From 82210ab480c1e508de69722b7391ed2916ae2fe5 Mon Sep 17 00:00:00 2001 From: jphalimi Date: Wed, 23 Nov 2016 20:10:34 -0800 Subject: Cache Vertices instead of Output registers (#2165) This patch brings +3% performance improvement on average. It removes ToVertex() as an important hotspot of the emulator. --- src/video_core/command_processor.cpp | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) (limited to 'src/video_core/command_processor.cpp') diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp index 6bd5b281c..b7c32035e 100644 --- a/src/video_core/command_processor.cpp +++ b/src/video_core/command_processor.cpp @@ -235,7 +235,8 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) { // 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 vertex_cache_ids; - std::array vertex_cache; + std::array vertex_cache; + Shader::OutputVertex output_vertex; unsigned int vertex_cache_pos = 0; vertex_cache_ids.fill(-1); @@ -265,7 +266,7 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) { for (unsigned int i = 0; i < VERTEX_CACHE_SIZE; ++i) { if (vertex == vertex_cache_ids[i]) { - output_registers = vertex_cache[i]; + output_vertex = vertex_cache[i]; vertex_cache_hit = true; break; } @@ -284,16 +285,16 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) { g_state.vs.Run(shader_unit, input, loader.GetNumTotalAttributes()); output_registers = shader_unit.output_registers; + // Retrieve vertex from register data + output_vertex = output_registers.ToVertex(regs.vs); + if (is_indexed) { - vertex_cache[vertex_cache_pos] = output_registers; + vertex_cache[vertex_cache_pos] = output_vertex; vertex_cache_ids[vertex_cache_pos] = vertex; vertex_cache_pos = (vertex_cache_pos + 1) % VERTEX_CACHE_SIZE; } } - // Retrieve vertex from register data - Shader::OutputVertex output_vertex = output_registers.ToVertex(regs.vs); - // Send to renderer using Pica::Shader::OutputVertex; auto AddTriangle = [](const OutputVertex& v0, const OutputVertex& v1, -- cgit v1.2.3