#if SPECIAL
#include "main.h"
#endif

#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"
"out vec2 TexCoord;\n"
"uniform int VertexMode;\n"
"uniform vec2 CompDimensions;\n"
"uniform vec2 LayerDimensions;\n"
"uniform vec2 ScreenDimensions;\n"
"uniform vec2 Pos;\n"
"uniform vec2 UIOffset;\n"
"uniform vec2 UIZoom;\n"
"uniform vec2 Anchor;\n"
"uniform float Rad;\n"
"uniform float Scale;\n"
"void main()\n"
"{\n"
" if (VertexMode == 0) {\n"
"   gl_Position = vec4(Point, 1.0);\n"
"} else {\n"
"   vec2 XRotation = vec2(cos(Rad),  sin(Rad));\n"
"   vec2 YRotation = vec2(sin(Rad), -cos(Rad));\n"
"   vec2 XAxis = (Point.x - (Anchor.x *  LayerDimensions.x))  * Scale * XRotation;\n"
"   vec2 YAxis = (Point.y - (Anchor.y * LayerDimensions.y)) * -Scale * YRotation;\n"
"   vec2 CompPoint = Pos + vec2(XAxis + YAxis);\n"
"   vec2 CompUV = CompPoint / CompDimensions;\n"
"   vec2 ScreenPoint = UIOffset + (CompUV * UIZoom);\n"
"   vec2 ScreenUV = ScreenPoint / ScreenDimensions;\n"
"   vec2 GL_Space = ScreenUV * 2 - vec2(1.0f, 1.0f);\n"
"   gl_Position = vec4(vec2(GL_Space.x, -GL_Space.y), 0.0f, 1.0);\n"
"}\n"
" TexCoord = aTexCoord;\n"
"}\0";
const char *DefaultFragmentShaderSource = "#version 330 core\n"
"out vec4 FragColor;\n"
"in vec2 TexCoord;\n"
"uniform sampler2D Texture;\n"
"uniform sampler2D Texture1;\n"
"uniform int FragmentMode;\n"
"uniform vec3 InputCol;\n"
"void main()\n"
"{\n"
"vec4 Col = texture(Texture, TexCoord);\n"
" if (FragmentMode == 0) {\n"
"   FragColor = Col;\n"
"} else if (FragmentMode == 1) {\n"
"   FragColor = vec4(InputCol, Col.a);\n"
"} else if (FragmentMode == 2) {\n"
"   FragColor = Col;\n"
"}\n"
"}\0";

