[vk] Fix query cache leak on missed sync (#131)

Provided by community member, elementary-particle. Submitted as PR by MaranBR.

Fixed issues:

1. The queue cache forgot to put the host query into unregister queue if they weren't synced. This will block all the banks from freeing causing a major leak.

2. SamplesQueryCounter is not aligned with renderpass begin/end. This creates invalid queries.

3. Conditional rendering is not turned on/off at the correct location making them invalid.

Co-authored-by: Maufeat <sahyno1996@gmail.com>
Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/131
Co-authored-by: Shinmegumi <shinmegumi@eden-emu.dev>
Co-committed-by: Shinmegumi <shinmegumi@eden-emu.dev>

4 files changed, 82 insertions, 98 deletions

diff --git a/src/video_core/query_cache/query_cache.h b/src/video_core/query_cache/query_cache.h
index 08b7790555..6e084cc079 100644
--- a/src/video_core/query_cache/query_cache.h
+++ b/src/video_core/query_cache/query_cache.h
@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
@@ -115,8 +118,8 @@ struct QueryCacheBase<Traits>::QueryCacheBaseImpl {
     QueryCacheBaseImpl(QueryCacheBase<Traits>* owner_, VideoCore::RasterizerInterface& rasterizer_,
                        Tegra::MaxwellDeviceMemoryManager& device_memory_, RuntimeType& runtime_,
                        Tegra::GPU& gpu_)
-        : owner{owner_}, rasterizer{rasterizer_},
-          device_memory{device_memory_}, runtime{runtime_}, gpu{gpu_} {
+        : owner{owner_}, rasterizer{rasterizer_}, device_memory{device_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));
@@ -267,7 +270,11 @@ void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type
             return;
         }
         if (False(query_base->flags & QueryFlagBits::IsFinalValueSynced)) [[unlikely]] {
-            ASSERT(false);
+            LOG_ERROR(HW_GPU,
+                      "Query report value not synchronized. Consider increasing GPU accuracy.");
+            if (!is_synced) [[likely]] {
+                impl->pending_unregister.push_back(query_location);
+            }
             return;
         }
         query_base->value += streamer->GetAmendValue();
@@ -370,8 +377,6 @@ void QueryCacheBase<Traits>::NotifySegment(bool resume) {
     if (resume) {
         impl->runtime.ResumeHostConditionalRendering();
     } else {
-        CounterClose(VideoCommon::QueryType::ZPassPixelCount64);
-        CounterClose(VideoCommon::QueryType::StreamingByteCount);
         impl->runtime.PauseHostConditionalRendering();
     }
 }
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index 44c06eddf3..1f71bc68c6 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
@@ -1161,10 +1164,9 @@ struct QueryCacheRuntimeImpl {
                           StagingBufferPool& staging_pool_,
                           ComputePassDescriptorQueue& compute_pass_descriptor_queue,
                           DescriptorPool& descriptor_pool)
-        : rasterizer{rasterizer_}, device_memory{device_memory_},
-          buffer_cache{buffer_cache_}, device{device_},
-          memory_allocator{memory_allocator_}, scheduler{scheduler_}, staging_pool{staging_pool_},
-          guest_streamer(0, runtime),
+        : rasterizer{rasterizer_}, device_memory{device_memory_}, buffer_cache{buffer_cache_},
+          device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
+          staging_pool{staging_pool_}, guest_streamer(0, runtime),
           sample_streamer(static_cast<size_t>(QueryType::ZPassPixelCount64), runtime, rasterizer,
                           device, scheduler, memory_allocator, compute_pass_descriptor_queue,
                           descriptor_pool),
@@ -1300,9 +1302,11 @@ void QueryCacheRuntime::HostConditionalRenderingCompareValueImpl(VideoCommon::Lo
     if (impl->hcr_is_set) {
         if (impl->hcr_setup.buffer == impl->hcr_buffer &&
             impl->hcr_setup.offset == impl->hcr_offset) {
-            ResumeHostConditionalRendering();
             return;
         }
+    }
+    bool was_running = impl->is_hcr_running;
+    if (was_running) {
         PauseHostConditionalRendering();
     }
     impl->hcr_setup.buffer = impl->hcr_buffer;
@@ -1310,7 +1314,9 @@ void QueryCacheRuntime::HostConditionalRenderingCompareValueImpl(VideoCommon::Lo
     impl->hcr_setup.flags = is_equal ? VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT : 0;
     impl->hcr_is_set = true;
     impl->is_hcr_running = false;
-    ResumeHostConditionalRendering();
+    if (was_running) {
+        ResumeHostConditionalRendering();
+    }
 }
 
