matrices.cpp

/** @file
 * Definitions of API matrix data structures, and their getters
 * and setters, as well as reporting utilities.
 * 
 * This file defines several layers of initialisation of complex
 * matrices, as explained in the header file.
 * 
 * @author Tyson Jones
 */
 
#include "quest/include/matrices.h"
#include "quest/include/modes.h"
#include "quest/include/types.h"
#include "quest/include/environment.h"
 
#include "quest/src/core/validation.hpp"
#include "quest/src/core/autodeployer.hpp"
#include "quest/src/core/utilities.hpp"
#include "quest/src/core/localiser.hpp"
#include "quest/src/core/paulilogic.hpp"
#include "quest/src/core/printer.hpp"
#include "quest/src/core/bitwise.hpp"
#include "quest/src/core/fastmath.hpp"
#include "quest/src/core/memory.hpp"
#include "quest/src/comm/comm_config.hpp"
#include "quest/src/comm/comm_routines.hpp"
#include "quest/src/cpu/cpu_config.hpp"
#include "quest/src/gpu/gpu_config.hpp"
#include "quest/src/cpu/cpu_subroutines.hpp"
 
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
 
using std::vector;
 
 
 
/*
 * FIZED-SIZE MATRIX VECTOR GETTERS
 *
 * enabling getCompMatr1 (etc) to receive vectors, in addition
 * to the existing pointer and array overloads in the header.
 * These are necessarily defined here in the source, unlike the
 * other header functions, so that they may use private valdiation.
 */
 
 
CompMatr1 getCompMatr1(vector<vector<qcomp>> in) {
    validate_matrixNumNewElems(1, in, __func__);
 
    qcomp* rowPtrs[] = {in[0].data(), in[1].data()};
    return getCompMatr1(rowPtrs);
}
 
CompMatr2 getCompMatr2(vector<vector<qcomp>> in) {
    validate_matrixNumNewElems(2, in, __func__);
 
    qcomp* rowPtrs[] = {in[0].data(), in[1].data(), in[2].data(), in[3].data()};
    return getCompMatr2(rowPtrs);
}
 
 
DiagMatr1 getDiagMatr1(vector<qcomp> in) {
    validate_matrixNumNewElems(1, in, __func__);
 
    return getDiagMatr1(in.data());
}
 
DiagMatr2 getDiagMatr2(vector<qcomp> in) {
    validate_matrixNumNewElems(2, in, __func__);
 
    return getDiagMatr2(in.data());
}
 
 
 
/*
 * PRIVATE VARIABLE-SIZE MATRIX UTILITIES
 */
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void freeHeapMatrix(T matr) {
 
    // WARNING: this is not overwriting any freed pointers with null, 
    // since the caller's struct is not changed. This is fine here,
    // but would be an issue if the struct contained nested pointers,
    // since caller would not know an outer pointer was freed and
    // ergo that it should not be enumerated (to check/free inner ptr)
 
    // free the 1D or 2D matrix - safe even if nullptr
    if constexpr (util_isDenseMatrixType<T>()) {
        cpu_deallocMatrixWrapper(matr.cpuElems);
        cpu_deallocArray(matr.cpuElemsFlat);
    } else
        cpu_deallocArray(matr.cpuElems);
 
    // we avoid invoking a GPU function in non-GPU mode
    auto gpuPtr = util_getGpuMemPtr(matr);
    if (mem_isAllocated(gpuPtr))
        gpu_deallocArray(gpuPtr);
 
    // free the teeny tiny heap flags
    util_deallocEpsilonSensitiveHeapFlag(matr.isApproxUnitary);
    util_deallocEpsilonSensitiveHeapFlag(matr.isApproxHermitian);
    cpu_deallocHeapFlag(matr.wasGpuSynced);
 
    // only diagonal matrices (which can be raised to
    // exponents) need their negativity/zeroness checked
    if constexpr (!util_isDenseMatrixType<T>()) {
        util_deallocEpsilonSensitiveHeapFlag(matr.isApproxNonZero);
        cpu_deallocHeapFlag(matr.isStrictlyNonNegative);
    }
}
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
bool didAnyLocalAllocsFail(T matr) {
 
    // god help us if these single-integer malloc failed
    if (!mem_isAllocated(matr.isApproxUnitary))   return true;
    if (!mem_isAllocated(matr.isApproxHermitian)) return true;
    if (!mem_isAllocated(matr.wasGpuSynced))      return true;
 
