Add HIP to make and cmake files

jakurzak · jakurzak · commit 5edafd6893f9 · 2023-08-16T15:44:17.000-04:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -2,7 +2,13 @@ cmake_minimum_required(VERSION 3.11)
 
 execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
 if(has_nvcc STREQUAL "")
-    project(qsim)
+    execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
+    if(has_hipcc STREQUAL "")
+        project(qsim)
+    else()
+        project(qsim LANGUAGES CXX HIP)
+        ADD_SUBDIRECTORY(pybind_interface/hip)
+    endif()
 else()
     project(qsim LANGUAGES CXX CUDA)
     ADD_SUBDIRECTORY(pybind_interface/cuda)
diff --git a/Makefile b/Makefile
@@ -6,10 +6,12 @@ TESTS = run-cxx-tests
 
 CXX=g++
 NVCC=nvcc
+HIPCC=hipcc
 
 CXXFLAGS = -O3 -fopenmp
 ARCHFLAGS = -march=native
 NVCCFLAGS = -O3
+HIPCCFLAGS = -O3
 
 # CUQUANTUM_ROOT should be set.
 CUSTATEVECFLAGS = -I$(CUQUANTUM_ROOT)/include -L${CUQUANTUM_ROOT}/lib -L$(CUQUANTUM_ROOT)/lib64 -lcustatevec -lcublas
@@ -22,6 +24,8 @@ export ARCHFLAGS
 export NVCC
 export NVCCFLAGS
 export CUSTATEVECFLAGS
+export HIPCC
+export HIPCCFLAGS
 
 ifeq ($(PYBIND11), true)
   TARGETS += pybind
@@ -43,6 +47,10 @@ qsim-cuda:
 qsim-custatevec:
 	$(MAKE) -C apps/ qsim-custatevec
 
+.PHONY: qsim-hip
+qsim-hip:
+	$(MAKE) -C apps/ qsim-hip
+
 .PHONY: pybind
 pybind:
 	$(MAKE) -C pybind_interface/ pybind
@@ -59,6 +67,10 @@ cuda-tests:
 custatevec-tests:
 	$(MAKE) -C tests/ custatevec-tests
 
+.PHONY: hip-tests
+hip-tests:
+	$(MAKE) -C tests/ hip-tests
+
 .PHONY: run-cxx-tests
 run-cxx-tests: cxx-tests
 	$(MAKE) -C tests/ run-cxx-tests
@@ -71,6 +83,10 @@ run-cuda-tests: cuda-tests
 run-custatevec-tests: custatevec-tests
 	$(MAKE) -C tests/ run-custatevec-tests
 
+.PHONY: run-hip-tests
+run-hip-tests: hip-tests
+	$(MAKE) -C tests/ run-hip-tests
+
 PYTESTS = $(shell find qsimcirq_tests/ -name '*_test.py')
 
 .PHONY: run-py-tests
diff --git a/apps/Makefile b/apps/Makefile
@@ -7,6 +7,9 @@ CUDA_TARGETS := $(CUDA_TARGETS:%cuda.cu=%cuda.x)
 CUSTATEVEC_TARGETS = $(shell find . -maxdepth 1 -name "*custatevec.cu")
 CUSTATEVEC_TARGETS := $(CUSTATEVEC_TARGETS:%custatevec.cu=%custatevec.x)
 
+HIP_TARGETS = $(shell find . -maxdepth 1 -name '*cuda.cu')
+HIP_TARGETS := $(HIP_TARGETS:%cuda.cu=%hip.x)
+
 .PHONY: qsim
 qsim: $(CXX_TARGETS)
 
@@ -16,6 +19,9 @@ qsim-cuda: $(CUDA_TARGETS)
 .PHONY: qsim-custatevec
 qsim-custatevec: $(CUSTATEVEC_TARGETS)
 