 void QueryCacheRuntime::HostConditionalRenderingCompareBCImpl(DAddr address, bool is_equal) {
@@ -1325,7 +1331,8 @@ void QueryCacheRuntime::HostConditionalRenderingCompareBCImpl(DAddr address, boo
         to_resolve = buffer->Handle();
         to_resolve_offset = static_cast<u32>(offset);
     }
-    if (impl->is_hcr_running) {
+    bool was_running = impl->is_hcr_running;
+    if (was_running) {
         PauseHostConditionalRendering();
     }
     impl->conditional_resolve_pass->Resolve(*impl->hcr_resolve_buffer, to_resolve,
@@ -1335,7 +1342,9 @@ void QueryCacheRuntime::HostConditionalRenderingCompareBCImpl(DAddr address, boo
     impl->hcr_setup.flags = is_equal ? 0 : VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT;
     impl->hcr_is_set = true;
     impl->is_hcr_running = false;
-    ResumeHostConditionalRendering();
+    if (was_running) {
+        ResumeHostConditionalRendering();
+    }
 }
 
 bool QueryCacheRuntime::HostConditionalRenderingCompareValue(VideoCommon::LookupData object_1,
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 92992c1ee7..c803b50e24 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -217,8 +217,6 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     FlushWork();
     gpu_memory->FlushCaching();
 
-    query_cache.NotifySegment(true);
-
     GraphicsPipeline* const pipeline{pipeline_cache.CurrentGraphicsPipeline()};
     if (!pipeline) {
         return;
@@ -232,9 +230,13 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     UpdateDynamicStates();
 
     HandleTransformFeedback();
+    query_cache.NotifySegment(true);
     query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
                               maxwell3d->regs.zpass_pixel_count_enable);
+
     draw_func();
+
+    query_cache.CounterEnable(VideoCommon::QueryType::StreamingByteCount, false);
 }
 
 void RasterizerVulkan::Draw(bool is_indexed, u32 instance_count) {
@@ -311,8 +313,6 @@ void RasterizerVulkan::DrawTexture() {
     };
     FlushWork();
 
-    query_cache.NotifySegment(true);
-
     std::scoped_lock l{texture_cache.mutex};
     texture_cache.SynchronizeGraphicsDescriptors();
     texture_cache.UpdateRenderTargets(false);
@@ -359,10 +359,6 @@ void RasterizerVulkan::Clear(u32 layer_count) {
     FlushWork();
     gpu_memory->FlushCaching();
 
-    query_cache.NotifySegment(true);
-    query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
-                              maxwell3d->regs.zpass_pixel_count_enable);
-
     auto& regs = maxwell3d->regs;
     const bool use_color = regs.clear_surface.R || regs.clear_surface.G || regs.clear_surface.B ||
                            regs.clear_surface.A;
@@ -378,6 +374,10 @@ void RasterizerVulkan::Clear(u32 layer_count) {
     const VkExtent2D render_area = framebuffer->RenderArea();
     scheduler.RequestRenderpass(framebuffer);
 
+    query_cache.NotifySegment(true);
+    query_cache.CounterEnable(VideoCommon::QueryType::ZPassPixelCount64,
+                              maxwell3d->regs.zpass_pixel_count_enable);
+
     u32 up_scale = 1;
     u32 down_shift = 0;
     if (texture_cache.IsRescaling()) {
@@ -832,6 +832,7 @@ std::optional<FramebufferTextureInfo> RasterizerVulkan::AccelerateDisplay(
     if (!image_view) {
         return {};
     }
+
     query_cache.NotifySegment(false);
 
     const auto& resolution = Settings::values.resolution_info;
@@ -943,22 +944,20 @@ void RasterizerVulkan::UpdateDynamicStates() {
     UpdateDepthBounds(regs);
     UpdateStencilFaces(regs);
     UpdateLineWidth(regs);
-    // TODO: updating line stipple causes the cmdbuf to die
-    // UpdateLineStipple(regs);
 
     const u8 dynamic_state = Settings::values.dyna_state.GetValue();
 