    // only diagonal matrices (which can be raised to
    // exponents) have these addtional fields
    if constexpr (!util_isDenseMatrixType<T>()) {
        if (!mem_isAllocated(matr.isApproxNonZero))       return true;
        if (!mem_isAllocated(matr.isStrictlyNonNegative)) return true;
    }
 
    // outer CPU memory should always be allocated
    if constexpr (util_isDenseMatrixType<T>()) {
        if (!mem_isAllocated(matr.cpuElemsFlat))  return true;
        if (!mem_isOuterAllocated(matr.cpuElems)) return true;
    } else
        if (!mem_isAllocated(matr.cpuElems)) return true;
 
    // if GPU memory is not allocated in a GPU environment...
    bool isGpuAlloc = mem_isAllocated(util_getGpuMemPtr(matr));
    if (getQuESTEnv().isGpuAccelerated && !isGpuAlloc) {
 
        // then FullStateDiagMatr GPU alloc failed only if it tried...
        if constexpr (util_isFullStateDiagMatr<T>()) {
            if (matr.isGpuAccelerated)
                return true;
            
        // but all other matrices always try to alloc, so must have failed
        } else
            return true;
    }
 
    // otherwise, all pointers were non-NULL and ergo all allocs were successful
    return false;
}
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void freeAllMemoryIfAnyAllocsFailed(T matr) {
 
    // ascertain whether any allocs failed on any node
    bool anyFail = didAnyLocalAllocsFail(matr);
    if (comm_isInit())
        anyFail = comm_isTrueOnAllNodes(anyFail);
 
    // if so, free all heap fields
    if (anyFail)
        freeHeapMatrix(matr);
}
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void validateMatrixAllocs(T matr, const char* caller) {
 
    // free memory before throwing validation error to avoid memory leaks
    freeAllMemoryIfAnyAllocsFailed(matr);
    validate_newMatrixAllocs(matr, caller);
}
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void setInitialHeapFlags(T matr) {
 
    // set initial propreties of the newly created matrix to unknown
    util_setFlagToUnknown(matr.isApproxUnitary);
    util_setFlagToUnknown(matr.isApproxHermitian);
 
    // only diagonal matrices (which can be exponentiated)
    // have these additional fields
    if constexpr (!util_isDenseMatrixType<T>()) {
        util_setFlagToUnknown(matr.isApproxNonZero);
        util_setFlagToUnknown(matr.isStrictlyNonNegative);
    }
 
    // indicate that GPU memory has not yet been synchronised
    *(matr.wasGpuSynced) = 0;
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX CONSTRUCTORS
 */
 
 
extern "C" CompMatr createCompMatr(int numQubits) {
    validate_envIsInit(__func__);
    validate_newCompMatrParams(numQubits, __func__);
 
    // validation ensures these never overflow
    qindex numRows = powerOf2(numQubits);
    qindex numElems = numRows * numRows;
 
    // attempt to allocate 1D memory
    qcomp* cpuMem = cpu_allocArray(numElems); // nullptr if failed
    qcomp* gpuMem = nullptr;
    if (getQuESTEnv().isGpuAccelerated)
        gpuMem = gpu_allocArray(numElems); // nullptr if failed
 
    // prepare output CompMatr (avoiding C++20 designated initialiser)
    CompMatr out;
    out.numQubits = numQubits;
    out.numRows = numRows;
 
    // attemptedly allocate (un-initialised) flags in the heap so that struct copies are mutable
    out.isApproxUnitary    = util_allocEpsilonSensitiveHeapFlag(); // nullptr if failed
    out.isApproxHermitian  = util_allocEpsilonSensitiveHeapFlag();
    out.wasGpuSynced = cpu_allocHeapFlag(); // nullptr if failed
 
    // attemptedly allocate 2D alias for 1D CPU memory
    out.cpuElems = cpu_allocAndInitMatrixWrapper(cpuMem, numRows); // nullptr if failed
    out.cpuElemsFlat = cpuMem;
    out.gpuElemsFlat = gpuMem;
 
    validateMatrixAllocs(out, __func__);
    setInitialHeapFlags(out);
    return out;
}
 
 
extern "C" DiagMatr createDiagMatr(int numQubits) {
    validate_envIsInit(__func__);
    validate_newDiagMatrParams(numQubits, __func__);
 