+.PHONY: qsim-hip
+qsim-hip: $(HIP_TARGETS)
+
 %.x: %.cc
 	$(CXX) -o ./$@ $< $(CXXFLAGS) $(ARCHFLAGS)
 
@@ -25,6 +31,9 @@ qsim-custatevec: $(CUSTATEVEC_TARGETS)
 %custatevec.x: %custatevec.cu
 	$(NVCC) -o ./$@ $< $(NVCCFLAGS) $(CUSTATEVECFLAGS)
 
+%hip.x: %cuda.cu
+	$(HIPCC) -o ./$@ $< $(HIPCCFLAGS)
+
 .PHONY: clean
 clean:
 	-rm -f ./*.x ./*.a ./*.so ./*.mod
diff --git a/apps/make.sh b/apps/make.sh
@@ -23,9 +23,15 @@ g++ -O3 -march=native -fopenmp -o qsim_amplitudes.x qsim_amplitudes.cc
 g++ -O3 -march=native -fopenmp -o qsimh_base.x qsimh_base.cc
 g++ -O3 -march=native -fopenmp -o qsimh_amplitudes.x qsimh_amplitudes.cc
 
-nvcc -O3 -o qsim_base_cuda.x qsim_base_cuda.cu
-nvcc -O3 -o qsim_qtrajectory_cuda.x qsim_qtrajectory_cuda.cu
+if command -v nvcc &>/dev/null; then
+    nvcc -O3 -o qsim_base_cuda.x qsim_base_cuda.cu
+    nvcc -O3 -o qsim_qtrajectory_cuda.x qsim_qtrajectory_cuda.cu
 
-# CUQUANTUM_ROOT should be set.
-CUSTATEVECFLAGS="-I${CUQUANTUM_ROOT}/include -L${CUQUANTUM_ROOT}/lib -L${CUQUANTUM_ROOT}/lib64 -lcustatevec -lcublas"
-nvcc -O3 $CUSTATEVECFLAGS -o qsim_base_custatevec.x qsim_base_custatevec.cu
+    if [ -n "$CUQUANTUM_ROOT" ]; then
+        CUSTATEVECFLAGS="-I${CUQUANTUM_ROOT}/include -L${CUQUANTUM_ROOT}/lib -L${CUQUANTUM_ROOT}/lib64 -lcustatevec -lcublas"
+        nvcc -O3 $CUSTATEVECFLAGS -o qsim_base_custatevec.x qsim_base_custatevec.cu
+    fi
+elif command -v hipcc &>/dev/null; then
+    hipcc -O3 -o qsim_base_hip.x qsim_base_cuda.cu
+    hipcc -O3 -o qsim_qtrajectory_hip.x qsim_qtrajectory_cuda.cu
+fi
diff --git a/pybind_interface/Makefile b/pybind_interface/Makefile
@@ -5,6 +5,7 @@ QSIMLIB_AVX2 = ../qsimcirq/qsim_avx2`python3-config --extension-suffix`
 QSIMLIB_AVX512 = ../qsimcirq/qsim_avx512`python3-config --extension-suffix`
 QSIMLIB_CUDA = ../qsimcirq/qsim_cuda`python3-config --extension-suffix`
 QSIMLIB_CUSTATEVEC = ../qsimcirq/qsim_custatevec`python3-config --extension-suffix`
+QSIMLIB_HIP = ../qsimcirq/qsim_hip`python3-config --extension-suffix`
 QSIMLIB_DECIDE = ../qsimcirq/qsim_decide`python3-config --extension-suffix`
 
 # The flags for the compilation of the simd-specific Pybind11 interfaces
@@ -13,21 +14,29 @@ PYBINDFLAGS_SSE = -msse4.1 -Wall -shared -std=c++17 -fPIC `python3 -m pybind11 -
 PYBINDFLAGS_AVX2 = -mavx2 -mfma -Wall -shared -std=c++17 -fPIC `python3 -m pybind11 --includes`
 PYBINDFLAGS_AVX512 = -mavx512f -mbmi2 -Wall -shared -std=c++17 -fPIC `python3 -m pybind11 --includes`
 
