Fix resource managment and a bug in the WProjection gridder's kernel …

…correction function.
lwa-project · Jan 29, 2025 · e74c449 · e74c449
1 parent 7f20444
commit e74c449
Show file tree

Hide file tree

Showing 7 changed files with 73 additions and 71 deletions.
diff --git a/lsl/common/fir.cpp b/lsl/common/fir.cpp
@@ -430,10 +430,10 @@ static PyObject *integerBeamformer(PyObject *self, PyObject *args, PyObject *kwd
             /*
             Cleanup
             */
-//             free(t1);
-//             free(t2);
-//             free(tX);
-//             free(tY);
+            apool.release(t1);
+            apool.release(t2);
+            apool.release(tX);
+            apool.release(tY);
         }
     }
 

diff --git a/lsl/correlator/core.cpp b/lsl/correlator/core.cpp
@@ -227,14 +227,14 @@ void compute_fengine_real(long nStand,
             *(valid + nFFT*i + j) = (unsigned char) cleanFactor;
         }
 
-//         fftwf_free(in);
-//         fftwf_free(out);
+        mpool.release(in);
+        mpool.release(out);
     }
-//     aligned64_free(rot);
+    apool.release(rot);
 
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
-//     fftwf_free(outP);
+    pcache.release_plan(p);
+    mpool.release(inP);
+    mpool.release(outP);
 
     Py_END_ALLOW_THREADS
 
@@ -286,7 +286,7 @@ void compute_fengine_complex(long nStand,
         #pragma omp parallel default(shared) private(in, i, j, k, l, secStart, cleanFactor)
     #endif
     {
-        in = (Complex32*) fftwf_malloc(sizeof(Complex32) * nChan*nTap);
+        in = (Complex32*) mpool.acquire<Complex32>(nChan*nTap);
 
         #ifdef _OPENMP
             #pragma omp for schedule(OMP_SCHEDULER)
@@ -337,12 +337,12 @@ void compute_fengine_complex(long nStand,
             *(valid + nFFT*i + j) = (unsigned char) cleanFactor;
         }
 
-//         fftwf_free(in);
+        mpool.release(in);
     }
-//     aligned64_free(rot);
+    apool.release(rot);
 
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
+    pcache.release_plan(p);
+    mpool.release(inP);
 
     Py_END_ALLOW_THREADS
 }
@@ -446,8 +446,8 @@ static PyObject *FEngine(PyObject *self, PyObject *args, PyObject *kwds) {
     dataF = (PyArrayObject*) PyArray_ZEROS(3, dims, NPY_COMPLEX64, 0);
     if(dataF == NULL) {
         PyErr_Format(PyExc_MemoryError, "Cannot create output array");
-//         aligned64_free(fifo);
-//         aligned64_free(frac);
+        apool.release(fifo);
+        apool.release(frac);
         goto fail;
     }
 
@@ -458,8 +458,8 @@ static PyObject *FEngine(PyObject *self, PyObject *args, PyObject *kwds) {
     validF = (PyArrayObject*) PyArray_ZEROS(2, dimsV, NPY_UINT8, 0);
     if(validF == NULL) {
         PyErr_Format(PyExc_MemoryError, "Cannot create valid index array");
-//         aligned64_free(fifo);
-//         aligned64_free(frac);
+        apool.release(fifo);
+        apool.release(frac);
         goto fail;
     }
 
@@ -496,8 +496,8 @@ static PyObject *FEngine(PyObject *self, PyObject *args, PyObject *kwds) {
 #undef LAUNCH_FENGINE_REAL
 #undef LAUNCH_FENGINE_COMPLEX
 
-//     aligned64_free(frac);
-//     aligned64_free(fifo);
+    apool.release(frac);
+    apool.release(fifo);
 
     signalsF = Py_BuildValue("(OO)", PyArray_Return(dataF), PyArray_Return(validF));
 
@@ -648,8 +648,8 @@ static PyObject *PFBEngine(PyObject *self, PyObject *args, PyObject *kwds) {
     dataF = (PyArrayObject*) PyArray_ZEROS(3, dims, NPY_COMPLEX64, 0);
     if(dataF == NULL) {
         PyErr_Format(PyExc_MemoryError, "Cannot create output array");
-//         aligned64_free(fifo);
-//         aligned64_free(frac);
+        apool.release(fifo);
+        apool.release(frac);
         goto fail;
     }
 
