Added support for dumping CUDA arguments.

cuda/cuda_export_tables.yaml

@@ -12,8 +12,8 @@
       }
   functions:
   - offset: 0x108
-    name: CUfunction_get_arg_count
-    declaration: "CUresult CUfunction_get_arg_count(CUfunction f, size_t * pArgCount)"
+    name: cuFunction_get_arg_count
+    declaration: "CUresult cuFunction_get_arg_count(CUfunction f, size_t * pArgCount)"
   - offset: 0x110
-    name: CUfunction_get_arg_descriptor
-    declaration: "CUresult CUfunction_get_arg_descriptor(CUfunction f, size_t argIndex, CUfunction_arg_desc_query *pArgDescQuery)"
+    name: cuFunction_get_arg_descriptor
+    declaration: "CUresult cuFunction_get_arg_descriptor(CUfunction f, size_t argIndex, CUfunction_arg_desc_query *pArgDescQuery)"

cuda/cuda_model.rb

@@ -748,3 +748,10 @@ end
 CUDA_POINTER_NAMES = ($cuda_commands).collect { |c|
   [c, upper_snake_case(c.pointer_name)]
 }.to_h
+
+dump_args = <<EOF
+  _dump_kernel_args(f, kernelParams, extra);
+EOF
+
+register_prologue "cuLaunchKernel", dump_args
+register_prologue "cuLaunchKernel_ptsz", dump_args

cuda/gen_babeltrace_cuda_lib.rb

@@ -16,6 +16,8 @@ meta_parameter_lambda = lambda { |m, dir|
     when YAMLCAst::CustomType
       if $objects.include?(t.name)
         "s << \"#{name}: \#{\"0x%016x\" % defi[\"#{name}\"]}\""
+      elsif t.name == "CUdeviceptr"
+        "s << \"#{name}: \#{\"0x%016x\" % defi[\"#{name}\"]}\""
       elsif $all_enum_names.include?(t.name)
         "s << \"#{name}: \#{CUDA::#{to_class_name(t.name)}.from_native(defi[\"#{name}\"], nil)}\""
       elsif $all_bitfield_names.include?(t.name)

cuda/gen_cuda.rb

@@ -5,6 +5,7 @@ puts <<EOF
 #include <cuda.h>
 #include <pthread.h>
 #include "cuda_tracepoints.h"
+#include "cuda_args.h"
 EOF
 
 $cuda_commands.each { |c|
@@ -74,7 +75,7 @@ EOF
   table["functions"].each { |func|
     puts <<EOF
       if (0x#{func["offset"].to_s(16)} < tableSize) {
-        #{upper_snake_case(func["name"]+"_ptr")} = (#{func["name"]}_t)((intptr_t)pExportTable + 0x#{func["offset"].to_s(16)});
+        #{upper_snake_case(func["name"]+"_ptr")} = *(#{func["name"]}_t *)((intptr_t)pExportTable + 0x#{func["offset"].to_s(16)});
       }
 EOF
   }

cuda/tracer_cuda.sh.in

@@ -20,7 +20,8 @@ whichlib() {
 while true; do
   case "$1" in
     --cudart) shift; cudart=1;;
-    -v | --visualize) shift; lttng_view=1;;
+    -a | --arguments ) shift; arguments=1;;
+    -v | --visualize ) shift; lttng_view=1;;
     -- ) shift; break ;;
     * ) break ;;
   esac
@@ -40,6 +41,10 @@ if [ ! -z "$cudart" ]
 then
   lttng enable-event --channel=blocking-channel --userspace lttng_ust_cudart:*
 fi
+if [ ! -z "$arguments" ]
+then
+  lttng enable-event --channel=blocking-channel --userspace lttng_ust_cuda_args:*
+fi
 if [ -z "$LTTNG_UST_CUDA_LIBCUDA" ]
 then
   export LTTNG_UST_CUDA_LIBCUDA=$(whichlib libcuda.so | head -n 1)

cuda/tracer_cuda_helpers.include.c

+
+static inline void _dump_kernel_args(CUfunction f, void **kernelParams, void** extra) {
+  (void)extra;
+  size_t argCount;
+  CUresult res;
+  int    count = 0;
+  CUfunction_arg_desc_query q;
+
+  if (tracepoint_enabled(lttng_ust_cuda_args, arg_count)) {
+    res = CU_FUNCTION_GET_ARG_COUNT_PTR(f, &argCount);
+    if (res != CUDA_SUCCESS)
+      return;
+    count = 1;
+    do_tracepoint(lttng_ust_cuda_args, arg_count, f, argCount);
+  }
+  if (tracepoint_enabled(lttng_ust_cuda_args, arg_value)) {
+    if (!count) {
+      res = CU_FUNCTION_GET_ARG_COUNT_PTR(f, &argCount);
+      if (res != CUDA_SUCCESS)
+        return;
+    }
+    if (kernelParams) {
+      for (size_t i = 0; i < argCount; i++) {
+        q.sz = 0x28;
+        res = CU_FUNCTION_GET_ARG_DESCRIPTOR_PTR(f, i, &q);
+        if (res == CUDA_SUCCESS) {
+          do_tracepoint(lttng_ust_cuda_args, arg_value, f, i, kernelParams[i], q.argSize);
+        }
+      }
+    }
+    if (extra) {
+      void *argBuffer = NULL;
+      size_t argBufferSize = 0;
+      int indx = 0;
+      while (extra[indx] != CU_LAUNCH_PARAM_END) {
+        if ((void *)CU_LAUNCH_PARAM_BUFFER_POINTER == extra[indx]) {
+          argBuffer = extra[indx+1];
+        } else if ((void *)CU_LAUNCH_PARAM_BUFFER_SIZE == extra[indx]) {
+          argBufferSize = *(size_t*)(extra + indx + 1);
+        }
+        indx += 2;
+      }
+      if (argBuffer && argBufferSize) {
+        for (size_t i = 0; i < argCount; i++) {
+          q.sz = 0x28;
+          res = CU_FUNCTION_GET_ARG_DESCRIPTOR_PTR(f, i, &q);
+          if (res == CUDA_SUCCESS && argBufferSize >= q.argOffset + q.argSize) {
+            do_tracepoint(lttng_ust_cuda_args, arg_value, f, i, (void *)((intptr_t)argBuffer + q.argOffset), q.argSize);
+          }
+        }
+      }
+    }
+  }
+}
+
 static pthread_once_t _init = PTHREAD_ONCE_INIT;
 static __thread volatile int in_init = 0;
 static volatile int _initialized = 0;

ze/tracer_ze.sh.in

@@ -20,7 +20,7 @@ whichlib() {
 while true; do
   case "$1" in
     -p | --profiling ) shift; profiling=1; export LTTNG_UST_ZE_PROFILE=1;;
-    -v | --visualize) shift; lttng_view=1;;
+    -v | --visualize ) shift; lttng_view=1;;
     --properties ) shift; properties=1;;
     -- ) shift; break ;;
     * ) break ;;