BuildMask(std::span<const QueryType> types) {
+        u64 mask = 0;
+        for (auto query_type : types) {
+            mask |= 1ULL << (static_cast<u64>(query_type));
+        }
+        return mask;
+    }
+
+    /// Return true when a CPU region is modified from the GPU
+    [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size) {
+        bool result = false;
+        IterateCache<false>(addr, size, [this, &result](QueryLocation location) {
+            result |= IsQueryDirty(location);
+            return result;
+        });
+        return result;
+    }
+
+    void CounterEnable(QueryType counter_type, bool is_enabled);
+
+    void CounterReset(QueryType counter_type);
+
+    void CounterClose(QueryType counter_type);
+
+    void CounterReport(GPUVAddr addr, QueryType counter_type, QueryPropertiesFlags flags,
+                       u32 payload, u32 subreport);
+
+    void NotifyWFI();
+
+    bool AccelerateHostConditionalRendering();
+
+    // Async downloads
+    void CommitAsyncFlushes();
+
+    bool HasUncommittedFlushes() const;
+
+    bool ShouldWaitAsyncFlushes();
+
+    void PopAsyncFlushes();
+
+    void NotifySegment(bool resume);
+
+    void BindToChannel(s32 id) override;
+
+protected:
+    template <bool remove_from_cache, typename Func>
+    void IterateCache(VAddr addr, std::size_t size, Func&& func) {
+        static constexpr bool RETURNS_BOOL =
+            std::is_same_v<std::invoke_result<Func, QueryLocation>, bool>;
+        const u64 addr_begin = addr;
+        const u64 addr_end = addr_begin + size;
+
+        const u64 page_end = addr_end >> Core::Memory::YUZU_PAGEBITS;
+        std::scoped_lock lock(cache_mutex);
+        for (u64 page = addr_begin >> Core::Memory::YUZU_PAGEBITS; page <= page_end; ++page) {
+            const u64 page_start = page << Core::Memory::YUZU_PAGEBITS;
+            const auto in_range = [page_start, addr_begin, addr_end](const u32 query_location) {
+                const u64 cache_begin = page_start + query_location;
+                const u64 cache_end = cache_begin + sizeof(u32);
+                return cache_begin < addr_end && addr_begin < cache_end;
+            };
+            const auto& it = cached_queries.find(page);
+            if (it == std::end(cached_queries)) {
+                continue;
+            }
+            auto& contents = it->second;
+            for (auto& query : contents) {
+                if (!in_range(query.first)) {
+                    continue;
+                }
+                if constexpr (RETURNS_BOOL) {
+                    if (func(query.second)) {
+                        return;
+                    }
+                } else {
+                    func(query.second);
+                }
+            }
+            if constexpr (remove_from_cache) {
+                const auto in_range2 = [&](const std::pair<u32, QueryLocation>& pair) {
+                    return in_range(pair.first);
+                };
+                std::erase_if(contents, in_range2);
+            }
+        }
+    }
+
+    using ContentCache = std::unordered_map<u64, std::unordered_map<u32, QueryLocation>>;
+
+    void InvalidateQuery(QueryLocation location);
+    bool IsQueryDirty(QueryLocation location);
+    bool SemiFlushQueryDirty(QueryLocation location);
+    void RequestGuestHostSync();
+    void UnregisterPending();
+
+    std::unordered_map<u64, std::unordered_map<u32, QueryLocation>> cached_queries;
+    std::mutex cache_mutex;
+
+    struct QueryCacheBaseImpl;
+    friend struct QueryCacheBaseImpl;
+    friend RuntimeType;
+
+    std::unique_ptr<QueryCacheBaseImpl> impl;
+};
+
+} // namespace VideoCommon

src/video_core/query_cache/query_stream.h

@@ -0,0 +1,149 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <deque>
+#include <optional>
+#include <vector>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "video_core/query_cache/bank_base.h"
+#include "video_core/query_cache/query_base.h"
+
+namespace VideoCommon {
+
+class StreamerInterface {
+public:
+    explicit StreamerInterface(size_t id_) : id{id_}, dependence_mask{}, dependent_mask{} {}
+    virtual ~StreamerInterface() = default;
+
+    virtual QueryBase* GetQuery(size_t id) = 0;
+
+    virtual void StartCounter() {
+        /* Do Nothing */
+    }
+
+    virtual void PauseCounter() {
+        /* Do Nothing */
+    }
+
+    virtual void ResetCounter() {
+        /* Do Nothing */
+    }
+
+    virtual void CloseCounter() {
+        /* Do Nothing */
+    }
+
+    virtual bool HasPendingSync() const {
+        return false;
+    }
+
+    virtual void PresyncWrites() {
+        /* Do Nothing */
+    }
+
+    virtual void SyncWrites() {
+        /* Do Nothing */
+    }
+
+    virtual size_t WriteCounter(VAddr address, bool has_timestamp, u32 value,
+                                std::optional<u32> subreport = std::nullopt) = 0;
+
+    virtual bool HasUnsyncedQueries() const {
+        return false;
+    }
+
+    virtual void PushUnsyncedQueries() {
+        /* Do Nothing */
+    }
+
+    virtual void PopUnsyncedQueries() {
+        /* Do Nothing */
+    }
+
+    virtual void Free(size_t query_id) = 0;
+
+    size_t GetId() const {
+        return id;
+    }
+
+    u64 GetDependenceMask() const {
+        return dependence_mask;
+    }
+
+    u64 GetDependentMask() const {
+        return dependence_mask;
+    }
+
+    u64 GetAmmendValue() const {
+        return ammend_value;
+    }
+
+    void SetAccumulationValue(u64 new_value) {
+        acumulation_value = new_value;
+    }
+
+protected:
+    void MakeDependent(StreamerInterface* depend_on) {
+        dependence_mask |= 1ULL << depend_on->id;
+        depend_on->dependent_mask |= 1ULL << id;
+    }
+
+    const size_t id;
+    u64 dependence_mask;
+    u64 dependent_mask;
+    u64 ammend_value{};
+    u64 acumulation_value{};
+};
+
+template <typename QueryType>
+class SimpleStreamer : public StreamerInterface {
+public:
+    explicit SimpleStreamer(size_t id_) : StreamerInterface{id_} {}
+    virtual ~SimpleStreamer() = default;
+
+protected:
+    virtual QueryType* GetQuery(size_t query_id) override {
+        if (query_id < slot_queries.size()) {
+            return &slot_queries[query_id];
+        }
+        return nullptr;
+    }
+
+    virtual void Free(size_t query_id) override {
+        std::scoped_lock lk(guard);
+        ReleaseQuery(query_id);
+    }
+
+    template <typename... Args, typename = decltype(QueryType(std::declval<Args>()...))>
+    size_t BuildQuery(Args&&... args) {
+        std::scoped_lock lk(guard);
+        if (!old_queries.empty()) {
+            size_t new_id = old_queries.front();
+            old_queries.pop_front();
+            new (&slot_queries[new_id]) QueryType(std::forward<Args>(args)...);
+            return new_id;
+        }
+        size_t new_id = slot_queries.size();
+        slot_queries.emplace_back(std::forward<Args>(args)...);
+        return new_id;
+    }
+
+    void ReleaseQuery(size_t query_id) {
+
+        if (query_id < slot_queries.size()) {
+            old_queries.push_back(query_id);
+            return;
+        }
+        UNREACHABLE();
+    }
+
+    std::mutex guard;
+    std::deque<QueryType> slot_queries;
+    std::deque<size_t> old_queries;
+};
+
+} // namespace VideoCommon

