Test improvements with realistic CUDA API stack traces

Makefile

@@ -8,6 +8,11 @@ LIB_NAME = libparcagpucupti.so.$(CUDA_MAJOR)
 # Default target: build all CUDA versions (12 & 13) for both architectures and test infrastructure
 all: cupti-all-versions test-infra
 
+local:
+	mkdir -p build && \
+    cmake -DCMAKE_BUILD_TYPE=RelWithDebInfo -B build -S cupti && \
+    cd build && make -j$(nproc)
+
 # Build libparcagpucupti.so for AMD64 using Docker
 cupti-amd64:
 	@echo "=== Building $(LIB_NAME) for AMD64 with Docker (CUDA $(CUDA_MAJOR)) ==="
@@ -104,21 +109,28 @@ push-cuda-headers:
 		.
 	@echo "CUDA header images pushed to $(CUDA_HEADERS_REGISTRY):12 and :13"
 
-# Build test infrastructure with Zig
+# Build test infrastructure with CMake
 test-infra:
-	@echo "=== Building test infrastructure with Zig ==="
-	@zig build
+	@echo "=== Building test infrastructure with CMake ==="
+	@mkdir -p test/build
+	@cd test/build && cmake .. && make
 
 # Run tests (using AMD64 library)
 test: cupti-amd64 test-infra
 	@./test.sh
 
+# Run advanced test (8 GPUs @ 2500 launches/s, multi-threaded)
+test-advanced: cupti-amd64 test-infra
+	@echo "=== Running advanced test (8 GPUs @ 2500 launches/s) ==="
+	@cd test/build && LD_LIBRARY_PATH=.:$$LD_LIBRARY_PATH \
+		./test_cupti_prof ../../build/$(CUDA_MAJOR)/amd64/libparcagpucupti.so \
+		--threads=4 --num-gpus=8 --launch-rate=2500 --duration=10
+
 # Clean build artifacts
 clean:
 	@echo "=== Cleaning build artifacts ==="
 	@rm -rf cupti/build cupti/build-amd64 cupti/build-arm64 build
-	@rm -rf zig-out
-	@rm -rf .zig-cache
+	@rm -rf test/build
 	@echo "Clean complete"
 
 # Build and push multi-arch Docker images to ghcr.io
@@ -155,4 +167,4 @@ docker-test-run: docker-test-build
 	@docker run --rm parcagpu-test:latest $(ARGS)
 
 format:
-	clang-format -i -style=file cupti/*.[ch]
+	clang-format -i -style=file cupti/*.[ch]

