Merge pull request #386 from QuEST-Kit/cuquantum

integrated a new cuQuantum and Thrust GPU backend

QuEST/CMakeLists.txt

@@ -37,6 +37,8 @@ option(USE_HIP "Whether to use HIP for GPU code compilation for AMD GPUs. Set to
 
 set(GPU_ARCH gfx90 CACHE STRING "GPU hardware dependent, used for AMD GPUs when USE_HIP=1. Lookup at https://llvm.org/docs/AMDGPUUsage.html#amdgpu-processor-table. Write without fullstop")
 
+option(USE_CUQUANTUM "Whether to use NVIDIA's cuQuantum library (requires prior installation) in lieu of QuEST's bespoke GPU kernel. Set to 1 to enable." 0)
+
 
 # *****************************************************************************
 # ***** NO CHANGES SHOULD BE REQUIRED FROM THE USER BEYOND THIS POINT *********
@@ -49,6 +51,7 @@ message(STATUS "OMP acceleration is ${MULTITHREADED}")
 message(STATUS "MPI distribution is ${DISTRIBUTED}")
 if (${GPUACCELERATED})
   message(STATUS "HIP compilation is ${USE_HIP}")
+  message(STATUS "cuQuantum compilation is ${USE_CUQUANTUM}")
 endif()
 
 
@@ -119,25 +122,28 @@ endif()
 if (GPUACCELERATED)
   if (USE_HIP)
 
-  if(NOT DEFINED HIP_PATH)
-    if(NOT DEFINED ENV{HIP_PATH})
-      message(WARNING "WARNING: HIP_PATH is not defiend. Using default HIP_PATH=/opt/rocm/hip    " ${HIP_VERSION})
-      set(HIP_PATH "/opt/rocm/hip" CACHE PATH "Path to which HIP has been installed")
-    else()
-      set(HIP_PATH $ENV{HIP_PATH}	CACHE PATH "Path to which HIP has been installed")
+    if(NOT DEFINED HIP_PATH)
+      if(NOT DEFINED ENV{HIP_PATH})
+        message(WARNING "WARNING: HIP_PATH is not defiend. Using default HIP_PATH=/opt/rocm/hip    " ${HIP_VERSION})
+        set(HIP_PATH "/opt/rocm/hip" CACHE PATH "Path to which HIP has been installed")
+      else()
+        set(HIP_PATH $ENV{HIP_PATH}	CACHE PATH "Path to which HIP has been installed")
+      endif()
     endif()
-  endif()
 
-  if(EXISTS "${HIP_PATH}")
-    set(CMAKE_MODULE_PATH "${HIP_PATH}/cmake" ${CMAKE_MODULE_PATH})
-    find_package(HIP REQUIRED)
-    message(STATUS "Found HIP: " ${HIP_VERSION})
-    message(STATUS "HIP PATH: " ${HIP_PATH})
-  endif()
-
-  ADD_DEFINITIONS( -DUSE_HIP )
-  ADD_DEFINITIONS( -D__HIP_PLATFORM_AMD__ )
+    if(EXISTS "${HIP_PATH}")
+      set(CMAKE_MODULE_PATH "${HIP_PATH}/cmake" ${CMAKE_MODULE_PATH})
+      find_package(HIP REQUIRED)
+      message(STATUS "Found HIP: " ${HIP_VERSION})
+      message(STATUS "HIP PATH: " ${HIP_PATH})
+    endif()
+      
+    ADD_DEFINITIONS( -DUSE_HIP )
+    ADD_DEFINITIONS( -D__HIP_PLATFORM_AMD__ )
 
+  elseif (USE_CUQUANTUM)
+    find_package(CUDA REQUIRED)
+    ADD_DEFINITIONS( -DUSE_CUQUANTUM )
   else()
     find_package(CUDA REQUIRED)
   endif()  
@@ -280,7 +286,12 @@ endif()
 # ----- C++ COMPILER FLAGS --------------------------------------------------
 
 # set C++ flags that are common between compilers and build types
-set (CMAKE_CXX_STANDARD 98)
+if (USE_CUQUANTUM)
+  set(CMAKE_CXX_STANDARD 14)
+  set(CMAKE_CXX_STANDARD_REQUIRED ON)
+else ()
+  set (CMAKE_CXX_STANDARD 98)
+endif ()
 
 # Use -O2 for all but debug mode by default 
 if (NOT("${CMAKE_BUILD_TYPE}" STREQUAL "Debug"))
@@ -412,6 +423,14 @@ target_link_libraries(QuEST PUBLIC ${MPI_C_LIBRARIES})
 # ----- GPU -------------------------------------------------------------------
 if (USE_HIP)
   target_link_libraries(QuEST PUBLIC  ${HIP_PATH}/lib/libamdhip64.so )
+elseif (USE_CUQUANTUM)
+  find_library(CUQUANTUM_LIBRARIES custatevec)
+  if (NOT CUQUANTUM_LIBRARIES)
+    message(FATAL_ERROR "cuQuantum library (specifically custatevec) not found")
+  endif ()
+
+  target_link_libraries(QuEST ${CUDA_LIBRARIES} ${CUQUANTUM_LIBRARIES})
+  target_include_directories(QuEST PUBLIC "/usr/local/cuda/include")
 else()
   target_link_libraries(QuEST ${CUDA_LIBRARIES})
 endif()