src/video_core/query_cache/types.h

@@ -0,0 +1,74 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/common_funcs.h"
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+enum class QueryPropertiesFlags : u32 {
+    HasTimeout = 1 << 0,
+    IsAFence = 1 << 1,
+};
+DECLARE_ENUM_FLAG_OPERATORS(QueryPropertiesFlags)
+
+// This should always be equivalent to maxwell3d Report Semaphore Reports
+enum class QueryType : u32 {
+    Payload = 0, // "None" in docs, but confirmed via hardware to return the payload
+    VerticesGenerated = 1,
+    ZPassPixelCount = 2,
+    PrimitivesGenerated = 3,
+    AlphaBetaClocks = 4,
+    VertexShaderInvocations = 5,
+    StreamingPrimitivesNeededMinusSucceeded = 6,
+    GeometryShaderInvocations = 7,
+    GeometryShaderPrimitivesGenerated = 9,
+    ZCullStats0 = 10,
+    StreamingPrimitivesSucceeded = 11,
+    ZCullStats1 = 12,
+    StreamingPrimitivesNeeded = 13,
+    ZCullStats2 = 14,
+    ClipperInvocations = 15,
+    ZCullStats3 = 16,
+    ClipperPrimitivesGenerated = 17,
+    VtgPrimitivesOut = 18,
+    PixelShaderInvocations = 19,
+    ZPassPixelCount64 = 21,
+    IEEECleanColorTarget = 24,
+    IEEECleanZetaTarget = 25,
+    StreamingByteCount = 26,
+    TessellationInitInvocations = 27,
+    BoundingRectangle = 28,
+    TessellationShaderInvocations = 29,
+    TotalStreamingPrimitivesNeededMinusSucceeded = 30,
+    TessellationShaderPrimitivesGenerated = 31,
+    // max.
+    MaxQueryTypes,
+};
+
+// Comparison modes for Host Conditional Rendering
+enum class ComparisonMode : u32 {
+    False = 0,
+    True = 1,
+    Conditional = 2,
+    IfEqual = 3,
+    IfNotEqual = 4,
+    MaxComparisonMode,
+};
+
+// Reduction ops.
+enum class ReductionOp : u32 {
+    RedAdd = 0,
+    RedMin = 1,
+    RedMax = 2,
+    RedInc = 3,
+    RedDec = 4,
+    RedAnd = 5,
+    RedOr = 6,
+    RedXor = 7,
+    MaxReductionOp,
+};
+
+} // namespace VideoCommon

src/video_core/rasterizer_interface.h

@@ -12,6 +12,7 @@
 #include "video_core/cache_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
+#include "video_core/query_cache/types.h"
 #include "video_core/rasterizer_download_area.h"
 
 namespace Tegra {
@@ -26,11 +27,6 @@ struct ChannelState;
 
 namespace VideoCore {
 
-enum class QueryType {
-    SamplesPassed,
-};
-constexpr std::size_t NumQueryTypes = 1;
-
 enum class LoadCallbackStage {
     Prepare,
     Build,
@@ -58,10 +54,11 @@ public:
     virtual void DispatchCompute() = 0;
 
     /// Resets the counter of a query
-    virtual void ResetCounter(QueryType type) = 0;
+    virtual void ResetCounter(VideoCommon::QueryType type) = 0;
 
     /// Records a GPU query and caches it
-    virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
+    virtual void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+                       VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) = 0;
 
     /// Signal an uniform buffer binding
     virtual void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
@@ -83,7 +80,7 @@ public:
     virtual void SignalReference() = 0;
 
     /// Release all pending fences.
-    virtual void ReleaseFences() = 0;
+    virtual void ReleaseFences(bool force = true) = 0;
 
     /// Notify rasterizer that all caches should be flushed to Switch memory
     virtual void FlushAll() = 0;

src/video_core/renderer_null/null_rasterizer.cpp

@@ -26,16 +26,18 @@ void RasterizerNull::Draw(bool is_indexed, u32 instance_count) {}
 void RasterizerNull::DrawTexture() {}
 void RasterizerNull::Clear(u32 layer_count) {}
 void RasterizerNull::DispatchCompute() {}
-void RasterizerNull::ResetCounter(VideoCore::QueryType type) {}
+void RasterizerNull::ResetCounter(VideoCommon::QueryType type) {}
-void RasterizerNull::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
+void RasterizerNull::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
-                           std::optional<u64> timestamp) {
+                           VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
     if (!gpu_memory) {
         return;
     }
-
+    if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
-    gpu_memory->Write(gpu_addr, u64{0});
+        u64 ticks = m_gpu.GetTicks();
-    if (timestamp) {
+        gpu_memory->Write<u64>(gpu_addr + 8, ticks);
-        gpu_memory->Write(gpu_addr + 8, *timestamp);
+        gpu_memory->Write<u64>(gpu_addr, static_cast<u64>(payload));
+    } else {
+        gpu_memory->Write<u32>(gpu_addr, payload);
     }
 }
 void RasterizerNull::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
@@ -74,7 +76,7 @@ void RasterizerNull::SignalSyncPoint(u32 value) {
     syncpoint_manager.IncrementHost(value);
 }
 void RasterizerNull::SignalReference() {}
-void RasterizerNull::ReleaseFences() {}
+void RasterizerNull::ReleaseFences(bool) {}
 void RasterizerNull::FlushAndInvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType) {}
 void RasterizerNull::WaitForIdle() {}
 void RasterizerNull::FragmentBarrier() {}

src/video_core/renderer_null/null_rasterizer.h

@@ -42,8 +42,9 @@ public:
     void DrawTexture() override;
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
-    void ResetCounter(VideoCore::QueryType type) override;
+    void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
@@ -63,7 +64,7 @@ public:
     void SyncOperation(std::function<void()>&& func) override;
     void SignalSyncPoint(u32 value) override;
     void SignalReference() override;
-    void ReleaseFences() override;
+    void ReleaseFences(bool force) override;
     void FlushAndInvalidateRegion(
         VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void WaitForIdle() override;

src/video_core/renderer_opengl/gl_query_cache.cpp

@@ -27,7 +27,7 @@ constexpr GLenum GetTarget(VideoCore::QueryType type) {
 } // Anonymous namespace
 