#if 1
const char *BlendVertexShaderSource = "#version 330 core\n"
"layout (location = 0) in vec2 Point;\n"
"layout (location = 1) in vec2 aTexCoord;\n"
"layout (location = 2) in vec2 aTexCoordBlend;\n"
"out vec2 TexCoord;\n"
"out vec2 TexCoordBlend;\n"
"uniform vec2 CompDimensions;\n"
"uniform vec2 ScreenDimensions;\n"
"uniform vec2 UIOffset;\n"
"uniform vec2 UIZoom;\n"
"void main()\n"
"{\n"
" vec2 GL_Space = Point * 2 - vec2(1.0f, 1.0f);\n"
" gl_Position = vec4(vec2(GL_Space.x, -GL_Space.y), 0.0f, 1.0);\n"
" TexCoord = aTexCoord;\n"
" TexCoordBlend = aTexCoordBlend;\n"
"}\0";
#else
const char *BlendVertexShaderSource = "#version 330 core\n"
"layout (location = 0) in vec2 Point;\n"
"layout (location = 1) in vec2 aTexCoord;\n"
"layout (location = 2) in vec2 aTexCoordBlend;\n"
"out vec2 TexCoord;\n"
"out vec2 TexCoordBlend;\n"
"uniform vec2 CompDimensions;\n"
"uniform vec2 LayerDimensions;\n"
"uniform vec2 ScreenDimensions;\n"
"uniform vec2 Pos;\n"
"uniform vec2 UIOffset;\n"
"uniform vec2 UIZoom;\n"
"uniform vec2 Anchor;\n"
"uniform float Rad;\n"
"uniform float Scale;\n"
"void main()\n"
"{\n"
" vec2 XRotation = vec2(cos(Rad),  sin(Rad));\n"
" vec2 YRotation = vec2(sin(Rad), -cos(Rad));\n"
" vec2 XAxis = (Point.x - (Anchor.x *  LayerDimensions.x))  * Scale * XRotation;\n"
" vec2 YAxis = (Point.y - (Anchor.y * LayerDimensions.y)) * -Scale * YRotation;\n"
" vec2 CompPoint = Pos + vec2(XAxis + YAxis);\n"
" vec2 CompUV = CompPoint / CompDimensions;\n"
" vec2 ScreenPoint = UIOffset + (CompUV * UIZoom);\n"
" vec2 ScreenUV = ScreenPoint / ScreenDimensions;\n"
" vec2 GL_Space = ScreenUV * 2 - vec2(1.0f, 1.0f);\n"
" gl_Position = vec4(vec2(GL_Space.x, -GL_Space.y), 0.0f, 1.0);\n"
" TexCoord = aTexCoord;\n"
" TexCoordBlend = aTexCoordBlend;\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"
"{\n"
" vec2 NewCoord = gl_FragCoord.xy / ScreenDimensions;\n"
" vec4 Col = texture(Texture, TexCoord);\n"
" vec4 Dest = texture(Texture1, NewCoord);\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";

static void GL_InitDefaultShader(uint32 *VertexShader,   const char *VertexShaderSource, 
                                 uint32 *FragmentShader, const char *FragmentShaderSource,
                                 uint32 *ShaderProgram) 
{
    *VertexShader = glCreateShader(GL_VERTEX_SHADER);

    glShaderSource(*VertexShader, 1, &VertexShaderSource, NULL);
    glCompileShader(*VertexShader);

    int success;
    char infoLog[512];
    glGetShaderiv(*VertexShader, GL_COMPILE_STATUS, &success);

    if(!success)
    {
        glGetShaderInfoLog(*VertexShader, 512, NULL, infoLog);
        printf("Vertex shader fail:\n %s", infoLog);
    }

    *FragmentShader = glCreateShader(GL_FRAGMENT_SHADER);

    glShaderSource(*FragmentShader, 1, &FragmentShaderSource, NULL);
    glCompileShader(*FragmentShader);

    glGetShaderiv(*FragmentShader, GL_COMPILE_STATUS, &success);

    if(!success)
    {
        glGetShaderInfoLog(*FragmentShader, 512, NULL, infoLog);
        printf("Fragment shader fail:\n %s", infoLog);
    }

    // Shader programs link both types of shaders together.
    *ShaderProgram = glCreateProgram();

    glAttachShader(*ShaderProgram, *VertexShader);
    glAttachShader(*ShaderProgram, *FragmentShader);
    glLinkProgram(*ShaderProgram);

    glGetProgramiv(*ShaderProgram, GL_LINK_STATUS, &success);
    if(!success) {
        glGetProgramInfoLog(*ShaderProgram, 512, NULL, infoLog);
        printf("Shader linkage fail:\n %s", infoLog);
    }
}

static void GL_InitDefaultVerts() {

    unsigned int GLIndices[] = {
        0, 1, 3,
        1, 2, 3
    };

    glGenVertexArrays(1, &DefaultVerts.VertexArrayObject);
    glGenBuffers(1, &DefaultVerts.ElementBufferObject);

