Dynamic trace support in Telluride

Signed-off-by: rbramand <[email protected]>

src/runtime_src/core/common/api/bo_int.h

@@ -28,6 +28,17 @@ XRT_CORE_COMMON_EXPORT
 xrt::bo
 create_debug_bo(const xrt::hw_context& hwctx, size_t sz);
 
+// create_dtrace_bo() - Create a trace buffer object within a hwctx
+//
+// Allocates a buffer object within a hwctx used for dynamic tracing.
+// The BO is used by driver / firmware to fill dynamic tracing data
+// and is shared with user space XRT.
+// The shim allocation is through hwctx_handle::alloc_bo with
+// the XRT_BO_USE_DTRACE flag captured in extension flags.
+XRT_CORE_COMMON_EXPORT
+xrt::bo
+create_dtrace_bo(const xrt::hw_context& hwctx, size_t sz);
+
 } // bo_int, xrt_core
 
 #endif

src/runtime_src/core/common/api/module_int.h

@@ -75,6 +75,11 @@ get_kernel_info(const xrt::module& module);
 uint32_t
 get_partition_size(const xrt::module& module);
 
+// Dump dynamic trace buffer
+// Buffer is dumped after the kernel run is finished
+void
+dump_dtrace_buffer(const xrt::module& module);
+
 } // xrt_core::module_int
 
 #endif

src/runtime_src/core/common/api/xrt_bo.cpp

@@ -1689,21 +1689,34 @@ get_offset(const xrt::bo& bo)
   return handle->get_offset();
 }
 
-xrt::bo
-create_debug_bo(const xrt::hw_context& hwctx, size_t sz)
+static xrt::bo
+create_bo_helper(const xrt::hw_context& hwctx, size_t sz, uint32_t use_flag)
 {
   xcl_bo_flags flags {0};  // see xrt_mem.h
   flags.flags = XRT_BO_FLAGS_CACHEABLE;
   flags.access = XRT_BO_ACCESS_LOCAL;
   flags.dir = XRT_BO_ACCESS_READ_WRITE;
-  flags.use = XRT_BO_USE_DEBUG;
+  flags.use = use_flag;
 
-  // While the memory group should be ignored (inferred) for debug
-  // buffers, it is still passed in as a default group 1 with no
-  // implied correlation to xclbin connectivity or memory group.
+  // While the memory group should be ignored (inferred) for
+  // debug / trace buffers, it is still passed in as a default
+  // group 1 with no implied correlation to xclbin connectivity
+  // or memory group.
   return xrt::bo{alloc(device_type{hwctx}, sz, flags.all, 1)};
 }
 
+xrt::bo
+create_debug_bo(const xrt::hw_context& hwctx, size_t sz)
+{
+  return create_bo_helper(hwctx, sz, XRT_BO_USE_DEBUG);
+}
+
+xrt::bo
+create_dtrace_bo(const xrt::hw_context& hwctx, size_t sz)
+{
+  return create_bo_helper(hwctx, sz, XRT_BO_USE_DTRACE);
+}
+
 } // xrt_core::bo_int
 
 ////////////////////////////////////////////////////////////////

src/runtime_src/core/common/api/xrt_kernel.cpp

@@ -2522,6 +2522,11 @@ class run_impl
     static bool dump = xrt_core::config::get_feature_toggle("Debug.dump_scratchpad_mem");
     if (dump)
       xrt_core::module_int::dump_scratchpad_mem(m_module);
+
+    // dump dtrace buffer if ini option is enabled
+    static auto dtrace_lib_path = xrt_core::config::get_dtrace_lib_path();
+    if (!dtrace_lib_path.empty())
+      xrt_core::module_int::dump_dtrace_buffer(m_module);
 
     return state;
   }
@@ -2546,6 +2551,12 @@ class run_impl
       state = cmd->wait();
     }
 
+    // dump dtrace buffer if ini option is enabled
+    // here dtrace is dumped in both passing and timeout cases
+    static auto dtrace_lib_path = xrt_core::config::get_dtrace_lib_path();
+    if (!dtrace_lib_path.empty())
+      xrt_core::module_int::dump_dtrace_buffer(m_module);
+
     if (state == ERT_CMD_STATE_COMPLETED) {
       m_usage_logger->log_kernel_run_info(kernel.get(), this, state);
       static bool dump = xrt_core::config::get_feature_toggle("Debug.dump_scratchpad_mem");