From ad25c9ee1156cbad1a6591f2bbb74c53cf171695 Mon Sep 17 00:00:00 2001
From: Celegans12 <53881228+Celegans12@users.noreply.github.com>
Date: Wed, 21 Oct 2026 05:59:34 +0200
Subject: [PATCH] perf(agc): prefer a pooled buffer of the requested exact size

A pooled buffer is retained by size class, and a hit whose VkBuffer was
created for another exact size destroys it and binds a new one to the
same memory (BindExactBuffer). Copying draw inputs on NVIDIA's Windows
driver takes one such buffer per index or vertex range per draw, so most
hits in that path pay a vkDestroyBuffer, vkCreateBuffer and bind.

Take now scans the newest 64 slots of the class for one whose buffer has
the requested size and takes the newest slot as before when none has.
The memory class, usage and properties are the same, so only the choice
among equivalent slots changes.

AI-assisted: yes
---
 .../Graphics/include/BufferPool.hpp           |  1 +
 .../Graphics/src/BufferPool.cpp               | 12 ++++++---
 .../prx/libSceAgcDriver/tests/BufferPool.cpp  | 27 +++++++++++++++++++
 3 files changed, 37 insertions(+), 3 deletions(-)

diff --git a/core/libs/prx/libSceAgcDriver/Graphics/include/BufferPool.hpp b/core/libs/prx/libSceAgcDriver/Graphics/include/BufferPool.hpp
index 657c5351c4..22b8c85f69 100644
--- a/core/libs/prx/libSceAgcDriver/Graphics/include/BufferPool.hpp
+++ b/core/libs/prx/libSceAgcDriver/Graphics/include/BufferPool.hpp
@@ -178,6 +178,7 @@ private:
     static constexpr std::size_t classLimit = std::size_t{1} << 20u;
     static constexpr unsigned deviceClassBits = 3;
     static constexpr std::size_t defaultSlots = 4096;
+    static constexpr std::size_t exactSizeScan = 64;
 };
 
 std::shared_ptr<BufferPool> GetBufferPool(const Context& context);
diff --git a/core/libs/prx/libSceAgcDriver/Graphics/src/BufferPool.cpp b/core/libs/prx/libSceAgcDriver/Graphics/src/BufferPool.cpp
index 128970a898..eef879399f 100644
--- a/core/libs/prx/libSceAgcDriver/Graphics/src/BufferPool.cpp
+++ b/core/libs/prx/libSceAgcDriver/Graphics/src/BufferPool.cpp
@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <iterator>
 #include <limits>
 
 namespace AgcDriver::Graphics {
@@ -409,9 +410,14 @@ std::optional<BufferAllocation> BufferPool::Take(std::size_t bytes, VkBufferUsag
         ++tier.misses;
         return std::nullopt;
     }
-    auto result = found->second.back().allocation;
-    found->second.pop_back();
-    if (found->second.empty()) tier.free.erase(found);
+    auto& slots = found->second;
+    const auto newest = std::prev(slots.end());
+    auto chosen = newest;
+    for (std::size_t scanned = 1; chosen->allocation.bufferBytes != bytes && scanned < exactSizeScan && chosen != slots.begin(); ++scanned) --chosen;
+    if (chosen->allocation.bufferBytes != bytes) chosen = newest;
+    auto result = chosen->allocation;
+    slots.erase(chosen);
+    if (slots.empty()) tier.free.erase(found);
     --tier.slots;
     tier.retainedBytes -= result.allocationBytes;
     ++tier.hits;
diff --git a/core/libs/prx/libSceAgcDriver/tests/BufferPool.cpp b/core/libs/prx/libSceAgcDriver/tests/BufferPool.cpp
index abdf8f0166..4d63ee7125 100644
--- a/core/libs/prx/libSceAgcDriver/tests/BufferPool.cpp
+++ b/core/libs/prx/libSceAgcDriver/tests/BufferPool.cpp
@@ -349,6 +349,32 @@ void LargerClass() {
     Expect(MemoryOf(best.Handle()) != twelve && ClassOf(best.Handle()) == 5 * MiB, "a 4 MiB + 1 request did not take the smallest retained class that fits (5 MiB)");
 }
 
+void ExactSizeReuse() {
+    mock = MockDevice{};
+    {
+        auto context = mockContext();
+        VkBuffer first = VK_NULL_HANDLE;
+        VkBuffer second = VK_NULL_HANDLE;
+        {
+            auto a = std::make_unique<Buffer>(context, 3000, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, HostVisible);
+            auto b = std::make_unique<Buffer>(context, 4000, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, HostVisible);
+            first = a->Handle();
+            second = b->Handle();
+            a.reset();
+            b.reset();
+        }
+        const auto destroyed = mock.destroyedBuffers;
+        {
+            Buffer again(context, 3000, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, HostVisible);
+            Expect(again.Handle() == first && mock.destroyedBuffers == destroyed, "a pooled buffer of the requested exact size was passed over for a newer one that had to be rebuilt");
+            Expect(mock.sizes.at(again.Handle()) == 3000 && again.Bytes().size() == 3000, "a reused buffer does not have the requested size");
+        }
+        Buffer other(context, 4000, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, HostVisible);
+        Expect(other.Handle() == second && mock.sizes.at(other.Handle()) == 4000, "the other exact size was not kept for its own request");
+    }
+    Expect(mock.liveBytes == 0 && mock.liveBuffers.empty(), "exact-size reuse leaked Vulkan objects");
+}
+
 void KeptUntilFence() {
     mock = MockDevice{};
     auto context = mockContext();
@@ -1019,6 +1045,7 @@ int main(int argc, char** argv) {
         Expect(mock.liveBytes == 0, "slab return test leaked memory");
         SizeClassesAndDirections();
         LargerClass();
+        ExactSizeReuse();
         KeptUntilFence();
         SmallDeviceRetention();
         Budget();