    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, DefaultVerts.ElementBufferObject);
    glBufferData(GL_ELEMENT_ARRAY_BUFFER, sizeof(GLIndices), GLIndices,
                 GL_STATIC_DRAW);
    glGenBuffers(1, &DefaultVerts.VertexBufferObject);

    glBindBuffer(GL_ARRAY_BUFFER, DefaultVerts.VertexBufferObject);
    glBufferData(GL_ARRAY_BUFFER, sizeof(GL_DefaultVertices), GL_DefaultVertices, GL_STATIC_DRAW);

    glGenVertexArrays(1, &ShapeVerts.VertexArrayObject);
    glGenBuffers(1, &ShapeVerts.VertexBufferObject);
}

void
GL_GenAndBindTexture(GLuint *GLTexture, int Width, int Height, int BytesPerPixel, void *BufferAddress)
{
    glGenTextures(1, GLTexture);
    glBindTexture(GL_TEXTURE_2D, *GLTexture);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
    int ByteFlag = (BytesPerPixel == 4) ? GL_RGBA : GL_RGBA16;
    int ByteFlag2 = (BytesPerPixel == 4) ? GL_UNSIGNED_BYTE : GL_UNSIGNED_SHORT;
    glTexImage2D(GL_TEXTURE_2D, 0, ByteFlag, Width, Height, 0, GL_RGBA, ByteFlag2, BufferAddress);
}

static void
GL_BindDefaultVertexArrays()
{
    glBindVertexArray(DefaultVerts.VertexArrayObject);
    // Switch to main buffer
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, DefaultVerts.ElementBufferObject);
    glBindBuffer(GL_ARRAY_BUFFER, DefaultVerts.VertexBufferObject);
    glBufferData(GL_ARRAY_BUFFER, sizeof(GL_DefaultVertices), GL_DefaultVertices, GL_STATIC_DRAW);
    // position attribute
    glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 5 * sizeof(float), (void*)0);
    glEnableVertexAttribArray(0);
    // texture coordinate (note the last parameter's offset)
    glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 5 * sizeof(float), (void*)(3 * sizeof(float)));
    glEnableVertexAttribArray(1);

}

void
GL_InitHWBuffer(gl_effect_layer *Test)
{
    glGenFramebuffers(1, &Test->FramebufferObject);
    glGenTextures(1, &Test->Texture);
    glGenRenderbuffers(1, &Test->Color_Renderbuffer);
    glGenRenderbuffers(1, &Test->Stencil_Renderbuffer);
    Test->Initialized = true;
}

void
GL_DeleteHWBuffer(gl_effect_layer *Test)
{
    glDeleteFramebuffers(1, &Test->FramebufferObject);
    glDeleteTextures(1, &Test->Texture);
    glDeleteRenderbuffers(1, &Test->Color_Renderbuffer);
    glDeleteRenderbuffers(1, &Test->Stencil_Renderbuffer);
    Test->Initialized = true;
}

static void
GL_BlitStencil(gl_effect_layer *TestM, void *StrokeData, void *FillData, uint32 StrokeCount, uint32 FillCount,
                  layer_transforms T, int Width, int Height, int BytesPerPixel, 
                  int L_Width, int L_Height,v4 StrokeCol, v4 FillCol, int RenderMode, int Vector,
                  ImVec2 ViewportSize, ImVec2 UIPos, ImVec2 UIZoom, int StencilLayer, int Mode)
{
    int Uniform = 0;

    Uniform = glGetUniformLocation(DefaultShaderProgram, "CompDimensions");
    glUniform2f(Uniform, Width, Height);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "LayerDimensions");
    glUniform2f(Uniform, L_Width, L_Height);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "ScreenDimensions");
    glUniform2f(Uniform, ViewportSize.x, ViewportSize.y);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "UIOffset");
    glUniform2f(Uniform, UIPos.x, UIPos.y);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "UIZoom");
    glUniform2f(Uniform, UIZoom.x, UIZoom.y);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "Pos");
    glUniform2f(Uniform, T.x, T.y);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "Anchor");
    glUniform2f(Uniform, T.ax, T.ay);
    real32 Rad = (T.rotation * (PI / 180));
    Uniform = glGetUniformLocation(DefaultShaderProgram, "Rad");
    glUniform1f(Uniform, Rad);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "Scale");
    glUniform1f(Uniform, T.scale);

    Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
    glUniform1i(Uniform, 1);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
    glUniform1i(Uniform, 1);
    Uniform = glGetUniformLocation(DefaultShaderProgram, "InputCol");
    glUniform3f(Uniform, 0, 0, 0);

    glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
    glBindVertexArray(ShapeVerts.VertexArrayObject);
    glStencilFunc(GL_EQUAL, StencilLayer, 0xFF);
    glStencilOp(GL_KEEP, GL_KEEP, Mode);

    glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
    glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * 4 * FillCount, FillData, GL_STATIC_DRAW);
    glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)0);
    glEnableVertexAttribArray(0);
    glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)(2 * sizeof(float)));

	glDrawArrays(GL_TRIANGLE_FAN, 0, FillCount);

	glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
    glBindVertexArray(0);

}

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, blend_mode BlendMode, int RenderMode, int Vector,
                  ImVec2 ViewportSize, ImVec2 UIPos, ImVec2 UIZoom, v2 BlendMin, v2 BlendMax, int StencilLayer)
{
    int Uniform = 0;

    uint32 ShaderProgram = DefaultShaderProgram;
    if (RenderMode & gl_renderflag_blend) {
        ShaderProgram = BlendShaderProgram;
        glUseProgram(BlendShaderProgram);
    } 
    Uniform = glGetUniformLocation(ShaderProgram, "CompDimensions");
    glUniform2f(Uniform, Width, Height);
    Uniform = glGetUniformLocation(ShaderProgram, "ScreenDimensions");
    glUniform2f(Uniform, ViewportSize.x, ViewportSize.y);
    Uniform = glGetUniformLocation(ShaderProgram, "UIOffset");
    glUniform2f(Uniform, UIPos.x, UIPos.y);
    Uniform = glGetUniformLocation(ShaderProgram, "UIZoom");
    glUniform2f(Uniform, UIZoom.x, UIZoom.y);
    Uniform = glGetUniformLocation(ShaderProgram, "LayerDimensions");
    glUniform2f(Uniform, L_Width, L_Height);
    Uniform = glGetUniformLocation(ShaderProgram, "Pos");
    glUniform2f(Uniform, T.x, T.y);
    Uniform = glGetUniformLocation(ShaderProgram, "Anchor");
    glUniform2f(Uniform, T.ax, T.ay);
    real32 Rad = (T.rotation * (PI / 180));
    Uniform = glGetUniformLocation(ShaderProgram, "Rad");
    glUniform1f(Uniform, Rad);
    Uniform = glGetUniformLocation(ShaderProgram, "Scale");
    glUniform1f(Uniform, T.scale);

    // Concave shapes are fairly more costly than convex: we have to write to
    // the stencil buffer, but since we also are using it to mask out precomps,
    // every shape has to be drawn a second time to "clean up" the buffer.

    if (RenderMode & gl_renderflag_fill && !(RenderMode & gl_renderflag_convex))
    {
        // disable color component writing and allow stencil writing using the shape layer's vertices
        glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
        glBindVertexArray(ShapeVerts.VertexArrayObject);
        glStencilFunc(GL_ALWAYS, 0, 0xFF);

	    glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_INCR_WRAP);
	    glStencilOpSeparate(GL_BACK,  GL_KEEP, GL_KEEP, GL_DECR_WRAP);

        Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
        glUniform1i(Uniform, 1);

        glBindVertexArray(ShapeVerts.VertexArrayObject);

        glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * 4 * FillCount, FillData, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)(2 * sizeof(float)));

	    glDrawArrays(GL_TRIANGLE_FAN, 0, FillCount);

        // ---

        // allow color component writing 
	    glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
        glBindVertexArray(DefaultVerts.VertexArrayObject);
        glStencilFunc(GL_EQUAL, StencilLayer + 1, 0xFF);
	    glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);

        Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
        glUniform1i(Uniform, 0);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "InputCol");
        glUniform3f(Uniform, FillCol.r, FillCol.g, FillCol.b);

        glBindBuffer(GL_ARRAY_BUFFER, DefaultVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(GL_DefaultVertices), GL_DefaultVertices, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 5 * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 5 * sizeof(float), (void*)(3 * sizeof(float)));
        glEnableVertexAttribArray(1);

        glDrawElements(GL_TRIANGLE_STRIP, 6, GL_UNSIGNED_INT, 0);

        // ---

        // TODO(fox): If this is the final method, add this optimization: shapes with
        // no precomps below them don't have to be drawn twice and can instead
        // get cleaned by setting StencilOp to GL_DECR. 

        glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
        glBindVertexArray(DefaultVerts.VertexArrayObject);
        glStencilFunc(GL_ALWAYS, 0, 0xFF);

	    glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_DECR_WRAP);
	    glStencilOpSeparate(GL_BACK,  GL_KEEP, GL_KEEP, GL_INCR_WRAP);

        Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
        glUniform1i(Uniform, 1);

        glBindVertexArray(ShapeVerts.VertexArrayObject);

        glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * 4 * FillCount, FillData, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)(2 * sizeof(float)));

	    glDrawArrays(GL_TRIANGLE_FAN, 0, FillCount);
	    glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
    } 
    else if (RenderMode & gl_renderflag_texture) 
    {
        GLuint Texture;
        int InitialTex1 = 0;
        if (RenderMode & gl_renderflag_blend) {
            
            glGenTextures(1, &Texture);
            glActiveTexture(GL_TEXTURE0);
            glBindTexture(GL_TEXTURE_2D, Texture);
            glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, Width, Height, 0, GL_RGBA,
                         GL_UNSIGNED_BYTE, Bitmap);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);

            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, TestM2->FramebufferObject);
            glBlitFramebuffer(0, 0, 3200, 1720,
                              0, 0, 3200, 1720,
                              GL_COLOR_BUFFER_BIT, GL_LINEAR);
            glBindFramebuffer(GL_FRAMEBUFFER, TestM->FramebufferObject);

            glGetIntegerv(GL_TEXTURE_BINDING_2D, &InitialTex1);
            glActiveTexture(GL_TEXTURE1);
            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);
            glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, Width, Height, 0, GL_RGBA,
                    GL_UNSIGNED_BYTE, Bitmap);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
        }

        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
        glBindVertexArray(ShapeVerts.VertexArrayObject);
        glStencilFunc(GL_EQUAL, StencilLayer, 0xFF);
	    glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);

        int VertRowSize = 4;
        if (!(RenderMode & gl_renderflag_blend)) {
            Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
            glUniform1i(Uniform, 1);
            Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
            glUniform1i(Uniform, 2);
        } else {
            VertRowSize = 6;
        }

        glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * VertRowSize * FillCount, FillData, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, VertRowSize * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, VertRowSize * sizeof(float), (void*)(2 * sizeof(float)));
        glEnableVertexAttribArray(1);
        if (RenderMode & gl_renderflag_blend) {
            glVertexAttribPointer(2, 2, GL_FLOAT, GL_FALSE, VertRowSize * sizeof(float), (void*)(4 * sizeof(float)));
            glEnableVertexAttribArray(2);
        }

	    glDrawArrays(GL_TRIANGLE_FAN, 0, FillCount);

        if (RenderMode & gl_renderflag_blend) {
            glActiveTexture(GL_TEXTURE1);
            glBindTexture(GL_TEXTURE_2D, InitialTex1);
            glUseProgram(DefaultShaderProgram);
        }
        glActiveTexture(GL_TEXTURE0);
        glDeleteTextures(1, &Texture);
    }
    else if (RenderMode & gl_renderflag_fill) 
    {
	    glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
        glBindVertexArray(ShapeVerts.VertexArrayObject);
        glStencilFunc(GL_EQUAL, StencilLayer, 0xFF);
	    glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);

        Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "InputCol");
        glUniform3f(Uniform, FillCol.r, FillCol.g, FillCol.b);

        glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * 4 * FillCount, FillData, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)(2 * sizeof(float)));

