From 57c9fbd86e202668f371fa776c07770ff9dfceb2 Mon Sep 17 00:00:00 2001
From: Brian Lach <32017152+lachbr@users.noreply.github.com>
Date: Wed, 9 Sep 2020 09:59:40 -0400
Subject: [PATCH] glgsg: fix identical states never short-circuiting
 set_state_and_transform

Closes #1012
---
 panda/src/display/graphicsStateGuardian.cxx   |  1 +
 panda/src/dxgsg9/dxGraphicsStateGuardian9.cxx |  5 +-
 .../glstuff/glGraphicsStateGuardian_src.cxx   | 49 +++++++++++++------
 3 files changed, 40 insertions(+), 15 deletions(-)

diff --git a/panda/src/display/graphicsStateGuardian.cxx b/panda/src/display/graphicsStateGuardian.cxx
index f26104449a..b87a6401e6 100644
--- a/panda/src/display/graphicsStateGuardian.cxx
+++ b/panda/src/display/graphicsStateGuardian.cxx
@@ -2616,6 +2616,7 @@ reset() {
   _state_rs = RenderState::make_empty();
   _target_rs = nullptr;
   _state_mask.clear();
+  _inv_state_mask = RenderState::SlotMask::all_on();
   _internal_transform = _cs_transform;
   _scene_null = new SceneSetup;
   _scene_setup = _scene_null;
diff --git a/panda/src/dxgsg9/dxGraphicsStateGuardian9.cxx b/panda/src/dxgsg9/dxGraphicsStateGuardian9.cxx
index e1efd0374b..aa466201f7 100644
--- a/panda/src/dxgsg9/dxGraphicsStateGuardian9.cxx
+++ b/panda/src/dxgsg9/dxGraphicsStateGuardian9.cxx
@@ -3223,11 +3223,14 @@ set_state_and_transform(const RenderState *target,
     _state_mask.set_bit(color_blend_slot);
   }
 
-  if (_target_shader != _state_shader) {
+  int shader_slot = ShaderAttrib::get_class_slot();
+  if (_target_shader != _state_shader ||
+      !_state_mask.get_bit(shader_slot)) {
     // PStatTimer timer(_draw_set_state_shader_pcollector);
     do_issue_shader();
     _state_shader = _target_shader;
     _state_mask.clear_bit(TextureAttrib::get_class_slot());
+    _state_mask.set_bit(shader_slot);
   }
 
   int texture_slot = TextureAttrib::get_class_slot();
diff --git a/panda/src/glstuff/glGraphicsStateGuardian_src.cxx b/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
index 0c60dcf806..2e5d32b859 100644
--- a/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
+++ b/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
@@ -608,10 +608,10 @@ reset() {
   GraphicsStateGuardian::reset();
 
   // Build _inv_state_mask as a mask of 1's where we don't care, and 0's where
-  // we do care, about the state.  _inv_state_mask =
-  // RenderState::SlotMask::all_on();
+  // we do care, about the state.
+#ifndef OPENGLES_1
   _inv_state_mask.clear_bit(ShaderAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(AlphaTestAttrib::get_class_slot());
+#endif
   _inv_state_mask.clear_bit(AntialiasAttrib::get_class_slot());
   _inv_state_mask.clear_bit(ClipPlaneAttrib::get_class_slot());
   _inv_state_mask.clear_bit(ColorAttrib::get_class_slot());
@@ -621,21 +621,30 @@ reset() {
   _inv_state_mask.clear_bit(DepthTestAttrib::get_class_slot());
   _inv_state_mask.clear_bit(DepthWriteAttrib::get_class_slot());
   _inv_state_mask.clear_bit(RenderModeAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(RescaleNormalAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(ShadeModelAttrib::get_class_slot());
   _inv_state_mask.clear_bit(TransparencyAttrib::get_class_slot());
   _inv_state_mask.clear_bit(ColorWriteAttrib::get_class_slot());
   _inv_state_mask.clear_bit(ColorBlendAttrib::get_class_slot());
+#if !defined(OPENGLES) || defined(OPENGLES_1)
   _inv_state_mask.clear_bit(LogicOpAttrib::get_class_slot());
+#endif
   _inv_state_mask.clear_bit(TextureAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(TexGenAttrib::get_class_slot());
   _inv_state_mask.clear_bit(TexMatrixAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(MaterialAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(LightAttrib::get_class_slot());
   _inv_state_mask.clear_bit(StencilAttrib::get_class_slot());
-  _inv_state_mask.clear_bit(FogAttrib::get_class_slot());
   _inv_state_mask.clear_bit(ScissorAttrib::get_class_slot());
 
+  // We only care about this state if we are using the fixed-function pipeline.
+#ifdef SUPPORT_FIXED_FUNCTION
+  if (has_fixed_function_pipeline()) {
+    _inv_state_mask.clear_bit(AlphaTestAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(RescaleNormalAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(ShadeModelAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(TexGenAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(MaterialAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(LightAttrib::get_class_slot());
+    _inv_state_mask.clear_bit(FogAttrib::get_class_slot());
+  }
+#endif
+
   // Output the vendor and version strings.
   query_gl_version();
 
@@ -11392,17 +11401,27 @@ set_state_and_transform(const RenderState *target,
   _state_pcollector.add_level(1);
   PStatGPUTimer timer1(this, _draw_set_state_pcollector);
 
-  if (transform != _internal_transform) {
+  bool transform_changed = transform != _internal_transform;
+  if (transform_changed) {
     // PStatGPUTimer timer(this, _draw_set_state_transform_pcollector);
     _transform_state_pcollector.add_level(1);
     _internal_transform = transform;
     do_issue_transform();
   }
 
-  //XXX the _inv_state_mask system does not appear to be used at the moment.
-  //if (target == _state_rs && (_state_mask | _inv_state_mask).is_all_on()) {
-  //  return;
-  //}
+  if (target == _state_rs && (_state_mask | _inv_state_mask).is_all_on()) {
+#ifndef OPENGLES_1
+    if (transform_changed) {
+      // The state has not changed, but the transform has. Set the new
+      // transform on the shader, if we have one.
+      if (_current_shader_context != nullptr) {
+        _current_shader_context->set_state_and_transform(_state_rs, transform,
+          _scene_setup->get_camera_transform(), _projection_mat);
+      }
+    }
+#endif
+    return;
+  }
   _target_rs = target;
 
 #ifndef OPENGLES_1
@@ -11413,10 +11432,12 @@ set_state_and_transform(const RenderState *target,
     do_issue_shader();
     _state_shader = _target_shader;
     _state_mask.clear_bit(TextureAttrib::get_class_slot());
+    _state_mask.set_bit(ShaderAttrib::get_class_slot());
   }
   else if (!has_fixed_function_pipeline() && _current_shader == nullptr) { // In the case of OpenGL ES 2.x, we need to glUseShader before we draw anything.
     do_issue_shader();
     _state_mask.clear_bit(TextureAttrib::get_class_slot());
+    _state_mask.set_bit(ShaderAttrib::get_class_slot());
   }
 
   // Update all of the state that is bound to the shader program.