Remade Vertex Cache

src/video_core/CMakeLists.txt

@@ -44,6 +44,7 @@ set(HEADERS
             shader/shader_interpreter.h
             swrasterizer.h
             utils.h
+            vertex_cache.h
             vertex_loader.h
             video_core.h
             )

src/video_core/command_processor.cpp

@@ -23,6 +23,7 @@
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 #include "video_core/shader/shader.h"
+#include "video_core/vertex_cache.h"
 #include "video_core/vertex_loader.h"
 #include "video_core/video_core.h"
 
@@ -235,16 +236,8 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
 
         DebugUtils::MemoryAccessTracker memory_accesses;
 
-        // Simple circular-replacement vertex cache
-        // The size has been tuned for optimal balance between hit-rate and the cost of lookup
-        const size_t VERTEX_CACHE_SIZE = 32;
-        std::array<u16, VERTEX_CACHE_SIZE> vertex_cache_ids;
-        std::array<Shader::OutputVertex, VERTEX_CACHE_SIZE> vertex_cache;
         Shader::OutputVertex output_vertex;
 
-        unsigned int vertex_cache_pos = 0;
-        vertex_cache_ids.fill(-1);
-
         auto* shader_engine = Shader::GetEngine();
         Shader::UnitState shader_unit;
 
@@ -260,25 +253,15 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
             // the PICA supports it, and it would mess up the caching, guard against it here.
             ASSERT(vertex != -1);
 
-            bool vertex_cache_hit = false;
-
             if (is_indexed) {
                 if (g_debug_context && Pica::g_debug_context->recorder) {
                     int size = index_u16 ? 2 : 1;
                     memory_accesses.AddAccess(base_address + index_info.offset + size * index,
                                               size);
                 }
-
-                for (unsigned int i = 0; i < VERTEX_CACHE_SIZE; ++i) {
-                    if (vertex == vertex_cache_ids[i]) {
-                        output_vertex = vertex_cache[i];
-                        vertex_cache_hit = true;
-                        break;
-                    }
-                }
             }
 
-            if (!vertex_cache_hit) {
+            if (!VertexCache::contains(vertex)) {
                 // Initialize data for the current vertex
                 Shader::InputVertex input;
                 loader.LoadVertex(base_address, index, vertex, input, memory_accesses);
@@ -294,12 +277,10 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                 output_vertex = Shader::OutputVertex::FromRegisters(shader_unit.registers.output,
                                                                     regs, regs.vs.output_mask);
 
-                if (is_indexed) {
+                if (is_indexed)
-                    vertex_cache[vertex_cache_pos] = output_vertex;
+                    VertexCache::store(vertex, output_vertex);
-                    vertex_cache_ids[vertex_cache_pos] = vertex;
+            } else
-                    vertex_cache_pos = (vertex_cache_pos + 1) % VERTEX_CACHE_SIZE;
+                output_vertex = VertexCache::obtain(vertex);
-                }
-            }
 
             // Send to renderer
             using Pica::Shader::OutputVertex;
@@ -316,6 +297,8 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                                                       range.second, range.first);
         }
 
+        if (is_indexed)
+            VertexCache::clear();
         break;
     }
 

src/video_core/vertex_cache.h

@@ -0,0 +1,81 @@
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include "common/common_types.h"
+#include "video_core/shader/shader.h"
+
+// This little module emulates PICA's vertex shader's post transform vertex cache.
+namespace VertexCache {
+
+namespace {
+constexpr size_t cache_max_size = 256 * 256;
+// testing has proven that this size produces few misses in most games.
+// in total the cache size is 64Kb
+constexpr size_t cache_size = 512;
+
+// this is used to find if a vertex is in the cache
+static std::array<bool, cache_max_size> index_table = {};
+
+#ifdef _DEBUG
+static std::array<bool, cache_max_size> seen_table = {};
+u32 cache_misses = 0;
+
+// The number of misses we can tolerate.
+const u32 tolerated_misses = 16;
+#endif
+
+// le cache
+static std::array<Pica::Shader::OutputVertex, cache_size> cache;
+
+// stores positions that are cached;
+static std::array<u16, cache_size> remapper = {};
+
+// used for clearing the cache;
+u32 min_index = cache_size;
+u32 max_index = 0;
+} // Anonymous namespace
+
+inline bool contains(u32 position) {
+#ifdef _DEBUG
+    if (seen_table[position] != index_table[position])
+        cache_misses++;
+#endif
+    return index_table[position];
+}
+
+inline Pica::Shader::OutputVertex& obtain(u32 position) {
+    return cache[position % cache_size];
+}
+
+inline void store(u32 position, Pica::Shader::OutputVertex& vertex) {
+    const u32 remap_index = position % cache_size;
+    index_table[remapper[remap_index]] = false;
+    remapper[remap_index] = position;
+    index_table[position] = true;
+    cache[remap_index] = vertex;
+    max_index = std::max(max_index, remap_index);
+    min_index = std::min(min_index, remap_index);
+#ifdef _DEBUG
+    seen_table[position] = true;
+#endif
+}
+
+void clear() {
+    max_index++;
+    auto result = std::minmax_element(remapper.begin() + min_index, remapper.begin() + max_index);
+    const u32 min = *result.first;
+    const u32 max = *result.second;
+    std::fill(index_table.begin() + min, index_table.begin() + max + 1, 0);
+    max_index = 0;
+    min_index = cache_size;
+#ifdef _DEBUG
+    seen_table.fill(0);
+    if (cache_misses > tolerated_misses)
+        LOG_TRACE(HW_GPU, "The vertex cache had %d misses", cache_misses);
+    cache_misses = 0;
+#endif
+}
+
+} // VertexCache