Upload the texture via PBOs. Slight speedup on nVidia, rather neutral on Intel.

[movit] / effect_chain.cpp

diff --git a/effect_chain.cpp b/effect_chain.cpp
index 623d64aad936e081e0b6d9e7f7ee26346a82107f..9eee1f66faaf1622715002d44edd195708cfc9fc 100644 (file)
--- a/effect_chain.cpp
+++ b/effect_chain.cpp
@@ -176,8 +176,8 @@ void EffectChain::finalize()
        printf("%s\n", frag_shader.c_str());
        
        glsl_program_num = glCreateProgram();
-       GLhandleARB vs_obj = compile_shader(read_file("vs.glsl"), GL_VERTEX_SHADER);
-       GLhandleARB fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
+       GLuint vs_obj = compile_shader(read_file("vs.glsl"), GL_VERTEX_SHADER);
+       GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
        glAttachShader(glsl_program_num, vs_obj);
        check_error();
        glAttachShader(glsl_program_num, fs_obj);
@@ -185,6 +185,46 @@ void EffectChain::finalize()
        glLinkProgram(glsl_program_num);
        check_error();
 
+       // Translate the format to OpenGL's enums.
+       GLenum internal_format;
+       if (use_srgb_texture_format) {
+               internal_format = GL_SRGB8;
+       } else {
+               internal_format = GL_RGBA8;
+       }
+       if (input_format.pixel_format == FORMAT_RGB) {
+               format = GL_RGB;
+               bytes_per_pixel = 3;
+       } else if (input_format.pixel_format == FORMAT_RGBA) {
+               format = GL_RGBA;
+               bytes_per_pixel = 4;
+       } else if (input_format.pixel_format == FORMAT_BGR) {
+               format = GL_BGR;
+               bytes_per_pixel = 3;
+       } else if (input_format.pixel_format == FORMAT_BGRA) {
+               format = GL_BGRA;
+               bytes_per_pixel = 4;
+       } else {
+               assert(false);
+       }
+
+       // Create PBO to hold the texture, and then the texture itself.
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 2);
+       check_error();
+       glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, width * height * bytes_per_pixel, NULL, GL_STREAM_DRAW);
+       check_error();
+
+       void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
+       memset(mapped_pbo, 0, width * height * bytes_per_pixel);
+       glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
+       
+       glBindTexture(GL_TEXTURE_2D, SOURCE_IMAGE);
+       check_error();
+       glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
+       check_error();
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
+       check_error();
+
        finalized = true;
 }
 
@@ -192,25 +232,27 @@ void EffectChain::render_to_screen(unsigned char *src)
 {
        assert(finalized);
 
+       // Copy the pixel data into the PBO.
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 2);
        check_error();
-       glUseProgram(glsl_program_num);
+       void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
+       memcpy(mapped_pbo, src, width * height * bytes_per_pixel);
+       glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
        check_error();
 
+       // Re-upload the texture from the PBO.
        glActiveTexture(GL_TEXTURE0);
+       check_error();
        glBindTexture(GL_TEXTURE_2D, SOURCE_IMAGE);
+       check_error();
+       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
+       check_error();
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
+       check_error();
 
-       GLenum internal_format = GL_RGBA8;
-       if (use_srgb_texture_format) {
-               internal_format = GL_SRGB8;
-       }
+       glUseProgram(glsl_program_num);
+       check_error();
 
-       if (input_format.pixel_format == FORMAT_RGB) {
-               glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, GL_RGB, GL_UNSIGNED_BYTE, src);
-       } else if (input_format.pixel_format == FORMAT_RGBA) {
-               glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, src);
-       } else {
-               assert(false);
-       }
        check_error();
        glUniform1i(glGetUniformLocation(glsl_program_num, "input_tex"), 0);