pyvpmr 250526__tar.gz → 250527__tar.gz

{pyvpmr-250526 → pyvpmr-250527}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyvpmr
-Version: 250526
+Version: 250527
 Summary: The VPMR Algorithm
 Author-email: Theodore Chang <tlcfem@gmail.com>
 Maintainer-email: Theodore Chang <tlcfem@gmail.com>
@@ -813,7 +813,7 @@ Options:
     -c, --max-exponent          <int>     maximum exponent (default: 4)
     -d, --precision-bits        <int>     number of precision bits (default: 512)
     -q, --quadrature-order      <int>     quadrature order (default: 500)
-    -m, --precision-multiplier  <float>   precision multiplier (default: 1.5)
+    -m, --precision-multiplier  <float>   precision multiplier (default: 1.05)
     -e, --tolerance             <float>   tolerance (default: 1E-8)
     -k, --kernel                <string>  file name of kernel function (default uses: exp(-t^2/4))
     -s, --singular-values                 print singular values
@@ -842,7 +842,7 @@ Using the following parameters:
     exponent = 4.
    precision = 355.
  quad. order = 500.
-  multiplier = 1.5000e+00.
+  multiplier = 1.0500e+00.
    tolerance = 1.0000e-08.
       kernel = exp(-t*t/4).
 
@@ -854,7 +854,7 @@ Using the following parameters:
 [6/6] Done.
 
 M = 
-+1.1745193571738943e+01+6.4089561283054790e-107j
++1.1745193571738943e+01+6.4196555242161141e-106j
 -5.5143304351134397e+00+5.7204056791636839e+00j
 -5.5143304351134397e+00-5.7204056791636839e+00j
 -1.6161617424833762e-02+2.3459542440459513e+00j
@@ -874,7 +874,7 @@ S =
 +1.7655956664692953e+00-2.7555720406099038e+00j
 +1.7655956664692953e+00+2.7555720406099038e+00j
 
-Running time: 3112 ms.
+Running time: 2044 ms.
 ```
 
 ![exp(-t^2/4)](https://raw.githubusercontent.com/TLCFEM/vpmr/master/resource/example.png)
@@ -906,8 +906,8 @@ echo "exp(-t*t/10)" > kernel.txt
 
 ## Performance
 
-The computation of weights, that involves integrals, and SVD are parallelised.
-A typical profiling would yield something similar to the following.
+The majority of the algorithm is parallelised to extract the maximum performance.
+The following is a typical performance profile on a i7-10750H platform using the `./vpmr -n 80`.
 
 ![profiling](https://raw.githubusercontent.com/TLCFEM/vpmr/master/resource/profile.png)
 

{pyvpmr-250526 → pyvpmr-250527}/README.md

@@ -106,7 +106,7 @@ Options:
     -c, --max-exponent          <int>     maximum exponent (default: 4)
     -d, --precision-bits        <int>     number of precision bits (default: 512)
     -q, --quadrature-order      <int>     quadrature order (default: 500)
-    -m, --precision-multiplier  <float>   precision multiplier (default: 1.5)
+    -m, --precision-multiplier  <float>   precision multiplier (default: 1.05)
     -e, --tolerance             <float>   tolerance (default: 1E-8)
     -k, --kernel                <string>  file name of kernel function (default uses: exp(-t^2/4))
     -s, --singular-values                 print singular values
@@ -135,7 +135,7 @@ Using the following parameters:
     exponent = 4.
    precision = 355.
  quad. order = 500.
-  multiplier = 1.5000e+00.
+  multiplier = 1.0500e+00.
    tolerance = 1.0000e-08.
       kernel = exp(-t*t/4).
 
@@ -147,7 +147,7 @@ Using the following parameters:
 [6/6] Done.
 
 M = 
-+1.1745193571738943e+01+6.4089561283054790e-107j
++1.1745193571738943e+01+6.4196555242161141e-106j
 -5.5143304351134397e+00+5.7204056791636839e+00j
 -5.5143304351134397e+00-5.7204056791636839e+00j
 -1.6161617424833762e-02+2.3459542440459513e+00j
@@ -167,7 +167,7 @@ S =
 +1.7655956664692953e+00-2.7555720406099038e+00j
 +1.7655956664692953e+00+2.7555720406099038e+00j
 
-Running time: 3112 ms.
+Running time: 2044 ms.
 ```
 
 ![exp(-t^2/4)](resource/example.png)
@@ -199,8 +199,8 @@ echo "exp(-t*t/10)" > kernel.txt
 
 ## Performance
 
-The computation of weights, that involves integrals, and SVD are parallelised.
-A typical profiling would yield something similar to the following.
+The majority of the algorithm is parallelised to extract the maximum performance.
+The following is a typical performance profile on a i7-10750H platform using the `./vpmr -n 80`.
 
