Skip to content

Latest commit

 

History

History
265 lines (199 loc) · 12.4 KB

debugging_gpu_related_code.md

File metadata and controls

265 lines (199 loc) · 12.4 KB

Debugging GPU related code

Chromium's GPU system is multi-process, which can make debugging it rather difficult. See GPU Command Buffer for some of the nitty gitty. These are just a few notes to help with debugging.

[TOC]

Renderer Process Code

--enable-gpu-client-logging

If you are trying to track down a bug in a GPU client process (compositing, WebGL, Skia/Ganesh, Aura), then in a debug build you can use the --enable-gpu-client-logging flag, which will show every GL call sent to the GPU service process. (From the point of view of a GPU client, it's calling OpenGL ES functions - but the real driver calls are made in the GPU process.)

You can also use this flag in a release build by specifying the GN argument:

enable_gpu_client_logging=true

It's typically necessary to specify the --enable-logging=stderr flag as well:

--enable-gpu-client-logging --enable-logging=stderr

The output looks like this:

[4782:4782:1219/141706:INFO:gles2_implementation.cc(1026)] [.WebGLRenderingContext] glUseProgram(3)
[4782:4782:1219/141706:INFO:gles2_implementation_impl_autogen.h(401)] [.WebGLRenderingContext] glGenBuffers(1, 0x7fffc9e1269c)
[4782:4782:1219/141706:INFO:gles2_implementation_impl_autogen.h(416)]   0: 1
[4782:4782:1219/141706:INFO:gles2_implementation_impl_autogen.h(23)] [.WebGLRenderingContext] glBindBuffer(GL_ARRAY_BUFFER, 1)
[4782:4782:1219/141706:INFO:gles2_implementation.cc(1313)] [.WebGLRenderingContext] glBufferData(GL_ARRAY_BUFFER, 36, 0x7fd268580120, GL_STATIC_DRAW)
[4782:4782:1219/141706:INFO:gles2_implementation.cc(2480)] [.WebGLRenderingContext] glEnableVertexAttribArray(0)
[4782:4782:1219/141706:INFO:gles2_implementation.cc(1140)] [.WebGLRenderingContext] glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 0, 0)
[4782:4782:1219/141706:INFO:gles2_implementation_impl_autogen.h(135)] [.WebGLRenderingContext] glClear(16640)
[4782:4782:1219/141706:INFO:gles2_implementation.cc(2490)] [.WebGLRenderingContext] glDrawArrays(GL_TRIANGLES, 0, 3)

Checking about:gpu

The GPU process logs many errors and warnings. You can see these by navigating to about:gpu. Logs appear at the bottom of the page. You can also see them on standard output if Chromium is run from the command line on Linux/Mac. On Windows, you need debugging tools (like VS, WinDbg, etc.) to connect to the debug output stream.

Note: If about:gpu is telling you that your GPU is disabled and hardware acceleration is unavailable, it might be a problem with your GPU being unsupported. To override this and turn on hardware acceleration anyway, you can use the --ignore-gpu-blocklist command line option when starting Chromium.

Breaking on GL Error

In gles2_implementation.h, there is some code like this:

// Set to 1 to have the client fail when a GL error is generated.
// This helps find bugs in the renderer since the debugger stops on the error.
#if DCHECK_IS_ON()
#if 0
#define GL_CLIENT_FAIL_GL_ERRORS
#endif
#endif

Change that #if 0 to #if 1, build a debug build, then run in a debugger. The debugger will break when any renderer code sees a GL error, and you should be able to examine the call stack to find the issue.

Labeling your calls

The output of all of the errors, warnings and debug logs are prefixed. You can set this prefix by calling glPushGroupMarkerEXT, glPopGroupMarkerEXT and glInsertEventMarkerEXT. glPushGroupMarkerEXT appends a string to the end of the current log prefix (think namespace in C++). glPopGroupmarkerEXT pops off the last string appended. glInsertEventMarkerEXT sets a suffix for the current string. Example:

glPushGroupMarkerEXT(0, "Foo");        // -> log prefix = "Foo"
glInsertEventMarkerEXT(0, "This");     // -> log prefix = "Foo.This"
glInsertEventMarkerEXT(0, "That");     // -> log prefix = "Foo.That"
glPushGroupMarkerEXT(0, "Bar");        // -> log prefix = "Foo.Bar"
glInsertEventMarkerEXT(0, "Orange");   // -> log prefix = "Foo.Bar.Orange"
glInsertEventMarkerEXT(0, "Banana");   // -> log prefix = "Foo.Bar.Banana"
glPopGroupMarkerEXT();                 // -> log prefix = "Foo.That"

Making a reduced test case.

You can often make a simple OpenGL-ES-2.0-only C++ reduced test case that is relatively quick to compile and test, by adding tests to the gl_tests target. Those tests exist in src/gpu/command_buffer/tests and are made part of the build in src/gpu/BUILD.gn. Build with ninja -C out/Debug gl_tests. All the same command line options listed on this page will work with the gl_tests, plus --gtest_filter=NameOfTest to run a specific test. Note the gl_tests are not multi-process, so they probably won't help with race conditions, but they do go through most of the same code and are much easier to debug.

