v4.3 blend modes added

5 files changed, 81 insertions, 17 deletions

diff --git a/src/gl_calls.cpp b/src/gl_calls.cpp
index 82e8be2..68380aa 100644
--- a/src/gl_calls.cpp
+++ b/src/gl_calls.cpp
@@ -4,6 +4,9 @@
 
 #include "gl_calls.h"
 
+
+// TODO(fox): Combine into one after fixing Mac/Windows blend mode slowdown.
+
 const char *DefaultVertexShaderSource = "#version 330 core\n"
 "layout (location = 0) in vec3 Point;\n"
 "layout (location = 1) in vec2 aTexCoord;\n"
@@ -52,11 +55,6 @@ const char *DefaultFragmentShaderSource = "#version 330 core\n"
 "   FragColor = vec4(InputCol, Col.a);\n"
 "} else if (FragmentMode == 2) {\n"
 "   FragColor = Col;\n"
-"} else {\n"
-"   vec4 Dest = texture(Texture1, TexCoord);\n"
-"   FragColor = Dest + Col;\n"
-"   FragColor = ((1.0f - Dest * 2) * Col * Col) + (Dest * 2 * Col);\n"
-"   FragColor.a = Col.a;\n"
 "}\n"
 "}\0";
 
@@ -111,12 +109,15 @@ const char *BlendVertexShaderSource = "#version 330 core\n"
 "}\0";
 #endif
 
+// NOTE(fox): Is this slower than using multiple fragment shaders with less source code?
+// TODO(fox): Multiply and add can be optimized without the texture copy.
 const char *BlendFragmentShaderSource = "#version 330 core\n"
 "out vec4 FragColor;\n"
 "in vec2 TexCoord;\n"
 "in vec2 TexCoordBlend;\n"
 "uniform sampler2D Texture;\n"
 "uniform sampler2D Texture1;\n"
+"uniform int BlendMode;\n"
 "uniform vec3 InputCol;\n"
 "uniform vec2 ScreenDimensions;\n"
 "void main()\n"
@@ -124,12 +125,68 @@ const char *BlendFragmentShaderSource = "#version 330 core\n"
 " vec2 NewCoord = gl_FragCoord.xy / ScreenDimensions;\n"
 " vec4 Col = texture(Texture, TexCoord);\n"
 " vec4 Dest = texture(Texture1, NewCoord);\n"
-" FragColor = Dest + Col;\n"
-" FragColor = ((1.0f - Dest * 2) * Col * Col) + (Dest * 2 * Col);\n"
+" vec4 Blend = Dest;\n"
+" FragColor.a = Dest.a;\n"
+" switch (BlendMode)\n"
+" {\n"
+"   case 0:\n"
+"   {\n"
+"       FragColor.a = Col.a;\n"
+"   } break;\n"
+"   case 1:\n"
+"   {\n"
+"       Blend = Dest * Col;\n"
+"   } break;\n"
+"   case 2:\n"
+"   {\n"
+"       Blend = 1.0f - ((1.0f - Dest) / Col);\n"
+"   } break;\n"
+"   case 3:\n"
+"   {\n"
+"       Blend = (Dest + Col) - 1.0f;\n"
+"   } break;\n"
+"   case 4:\n"
+"   {\n"
+"       Blend = Dest + Col;\n"
+"   } break;\n"
+"   case 5:\n"
+"   {\n"
+"       Blend  = 1.0f - ((1.0f - Dest) * (1.0f - Col));\n"
+"   } break;\n"
+"   case 6:\n"
+"   {\n"
+"       Blend = 2.0f * Dest * Col;\n"
+"       vec4 BlendSecond = 1.0f - (2.0f * (1.0f - Dest) * (1.0f - Col));\n"
+"       Blend = mix(BlendSecond, Blend, lessThan(Dest, vec4(0.5)));\n"
+"   } break;\n"
+"   case 7:\n"
+"   {\n"
+"       Blend = ((1.0f - Col * 2) * Dest * Dest) + (Col * 2 * Dest);\n"
+"   } break;\n"
+"   case 8:\n"
+"   {\n"
+"       Blend = 2.0f * Dest * Col;\n"
+"       vec4 BlendSecond = 1.0f - (2.0f * (1.0f - Dest) * (1.0f - Col));\n"
+"       Blend = mix(BlendSecond, Blend, greaterThan(Dest, vec4(0.5)));\n"
+"   } break;\n"
+"   case 9:\n"
+"   {\n"
+"       Blend = Dest - Col;\n"
+"   } break;\n"
+"   case 10:\n"
+"   {\n"
+"       Blend = Dest / (Col + 0.001f);\n"
+"   } break;\n"
+"   case 11:\n"
+"   {\n"
+"       Blend = Col - Dest;\n"
+"       vec4 BlendSecond = Dest - Col;\n"
+"       Blend = mix(BlendSecond, Blend, greaterThan(Col - Dest, vec4(0.0f)));\n"
+"   } break;\n"
+" }\n"
+" FragColor.rgb = Blend.rgb ;\n"
 "}\0";
 