     auto features = DynamicFeatures{
-        .has_extended_dynamic_state = device.IsExtExtendedDynamicStateSupported()
-                                      && dynamic_state > 0,
-        .has_extended_dynamic_state_2 = device.IsExtExtendedDynamicState2Supported()
-                                        && dynamic_state > 1,
-        .has_extended_dynamic_state_2_extra = device.IsExtExtendedDynamicState2ExtrasSupported()
-                                              && dynamic_state > 1,
-        .has_extended_dynamic_state_3_blend = device.IsExtExtendedDynamicState3BlendingSupported()
-                                              && dynamic_state > 2,
-        .has_extended_dynamic_state_3_enables = device.IsExtExtendedDynamicState3EnablesSupported()
-                                                && dynamic_state > 2,
+        .has_extended_dynamic_state =
+            device.IsExtExtendedDynamicStateSupported() && dynamic_state > 0,
+        .has_extended_dynamic_state_2 =
+            device.IsExtExtendedDynamicState2Supported() && dynamic_state > 1,
+        .has_extended_dynamic_state_2_extra =
+            device.IsExtExtendedDynamicState2ExtrasSupported() && dynamic_state > 1,
+        .has_extended_dynamic_state_3_blend =
+            device.IsExtExtendedDynamicState3BlendingSupported() && dynamic_state > 2,
+        .has_extended_dynamic_state_3_enables =
+            device.IsExtExtendedDynamicState3EnablesSupported() && dynamic_state > 2,
         .has_dynamic_vertex_input = device.IsExtVertexInputDynamicStateSupported(),
     };
 
@@ -983,16 +982,12 @@ void RasterizerVulkan::UpdateDynamicStates() {
             if (features.has_extended_dynamic_state_3_enables) {
                 using namespace Tegra::Engines;
 
-                if (device.GetDriverID() == VkDriverIdKHR::VK_DRIVER_ID_AMD_OPEN_SOURCE
-                    || device.GetDriverID() == VkDriverIdKHR::VK_DRIVER_ID_AMD_PROPRIETARY) {
-                    struct In
-                    {
+                if (device.GetDriverID() == VkDriverIdKHR::VK_DRIVER_ID_AMD_OPEN_SOURCE ||
+                    device.GetDriverID() == VkDriverIdKHR::VK_DRIVER_ID_AMD_PROPRIETARY) {
+                    struct In {
                         const Maxwell3D::Regs::VertexAttribute::Type d;
-                        In(Maxwell3D::Regs::VertexAttribute::Type n)
-                            : d(n)
-                        {}
-                        bool operator()(Maxwell3D::Regs::VertexAttribute n) const
-                        {
+                        In(Maxwell3D::Regs::VertexAttribute::Type n) : d(n) {}
+                        bool operator()(Maxwell3D::Regs::VertexAttribute n) const {
                             return n.type == d;
                         }
                     };
@@ -1143,36 +1138,36 @@ void RasterizerVulkan::UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs) {
     if (is_d24 && !device.SupportsD24DepthBuffer()) {
         static constexpr const size_t length = sizeof(NEEDS_D24) / sizeof(NEEDS_D24[0]);
 
-        static constexpr const u64 *start = NEEDS_D24;
-        static constexpr const u64 *end = NEEDS_D24 + length;
+        static constexpr const u64* start = NEEDS_D24;
+        static constexpr const u64* end = NEEDS_D24 + length;
 
-        const u64 *it = std::find(start, end, program_id);
+        const u64* it = std::find(start, end, program_id);
 
         if (it != end) {
             // the base formulas can be obtained from here:
             //   https://docs.microsoft.com/en-us/windows/win32/direct3d11/d3d10-graphics-programming-guide-output-merger-stage-depth-bias
-            const double rescale_factor = static_cast<double>(1ULL << (32 - 24))
-                                          / (static_cast<double>(0x1.ep+127));
+            const double rescale_factor =
+                static_cast<double>(1ULL << (32 - 24)) / (static_cast<double>(0x1.ep+127));
             units = static_cast<float>(static_cast<double>(units) * rescale_factor);
         }
     }
 
-    scheduler.Record(
-        [constant = units, clamp = regs.depth_bias_clamp, factor = regs.slope_scale_depth_bias, this](
-            vk::CommandBuffer cmdbuf) {
-            if (device.IsExtDepthBiasControlSupported()) {
-                static VkDepthBiasRepresentationInfoEXT bias_info{
-                    .sType = VK_STRUCTURE_TYPE_DEPTH_BIAS_REPRESENTATION_INFO_EXT,
-                    .pNext = nullptr,
-                    .depthBiasRepresentation = VK_DEPTH_BIAS_REPRESENTATION_LEAST_REPRESENTABLE_VALUE_FORCE_UNORM_EXT,
-                    .depthBiasExact = VK_FALSE,
-                };
+    scheduler.Record([constant = units, clamp = regs.depth_bias_clamp,
+                      factor = regs.slope_scale_depth_bias, this](vk::CommandBuffer cmdbuf) {
+        if (device.IsExtDepthBiasControlSupported()) {
+            static VkDepthBiasRepresentationInfoEXT bias_info{
+                .sType = VK_STRUCTURE_TYPE_DEPTH_BIAS_REPRESENTATION_INFO_EXT,
+                .pNext = nullptr,
+                .depthBiasRepresentation =
+                    VK_DEPTH_BIAS_REPRESENTATION_LEAST_REPRESENTABLE_VALUE_FORCE_UNORM_EXT,
+                .depthBiasExact = VK_FALSE,
+            };
 