kernel_names.txt

@@ -0,0 +1,63 @@
+sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x16x32_stage1_warpsize4x1x1_g1_tensor16x8x8_aligna4_alignc4_execute_kernel__5x_cudnn
+sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x32x16_stage1_warpsize4x1x1_g1_tensor16x8x8_aligna4_execute_kernel__5x_cudnn
+sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x32x16_stage1_warpsize4x1x1_g1_tensor16x8x8_alignc8_execute_kernel__5x_cudnn
+sm80_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x128x16_stage4_warpsize2x2x1_g1_tensor16x8x8_alignc4_execute_kernel__5x_cudnn
+_Z17wgrad_alg0_engineIfLi128ELi5ELi5ELi3ELi3ELi3ELb0ELi512EEviiiPKT_iPS0_S2_18kernel_grad_paramsyifiiii
+_Z20magma_sgemmEx_kernelIfffLb0ELb0ELi6ELi3ELi5ELi3ELi3EEviii6TensoriS0_iS0_iS0_iiiPKT_S3_S1_S1_i18cublasLtEpilogue_tiPKvl
+_Z20magma_sgemmEx_kernelIfffLb0ELb1ELi6ELi4ELi6ELi3ELi4EEviii6TensoriS0_iS0_iS0_iiiPKT_S3_S1_S1_i18cublasLtEpilogue_tiPKvl
+_Z22foldedNhwcToNchwKernelIfffLb1EL21cudnnKernelDataType_t0EEviiiiiiiPKT_PT0_iiiiiiiiiiT1_S6_N5cudnn15reduced_divisorES8_S8_
+_Z22nchwToFoldedNhwcKernelIfffLb1EL21cudnnKernelDataType_t2EEviiiiPKT_PT0_iiiiiiiiiiT1_S6_N5cudnn15reduced_divisorES8_S8_
+_Z22tensorTransformGenericIfffLb1ELb0ELb0EL21cudnnKernelDataType_t2EEv26cudnnTensorTransformStruct21tensorTransformParamsimPKT_PT0_T1_S8_
+_Z23implicit_convolve_sgemmIffLi1024ELi5ELi5ELi3ELi3ELi3ELi1ELb0ELb0ELb1EEviiiPKT_iPT0_S2_18kernel_conv_paramsyiffiPKS3_S7_bbii
+_Z23implicit_convolve_sgemmIffLi128ELi5ELi5ELi3ELi3ELi3ELi1ELb0ELb0ELb1EEviiiPKT_iPT0_S2_18kernel_conv_paramsyiffiPKS3_S7_bbii
+_ZN17cutlass__5x_cudnn6KernelI65cutlass_tensorop_s1688wgrad_optimized_tf32_64x64_32x5_nhwc_align4EEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelI66cutlass_tensorop_s1688dgrad_optimized_tf32_256x64_16x4_nhwc_align4EEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelI66cutlass_tensorop_s1688fprop_optimized_tf32_64x64_16x10_nhwc_align4EEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelI79cutlass_tensorop_s1688dgrad_optimized_tf32_256x64_16x4_nhwc_unity_stride_align4EEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelI79cutlass_tensorop_s1688dgrad_optimized_tf32_64x64_16x10_nhwc_unity_stride_align4EEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelINS_4conv6kernel23ImplicitGemmConvolutionINS1_11threadblock22ImplicitGemmMultistageINS_4gemm9GemmShapeILi64ELi64ELi32EEENS4_52Conv2dWgradOutputGradientTileAccessIteratorOptimizedINS_11MatrixShapeILi64ELi32EEENS_10tfloat32_tENS_9transform29PitchLinearWarpRakedThreadMapINS_16PitchLinearShapeILi64ELi32EEELi128ENSF_ILi8ELi4EEELi4EEENS_12AlignedArrayISC_Li4ELi16EEEEENSD_11threadblock25RegularTileAccessIteratorISB_SC_NS_6layout40ColumnMajorTensorOpMultiplicandCongruousILi32ELi32EEELi1ESI_Li16EEELNS_4arch14CacheOperation4KindE0ENS4_48Conv2dWgradActivationTileAccessIteratorOptimizedINSA_ILi32ELi64EEESC_SI_SK_EENSN_ISW_SC_NSO_37RowMajorTensorOpMultiplicandCongruousILi32ELi32EEELi0ESI_Li16EEELSU_0ENS6_11threadblock9MmaPolicyINS6_4warp11MmaTensorOpINS7_ILi32ELi32ELi32EEESC_SQ_SC_SZ_fNSO_8RowMajorENS13_17MmaTensorOpPolicyINSS_3MmaINS7_ILi16ELi8ELi8EEELi32ESC_S16_SC_NSO_11ColumnMajorEfS16_NSS_13OpMultiplyAddEEENSA_ILi1ELi1EEEEELi1ELb0EbEENSA_ILi0ELi0EEES1G_Li1EEELi5EbEENS_8epilogue11threadblock8EpilogueIS8_S1F_Li1ENS1K_22PredicatedTileIteratorINS1K_26OutputTileOptimalThreadMapINS1K_15OutputTileShapeILi64ELi8ELi2ELi1ELi1EEENS1O_ILi1ELi4ELi1ELi1ELi4EEELi128ELi4ELi32EEEfLb0ENSO_9NoPermuteELb0EEENS1J_4warp24FragmentIteratorTensorOpIS15_S19_fNS_5ArrayIfLi4ELb1EEES16_EENS1U_20TileIteratorTensorOpIS15_S19_fS16_EENS1K_18SharedLoadIteratorINS1R_18CompactedThreadMapEfLi16EEENS1J_6thread17LinearCombinationIfLi4EffLNS24_9ScaleType4KindE0ELNS_15FloatRoundStyleE2EEENSA_ILi0ELi8EEELi2ELi1EEENS11_30GemmIdentityThreadblockSwizzleILi4EEELNS1_8OperatorE2ENS1_17Conv2dProblemSizeELNS1_9GroupModeE0EEEEEvNT_6ParamsE
+_ZN17cutlass__5x_cudnn6KernelINS_9reduction6kernel12ReduceSplitKINS_11MatrixShapeILi4ELi128EEENS_8epilogue6thread17LinearCombinationIfLi4EffLNS7_9ScaleType4KindE0ELNS_15FloatRoundStyleE2EEENS1_6thread9ReduceAddIffLi4EEELi4EEEEEvNT_6ParamsE
+_ZN2at6native13reduce_kernelILi128ELi4ENS0_8ReduceOpIfNS0_14func_wrapper_tIfZNS0_11sum_functorIfffEclERNS_14TensorIteratorEEUlffE_EEjfLi4ELi4EEEEEvT1_
+_ZN2at6native13reduce_kernelILi512ELi1ENS0_8ReduceOpIfNS0_14func_wrapper_tIfZNS0_11sum_functorIfffEclERNS_14TensorIteratorEEUlffE_EEjfLi4ELi4EEEEEvT1_
+_ZN2at6native13reduce_kernelILi512ELi1ENS0_8ReduceOpIfNS0_7MeanOpsIffffEEjfLi4ELi4EEEEEvT1_
+_ZN2at6native18elementwise_kernelILi128ELi2EZNS0_22gpu_kernel_impl_nocastINS0_13BinaryFunctorIfffNS0_15binary_internal10DivFunctorIfEEEEEEvRNS_18TensorIteratorBaseERKT_EUliE_EEviT1_
+_ZN2at6native18elementwise_kernelILi128ELi2EZNS0_22gpu_kernel_impl_nocastINS0_15CUDAFunctor_addIfEEEEvRNS_18TensorIteratorBaseERKT_EUliE_EEviT1_
+_ZN2at6native18elementwise_kernelILi128ELi2EZNS0_22gpu_kernel_impl_nocastIZZZNS0_23direct_copy_kernel_cudaERNS_18TensorIteratorBaseEENKUlvE1_clEvENKUlvE5_clEvEUlfE_EEvS4_RKT_EUliE_EEviT1_
