current state

pypocketfft/demos/bench.py

@@ -118,25 +118,26 @@ def bench_nd(ndim, nmax, nthr, ntry, tp, funcs, nrepeat, ttl="", filename="",
             tmp = func(a, nrepeat, nthr)
             res.append(tmp[0])
             output.append(tmp[1])
-        print("{0:5.2e}/{1:5.2e} = {2:5.2f}  L2 error={3}".format(results[0][n],results[1][n],results[0][n]/results[1][n],_l2error(output[0],output[1])))
-    results = np.array(results)
-    plt.title("{}: {}D, {}, max_extent={}".format(
-        ttl, ndim, str(tp), nmax))
-    plt.xlabel("time ratio")
-    plt.ylabel("counts")
-    plt.hist(results[0, :]/results[1, :], bins="auto")
-    if filename != "":
-        plt.savefig(filename)
-    plt.show()
-
-
-funcs = (measure_pypocketfft, measure_fftw)
+#         print("{0:5.2e}/{1:5.2e} = {2:5.2f}  L2 error={3}".format(results[0][n],results[1][n],results[0][n]/results[1][n],_l2error(output[0],output[1])))
+#     results = np.array(results)
+#     plt.title("{}: {}D, {}, max_extent={}".format(
+#         ttl, ndim, str(tp), nmax))
+#     plt.xlabel("time ratio")
+#     plt.ylabel("counts")
+#     plt.hist(results[0, :]/results[1, :], bins="auto")
+#     if filename != "":
+#         plt.savefig(filename)
+#     plt.show()
+
+
+funcs = (measure_pypocketfft,)
 ttl = "pypocketfft/FFTW()"
+ntry=100
 nthr = 1
 nice_sizes = True
-bench_nd(1, 8192, nthr, 100, "c16", funcs, 10, ttl, "1d.png", nice_sizes)
-bench_nd(2, 2048, nthr, 100, "c16", funcs, 2, ttl, "2d.png", nice_sizes)
-bench_nd(3, 256, nthr, 100, "c16", funcs, 2, ttl, "3d.png", nice_sizes)
-bench_nd(1, 8192, nthr, 100, "c8", funcs, 10, ttl, "1d_single.png", nice_sizes)
-bench_nd(2, 2048, nthr, 100, "c8", funcs, 2, ttl, "2d_single.png", nice_sizes)
-bench_nd(3, 256, nthr, 100, "c8", funcs, 2, ttl, "3d_single.png", nice_sizes)
+#bench_nd(1, 8192, nthr, ntry, "c16", funcs, 10, ttl, "1d.png", nice_sizes)
+bench_nd(2, 2048, nthr, ntry, "c16", funcs, 2, ttl, "2d.png", nice_sizes)
+# bench_nd(3, 256, nthr, ntry, "c16", funcs, 2, ttl, "3d.png", nice_sizes)
+# bench_nd(1, 8192, nthr, ntry, "c8", funcs, 10, ttl, "1d_single.png", nice_sizes)
+# bench_nd(2, 2048, nthr, ntry, "c8", funcs, 2, ttl, "2d_single.png", nice_sizes)
+# bench_nd(3, 256, nthr, ntry, "c8", funcs, 2, ttl, "3d_single.png", nice_sizes)

src/mr_util/infra/communication.h

@@ -24,6 +24,8 @@
 #ifndef MRUTIL_COMMUNICATION_H
 #define MRUTIL_COMMUNICATION_H
 
+#define MRUTIL_USE_MPI
+
 #include <vector>
 #ifdef MRUTIL_USE_MPI
 #include <mpi.h>
@@ -95,8 +97,7 @@ class Communicator
     template<typename T> void sendrecvRaw (const T *sendbuf, size_t sendcnt,
       size_t dest, T *recvbuf, size_t recvcnt, size_t src) const
       {
-      sendrecvRawVoid(sendbuf, sendcnt, dest, recvbuf, recvcnt, src,
-                      tidx<T>());
+      sendrecvRawVoid(sendbuf, sendcnt, dest, recvbuf, recvcnt, src, tidx<T>());
       }
     template<typename T> void sendrecv_replaceRaw (T *data, size_t num,
       size_t dest, size_t src) const
@@ -110,6 +111,12 @@ class Communicator
     template<typename T> void allgathervRaw (const T *in, int numin, T *out,
       const int *numout, const int *disout) const
       { allgathervRawVoid (in, numin, out, numout, disout, tidx<T>()); }
+    template<typename T> vector<T> allgatherVec (const T &in) const
+      {
+      vector<T> res(num_ranks_);
+      allgatherRaw(&in, res.data(), 1);
+      return res;
+      }
 
     template<typename T> T allreduce(const T &in, redOp op) const
       {
@@ -117,13 +124,18 @@ class Communicator
       allreduceRaw (&in, &out, 1, op);
       return out;
       }
-    template<typename T> std::vector<T> allreduce
+    template<typename T> std::vector<T> allreduceVec
       (const std::vector<T> &in, redOp op) const
       {
       std::vector<T> out(in.size());
       allreduceRaw (in.data(), out.data(), in.size(), op);
       return out;
       }
+    template<typename T> void sendrecvVec(const vector<T> &sendbuf, size_t dest,
+      vector<T> &recvbuf, size_t src) const
+      {
+      sendrecvRaw(sendbuf.data(), sendbuf.size(), dest, recvbuf.data(), recvbuf.size(), src);
+      }
     /*! NB: \a num refers to the <i>total</i> number of items in the arrays;
         the individual message size is \a num/num_ranks(). */
     template<typename T> void all2allRaw (const T *in, T *out, size_t num) const

src/mr_util/infra/simd.h

@@ -52,6 +52,10 @@ namespace mr {
 
 namespace detail_simd {
 
+template<typename T> T myexp(T);// {return -42;}
+template<> inline double myexp(double v) {return std::exp(v);}
+template<> inline float myexp(float v) {return std::exp(v);}
+
 template<typename T> constexpr inline bool vectorizable = false;
 template<> constexpr inline bool vectorizable<float> = true;
 template<> constexpr inline bool vectorizable<double> = true;
@@ -114,6 +118,13 @@ template<typename T, size_t len> class vtp
     vtp &operator*=(vtp other) { v*=other.v; return *this; }
     vtp &operator/=(vtp other) { v/=other.v; return *this; }
     vtp abs() const { return hlp::abs(v); }
+    template<typename Func> vtp apply(Func func) const
+      {
+      vtp res;
+      for (size_t i=0; i<len; ++i)
+        res[i] = func(v[i]);
+      return res;
+      }
     inline vtp sqrt() const
       { return hlp::sqrt(v); }
     vtp max(const vtp &other) const
@@ -188,6 +199,8 @@ template<typename Op, typename T, size_t len> T reduce(const vtp<T, len> &v, Op
     res = op(res, v[i]);
   return res;
   }
+template<typename T, size_t len> vtp<T, len> exp(const vtp<T, len> &v)
+  { return v.apply(myexp<T>); }
 template<typename T> class pseudoscalar
   {
   private:
@@ -407,6 +420,7 @@ using detail_simd::native_simd;
 using detail_simd::reduce;
 using detail_simd::max;
 using detail_simd::abs;
+using detail_simd::exp;
 using detail_simd::sqrt;
 using detail_simd::any_of;
 using detail_simd::none_of;