    // validation ensures this never overflows
    qindex numElems = powerOf2(numQubits);
 
    // prepare output DiagMatr (avoiding C++20 designated initialiser)
    DiagMatr out;
    out.numQubits = numQubits,
    out.numElems = numElems,
 
    // attempt to allocate (uninitialised) flags in the heap so that struct copies are mutable
    out.isApproxUnitary       = util_allocEpsilonSensitiveHeapFlag(); // nullptr if failed
    out.isApproxHermitian     = util_allocEpsilonSensitiveHeapFlag();
    out.isApproxNonZero       = util_allocEpsilonSensitiveHeapFlag();
    out.isStrictlyNonNegative = cpu_allocHeapFlag(); // nullptr if failed
    out.wasGpuSynced          = cpu_allocHeapFlag();
 
    // attempt to allocate 1D memory (nullptr if failed or not allocated)
    out.cpuElems = cpu_allocArray(numElems);
    out.gpuElems = (getQuESTEnv().isGpuAccelerated)? gpu_allocArray(numElems) : nullptr;
 
    validateMatrixAllocs(out, __func__);
    setInitialHeapFlags(out);
    return out;
}
 
 
FullStateDiagMatr validateAndCreateCustomFullStateDiagMatr(int numQubits, int useDistrib, int useGpuAccel, int useMultithread, const char* caller) {
    validate_envIsInit(caller);
    QuESTEnv env = getQuESTEnv();
 
    // validate parameters before passing them to autodeployer
    validate_newFullStateDiagMatrParams(numQubits, useDistrib, useGpuAccel, useMultithread, caller);
 
    // overwrite all args left as AUTO_FLAG
    autodep_chooseFullStateDiagMatrDeployment(numQubits, useDistrib, useGpuAccel, useMultithread, env);
 
    // validation ensures this never overflows
    qindex numElems = powerOf2(numQubits);
    qindex numElemsPerNode = numElems / (useDistrib? env.numNodes : 1); // divides evenly
 
    // prepare output FullStateDiagMatr (avoiding C++20 designated initialiser)
    FullStateDiagMatr out;
    out.numQubits = numQubits;
    out.numElems = numElems;
 
    // bind deployments, disabling distribution if using a single MPI node
    out.isGpuAccelerated = useGpuAccel;
    out.isMultithreaded = useMultithread;
    out.isDistributed = useDistrib && (env.numNodes > 1);
    out.numElemsPerNode = numElemsPerNode;
 
    // allocate (unitialised) flags in the heap so that struct copies are mutable
    out.isApproxUnitary       = util_allocEpsilonSensitiveHeapFlag(); // nullptr if failed
    out.isApproxHermitian     = util_allocEpsilonSensitiveHeapFlag();
    out.isApproxNonZero       = util_allocEpsilonSensitiveHeapFlag();
    out.isStrictlyNonNegative = cpu_allocHeapFlag(); // nullptr if failed
    out.wasGpuSynced          = cpu_allocHeapFlag();
 
    // allocate 1D memory (nullptr if failed or not allocated)
    out.cpuElems = cpu_allocArray(numElemsPerNode);
    out.gpuElems = (useGpuAccel)? gpu_allocArray(numElemsPerNode) : nullptr;
 
    validateMatrixAllocs(out, __func__);
    setInitialHeapFlags(out);
    return out;
}
 
extern "C" FullStateDiagMatr createCustomFullStateDiagMatr(int numQubits, int useDistrib, int useGpuAccel, int useMultithread) {
 
    return validateAndCreateCustomFullStateDiagMatr(numQubits, useDistrib, useGpuAccel, useMultithread, __func__);
}
 
extern "C" FullStateDiagMatr createFullStateDiagMatr(int numQubits) {
 
    return validateAndCreateCustomFullStateDiagMatr(numQubits, modeflag::USE_AUTO, modeflag::USE_AUTO, modeflag::USE_AUTO, __func__);
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX SYNC
 */
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void markMatrixAsSynced(T matr) {
 
    // indicate that the matrix is now permanently GPU synchronised, even
    // if we are not in GPU-accelerated mode (in which case it's never consulted)
    *(matr.wasGpuSynced) = 1;
 
    // indicate that we do not know the revised matrix properties;
    // we defer establishing that until validation needs to check them
    util_setFlagToUnknown(matr.isApproxUnitary);
    util_setFlagToUnknown(matr.isApproxHermitian);
 
