Query Cache: Implement host side sample counting.

5 files changed, 348 insertions, 48 deletions

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index fb24b6532..8218ec4c8 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -41,6 +41,7 @@ set(SHADER_FILES
     pitch_unswizzle.comp
     present_bicubic.frag
     present_gaussian.frag
+    queries_prefix_scan_sum.comp
     resolve_conditional_render.comp
     smaa_edge_detection.vert
     smaa_edge_detection.frag
diff --git a/src/video_core/host_shaders/queries_prefix_scan_sum.comp b/src/video_core/host_shaders/queries_prefix_scan_sum.comp
new file mode 100644
index 000000000..dce1279fe
--- /dev/null
+++ b/src/video_core/host_shaders/queries_prefix_scan_sum.comp
@@ -0,0 +1,124 @@
+// 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 and optimize for subgroup
+
+#version 460 core
+
+#ifdef VULKAN
+
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#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_KHR_shader_subgroup_arithmetic : enable
+#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 max_accumulation_base;
+UNIFORM(1) uint accumulation_limit;
+END_PUSH_CONSTANTS
+
+layout(local_size_x = 32) in;
+
+layout(std430, binding = 0) readonly buffer block1 {
+    uvec2 input_data[gl_WorkGroupSize.x];
+};
+
+layout(std430, binding = 1) writeonly coherent buffer block2 {
+    uvec2 output_data[gl_WorkGroupSize.x];
+};
+
+layout(std430, binding = 2) coherent buffer block3 {
+    uvec2 accumulated_data;
+};
+
+shared uvec2 shared_data[gl_WorkGroupSize.x * 2];
+
+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_1 = (id * 2) < max_accumulation_base ? accumulated_data : uvec2(0);
+    uvec2 base_value_2 = (id * 2 + 1) < max_accumulation_base ? accumulated_data : uvec2(0);
+    uint work_size = gl_WorkGroupSize.x;
+    uint rd_id;
+    uint wr_id;
+    uint mask;
+    uvec2 input_1 = input_data[id * 2];
+    uvec2 input_2 = input_data[id * 2 + 1];
+    // 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)) + 1;
+    uint step = 0;
+
+    // Each invocation is responsible for the content of
+    // two elements of the output array
+    shared_data[id * 2] = input_1;
+    shared_data[id * 2 + 1] = input_2;
+    // 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
+    shared_data[id * 2] = AddUint64(shared_data[id * 2], base_value_1);
+    shared_data[id * 2 + 1] = AddUint64(shared_data[id * 2 + 1], base_value_2);
+    barrier();
+    memoryBarrierShared();
+
+    // Finally write our data back to the output buffer
+    output_data[id * 2] = shared_data[id * 2];
+    output_data[id * 2 + 1] = shared_data[id * 2 + 1];
+    if (id == 0) {
+        if (max_accumulation_base >= accumulation_limit + 1) {
+            accumulated_data = shared_data[accumulation_limit];
+            return;
+        }
+        uvec2 value_1 = shared_data[max_accumulation_base];
+        uvec2 value_2 = shared_data[accumulation_limit];
+        accumulated_data = AddUint64(value_1, -value_2);
+    }
+}
\ No newline at end of file
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
index 039dc95e1..a1af08cda 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
@@ -12,6 +12,7 @@
 #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/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"
@@ -58,6 +59,30 @@ 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,
@@ -68,6 +93,16 @@ 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,
@@ -104,6 +139,15 @@ 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{{
         {
@@ -132,6 +176,11 @@ struct AstcPushConstants {
     u32 block_height;
     u32 block_height_mask;
 };
+
+struct QueriesPrefixScanPushConstants {
+    u32 max_accumulation_base;
+    u32 accumulation_limit;
+};
 } // Anonymous namespace
 
 ComputePass::ComputePass(const Device& device_, DescriptorPool& descriptor_pool,
@@ -313,8 +362,6 @@ ConditionalRenderingResolvePass::ConditionalRenderingResolvePass(
 
 void ConditionalRenderingResolvePass::Resolve(VkBuffer dst_buffer, VkBuffer src_buffer,
                                               u32 src_offset, bool compare_to_zero) {
-    scheduler.RequestOutsideRenderPassOperationContext();
-
     const size_t compare_size = compare_to_zero ? 8 : 24;
 
     compute_pass_descriptor_queue.Acquire();
@@ -327,7 +374,7 @@ void ConditionalRenderingResolvePass::Resolve(VkBuffer dst_buffer, VkBuffer src_
         static constexpr VkMemoryBarrier read_barrier{
             .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
             .pNext = nullptr,
-            .srcAccessMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+            .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{
@@ -349,6 +396,63 @@ void ConditionalRenderingResolvePass::Resolve(VkBuffer dst_buffer, VkBuffer src_
     });
 }
 
