]> git.cworth.org Git - apitrace/blobdiff - retrace/glretrace_main.cpp
Improved profiling capabilities.
[apitrace] / retrace / glretrace_main.cpp
index 313a563459216f4508edfa2ae836af2a0c6c3a84..ce064021bfa28ba74b99884f92ad5de49c52e28c 100644 (file)
 
 #include <string.h>
 
-#include "os_binary.hpp"
-#include "os_string.hpp"
-#include "os_time.hpp"
-#include "image.hpp"
 #include "retrace.hpp"
-#include "trace_callset.hpp"
 #include "glproc.hpp"
 #include "glstate.hpp"
 #include "glretrace.hpp"
+#include "os_time.hpp"
 
 
 namespace glretrace {
 
-bool double_buffer = true;
+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;
 
-unsigned frame = 0;
-long long startTime = 0;
-bool wait = false;
+struct CallQuery
+{
+    GLuint ids[3];
+    unsigned call;
+    GLuint program;
+    const trace::FunctionSig *sig;
+    uint64_t start;
+    uint64_t duration;
+};
 
-bool benchmark = false;
-static const char *compare_prefix = NULL;
-static const char *snapshot_prefix = NULL;
-static trace::CallSet snapshot_frequency;
-static trace::CallSet compare_frequency;
+static bool firstFrame = true;
+static std::list<CallQuery> callQueries;
+static const int maxActiveCallQueries = 128;
+static std::map<glws::Context*, GLuint> activePrograms;
 
-unsigned dump_state = ~0;
 
 void
 checkGlError(trace::Call &call) {
@@ -103,288 +99,217 @@ 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.
- */
-void
-updateDrawable(int width, int height) {
-    if (!drawable) {
-        return;
+static GLuint64
+getGpuTimestamp() {
+    GLuint query = 0;
+    GLuint64 timestamp = 0;
+
+    if (retrace::profilingGpuTimes) {
+        glGenQueries(1, &query);
+        glQueryCounter(query, GL_TIMESTAMP);
+        glGetQueryObjectui64vEXT(query, GL_QUERY_RESULT, &timestamp);
+        glDeleteQueries(1, &query);
     }
 
-    if (drawable->visible &&
-        width  <= drawable->width &&
-        height <= drawable->height) {
-        return;
+    return timestamp;
+}
+
+static GLuint64
+getCpuTimestamp() {
+    if (retrace::profilingCpuTimes) {
+        return os::getTime() * (1.0E9 / os::timeFrequency);
+    } else {
+        return 0;
     }
+}
 
-    // Ignore zero area viewports
-    if (width == 0 || height == 0) {
-        return;
+static void
+completeCallQuery(CallQuery& query) {
+    /* Get call start and duration */
+    GLuint64 timestamp = 0, duration = 0, samples = 0;
+
+    if (retrace::profilingGpuTimes) {
+        glGetQueryObjectui64vEXT(query.ids[0], GL_QUERY_RESULT, &timestamp);
+        glGetQueryObjectui64vEXT(query.ids[1], GL_QUERY_RESULT, &duration);
     }
 
-    // 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) {
+        glGetQueryObjectui64vEXT(query.ids[2], GL_QUERY_RESULT, &samples);
     }
 
-    drawable->resize(width, height);
-    drawable->show();
+    glDeleteQueries(3, query.ids);
 
-    glScissor(0, 0, width, height);
+    /* Add call to profile */
+    retrace::profiler.addCall(query.call, query.sig->name, query.program, samples, timestamp, duration, query.start, query.duration);
 }
 
+void
+flushQueries() {
+    for (std::list<CallQuery>::iterator itr = callQueries.begin(); itr != callQueries.end(); ++itr) {
+        completeCallQuery(*itr);
+    }
 
-static void
-snapshot(unsigned call_no) {
-    assert(snapshot_prefix || compare_prefix);
+    callQueries.clear();
+}
 
