Fixed CUDA tracer to work on basic tests.

configure.ac

@@ -80,6 +80,7 @@ AC_CONFIG_FILES([opencl/babeltrace_opencl], [chmod +x opencl/babeltrace_opencl])
 AC_CONFIG_FILES([ze/babeltrace_ze], [chmod +x ze/babeltrace_ze])
 AC_CONFIG_FILES([ze/tracer_ze.sh], [chmod +x ze/tracer_ze.sh])
 AC_CONFIG_FILES([ze/zeprof.sh], [chmod +x ze/zeprof.sh])
+AC_CONFIG_FILES([cuda/babeltrace_cuda], [chmod +x cuda/babeltrace_cuda])
 AC_CONFIG_FILES([cuda/tracer_cuda.sh], [chmod +x cuda/tracer_cuda.sh])
 AC_CONFIG_FILES([xprof/iprof], [chmod +x xprof/iprof])
 AC_OUTPUT

cuda/Makefile.am

@@ -11,6 +11,15 @@ CUDA_PROBES_INCL = \
 CUDA_PROBES_SRC = \
 	cuda_tracepoints.c
 
+CUDA_STATIC_PROBES_TP = \
+	cuda_args.tp
+
+CUDA_STATIC_PROBES_INCL = \
+	cuda_args.h
+
+CUDA_STATIC_PROBES_SRC = \
+	cuda_args.c
+
 CUDART_PROBES_TP = \
 	cudart_tracepoints.tp
 
@@ -24,6 +33,9 @@ CLEANFILES = \
 	$(CUDA_PROBES_TP) \
 	$(CUDA_PROBES_INCL) \
 	$(CUDA_PROBES_SRC) \
+	$(CUDA_STATIC_PROBES_TP) \
+	$(CUDA_STATIC_PROBES_INCL) \
+	$(CUDA_STATIC_PROBES_SRC) \
 	$(CUDART_PROBES_TP) \
 	$(CUDART_PROBES_INCL) \
 	$(CUDART_PROBES_SRC)
@@ -131,17 +143,21 @@ cuda_tracepoints.tp: $(srcdir)/gen_cuda_tracepoints.rb $(CUDA_MODEL) gen_probe_b
 cudart_tracepoints.tp: $(srcdir)/gen_cudart_tracepoints.rb $(CUDART_MODEL) gen_probe_base.rb
 	SRC_DIR=$(srcdir) $(RUBY) $(srcdir)/gen_cudart_tracepoints.rb > $@
 
+cuda_args.tp: $(srcdir)/gen_cuda_custom_probes.rb $(CUDA_MODEL)
+	SRC_DIR=$(srcdir) $(RUBY) $(srcdir)/gen_cuda_custom_probes.rb lttng_ust_cuda_args > $@
+
 %.h %.c: %.tp
 	$(LTTNG_GEN_TP) $< -o $*.c -o $*.h
 
 EXTRA_DIST += \
 	gen_cuda_tracepoints.rb \
 	gen_cudart_tracepoints.rb \
-	gen_probe_base.rb
+	gen_probe_base.rb \
+	gen_cuda_custom_probes.rb
 
-BUILT_SOURCES = $(CUDA_PROBES_INCL) $(CUDART_PROBES_INCL)
+BUILT_SOURCES = $(CUDA_PROBES_INCL) $(CUDA_STATIC_PROBES_INCL) $(CUDART_PROBES_INCL)
 
-tracer_cuda.c: $(srcdir)/gen_cuda.rb $(srcdir)/tracer_cuda_helpers.include.c $(CUDA_MODEL) $(CUDA_PROBES_INCL) cuda_export_tables.yaml
+tracer_cuda.c: $(srcdir)/gen_cuda.rb $(srcdir)/tracer_cuda_helpers.include.c $(CUDA_MODEL) $(CUDA_PROBES_INCL) $(CUDA_STATIC_PROBES_INCL) cuda_export_tables.yaml
 	SRC_DIR=$(srcdir) $(RUBY) $< > $@
 
 tracer_cudart.c: $(srcdir)/gen_cudart.rb $(srcdir)/tracer_cudart_helpers.include.c $(CUDART_MODEL) $(CUDART_PROBES_INCL)
@@ -158,13 +174,15 @@ CLEANFILES += \
 	tracer_cuda.c \
 	tracer_cudart.c
 