        // glDrawElements(GL_TRIANGLE_STRIP, 6, GL_UNSIGNED_INT, 0);
	    glDrawArrays(GL_TRIANGLE_FAN, 0, FillCount);
    } 
    else {
	    glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
    }

    glBindVertexArray(0);

    // stroke component
    if (RenderMode & gl_renderflag_stroke)
    {
        glBindVertexArray(ShapeVerts.VertexArrayObject);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "VertexMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "FragmentMode");
        glUniform1i(Uniform, 1);
        Uniform = glGetUniformLocation(DefaultShaderProgram, "InputCol");
        glUniform3f(Uniform, StrokeCol.r, StrokeCol.g, StrokeCol.b);

        glBindBuffer(GL_ARRAY_BUFFER, ShapeVerts.VertexBufferObject);
        glBufferData(GL_ARRAY_BUFFER, sizeof(real32) * 4 * StrokeCount, StrokeData, GL_STATIC_DRAW);
        glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)0);
        glEnableVertexAttribArray(0);
        glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (void*)(2 * sizeof(float)));
        glEnableVertexAttribArray(1);

        glDrawArrays(GL_TRIANGLE_STRIP, 0, StrokeCount);
    }
}

void
GL_UpdateTexture2(gl_effect_layer *Test, void *Data, uint16 Width, uint16 Height, uint16 BytesPerPixel, bool32 Multisample)
{
    glViewport(0, 0, Width, Height);

    // int err;
    // err = glGetError(); Assert(err == 0);

    if (!Test->Initialized) {
        GL_InitHWBuffer(Test);
    }

    GLenum Target = GL_TEXTURE_2D;
    if (Multisample)
        Target = GL_TEXTURE_2D_MULTISAMPLE;


    int Depth = 0, StencilDepth = 0;
    if (BytesPerPixel == 4) {
        Depth = GL_RGBA8;
        StencilDepth = GL_STENCIL_INDEX8;
    } else if (BytesPerPixel == 8) {
        Depth = GL_RGBA16;
        StencilDepth = GL_STENCIL_INDEX16;
    }

    glBindTexture(Target, Test->Texture);
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, Width, Height, 0, GL_RGBA, GL_UNSIGNED_BYTE, Data);
    glTexParameteri(Target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glTexParameteri(Target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
    glTexParameteri(Target, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
    glTexParameteri(Target, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
    glBindTexture(GL_TEXTURE_2D, 0);

    glBindRenderbuffer(GL_RENDERBUFFER, (GLuint)Test->Stencil_Renderbuffer );
    glRenderbufferStorage(GL_RENDERBUFFER, StencilDepth, Width, Height);

    glBindFramebuffer(GL_FRAMEBUFFER, Test->FramebufferObject);

    glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, Test->Texture, 0);
    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_RENDERBUFFER, Test->Stencil_Renderbuffer);

    GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
    if (status != GL_FRAMEBUFFER_COMPLETE) {
        printf("incomplete framebuffer");
        Assert(0);
    }

    glBindFramebuffer(GL_FRAMEBUFFER, 0);
}

void
GL_UpdateTexture(gl_effect_layer *Test, void *Data, uint16 Width, uint16 Height, uint16 BytesPerPixel, bool32 Multisample)
{
    glViewport(0, 0, Width, Height);