 QueryCache::QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_)
-    : QueryCacheBase(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
+    : QueryCacheLegacy(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
 
 QueryCache::~QueryCache() = default;
 

src/video_core/renderer_opengl/gl_query_cache.h

@@ -26,7 +26,7 @@ class RasterizerOpenGL;
 using CounterStream = VideoCommon::CounterStreamBase<QueryCache, HostCounter>;
 
 class QueryCache final
-    : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
+    : public VideoCommon::QueryCacheLegacy<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
     explicit QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_);
     ~QueryCache();

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -396,13 +396,39 @@ void RasterizerOpenGL::DispatchCompute() {
     has_written_global_memory |= pipeline->WritesGlobalMemory();
 }
 
-void RasterizerOpenGL::ResetCounter(VideoCore::QueryType type) {
+void RasterizerOpenGL::ResetCounter(VideoCommon::QueryType type) {
-    query_cache.ResetCounter(type);
+    if (type == VideoCommon::QueryType::ZPassPixelCount64) {
+        query_cache.ResetCounter(VideoCore::QueryType::SamplesPassed);
+    }
 }
 
-void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
+void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
-                             std::optional<u64> timestamp) {
+                             VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
-    query_cache.Query(gpu_addr, type, timestamp);
+    if (type == VideoCommon::QueryType::ZPassPixelCount64) {
+        if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
+            query_cache.Query(gpu_addr, VideoCore::QueryType::SamplesPassed, {gpu.GetTicks()});
+        } else {
+            query_cache.Query(gpu_addr, VideoCore::QueryType::SamplesPassed, std::nullopt);
+        }
+        return;
+    }
+    if (type != VideoCommon::QueryType::Payload) {
+        payload = 1u;
+    }
+    std::function<void()> func([this, gpu_addr, flags, memory_manager = gpu_memory, payload]() {
+        if (True(flags & VideoCommon::QueryPropertiesFlags::HasTimeout)) {
+            u64 ticks = gpu.GetTicks();
+            memory_manager->Write<u64>(gpu_addr + 8, ticks);
+            memory_manager->Write<u64>(gpu_addr, static_cast<u64>(payload));
+        } else {
+            memory_manager->Write<u32>(gpu_addr, payload);
+        }
+    });
+    if (True(flags & VideoCommon::QueryPropertiesFlags::IsAFence)) {
+        SignalFence(std::move(func));
+        return;
+    }
+    func();
 }
 
 void RasterizerOpenGL::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
@@ -573,8 +599,8 @@ void RasterizerOpenGL::SignalReference() {
     fence_manager.SignalOrdering();
 }
 
-void RasterizerOpenGL::ReleaseFences() {
+void RasterizerOpenGL::ReleaseFences(bool force) {
-    fence_manager.WaitPendingFences();
+    fence_manager.WaitPendingFences(force);
 }
 
 void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size,

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -86,8 +86,9 @@ public:
     void DrawTexture() override;
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
-    void ResetCounter(VideoCore::QueryType type) override;
+    void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
@@ -107,7 +108,7 @@ public:
     void SyncOperation(std::function<void()>&& func) override;
     void SignalSyncPoint(u32 value) override;
     void SignalReference() override;
-    void ReleaseFences() override;
+    void ReleaseFences(bool force = true) override;
     void FlushAndInvalidateRegion(
         VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void WaitForIdle() override;

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -61,6 +61,9 @@ vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allo
     if (device.IsExtTransformFeedbackSupported()) {
         flags |= VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
     }
+    if (device.IsExtConditionalRendering()) {
+        flags |= VK_BUFFER_USAGE_CONDITIONAL_RENDERING_BIT_EXT;
+    }
     const VkBufferCreateInfo buffer_ci = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .pNext = nullptr,

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -12,6 +12,9 @@
 #include "common/common_types.h"
 #include "common/div_ceil.h"
 #include "video_core/host_shaders/astc_decoder_comp_spv.h"
+#include "video_core/host_shaders/queries_prefix_scan_sum_comp_spv.h"
+#include "video_core/host_shaders/queries_prefix_scan_sum_nosubgroups_comp_spv.h"
+#include "video_core/host_shaders/resolve_conditional_render_comp_spv.h"
 #include "video_core/host_shaders/vulkan_quad_indexed_comp_spv.h"
 #include "video_core/host_shaders/vulkan_uint8_comp_spv.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
@@ -57,6 +60,30 @@ constexpr std::array<VkDescriptorSetLayoutBinding, 2> INPUT_OUTPUT_DESCRIPTOR_SE
     },
 }};
 
+constexpr std::array<VkDescriptorSetLayoutBinding, 3> QUERIES_SCAN_DESCRIPTOR_SET_BINDINGS{{
+    {
+        .binding = 0,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    },
+    {
+        .binding = 1,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    },
+    {
+        .binding = 2,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    },
+}};
+
 constexpr DescriptorBankInfo INPUT_OUTPUT_BANK_INFO{
     .uniform_buffers = 0,
     .storage_buffers = 2,
@@ -67,6 +94,16 @@ constexpr DescriptorBankInfo INPUT_OUTPUT_BANK_INFO{
     .score = 2,
 };
 
+constexpr DescriptorBankInfo QUERIES_SCAN_BANK_INFO{
+    .uniform_buffers = 0,
+    .storage_buffers = 3,
+    .texture_buffers = 0,
+    .image_buffers = 0,
+    .textures = 0,
+    .images = 0,
+    .score = 3,
+};
+
 constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> ASTC_DESCRIPTOR_SET_BINDINGS{{
     {
         .binding = ASTC_BINDING_INPUT_BUFFER,
@@ -103,6 +140,15 @@ constexpr VkDescriptorUpdateTemplateEntry INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLAT
     .stride = sizeof(DescriptorUpdateEntry),
 };
 
+constexpr VkDescriptorUpdateTemplateEntry QUERIES_SCAN_DESCRIPTOR_UPDATE_TEMPLATE{
+    .dstBinding = 0,
+    .dstArrayElement = 0,
+    .descriptorCount = 3,
+    .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+    .offset = 0,
+    .stride = sizeof(DescriptorUpdateEntry),
+};
+
 constexpr std::array<VkDescriptorUpdateTemplateEntry, ASTC_NUM_BINDINGS>
     ASTC_PASS_DESCRIPTOR_UPDATE_TEMPLATE_ENTRY{{
         {
@@ -131,13 +177,21 @@ struct AstcPushConstants {
     u32 block_height;
     u32 block_height_mask;
 };
+
+struct QueriesPrefixScanPushConstants {
+    u32 min_accumulation_base;
+    u32 max_accumulation_base;
+    u32 accumulation_limit;
+    u32 buffer_offset;
+};
 } // Anonymous namespace
 
