Merge branch 'device-timer' into 'master'

Added new event to dump device timer when available.

See merge request !7

opencl/opencl_events.yaml

@@ -243,3 +243,12 @@ lttng_ust_opencl_devices:
     fields:
     - [ ctf_integer_hex, uintptr_t, device, "(uintptr_t)device" ]
     - [ ctf_string, name, name ]
+  - name: device_timer
+    args:
+    - [ cl_device_id, device ]
+    - [ cl_ulong, device_timestamp ]
+    - [ cl_ulong, host_timestamp ]
+    fields:
+    - [ ctf_integer_hex, uintptr_t, device, "(uintptr_t)device" ]
+    - [ ctf_integer, cl_ulong, device_timestamp, device_timestamp ]
+    - [ ctf_integer, cl_ulong, host_timestamp, host_timestamp ]

opencl/opencl_model.rb

@@ -1150,8 +1150,14 @@ EOF
 register_prologue "clGetDeviceIDs", <<EOF
   cl_uint n_e;
   if (tracepoint_enabled(lttng_ust_opencl_devices, device_name) && !num_devices && devices) {
-  num_devices = &n_e;
-}
+    num_devices = &n_e;
+  }
+EOF
+
+register_prologue "clGetDeviceIDs", <<EOF
+  if (tracepoint_enabled(lttng_ust_opencl_devices, device_timer) && !num_devices && devices) {
+    num_devices = &n_e;
+  }
 EOF
 
 register_epilogue "clGetDeviceIDs", <<EOF
@@ -1162,6 +1168,14 @@ register_epilogue "clGetDeviceIDs", <<EOF
   }
 EOF
 
+register_epilogue "clGetDeviceIDs", <<EOF
+  if (tracepoint_enabled(lttng_ust_opencl_devices, device_timer) && _retval == CL_SUCCESS && devices) {
+    for (cl_uint i = 0; i < *num_devices; i++) {
+      dump_device_timer(devices[i]);
+    }
+  }
+EOF
+
 str = $opencl_commands.select{ |c| c.extension? }.collect { |c|
   <<EOF
   if (strcmp(func_name, "#{c.prototype.name}") == 0) {

opencl/tracer_opencl_helpers.include.c

@@ -23,7 +23,7 @@ struct opencl_version {
 //static const struct opencl_version opencl_version_1_1 = {1, 1};
 static const struct opencl_version opencl_version_1_2 = {1, 2};
 //static const struct opencl_version opencl_version_2_0 = {2, 0};
-//static const struct opencl_version opencl_version_2_1 = {2, 1};
+static const struct opencl_version opencl_version_2_1 = {2, 1};
 //static const struct opencl_version opencl_version_2_2 = {2, 2};
 
 static inline int compare_opencl_version(const struct opencl_version *v1, const struct opencl_version *v2) {
@@ -759,6 +759,19 @@ cleanup:
   free(name);
 }
 
+static inline void dump_device_timer(cl_device_id device) {
+  struct opencl_version version = {1, 0};
+
+  get_device_platform_version(device, &version);
+  if (compare_opencl_version(&version, &opencl_version_2_1) >= 0) {
+    cl_ulong device_timestamp;
+    cl_ulong host_timestamp;
+    if (CL_GET_DEVICE_AND_HOST_TIMER_PTR(device, &device_timestamp, &host_timestamp) != CL_SUCCESS)
+      return;
+    do_tracepoint(lttng_ust_opencl_devices, device_timer, device, device_timestamp, host_timestamp);
+  }
+}
+
 static inline void dump_program_device_build_infos(cl_program program, cl_device_id device) {
   cl_build_status build_status;
   char *build_options = "";

utils/xprof.sh.erb.in

@@ -73,7 +73,7 @@ trace_and_summary() {
     lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_opencl_build:infos*
     lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_opencl_profiling:*
     lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_opencl_arguments:kernel_info
-    lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_opencl_devices:device_name
+    lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_opencl_devices:*
 <% end %>
 <% if languages.include?("ze") %>
     lttng $quiet enable-event --channel=blocking-channel --userspace lttng_ust_ze:*