Merge pull request #334 from snowstormxd/codex/upstream-mkw-test

MKW performance improvements

3 files changed, 22 insertions, 5 deletions

diff --git a/src/video_core/engines/draw_manager.cpp b/src/video_core/engines/draw_manager.cpp
index 974a9a6e9a..564220b71c 100644
--- a/src/video_core/engines/draw_manager.cpp
+++ b/src/video_core/engines/draw_manager.cpp
@@ -313,6 +313,8 @@ void Maxwell3D::DrawManager::UpdateTopology(Maxwell3D& maxwell3d) {
 }
 
 void Maxwell3D::DrawManager::ProcessDraw(Maxwell3D& maxwell3d, bool draw_indexed, u32 instance_count) {
+    draw_state.draw_element_count =
+        draw_indexed ? draw_state.index_buffer.count : draw_state.vertex_buffer.count;
     LOG_TRACE(HW_GPU, "called, topology={}, count={}", draw_state.topology, draw_indexed ? draw_state.index_buffer.count : draw_state.vertex_buffer.count);
     UpdateTopology(maxwell3d);
     if (maxwell3d.ShouldExecute()) {
@@ -321,6 +323,7 @@ void Maxwell3D::DrawManager::ProcessDraw(Maxwell3D& maxwell3d, bool draw_indexed
 }
 
 void Maxwell3D::DrawManager::ProcessDrawIndirect(Maxwell3D& maxwell3d) {
+    draw_state.draw_element_count = 0;
     LOG_TRACE(HW_GPU, "called, topology={}, is_indexed={}, includes_count={}, buffer_size={}, max_draw_count={}", draw_state.topology, indirect_state.is_indexed, indirect_state.include_count, indirect_state.buffer_size, indirect_state.max_draw_counts);
     UpdateTopology(maxwell3d);
     if (maxwell3d.ShouldExecute()) {
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 4ef7d42979..92313988bd 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -518,6 +518,10 @@ void Maxwell3D::DumpRecentWrites(u32 method) {
     }
     LOG_ERROR(HW_GPU, "Write to {:#x} while macro {:#x} collects parameters ({} so far); last writes:",
               method, executing_macro, macro_params.size());
+    if (!record_write_history) {
+        LOG_ERROR(HW_GPU, "Register write history disabled; EDEN_NX2_3D_HISTORY=1 enables it");
+        return;
+    }
     for (size_t i = 0; i < recent_writes.size(); ++i) {
         const auto& w = recent_writes[(recent_write_index + i) % recent_writes.size()];
         if (w.amount != 0) {
@@ -818,7 +822,6 @@ struct Maxwell3D::FrontState {
     Regs::ShadowRamControl control{};
     u32 executing_macro{};
     std::vector<u32> parameters;
-    std::vector<u32> fifo;
     u64 calls{};
     /// A method the draw manager handles was sent since the back last ran DrawDeferred: only
     /// then can it have a deferred instanced draw to issue at the end of a macro.
@@ -1095,9 +1098,9 @@ void Maxwell3D::FrontRunMacro(Core::System& system, VideoCommon::GpuSplit& split
             engine.native_jit = std::make_shared<TuringMacroJit>();
         }
         engine.native_jit->Prepare(engine.native_code, engine.native_code_generation);
-        fs.fifo.assign(fs.parameters.begin(), fs.parameters.end());
+        const size_t parameter_count = fs.parameters.size();
         const auto result = engine.native_jit->Execute(
-            start, fs.fifo, std::span<u32>{native_data_ram},
+            start, fs.parameters, std::span<u32>{native_data_ram},
             [&fs](u32 address) { return address < Regs::NUM_REGS ? fs.regs[address] : 0U; },
             [&](u32 address, u32 value) {
                 if (address >= MacroRegistersStart) {
@@ -1113,7 +1116,7 @@ void Maxwell3D::FrontRunMacro(Core::System& system, VideoCommon::GpuSplit& split
         if (!result.success) {
             engine.native_execution_failed = true;
             LOG_ERROR(HW_GPU, "Native MME (JIT, split front) failed: start={:#x}, consumed={}/{}, error={}",
-                      start, result.consumed, fs.parameters.size(), result.error);
+                      start, result.consumed, parameter_count, result.error);
         }
     }
     fs.parameters.clear();
@@ -1148,7 +1151,7 @@ void Maxwell3D::FrontDmaRead(VideoCommon::GpuSplit& split, FrontState& fs, u32 m
     for (u32 i = 0; i < words; ++i) {
         const u32 word = memory_manager.Read<u32>(source + u64{i} * sizeof(u32));
         if (method == 0x158) {
-            fs.fifo.push_back(word);
+            fs.parameters.push_back(word);
         } else if (ram + i < native_data_ram.size()) {
             native_data_ram[ram + i] = word;
         }
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index c5d1eb3230..7646ddbbe3 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -3064,6 +3064,7 @@ public:
             Maxwell3D::Regs::PrimitiveTopology topology{};
             DrawMode draw_mode{};
             bool draw_indexed{};
+            u32 draw_element_count{}; // Zero for indirect draws whose count is not known yet.
             u32 base_index{};
             Maxwell3D::Regs::VertexBuffer vertex_buffer;
             Maxwell3D::Regs::IndexBuffer index_buffer;
@@ -3352,7 +3353,17 @@ private:
     };
     std::array<RecentWrite, 48> recent_writes{};
     size_t recent_write_index = 0;
+    const bool record_write_history{
+#ifdef NDEBUG
+        std::getenv("EDEN_NX2_3D_HISTORY") != nullptr
+#else
+        true
+#endif
+    };
     void RecordWrite(u32 method, u32 argument, u32 amount, u32 pending) {
+        if (!record_write_history) {
+            return;
+        }
         recent_writes[recent_write_index++ % recent_writes.size()] = {method, argument, amount, pending};
     }
     void DumpRecentWrites(u32 method);