Miscellaneous bug patches

quest/src/api/calculations.cpp

@@ -428,6 +428,11 @@ qreal calcProbOfMultiQubitOutcome(Qureg qureg, vector<int> qubits, vector<int> o
 
 vector<qreal> calcProbsOfAllMultiQubitOutcomes(Qureg qureg, vector<int> qubits) {
 
+    // validate before exponentiating the caller-controlled vector size
+    validate_quregFields(qureg, __func__);
+    validate_targets(qureg, qubits.data(), qubits.size(), __func__);
+    validate_measurementOutcomesFitInGpuMem(qureg, qubits.size(), __func__);
+
     // allocate temp vector, and validate successful (since it's exponentially large!)
     vector<qreal> out;
     qindex numOut = powerOf2(qubits.size());

quest/src/api/channels.cpp

@@ -108,7 +108,7 @@ void freeAllMemoryIfAnyAllocsFailed(T& obj) {
     // determine whether any node experienced a failure
     bool anyFail = didAnyLocalAllocsFail(obj);
     if (comm_isActive())
-        anyFail = comm_isTrueOnAllNodes(anyFail);
+        anyFail = comm_isTrueOnAnyNode(anyFail);
 
     // if so, free all memory before subsequent validation
     if (anyFail)

quest/src/api/decoherence.cpp

@@ -143,6 +143,7 @@ void mixQureg(Qureg outQureg, Qureg inQureg, qreal inProb) {
 
 void mixSuperOp(Qureg qureg, int* targets, int numTargets, SuperOp superop) {
     validate_quregFields(qureg, __func__);
+    validate_quregIsDensityMatrix(qureg, __func__);
     validate_targets(qureg, targets, numTargets, __func__);
     validate_superOpFields(superop, __func__);
     validate_superOpIsSynced(superop, __func__);

quest/src/api/environment.cpp

@@ -375,7 +375,7 @@ void printQuregAutoDeployments(bool isDensMatr) {
     // auto-deploy will still try to deploy there to (then subsequent validation will fail)
     int maxQubits = mem_getMaxNumQuregQubitsBeforeGlobalMemSizeofOverflow(isDensMatr, global_envPtr->numNodes);
 
-    for (int numQubits=1; numQubits<maxQubits; numQubits++) {
+    for (int numQubits=1; numQubits<=maxQubits; numQubits++) {
 
         // re-choose auto deployment
         useDistrib  = modeflag::USE_AUTO;

quest/src/api/initialisations.cpp

@@ -293,7 +293,8 @@ void setDensityQuregAmps(Qureg qureg, qindex startRow, qindex startCol, vector<v
         ptrs[i] = amps[i].data();
 
     // C function performs main validation
-    setDensityQuregAmps(qureg, startRow, startCol, ptrs.data(), len, (len>0)? len : 0); // avoid seg-fault
+    size_t numCols = (len > 0)? amps[0].size() : 0; // avoid seg-fault
+    setDensityQuregAmps(qureg, startRow, startCol, ptrs.data(), len, numCols);
 }
 
 void setQuregAmps(Qureg qureg, qindex startInd, vector<qcomp> amps) {

quest/src/api/matrices.cpp

@@ -166,7 +166,7 @@ void freeAllMemoryIfAnyAllocsFailed(T matr) {
     // ascertain whether any allocs failed on any node
     bool anyFail = didAnyLocalAllocsFail(matr);
     if (comm_isActive())
-        anyFail = comm_isTrueOnAllNodes(anyFail);
+        anyFail = comm_isTrueOnAnyNode(anyFail);
 
     // if so, free all heap fields
     if (anyFail)
@@ -310,7 +310,7 @@ FullStateDiagMatr validateAndCreateCustomFullStateDiagMatr(int numQubits, int us
     out.cpuElems = cpu_allocArray(numElemsPerNode);
     out.gpuElems = (useGpuAccel)? gpu_allocArray(numElemsPerNode) : nullptr;
 
-    validateMatrixAllocs(out, __func__);
+    validateMatrixAllocs(out, caller);
     setInitialHeapFlags(out);
     return out;
 }
@@ -743,7 +743,7 @@ extern "C" void setFullStateDiagMatrFromMultiDimLists(FullStateDiagMatr out, voi
 template<class T> 
 void validateAndPrintMatrix(T matr, const char* caller) {
     validate_matrixFields(matr, caller);
-    validate_numReportedNewlinesAboveZero(__func__); // because trailing newline mandatory
+    validate_numReportedNewlinesAboveZero(caller); // because trailing newline mandatory
 
