#include "glproc.hpp"
#include "glstate.hpp"
#include "glretrace.hpp"
+#include "os_time.hpp"
+/* Synchronous debug output may reduce performance however,
+ * without it the callNo in the callback may be inaccurate
+ * as the callback may be called at any time.
+ */
+#define DEBUG_OUTPUT_SYNCHRONOUS 0
namespace glretrace {
+bool insideList = false;
bool insideGlBeginEnd = false;
-glws::Profile defaultProfile = glws::PROFILE_COMPAT;
-glws::Visual *visual[glws::PROFILE_MAX];
-glws::Drawable *drawable = NULL;
-glws::Context *context = NULL;
+
+struct CallQuery
+{
+ GLuint ids[3];
+ unsigned call;
+ GLuint program;
+ const trace::FunctionSig *sig;
+ uint64_t start;
+ uint64_t duration;
+};
+
+static bool supportsElapsed = true;
+static bool supportsTimestamp = true;
+static bool supportsOcclusion = true;
+
+static bool firstFrame = true;
+static std::list<CallQuery> callQueries;
+
+static void APIENTRY
+debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam);
void
checkGlError(trace::Call &call) {
os << "\n";
}
-/**
- * Grow the current drawble.
- *
- * We need to infer the drawable size from GL calls because the drawable sizes
- * are specified by OS specific calls which we do not trace.
- */
+static GLuint64
+getGpuTimestamp() {
+ GLuint query = 0;
+ GLuint64 timestamp = 0;
+
+ if (retrace::profilingGpuTimes && supportsTimestamp) {
+ glGenQueries(1, &query);
+ glQueryCounter(query, GL_TIMESTAMP);
+ glGetQueryObjectui64vEXT(query, GL_QUERY_RESULT, ×tamp);
+ glDeleteQueries(1, &query);
+ }
+
+ return timestamp;
+}
+
+static GLuint64
+getCpuTimestamp() {
+ if (retrace::profilingCpuTimes) {
+ return os::getTime();
+ } else {
+ return 0;
+ }
+}
+
+static void
+completeCallQuery(CallQuery& query) {
+ /* Get call start and duration */
+ GLuint64 timestamp = 0, duration = 0, samples = 0;
+
+ if (retrace::profilingGpuTimes) {
+ if (supportsTimestamp) {
+ glGetQueryObjectui64vEXT(query.ids[0], GL_QUERY_RESULT, ×tamp);
+ }
+
+ if (supportsElapsed) {
+ glGetQueryObjectui64vEXT(query.ids[1], GL_QUERY_RESULT, &duration);
+ }
+ }
+
+ if (retrace::profilingPixelsDrawn && supportsOcclusion) {
+ glGetQueryObjectui64vEXT(query.ids[2], GL_QUERY_RESULT, &samples);
+ }
+
+ glDeleteQueries(3, query.ids);
+
+ /* Add call to profile */
+ retrace::profiler.addCall(query.call, query.sig->name, query.program, samples, timestamp, duration, query.start, query.duration);
+}
+
void
-updateDrawable(int width, int height) {
- if (!drawable) {
- return;
+flushQueries() {
+ for (std::list<CallQuery>::iterator itr = callQueries.begin(); itr != callQueries.end(); ++itr) {
+ completeCallQuery(*itr);
}
- if (drawable->visible &&
- width <= drawable->width &&
- height <= drawable->height) {
- return;
+ callQueries.clear();
+}
+
+void
+beginProfile(trace::Call &call) {
+ if (firstFrame) {
+ frame_start();
}
- // Ignore zero area viewports
- if (width == 0 || height == 0) {
- return;
+ /* Create call query */
+ CallQuery query;
+ query.call = call.no;
+ query.sig = call.sig;
+ query.program = glretrace::currentContext ? glretrace::currentContext->activeProgram : 0;
+
+ glGenQueries(3, query.ids);
+
+ if (retrace::profilingGpuTimes) {
+ if (supportsTimestamp) {
+ glQueryCounter(query.ids[0], GL_TIMESTAMP);
+ }
+
+ if (supportsElapsed) {
+ glBeginQuery(GL_TIME_ELAPSED, query.ids[1]);
+ }
}
- // Check for bound framebuffer last, as this may have a performance impact.