-                cmdbuf.SetDepthBias(constant, clamp, factor, &bias_info);
-            } else {
-                cmdbuf.SetDepthBias(constant, clamp, factor);
-            }
-        });
+            cmdbuf.SetDepthBias(constant, clamp, factor, &bias_info);
+        } else {
+            cmdbuf.SetDepthBias(constant, clamp, factor);
+        }
+    });
 }
 
 void RasterizerVulkan::UpdateBlendConstants(Tegra::Engines::Maxwell3D::Regs& regs) {
@@ -1354,8 +1349,7 @@ void RasterizerVulkan::UpdateRasterizerDiscardEnable(Tegra::Engines::Maxwell3D::
     });
 }
 
-void RasterizerVulkan::UpdateConservativeRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs)
-{
+void RasterizerVulkan::UpdateConservativeRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs) {
     if (!state_tracker.TouchConservativeRasterizationMode()) {
         return;
     }
@@ -1367,8 +1361,7 @@ void RasterizerVulkan::UpdateConservativeRasterizationMode(Tegra::Engines::Maxwe
     });
 }
 
-void RasterizerVulkan::UpdateLineStippleEnable(Tegra::Engines::Maxwell3D::Regs& regs)
-{
+void RasterizerVulkan::UpdateLineStippleEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
     if (!state_tracker.TouchLineStippleEnable()) {
         return;
     }
@@ -1378,19 +1371,7 @@ void RasterizerVulkan::UpdateLineStippleEnable(Tegra::Engines::Maxwell3D::Regs&
     });
 }
 
-void RasterizerVulkan::UpdateLineStipple(Tegra::Engines::Maxwell3D::Regs& regs)
-{
-    if (!state_tracker.TouchLineStipple()) {
-        return;
-    }
-
-    scheduler.Record([params = regs.line_stipple_params](vk::CommandBuffer cmdbuf) {
-        cmdbuf.SetLineStippleEXT(params.factor, static_cast<uint16_t>(params.pattern));
-    });
-}
-
-void RasterizerVulkan::UpdateLineRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs)
-{
+void RasterizerVulkan::UpdateLineRasterizationMode(Tegra::Engines::Maxwell3D::Regs& regs) {
     // if (!state_tracker.TouchLi()) {
     //     return;
     // }
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.cpp b/src/video_core/renderer_vulkan/vk_scheduler.cpp
index 146923db4d..7c556588b6 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.cpp
+++ b/src/video_core/renderer_vulkan/vk_scheduler.cpp
@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -257,16 +260,6 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
 
 void Scheduler::AllocateNewContext() {
     // Enable counters once again. These are disabled when a command buffer is finished.
-    if (query_cache) {
-#if ANDROID
-        if (Settings::IsGPULevelHigh()) {
-            // This is problematic on Android, disable on GPU Normal.
-            query_cache->NotifySegment(true);
-        }
-#else
-        query_cache->NotifySegment(true);
-#endif
-    }
 }
 
 void Scheduler::InvalidateState() {
@@ -276,15 +269,7 @@ void Scheduler::InvalidateState() {
 }
 
 void Scheduler::EndPendingOperations() {
-#if ANDROID
-    if (Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        // query_cache->DisableStreams();
-    }
-#else
-    // query_cache->DisableStreams();
-#endif
-    query_cache->NotifySegment(false);
+    query_cache->CounterReset(VideoCommon::QueryType::ZPassPixelCount64);
     EndRenderPass();
 }
 
@@ -292,6 +277,10 @@ void Scheduler::EndRenderPass() {
     if (!state.renderpass) {
         return;
     }
+
+    query_cache->CounterEnable(VideoCommon::QueryType::ZPassPixelCount64, false);
+    query_cache->NotifySegment(false);
+
     Record([num_images = num_renderpass_images, images = renderpass_images,
             ranges = renderpass_image_ranges](vk::CommandBuffer cmdbuf) {
         std::array<VkImageMemoryBarrier, 9> barriers;