     // syncable matrices must be synced before reporting (though only CPU elems are printed)
     if constexpr (util_isHeapMatrixType<T>())

quest/src/api/paulis.cpp

@@ -39,7 +39,7 @@ bool didAnyAllocsFailOnAnyNode(PauliStrSum sum) {
         ! mem_isAllocated(sum.isApproxHermitian) );
 
     if (comm_isActive())
-        anyFail = comm_isTrueOnAllNodes(anyFail);
+        anyFail = comm_isTrueOnAnyNode(anyFail);
 
     return anyFail;
 }

quest/src/api/qureg.cpp

@@ -117,7 +117,7 @@ bool didAnyAllocsFailOnAnyNode(Qureg qureg) {
 
     bool anyFail = didAnyLocalAllocsFail(qureg);
     if (comm_isActive())
-        anyFail = comm_isTrueOnAllNodes(anyFail);
+        anyFail = comm_isTrueOnAnyNode(anyFail);
 
     return anyFail;
 }
@@ -130,7 +130,7 @@ void freeAllMemoryIfAnyAllocsFailed(Qureg qureg) {
         return;
 
     // otherwise, free everything that was successfully allocated (freeing nullptr is legal)
-    cpu_deallocArray(qureg.cpuAmps);
+    cpu_deallocNumaArray(qureg.cpuAmps, qureg.numAmpsPerNode);
     cpu_deallocArray(qureg.cpuCommBuffer);
 
     // although we avoid calling GPU deallocation in non-GPU mode
@@ -166,7 +166,7 @@ Qureg validateAndCreateCustomQureg(int numQubits, int isDensMatr, int useDistrib
 
     // if any of the above mallocs failed, below validation will memory leak; so free first (but don't set to nullptr)
     freeAllMemoryIfAnyAllocsFailed(qureg);
-    validate_newQuregAllocs(qureg, __func__);
+    validate_newQuregAllocs(qureg, caller);
 
     // TODO / DEBUG:
     // TEMPORARILY FORBID HIP GPU ACCELERATION OF > 2^32 AMPS PER NODE,
@@ -534,6 +534,11 @@ extern "C" void _wrap_getDensityQuregAmp(qcomp* out, Qureg qureg, qindex row, qi
 
 vector<qcomp> getQuregAmps(Qureg qureg, qindex startInd, qindex numAmps) {
 
+    // validate before interpreting numAmps as an allocation size
+    validate_quregFields(qureg, __func__);
+    validate_quregIsStateVector(qureg, __func__);
+    validate_basisStateIndices(qureg, startInd, numAmps, __func__);
+
     // allocate the output vector, and validate successful
     vector<qcomp> out;
     auto callback = [&]() { validate_tempListAllocSucceeded(false, numAmps, sizeof(qcomp), __func__); };
@@ -547,9 +552,14 @@ vector<qcomp> getQuregAmps(Qureg qureg, qindex startInd, qindex numAmps) {
 
 vector<vector<qcomp>> getDensityQuregAmps(Qureg qureg, qindex startRow, qindex startCol, qindex numRows, qindex numCols) {
 
+    // validate before multiplying dimensions or interpreting them as allocation sizes
+    validate_quregFields(qureg, __func__);
+    validate_quregIsDensityMatrix(qureg, __func__);
+    validate_basisStateRowCols(qureg, startRow, startCol, numRows, numCols, __func__);
+
     // allocate the output matrix, and validate successful
     vector<vector<qcomp>> out;
-    qindex numElems = numRows * numCols; // never overflows (else Qureg alloc would fail)
+    qindex numElems = numRows * numCols;
     auto callback1 = [&]() { validate_tempListAllocSucceeded(false, numElems, sizeof(qcomp), __func__); };
     util_tryAllocMatrix(out, numRows, numCols, callback1);
 

quest/src/api/trotterisation.cpp

@@ -36,7 +36,7 @@ void internal_applyFirstOrderTrotterRepetition(
 ) {
     // apply each sum term as a gadget, in forward or reverse order
     for (qindex i=0; i<sum.numTerms; i++) {
-        int j = reverse? sum.numTerms - i - 1 : i;
+        qindex j = reverse? sum.numTerms - i - 1 : i;
         qindex k = (sumOrdering.empty())? j : sumOrdering[j];
         qcomp coeff = sum.coeffs[k];
         PauliStr str = sum.strings[k];