    // int err;
    // err = glGetError(); Assert(err == 0);

    if (!Test->Initialized) {
        GL_InitHWBuffer(Test);
    }

    GLenum Target = GL_TEXTURE_2D;
    if (Multisample)
        Target = GL_TEXTURE_2D_MULTISAMPLE;

    glBindTexture(Target, Test->Texture);

    int Depth = 0, StencilDepth = 0;
    if (BytesPerPixel == 4) {
        Depth = GL_RGBA8;
        StencilDepth = GL_STENCIL_INDEX8;
    } else if (BytesPerPixel == 8) {
        Depth = GL_RGBA16;
        StencilDepth = GL_STENCIL_INDEX16;
    }

    if (Multisample) {
        // glTexImage2DMultisample(GL_TEXTURE_2D_MULTISAMPLE, 4, GL_RGB, Width, Height, GL_TRUE);
        glBindTexture(GL_TEXTURE_2D_MULTISAMPLE, 0);

        glBindRenderbuffer(GL_RENDERBUFFER, Test->Color_Renderbuffer);

        glRenderbufferStorageMultisample(GL_RENDERBUFFER, 4, Depth, Width, Height);

        glBindRenderbuffer(GL_RENDERBUFFER, (GLuint)Test->Stencil_Renderbuffer );
        glRenderbufferStorageMultisample(GL_RENDERBUFFER, 4, StencilDepth, Width, Height );
    } else {
        glTexParameteri(Target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
        glTexParameteri(Target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
        glTexParameteri(Target, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
        glTexParameteri(Target, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
        // glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, Width, Height, 0, GL_RGBA,
        //              GL_UNSIGNED_BYTE, Data);
        glBindTexture(GL_TEXTURE_2D, 0);

        glBindRenderbuffer(GL_RENDERBUFFER, Test->Color_Renderbuffer);
        glRenderbufferStorage(GL_RENDERBUFFER, Depth, Width, Height);

        glBindRenderbuffer(GL_RENDERBUFFER, (GLuint)Test->Stencil_Renderbuffer );
        glRenderbufferStorage(GL_RENDERBUFFER, StencilDepth, Width, Height );
    }

    // glBindTexture(GL_TEXTURE_2D, Test->Texture);
    //     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, Width, Height, 0, GL_RGBA,
    //                  GL_UNSIGNED_BYTE, Data);
    // glBindTexture(GL_TEXTURE_2D, 0);

    glBindFramebuffer(GL_FRAMEBUFFER, Test->FramebufferObject);

    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, Test->Color_Renderbuffer);
    // glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, GL_RGBA, Test->Texture, 0);
    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_RENDERBUFFER, Test->Stencil_Renderbuffer);

    GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
    if (status != GL_FRAMEBUFFER_COMPLETE) {
        printf("incomplete framebuffer");
        Assert(0);
    }

    glBindFramebuffer(GL_FRAMEBUFFER, 0);
}

static uint16
Effect_GL_InitShader(const char *FragmentShaderEffectSource)
{
    glShaderSource(DefaultVertexShader, 1, &DefaultVertexShaderSource, NULL);
    glCompileShader(DefaultVertexShader);

    int success;
    char infoLog[512];
    glGetShaderiv(DefaultVertexShader, GL_COMPILE_STATUS, &success);

    if(!success)
    {
        glGetShaderInfoLog(DefaultVertexShader, 512, NULL, infoLog);
        printf("Vertex shader fail:\n %s", infoLog);
    }

    uint32 FragmentShader = glCreateShader(GL_FRAGMENT_SHADER);

    glShaderSource(FragmentShader, 1, &FragmentShaderEffectSource, NULL);
    glCompileShader(FragmentShader);

    glGetShaderiv(FragmentShader, GL_COMPILE_STATUS, &success);

    if(!success)
    {
        glGetShaderInfoLog(FragmentShader, 512, NULL, infoLog);
        printf("Fragment shader fail:\n %s", infoLog);
    }

    uint16 ShaderProgram = glCreateProgram();

    glAttachShader(ShaderProgram, DefaultVertexShader);
    glAttachShader(ShaderProgram, FragmentShader);
    glLinkProgram(ShaderProgram);

    glGetProgramiv(ShaderProgram, GL_LINK_STATUS, &success);
    if(!success) {
        glGetProgramInfoLog(ShaderProgram, 512, NULL, infoLog);
        printf("Shader linkage fail:\n %s", infoLog);
    }

    glDeleteShader(FragmentShader);

    glUseProgram(ShaderProgram);

    return ShaderProgram;
}