mirror of
https://gitlab.com/libeigen/eigen.git
synced 2026-04-10 11:34:33 +08:00
014f12f11a
Add the operator interface needed for GPU iterative solvers: - BLAS Level-1 on DeviceMatrix: dot(), norm(), squaredNorm(), setZero(), noalias(), operator+=/-=/\*= dispatching to cuBLAS axpy/scal/dot/nrm2. - DeviceScalar<Scalar>: device-resident scalar returned by reductions. Defers host sync until value is read (implicit conversion). Device-side division via NPP for real types. - GpuContext: stream-borrowing constructor, setThreadLocal(), cublasLtHandle(), cusparseHandle(). - GEMM upgraded from cublasGemmEx to cublasLtMatmul with heuristic algorithm selection and plan caching. - GpuSparseContext: GpuContext& constructor for same-stream execution, deviceView() returning DeviceSparseView with operator* for device-resident SpMV (d_y = d_A * d_x). - geam expressions: d_C = d_A + alpha * d_B via cublasXgeam. - GpuSVD::matrixV() convenience wrapper. These additions make DeviceMatrix usable as a VectorType in Eigen algorithm templates. Conjugate gradient is the motivating example and is tested against CPU ConjugateGradient for correctness. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
92 lines
4.1 KiB
CMake
92 lines
4.1 KiB
CMake
# GPU benchmarks require CUDA runtime + cuSOLVER.
|
|
# Build separately from the main benchmark tree since they need CUDA toolchain.
|
|
#
|
|
# Usage:
|
|
# cmake -G Ninja -B build-bench-gpu -S benchmarks/GPU \
|
|
# -DCMAKE_CUDA_ARCHITECTURES=89
|
|
# cmake --build build-bench-gpu
|
|
#
|
|
# Profiling:
|
|
# nsys profile --trace=cuda ./build-bench-gpu/bench_gpu_solvers
|
|
# ncu --set full -o profile ./build-bench-gpu/bench_gpu_solvers --benchmark_filter=BM_GpuLLT_Compute/4096
|
|
|
|
cmake_minimum_required(VERSION 3.18)
|
|
project(EigenGpuBenchmarks CXX CUDA)
|
|
|
|
find_package(benchmark REQUIRED)
|
|
find_package(CUDAToolkit REQUIRED)
|
|
|
|
set(EIGEN_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../..")
|
|
|
|
function(eigen_add_gpu_benchmark name source)
|
|
cmake_parse_arguments(BENCH "" "" "LIBRARIES;DEFINITIONS" ${ARGN})
|
|
if(NOT IS_ABSOLUTE "${source}")
|
|
set(source "${CMAKE_CURRENT_SOURCE_DIR}/${source}")
|
|
endif()
|
|
add_executable(${name} ${source})
|
|
target_include_directories(${name} PRIVATE
|
|
${EIGEN_SOURCE_DIR}
|
|
${CUDAToolkit_INCLUDE_DIRS})
|
|
target_link_libraries(${name} PRIVATE
|
|
benchmark::benchmark benchmark::benchmark_main
|
|
CUDA::cudart CUDA::cusolver CUDA::cublas)
|
|
if(BENCH_LIBRARIES)
|
|
target_link_libraries(${name} PRIVATE ${BENCH_LIBRARIES})
|
|
endif()
|
|
target_compile_options(${name} PRIVATE -O3 -DNDEBUG)
|
|
target_compile_definitions(${name} PRIVATE EIGEN_USE_GPU)
|
|
if(BENCH_DEFINITIONS)
|
|
target_compile_definitions(${name} PRIVATE ${BENCH_DEFINITIONS})
|
|
endif()
|
|
endfunction()
|
|
|
|
# Solver benchmarks: LLT/LU compute + solve, host vs device paths, CPU baselines.
|
|
eigen_add_gpu_benchmark(bench_gpu_solvers bench_gpu_solvers.cpp)
|
|
eigen_add_gpu_benchmark(bench_gpu_solvers_float bench_gpu_solvers.cpp DEFINITIONS SCALAR=float)
