Also defer GpuContextName for on-demand profiling

Without this, a late-connecting client receives the deferred
GpuNewContext but not the GpuContextName, so the GPU context appears
unnamed in the profiler.

Add check_gpu_ctx_name tool to verify context names in captured traces.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

examples/RocprofOnDemandRepro/Makefile

@@ -4,12 +4,13 @@ CXX          := g++
 ROCM_LIB     := /opt/rocm/lib
 
 TRACY_SRCS   := $(TRACY_PUBLIC)/TracyClient.cpp
-INCLUDES     := -I$(TRACY_PUBLIC)
+ROCM_INC     := /opt/rocm/include
+INCLUDES     := -I$(TRACY_PUBLIC) -I$(ROCM_INC)
 LIBS         := -L$(ROCM_LIB) -lrocprofiler-sdk -lpthread -ldl
 
 # On-demand mode (the default for this repro) — profiling starts when
 # a client connects, not at program launch.
-DEFINES      := -DTRACY_ENABLE -DTRACY_ON_DEMAND -DTRACY_ROCPROF
+DEFINES      := -DTRACY_ENABLE -DTRACY_ON_DEMAND -DTRACY_ROCPROF -D__HIP_PLATFORM_AMD__
 CXXFLAGS     := -O2 $(DEFINES)
 HIPCCFLAGS   := -O2 $(DEFINES)
 
@@ -17,11 +18,14 @@ HIPCCFLAGS   := -O2 $(DEFINES)
 
 all: repro
 
-repro: repro.cpp tracy_client.o
-	$(HIPCC) $(HIPCCFLAGS) $(INCLUDES) -o $@ $< tracy_client.o $(LIBS)
+repro: repro.o tracy_client.o
+	$(HIPCC) -o $@ $^ $(LIBS)
+
+repro.o: repro.cpp
+	$(HIPCC) $(HIPCCFLAGS) $(INCLUDES) -c -o $@ $<
 
 tracy_client.o: $(TRACY_SRCS)
 	$(CXX) $(CXXFLAGS) $(INCLUDES) -c -o $@ $<
 
 clean:
-	rm -f repro tracy_client.o
+	rm -f repro repro.o tracy_client.o

examples/RocprofOnDemandRepro/README.md

@@ -5,7 +5,7 @@ HIP application built with `TRACY_ON_DEMAND` and `TRACY_ROCPROF`.
 
 ## Root cause
 
-Two bugs in `TracyRocprof.cpp` break on-demand profiling:
+Three bugs in `TracyRocprof.cpp` break on-demand profiling:
 
 1. **GpuNewContext not deferred.** `gpu_context_allocate()` writes a
    `GpuNewContext` queue item but does not call `DeferItem()`. When a
@@ -15,7 +15,12 @@ Two bugs in `TracyRocprof.cpp` break on-demand profiling:
 
        Assertion `ctx' failed in ProcessGpuZoneBeginImplCommon
 
-2. **Kernel symbols dropped before init.** The `data->init` guard at the
+2. **GpuContextName not deferred.** Same function writes the context
+   name ("rocprofv3") without calling `DeferItem()`. Even after fixing
+   bug 1, a late-connecting client sees the GPU context but it appears
+   unnamed in the profiler. Use `check_gpu_ctx_name` to verify.
+
+3. **Kernel symbols dropped before init.** The `data->init` guard at the
    top of `tool_callback_tracing_callback()` blocks all callbacks before
    the GPU context is allocated. Kernel symbol registrations
    (`CODE_OBJECT_DEVICE_KERNEL_SYMBOL_REGISTER`) happen at HIP init
@@ -36,9 +41,24 @@ make
 tracy-capture -o repro.tracy -s 5
 ```
 
+## Verifying the context name
+
+`check_gpu_ctx_name` loads a `.tracy` file and prints the GPU context
+names. Build it against the Tracy server library (e.g. from a capture
+build directory) and run:
+
+```bash
+./check_gpu_ctx_name repro.tracy
+# Expected (patched):   "GPU context 0: rocprofv3"
+# Expected (unpatched): "GPU context 0: (unnamed)"
+```
+
+Exit codes: 0 = all contexts named, 2 = unnamed context found.
+
 ## What to expect
 
 | Tracy version | Result |
 |---|---|
 | Unpatched | `tracy-capture` crashes: `Assertion 'ctx' failed` |
-| Patched   | Capture succeeds with ~50 GPU zones (`vectorAdd`) with kernel names |
+| Patched (GpuNewContext only) | Capture succeeds but GPU context is unnamed |
+| Fully patched | Capture succeeds with ~50 GPU zones and context named "rocprofv3" |

examples/RocprofOnDemandRepro/check_gpu_ctx_name.cpp

@@ -0,0 +1,65 @@
+// Loads a .tracy file and prints the GPU context names.
+// Used to verify that on-demand profiling correctly defers the
+// GpuContextName message so late-connecting clients see the name.
+//
+// Usage: ./check_gpu_ctx_name trace.tracy
+// Expected output: "GPU context 0: rocprofv3"
+// If name is missing: "GPU context 0: (unnamed)"
+
+#include <cstdio>
+#include <cstdlib>
+#include "server/TracyFileRead.hpp"
+#include "server/TracyWorker.hpp"
+
+int main( int argc, char** argv )
+{
+    if( argc != 2 )
+    {
+        fprintf( stderr, "Usage: %s <trace.tracy>\n", argv[0] );
+        return 1;
+    }
+
+    try
+    {
+        auto f = std::unique_ptr<tracy::FileRead>( tracy::FileRead::Open( argv[1] ) );
+        if( !f )
+        {
+            fprintf( stderr, "Cannot open %s\n", argv[1] );
+            return 1;
+        }
+
+        tracy::Worker worker( *f, tracy::EventType::None, false );
+
+        const auto& gpuData = worker.GetGpuData();
+        if( gpuData.empty() )
+        {
+            printf( "No GPU contexts found.\n" );
+            return 1;
+        }
+
+        bool all_named = true;
+        for( size_t i = 0; i < gpuData.size(); i++ )
+        {
+            const auto& ctx = gpuData[i];
+            if( ctx->name.Active() )
+            {
+                const char* name = worker.GetString( ctx->name );
+                bool has_name = name && name[0] != '\0';
+                printf( "GPU context %zu: %s\n", i, has_name ? name : "(unnamed)" );
+                if( !has_name ) all_named = false;
+            }
+            else
+            {
+                printf( "GPU context %zu: (unnamed)\n", i );
+                all_named = false;
+            }
+        }
+
+        return all_named ? 0 : 2;
+    }
+    catch( const std::exception& e )
+    {
+        fprintf( stderr, "Error: %s\n", e.what() );
+        return 1;
+    }
+}

public/client/TracyRocprof.cpp

@@ -124,6 +124,9 @@ uint8_t gpu_context_allocate( ToolData* data )
         tracy::MemWrite( &item->gpuContextNameFat.context, context_id );
         tracy::MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)cloned_name );
         tracy::MemWrite( &item->gpuContextNameFat.size, name_length );
+#ifdef TRACY_ON_DEMAND
+        GetProfiler().DeferItem( *item );
+#endif
         tracy::Profiler::QueueSerialFinish();
     }