+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)>,
+                  QUERIES_PREFIX_SCAN_SUM_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 max_accumulation_limit) {
+    size_t aligned_runs = Common::AlignUp(number_of_sums, 32);
+
+    compute_pass_descriptor_queue.Acquire();
+    compute_pass_descriptor_queue.AddBuffer(src_buffer, 0, aligned_runs * sizeof(u64));
+    compute_pass_descriptor_queue.AddBuffer(dst_buffer, 0, aligned_runs * sizeof(u64));
+    compute_pass_descriptor_queue.AddBuffer(accumulation_buffer, 0, sizeof(u64));
+    const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([this, descriptor_data, max_accumulation_limit, number_of_sums,
+                      aligned_runs](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{
+            .max_accumulation_base = static_cast<u32>(max_accumulation_limit),
+            .accumulation_limit = static_cast<u32>(number_of_sums - 1),
+        };
+        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(static_cast<u32>(aligned_runs / 32U), 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_,
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.h b/src/video_core/renderer_vulkan/vk_compute_pass.h
index c62f30d30..e6ff86e9a 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.h
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.h
@@ -95,6 +95,20 @@ private:
     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 max_accumulation_limit);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
 class ASTCDecoderPass final : public ComputePass {
 public:
     explicit ASTCDecoderPass(const Device& device_, Scheduler& scheduler_,
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index 2147776f8..ded190ae0 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -11,6 +11,7 @@
 #include <utility>
 #include <vector>
 
+#include "common/bit_util.h"
 #include "common/common_types.h"
 #include "core/memory.h"
 #include "video_core/engines/draw_manager.h"
@@ -112,14 +113,34 @@ class SamplesStreamer : public BaseStreamer {
 public:
     explicit SamplesStreamer(size_t id_, QueryCacheRuntime& runtime_,
                              VideoCore::RasterizerInterface* rasterizer_, const Device& device_,
-                             Scheduler& scheduler_, const MemoryAllocator& memory_allocator_)
+                             Scheduler& scheduler_, const MemoryAllocator& memory_allocator_,
+                             ComputePassDescriptorQueue& compute_pass_descriptor_queue,
+                             DescriptorPool& descriptor_pool)
         : BaseStreamer(id_), runtime{runtime_}, rasterizer{rasterizer_}, device{device_},
           scheduler{scheduler_}, memory_allocator{memory_allocator_} {
-        BuildResolveBuffer();
         current_bank = nullptr;
         current_query = nullptr;
         ammend_value = 0;
         acumulation_value = 0;
+        queries_prefix_scan_pass = std::make_unique<QueriesPrefixScanPass>(
+            device, scheduler, descriptor_pool, compute_pass_descriptor_queue);
+
+        const VkBufferCreateInfo buffer_ci = {
+            .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .size = 8,
+            .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
+                     VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
+            .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+            .queueFamilyIndexCount = 0,
+            .pQueueFamilyIndices = nullptr,
+        };
+        accumulation_buffer = memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal);
+        scheduler.RequestOutsideRenderPassOperationContext();
+        scheduler.Record([buffer = *accumulation_buffer](vk::CommandBuffer cmdbuf) {
+            cmdbuf.FillBuffer(buffer, 0, 8, 0);
+        });
     }
 
     ~SamplesStreamer() = default;
@@ -159,6 +180,8 @@ public:
             acumulation_value = 0;
         });
         rasterizer->SyncOperation(std::move(func));
+        accumulation_since_last_sync = false;
+        last_accumulation_checkpoint = std::min(last_accumulation_checkpoint, num_slots_used);
     }
 
     void CloseCounter() override {
@@ -175,7 +198,8 @@ public:
         }
 
         for (size_t i = 0; i < sync_values_stash.size(); i++) {
-            runtime.template SyncValues<HostSyncValues>(sync_values_stash[i], *resolve_buffers[i]);
+            runtime.template SyncValues<HostSyncValues>(sync_values_stash[i],
+                                                        *buffers[resolve_buffers[i]]);
         }
 
