[Vulkan] speed up bank lookup with LRU cache, batch growth to 32

Adds a tiny 8 entry LRU cache(should be tested for optimal number)
Increase batch growth to 32 (should be tested to determine the difference between 16 and 32)

src/video_core/renderer_vulkan/vk_descriptor_pool.cpp

@@ -20,7 +20,7 @@
 namespace Vulkan {
 
 // Prefer small grow rates to avoid saturating the descriptor pool with barely used pipelines
-constexpr size_t SETS_GROW_RATE = 16;
+constexpr size_t SETS_GROW_RATE = 32; //test difference between 16 and 32
 constexpr s32 SCORE_THRESHOLD = 3;
 
 struct DescriptorBank {
@@ -29,9 +29,12 @@ struct DescriptorBank {
 };
 
 bool DescriptorBankInfo::IsSuperset(const DescriptorBankInfo& subset) const noexcept {
-    return uniform_buffers >= subset.uniform_buffers && storage_buffers >= subset.storage_buffers &&
-           texture_buffers >= subset.texture_buffers && image_buffers >= subset.image_buffers &&
-           textures >= subset.textures && images >= subset.images;
+    return uniform_buffers >= subset.uniform_buffers &&
+           storage_buffers >= subset.storage_buffers &&
+           texture_buffers >= subset.texture_buffers &&
+           image_buffers >= subset.image_buffers &&
+           textures >= subset.textures &&
+           images >= subset.images;
 }
 
 template <typename Descriptors>
@@ -45,6 +48,19 @@ static u32 Accumulate(const Descriptors& descriptors) {
 
 static DescriptorBankInfo MakeBankInfo(std::span<const Shader::Info> infos) {
     DescriptorBankInfo bank;
+    if (infos.size() == 1) {
+        const auto& info = infos.front();
+        const auto acc = [](const auto& ds){ u32 c=0; for (const auto& d: ds) c+=d.count; return c; };
+        bank.uniform_buffers += acc(info.constant_buffer_descriptors);
+        bank.storage_buffers += acc(info.storage_buffers_descriptors);
+        bank.texture_buffers += acc(info.texture_buffer_descriptors);
+        bank.image_buffers   += acc(info.image_buffer_descriptors);
+        bank.textures        += acc(info.texture_descriptors);
+        bank.images          += acc(info.image_descriptors);
+        bank.score = bank.uniform_buffers + bank.storage_buffers + bank.texture_buffers +
+        bank.image_buffers + bank.textures + bank.images;
+        return bank;
+    }
     for (const Shader::Info& info : infos) {
         bank.uniform_buffers += Accumulate(info.constant_buffer_descriptors);
         bank.storage_buffers += Accumulate(info.storage_buffers_descriptors);
@@ -102,13 +118,22 @@ void DescriptorAllocator::Allocate(size_t begin, size_t end) {
 }
 
 vk::DescriptorSets DescriptorAllocator::AllocateDescriptors(size_t count) {
-    const std::vector<VkDescriptorSetLayout> layouts(count, layout);
+    std::array<VkDescriptorSetLayout, 64> stack{};
+    const VkDescriptorSetLayout* p_layouts = nullptr;
+    std::vector<VkDescriptorSetLayout> heap;
+    if (count <= stack.size()) {
+        stack.fill(layout);
+        p_layouts = stack.data();
+    } else {
+        heap.assign(count, layout);
+        p_layouts = heap.data();
+        }
     VkDescriptorSetAllocateInfo allocate_info{
         .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
         .pNext = nullptr,
         .descriptorPool = *bank->pools.back(),
         .descriptorSetCount = static_cast<u32>(count),
-        .pSetLayouts = layouts.data(),
+        .pSetLayouts = p_layouts,
     };
     vk::DescriptorSets new_sets = bank->pools.back().Allocate(allocate_info);
     if (!new_sets.IsOutOfPoolMemory()) {
@@ -146,21 +171,58 @@ DescriptorAllocator DescriptorPool::Allocator(VkDescriptorSetLayout layout,
 }
 
