Merge pull request #42 from ROCm/rogarcia.llamacpp-roofline-for-vlm

mtmd-cli: reset roofline profiler after warmup prefill

Author: roberteg16

tools/mtmd/CMakeLists.txt

@@ -134,6 +134,13 @@ if (LLAMA_BUILD_TOOLS)
     target_link_libraries  (${TARGET} PRIVATE llama-common mtmd Threads::Threads)
     target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
+    # Compile in the post-warmup roofline reset only when the HIP profiler is built
+    # (ggml-cuda-roofline.cpp, linked via ggml-hip). No-op in every other build.
+    if (GGML_HIP_ROOFLINE)
+        target_compile_definitions(${TARGET} PRIVATE GGML_HIP_ROOFLINE)
+        target_include_directories(${TARGET} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../ggml/src/ggml-cuda)
+    endif()
+
     # mtmd-debug tool
     add_executable(llama-mtmd-debug debug/mtmd-debug.cpp)
     set_target_properties(llama-mtmd-debug PROPERTIES OUTPUT_NAME llama-mtmd-debug)

tools/mtmd/mtmd-cli.cpp

@@ -10,6 +10,13 @@
 #include "mtmd.h"
 #include "mtmd-helper.h"
 
+// Optional per-op roofline profiler (HIP build with GGML_HIP_ROOFLINE). Used to
+// discard warmup so the report covers only the measured prefill. Compiled out
+// otherwise. See ggml/src/ggml-cuda/ggml-cuda-roofline.cpp.
+#ifdef GGML_HIP_ROOFLINE
+#include "ggml-cuda-roofline.h"
+#endif
+
 #include <vector>
 #include <limits.h>
 #include <cinttypes>
@@ -453,6 +460,11 @@ int main(int argc, char ** argv) {
                 return 1; // error is already printed by libmtmd
             }
         }
+        // Discard warmup from the per-op roofline profiler so the report covers
+        // only the measured prefill below. Compiled out unless GGML_HIP_ROOFLINE.
+#ifdef GGML_HIP_ROOFLINE
+        ggml_cuda_roofline_reset();
+#endif
         if (eval_message(ctx, msg)) {
             return 1;
         }