 ComputePass::ComputePass(const Device& device_, DescriptorPool& descriptor_pool,
                          vk::Span<VkDescriptorSetLayoutBinding> bindings,
                          vk::Span<VkDescriptorUpdateTemplateEntry> templates,
                          const DescriptorBankInfo& bank_info,
-                         vk::Span<VkPushConstantRange> push_constants, std::span<const u32> code)
+                         vk::Span<VkPushConstantRange> push_constants, std::span<const u32> code,
+                         std::optional<u32> optional_subgroup_size)
     : device{device_} {
     descriptor_set_layout = device.GetLogical().CreateDescriptorSetLayout({
         .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
@@ -178,13 +232,19 @@ ComputePass::ComputePass(const Device& device_, DescriptorPool& descriptor_pool,
         .pCode = code.data(),
     });
     device.SaveShader(code);
+    const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT subgroup_size_ci{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT,
+        .pNext = nullptr,
+        .requiredSubgroupSize = optional_subgroup_size ? *optional_subgroup_size : 32U,
+    };
+    bool use_setup_size = device.IsExtSubgroupSizeControlSupported() && optional_subgroup_size;
     pipeline = device.GetLogical().CreateComputePipeline({
         .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
         .stage{
             .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
-            .pNext = nullptr,
+            .pNext = use_setup_size ? &subgroup_size_ci : nullptr,
             .flags = 0,
             .stage = VK_SHADER_STAGE_COMPUTE_BIT,
             .module = *module,
@@ -302,6 +362,123 @@ std::pair<VkBuffer, VkDeviceSize> QuadIndexedPass::Assemble(
     return {staging.buffer, staging.offset};
 }
 
+ConditionalRenderingResolvePass::ConditionalRenderingResolvePass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, descriptor_pool_, INPUT_OUTPUT_DESCRIPTOR_SET_BINDINGS,
+                  INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE, INPUT_OUTPUT_BANK_INFO, nullptr,
+                  RESOLVE_CONDITIONAL_RENDER_COMP_SPV),
+      scheduler{scheduler_}, compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+void ConditionalRenderingResolvePass::Resolve(VkBuffer dst_buffer, VkBuffer src_buffer,
+                                              u32 src_offset, bool compare_to_zero) {
+    const size_t compare_size = compare_to_zero ? 8 : 24;
+
+    compute_pass_descriptor_queue.Acquire();
+    compute_pass_descriptor_queue.AddBuffer(src_buffer, src_offset, compare_size);
+    compute_pass_descriptor_queue.AddBuffer(dst_buffer, 0, sizeof(u32));
+    const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([this, descriptor_data](vk::CommandBuffer cmdbuf) {
+        static constexpr VkMemoryBarrier read_barrier{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+        };
+        static constexpr VkMemoryBarrier write_barrier{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT,
+        };
+        const VkDescriptorSet set = descriptor_allocator.Commit();
+        device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, read_barrier);
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
+        cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+        cmdbuf.Dispatch(1, 1, 1);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                               VK_PIPELINE_STAGE_CONDITIONAL_RENDERING_BIT_EXT, 0, write_barrier);
+    });
+}
+
+QueriesPrefixScanPass::QueriesPrefixScanPass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(
+          device_, descriptor_pool_, QUERIES_SCAN_DESCRIPTOR_SET_BINDINGS,
+          QUERIES_SCAN_DESCRIPTOR_UPDATE_TEMPLATE, QUERIES_SCAN_BANK_INFO,
+          COMPUTE_PUSH_CONSTANT_RANGE<sizeof(QueriesPrefixScanPushConstants)>,
+          device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_BASIC_BIT) &&
+                  device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_ARITHMETIC_BIT) &&
+                  device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_SHUFFLE_BIT) &&
+                  device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT)
+              ? std::span<const u32>(QUERIES_PREFIX_SCAN_SUM_COMP_SPV)
+              : std::span<const u32>(QUERIES_PREFIX_SCAN_SUM_NOSUBGROUPS_COMP_SPV)),
+      scheduler{scheduler_}, compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+void QueriesPrefixScanPass::Run(VkBuffer accumulation_buffer, VkBuffer dst_buffer,
+                                VkBuffer src_buffer, size_t number_of_sums,
+                                size_t min_accumulation_limit, size_t max_accumulation_limit) {
+    size_t current_runs = number_of_sums;
+    size_t offset = 0;
+    while (current_runs != 0) {
+        static constexpr size_t DISPATCH_SIZE = 2048U;
+        size_t runs_to_do = std::min<size_t>(current_runs, DISPATCH_SIZE);
+        current_runs -= runs_to_do;
+        compute_pass_descriptor_queue.Acquire();
+        compute_pass_descriptor_queue.AddBuffer(src_buffer, 0, number_of_sums * sizeof(u64));
+        compute_pass_descriptor_queue.AddBuffer(dst_buffer, 0, number_of_sums * sizeof(u64));
+        compute_pass_descriptor_queue.AddBuffer(accumulation_buffer, 0, sizeof(u64));
+        const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+        size_t used_offset = offset;
+        offset += runs_to_do;
+
+        scheduler.RequestOutsideRenderPassOperationContext();
+        scheduler.Record([this, descriptor_data, min_accumulation_limit, max_accumulation_limit,
+                          runs_to_do, used_offset](vk::CommandBuffer cmdbuf) {
+            static constexpr VkMemoryBarrier read_barrier{
+                .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            };
+            static constexpr VkMemoryBarrier write_barrier{
+                .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT |
+                                 VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT |
+                                 VK_ACCESS_INDIRECT_COMMAND_READ_BIT | VK_ACCESS_INDEX_READ_BIT |
+                                 VK_ACCESS_UNIFORM_READ_BIT |
+                                 VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT,
+            };
+            const QueriesPrefixScanPushConstants uniforms{
+                .min_accumulation_base = static_cast<u32>(min_accumulation_limit),
+                .max_accumulation_base = static_cast<u32>(max_accumulation_limit),
+                .accumulation_limit = static_cast<u32>(runs_to_do - 1),
+                .buffer_offset = static_cast<u32>(used_offset),
+            };
+            const VkDescriptorSet set = descriptor_allocator.Commit();
+            device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                                   VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, read_barrier);
+            cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+            cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, uniforms);
+            cmdbuf.Dispatch(1, 1, 1);
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                                   VK_PIPELINE_STAGE_CONDITIONAL_RENDERING_BIT_EXT, 0,
+                                   write_barrier);
+        });
+    }
+}
+
 ASTCDecoderPass::ASTCDecoderPass(const Device& device_, Scheduler& scheduler_,
                                  DescriptorPool& descriptor_pool_,
                                  StagingBufferPool& staging_buffer_pool_,