    // only diagonal matrices (which can be exponentiated)
    // have these additional fields
    if constexpr (!util_isDenseMatrixType<T>()) {
        util_setFlagToUnknown(matr.isApproxNonZero);
        util_setFlagToUnknown(matr.isStrictlyNonNegative);
    }
}
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void validateAndSyncMatrix(T matr, const char* caller) {
    validate_matrixFields(matr, caller);
 
    // optionally overwrite GPU elements with user-modified CPU elements
    if (mem_isAllocated(util_getGpuMemPtr(matr)))
        gpu_copyCpuToGpu(matr);
 
    markMatrixAsSynced(matr);
}
 
 
// de-mangled for both C++ and C compatibility
extern "C" {
 
    void syncCompMatr(CompMatr matr) { validateAndSyncMatrix(matr, __func__); }
    void syncDiagMatr(DiagMatr matr) { validateAndSyncMatrix(matr, __func__); }
    void syncFullStateDiagMatr(FullStateDiagMatr matr) { validateAndSyncMatrix(matr, __func__); }
 
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX DESTRUCTION
 */
 
 
// type T can be CompMatr, DiagMatr or FullStateDiagMatr
template <class T>
void validateAndDestroyMatrix(T matrix, const char* caller) {
    validate_matrixFields(matrix, caller);
 
    freeHeapMatrix(matrix);
}
 
 
// de-mangled for C++ and C compatibility
extern "C" {
    
    void destroyCompMatr(CompMatr matr) { validateAndDestroyMatrix(matr, __func__); }
    void destroyDiagMatr(DiagMatr matr) { validateAndDestroyMatrix(matr, __func__); }
    void destroyFullStateDiagMatr(FullStateDiagMatr matr) { validateAndDestroyMatrix(matr, __func__); }
 
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX SETTERS VIA POINTERS
 */
 
 
// type T can be qcomp** or vector<vector<qcomp>>, but qcomp(*)[] is handled by header
template <typename T> 
void setAndSyncDenseMatrElems(CompMatr out, T elems) {
    
    // copy elems into matrix's CPU memory
    cpu_copyMatrix(out.cpuElems, elems, out.numRows);
 
    // overwrite GPU elements; validation gauranteed to pass
    syncCompMatr(out); 
}
 
 
extern "C" void setCompMatr(CompMatr out, qcomp** in) {
    validate_matrixFields(out, __func__);
    validate_matrixNewElemsPtrNotNull(in, out.numRows, __func__);
 
    setAndSyncDenseMatrElems(out, in);
}
 
 
extern "C" void setDiagMatr(DiagMatr out, qcomp* in) {
    validate_matrixFields(out, __func__);
    validate_matrixNewElemsPtrNotNull(in, __func__);
 
    // overwrite CPU memory
    cpu_copyArray(out.cpuElems, in, out.numElems);
 
    // overwrite GPU elements; validation gauranteed to pass
    syncDiagMatr(out);
}
 
 
extern "C" void setFullStateDiagMatr(FullStateDiagMatr out, qindex startInd, qcomp* in, qindex numElems) {
    validate_matrixFields(out, __func__);
    validate_fullStateDiagMatrNewElems(out, startInd, numElems, __func__);
    validate_matrixNewElemsPtrNotNull(in, __func__);
 
    // overwrites both the CPU and GPU memory (if it exists), maintaining consistency.
    // note that cpu_copyArray() isn't called directly here like setDiagMatr() above
    // because we must handle when 'out' is and isn't distributed
    localiser_fullstatediagmatr_setElems(out, startInd, in, numElems);
 
    // even though we have not necessarily overwritten every element, we must mark
    // the matrix as synced so that it can be subsequently used without error
    markMatrixAsSynced(out);
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX SETTERS VIA VECTORS
 */
 
 
void setCompMatr(CompMatr out, vector<vector<qcomp>> in) {
    validate_matrixFields(out, __func__);
    validate_matrixNumNewElems(out.numQubits, in, __func__);
 
    setAndSyncDenseMatrElems(out, in);
}
 
 
void setDiagMatr(DiagMatr out, vector<qcomp> in) {
    validate_matrixFields(out, __func__);
    validate_matrixNumNewElems(out.numQubits, in, __func__); // validates 'in' dim
 
    setDiagMatr(out, in.data()); // harmessly re-validates
}
 
 
void setFullStateDiagMatr(FullStateDiagMatr out, qindex startInd, vector<qcomp> in) {
 
    setFullStateDiagMatr(out, startInd, in.data(), in.size());
}
 
 
// no bespoke array functions are necessary for diagonal matrices initialisation, 
// since passed arrays automatically decay to pointers
 