 ![profiling](resource/profile.png)
 

{pyvpmr-250526 → pyvpmr-250527}/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h

@@ -1466,8 +1466,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
     const Index actual_panel_rows =
         (3 * LhsProgress) * std::max<Index>(1, ((l1 - sizeof(ResScalar) * mr * nr - depth * nr * sizeof(RhsScalar)) /
                                                 (depth * sizeof(LhsScalar) * 3 * LhsProgress)));
-    tbb::parallel_for(Index(0), (peeled_mc3 + actual_panel_rows - 1) / actual_panel_rows, [&](Index i1) {
-      i1 *= actual_panel_rows;
+    for (Index i1 = 0; i1 < peeled_mc3; i1 += actual_panel_rows) {
       const Index actual_panel_end = (std::min)(i1 + actual_panel_rows, peeled_mc3);
 #if EIGEN_ARCH_ARM64 || EIGEN_ARCH_LOONGARCH64
       EIGEN_IF_CONSTEXPR(nr >= 8) {
@@ -1922,7 +1921,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
           r0.storePacket(2 * Traits::ResPacketSize, R2);
         }
       }
-    });
+    }
   }
 
   //---------- Process 2 * LhsProgress rows at once ----------
@@ -1935,8 +1934,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
         (2 * LhsProgress) * std::max<Index>(1, ((l1 - sizeof(ResScalar) * mr * nr - depth * nr * sizeof(RhsScalar)) /
                                                 (depth * sizeof(LhsScalar) * 2 * LhsProgress)));
 
-    tbb::parallel_for(peeled_mc3, (peeled_mc2 + actual_panel_rows - 1) / actual_panel_rows, [&](Index i1) {
-      i1 *= actual_panel_rows;
+    for (Index i1 = peeled_mc3; i1 < peeled_mc2; i1 += actual_panel_rows) {
       Index actual_panel_end = (std::min)(i1 + actual_panel_rows, peeled_mc2);
 #if EIGEN_ARCH_ARM64 || EIGEN_ARCH_LOONGARCH64
       EIGEN_IF_CONSTEXPR(nr >= 8) {
@@ -2110,7 +2108,8 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
         }
       }
 #endif
-      for (Index j2 = packet_cols8; j2 < packet_cols4; j2 += 4) {
+      tbb::parallel_for(packet_cols8 / 4, packet_cols4 / 4, [&](Index j2) {
+        j2 *= 4;
         for (Index i = i1; i < actual_panel_end; i += 2 * LhsProgress) {
           // We selected a 2*Traits::LhsProgress x nr micro block of res which is entirely
           // stored into 2 x nr registers.
@@ -2229,7 +2228,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
           r3.storePacket(0 * Traits::ResPacketSize, R2);
           r3.storePacket(1 * Traits::ResPacketSize, R3);
         }
-      }
+      });
 
       // Deal with remaining columns of the rhs
       for (Index j2 = packet_cols4; j2 < cols; j2++) {
@@ -2300,7 +2299,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
           r0.storePacket(1 * Traits::ResPacketSize, R1);
         }
       }
-    });
+    }
   }
   //---------- Process 1 * LhsProgress rows at once ----------
   if (mr >= 1 * Traits::LhsProgress) {