-bin_SCRIPTS = tracer_cuda.sh
+bin_SCRIPTS = tracer_cuda.sh babeltrace_cuda
 
 noinst_LTLIBRARIES = libcudatracepoints.la libcudarttracepoints.la
 
 nodist_libcudatracepoints_la_SOURCES = \
 	$(CUDA_PROBES_INCL) \
-	$(CUDA_PROBES_SRC)
+	$(CUDA_PROBES_SRC) \
+	$(CUDA_STATIC_PROBES_INCL) \
+	$(CUDA_STATIC_PROBES_SRC)
 
 libcudatracepoints_la_CFLAGS = $(LTTNG_FLAGS) $(LTTNG_UST_CFLAGS)
 libcudatracepoints_la_LDFLAGS = $(LTTNG_UST_LIBS)

cuda/babeltrace_cuda.in

+#!/usr/bin/env ruby
+DATADIR = File.join("@prefix@", "share")
+$:.unshift(DATADIR) if File::directory?(DATADIR)
+$:.unshift("./")
+require 'babeltrace'
+require 'babeltrace/ctf'
+require 'yaml'
+require 'pp'
+require 'babeltrace_cuda_lib'
+
+c = Babeltrace::Context::new
+traces = c.add_traces(path: ARGV[0])
+
+puts "#{c.get_timestamp_begin.strftime("%Y-%m-%d %H:%M:%S.%9L %z")} -- #{c.get_timestamp_end.strftime("%Y-%m-%d %H:%M:%S.%9L %z")}"
+
+events = {}
+traces.each { |t|
+  events[t.handle_id] = t.get_event_decl_list.collect(&:name)
+}
+
+it = c.iter_create
+it.each { |ev|
+  name = events[ev.handle_id][ev.id]
+  print "#{ev.timestamp.strftime("%H:%M:%S.%9L")} - #{name}: "
+  defi = ev.top_level_scope(:EVENT_FIELDS)
+  l = $event_lambdas[name]
+  if l
+    puts l.call(defi.value)
+  else
+    puts defi.value
+  end
+}
+

cuda/cuda_events.yaml

 ---
-{}
+lttng_ust_cuda_args:
+  events:
+  - name: arg_count
+    args:
+    - [ CUfunction, f ]
+    - [ size_t, argCount ]
+    fields:
+    - [ ctf_integer_hex, uintptr_t, f, "(uintptr_t)f" ]
+    - [ ctf_integer, size_t, argCount, argCount ]
+  - name: arg_value
+    args:
+    - [ CUfunction, f ]
+    - [ "unsigned int", argIndex ]
+    - [ "void *", kernelParam ]
+    - [ size_t, argSize ]
+    fields:
+    - [ ctf_integer_hex, uintptr_t, f, "(uintptr_t)f" ]
+    - [ ctf_integer_hex, uintptr_t, kernelParam, "(uintptr_t)kernelParam" ]
+    - [ ctf_sequence_text, uint8_t, kernelParam_val, "(uint8_t *)kernelParam", size_t, "kernelParam ? argSize : 0" ]

cuda/cuda_meta_parameters.yaml

@@ -226,7 +226,7 @@ meta_parameters:
   cuMipmappedArrayGetLevel:
     - [OutScalar, pLevelArray]
   cuPointerGetAttribute:
-    - [OutLTTng, data, ctf_sequence_text, uint8_t, data, data, size_t, "attribute == CU_POINTER_ATTRIBUTE_CONTEXT ? sizeof(CUcontext) : attribute == CU_POINTER_ATTRIBUTE_MEMORY_TYPE ? sizeof(unsigned int) : attribute == CU_POINTER_ATTRIBUTE_DEVICE_POINTER ? sizeof(CUdeviceptr) : attribute == CU_POINTER_ATTRIBUTE_HOST_POINTER ? sizeof(void *) : attribute == CU_POINTER_ATTRIBUTE_P2P_TOKENS ? sizeof(CUDA_POINTER_ATTRIBUTE_P2P_TOKENS) : attribute == CU_POINTER_ATTRIBUTE_SYNC_MEMOPS ? sizeof(uint8_t) : attribute == CU_POINTER_ATTRIBUTE_BUFFER_ID ? sizeof(unsigned long long) : attribute == CU_POINTER_ATTRIBUTE_IS_MANAGED ? sizeof(uint8_t) : attribute == CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL ? sizeof(int) : 0"]
+    - [OutLTTng, data, ctf_sequence_text, uint8_t, data_val, data, size_t, "attribute == CU_POINTER_ATTRIBUTE_CONTEXT ? sizeof(CUcontext) : attribute == CU_POINTER_ATTRIBUTE_MEMORY_TYPE ? sizeof(unsigned int) : attribute == CU_POINTER_ATTRIBUTE_DEVICE_POINTER ? sizeof(CUdeviceptr) : attribute == CU_POINTER_ATTRIBUTE_HOST_POINTER ? sizeof(void *) : attribute == CU_POINTER_ATTRIBUTE_P2P_TOKENS ? sizeof(CUDA_POINTER_ATTRIBUTE_P2P_TOKENS) : attribute == CU_POINTER_ATTRIBUTE_SYNC_MEMOPS ? sizeof(uint8_t) : attribute == CU_POINTER_ATTRIBUTE_BUFFER_ID ? sizeof(unsigned long long) : attribute == CU_POINTER_ATTRIBUTE_IS_MANAGED ? sizeof(uint8_t) : attribute == CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL ? sizeof(int) : 0"]
   cuMemRangeGetAttribute:
     - [OutArray, data, dataSize]
   cuMemRangeGetAttributes:
@@ -234,7 +234,7 @@ meta_parameters:
     - [InArray, dataSizes, numAttributes]
     - [OutArray, data, numAttributes]
   cuPointerSetAttribute:
-    - [InLTTng, value, ctf_sequence_text, uint8_t, value, value, size_t, "attribute == CU_POINTER_ATTRIBUTE_SYNC_MEMOPS ? sizeof(uint8_t) : 0"]
+    - [InLTTng, value, ctf_sequence_text, uint8_t, value_val, value, size_t, "attribute == CU_POINTER_ATTRIBUTE_SYNC_MEMOPS ? sizeof(uint8_t) : 0"]
   cuPointerGetAttributes:
     - [InArray, attributes, numAttributes]
     - [OutArray, data, numAttributes]