- GLint draw_framebuffer = 0;
- glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &draw_framebuffer);
- if (draw_framebuffer != 0) {
- return;
+ if (retrace::profilingPixelsDrawn && supportsOcclusion) {
+ glBeginQuery(GL_SAMPLES_PASSED, query.ids[2]);
}
- drawable->resize(width, height);
- drawable->show();
+ if (retrace::profilingCpuTimes) {
+ query.start = os::getTime();
+ }
- glScissor(0, 0, width, height);
+ callQueries.push_back(query);
}
+void
+endProfile(trace::Call &call) {
+ if (retrace::profilingCpuTimes) {
+ CallQuery& query = callQueries.back();
+ query.duration = os::getTime() - query.start;
+ }
+
+ if (retrace::profilingGpuTimes && supportsElapsed) {
+ glEndQuery(GL_TIME_ELAPSED);
+ }
+
+ if (retrace::profilingPixelsDrawn && supportsOcclusion) {
+ glEndQuery(GL_SAMPLES_PASSED);
+ }
+}
+
+void
+initContext() {
+ /* Check for extension support */
+ const char* extensions = (const char*)glGetString(GL_EXTENSIONS);
+
+ supportsTimestamp = glws::checkExtension("GL_ARB_timer_query", extensions);
+ supportsElapsed = glws::checkExtension("GL_EXT_timer_query", extensions) || supportsTimestamp;
+ supportsOcclusion = glws::checkExtension("GL_ARB_occlusion_query", extensions);
+
+ if (retrace::profilingGpuTimes) {
+ if (!supportsTimestamp && !supportsElapsed) {
+ std::cout << "Error: Cannot run profile, GL_EXT_timer_query extension is not supported." << std::endl;
+ exit(-1);
+ }
+
+ GLint bits = 0;
+ glGetQueryiv(GL_TIME_ELAPSED, GL_QUERY_COUNTER_BITS, &bits);
+
+ if (!bits) {
+ std::cout << "Error: Cannot run profile, GL_QUERY_COUNTER_BITS == 0." << std::endl;
+ exit(-1);
+ }
+ }
+
+ if (retrace::profilingPixelsDrawn && !supportsOcclusion) {
+ std::cout << "Error: Cannot run profile, GL_ARB_occlusion_query extension is not supported." << std::endl;
+ exit(-1);
+ }
+
+ if (retrace::debug) {
+ bool supportsDebugOutput = glws::checkExtension("GL_ARB_debug_output", extensions);
+
+ if (supportsDebugOutput) {
+ glDebugMessageCallbackARB(&debugOutputCallback, currentContext);
+
+ if (DEBUG_OUTPUT_SYNCHRONOUS) {
+ glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
+ }
+ }
+ }
+}
+
+void
+frame_start() {
+ firstFrame = false;
+
+ if (retrace::profiling) {
+ retrace::profiler.addFrameStart(retrace::frameNo, getGpuTimestamp(), getCpuTimestamp());
+ }
+}
+
+void
+frame_complete(trace::Call &call) {
+ if (retrace::profiling) {
+ /* Complete any remaining queries */
+ flushQueries();
+
+ /* Indicate end of current frame */
+ retrace::profiler.addFrameEnd(getGpuTimestamp(), getCpuTimestamp());
+ }
-void frame_complete(trace::Call &call) {
retrace::frameComplete(call);
- if (!drawable) {
+ /* Indicate start of next frame */
+ frame_start();
+
+ if (!currentDrawable) {
return;
}
- if (!drawable->visible) {
+ if (retrace::debug && !currentDrawable->visible) {
retrace::warning(call) << "could not infer drawable size (glViewport never called)\n";