         sync_values_stash.clear();
@@ -189,36 +213,21 @@ public:
         sync_values_stash.clear();
         sync_values_stash.emplace_back();
         std::vector<HostSyncValues>* sync_values = &sync_values_stash.back();
-        sync_values->reserve(resolve_slots * SamplesQueryBank::BANK_SIZE);
+        sync_values->reserve(num_slots_used);
         std::unordered_map<size_t, std::pair<size_t, size_t>> offsets;
-        size_t this_bank_slot = std::numeric_limits<size_t>::max();
-        size_t resolve_slots_remaining = resolve_slots;
-        size_t resolve_buffer_index = 0;
+        resolve_buffers.clear();
+        size_t resolve_buffer_index = ObtainBuffer<true>(num_slots_used);
+        resolve_buffers.push_back(resolve_buffer_index);
+        size_t base_offset = 0;
+
         ApplyBanksWideOp<true>(pending_sync, [&](SamplesQueryBank* bank, size_t start,
                                                  size_t amount) {
             size_t bank_id = bank->GetIndex();
-            if (this_bank_slot != bank_id) {
-                this_bank_slot = bank_id;
-                if (resolve_slots_remaining == 0) {
-                    resolve_buffer_index++;
-                    if (resolve_buffer_index >= resolve_buffers.size()) {
-                        BuildResolveBuffer();
-                    }
-                    resolve_slots_remaining = resolve_slots;
-                    sync_values_stash.emplace_back();
-                    sync_values = &sync_values_stash.back();
-                    sync_values->reserve(resolve_slots * SamplesQueryBank::BANK_SIZE);
-                }
-                resolve_slots_remaining--;
-            }
-            auto& resolve_buffer = resolve_buffers[resolve_buffer_index];
-            const size_t base_offset = SamplesQueryBank::QUERY_SIZE * SamplesQueryBank::BANK_SIZE *
-                                       (resolve_slots - resolve_slots_remaining - 1);
+            auto& resolve_buffer = buffers[resolve_buffer_index];
             VkQueryPool query_pool = bank->GetInnerPool();
             scheduler.RequestOutsideRenderPassOperationContext();
             scheduler.Record([start, amount, base_offset, query_pool,
                               buffer = *resolve_buffer](vk::CommandBuffer cmdbuf) {
-                size_t final_offset = base_offset + start * SamplesQueryBank::QUERY_SIZE;
                 const VkBufferMemoryBarrier copy_query_pool_barrier{
                     .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
                     .pNext = nullptr,
@@ -227,39 +236,60 @@ public:
                     .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
                     .buffer = buffer,
-                    .offset = final_offset,
+                    .offset = base_offset,
                     .size = amount * SamplesQueryBank::QUERY_SIZE,
                 };
 
                 cmdbuf.CopyQueryPoolResults(
                     query_pool, static_cast<u32>(start), static_cast<u32>(amount), buffer,
-                    static_cast<u32>(final_offset), SamplesQueryBank::QUERY_SIZE,
+                    static_cast<u32>(base_offset), SamplesQueryBank::QUERY_SIZE,
                     VK_QUERY_RESULT_WAIT_BIT | VK_QUERY_RESULT_64_BIT);
                 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
                                        VK_PIPELINE_STAGE_TRANSFER_BIT, 0, copy_query_pool_barrier);
             });
-            offsets[bank_id] = {sync_values_stash.size() - 1, base_offset};
+            offsets[bank_id] = {start, base_offset};
+            base_offset += amount * SamplesQueryBank::QUERY_SIZE;
         });
 
         // Convert queries
+        bool has_multi_queries = false;
         for (auto q : pending_sync) {
             auto* query = GetQuery(q);
+            size_t sync_value_slot = 0;
             if (True(query->flags & VideoCommon::QueryFlagBits::IsRewritten)) {
                 continue;
             }
             if (True(query->flags & VideoCommon::QueryFlagBits::IsInvalidated)) {
                 continue;
             }
-            if (query->size_slots > 1) {
-                // This is problematic.
-                // UNIMPLEMENTED();
+            if (accumulation_since_last_sync || query->size_slots > 1) {
+                if (!has_multi_queries) {
+                    has_multi_queries = true;
+                    sync_values_stash.emplace_back();
+                }
+                sync_value_slot = 1;
             }
             query->flags |= VideoCommon::QueryFlagBits::IsHostSynced;
             auto loc_data = offsets[query->start_bank_id];
-            sync_values_stash[loc_data.first].emplace_back(HostSyncValues{
+            sync_values_stash[sync_value_slot].emplace_back(HostSyncValues{
                 .address = query->guest_address,
                 .size = SamplesQueryBank::QUERY_SIZE,
-                .offset = loc_data.second + query->start_slot * SamplesQueryBank::QUERY_SIZE,
+                .offset =
+                    loc_data.second + (query->start_slot - loc_data.first + query->size_slots - 1) *
+                                          SamplesQueryBank::QUERY_SIZE,
+            });
+        }
+
+        if (has_multi_queries) {
+            size_t intermediary_buffer_index = ObtainBuffer<false>(num_slots_used);
+            resolve_buffers.push_back(intermediary_buffer_index);
+            queries_prefix_scan_pass->Run(*accumulation_buffer, *buffers[intermediary_buffer_index],
+                                          *buffers[resolve_buffer_index], num_slots_used,
+                                          std::min(last_accumulation_checkpoint, num_slots_used));
+        } else {
+            scheduler.RequestOutsideRenderPassOperationContext();
+            scheduler.Record([buffer = *accumulation_buffer](vk::CommandBuffer cmdbuf) {
+                cmdbuf.FillBuffer(buffer, 0, 8, 0);
             });
         }
 