-    if (!drawable) {
-        return;
-    }
+void setActiveProgram(GLuint program)
+{
+    activePrograms[glretrace::currentContext] = program;
+}
+
+static GLuint
+getActiveProgram()
+{
+    std::map<glws::Context*, GLuint>::iterator it;
+    it = activePrograms.find(glretrace::currentContext);
+    if (it == activePrograms.end())
+        return 0;
+
+    return it->second;
+}
 
-    image::Image *ref = NULL;
-
-    if (compare_prefix) {
-        os::String filename = os::String::format("%s%010u.png", compare_prefix, call_no);
-        ref = image::readPNG(filename);
-        if (!ref) {
-            return;
-        }
-        if (retrace::verbosity >= 0) {
-            std::cout << "Read " << filename << "\n";
-        }
+void
+beginProfile(trace::Call &call) {
+    if (firstFrame) {
+        frame_start();
     }
 
-    image::Image *src = glstate::getDrawBufferImage();
-    if (!src) {
-        return;
+    /* Ensure we don't have TOO many queries waiting for results */
+    if (callQueries.size() >= maxActiveCallQueries) {
+        completeCallQuery(callQueries.front());
+        callQueries.pop_front();
     }
 
-    if (snapshot_prefix) {
-        if (snapshot_prefix[0] == '-' && snapshot_prefix[1] == 0) {
-            char comment[21];
-            snprintf(comment, sizeof comment, "%u", call_no);
-            src->writePNM(std::cout, comment);
-        } else {
-            os::String filename = os::String::format("%s%010u.png", snapshot_prefix, call_no);
-            if (src->writePNG(filename) && retrace::verbosity >= 0) {
-                std::cout << "Wrote " << filename << "\n";
-            }
-        }
+    /* Create call query */
+    CallQuery query;
+    query.call = call.no;
+    query.sig = call.sig;
+    query.program = getActiveProgram();
+
+    glGenQueries(3, query.ids);
+
+    if (retrace::profilingGpuTimes) {
+        glQueryCounter(query.ids[0], GL_TIMESTAMP);
+        glBeginQuery(GL_TIME_ELAPSED, query.ids[1]);
     }
 
-    if (ref) {
-        std::cout << "Snapshot " << call_no << " average precision of " << src->compare(*ref) << " bits\n";
-        delete ref;
+    if (retrace::profilingPixelsDrawn) {
+        glBeginQuery(GL_SAMPLES_PASSED, query.ids[2]);
     }
 
-    delete src;
+    callQueries.push_back(query);
+
+    if (retrace::profilingCpuTimes) {
+        query.start = os::getTime();
+    }
 }
 
+void
+endProfile(trace::Call &call) {
+    if (retrace::profilingCpuTimes) {
+        CallQuery& query = callQueries.back();
+        query.duration = (os::getTime() - query.start) * (1.0E9 / os::timeFrequency);
+    }
 
-void frame_complete(trace::Call &call) {
-    ++frame;
+    if (retrace::profilingGpuTimes) {
+        glEndQuery(GL_TIME_ELAPSED);
+    }
 
-    if (!drawable) {
-        return;
+    if (retrace::profilingPixelsDrawn) {
+        glEndQuery(GL_SAMPLES_PASSED);
     }
+}
 
-    if (!drawable->visible) {
-        retrace::warning(call) << "could not infer drawable size (glViewport never called)\n";
+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();
 