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include <optional>
 #include <span>
 #include <utility>
 
@@ -31,7 +32,8 @@ public:
                          vk::Span<VkDescriptorSetLayoutBinding> bindings,
                          vk::Span<VkDescriptorUpdateTemplateEntry> templates,
                          const DescriptorBankInfo& bank_info,
-                         vk::Span<VkPushConstantRange> push_constants, std::span<const u32> code);
+                         vk::Span<VkPushConstantRange> push_constants, std::span<const u32> code,
+                         std::optional<u32> optional_subgroup_size = std::nullopt);
     ~ComputePass();
 
 protected:
@@ -82,6 +84,33 @@ private:
     ComputePassDescriptorQueue& compute_pass_descriptor_queue;
 };
 
+class ConditionalRenderingResolvePass final : public ComputePass {
+public:
+    explicit ConditionalRenderingResolvePass(
+        const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+        ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+
+    void Resolve(VkBuffer dst_buffer, VkBuffer src_buffer, u32 src_offset, bool compare_to_zero);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
+class QueriesPrefixScanPass final : public ComputePass {
+public:
+    explicit QueriesPrefixScanPass(const Device& device_, Scheduler& scheduler_,
+                                   DescriptorPool& descriptor_pool_,
+                                   ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+
+    void Run(VkBuffer accumulation_buffer, VkBuffer dst_buffer, VkBuffer src_buffer,
+             size_t number_of_sums, size_t min_accumulation_limit, size_t max_accumulation_limit);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
 class ASTCDecoderPass final : public ComputePass {
 public:
     explicit ASTCDecoderPass(const Device& device_, Scheduler& scheduler_,

src/video_core/renderer_vulkan/vk_fence_manager.h

@@ -7,6 +7,7 @@
 
 #include "video_core/fence_manager.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 
 namespace Core {
@@ -20,7 +21,6 @@ class RasterizerInterface;
 namespace Vulkan {
 
 class Device;
-class QueryCache;
 class Scheduler;
 
 class InnerFence : public VideoCommon::FenceBase {

src/video_core/renderer_vulkan/vk_query_cache.h

@@ -1,101 +1,75 @@
-// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
+// SPDX-License-Identifier: GPL-3.0-or-later
 
 #pragma once
 
-#include <cstddef>
 #include <memory>
-#include <utility>
-#include <vector>
 
-#include "common/common_types.h"
+#include "video_core/query_cache/query_cache_base.h"
-#include "video_core/query_cache.h"
+#include "video_core/renderer_vulkan/vk_buffer_cache.h"
-#include "video_core/renderer_vulkan/vk_resource_pool.h"
-#include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace VideoCore {
 class RasterizerInterface;
 }
 
+namespace VideoCommon {
+class StreamerInterface;
+}
+
 namespace Vulkan {
 
-class CachedQuery;
 class Device;
-class HostCounter;
-class QueryCache;
 class Scheduler;
+class StagingBufferPool;
 
-using CounterStream = VideoCommon::CounterStreamBase<QueryCache, HostCounter>;
+struct QueryCacheRuntimeImpl;
 
-class QueryPool final : public ResourcePool {
+class QueryCacheRuntime {
 public:
-    explicit QueryPool(const Device& device, Scheduler& scheduler, VideoCore::QueryType type);
+    explicit QueryCacheRuntime(VideoCore::RasterizerInterface* rasterizer,
-    ~QueryPool() override;
+                               Core::Memory::Memory& cpu_memory_,
+                               Vulkan::BufferCache& buffer_cache_, const Device& device_,
+                               const MemoryAllocator& memory_allocator_, Scheduler& scheduler_,
+                               StagingBufferPool& staging_pool_,
+                               ComputePassDescriptorQueue& compute_pass_descriptor_queue,
+                               DescriptorPool& descriptor_pool);
+    ~QueryCacheRuntime();
 
-    std::pair<VkQueryPool, u32> Commit();
+    template <typename SyncValuesType>
+    void SyncValues(std::span<SyncValuesType> values, VkBuffer base_src_buffer = nullptr);
 
-    void Reserve(std::pair<VkQueryPool, u32> query);
+    void Barriers(bool is_prebarrier);
 
-protected:
+    void EndHostConditionalRendering();
-    void Allocate(std::size_t begin, std::size_t end) override;
+
+    void PauseHostConditionalRendering();
+
+    void ResumeHostConditionalRendering();
+
+    bool HostConditionalRenderingCompareValue(VideoCommon::LookupData object_1, bool qc_dirty);
+
+    bool HostConditionalRenderingCompareValues(VideoCommon::LookupData object_1,
+                                               VideoCommon::LookupData object_2, bool qc_dirty,
+                                               bool equal_check);
+
+    VideoCommon::StreamerInterface* GetStreamerInterface(VideoCommon::QueryType query_type);
+
+    void Bind3DEngine(Tegra::Engines::Maxwell3D* maxwell3d);
+
+    template <typename Func>
+    void View3DRegs(Func&& func);
 
 private:
-    static constexpr std::size_t GROW_STEP = 512;
+    void HostConditionalRenderingCompareValueImpl(VideoCommon::LookupData object, bool is_equal);
-
+    void HostConditionalRenderingCompareBCImpl(VAddr address, bool is_equal);
-    const Device& device;
+    friend struct QueryCacheRuntimeImpl;
-    const VideoCore::QueryType type;
+    std::unique_ptr<QueryCacheRuntimeImpl> impl;
-
-    std::vector<vk::QueryPool> pools;
-    std::vector<bool> usage;
 };
 
-class QueryCache final
+struct QueryCacheParams {
-    : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
+    using RuntimeType = typename Vulkan::QueryCacheRuntime;
-public:
-    explicit QueryCache(VideoCore::RasterizerInterface& rasterizer_,
-                        Core::Memory::Memory& cpu_memory_, const Device& device_,
-                        Scheduler& scheduler_);
-    ~QueryCache();
-
-    std::pair<VkQueryPool, u32> AllocateQuery(VideoCore::QueryType type);
-
-    void Reserve(VideoCore::QueryType type, std::pair<VkQueryPool, u32> query);
-
-    const Device& GetDevice() const noexcept {
-        return device;
-    }
-
-    Scheduler& GetScheduler() const noexcept {
-        return scheduler;
-    }
-
-private:
-    const Device& device;
-    Scheduler& scheduler;
-    std::array<QueryPool, VideoCore::NumQueryTypes> query_pools;
 };
 