@@ -267,6 +297,9 @@ public:
         std::function<void()> func([this] { ammend_value = acumulation_value; });
         rasterizer->SyncOperation(std::move(func));
         AbandonCurrentQuery();
+        num_slots_used = 0;
+        last_accumulation_checkpoint = std::numeric_limits<size_t>::max();
+        accumulation_since_last_sync = has_multi_queries;
         pending_sync.clear();
     }
 
@@ -400,6 +433,7 @@ private:
     void ReserveHostQuery() {
         size_t new_slot = ReserveBankSlot();
         current_bank->AddReference(1);
+        num_slots_used++;
         if (current_query) {
             size_t bank_id = current_query->start_bank_id;
             size_t banks_set = current_query->size_banks - 1;
@@ -470,32 +504,50 @@ private:
         });
     }
 
-    void BuildResolveBuffer() {
+    template <bool is_resolve>
+    size_t ObtainBuffer(size_t num_needed) {
+        const size_t log_2 = std::max<size_t>(6U, Common::Log2Ceil64(num_needed));
+        if constexpr (is_resolve) {
+            if (resolve_table[log_2] != 0) {
+                return resolve_table[log_2] - 1;
+            }
+        } else {
+            if (intermediary_table[log_2] != 0) {
+                return intermediary_table[log_2] - 1;
+            }
+        }
         const VkBufferCreateInfo buffer_ci = {
             .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
             .pNext = nullptr,
             .flags = 0,
-            .size = SamplesQueryBank::QUERY_SIZE * SamplesQueryBank::BANK_SIZE * resolve_slots,
+            .size = SamplesQueryBank::QUERY_SIZE * (1ULL << log_2),
             .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
                      VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
             .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
             .queueFamilyIndexCount = 0,
             .pQueueFamilyIndices = nullptr,
         };
-        resolve_buffers.emplace_back(
-            memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal));
+        buffers.emplace_back(memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal));
+        if constexpr (is_resolve) {
+            resolve_table[log_2] = buffers.size();
+        } else {
+            intermediary_table[log_2] = buffers.size();
+        }
+        return buffers.size() - 1;
     }
 
-    static constexpr size_t resolve_slots = 8;
-
     QueryCacheRuntime& runtime;
     VideoCore::RasterizerInterface* rasterizer;
     const Device& device;
     Scheduler& scheduler;
     const MemoryAllocator& memory_allocator;
     VideoCommon::BankPool<SamplesQueryBank> bank_pool;
-    std::deque<vk::Buffer> resolve_buffers;
+    std::deque<vk::Buffer> buffers;
+    std::array<size_t, 32> resolve_table{};
+    std::array<size_t, 32> intermediary_table{};
+    vk::Buffer accumulation_buffer;
     std::deque<std::vector<HostSyncValues>> sync_values_stash;
+    std::vector<size_t> resolve_buffers;
 
     // syncing queue
     std::vector<size_t> pending_sync;
@@ -510,10 +562,14 @@ private:
     SamplesQueryBank* current_bank;
     VkQueryPool current_query_pool;
     size_t current_query_id;
+    size_t num_slots_used{};
+    size_t last_accumulation_checkpoint{};
+    bool accumulation_since_last_sync{};
     VideoCommon::HostQueryBase* current_query;
     bool has_started{};
-    bool current_unset{};
     std::mutex flush_guard;
+
+    std::unique_ptr<QueriesPrefixScanPass> queries_prefix_scan_pass;
 };
 
 // Transform feedback queries
@@ -1090,7 +1146,8 @@ struct QueryCacheRuntimeImpl {
           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),
+                          device, scheduler, memory_allocator, compute_pass_descriptor_queue,
+                          descriptor_pool),
           tfb_streamer(static_cast<size_t>(QueryType::StreamingByteCount), runtime, device,
                        scheduler, memory_allocator, staging_pool),
           primitives_succeeded_streamer(
@@ -1319,10 +1376,10 @@ bool QueryCacheRuntime::HostConditionalRenderingCompareValues(VideoCommon::Looku
             return true;
         }
     }
-    if (!is_in_bc[0] && !is_in_bc[1]) {
+    /*if (!is_in_bc[0] && !is_in_bc[1]) {
         // Both queries are in query cache, it's best to just flush.
-        return false;
-    }
+        return true;
+    }*/
     HostConditionalRenderingCompareBCImpl(object_1.address, equal_check);
     return true;
 }