-static void display(void) {
-    retrace::Retracer retracer;
-
-    retracer.addCallbacks(gl_callbacks);
-    retracer.addCallbacks(glx_callbacks);
-    retracer.addCallbacks(wgl_callbacks);
-    retracer.addCallbacks(cgl_callbacks);
-    retracer.addCallbacks(egl_callbacks);
-
-    startTime = os::getTime();
-    trace::Call *call;
-
-    while ((call = retrace::parser.parse_call())) {
-        bool swapRenderTarget = call->flags & trace::CALL_FLAG_SWAP_RENDERTARGET;
-        bool doSnapshot =
-            snapshot_frequency.contains(*call) ||
-            compare_frequency.contains(*call)
-        ;
-
-        // For calls which cause rendertargets to be swaped, we take the
-        // snapshot _before_ swapping the rendertargets.
-        if (doSnapshot && swapRenderTarget) {
-            if (call->flags & trace::CALL_FLAG_END_FRAME) {
-                // For swapbuffers/presents we still use this call number,
-                // spite not have been executed yet.
-                snapshot(call->no);
-            } else {
-                // Whereas for ordinate fbo/rendertarget changes we use the
-                // previous call's number.
-                snapshot(call->no - 1);
-            }
-        }
-
-        retracer.retrace(*call);
-
-        if (doSnapshot && !swapRenderTarget) {
-            snapshot(call->no);
-        }
-
-        if (!insideGlBeginEnd &&
-            drawable && context &&
-            call->no >= dump_state) {
-            glstate::dumpCurrentContext(std::cout);
-            exit(0);
-        }
-
-        delete call;
+        /* Indicate end of current frame */
+        retrace::profiler.addFrameEnd(getGpuTimestamp(), getCpuTimestamp());
     }
 
-    // Reached the end of trace
-    glFlush();
+    retrace::frameComplete(call);
 
-    long long endTime = os::getTime();
-    float timeInterval = (endTime - startTime) * (1.0 / os::timeFrequency);
+    /* Indicate start of next frame */
+    frame_start();
 
-    if ((retrace::verbosity >= -1) || (retrace::profiling)) {
-        std::cout << 
-            "Rendered " << frame << " frames"
-            " in " <<  timeInterval << " secs,"
-            " average of " << (frame/timeInterval) << " fps\n";
+    if (!currentDrawable) {
+        return;
     }
 
-    if (wait) {
-        while (glws::processEvents()) {}
-    } else {
-        exit(0);
+    if (retrace::debug && !currentDrawable->visible) {
+        retrace::warning(call) << "could not infer drawable size (glViewport never called)\n";
     }
 }
 
+} /* namespace glretrace */
+
 
-static void usage(void) {
-    std::cout << 
-        "Usage: glretrace [OPTION] TRACE\n"
-        "Replay TRACE.\n"
-        "\n"
-        "  -b           benchmark mode (no error checking or warning messages)\n"
-        "  -p           profiling mode (run whole trace, dump profiling info)\n"
-        "  -c PREFIX    compare against snapshots\n"
-        "  -C CALLSET   calls to compare (default is every frame)\n"
-        "  -core        use core profile\n"
-        "  -db          use a double buffer visual (default)\n"
-        "  -sb          use a single buffer visual\n"
-        "  -s PREFIX    take snapshots; `-` for PNM stdout output\n"
-        "  -S CALLSET   calls to snapshot (default is every frame)\n"
-        "  -v           increase output verbosity\n"
-        "  -D CALLNO    dump state at specific call no\n"
-        "  -w           wait on final frame\n";
+void
+retrace::setUp(void) {
+    glws::init();
 }
 