@@ -660,8 +660,8 @@ static PyObject *PFBEngine(PyObject *self, PyObject *args, PyObject *kwds) {
     validF = (PyArrayObject*) PyArray_ZEROS(2, dimsV, NPY_UINT8, 0);
     if(validF == NULL) {
         PyErr_Format(PyExc_MemoryError, "Cannot create valid index array");
-//         aligned64_free(fifo);
-//         aligned64_free(frac);
+        apool.release(fifo);
+        apool.release(frac);
         goto fail;
     }
 
@@ -698,9 +698,9 @@ static PyObject *PFBEngine(PyObject *self, PyObject *args, PyObject *kwds) {
 #undef LAUNCH_PFBENGINE_REAL
 #undef LAUNCH_PFBENGINE_COMPLEX
 
-//     aligned64_free(frac);
-//     aligned64_free(fifo);
-//     aligned64_free(pfb);
+    apool.release(frac);
+    apool.release(fifo);
+    apool.release(pfb);
 
     signalsF = Py_BuildValue("(OO)", PyArray_Return(dataF), PyArray_Return(validF));
 
@@ -714,7 +714,7 @@ static PyObject *PFBEngine(PyObject *self, PyObject *args, PyObject *kwds) {
 
 fail:
     if( pfb != NULL ) {
-//         aligned64_free(pfb);
+        apool.release(pfb);
     }
     Py_XDECREF(data);
     Py_XDECREF(freq);

diff --git a/lsl/correlator/pool.hpp b/lsl/correlator/pool.hpp
@@ -75,7 +75,7 @@ class Aligned64BufferPool {
                 return;
             }
         }
-        throw std::runtime_error("Attempted to release unmanaged aligned buffer");
+        throw std::runtime_error("Attempted to release unmanaged aligned64 buffer");
     }
 
     std::tuple<size_t, size_t> get_stats() {
@@ -163,7 +163,7 @@ class FFTWBufferPool {
                 return;
             }
         }
-        throw std::runtime_error("Attempted to release unmanaged buffer");
+        throw std::runtime_error("Attempted to release unmanaged fftw buffer");
     }
 
     std::tuple<size_t, size_t> get_stats() {

diff --git a/lsl/correlator/spec.cpp b/lsl/correlator/spec.cpp
@@ -143,12 +143,12 @@ void compute_spec_real(long nStand,
             blas_scal(nChan, 1.0/(2*nChan*nActFFT), (psd + i*nChan), 1);
         }
 
-//         fftwf_free(in);
-//         fftwf_free(out);
+        mpool.release(in);
+        mpool.release(out);
     }
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
-//     fftwf_free(outP);
+    pcache.release_plan(p);
+    mpool.release(inP);
+    mpool.release(outP);
 
     Py_END_ALLOW_THREADS
 }
@@ -246,11 +246,11 @@ void compute_spec_complex(long nStand,
             blas_scal(nChan, 1.0/(nActFFT*nChan), (psd + i*nChan), 1);
         }
 
-//         fftwf_free(in);
-//         aligned64_free(temp2);
+        mpool.release(in);
+        apool.release(temp2);
     }
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
+    pcache.release_plan(p);
+    mpool.release(inP);
 
     Py_END_ALLOW_THREADS
 }
@@ -471,7 +471,7 @@ static PyObject *PFBPSD(PyObject *self, PyObject *args, PyObject *kwds) {
 #undef LAUNCH_PFB_REAL
 #undef LAUNCH_PFB_COMPLEX
 
-//     aligned64_free(pfb);
+    apool.release(pfb);
 
     signalsF = Py_BuildValue("O", PyArray_Return(dataF));
 
@@ -482,7 +482,7 @@ static PyObject *PFBPSD(PyObject *self, PyObject *args, PyObject *kwds) {
 
 fail:
     if( pfb != NULL ) {
-//         aligned64_free(pfb);
+        apool.release(pfb);
     }
     Py_XDECREF(data);
     Py_XDECREF(dataF);

diff --git a/lsl/correlator/stokes.cpp b/lsl/correlator/stokes.cpp
@@ -172,14 +172,14 @@ void compute_stokes_real(long nStand,
             }
         }
 
-//         fftwf_free(inX);
-//         fftwf_free(inY);
-//         fftwf_free(outX);
-//         fftwf_free(outY);
+        mpool.release(inX);
+        mpool.release(inY);
+        mpool.release(outX);
+        mpool.release(outY);
     }
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
-//     fftwf_free(outP);
+    pcache.release_plan(p);
+    mpool.release(inP);
+    mpool.release(outP);
 
     Py_END_ALLOW_THREADS
 }
@@ -305,12 +305,12 @@ void compute_stokes_complex(long nStand,
             }
         }
 