Debugging the renderer process

Given that Chrome starts many renderer processes I find it's easier if I either have a remote webpage I can access or I make one locally and then use a local server to serve it like python -m SimpleHTTPServer. Then

On Linux this works for me:

  • out/Debug/chromium --no-sandbox --renderer-cmd-prefix="xterm -e gdb --args" http://localhost:8000/page-to-repro.html

On OSX this works for me:

  • out/Debug/Chromium.app/Contents/MacOSX/Chromium --no-sandbox --renderer-cmd-prefix="xterm -e gdb --args" http://localhost:8000/page-to-repro.html

On Windows I use --renderer-startup-dialog and then connect to the listed process.

Note 1: On Linux and OSX I use cgdb instead of gdb.

Note 2: GDB can take minutes to index symbol. To save time, you can precache that computation by running build/gdb-add-index out/Debug/chrome.

GPU Process Code

--enable-gpu-service-logging

In a debug build or a release build with dcheck_always_on=true in GN argument, this will print all actual calls into the GL driver.

To use it in Release builds without dcheck_always_on = true, specify GN argument enable_gpu_service_logging=true.

For non-rooted devices running production builds, we can not set the command line flags. Use about://flags 'Enable gpu service logging' instead.

[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kEnableVertexAttribArray
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(905)] glEnableVertexAttribArray(0)
[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kVertexAttribPointer
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(1573)] glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 0, 0)
[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kClear
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(746)] glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(840)] glDepthMask(GL_TRUE)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(900)] glEnable(GL_DEPTH_TEST)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(1371)] glStencilMaskSeparate(GL_FRONT, 4294967295)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(1371)] glStencilMaskSeparate(GL_BACK, 4294967295)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(860)] glDisable(GL_STENCIL_TEST)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(860)] glDisable(GL_CULL_FACE)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(860)] glDisable(GL_SCISSOR_TEST)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(900)] glEnable(GL_BLEND)
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(721)] glClear(16640)
[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kDrawArrays
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(870)] glDrawArrays(GL_TRIANGLES, 0, 3)

Note that GL calls into the driver are not currently prefixed (todo?). But, you can tell from the commands logged which command, from which context caused the following GL calls to be made.

Also note that client resource IDs are virtual IDs, so calls into the real GL driver will not match (though some commands print the mapping). Examples:

[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kBindTexture
[5497:5497:1219/142413:INFO:gles2_cmd_decoder.cc(837)] [.WebGLRenderingContext] glBindTexture: client_id = 2, service_id = 10
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(662)] glBindTexture(GL_TEXTURE_2D, 10)
[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [0052064A367F0000]cmd: kBindBuffer
[5497:5497:1219/142413:INFO:gles2_cmd_decoder.cc(837)] [0052064A367F0000] glBindBuffer: client_id = 2, service_id = 6
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(637)] glBindBuffer(GL_ARRAY_BUFFER, 6)
[5497:5497:1219/142413:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kBindFramebuffer
[5497:5497:1219/142413:INFO:gles2_cmd_decoder.cc(837)] [.WebGLRenderingContext] glBindFramebuffer: client_id = 1, service_id = 3
[5497:5497:1219/142413:INFO:gl_bindings_autogen_gl.cc(652)] glBindFramebufferEXT(GL_FRAMEBUFFER, 3)

etc... so that you can see renderer process code would be using the client IDs where as the gpu process is using the service IDs. This is useful for matching up calls if you're dumping both client and service GL logs.

--enable-gpu-debugging

In any build, this will call glGetError after each command

--enable-gpu-command-logging

This will print the name of each GPU command before it is executed.

[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kBindBuffer
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kBufferData
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: SetToken
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kEnableVertexAttribArray
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kVertexAttribPointer
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kClear
[5234:5234:1219/052139:ERROR:gles2_cmd_decoder.cc(3301)] [.WebGLRenderingContext]cmd: kDrawArrays

Debugging in the GPU Process

Given the multi-processness of chromium it can be hard to debug both sides. Turning on all the logging and having a small test case is useful. One minor suggestion, if you have some idea where the bug is happening a call to some obscure gl function like glHint() can give you a place to catch a command being processed in the GPU process (put a break point on gpu::gles2::GLES2DecoderImpl::HandleHint. Once in you can follow the commands after that. All of them go through gpu::gles2::GLES2DecoderImpl::DoCommand.

To actually debug the GPU process:

On Linux this works for me:

  • out/Debug/chromium --no-sandbox --gpu-launcher="xterm -e gdb --args" http://localhost:8000/page-to-repro.html

On OSX this works for me:

  • out/Debug/Chromium.app/Contents/MacOSX/Chromium --no-sandbox --gpu-launcher="xterm -e gdb --args" http://localhost:8000/page-to-repro.html

On Windows I use --gpu-startup-dialog and then connect to the listed process.

GPU PARSE ERROR

If you see this message in about:gpu or your console and you didn't cause it directly (by calling glLoseContextCHROMIUM) and it's something other than 5 that means there's likely a bug. Please file an issue at http://crbug.com/new.

Tracing OpenGL calls

Passing the command line flag --enable-gpu-service-tracing causes the GPU process to emit one trace event per OpenGL API call. (See "Debugging Performance", below.) This is useful when trying to understand where the expensive operations are in a given set of work sent from a renderer process to the GPU process, and processed underneath CommandBufferService::PutChanged.

Debugging Performance

If you have something to add here please add it. Most perf debugging is done using about:tracing (see Trace Event Profiling for details). Otherwise, be aware that, since the system is multi-process, calling:

start = GetTime()
DoSomething()
glFinish()
end = GetTime
printf("elapsedTime = %f\n", end - start);

will not give you meaningful results.