cuda/cuda_model.rb

@@ -50,7 +50,7 @@ all_types.select { |t| t.type.kind_of? YAMLCAst::Struct }.each { |t|
   end
 }
 
-INIT_FUNCTIONS = /cuInit/
+INIT_FUNCTIONS = /cuInit|cuDriverGetVersion/
 
 FFI_TYPE_MAP =  {
  "unsigned char" => "ffi_type_uint8",

cuda/gen_babeltrace_cuda_lib.rb

@@ -56,7 +56,7 @@ EOF
   fields = []
   if dir == :start
     if c.parameters
-      fields += (c.parameters+c.tracepoint_parameters).collect { |p|
+      fields += c.parameters.collect { |p|
         case p.type
         when YAMLCAst::Pointer
           "s << \"#{p.name}: \#{\"0x%016x\" % defi[\"#{p.name}\"]}\""
@@ -134,6 +134,21 @@ EOF
         else
           "s << \"#{field.name}: \#{defi[\"#{field.name}\"]}\""
         end
+      when :ctf_sequence_text
+        arg = e["args"].find { |type, name|
+          name == field.expression
+        }
+        name = field.name
+        if arg
+          t = arg[0].sub("*","").strip
+          if $all_struct_names.include?(t)
+            "s << \"#{name}: \#{defi[\"#{name}\"].size > 0 ? CUDA::#{to_class_name(t)}.new(FFI::MemoryPointer.from_string(defi[\"#{name}\"])) : nil}\""
+          else
+            "s << \"#{name}: \#{defi[\"#{name}\"].inspect}\""
+          end
+        else
+          "s << \"#{name}: \#{defi[\"#{name}\"].inspect}\""
+        end
       else
         raise "Unsupported LTTng macro #{field.macro}!"
       end

cuda/gen_cuda_custom_probes.rb

+require_relative 'cuda_model'
+
+namespace = ARGV[0]
+
+raise "No namespace provided!" unless namespace
+
+h = YAML::load_file(File.join(SRC_DIR,"cuda_events.yaml"))[namespace]
+
+raise "Invalid namespace!" unless h
+
+puts <<EOF
+#include <cuda.h>
+
+EOF
+
+enums = h["enums"]
+
+if enums
+  enums.each { |e|
+    LTTng.print_enum(namespace, e)
+  }
+end
+
+h["events"].each { |e|
+  LTTng.print_tracepoint(namespace, e)
+}

cuda/gen_cuda_library.rb

@@ -181,18 +181,18 @@ module CUDA
 
   module UUID
     def to_s
-      a = self[:bytes].to_a
+      a = self[:bytes].to_a.collect { |v| v < 0 ? 0x100 + v : v }
       s = "{ id: "
-      s << "%02x" % a[3]
-      s << "%02x" % a[2]
-      s << "%02x" % a[1]
       s << "%02x" % a[0]
+      s << "%02x" % a[1]
+      s << "%02x" % a[2]
+      s << "%02x" % a[3]
       s << "-"
-      s << "%02x" % a[5]
       s << "%02x" % a[4]
+      s << "%02x" % a[5]
       s << "-"
-      s << "%02x" % a[7]
       s << "%02x" % a[6]
+      s << "%02x" % a[7]
       s << "-"
       s << "%02x" % a[8]
       s << "%02x" % a[9]