+_ZN2at6native18elementwise_kernelILi128ELi2EZNS0_22gpu_kernel_impl_nocastIZZZNS0_24mse_backward_cuda_kernelERNS_14TensorIteratorERKN3c106ScalarEENKUlvE_clEvENKUlvE0_clEvEUlfffE_EEvRNS_18TensorIteratorBaseERKT_EUliE_EEviT1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_11FillFunctorIfEESt5arrayIPcLm1EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_13AUnaryFunctorIfffNS0_15binary_internal10MulFunctorIfEEEESt5arrayIPcLm2EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_13BinaryFunctorIfffZNS0_67_GLOBAL__N__06c19b41_28_ActivationThresholdKernel_cu_c1096568_2903721threshold_kernel_implIfEEvRNS_18TensorIteratorBaseET_S7_EUlffE_EESt5arrayIPcLm3EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_13BUnaryFunctorIfffNS0_15binary_internal10MulFunctorIfEEEESt5arrayIPcLm2EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_15CUDAFunctor_addIfEESt5arrayIPcLm3EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4ENS0_21CUDAFunctorOnSelf_addIfEESt5arrayIPcLm2EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4EZZZNS0_15mse_kernel_cudaERNS_18TensorIteratorBaseEENKUlvE_clEvENKUlvE0_clEvEUlffE_St5arrayIPcLm3EEEEviT0_T1_
+_ZN2at6native29vectorized_elementwise_kernelILi4EZZZNS0_49_GLOBAL__N__d2ba64fb_16_TensorCompare_cu_71e06f4e19launch_clamp_scalarERNS_18TensorIteratorBaseEN3c106ScalarES6_NS0_6detail11ClampLimitsEENKUlvE_clEvENKUlvE5_clEvEUlfE_St5arrayIPcLm2EEEEviT0_T1_
+_ZN2at6native49_GLOBAL__N__7861cdb6_16_ReflectionPad_cu_f800513927reflection_pad2d_out_kernelIfEEvPKT_PS3_lliiiiiii
+_ZN2at6native49_GLOBAL__N__7861cdb6_16_ReflectionPad_cu_f800513936reflection_pad2d_backward_out_kernelIfEEvPT_PKS3_lliiiiiii
+_ZN2at6native50_GLOBAL__N__b6ca21f8_17_ForeachUnaryOp_cu_77fe623d25multi_tensor_apply_kernelINS1_18TensorListMetadataILi2EEENS1_14UnaryOpFunctorIfLi2ELi1ELi1EEEJNS0_4SqrtIfEEEEEvT_T0_DpT1_
+_ZN2at6native52_GLOBAL__N__4671fab5_19_ForeachTernaryOp_cu_5285c63625multi_tensor_apply_kernelINS1_18TensorListMetadataILi2EEENS1_22TernaryOpScalarFunctorIfLi2ELi2ELi0EEEJNS0_11LerpFunctorIfEEfEEEvT_T0_DpT1_
+_ZN2at6native52_GLOBAL__N__e57809e0_19_DilatedMaxPool2d_cu_6258b57421max_pool_forward_nchwIfEEviPKT_llliiiiiiiiiiPS3_Pl
+_ZN2at6native52_GLOBAL__N__e57809e0_19_DilatedMaxPool2d_cu_6258b57422max_pool_backward_nchwIffEEvPKT_PKlillliiiiiiiiiiPS3_
+_ZN2at6native53_GLOBAL__N__222ba01c_20_UpSampleNearest2d_cu_198c5ce228upsample_nearest2d_out_frameIfXadL_ZNS0_37nearest_neighbor_compute_source_indexEfiiEEEEvPKT_PS3_mmmmmff
+_ZN2at6native53_GLOBAL__N__222ba01c_20_UpSampleNearest2d_cu_198c5ce237upsample_nearest2d_backward_out_frameIffXadL_ZNS0_40nearest_neighbor_bw_compute_source_indexEfiiEEEEvPKT_mmmmmmPS3_ff