 
 
/*
 * VARIABLE-SIZE MATRIX SETTERS VIA LITERALS
 *
 * Only the C++ versions are defined here, while the C versions are macros
 * defined in the header. Note the C++ versions themselves are entirely
 * superfluous and merely call the above vector setters, but we still define
 * them for API consistency, and we additionally validate the superfluous
 * additional parameters they pass.
 */
 
 
void setInlineCompMatr(CompMatr matr, int numQb, vector<vector<qcomp>> in) {
    validate_matrixFields(matr, __func__);
    validate_matrixNumQubitsMatchesParam(matr.numQubits, numQb, __func__);
    validate_matrixNumNewElems(matr.numQubits, in, __func__);
 
    setAndSyncDenseMatrElems(matr, in);
}
 
void setInlineDiagMatr(DiagMatr matr, int numQb, vector<qcomp> in) {
    validate_matrixFields(matr, __func__);
    validate_matrixNumQubitsMatchesParam(matr.numQubits, numQb, __func__);
    validate_matrixNumNewElems(matr.numQubits, in, __func__);
 
    setDiagMatr(matr, in.data()); // validation gauranteed to pass
}
 
void setInlineFullStateDiagMatr(FullStateDiagMatr matr, qindex startInd, qindex numElems, vector<qcomp> in) {
    validate_matrixFields(matr, __func__);
    validate_declaredNumElemsMatchesVectorLength(numElems, in.size(), __func__);
    validate_fullStateDiagMatrNewElems(matr, startInd, numElems, __func__);
 
    setFullStateDiagMatr(matr, startInd, in); // validation gauranteed to pass
}
 
 
 
/*
 * VARIABLE-SIZE MATRIX INLINE-SETTER CONSTRUCTORS 
 *
 * Only the C++ versions are defined here; the C versions are header macros
 */
 
 
CompMatr createInlineCompMatr(int numQb, vector<vector<qcomp>> elems) {
    validate_envIsInit(__func__);
    validate_newCompMatrParams(numQb, __func__);
    validate_matrixNumNewElems(numQb, elems, __func__);
 
    // pre-validation gauranteed to pass, but malloc failures will trigger an error 
    // message specific to 'createCompMatr', rather than this 'inline' version. Alas!
    CompMatr matr = createCompMatr(numQb);
    setAndSyncDenseMatrElems(matr, elems);
    return matr;
}
 
 
DiagMatr createInlineDiagMatr(int numQb, vector<qcomp> elems) {
    validate_envIsInit(__func__);
    validate_newDiagMatrParams(numQb, __func__);
    validate_matrixNumNewElems(numQb, elems, __func__);
 
    // pre-validation gauranteed to pass, but malloc failures will trigger an error 
    // message specific to 'createCompMatr', rather than this 'inline' version. Alas!
    DiagMatr matr = createDiagMatr(numQb);
    setDiagMatr(matr, elems.data()); // validation gauranteed to pass
    return matr;
}
 
 
 
/*
 * EXPOSING SOME SETTER VALIDATION TO HEADER
 *
 * Some setters are necessarily defined in the header, because they accept 
 * C-only VLAs which need to be cast into pointers before being passed to 
 * this C++ backend (which does not support VLA). These setters need their
 * validators exposed, though we cannot expose the entirety of validation.hpp 
 * because it cannot be parsed by C; so we here wrap specific functions.
 */
 
 
extern "C" {
 
    void _validateNewNestedElemsPtrNotNull(qcomp** ptrs, int numQubits, const char* caller) {
 
        validate_matrixNewElemsPtrNotNull(ptrs, powerOf2(numQubits), caller);
    }
    void _validateNewElemsPtrNotNull(qcomp* ptr, const char* caller) {
        
        validate_matrixNewElemsPtrNotNull(ptr, caller);
    }
 
    void _validateParamsToSetCompMatrFromArr(CompMatr matr) { 
 
        validate_matrixFields(matr, "setCompMatr");
    }
 
    void _validateParamsToSetInlineCompMatr(CompMatr matr, int numQb) {
 
        const char* caller = "setInlineCompMatr";
        validate_matrixFields(matr, caller);
        validate_matrixNumQubitsMatchesParam(matr.numQubits, numQb, caller);
    }
 