}
}
+static const char*
+getDebugOutputSource(GLenum source) {
+ switch(source) {
+ case GL_DEBUG_SOURCE_API_ARB:
+ return "API";
+ case GL_DEBUG_SOURCE_WINDOW_SYSTEM_ARB:
+ return "Window System";
+ case GL_DEBUG_SOURCE_SHADER_COMPILER_ARB:
+ return "Shader Compiler";
+ case GL_DEBUG_SOURCE_THIRD_PARTY_ARB:
+ return "Third Party";
+ case GL_DEBUG_SOURCE_APPLICATION_ARB:
+ return "Application";
+ case GL_DEBUG_SOURCE_OTHER_ARB:
+ default:
+ return "";
+ }
+}
+
+static const char*
+getDebugOutputType(GLenum type) {
+ switch(type) {
+ case GL_DEBUG_TYPE_ERROR_ARB:
+ return "error";
+ case GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR_ARB:
+ return "deprecated behaviour";
+ case GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR_ARB:
+ return "undefined behaviour";
+ case GL_DEBUG_TYPE_PORTABILITY_ARB:
+ return "portability issue";
+ case GL_DEBUG_TYPE_PERFORMANCE_ARB:
+ return "performance issue";
+ case GL_DEBUG_TYPE_OTHER_ARB:
+ default:
+ return "unknown issue";
+ }
+}
+
+static const char*
+getDebugOutputSeverity(GLenum severity) {
+ switch(severity) {
+ case GL_DEBUG_SEVERITY_HIGH_ARB:
+ return "High";
+ case GL_DEBUG_SEVERITY_MEDIUM_ARB:
+ return "Medium";
+ case GL_DEBUG_SEVERITY_LOW_ARB:
+ return "Low";
+ default:
+ return "usnknown";
+ }
+}
+
+static void APIENTRY
+debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam) {
+ std::cerr << retrace::callNo << ": ";
+ std::cerr << "glDebugOutputCallback: ";
+ std::cerr << getDebugOutputSeverity(severity) << " severity ";
+ std::cerr << getDebugOutputSource(source) << " " << getDebugOutputType(type);
+ std::cerr << " " << id;
+ std::cerr << ", " << message;
+ std::cerr << std::endl;
+}
} /* namespace glretrace */
void
retrace::setUp(void) {
- if (retrace::coreProfile) {
- glretrace::defaultProfile = glws::PROFILE_CORE;
- }
-
glws::init();
-
- glretrace::visual[glws::PROFILE_COMPAT] = glws::createVisual(retrace::doubleBuffer, glws::PROFILE_COMPAT);
- glretrace::visual[glws::PROFILE_CORE] = glws::createVisual(retrace::doubleBuffer, glws::PROFILE_CORE);
- glretrace::visual[glws::PROFILE_ES1] = glws::createVisual(retrace::doubleBuffer, glws::PROFILE_ES1);
- glretrace::visual[glws::PROFILE_ES2] = glws::createVisual(retrace::doubleBuffer, glws::PROFILE_ES2);
}
image::Image *
retrace::getSnapshot(void) {
- if (!glretrace::drawable) {
+ if (!glretrace::currentDrawable) {
return NULL;
}
retrace::dumpState(std::ostream &os)
{
if (glretrace::insideGlBeginEnd ||
- !glretrace::drawable ||
- !glretrace::context) {
+ !glretrace::currentDrawable ||
+ !glretrace::currentContext) {
return false;
}
void
retrace::flushRendering(void) {
+ glretrace::flushQueries();
glFlush();
}
void
retrace::cleanUp(void) {
- for (int n = 0; n < glws::PROFILE_MAX; n++) {
- delete glretrace::visual[n];
- }
-
glws::cleanup();
}
projects / apitrace / blobdiff