-// " vec2 NewCoord = vec2(0.5, 0.5) + vec2(XAxis + YAxis);\n"
-
 static void GL_InitDefaultShader(uint32 *VertexShader,   const char *VertexShaderSource, 
                                  uint32 *FragmentShader, const char *FragmentShaderSource,
                                  uint32 *ShaderProgram) 
@@ -305,7 +362,7 @@ GL_BlitStencil(gl_effect_layer *TestM, void *StrokeData, void *FillData, uint32
 static void
 GL_RasterizeShape2(gl_effect_layer *TestM, gl_effect_layer *TestM2, void *StrokeData, void *FillData, uint32 StrokeCount, uint32 FillCount,
                   layer_transforms T, int Width, int Height, int BytesPerPixel, void *Bitmap,
-                  int L_Width, int L_Height,v4 StrokeCol, v4 FillCol, int RenderMode, int Vector,
+                  int L_Width, int L_Height,v4 StrokeCol, v4 FillCol, blend_mode BlendMode, int RenderMode, int Vector,
                   ImVec2 ViewportSize, ImVec2 UIPos, ImVec2 UIZoom, v2 BlendMin, v2 BlendMax, int StencilLayer)
 {
     int Uniform = 0;
@@ -444,6 +501,7 @@ GL_RasterizeShape2(gl_effect_layer *TestM, gl_effect_layer *TestM2, void *Stroke
             glBindTexture(GL_TEXTURE_2D, TestM2->Texture);
             glUniform1i(glGetUniformLocation(BlendShaderProgram, "Texture"),  0); 
             glUniform1i(glGetUniformLocation(BlendShaderProgram, "Texture1"), 1); 
+            glUniform1i(glGetUniformLocation(BlendShaderProgram, "BlendMode"), BlendMode); 
         } else {
             glGenTextures(1, &Texture);
             glBindTexture(GL_TEXTURE_2D, Texture);
diff --git a/src/imgui_ui.cpp b/src/imgui_ui.cpp
index 57d3f79..546392a 100644
--- a/src/imgui_ui.cpp
+++ b/src/imgui_ui.cpp
@@ -596,11 +596,12 @@ ImGui_ProcessInputs(project_data *File, project_state *State, ui *UI, memory *Me
 
     if (ImGui::IsKeyPressed(ImGuiKey_Delete))
     {
-        Bezier_Delete_Selected(File, State, Memory, 
-                Sorted.LayerArray, Sorted.CompArray,
-                Sorted.PropertyStart, Sorted.PropertyArray);
+    State->HotkeyInput = hotkey_deletelayer;
 #if 0
         if (State->FocusedWindow == focus_timeline) { 
+            Bezier_Delete_Selected(File, State, Memory, 
+                    Sorted.LayerArray, Sorted.CompArray,
+                    Sorted.PropertyStart, Sorted.PropertyArray);
         } else {
             State->HotkeyInput = hotkey_deletelayer;
         }
diff --git a/src/include/main.h b/src/include/main.h
index 46aed98..68df39d 100644
--- a/src/include/main.h
+++ b/src/include/main.h
@@ -422,6 +422,7 @@ struct gl_data
 	real32 Width;
 	real32 Height;
 	void *BitmapData;
+	blend_mode BlendMode;
 	int RenderMode;
 
