tweak expected size variables; updated perf suite

jrenaud90 · Oct 5, 2023 · b3d800b · b3d800b
1 parent 6efb520
commit b3d800b
Show file tree

Hide file tree

Showing 19 changed files with 88 additions and 74 deletions.
diff --git a/Benchmarks/CyRK - SciPy Comparison.ipynb b/Benchmarks/CyRK - SciPy Comparison.ipynb
diff --git a/Benchmarks/CyRK_CySolver.pdf b/Benchmarks/CyRK_CySolver.pdf
diff --git a/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-3.png b/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-3.png
diff --git a/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-3a0-dev5.png b/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-3a0-dev5.png
diff --git a/Benchmarks/CyRK_cyrk_ode.pdf b/Benchmarks/CyRK_cyrk_ode.pdf
diff --git a/Benchmarks/CyRK_numba.pdf b/Benchmarks/CyRK_numba.pdf
diff --git a/Benchmarks/SciPy.pdf b/Benchmarks/SciPy.pdf
diff --git a/...ks/CyRK_SciPy_Compare_predprey_v0-8-0.png → ...ve/CyRK_SciPy_Compare_predprey_v0-8-0.png b/...ks/CyRK_SciPy_Compare_predprey_v0-8-0.png → ...ve/CyRK_SciPy_Compare_predprey_v0-8-0.png
diff --git a/CHANGES.md b/CHANGES.md
@@ -2,7 +2,7 @@
 
 ## 2023
 
-#### v0.8.3 (TBD)
+#### v0.8.3 (2023-10-05)
 
 New Features:
 - Added tests to check if memory access violations can occur when `CySolver` is resolved many times.
@@ -17,6 +17,7 @@ Performance:
 Other Changes:
 - Moved some common constants for both `CySolver` and `cyrk_ode` out of their files and into `cy.common`.
 - Added more meaningful memory error messages to `CySolver`.
+- Better future-proofed package structure (mainifests, gitignores, etc.).
 
 #### v0.8.2 (2023-09-25)
 

diff --git a/CyRK/cy/common.pxd b/CyRK/cy/common.pxd
@@ -18,8 +18,9 @@ cdef Py_ssize_t MAX_SIZET_SIZE
 cdef double CPU_CACHE_SIZE
 cdef double EXPECTED_SIZE_DBL
 cdef double EXPECTED_SIZE_DBLCMPLX
+cdef double MAX_ARRAY_PREALLOCATE_SIZE_DBL
+cdef double MAX_ARRAY_PREALLOCATE_SIZE_DBLCMPLX
 cdef double MIN_ARRAY_PREALLOCATE_SIZE
-cdef double MAX_ARRAY_PREALLOCATE_SIZE
 cdef double ARRAY_PREALLOC_TABS_SCALE
 cdef double ARRAY_PREALLOC_RTOL_SCALE
 

diff --git a/CyRK/cy/common.pyx b/CyRK/cy/common.pyx
@@ -20,15 +20,14 @@ cdef Py_ssize_t MAX_SIZET_SIZE = <Py_ssize_t>(0.95 * SIZE_MAX)
 cdef Py_ssize_t MAX_INT_SIZE   = <Py_ssize_t>(0.95 * INT32_MAX)
 
 # # Memory management constants
-# Assume that a cpu has a L1 of 300KB. Say that this progam will have access to 50% of that total.
-cdef double CPU_CACHE_SIZE = 0.5 * 300_000.
+# Assume that a cpu has a L1 of 300KB. Say that this progam will have access to 75% of that total.
+cdef double CPU_CACHE_SIZE = 0.75 * 300_000.
 # Number of entities we can fit into that size is based on the size of double (or double complex)
-cdef double EXPECTED_SIZE_DBL          = CPU_CACHE_SIZE / sizeof(double)
-cdef double EXPECTED_SIZE_DBLCMPLX     = CPU_CACHE_SIZE / sizeof(double complex)
-cdef double MIN_ARRAY_PREALLOCATE_SIZE = 100.
-cdef double MAX_ARRAY_PREALLOCATE_SIZE = 100_000.
+cdef double MAX_ARRAY_PREALLOCATE_SIZE_DBL      = 600_000.
+cdef double MAX_ARRAY_PREALLOCATE_SIZE_DBLCMPLX = 300_000.
+cdef double MIN_ARRAY_PREALLOCATE_SIZE = 10.
 cdef double ARRAY_PREALLOC_TABS_SCALE  = 1000.  # A delta_t_abs higher than this value will start to grow array size.
-cdef double ARRAY_PREALLOC_RTOL_SCALE  = 1.0e-6  # A rtol lower than this value will start to grow array size.
+cdef double ARRAY_PREALLOC_RTOL_SCALE  = 1.0e-5  # A rtol lower than this value will start to grow array size.
 
 
 cdef void interpolate(

diff --git a/CyRK/cy/cyrk.pyx b/CyRK/cy/cyrk.pyx
@@ -15,7 +15,7 @@ from libc.math cimport sqrt, fabs, nextafter, fmax, fmin, NAN
 from CyRK.rk.rk cimport find_rk_properties
 from CyRK.cy.common cimport double_numeric, interpolate, SAFETY, MIN_FACTOR, MAX_FACTOR, MAX_STEP, INF, EPS, \
     EPS_10, EPS_100, \
-    MAX_INT_SIZE, EXPECTED_SIZE_DBL, EXPECTED_SIZE_DBLCMPLX, MIN_ARRAY_PREALLOCATE_SIZE, MAX_ARRAY_PREALLOCATE_SIZE, \
+    MAX_INT_SIZE, MIN_ARRAY_PREALLOCATE_SIZE, MAX_ARRAY_PREALLOCATE_SIZE_DBL, MAX_ARRAY_PREALLOCATE_SIZE_DBLCMPLX, \
     ARRAY_PREALLOC_TABS_SCALE, ARRAY_PREALLOC_RTOL_SCALE
 