+_ZN2at6native54_GLOBAL__N__f8a10d72_21_ForeachPointwiseOp_cu_ef478fac25multi_tensor_apply_kernelINS1_18TensorListMetadataILi3EEENS1_24PointwiseOpScalarFunctorIfLi3ELi3ELi0EEEJSt10multipliesIfEfEEEvT_T0_DpT1_
+_ZN2at6native54_GLOBAL__N__f8a10d72_21_ForeachPointwiseOp_cu_ef478fac25multi_tensor_apply_kernelINS1_28TensorListScalarListMetadataIfLi3EEENS1_28PointwiseOpScalarListFunctorIfLi3ELi3ELi0EEEJSt7dividesIfEEEEvT_T0_DpT1_
+_ZN2at6native57_GLOBAL__N__e658eeb9_24_ForeachBinaryOpScalar_cu_86b9896c25multi_tensor_apply_kernelINS1_18TensorListMetadataILi1EEENS1_21BinaryOpScalarFunctorIfLi1ELi1ELi0EEEJSt10multipliesIfEfEEEvT_T0_DpT1_
+_ZN2at6native57_GLOBAL__N__e658eeb9_24_ForeachBinaryOpScalar_cu_86b9896c25multi_tensor_apply_kernelINS1_18TensorListMetadataILi1EEENS1_21BinaryOpScalarFunctorIfLi1ELi1ELi0EEEJSt4plusIfEfEEEvT_T0_DpT1_
+_ZN2at6native61_GLOBAL__N__538ad94b_28_ForeachBinaryOpScalarList_cu_dda10a6925multi_tensor_apply_kernelINS1_28TensorListScalarListMetadataIfLi1EEENS1_25BinaryOpScalarListFunctorIfLi1ELi1ELi0EEEJSt7dividesIfEEEEvT_T0_DpT1_
+_ZN5cudnn17winograd_nonfused20winogradWgradData4x4IffEEvNS0_18WinogradDataParamsIT_T0_EE
+_ZN5cudnn17winograd_nonfused21winogradWgradDelta4x4IffEEvNS0_19WinogradDeltaParamsIT_T0_EE
+_ZN5cudnn17winograd_nonfused22winogradWgradOutput4x4IffEEvNS0_25WinogradWgradOutputParamsIT_T0_EE
+_ZN5cudnn19engines_precompiled16nchwToNhwcKernelIfffLb0ELb1EL21cudnnKernelDataType_t2EEEvNS0_18nchw2nhwc_params_tIT1_EEPKT_PT0_
+_ZN5cudnn19engines_precompiled16nhwcToNchwKernelIfffLb1ELb0EL21cudnnKernelDataType_t0EEEvNS0_18nhwc2nchw_params_tIT1_EEPKT_PT0_
+_ZN5cudnn19engines_precompiled24scalePackedTensor_kernelIffEEvlPT_T0_
+_ZN5cudnn21bn_bw_1C11_kernel_newIff6float2Li512ELb1ELi1EEEvT0_S2_S2_S2_17cudnnTensorStructPKT_S3_S6_S3_PS4_PKS2_PS2_SA_S9_S9_S2_
+_ZN5cudnn21bn_bw_1C11_singlereadIfLi512ELb1ELi1ELi2ELi0EEEvNS_15bn_bw_1C11_argsIT_EE
+_ZN5cudnn24bn_fw_tr_1C11_singlereadIfLi512ELb1ELi1ELi2ELi0EEEvNS_18bn_fw_tr_1C11_argsIT_EE
+_ZN5cudnn25bn_fw_tr_1C11_kernel_NCHWIffiLi512ELb1ELi1ELb1EEEv17cudnnTensorStructPKT_S1_PS2_PKT0_S8_S6_S6_PS6_S9_S9_S9_S6_S6_
+_ZN5cudnn3cnn17wgrad_alg1_engineIffLi512ELi6ELi5ELi3ELi3ELi3ELb0ELb1EEEviiiPKT_iPT0_S4_18kernel_grad_paramsyiffiiPiS8_ii
+_ZN5cudnn6detail12dgrad_engineIfLi128ELi6ELi8ELi3ELi3ELi5ELb0EEEviiiPKT_iS4_iPS2_18kernel_grad_paramsyiyifiii
+_ZN7cutlass7Kernel2I41cutlass_80_simt_sgemm_64x64_8x5_tn_align1EEvNT_6ParamsE
+_ZN7cutlass7Kernel2I42cutlass_80_simt_sgemm_128x32_8x5_nn_align1EEvNT_6ParamsE
+_ZN7cutlass7Kernel2I42cutlass_80_simt_sgemm_128x32_8x5_nt_align1EEvNT_6ParamsE
+_ZN7cutlass7Kernel2I42cutlass_80_simt_sgemm_128x32_8x5_tn_align1EEvNT_6ParamsE