{pyvpmr-250526 → pyvpmr-250527}/patch_size.patch

@@ -1,45 +1,26 @@
 diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
-index e72c6b48e..a72e7e53e 100644
+index e72c6b48e..4f918cd20 100644
 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
 +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
-@@ -1466,7 +1466,8 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
-     const Index actual_panel_rows =
-         (3 * LhsProgress) * std::max<Index>(1, ((l1 - sizeof(ResScalar) * mr * nr - depth * nr * sizeof(RhsScalar)) /
-                                                 (depth * sizeof(LhsScalar) * 3 * LhsProgress)));
--    for (Index i1 = 0; i1 < peeled_mc3; i1 += actual_panel_rows) {
-+    tbb::parallel_for(Index(0), (peeled_mc3 + actual_panel_rows - 1) / actual_panel_rows, [&](Index i1) {
-+      i1 *= actual_panel_rows;
-       const Index actual_panel_end = (std::min)(i1 + actual_panel_rows, peeled_mc3);
- #if EIGEN_ARCH_ARM64 || EIGEN_ARCH_LOONGARCH64
-       EIGEN_IF_CONSTEXPR(nr >= 8) {
-@@ -1921,7 +1922,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
-           r0.storePacket(2 * Traits::ResPacketSize, R2);
+@@ -2108,7 +2108,8 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
          }
        }
--    }
-+    });
-   }
- 
-   //---------- Process 2 * LhsProgress rows at once ----------
-@@ -1934,7 +1935,8 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
-         (2 * LhsProgress) * std::max<Index>(1, ((l1 - sizeof(ResScalar) * mr * nr - depth * nr * sizeof(RhsScalar)) /
-                                                 (depth * sizeof(LhsScalar) * 2 * LhsProgress)));
- 
--    for (Index i1 = peeled_mc3; i1 < peeled_mc2; i1 += actual_panel_rows) {
-+    tbb::parallel_for(peeled_mc3, (peeled_mc2 + actual_panel_rows - 1) / actual_panel_rows, [&](Index i1) {
-+      i1 *= actual_panel_rows;
-       Index actual_panel_end = (std::min)(i1 + actual_panel_rows, peeled_mc2);
- #if EIGEN_ARCH_ARM64 || EIGEN_ARCH_LOONGARCH64
-       EIGEN_IF_CONSTEXPR(nr >= 8) {
-@@ -2298,7 +2300,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
-           r0.storePacket(1 * Traits::ResPacketSize, R1);
+ #endif
+-      for (Index j2 = packet_cols8; j2 < packet_cols4; j2 += 4) {
++      tbb::parallel_for(packet_cols8 / 4, packet_cols4 / 4, [&](Index j2) {
++        j2 *= 4;
+         for (Index i = i1; i < actual_panel_end; i += 2 * LhsProgress) {
+           // We selected a 2*Traits::LhsProgress x nr micro block of res which is entirely
+           // stored into 2 x nr registers.
+@@ -2227,7 +2228,7 @@ EIGEN_DONT_INLINE void gebp_kernel<LhsScalar, RhsScalar, Index, DataMapper, mr,
+           r3.storePacket(0 * Traits::ResPacketSize, R2);
+           r3.storePacket(1 * Traits::ResPacketSize, R3);
          }
-       }
--    }
-+    });
-   }
-   //---------- Process 1 * LhsProgress rows at once ----------
-   if (mr >= 1 * Traits::LhsProgress) {
+-      }
++      });
+ 
+       // Deal with remaining columns of the rhs
+       for (Index j2 = packet_cols4; j2 < cols; j2++) {
 diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h
 index ba72a8a4f..4ce9f99ab 100644
 --- a/Eigen/src/Core/products/GeneralMatrixVector.h

{pyvpmr-250526 → pyvpmr-250527}/pyvpmr.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyvpmr
-Version: 250526
+Version: 250527
 Summary: The VPMR Algorithm
 Author-email: Theodore Chang <tlcfem@gmail.com>
 Maintainer-email: Theodore Chang <tlcfem@gmail.com>
@@ -813,7 +813,7 @@ Options:
     -c, --max-exponent          <int>     maximum exponent (default: 4)
     -d, --precision-bits        <int>     number of precision bits (default: 512)
     -q, --quadrature-order      <int>     quadrature order (default: 500)
-    -m, --precision-multiplier  <float>   precision multiplier (default: 1.5)
+    -m, --precision-multiplier  <float>   precision multiplier (default: 1.05)
     -e, --tolerance             <float>   tolerance (default: 1E-8)
     -k, --kernel                <string>  file name of kernel function (default uses: exp(-t^2/4))
     -s, --singular-values                 print singular values
@@ -842,7 +842,7 @@ Using the following parameters:
     exponent = 4.
    precision = 355.
  quad. order = 500.
-  multiplier = 1.5000e+00.
+  multiplier = 1.0500e+00.
    tolerance = 1.0000e-08.
       kernel = exp(-t*t/4).
 