    void _validateParamsToSetInlineDiagMatr(DiagMatr matr, int numQb) {
 
        const char* caller = "setInlineDiagMatr";
        validate_matrixFields(matr, caller);
        validate_matrixNumQubitsMatchesParam(matr.numQubits, numQb, caller);
    }
 
    void _validateParamsToSetInlineFullStateDiagMatr(FullStateDiagMatr matr, qindex startInd, qindex numElems) {
 
        const char* caller = "setInlineFullStateDiagMatr";
        validate_matrixFields(matr, caller);
        validate_fullStateDiagMatrNewElems(matr, startInd, numElems, caller);
    }
 
    void _validateParamsToCreateInlineCompMatr(int numQb) {
 
        const char* caller = "createInlineCompMatr";
        validate_envIsInit(caller);
        validate_newCompMatrParams(numQb, caller);
    }
 
    void _validateParamsToCreateInlineDiagMatr(int numQb) {
 
        const char* caller = "createInlineDiagMatr";
        validate_envIsInit(caller);
        validate_newDiagMatrParams(numQb, caller);
    }
 
}
 
 
 
/*
 * SPECIAL CREATORS AND SETTERS
 */
 
 
extern "C" void setFullStateDiagMatrFromPauliStrSum(FullStateDiagMatr out, PauliStrSum in) {
    validate_matrixFields(out, __func__);
    validate_pauliStrSumFields(in, __func__);
    validate_pauliStrSumCanInitMatrix(out, in, __func__);
 
    // permit 'in' to be non-Hermitian since it does not determine 'out' unitarity
 
    // unlike other FullStateDiagMatr initialisers, we employ an accelerated
    // backend since the input data 'in' is expectedly significantly smaller
    // than the created data in 'out', making parallelisation worthwhile as
    // the memory-movement costs of copying 'in' to a GPU are small
    localiser_fullstatediagmatr_setElemsToPauliStrSum(out, in);
 
    markMatrixAsSynced(out);
}
 
 
extern "C" FullStateDiagMatr createFullStateDiagMatrFromPauliStrSum(PauliStrSum in) {
    validate_pauliStrSumFields(in, __func__);
    
    // ensure createFullStateDiagMatr() below succeeds (so if not, that thrower name is correct)
    int numQubits = 1 + paulis_getIndOfLefmostNonIdentityPauli(in);
    validate_newFullStateDiagMatrParams(numQubits, modeflag::USE_AUTO, modeflag::USE_AUTO, modeflag::USE_AUTO, __func__);
 
    // permit 'in' to be non-Hermitian since it does not determine 'out' unitarity
 
    FullStateDiagMatr out = createFullStateDiagMatr(numQubits);
    localiser_fullstatediagmatr_setElemsToPauliStrSum(out, in);
    markMatrixAsSynced(out);
    return out;
}
 
 
extern "C" void setDiagMatrFromMultiVarFunc(DiagMatr out, qcomp (*callbackFunc)(qindex*), int* numQubitsPerVar, int numVars, int areSigned) {
    validate_matrixFields(out, __func__);
    validate_multiVarFuncQubits(out.numQubits, numQubitsPerVar, numVars, __func__);
    validate_funcVarSignedFlag(areSigned, __func__);
 
    vector<qindex> varValues(numVars);
 
    // set each element of the diagonal in-turn; user's callback might not be thread-safe
    for (qindex elemInd=0; elemInd<out.numElems; elemInd++) {
        fast_getSubQuregValues(elemInd, numQubitsPerVar, numVars, areSigned, varValues.data());
 
        // call user function, and update only the CPU elems
        out.cpuElems[elemInd] = callbackFunc(varValues.data());
    }
 
    // overwrite all GPU elems
    syncDiagMatr(out);
}
 
 
extern "C" void setFullStateDiagMatrFromMultiVarFunc(FullStateDiagMatr out, qcomp (*callbackFunc)(qindex*), int* numQubitsPerVar, int numVars, int areSigned) {
    validate_matrixFields(out, __func__);
    validate_multiVarFuncQubits(out.numQubits, numQubitsPerVar, numVars, __func__);
    validate_funcVarSignedFlag(areSigned, __func__);
 