|
|
|
|
# Chaining benchmarks: async pipeline efficiency, host-roundtrip vs device chain.
|
|
eigen_add_gpu_benchmark(bench_gpu_chaining bench_gpu_chaining.cpp)
|
|
eigen_add_gpu_benchmark(bench_gpu_chaining_float bench_gpu_chaining.cpp DEFINITIONS SCALAR=float)
|
|
|
|
# Batching benchmarks: multi-stream concurrency for many small systems.
|
|
eigen_add_gpu_benchmark(bench_gpu_batching bench_gpu_batching.cpp)
|
|
eigen_add_gpu_benchmark(bench_gpu_batching_float bench_gpu_batching.cpp DEFINITIONS SCALAR=float)
|
|
|
|
# FFT benchmarks: 1D/2D C2C, R2C, C2R throughput and plan reuse.
|
|
eigen_add_gpu_benchmark(bench_gpu_fft bench_gpu_fft.cpp LIBRARIES CUDA::cufft)
|
|
eigen_add_gpu_benchmark(bench_gpu_fft_double bench_gpu_fft.cpp LIBRARIES CUDA::cufft DEFINITIONS SCALAR=double)
|
|
|
|
# CG sync overhead benchmark: host vs device pointer mode for reductions.
|
|
# Uses CUDA kernels for device scalar arithmetic.
|
|
add_executable(bench_gpu_cg_sync bench_gpu_cg_sync.cu)
|
|
target_include_directories(bench_gpu_cg_sync PRIVATE
|
|
${EIGEN_SOURCE_DIR}
|
|
${CUDAToolkit_INCLUDE_DIRS})
|
|
target_link_libraries(bench_gpu_cg_sync PRIVATE
|
|
benchmark::benchmark benchmark::benchmark_main
|
|
CUDA::cudart CUDA::cusolver CUDA::cublas CUDA::cusparse CUDA::npps CUDA::nppc)
|
|
target_compile_options(bench_gpu_cg_sync PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-O3 --expt-relaxed-constexpr>)
|
|
target_compile_definitions(bench_gpu_cg_sync PRIVATE EIGEN_USE_GPU)
|
|
|
|
# GPU CG vs CPU CG comparison benchmark.
|
|
add_executable(bench_gpu_cg_vs_cpu bench_gpu_cg_vs_cpu.cu)
|
|
target_include_directories(bench_gpu_cg_vs_cpu PRIVATE
|
|
${EIGEN_SOURCE_DIR}
|
|
${CUDAToolkit_INCLUDE_DIRS})
|
|
target_link_libraries(bench_gpu_cg_vs_cpu PRIVATE
|
|
benchmark::benchmark benchmark::benchmark_main
|
|
CUDA::cudart CUDA::cusolver CUDA::cublas CUDA::cusparse CUDA::npps CUDA::nppc)
|
|
target_compile_options(bench_gpu_cg_vs_cpu PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-O3 --expt-relaxed-constexpr>)
|
|
target_compile_definitions(bench_gpu_cg_vs_cpu PRIVATE EIGEN_USE_GPU)
|
|
|
|
# Bundle Adjustment benchmark: GPU CG vs CPU CG on real BAL datasets.
|
|
add_executable(bench_gpu_ba bench_gpu_ba.cu)
|
|
target_include_directories(bench_gpu_ba PRIVATE
|
|
${EIGEN_SOURCE_DIR}
|
|
${CUDAToolkit_INCLUDE_DIRS})
|
|
target_link_libraries(bench_gpu_ba PRIVATE
|
|
benchmark::benchmark
|
|
CUDA::cudart CUDA::cusolver CUDA::cublas CUDA::cusparse CUDA::npps CUDA::nppc)
|
|
target_compile_options(bench_gpu_ba PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-O3 --expt-relaxed-constexpr>)
|
|
target_compile_definitions(bench_gpu_ba PRIVATE EIGEN_USE_GPU)
|