-class HostCounter final : public VideoCommon::HostCounterBase<QueryCache, HostCounter> {
+using QueryCache = VideoCommon::QueryCacheBase<QueryCacheParams>;
-public:
-    explicit HostCounter(QueryCache& cache_, std::shared_ptr<HostCounter> dependency_,
-                         VideoCore::QueryType type_);
-    ~HostCounter();
-
-    void EndQuery();
-
-private:
-    u64 BlockingQuery(bool async = false) const override;
-
-    QueryCache& cache;
-    const VideoCore::QueryType type;
-    const std::pair<VkQueryPool, u32> query;
-    const u64 tick;
-};
-
-class CachedQuery : public VideoCommon::CachedQueryBase<HostCounter> {
-public:
-    explicit CachedQuery(QueryCache&, VideoCore::QueryType, VAddr cpu_addr_, u8* host_ptr_)
-        : CachedQueryBase{cpu_addr_, host_ptr_} {}
-};
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -24,6 +24,7 @@
 #include "video_core/renderer_vulkan/vk_compute_pipeline.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
+#include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/renderer_vulkan/vk_rasterizer.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
@@ -170,9 +171,11 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache_runtime(device, memory_allocator, scheduler, staging_pool,
                            guest_descriptor_queue, compute_pass_descriptor_queue, descriptor_pool),
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
+      query_cache_runtime(this, cpu_memory_, buffer_cache, device, memory_allocator, scheduler,
+                          staging_pool, compute_pass_descriptor_queue, descriptor_pool),
+      query_cache(gpu, *this, cpu_memory_, query_cache_runtime),
       pipeline_cache(*this, device, scheduler, descriptor_pool, guest_descriptor_queue,
                      render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
-      query_cache{*this, cpu_memory_, device, scheduler},
       accelerate_dma(buffer_cache, texture_cache, scheduler),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
@@ -189,14 +192,7 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     FlushWork();
     gpu_memory->FlushCaching();
 
-#if ANDROID
+    query_cache.NotifySegment(true);
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        query_cache.UpdateCounters();
-    }
-#else
-    query_cache.UpdateCounters();
-#endif
 
     GraphicsPipeline* const pipeline{pipeline_cache.CurrentGraphicsPipeline()};
     if (!pipeline) {
@@ -207,13 +203,12 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     pipeline->SetEngine(maxwell3d, gpu_memory);
     pipeline->Configure(is_indexed);
 
-    BeginTransformFeedback();
-
     UpdateDynamicStates();
 
+    HandleTransformFeedback();
+    query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
+                              maxwell3d->regs.zpass_pixel_count_enable);
     draw_func();
-
-    EndTransformFeedback();
 }
 
 void RasterizerVulkan::Draw(bool is_indexed, u32 instance_count) {
@@ -241,6 +236,14 @@ void RasterizerVulkan::DrawIndirect() {
         const auto indirect_buffer = buffer_cache.GetDrawIndirectBuffer();
         const auto& buffer = indirect_buffer.first;
         const auto& offset = indirect_buffer.second;
+        if (params.is_byte_count) {
+            scheduler.Record([buffer_obj = buffer->Handle(), offset,
+                              stride = params.stride](vk::CommandBuffer cmdbuf) {
+                cmdbuf.DrawIndirectByteCountEXT(1, 0, buffer_obj, offset, 0,
+                                                static_cast<u32>(stride));
+            });
+            return;
+        }
         if (params.include_count) {
             const auto count = buffer_cache.GetDrawIndirectCount();
             const auto& draw_buffer = count.first;
@@ -280,20 +283,15 @@ void RasterizerVulkan::DrawTexture() {
     SCOPE_EXIT({ gpu.TickWork(); });
     FlushWork();
 
-#if ANDROID
+    query_cache.NotifySegment(true);
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        query_cache.UpdateCounters();
-    }
-#else
-    query_cache.UpdateCounters();
-#endif
 
     texture_cache.SynchronizeGraphicsDescriptors();
     texture_cache.UpdateRenderTargets(false);
 
     UpdateDynamicStates();
 
+    query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
+                              maxwell3d->regs.zpass_pixel_count_enable);
     const auto& draw_texture_state = maxwell3d->draw_manager->GetDrawTextureState();
     const auto& sampler = texture_cache.GetGraphicsSampler(draw_texture_state.src_sampler);
     const auto& texture = texture_cache.GetImageView(draw_texture_state.src_texture);
@@ -316,14 +314,9 @@ void RasterizerVulkan::Clear(u32 layer_count) {
     FlushWork();
     gpu_memory->FlushCaching();
 
-#if ANDROID
+    query_cache.NotifySegment(true);
-    if (Settings::IsGPULevelHigh()) {
+    query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
-        // This is problematic on Android, disable on GPU Normal.
+                              maxwell3d->regs.zpass_pixel_count_enable);
-        query_cache.UpdateCounters();
-    }
-#else
-    query_cache.UpdateCounters();
-#endif
 
     auto& regs = maxwell3d->regs;
     const bool use_color = regs.clear_surface.R || regs.clear_surface.G || regs.clear_surface.B ||
@@ -482,13 +475,13 @@ void RasterizerVulkan::DispatchCompute() {
     scheduler.Record([dim](vk::CommandBuffer cmdbuf) { cmdbuf.Dispatch(dim[0], dim[1], dim[2]); });
 }
 
-void RasterizerVulkan::ResetCounter(VideoCore::QueryType type) {
+void RasterizerVulkan::ResetCounter(VideoCommon::QueryType type) {
-    query_cache.ResetCounter(type);
+    query_cache.CounterReset(type);
 }
 
-void RasterizerVulkan::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
+void RasterizerVulkan::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
-                             std::optional<u64> timestamp) {
+                             VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
-    query_cache.Query(gpu_addr, type, timestamp);
+    query_cache.CounterReport(gpu_addr, type, flags, payload, subreport);
 }
 
 void RasterizerVulkan::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
@@ -669,8 +662,8 @@ void RasterizerVulkan::SignalReference() {
     fence_manager.SignalReference();
 }
 
-void RasterizerVulkan::ReleaseFences() {
+void RasterizerVulkan::ReleaseFences(bool force) {
-    fence_manager.WaitPendingFences();
+    fence_manager.WaitPendingFences(force);
 }
 
 void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size,
@@ -694,6 +687,8 @@ void RasterizerVulkan::WaitForIdle() {
         flags |= VK_PIPELINE_STAGE_TRANSFORM_FEEDBACK_BIT_EXT;
     }
 
