add doc for rambo.py & change unpack in idtr.cpp

Author: AllanZyne

examples/rambo.py

@@ -1,8 +1,13 @@
 """
+Rambo benchmark
 
 Examples:
-    python rambo.py -nevts 10 -nout 10 -b sharpy -i 10000
-    mpiexec -n 3 python rambo.py -nevts 64 -nout 64 -b sharpy -i 100
+
+    # run 1000 iterations of 10 events and 100 outputs on sharpy backend
+    python rambo.py -nevts 10 -nout 100 -b sharpy -i 1000
+
+    # MPI parallel run
+    mpiexec -n 3 python rambo.py -nevts 64 -nout 64 -b sharpy -i 1000
 
 """
 
@@ -31,27 +36,6 @@ def info(s):
         print(s)
 
 
-def naive_erf(x):
-    """
-    Error function (erf) implementation
-
-    Adapted from formula 7.1.26 in
-    Abramowitz and Stegun, "Handbook of Mathematical Functions", 1965.
-    """
-    y = numpy.abs(x)
-
-    a1 = 0.254829592
-    a2 = -0.284496736
-    a3 = 1.421413741
-    a4 = -1.453152027
-    a5 = 1.061405429
-    p = 0.3275911
-
-    t = 1.0 / (1.0 + p * y)
-    f = (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t
-    return numpy.sign(x) * (1.0 - f * numpy.exp(-y * y))
-
-
 def sp_rambo(sp, sp_C1, sp_F1, sp_Q1, sp_output, nevts, nout):
     sp_C = 2.0 * sp_C1 - 1.0
     sp_S = sp.sqrt(1 - sp.square(sp_C))
@@ -159,14 +143,12 @@ def eval():
     t_min = numpy.min(time_list)
     t_max = numpy.max(time_list)
     t_med = numpy.median(time_list)
-    # perf_rate = nopt / t_med / 1e6  # million options per second
     init_overhead = t_warm - t_med
     if backend == "sharpy":
         info(f"Estimated initialization overhead: {init_overhead:.5f} s")
     info(f"Min.   duration: {t_min:.5f} s")
     info(f"Max.   duration: {t_max:.5f} s")
     info(f"Median duration: {t_med:.5f} s")
-    # info(f"Median rate: {perf_rate:.5f} Mopts/s")
 
     fini()
 

src/idtr.cpp

@@ -241,10 +241,11 @@ void bufferize(void *cptr, SHARPY::DTypeId dtype, const int64_t *sizes,
            });
 }
 
-/// copy contiguous block of data into a possibly strided array
-void unpack(void *in, SHARPY::DTypeId dtype, const int64_t *sizes,
-            const int64_t *strides, const int64_t *tStarts,
-            const int64_t *tSizes, uint64_t nd, uint64_t N, void *out) {
+/// copy contiguous block of data into a possibly strided array distributed to N
+/// ranks
+void unpackN(void *in, SHARPY::DTypeId dtype, const int64_t *sizes,
+             const int64_t *strides, const int64_t *tStarts,
+             const int64_t *tSizes, uint64_t nd, uint64_t N, void *out) {
   if (!in || !sizes || !strides || !tStarts || !tSizes || !out) {
     return;
   }
@@ -269,8 +270,8 @@ void unpack(void *in, SHARPY::DTypeId dtype, const int64_t *sizes,
 }
 
 /// copy contiguous block of data into a possibly strided array
-void unpack1(void *in, SHARPY::DTypeId dtype, const int64_t *sizes,
-             const int64_t *strides, uint64_t ndim, void *out) {
+void unpack(void *in, SHARPY::DTypeId dtype, const int64_t *sizes,
+            const int64_t *strides, uint64_t ndim, void *out) {
   if (!in || !sizes || !strides || !out) {
     return;
   }
@@ -522,8 +523,8 @@ WaitHandleBase *_idtr_copy_reshape(SHARPY::DTypeId sharpytype,
                roffs = std::move(roffs)]() {
     tc->wait(hdl);
     if (isStrided) {
-      unpack1(rBuff, sharpytype, oDataShapePtr, oDataStridesPtr, oNDims,
-              oDataPtr);
+      unpack(rBuff, sharpytype, oDataShapePtr, oDataStridesPtr, oNDims,
+             oDataPtr);
       delete[](char *) rBuff;
     }
   };
@@ -931,15 +932,15 @@ void *_idtr_update_halo(SHARPY::DTypeId sharpytype, int64_t ndims,
     tc->wait(lwh);
     std::vector<int64_t> recvBufferStart(nworkers * ndims, 0);
     if (cache->_bufferizeLRecv) {
-      unpack(lRecvData, sharpytype, leftHaloShape, leftHaloStride,
-             recvBufferStart.data(), cache->_lRecvBufferSize.data(), ndims,
-             nworkers, leftHaloData);
+      unpackN(lRecvData, sharpytype, leftHaloShape, leftHaloStride,
+              recvBufferStart.data(), cache->_lRecvBufferSize.data(), ndims,
+              nworkers, leftHaloData);
     }
     tc->wait(rwh);
     if (cache->_bufferizeRRecv) {
-      unpack(rRecvData, sharpytype, rightHaloShape, rightHaloStride,
-             recvBufferStart.data(), cache->_rRecvBufferSize.data(), ndims,
-             nworkers, rightHaloData);
+      unpackN(rRecvData, sharpytype, rightHaloShape, rightHaloStride,
+              recvBufferStart.data(), cache->_rRecvBufferSize.data(), ndims,
+              nworkers, rightHaloData);
     }
   };