 DescriptorBank& DescriptorPool::Bank(const DescriptorBankInfo& reqs) {
+    {
+        std::scoped_lock lk(cache_mutex);
+        DescriptorBank* best = nullptr; u64 best_stamp = 0;
+        for (const auto& e : cache_) {
+            if (!e.bank) continue;
+            if (std::abs(e.info.score - reqs.score) < SCORE_THRESHOLD && e.info.IsSuperset(reqs)) {
+                if (e.stamp >= best_stamp) { best_stamp = e.stamp; best = e.bank; }
+            }
+        }
+        if (best) return *best;
+    }
     std::shared_lock read_lock{banks_mutex};
     const auto it = std::ranges::find_if(bank_infos, [&reqs](const DescriptorBankInfo& bank) {
         return std::abs(bank.score - reqs.score) < SCORE_THRESHOLD && bank.IsSuperset(reqs);
     });
     if (it != bank_infos.end()) {
-        return *banks[std::distance(bank_infos.begin(), it)].get();
+        DescriptorBank& found = *banks[std::distance(bank_infos.begin(), it)].get();
+        read_lock.unlock();
+        // update cache
+        std::scoped_lock lk(cache_mutex);
+        size_t victim = 0; u64 oldest = UINT64_MAX;
+        for (size_t i=0;i<cache_.size();++i) if (cache_[i].stamp < oldest) { oldest = cache_[i].stamp; victim = i; }
+             cache_[victim] = CacheEntry{found.info, &found, ++cache_tick_};
+        return found;
     }
     read_lock.unlock();
 
     std::unique_lock write_lock{banks_mutex};
+    auto it2 = std::ranges::find_if(bank_infos, [&reqs](const DescriptorBankInfo& bank) {
+          return std::abs(bank.score - reqs.score) < SCORE_THRESHOLD && bank.IsSuperset(reqs);
+    });
+    if (it2 != bank_infos.end()) {
+        DescriptorBank& found = *banks[std::distance(bank_infos.begin(), it2)].get();
+        // update cache
+        std::scoped_lock lk(cache_mutex);
+        size_t victim = 0; u64 oldest = UINT64_MAX;
+        for (size_t i=0;i<cache_.size();++i) if (cache_[i].stamp < oldest) { oldest = cache_[i].stamp; victim = i; }
+        cache_[victim] = CacheEntry{found.info, &found, ++cache_tick_};
+        return found;
+    }
     bank_infos.push_back(reqs);
 
     auto& bank = *banks.emplace_back(std::make_unique<DescriptorBank>());
     bank.info = reqs;
     AllocatePool(device, bank);
+    // update cache
+    {
+         std::scoped_lock lk(cache_mutex);
+         size_t victim = 0; u64 oldest = UINT64_MAX;
+         for (size_t i=0;i<cache_.size();++i) if (cache_[i].stamp < oldest) { oldest = cache_[i].stamp; victim = i; }
+         cache_[victim] = CacheEntry{bank.info, &bank, ++cache_tick_};
+    }
     return bank;
 }
 

src/video_core/renderer_vulkan/vk_descriptor_pool.h

@@ -6,7 +6,8 @@
 #include <shared_mutex>
 #include <span>
 #include <vector>
-
+#include <array>
+#include <mutex>
 #include "shader_recompiler/shader_info.h"
 #include "video_core/renderer_vulkan/vk_resource_pool.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -75,6 +76,14 @@ public:
 
 private:
     DescriptorBank& Bank(const DescriptorBankInfo& reqs);
+    struct CacheEntry {
+        DescriptorBankInfo info{};
+        DescriptorBank* bank{nullptr};
+        u64 stamp{0};
+    };
+    std::mutex cache_mutex{};
+    std::array<CacheEntry, 8> cache_{}; //test and then adjust
+    u64 cache_tick_{0};
 
     const Device& device;
     MasterSemaphore& master_semaphore;