+    query_cache.NotifyWFI();
+
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([event = *wfi_event, flags](vk::CommandBuffer cmdbuf) {
         cmdbuf.SetEvent(event, flags);
@@ -737,19 +732,7 @@ void RasterizerVulkan::TickFrame() {
 
 bool RasterizerVulkan::AccelerateConditionalRendering() {
     gpu_memory->FlushCaching();
-    if (Settings::IsGPULevelHigh()) {
+    return query_cache.AccelerateHostConditionalRendering();
-        // TODO(Blinkhawk): Reimplement Host conditional rendering.
-        return false;
-    }
-    // Medium / Low Hack: stub any checks on queries written into the buffer cache.
-    const GPUVAddr condition_address{maxwell3d->regs.render_enable.Address()};
-    Maxwell::ReportSemaphore::Compare cmp;
-    if (gpu_memory->IsMemoryDirty(condition_address, sizeof(cmp),
-                                  VideoCommon::CacheType::BufferCache |
-                                      VideoCommon::CacheType::QueryCache)) {
-        return true;
-    }
-    return false;
 }
 
 bool RasterizerVulkan::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
@@ -795,6 +778,7 @@ bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config,
     if (!image_view) {
         return false;
     }
+    query_cache.NotifySegment(false);
     screen_info.image = image_view->ImageHandle();
     screen_info.image_view = image_view->Handle(Shader::TextureType::Color2D);
     screen_info.width = image_view->size.width;
@@ -933,31 +917,18 @@ void RasterizerVulkan::UpdateDynamicStates() {
     }
 }
 
-void RasterizerVulkan::BeginTransformFeedback() {
+void RasterizerVulkan::HandleTransformFeedback() {
     const auto& regs = maxwell3d->regs;
-    if (regs.transform_feedback_enabled == 0) {
-        return;
-    }
     if (!device.IsExtTransformFeedbackSupported()) {
         LOG_ERROR(Render_Vulkan, "Transform feedbacks used but not supported");
         return;
     }
+    query_cache.CounterEnable(VideoCommon::QueryType::StreamingByteCount,
+                              regs.transform_feedback_enabled);
+    if (regs.transform_feedback_enabled != 0) {
         UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderType::TessellationInit) ||
                          regs.IsShaderConfigEnabled(Maxwell::ShaderType::Tessellation));
-    scheduler.Record(
-        [](vk::CommandBuffer cmdbuf) { cmdbuf.BeginTransformFeedbackEXT(0, 0, nullptr, nullptr); });
     }
-
-void RasterizerVulkan::EndTransformFeedback() {
-    const auto& regs = maxwell3d->regs;
-    if (regs.transform_feedback_enabled == 0) {
-        return;
-    }
-    if (!device.IsExtTransformFeedbackSupported()) {
-        return;
-    }
-    scheduler.Record(
-        [](vk::CommandBuffer cmdbuf) { cmdbuf.EndTransformFeedbackEXT(0, 0, nullptr, nullptr); });
 }
 
 void RasterizerVulkan::UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& regs) {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -84,8 +84,9 @@ public:
     void DrawTexture() override;
     void Clear(u32 layer_count) override;
     void DispatchCompute() override;
-    void ResetCounter(VideoCore::QueryType type) override;
+    void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
+               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
@@ -106,7 +107,7 @@ public:
     void SyncOperation(std::function<void()>&& func) override;
     void SignalSyncPoint(u32 value) override;
     void SignalReference() override;
-    void ReleaseFences() override;
+    void ReleaseFences(bool force = true) override;
     void FlushAndInvalidateRegion(
         VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void WaitForIdle() override;
@@ -146,9 +147,7 @@ private:
 
     void UpdateDynamicStates();
 
-    void BeginTransformFeedback();
+    void HandleTransformFeedback();
-
-    void EndTransformFeedback();
 
     void UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateScissorsState(Tegra::Engines::Maxwell3D::Regs& regs);
@@ -195,8 +194,9 @@ private:
     TextureCache texture_cache;
     BufferCacheRuntime buffer_cache_runtime;
     BufferCache buffer_cache;
-    PipelineCache pipeline_cache;
+    QueryCacheRuntime query_cache_runtime;
     QueryCache query_cache;
+    PipelineCache pipeline_cache;
     AccelerateDMA accelerate_dma;
     FenceManager fence_manager;
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -243,10 +243,10 @@ void Scheduler::AllocateNewContext() {
 #if ANDROID
         if (Settings::IsGPULevelHigh()) {
             // This is problematic on Android, disable on GPU Normal.
-            query_cache->UpdateCounters();
+            query_cache->NotifySegment(true);
         }
 #else
-        query_cache->UpdateCounters();
+        query_cache->NotifySegment(true);
 #endif
     }
 }
@@ -261,11 +261,12 @@ void Scheduler::EndPendingOperations() {
 #if ANDROID
     if (Settings::IsGPULevelHigh()) {
         // This is problematic on Android, disable on GPU Normal.
-        query_cache->DisableStreams();
+        // query_cache->DisableStreams();
     }
 #else
-    query_cache->DisableStreams();
+    // query_cache->DisableStreams();
 #endif
+    query_cache->NotifySegment(false);
     EndRenderPass();
 }
 

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -17,6 +17,11 @@
 #include "video_core/renderer_vulkan/vk_master_semaphore.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
+namespace VideoCommon {
+template <typename Trait>
+class QueryCacheBase;
+}
+
 namespace Vulkan {
 
 class CommandPool;
@@ -24,7 +29,8 @@ class Device;
 class Framebuffer;
 class GraphicsPipeline;
 class StateTracker;
-class QueryCache;
+
+struct QueryCacheParams;
 
 /// The scheduler abstracts command buffer and fence management with an interface that's able to do
 /// OpenGL-like operations on Vulkan command buffers.
@@ -63,7 +69,7 @@ public:
     void InvalidateState();
 
     /// Assigns the query cache.
-    void SetQueryCache(QueryCache& query_cache_) {
+    void SetQueryCache(VideoCommon::QueryCacheBase<QueryCacheParams>& query_cache_) {
         query_cache = &query_cache_;
     }
 
@@ -219,7 +225,7 @@ private:
     std::unique_ptr<MasterSemaphore> master_semaphore;
     std::unique_ptr<CommandPool> command_pool;
 
-    QueryCache* query_cache = nullptr;
+    VideoCommon::QueryCacheBase<QueryCacheParams>* query_cache = nullptr;
 
     vk::CommandBuffer current_cmdbuf;
 

src/video_core/vulkan_common/vulkan_device.h

@@ -61,6 +61,7 @@ VK_DEFINE_HANDLE(VmaAllocator)
 