test.sh

@@ -21,8 +21,9 @@ for arg in "$@"; do
 done
 
 echo ""
-echo "=== Building test infrastructure with Zig ==="
-zig build
+echo "=== Building test infrastructure with CMake ==="
+mkdir -p test/build
+cd test/build && cmake .. && make && cd ../..
 
 # Start bpftrace if requested
 if [ "$USE_BPFTRACE" -eq 1 ]; then
@@ -48,11 +49,11 @@ fi
 echo ""
 echo "=== Running test program ==="
 # Set LD_LIBRARY_PATH so the test can find libcupti.so at runtime
-# Set PARCAGPU_DEBUG to enable debug output
-export LD_LIBRARY_PATH="$(pwd)/zig-out/lib:$LD_LIBRARY_PATH"
-export PARCAGPU_DEBUG=1
+# Set PARCAGPU_DEBUG externally to enable debug output
+# Set PARCAGPU_RATE_LIMIT externally to override default (100/s)
+export LD_LIBRARY_PATH="$(pwd)/test/build:$LD_LIBRARY_PATH"
 # Use the CMake-built library with real CUPTI
-zig-out/bin/test_cupti_prof build/$CUDA_MAJOR/$ARCH/libparcagpucupti.so "$@"
+test/build/test_cupti_prof build/libparcagpucupti.so --kernel-names=kernel_names.txt "$@"
 
 # If bpftrace was started, stop it and show results
 if [ "$USE_BPFTRACE" -eq 1 ]; then

test/CMakeLists.txt

@@ -0,0 +1,59 @@
+cmake_minimum_required(VERSION 3.18)
+project(parcagpu_test C)
+
+# CUDA include paths
+if(NOT DEFINED CUDA_ROOT)
+    set(CUDA_ROOT "/usr/local/cuda" CACHE PATH "CUDA installation directory")
+endif()
+
+set(CUDA_INCLUDE_DIRS
+    "${CUDA_ROOT}/include"
+    "${CUDA_ROOT}/extras/CUPTI/include"
+)
+
+# Mock CUPTI library
+add_library(cupti SHARED mock_cupti.c)
+target_include_directories(cupti PRIVATE ${CUDA_INCLUDE_DIRS})
+set_target_properties(cupti PROPERTIES
+    C_STANDARD 11
+    C_STANDARD_REQUIRED ON
+    POSITION_INDEPENDENT_CODE ON
+)
+
+# Detect CUDA major version from CUPTI library
+execute_process(
+    COMMAND bash -c "ls ${CUDA_ROOT}/lib64/libcupti.so.* 2>/dev/null | grep -oE 'libcupti\\.so\\.[0-9]+$' | head -1 | grep -oE '[0-9]+$'"
+    OUTPUT_VARIABLE CUDA_MAJOR_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+if(NOT CUDA_MAJOR_VERSION)
+    set(CUDA_MAJOR_VERSION "12")
+endif()
+message(STATUS "Detected CUDA major version: ${CUDA_MAJOR_VERSION}")
+
+# Create versioned symlink matching installed CUDA
+add_custom_command(TARGET cupti POST_BUILD
+    COMMAND ${CMAKE_COMMAND} -E create_symlink
+    $<TARGET_FILE_NAME:cupti>
+    ${CMAKE_CURRENT_BINARY_DIR}/libcupti.so.${CUDA_MAJOR_VERSION}
+)
+
+# Test executable
+add_executable(test_cupti_prof test_cupti_prof.c)
+target_include_directories(test_cupti_prof PRIVATE ${CUDA_INCLUDE_DIRS})
+target_compile_options(test_cupti_prof PRIVATE
+    -D_POSIX_C_SOURCE=199309L
+    -Wall
+    -Wextra
+)
+target_link_libraries(test_cupti_prof PRIVATE dl pthread)
+set_target_properties(test_cupti_prof PROPERTIES
+    C_STANDARD 11
+    C_STANDARD_REQUIRED ON
+)
+
+# Installation
+install(TARGETS cupti test_cupti_prof
+    RUNTIME DESTINATION bin
+    LIBRARY DESTINATION lib
+)