Optimize Decompose()/Dispose()

unitaryfund · Nov 1, 2024 · 399e28b · 399e28b
1 parent 0dd1e6c
commit 399e28b
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 29 deletions.
diff --git a/src/common/qengine.cu b/src/common/qengine.cu
@@ -671,9 +671,7 @@ __global__ void decomposeprob(qCudaCmplx* stateVec, bitCapIntOcl* bitCapIntOclPt
         qCudaReal1_f partProb = ZERO_R1_CUDA;
 
         for (bitCapIntOcl k = 0U; k < partPower; k++) {
-            bitCapIntOcl l = j | (k << start);
-
-            const qCudaCmplx amp = stateVec[l];
+            const qCudaCmplx amp = stateVec[j | (k << start)];
             const qCudaReal1_f nrm = (qCudaReal1_f)qCudaDot(amp, amp);
             partProb += nrm;
             partStateAngle[k] += qCudaArg(amp) * (qCudaReal1)nrm;
@@ -734,9 +732,7 @@ __global__ void disposeprob(qCudaCmplx* stateVec, bitCapIntOcl* bitCapIntOclPtr,
         qCudaReal1 partProb = ZERO_R1_CUDA;
 
         for (bitCapIntOcl k = 0U; k < partPower; k++) {
-            bitCapIntOcl l = j | (k << start);
-
-            qCudaCmplx amp = stateVec[l];
+            qCudaCmplx amp = stateVec[j | (k << start)];
             qCudaReal1 nrm = qCudaDot(amp, amp);
             partProb += nrm;
         }
@@ -754,16 +750,9 @@ __global__ void disposeprob(qCudaCmplx* stateVec, bitCapIntOcl* bitCapIntOclPtr,
             l |= (k ^ l) << len;
             l = j | l;
 
-            qCudaCmplx amp = stateVec[l];
-            qCudaReal1_f nrm = (qCudaReal1_f)qCudaDot(amp, amp);
-
-            if (nrm >= REAL1_EPSILON_CUDA) {
-                qCudaReal1 currentAngle = qCudaArg(amp);
-                if (firstAngle < angleThresh) {
-                    firstAngle = currentAngle;
-                }
-                remainderStateAngle[k] = currentAngle - (qCudaReal1)firstAngle;
-            }
+            const qCudaCmplx amp = stateVec[l];
+            const qCudaReal1_f nrm = (qCudaReal1_f)qCudaDot(amp, amp);
+            remainderStateAngle[k] += qCudaArg(amp) * (qCudaReal1)nrm;
         }
     }
 }

diff --git a/src/qengine/state.cpp b/src/qengine/state.cpp
@@ -1206,26 +1206,23 @@ void QEngineCPU::DecomposeDispose(bitLenInt start, bitLenInt length, QEngineCPUP
 
     if (destination) {
         par_for(0U, remainderPower, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
-            bitCapIntOcl j;
-            j = lcv & pow2MaskOcl(start);
+            bitCapIntOcl j = lcv & pow2MaskOcl(start);
             j |= (lcv ^ j) << length;
 
             for (bitCapIntOcl k = 0U; k < partPower; ++k) {
-                bitCapIntOcl l = j | (k << start);
-
-                const complex amp = stateVec->read(l);
+                const complex amp = stateVec->read(j | (k << start));
                 const real1 nrm = norm(amp);
                 remainderStateProb[lcv] += nrm;
                 partStateAngle[k] += arg(amp) * nrm;
             }
         });
 
         par_for(0U, partPower, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
-            bitCapIntOcl j;
-            j = lcv << start;
+            const bitCapIntOcl startMask = pow2MaskOcl(start);
+            const bitCapIntOcl j = lcv << start;
 
             for (bitCapIntOcl k = 0U; k < remainderPower; ++k) {
-                bitCapIntOcl l = k & pow2MaskOcl(start);
+                bitCapIntOcl l = k & startMask;
                 l |= j | ((k ^ l) << length);
 
                 const complex amp = stateVec->read(l);
@@ -1253,17 +1250,16 @@ void QEngineCPU::DecomposeDispose(bitLenInt start, bitLenInt length, QEngineCPUP
             j |= (lcv ^ j) << length;
 
             for (bitCapIntOcl k = 0U; k < partPower; ++k) {
-                bitCapIntOcl l = j | (k << start);
-                remainderStateProb[lcv] += norm(stateVec->read(l));
+                remainderStateProb[lcv] += norm(stateVec->read(j | (k << start)));
             }
         });
 
         par_for(0U, partPower, [&](const bitCapIntOcl& lcv, const unsigned& cpu) {
-            bitCapIntOcl j;
-            j = lcv << start;
+            const bitCapIntOcl startMask = pow2MaskOcl(start);
+            const bitCapIntOcl j = lcv << start;
 
             for (bitCapIntOcl k = 0U; k < remainderPower; ++k) {
-                bitCapIntOcl l = k & pow2MaskOcl(start);
+                bitCapIntOcl l = k & startMask;
                 l |= j | ((k ^ l) << length);
 
                 const complex amp = stateVec->read(l);