 // Define miscellaneous extensions which may be used by the implementation here.
 #define FOR_EACH_VK_EXTENSION(EXTENSION)                                                           \
+    EXTENSION(EXT, CONDITIONAL_RENDERING, conditional_rendering)                                   \
     EXTENSION(EXT, CONSERVATIVE_RASTERIZATION, conservative_rasterization)                         \
     EXTENSION(EXT, DEPTH_RANGE_UNRESTRICTED, depth_range_unrestricted)                             \
     EXTENSION(EXT, MEMORY_BUDGET, memory_budget)                                                   \
@@ -93,6 +94,7 @@ VK_DEFINE_HANDLE(VmaAllocator)
 
 // Define extensions where the absence of the extension may result in a degraded experience.
 #define FOR_EACH_VK_RECOMMENDED_EXTENSION(EXTENSION_NAME)                                          \
+    EXTENSION_NAME(VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME)                                    \
     EXTENSION_NAME(VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME)                               \
     EXTENSION_NAME(VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME)                                 \
     EXTENSION_NAME(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME)                                   \
@@ -541,6 +543,10 @@ public:
         return extensions.shader_atomic_int64;
     }
 
+    bool IsExtConditionalRendering() const {
+        return extensions.conditional_rendering;
+    }
+
     bool HasTimelineSemaphore() const;
 
     /// Returns the minimum supported version of SPIR-V.

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -75,6 +75,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkBeginCommandBuffer);
     X(vkBindBufferMemory);
     X(vkBindImageMemory);
+    X(vkCmdBeginConditionalRenderingEXT);
     X(vkCmdBeginQuery);
     X(vkCmdBeginRenderPass);
     X(vkCmdBeginTransformFeedbackEXT);
@@ -91,6 +92,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdCopyBufferToImage);
     X(vkCmdCopyImage);
     X(vkCmdCopyImageToBuffer);
+    X(vkCmdCopyQueryPoolResults);
     X(vkCmdDispatch);
     X(vkCmdDispatchIndirect);
     X(vkCmdDraw);
@@ -99,6 +101,8 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdDrawIndexedIndirect);
     X(vkCmdDrawIndirectCount);
     X(vkCmdDrawIndexedIndirectCount);
+    X(vkCmdDrawIndirectByteCountEXT);
+    X(vkCmdEndConditionalRenderingEXT);
     X(vkCmdEndQuery);
     X(vkCmdEndRenderPass);
     X(vkCmdEndTransformFeedbackEXT);

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -185,6 +185,7 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkBeginCommandBuffer vkBeginCommandBuffer{};
     PFN_vkBindBufferMemory vkBindBufferMemory{};
     PFN_vkBindImageMemory vkBindImageMemory{};
+    PFN_vkCmdBeginConditionalRenderingEXT vkCmdBeginConditionalRenderingEXT{};
     PFN_vkCmdBeginDebugUtilsLabelEXT vkCmdBeginDebugUtilsLabelEXT{};
     PFN_vkCmdBeginQuery vkCmdBeginQuery{};
     PFN_vkCmdBeginRenderPass vkCmdBeginRenderPass{};
@@ -202,6 +203,7 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdCopyBufferToImage vkCmdCopyBufferToImage{};
     PFN_vkCmdCopyImage vkCmdCopyImage{};
     PFN_vkCmdCopyImageToBuffer vkCmdCopyImageToBuffer{};
+    PFN_vkCmdCopyQueryPoolResults vkCmdCopyQueryPoolResults{};
     PFN_vkCmdDispatch vkCmdDispatch{};
     PFN_vkCmdDispatchIndirect vkCmdDispatchIndirect{};
     PFN_vkCmdDraw vkCmdDraw{};
@@ -210,6 +212,8 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdDrawIndexedIndirect vkCmdDrawIndexedIndirect{};
     PFN_vkCmdDrawIndirectCount vkCmdDrawIndirectCount{};
     PFN_vkCmdDrawIndexedIndirectCount vkCmdDrawIndexedIndirectCount{};
+    PFN_vkCmdDrawIndirectByteCountEXT vkCmdDrawIndirectByteCountEXT{};
+    PFN_vkCmdEndConditionalRenderingEXT vkCmdEndConditionalRenderingEXT{};
     PFN_vkCmdEndDebugUtilsLabelEXT vkCmdEndDebugUtilsLabelEXT{};
     PFN_vkCmdEndQuery vkCmdEndQuery{};
     PFN_vkCmdEndRenderPass vkCmdEndRenderPass{};
@@ -1182,6 +1186,13 @@ public:
                                            count_offset, draw_count, stride);
     }
 
+    void DrawIndirectByteCountEXT(u32 instance_count, u32 first_instance, VkBuffer counter_buffer,
+                                  VkDeviceSize counter_buffer_offset, u32 counter_offset,
+                                  u32 stride) {
+        dld->vkCmdDrawIndirectByteCountEXT(handle, instance_count, first_instance, counter_buffer,
+                                           counter_buffer_offset, counter_offset, stride);
+    }
+
     void ClearAttachments(Span<VkClearAttachment> attachments,
                           Span<VkClearRect> rects) const noexcept {
         dld->vkCmdClearAttachments(handle, attachments.size(), attachments.data(), rects.size(),
@@ -1270,6 +1281,13 @@ public:
                                     regions.data());
     }
 
+    void CopyQueryPoolResults(VkQueryPool query_pool, u32 first_query, u32 query_count,
+                              VkBuffer dst_buffer, VkDeviceSize dst_offset, VkDeviceSize stride,
+                              VkQueryResultFlags flags) const noexcept {
+        dld->vkCmdCopyQueryPoolResults(handle, query_pool, first_query, query_count, dst_buffer,
+                                       dst_offset, stride, flags);
+    }
+
     void FillBuffer(VkBuffer dst_buffer, VkDeviceSize dst_offset, VkDeviceSize size,
                     u32 data) const noexcept {
         dld->vkCmdFillBuffer(handle, dst_buffer, dst_offset, size, data);
@@ -1448,6 +1466,15 @@ public:
                                           counter_buffers, counter_buffer_offsets);
     }
 
+    void BeginConditionalRenderingEXT(
+        const VkConditionalRenderingBeginInfoEXT& info) const noexcept {
+        dld->vkCmdBeginConditionalRenderingEXT(handle, &info);
+    }
+
+    void EndConditionalRenderingEXT() const noexcept {
+        dld->vkCmdEndConditionalRenderingEXT(handle);
+    }
+
     void BeginDebugUtilsLabelEXT(const char* label, std::span<float, 4> color) const noexcept {
         const VkDebugUtilsLabelEXT label_info{
             .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,