In the development of Shou, I've been using GLSL with NEON to manipulate image rotation, scaling and color conversion, before send them to video encoder.
So I need a very efficient way to transfer pixels between OpenGL and memory space. TheglTexImage2D
and glReadPixels
performance are very unacceptable, especially for some specific vendors, e.g. Samsung Galaxy devices with Exynos chip.
Compared to glTex(Sub)Image2D
, the glReadPixels
is the real bottleneck, which blocks all OpenGL pipeline and results in about 100ms delay for a standard 720P frame read back.
Here I will share two standard OpenGL approaches to achieve really fast pixels pack, which should be available on all OpenGL implementations. Only glReadPixels
will be discussed, as the glTexImage2D
should have the same usage.
PBO is not introduced until OpenGL ES 3.0, which is available since Android 4.3. The pixels pack operation will be reduced to about 5ms using PBO.
PBO is created just like any other buffer objects:
glGenBuffers(1, &pbo_id);
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
glBufferData(GL_PIXEL_PACK_BUFFER, pbo_size, 0, GL_DYNAMIC_READ);
glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
According to the reference of glReadPixels
:
If a non-zero named buffer object is bound to the GL_PIXEL_PACK_BUFFER target (see glBindBuffer) while a block of pixels is requested, data is treated as a byte offset into the buffer object's data store rather than a pointer to client memory.
When we need to read pixels from an FBO:
glReadBuffer(GL_COLOR_ATTACHMENT0);
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, 0);
GLubyte *ptr = glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, pbo_size, GL_MAP_READ_BIT);
memcpy(pixels, ptr, pbo_size);
glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
In a real project, we may consider using double or triple PBOs to improve the performance.
EGL_KHR_image_base is a completed EGL extension, which achieves the same performance as PBO, but only require OpenGL-ES 1.1 or 2.0.
The function to create an EGLImageKHR
is
EGLImageKHR eglCreateImageKHR(EGLDisplay dpy,
EGLContext ctx,
EGLenum target,
EGLClientBuffer buffer,
const EGLint *attrib_list)
The Android EGL implementation frameworks/native/opengl/libagl/egl.cpp implies that theEGLDisplay
should be a valid display, the EGLClientBuffer
type should be ANativeWindowBuffer
, theEGLContext
can only be EGL_NO_CONTEXT
, and the target can only be EGL_NATIVE_BUFFER_ANDROID
.
All the other parameters are obvious, except for the ANativeWindowBuffer
, which is defined insystem/core/include/system/window.h
.
To allocate an ANativeWindowBuffer
, Android has a simple wrapper called GraphicBuffer
, defined inframeworks/native/include/ui/GraphicBuffer.h
.
GraphicBuffer *window = new GraphicBuffer(width, height, PIXEL_FORMAT_RGBA_8888, GraphicBuffer::USAGE_SW_READ_OFTEN | GraphicBuffer::USAGE_HW_TEXTURE);
struct ANativeWindowBuffer *buffer = window->getNativeBuffer();
EGLImageKHR *image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_NATIVE_BUFFER_ANDROID, *attribs);
Then anytime we want to read pixels from an FBO, we should use one of the two methods below:
void EGLImageTargetTexture2DOES(enum target, eglImageOES image)
void EGLImageTargetRenderbufferStorageOES(enum target, eglImageOES image)
These two methods will establishes all the properties of the target GL_TEXTURE_2D
orGL_RENDERBUFFER
.
uint8_t *ptr;
glBindTexture(GL_TEXTURE_2D, texture_id);
glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, image);
window->lock(GraphicBuffer::USAGE_SW_READ_OFTEN, &ptr);
memcpy(pixels, ptr, width * height * 4);
window->unlock();