-# The flags for the compilation of GPU-specific Pybind11 interfaces
+# The flags for the compilation of CUDA-specific Pybind11 interfaces
 PYBINDFLAGS_CUDA = -std=c++17 -x cu -Xcompiler "-Wall -shared -fPIC `python3 -m pybind11 --includes`"
 
 # The flags for the compilation of cuStateVec-specific Pybind11 interfaces
 PYBINDFLAGS_CUSTATEVEC = $(CUSTATEVECFLAGS) $(PYBINDFLAGS_CUDA)
 
+# The flags for the compilation of HIP-specific Pybind11 interfaces
+PYBINDFLAGS_HIP = -std=c++17 -Wall -shared -fPIC `python3 -m pybind11 --includes`
+
 # Check for nvcc to decide compilation mode.
 ifeq ($(shell which $(NVCC)),)
+# Check for hipcc to decide compilation mode.
+ifeq ($(shell which $(HIPCC)),)
 pybind: pybind-cpu decide-cpu
 else
+pybind: pybind-cpu pybind-hip decide-hip
+endif
+else
 # Check for the cuStateVec library.
 ifeq ($(CUQUANTUM_ROOT),)
-pybind: pybind-cpu pybind-gpu decide-gpu
+pybind: pybind-cpu pybind-cuda decide-cuda
 else
-pybind: pybind-cpu pybind-gpu pybind-custatevec decide-custatevec
+pybind: pybind-cpu pybind-cuda pybind-custatevec decide-custatevec
 endif
 endif
 
@@ -40,14 +49,15 @@ pybind-cpu:
 
 .PHONY: decide-cpu
 decide-cpu:
+	echo "building decide-cpu"
 	$(CXX) decide/decide.cpp -o $(QSIMLIB_DECIDE) $(CXXFLAGS) $(PYBINDFLAGS_BASIC)
 
-.PHONY: pybind-gpu
-pybind-gpu:
+.PHONY: pybind-cuda
+pybind-cuda:
 	$(NVCC) cuda/pybind_main_cuda.cpp -o $(QSIMLIB_CUDA) $(NVCCFLAGS) $(PYBINDFLAGS_CUDA)
 
-.PHONY: decide-gpu
-decide-gpu:
+.PHONY: decide-cuda
+decide-cuda:
 	$(NVCC) decide/decide.cpp -o $(QSIMLIB_DECIDE) $(NVCCFLAGS) $(PYBINDFLAGS_CUDA)
 
 .PHONY: pybind-custatevec
@@ -58,6 +68,14 @@ pybind-custatevec:
 decide-custatevec:
 	$(NVCC) decide/decide.cpp -D__CUSTATEVEC__ -o $(QSIMLIB_DECIDE) $(NVCCFLAGS) $(PYBINDFLAGS_CUDA)
 
+.PHONY: pybind-hip
+pybind-hip:
+	$(HIPCC) hip/pybind_main_hip.cpp -o $(QSIMLIB_HIP) $(HIPCCFLAGS) $(PYBINDFLAGS_HIP)
+
+.PHONY: decide-hip
+decide-hip:
+	$(HIPCC) decide/decide.cpp -o $(QSIMLIB_DECIDE) $(HIPCCFLAGS) $(PYBINDFLAGS_HIP)
+
 .PHONY: clean
 clean:
 	-rm -f ./basic/*.x ./basic/*.a ./basic/*.so ./basic/*.mod $(QSIMLIB_BASIC)
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
@@ -2,7 +2,12 @@ cmake_minimum_required(VERSION 3.11)
 
 execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
 if(has_nvcc STREQUAL "")
-  project(qsim)
+  execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
+  if(has_hipcc STREQUAL "")
+    project(qsim)
+  else()
+    project(qsim LANGUAGES CXX HIP)
+  endif()
 else()
   project(qsim LANGUAGES CXX CUDA)
 endif()
@@ -22,7 +27,22 @@ endif()
 INCLUDE(../GetPybind11.cmake)
 
 if(has_nvcc STREQUAL "")
-  pybind11_add_module(qsim_decide decide.cpp)
+  if(has_hipcc STREQUAL "")
+    pybind11_add_module(qsim_decide decide.cpp)
+  else()
+    find_package(HIP REQUIRED)
+    find_package(PythonLibs 3.7 REQUIRED)
+
+    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
+
+    hip_add_library(qsim_decide MODULE decide.cpp)
+
+    set_target_properties(qsim_decide PROPERTIES
+        PREFIX "${PYTHON_MODULE_PREFIX}"
+        SUFFIX "${PYTHON_MODULE_EXTENSION}"
+    )
+    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
+  endif()
 else()
   find_package(PythonLibs 3.7 REQUIRED)
   find_package(CUDA REQUIRED)
diff --git a/tests/Makefile b/tests/Makefile
@@ -14,6 +14,9 @@ CUDA_TARGETS := $(CUDA_TARGETS:%cuda_test.cu=%cuda_test.x)
 CUSTATEVEC_TARGETS = $(shell find . -maxdepth 1 -name "*custatevec_test.cu")
 CUSTATEVEC_TARGETS := $(CUSTATEVEC_TARGETS:%custatevec_test.cu=%custatevec_test.x)
 
+HIP_TARGETS = $(shell find . -maxdepth 1 -name "*cuda_test.cu")
+HIP_TARGETS := $(HIP_TARGETS:%cuda_test.cu=%hip_test.x)
+
 GTEST_DIR = $(CURDIR)/googletest/googletest
 GMOCK_DIR = $(CURDIR)/googletest/googlemock
 
@@ -30,6 +33,9 @@ cuda-tests: $(CUDA_TARGETS)
 .PHONY: custatevec-tests
 custatevec-tests: $(CUSTATEVEC_TARGETS)
 
+.PHONY: hip-tests
+hip-tests: $(HIP_TARGETS)
+
 .PHONY: run-cxx-tests
 run-cxx-tests: cxx-tests
 	for exe in $(CXX_TARGETS); do if ! ./$$exe; then exit 1; fi; done
@@ -42,6 +48,10 @@ run-cuda-tests: cuda-tests
 run-custatevec-tests: custatevec-tests
 	for exe in $(CUSTATEVEC_TARGETS); do if ! ./$$exe; then exit 1; fi; done
 
+.PHONY: run-hip-tests
+run-hip-tests: hip-tests
+	for exe in $(HIP_TARGETS); do if ! ./$$exe; then exit 1; fi; done
+
 $(GTEST_DIR)/make:
 	-git submodule update --init --recursive googletest
 	mkdir -p $(GTEST_DIR)/make
@@ -56,6 +66,9 @@ $(GTEST_DIR)/make:
 %custatevec_test.x: %custatevec_test.cu $(GTEST_DIR)/make
 	$(NVCC) -o ./$@ $< $(TESTFLAGS) $(NVCCFLAGS) $(CUSTATEVECFLAGS)
 
+%hip_test.x: %cuda_test.cu $(GTEST_DIR)/make
+	$(HIPCC) -o ./$@ $< $(TESTFLAGS) $(HIPCCFLAGS)
+
 .PHONY: clean
 clean:
 	-rm -f ./*.x ./*.a ./*.so ./*.mod
diff --git a/tests/make.sh b/tests/make.sh
@@ -55,14 +55,22 @@ g++ -O3 -I$path_to_include -L$path_to_lib -fopenmp -o unitaryspace_basic_test.x
 g++ -O3 -I$path_to_include -L$path_to_lib -msse4 -fopenmp -o unitaryspace_sse_test.x unitaryspace_sse_test.cc -lgtest -lpthread
 g++ -O3 -I$path_to_include -L$path_to_lib -o vectorspace_test.x vectorspace_test.cc -lgtest -lpthread
 