-extern "C"
-int main(int argc, char **argv)
+
+void
+retrace::addCallbacks(retrace::Retracer &retracer)
 {
-    assert(compare_frequency.empty());
-    assert(snapshot_frequency.empty());
-
-    int i;
-    for (i = 1; i < argc; ++i) {
-        const char *arg = argv[i];
-
-        if (arg[0] != '-') {
-            break;
-        }
-
-        if (!strcmp(arg, "--")) {
-            break;
-        } else if (!strcmp(arg, "-b")) {
-            benchmark = true;
-            retrace::verbosity = -1;
-            glws::debug = false;
-        } else if (!strcmp(arg, "-p")) {
-            retrace::profiling = true;
-            retrace::verbosity = -1;
-            glws::debug = false;
-        } else if (!strcmp(arg, "-c")) {
-            compare_prefix = argv[++i];
-            if (compare_frequency.empty()) {
-                compare_frequency = trace::CallSet(trace::FREQUENCY_FRAME);
-            }
-        } else if (!strcmp(arg, "-C")) {
-            compare_frequency = trace::CallSet(argv[++i]);
-            if (compare_prefix == NULL) {
-                compare_prefix = "";
-            }
-        } else if (!strcmp(arg, "-D")) {
-            dump_state = atoi(argv[++i]);
-            retrace::verbosity = -2;
-        } else if (!strcmp(arg, "-core")) {
-            defaultProfile = glws::PROFILE_CORE;
-        } else if (!strcmp(arg, "-db")) {
-            double_buffer = true;
-        } else if (!strcmp(arg, "-sb")) {
-            double_buffer = false;
-        } else if (!strcmp(arg, "--help")) {
-            usage();
-            return 0;
-        } else if (!strcmp(arg, "-s")) {
-            snapshot_prefix = argv[++i];
-            if (snapshot_frequency.empty()) {
-                snapshot_frequency = trace::CallSet(trace::FREQUENCY_FRAME);
-            }
-            if (snapshot_prefix[0] == '-' && snapshot_prefix[1] == 0) {
-                os::setBinaryMode(stdout);
-                retrace::verbosity = -2;
-            }
-        } else if (!strcmp(arg, "-S")) {
-            snapshot_frequency = trace::CallSet(argv[++i]);
-            if (snapshot_prefix == NULL) {
-                snapshot_prefix = "";
-            }
-        } else if (!strcmp(arg, "-v")) {
-            ++retrace::verbosity;
-        } else if (!strcmp(arg, "-w")) {
-            wait = true;
-        } else {
-            std::cerr << "error: unknown option " << arg << "\n";
-            usage();
-            return 1;
-        }
-    }
+    retracer.addCallbacks(glretrace::gl_callbacks);
+    retracer.addCallbacks(glretrace::glx_callbacks);
+    retracer.addCallbacks(glretrace::wgl_callbacks);
+    retracer.addCallbacks(glretrace::cgl_callbacks);
+    retracer.addCallbacks(glretrace::egl_callbacks);
+}
 
-    glws::init();
-    visual[glws::PROFILE_COMPAT] = glws::createVisual(double_buffer, glws::PROFILE_COMPAT);
-    visual[glws::PROFILE_CORE] = glws::createVisual(double_buffer, glws::PROFILE_CORE);
-    visual[glws::PROFILE_ES1] = glws::createVisual(double_buffer, glws::PROFILE_ES1);
-    visual[glws::PROFILE_ES2] = glws::createVisual(double_buffer, glws::PROFILE_ES2);
 
-    for ( ; i < argc; ++i) {
-        if (!retrace::parser.open(argv[i])) {
-            std::cerr << "error: failed to open " << argv[i] << "\n";
-            return 1;
-        }
+image::Image *
+retrace::getSnapshot(void) {
+    if (!glretrace::currentDrawable) {
+        return NULL;
+    }
 
-        display();
+    return glstate::getDrawBufferImage();
+}
 
-        retrace::parser.close();
-    }
 
-    for (int n = 0; n < glws::PROFILE_MAX; n++) {
-        delete visual[n];
+bool
+retrace::dumpState(std::ostream &os)
+{
+    if (glretrace::insideGlBeginEnd ||
+        !glretrace::currentDrawable ||
+        !glretrace::currentContext) {
+        return false;
     }
 
-    glws::cleanup();
+    glstate::dumpCurrentContext(os);
 
-    return 0;
+    return true;
 }
 
-} /* namespace glretrace */
+void
+retrace::flushRendering(void) {
+    glretrace::flushQueries();
+    glFlush();
+}
+
+void
+retrace::waitForInput(void) {
+    while (glws::processEvents()) {
+    }
+}
+
+void
+retrace::cleanUp(void) {
+    glws::cleanup();
+}