 	v2 BlendMin;
diff --git a/src/main.cpp b/src/main.cpp
index 91a8ef2..894238a 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -353,7 +353,7 @@ GL_Test(const ImDrawList* parent_list, const ImDrawCmd* cmd)
             glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
             GL_RasterizeShape2(&MSBuffer, &MSBuffer2, Data->StrokeData, Data->FillData, Data->StrokeCount, Data->FillCount,
                                Data->T, RenderData->Width, RenderData->Height, RenderData->BytesPerPixel, Data->BitmapData,
-                               Data->Width, Data->Height, Data->StrokeCol, Data->FillCol, Data->RenderMode, 0,
+                               Data->Width, Data->Height, Data->StrokeCol, Data->FillCol, Data->BlendMode, Data->RenderMode, 0,
                                RenderData->ViewportSize, RenderData->UIPos, RenderData->UIZoom, Data->BlendMin, Data->BlendMax, StencilLayer);
         } else if (Data->Type == 1) {
             GL_BlitStencil(&MSBuffer, Data->StrokeData, Data->FillData, Data->StrokeCount, Data->FillCount,
@@ -709,7 +709,7 @@ Render_UI(project_data *File, project_state *State, memory *Memory, ui *UI, ImDr
 
                 *GL_Data = { 0, Data_Stroke, StrokeCount, ShapeOpt.StrokeCol, 
                              Data_Fill, NumberOfVerts, ShapeOpt.FillCol,
-                             T, Shape->Width, Shape->Height, NULL, RenderFlags };
+                             T, Shape->Width, Shape->Height, NULL, Layer->BlendMode, RenderFlags };
             } else if (Layer->IsPrecomp) {
 
                 layer_transforms NewExtraT = Layer_GetTransforms(Layer);
@@ -892,6 +892,7 @@ Render_UI(project_data *File, project_state *State, memory *Memory, ui *UI, ImDr
                     GL_Data->BlendMin = Min;
                     GL_Data->BlendMax = Max;
                 }
+                GL_Data->BlendMode = Layer->BlendMode;
                 GL_Data->BitmapData = BitmapAddress;
                 GL_Data->T = T;
                 GL_Data->RenderMode = gl_renderflag_convex | gl_renderflag_fill | gl_renderflag_texture | Blend;
@@ -1691,7 +1692,7 @@ int main(int argc, char *argv[]) {
         uint64 TotalTime = GetCPUTime() - StartTime;
         // printf("TOTAL: %.2lu, (%.2f ms) - INPUTS: %.2lu - RENDERING: %.2lu\n", PerfTime, FrameMS, InputTime, RenderTime);
 #if DEBUG
-        printf("TOTAL: %.2lu (%.2f ms) - INPUTS: %.2lu (%.2f) - RENDERING: %.2lu (%.2f)\n", TotalTime, FrameMS, InputTime, (real64)InputTime / TotalTime, RenderTime, (real64)RenderTime / TotalTime);
+        // printf("TOTAL: %.2lu (%.2f ms) - INPUTS: %.2lu (%.2f) - RENDERING: %.2lu (%.2f)\n", TotalTime, FrameMS, InputTime, (real64)InputTime / TotalTime, RenderTime, (real64)RenderTime / TotalTime);
 #endif
         // printf("TOTAL: %.2lu, (%.2f ms) - RENDERING: %.2lu\n", PerfTime, FrameMS, PerfTime);
     }
diff --git a/src/memory.cpp b/src/memory.cpp
index fc82a13..9503b00 100644
--- a/src/memory.cpp
+++ b/src/memory.cpp
@@ -291,8 +291,10 @@ void Memory_Fill(uint8 *Address_Write, uint8 *Address_Read, uint64 WriteSize, ui
 
 void Arbitrary_Zero(uint8 *Address_Write, uint64 Size)
 {
-    __m256i Zero256 = _mm256_setzero_si256();
     uint64 i = 0;
+#if ARM
+#else
+    __m256i Zero256 = _mm256_setzero_si256();
     if (Size > 64 && InstructionMode == instruction_mode_avx) {
         uint64 Size_Lane = Size - (Size % 64);
         while (i < Size_Lane) {
@@ -301,6 +303,7 @@ void Arbitrary_Zero(uint8 *Address_Write, uint64 Size)
             i += 64;
         }
     }
+#endif
     while (i < Size) {
         *(Address_Write + i) = 0;
         i++;