-nvcc -O3 -I$path_to_include -L$path_to_lib -o hybrid_cuda_test.x hybrid_cuda_test.cu -lgtest -lpthread
-nvcc -O3 -I$path_to_include -L$path_to_lib -o qtrajectory_cuda_test.x qtrajectory_cuda_test.cu -lgtest -lpthread
-nvcc -O3 -I$path_to_include -L$path_to_lib -o simulator_cuda_test.x simulator_cuda_test.cu -lgtest -lpthread
-nvcc -O3 -I$path_to_include -L$path_to_lib -o statespace_cuda_test.x statespace_cuda_test.cu -lgtest -lpthread
+if command -v nvcc &>/dev/null; then
+    nvcc -O3 -I$path_to_include -L$path_to_lib -o hybrid_cuda_test.x hybrid_cuda_test.cu -lgtest -lpthread
+    nvcc -O3 -I$path_to_include -L$path_to_lib -o qtrajectory_cuda_test.x qtrajectory_cuda_test.cu -lgtest -lpthread
+    nvcc -O3 -I$path_to_include -L$path_to_lib -o simulator_cuda_test.x simulator_cuda_test.cu -lgtest -lpthread
+    nvcc -O3 -I$path_to_include -L$path_to_lib -o statespace_cuda_test.x statespace_cuda_test.cu -lgtest -lpthread
 
-# CUQUANTUM_ROOT should be set.
-CUSTATEVECFLAGS="-I${CUQUANTUM_ROOT}/include -L${CUQUANTUM_ROOT}/lib -L${CUQUANTUM_ROOT}/lib64 -lcustatevec -lcublas"
-nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o hybrid_custatevec_test.x hybrid_custatevec_test.cu -lgtest -lpthread
-nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o qtrajectory_custatevec_test.x qtrajectory_custatevec_test.cu -lgtest -lpthread
-nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o simulator_custatevec_test.x simulator_custatevec_test.cu -lgtest -lpthread
-nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o statespace_custatevec_test.x statespace_custatevec_test.cu -lgtest -lpthread
+    if [ -n "$CUQUANTUM_ROOT" ]; then
+        CUSTATEVECFLAGS="-I${CUQUANTUM_ROOT}/include -L${CUQUANTUM_ROOT}/lib -L${CUQUANTUM_ROOT}/lib64 -lcustatevec -lcublas"
+        nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o hybrid_custatevec_test.x hybrid_custatevec_test.cu -lgtest -lpthread
+        nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o qtrajectory_custatevec_test.x qtrajectory_custatevec_test.cu -lgtest -lpthread
+        nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o simulator_custatevec_test.x simulator_custatevec_test.cu -lgtest -lpthread
+        nvcc -O3 $CUSTATEVECFLAGS -I$path_to_include -L$path_to_lib -o statespace_custatevec_test.x statespace_custatevec_test.cu -lgtest -lpthread
+    fi
+elif command -v hipcc &>/dev/null; then
+    hipcc -O3 -I$path_to_include -L$path_to_lib -o hybrid_hip_test.x hybrid_cuda_test.cu -lgtest -lpthread
+    hipcc -O3 -I$path_to_include -L$path_to_lib -o qtrajectory_hip_test.x qtrajectory_cuda_test.cu -lgtest -lpthread
+    hipcc -O3 -I$path_to_include -L$path_to_lib -o simulator_hip_test.x simulator_cuda_test.cu -lgtest -lpthread
+    hipcc -O3 -I$path_to_include -L$path_to_lib -o statespace_hip_test.x statespace_cuda_test.cu -lgtest -lpthread
+fi
