Optimize CPU simulation

unitaryfund · Nov 2, 2024 · ad12e08 · ad12e08
1 parent e9fcdd0
commit ad12e08
Show file tree

Hide file tree

Showing 5 changed files with 66 additions and 132 deletions.
diff --git a/include/common/qrack_functions.hpp b/include/common/qrack_functions.hpp
@@ -20,6 +20,9 @@
 
 #include <set>
 #include <vector>
+#if CPP_STD >= 20
+#include <bit>
+#endif
 
 #define _bi_div_mod(left, right, quotient, rmndr)                                                                      \
     if (quotient) {                                                                                                    \
@@ -83,7 +86,9 @@ namespace Qrack {
 inline bitLenInt log2Ocl(bitCapIntOcl n)
 {
 // Source: https://stackoverflow.com/questions/11376288/fast-computing-of-log2-for-64-bit-integers#answer-11376759
-#if ENABLE_INTRINSICS && defined(_WIN32) && !defined(__CYGWIN__)
+#if CPP_STD >= 20
+    return std::bit_width(n) - 1U;
+#elif ENABLE_INTRINSICS && defined(_WIN32) && !defined(__CYGWIN__)
 #if UINTPOW < 6
     return (bitLenInt)(bitsInByte * sizeof(unsigned int) - _lzcnt_u32((unsigned int)n) - 1U);
 #else
@@ -95,8 +100,6 @@ inline bitLenInt log2Ocl(bitCapIntOcl n)
 #else
     return (bitLenInt)(bitsInByte * sizeof(unsigned long long) - __builtin_clzll((unsigned long long)n) - 1U);
 #endif
-#elif CPP_STD >= 20
-    return std::bit_width(n) - 1U;
 #else
     bitLenInt pow = 0U;
     bitCapIntOcl p = n >> 1U;
@@ -108,6 +111,21 @@ inline bitLenInt log2Ocl(bitCapIntOcl n)
 #endif
 }
 
+inline bitLenInt popCountOcl(bitCapIntOcl n)
+{
+#if CPP_STD >= 20
+    return std::popcount(n);
+#elif defined(__GNUC__) || defined(__clang__)
+    return __builtin_popcount(n);
+#else
+    bitCapIntOcl popCount;
+    for (popCount = 0U; n; ++popCount) {
+        n &= n - 1U;
+    }
+    return popCount;
+#endif
+}
+
 #if (QBCAPPOW < 7) || ((QBCAPPOW < 8) && defined(__SIZEOF_INT128__))
 inline int bi_log2(const bitCapInt& n) { return log2Ocl(n); }
 #endif

diff --git a/include/statevector.hpp b/include/statevector.hpp
@@ -40,11 +40,8 @@ class StateVector : public ParallelFor {
     bitCapIntOcl capacity;
 
 public:
-    bool isReadLocked;
-
     StateVector(bitCapIntOcl cap)
         : capacity(cap)
-        , isReadLocked(true)
     {
     }
     virtual ~StateVector()

diff --git a/src/common/qengine.cu b/src/common/qengine.cu
@@ -60,7 +60,10 @@ __device__ inline qCudaReal1 qCudaDot(qCudaReal4 a, qCudaReal4 b)
 #if FPPOW > 4
 __device__ inline qCudaCmplx polar_unit(const qCudaReal1 theta) { return make_qCudaCmplx(cos(theta), sin(theta)); }
 #else
-__device__ inline qCudaCmplx polar_unit(const qCudaReal1 theta) { return make_qCudaCmplx((qCudaReal1)cos((qCudaReal1_f)theta), (qCudaReal1)sin((qCudaReal1_f)theta)); }
+__device__ inline qCudaCmplx polar_unit(const qCudaReal1 theta)
+{
+    return make_qCudaCmplx((qCudaReal1)cos((qCudaReal1_f)theta), (qCudaReal1)sin((qCudaReal1_f)theta));
+}
 #endif
 