@@ -854,7 +854,7 @@ Using the following parameters:
 [6/6] Done.
 
 M = 
-+1.1745193571738943e+01+6.4089561283054790e-107j
++1.1745193571738943e+01+6.4196555242161141e-106j
 -5.5143304351134397e+00+5.7204056791636839e+00j
 -5.5143304351134397e+00-5.7204056791636839e+00j
 -1.6161617424833762e-02+2.3459542440459513e+00j
@@ -874,7 +874,7 @@ S =
 +1.7655956664692953e+00-2.7555720406099038e+00j
 +1.7655956664692953e+00+2.7555720406099038e+00j
 
-Running time: 3112 ms.
+Running time: 2044 ms.
 ```
 
 ![exp(-t^2/4)](https://raw.githubusercontent.com/TLCFEM/vpmr/master/resource/example.png)
@@ -906,8 +906,8 @@ echo "exp(-t*t/10)" > kernel.txt
 
 ## Performance
 
-The computation of weights, that involves integrals, and SVD are parallelised.
-A typical profiling would yield something similar to the following.
+The majority of the algorithm is parallelised to extract the maximum performance.
+The following is a typical performance profile on a i7-10750H platform using the `./vpmr -n 80`.
 
 ![profiling](https://raw.githubusercontent.com/TLCFEM/vpmr/master/resource/profile.png)
 

{pyvpmr-250526 → pyvpmr-250527}/src/VPMR.cpp

@@ -39,7 +39,7 @@ struct Config {
     int max_exponent = 4;
     int precision_bits = 512;
     int quadrature_order = 500;
-    double precision_multiplier = 1.5;
+    double precision_multiplier = 1.05;
     mpreal tolerance{"1E-8", precision_bits};
     std::string kernel = "exp(-t*t/4)";
 
@@ -432,7 +432,7 @@ int print_helper() {
     std::cout << "    -c, --max-exponent          <int>     maximum exponent (default: 4)\n";
     std::cout << "    -d, --precision-bits        <int>     number of precision bits (default: 512)\n";
     std::cout << "    -q, --quadrature-order      <int>     quadrature order (default: 500)\n";
-    std::cout << "    -m, --precision-multiplier  <float>   precision multiplier (default: 1.5)\n";
+    std::cout << "    -m, --precision-multiplier  <float>   precision multiplier (default: 1.05)\n";
     std::cout << "    -e, --tolerance             <float>   tolerance (default: 1E-8)\n";
     std::cout << "    -k, --kernel                <string>  file name of kernel function (default uses: exp(-t^2/4))\n";
     std::cout << "    -s, --singular-values                 print singular values\n";
@@ -554,7 +554,7 @@ std::tuple<std::vector<std::complex<double>>, std::vector<std::complex<double>>>
     config.max_exponent = std::max(1, c);
     config.precision_bits = std::max(1, d);
     config.quadrature_order = std::max(1, q);
-    config.precision_multiplier = std::max(1.5, m);
+    config.precision_multiplier = std::max(1.05, m);
     config.tolerance = mpreal(e);
     if(!k.empty()) config.kernel = k;
 
@@ -606,11 +606,11 @@ PYBIND11_MODULE(_pyvpmr, m) {
         ":param c: maximum exponent (default: 4)\n"
         ":param d: number of precision bits (default: 512)\n"
         ":param q: quadrature order (default: 500)\n"
-        ":param m: precision multiplier (default: 1.5)\n"
+        ":param m: precision multiplier (default: 1.05)\n"
         ":param e: tolerance (default: 1E-8)\n"
         ":param k: kernel function (default: exp(-t^2/4))\n"
         ":return: M, S\n",
         pybind11::call_guard<pybind11::gil_scoped_release>(),
-        pybind11::kw_only(), pybind11::arg("n") = 10, pybind11::arg("c") = 4, pybind11::arg("d") = 0, pybind11::arg("q") = 500, pybind11::arg("m") = 1.5, pybind11::arg("e") = 1E-8, pybind11::arg("k") = "");
+        pybind11::kw_only(), pybind11::arg("n") = 10, pybind11::arg("c") = 4, pybind11::arg("d") = 0, pybind11::arg("q") = 500, pybind11::arg("m") = 1.05, pybind11::arg("e") = 1E-8, pybind11::arg("k") = "");
 }
 #endif