-//         fftwf_free(inX);
-//         fftwf_free(inY);
-//         aligned64_free(temp2);
+        mpool.release(inX);
+        mpool.release(inY);
+        apool.release(temp2);
     }
-//     fftwf_destroy_plan(p);
-//     fftwf_free(inP);
+    pcache.release_plan(p);
+    mpool.release(inP);
 
     Py_END_ALLOW_THREADS
 }
@@ -536,7 +536,7 @@ static PyObject *PFBPSD(PyObject *self, PyObject *args, PyObject *kwds) {
     }
 
     // Calculate the windowing function for the PFB
-    pfb = (double*) aligned64_malloc(sizeof(double) * (1+isReal)*nChan*nTap);
+    pfb = (double*) apool.acquire<double>((1+isReal)*nChan*nTap);
     for(int i=0; i<(1+isReal)*nChan*nTap; i++) {
         *(pfb + i) = sinc((i - (1+isReal)*nChan*nTap/2.0 + 0.5)/((1+isReal)*nChan));
         *(pfb + i) *= hamming(2*NPY_PI*i/((1+isReal)*nChan*nTap));
@@ -583,7 +583,7 @@ static PyObject *PFBPSD(PyObject *self, PyObject *args, PyObject *kwds) {
 #undef LAUNCH_PFB_REAL
 #undef LAUNCH_PFB_COMPLEX
 
-    aligned64_free(pfb);
+    apool.release(pfb);
 
     signalsF = Py_BuildValue("O", PyArray_Return(dataF));
 
@@ -595,7 +595,7 @@ static PyObject *PFBPSD(PyObject *self, PyObject *args, PyObject *kwds) {
 
 fail:
     if( pfb != NULL ) {
-        aligned64_free(pfb);
+        apool.release(pfb);
     }
     Py_XDECREF(dataX);
     Py_XDECREF(dataY);

diff --git a/lsl/imaging/gridder.cpp b/lsl/imaging/gridder.cpp
@@ -164,11 +164,13 @@ void compute_kernel_correction(long nPixSide,
 
     // Inverse transform
     fftwf_plan pB;
+    float *inP = mpool.acquire<float>(nPixSide*GRID_KERNEL_OVERSAMPLE);
     pB = pcache.plan_r2r_1d(nPixSide*GRID_KERNEL_OVERSAMPLE/2,
-                            corr_full, corr_full,
+                            inP, inP,
                             FFTW_REDFT01, FFTW_MEASURE);
-    fftwf_execute(pB);
-//     fftwf_destroy_plan(pB);
+    fftwf_execute_r2r(pB, corr_full, corr_full);
+    pcache.release_plan(pB);
+    mpool.release(inP);
 
     // Select what to keep
     for(i=0; i<nPixSide; i++) {
@@ -190,7 +192,7 @@ void compute_kernel_correction(long nPixSide,
     }
 
     // Cleanup
-//     aligned64_free(corr_full);
+    mpool.release(corr_full);
 }
 
 
@@ -349,23 +351,23 @@ void compute_gridding(long nVis,
             }
         }
 
-//         fftwf_free(suv);
-//         fftwf_free(sbm);
-//         fftwf_free(kern);
+        mpool.release(suv);
+        mpool.release(sbm);
+        mpool.release(kern);
     }
 
     // Correct for the kernel
     compute_kernel_correction(nPixSide, kernel1D, corr);
 
     // Cleanup
-//     fftwf_destroy_plan(pF);
-//     fftwf_destroy_plan(pR);
-//     fftwf_free(inP);
+    pcache.release_plan(pF);
+    pcache.release_plan(pR);
+    mpool.release(inP);
 
-//     aligned64_free(kernel1D);
+    apool.release(kernel1D);
 
-//     free(planeStart);
-//     free(planeStop);
+    apool.release(planeStart);
+    apool.release(planeStop);
 
     Py_END_ALLOW_THREADS
 }

diff --git a/lsl/sim/simfast.cpp b/lsl/sim/simfast.cpp
@@ -148,7 +148,7 @@ void compute_visibility(long nBL,
             }
         }
 
-//         aligned64_free(tempVis);
+        apool.release(tempVis);
     }
 
     Py_END_ALLOW_THREADS
-Original file line number
+Diff line change
@@ Expand Up / @@ -148,7 +148,7 @@ void compute_visibility(long nBL, @@
                 }
             }
-    //         aligned64_free(tempVis);
+            apool.release(tempVis);
         }
         Py_END_ALLOW_THREADS
@@ Expand Down @@