 __device__ inline qCudaCmplx qCudaConj(qCudaCmplx a) { return make_qCudaCmplx(a.x, -a.y); }

diff --git a/src/qengine/arithmetic.cpp b/src/qengine/arithmetic.cpp
@@ -45,7 +45,6 @@ void QEngineCPU::ROL(bitLenInt shift, bitLenInt start, bitLenInt length)
     Finish();
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
-    stateVec->isReadLocked = false;
 
     par_for(0, maxQPowerOcl, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -83,7 +82,6 @@ void QEngineCPU::INC(const bitCapInt& toAdd, bitLenInt inOutStart, bitLenInt len
     Finish();
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
-    stateVec->isReadLocked = false;
 
     par_for(0, maxQPowerOcl, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -138,7 +136,6 @@ void QEngineCPU::CINC(
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->copy(stateVec);
-    stateVec->isReadLocked = false;
 
     par_for_mask(0, maxQPowerOcl, controlPowers, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -182,7 +179,6 @@ void QEngineCPU::INCDECC(const bitCapInt& toMod, bitLenInt inOutStart, bitLenInt
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, pow2Ocl(carryIndex), 1U, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -235,7 +231,6 @@ void QEngineCPU::INCS(const bitCapInt& toAdd, bitLenInt inOutStart, bitLenInt le
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for(0, maxQPowerOcl, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -286,7 +281,6 @@ void QEngineCPU::INCDECSC(const bitCapInt& toMod, bitLenInt inOutStart, bitLenIn
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, carryMask, 1U, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -344,7 +338,6 @@ void QEngineCPU::INCDECSC(
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, carryMask, 1U, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -388,7 +381,6 @@ void QEngineCPU::MULDIV(const IOFn& inFn, const IOFn& outFn, const bitCapInt& to
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, pow2Ocl(carryStart), length, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -470,7 +462,6 @@ void QEngineCPU::CMULDIV(const IOFn& inFn, const IOFn& outFn, const bitCapInt& t
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_mask(0, maxQPowerOcl, skipPowers, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -563,7 +554,6 @@ void QEngineCPU::ModNOut(const MFn& kernelFn, const bitCapInt& modN, const bitLe
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, pow2Ocl(outStart), length, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -656,7 +646,6 @@ void QEngineCPU::CModNOut(const MFn& kernelFn, const bitCapInt& modN, const bitL
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_mask(0, maxQPowerOcl, skipPowers, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -758,7 +747,6 @@ void QEngineCPU::INCBCD(const bitCapInt& toAdd, bitLenInt inOutStart, bitLenInt
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for(0, maxQPowerOcl, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -833,7 +821,6 @@ void QEngineCPU::INCDECBCDC(const bitCapInt& toMod, bitLenInt inOutStart, bitLen
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for_skip(0, maxQPowerOcl, pow2Ocl(carryIndex), ONE_BCI, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl otherRes = lcv & otherMask;
@@ -919,7 +906,6 @@ bitCapInt QEngineCPU::IndexedLDA(bitLenInt indexStart, bitLenInt indexLength, bi
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     ParallelFunc fn;
     if (valueBytes == 1) {
@@ -1004,7 +990,6 @@ bitCapInt QEngineCPU::IndexedADC(bitLenInt indexStart, bitLenInt indexLength, bi
     // We calloc a new stateVector for output.
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     // We're going to loop over every eigenstate in the vector, (except, we
     // already know the carry is zero).  This bit masks let us quickly
@@ -1121,7 +1106,6 @@ bitCapInt QEngineCPU::IndexedSBC(bitLenInt indexStart, bitLenInt indexLength, bi
     // We calloc a new stateVector for output.
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     // We're going to loop over every eigenstate in the vector, (except, we already know the carry is zero).
     // This bit masks let us quickly distinguish the different values of the input register, output register, carry, and
@@ -1214,7 +1198,6 @@ void QEngineCPU::Hash(bitLenInt start, bitLenInt length, const unsigned char* va
 
     StateVectorPtr nStateVec = AllocStateVec(maxQPowerOcl);
     nStateVec->clear();
-    stateVec->isReadLocked = false;
 
     par_for(0, maxQPowerOcl, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
         const bitCapIntOcl inputRes = lcv & inputMask;