QuEST/include/QuEST.h

@@ -34,6 +34,23 @@
 
 # include "QuEST_precision.h"
 
+
+
+// ensure custatevecHandle_t is defined, even if no GPU
+# ifdef USE_CUQUANTUM
+# include <custatevec.h>
+typedef struct CuQuantumConfig {
+    cudaMemPool_t cuMemPool;
+    cudaStream_t cuStream;
+    custatevecHandle_t cuQuantumHandle;
+    custatevecDeviceMemHandler_t cuMemHandler;
+} CuQuantumConfig;
+# else
+# define CuQuantumConfig void*
+# endif
+
+
+
 // prevent C++ name mangling
 #ifdef __cplusplus
 extern "C" {
@@ -368,6 +385,11 @@ typedef struct Qureg
     //! Storage for reduction of probabilities on GPU
     qreal *firstLevelReduction, *secondLevelReduction;
 
+    //! Storage for wavefunction amplitues and config (copy of QuESTEnv's handle) in cuQuantum deployment
+    cuAmp* cuStateVec;
+    cuAmp* deviceCuStateVec;
+    CuQuantumConfig* cuConfig;
+
     //! Storage for generated QASM output
     QASMLogger* qasmLog;
 
@@ -386,6 +408,10 @@ typedef struct QuESTEnv
     int numRanks;
     unsigned long int* seeds;
     int numSeeds;
+
+    // a copy of the QuESTEnv's config, used only in cuQuantum deployment
+    CuQuantumConfig* cuConfig;
+
 } QuESTEnv;
 
 
@@ -4236,6 +4262,10 @@ qreal calcPurity(Qureg qureg);
  * linear algebra calculation.
  *
  * The number of qubits represented in \p qureg and \p pureState must match.
+ *
+ * > In the GPU-accelerated cuQuantum backend, this function further assumes that
+ * > the density matrix \p qureg is correctly normalised, and otherwise returns the 
+ * > fidelity of the conjugate-transpose of \p qureg.
  * 
  * @see
  * - calcHilbertSchmidtDistance()

QuEST/include/QuEST_precision.h

@@ -10,15 +10,27 @@
  * @author Tyson Jones (doc)
  */
 
-# include <math.h>
-
 # ifndef QUEST_PRECISION_H
 # define QUEST_PRECISION_H
 
+# include <math.h>
+
+
+// define CUDA complex types as void if not using cuQuantum.
+// note we used cuComplex.h for complex numbers, in lieu of
+// Thrust's complex<qreal>, so that the QuEST.h header can
+// always be compiled with C99, rather than C++14.
+# ifdef USE_CUQUANTUM
+    # include <cuComplex.h>
+# else
+    # define cuFloatComplex void
+    # define cuDoubleComplex void
+# endif
+
 
 // set default double precision if not set during compilation
 # ifndef QuEST_PREC
-# define QuEST_PREC 2
+    # define QuEST_PREC 2
 # endif
 
 
@@ -28,6 +40,7 @@
 # if QuEST_PREC==1
     # define qreal float
     // \cond HIDDEN_SYMBOLS   
+    # define cuAmp cuFloatComplex
     # define MPI_QuEST_REAL MPI_FLOAT
     # define MPI_MAX_AMPS_IN_MSG (1LL<<29) // must be 2^int
     # define REAL_STRING_FORMAT "%.8f"
@@ -41,7 +54,8 @@
  */
 # elif QuEST_PREC==2
     # define qreal double
-    // \cond HIDDEN_SYMBOLS   
+    // \cond HIDDEN_SYMBOLS
+    # define cuAmp cuDoubleComplex
     # define MPI_QuEST_REAL MPI_DOUBLE
     # define MPI_MAX_AMPS_IN_MSG (1LL<<28) // must be 2^int
     # define REAL_STRING_FORMAT "%.14f"
@@ -57,6 +71,7 @@
 # elif QuEST_PREC==4
     # define qreal long double
     // \cond HIDDEN_SYMBOLS   
+    # define cuAmp void // invalid
     # define MPI_QuEST_REAL MPI_LONG_DOUBLE
     # define MPI_MAX_AMPS_IN_MSG (1LL<<27) // must be 2^int
     # define REAL_STRING_FORMAT "%.17Lf"

QuEST/src/GPU/CMakeLists.txt

@@ -1,6 +1,11 @@
+if (USE_CUQUANTUM)
+    set(GPU_CORE QuEST_cuQuantum.cu)
+else ()
+    set(GPU_CORE QuEST_gpu.cu)
+endif()
+
 set(QuEST_SRC_ARCHITECTURE_DEPENDENT
-    ${CMAKE_CURRENT_SOURCE_DIR}/QuEST_gpu.cu
+    ${CMAKE_CURRENT_SOURCE_DIR}/${GPU_CORE}
+    ${CMAKE_CURRENT_SOURCE_DIR}/QuEST_gpu_common.cu
     PARENT_SCOPE
 )
-
-