    // we assume callbackFunc is thread-safe (!!!!) and possibly use multithreading, but never 
    // GPU acceleration, since we cannot invoke user callback functions from GPU kernels
    cpu_fullstatediagmatr_setElemsFromMultiVarFunc(out, callbackFunc, numQubitsPerVar, numVars, areSigned);
 
    // overwrite all GPU elems
    syncFullStateDiagMatr(out);
}
 
 
extern "C" void setDiagMatrFromMultiDimLists(DiagMatr out, void* lists, int* numQubitsPerDim, int numDims) {
    validate_matrixFields(out, __func__);
    validate_multiVarFuncQubits(out.numQubits, numQubitsPerDim, numDims, __func__);
 
    vector<qindex> listInds(numDims);
 
    // set each element of the diagonal in-turn, which is embarrassingly parallel, 
    // although we do not parallelise - the DiagMatr is intendedly small
    for (qindex elemInd=0; elemInd<out.numElems; elemInd++) {
 
        // nested list indices = unsigned integer values of variables
        fast_getSubQuregValues(elemInd, numQubitsPerDim, numDims, false, listInds.data());
 
        // update only the CPU elems
        out.cpuElems[elemInd] = util_getElemFromNestedPtrs(lists, listInds.data(), numDims);
    }
 
    // overwrite all GPU elems
    syncDiagMatr(out);
}
 
 
extern "C" void setFullStateDiagMatrFromMultiDimLists(FullStateDiagMatr out, void* lists, int* numQubitsPerDim, int numDims) {
    validate_matrixFields(out, __func__);
    validate_multiVarFuncQubits(out.numQubits, numQubitsPerDim, numDims, __func__);
 
    // possibly use multithreading, but never GPU acceleration, due to the
    // arbitrarily nested nature of the input lists
    cpu_fullstatediagmatr_setElemsFromMultiDimLists(out, lists, numQubitsPerDim, numDims);
 
    // overwrite all GPU elems
    syncFullStateDiagMatr(out);
}
 
 
 
/*
 * MATRIX REPORTERS
 */
 
 
// type T can be CompMatr1, CompMatr2, CompMatr, DiagMatr1, DiagMatr2, DiagMatr, FullStateDiagMatr
template<class T> 
void validateAndPrintMatrix(T matr, const char* caller) {
    validate_matrixFields(matr, caller);
    validate_numReportedNewlinesAboveZero(__func__); // because trailing newline mandatory
 
    // syncable matrices must be synced before reporting (though only CPU elems are printed)
    if constexpr (util_isHeapMatrixType<T>())
        validate_matrixIsSynced(matr, caller);
 
    // calculate the total memory (in bytes) consumed by the matrix on each
    // node, which will depend on whether the matrix is distributed, and
    // includes the size of the matrix struct itself. Note that GPU memory
    // is not included since it is less than or equal to the CPU memory, and
    // occupies different memory spaces, confusing capacity accounting
    int numNodes = (util_isDistributedMatrix(matr))? comm_getNumNodes() : 1;
    size_t elemMem = mem_getLocalMatrixMemoryRequired(matr.numQubits, util_isDenseMatrixType<T>(), numNodes);
    size_t structMem = sizeof(matr);
 
    // struct memory includes fixed-size arrays (qcomp[][]), so we undo double-counting
    if (util_isFixedSizeMatrixType<T>())
        structMem -= elemMem;
 
    size_t numBytesPerNode = elemMem + structMem;
    print_header(matr, numBytesPerNode);
    print_elems(matr);
 
    // exclude mandatory newline above
    print_oneFewerNewlines();
}
 
 
// all reporters are C and C++ accessible, so are de-mangled
extern "C" {
    
    void reportCompMatr1(CompMatr1 matr) { validateAndPrintMatrix(matr, __func__); }
    void reportCompMatr2(CompMatr2 matr) { validateAndPrintMatrix(matr, __func__); }
    void reportCompMatr (CompMatr  matr) { validateAndPrintMatrix(matr, __func__); }
    void reportDiagMatr1(DiagMatr1 matr) { validateAndPrintMatrix(matr, __func__); }
    void reportDiagMatr2(DiagMatr2 matr) { validateAndPrintMatrix(matr, __func__); }
    void reportDiagMatr (DiagMatr  matr) { validateAndPrintMatrix(matr, __func__); }
    void reportFullStateDiagMatr(FullStateDiagMatr matr) { validateAndPrintMatrix(matr, __func__); }
 
}