 
@@ -284,30 +284,33 @@ def cyrk_ode(
     # Expected size of output arrays.
     cdef Py_ssize_t expected_size_to_use, num_concats, current_size
     cdef double temp_expected_size
+    cdef double max_expected
     if expected_size == 0:
         # CySolver will attempt to guess on a best size for the arrays.
-        # Pick starting value that makes sense with the anticipated CPU cache and the data type's size.
-        if y_is_complex:
-            temp_expected_size = EXPECTED_SIZE_DBLCMPLX
-        else:
-            temp_expected_size = EXPECTED_SIZE_DBL
-        # Then there are going to be (y + num_extra) number of doubles
-        if capture_extra:
-            temp_expected_size = temp_expected_size / (y_size_dbl + num_extra)
-        else:
-            temp_expected_size = temp_expected_size / y_size_dbl
+        # Pick starting value that works with most problems
+        temp_expected_size = 500.0
         # If t_delta_abs is very large or rtol is very small, then we may need more. 
         temp_expected_size = \
             fmax(
                 temp_expected_size,
                 fmax(
-                    fmax(1., t_delta_abs / ARRAY_PREALLOC_TABS_SCALE),
-                    fmax(1., (ARRAY_PREALLOC_RTOL_SCALE / rtol_min))
+                    t_delta_abs / ARRAY_PREALLOC_TABS_SCALE,
+                    ARRAY_PREALLOC_RTOL_SCALE / rtol_min
                     )
                 )
         # Fix values that are very small/large
         temp_expected_size = fmax(temp_expected_size, MIN_ARRAY_PREALLOCATE_SIZE)
-        temp_expected_size = fmin(temp_expected_size, MAX_ARRAY_PREALLOCATE_SIZE)
+
+        if double_numeric is cython.double:
+            max_expected = MAX_ARRAY_PREALLOCATE_SIZE_DBL
+        else:
+            max_expected = MAX_ARRAY_PREALLOCATE_SIZE_DBLCMPLX
+        if capture_extra:
+            max_expected /= (y_size + num_extra)
+        else:
+            max_expected /= y_size
+
+        temp_expected_size = fmin(temp_expected_size, max_expected)
         # Store result as int
         expected_size_to_use = <Py_ssize_t>temp_expected_size
     else:

diff --git a/CyRK/cy/cysolver.pyx b/CyRK/cy/cysolver.pyx
@@ -13,7 +13,7 @@ np.import_array()
 
 from CyRK.rk.rk cimport find_rk_properties
 from CyRK.cy.common cimport interpolate, SAFETY, MIN_FACTOR, MAX_FACTOR, MAX_STEP, INF, EPS, EPS_10, EPS_100, \
-    MAX_INT_SIZE, EXPECTED_SIZE_DBL, MIN_ARRAY_PREALLOCATE_SIZE, MAX_ARRAY_PREALLOCATE_SIZE, \
+    MAX_INT_SIZE, MIN_ARRAY_PREALLOCATE_SIZE, MAX_ARRAY_PREALLOCATE_SIZE_DBL, \
     ARRAY_PREALLOC_TABS_SCALE, ARRAY_PREALLOC_RTOL_SCALE
 
 cdef (double, double) EMPTY_T_SPAN = (NAN, NAN)
@@ -414,27 +414,27 @@ cdef class CySolver:
 
         # Expected size of output arrays.
         cdef double temp_expected_size
+        cdef double max_expected
         if expected_size == 0:
             # CySolver will attempt to guess on a best size for the arrays.
-            # Pick starting value that makes sense with the anticipated CPU cache and the data type's size.
-            temp_expected_size = EXPECTED_SIZE_DBL
-            # Then there are going to be (y + num_extra) number of doubles
-            if self.capture_extra:
-                temp_expected_size = temp_expected_size / (self.y_size_dbl + self.num_extra)
-            else:
-                temp_expected_size = temp_expected_size / self.y_size_dbl
+            # Pick starting value that works with most problems
+            temp_expected_size = 500.0
             # If t_delta_abs is very large or rtol is very small, then we may need more. 
             temp_expected_size = \
             fmax(
                 temp_expected_size,
                 fmax(
-                    fmax(1., self.t_delta_abs / ARRAY_PREALLOC_TABS_SCALE),
-                    fmax(1., (ARRAY_PREALLOC_RTOL_SCALE / rtol_min))
+                    self.t_delta_abs / ARRAY_PREALLOC_TABS_SCALE,
+                    ARRAY_PREALLOC_RTOL_SCALE / rtol_min
                     )
                 )
             # Fix values that are very small/large
             temp_expected_size = fmax(temp_expected_size, MIN_ARRAY_PREALLOCATE_SIZE)
-            temp_expected_size = fmin(temp_expected_size, MAX_ARRAY_PREALLOCATE_SIZE)
+            if capture_extra:
+                max_expected = MAX_ARRAY_PREALLOCATE_SIZE_DBL / (self.y_size + self.num_extra)
+            else:
+                max_expected = MAX_ARRAY_PREALLOCATE_SIZE_DBL / self.y_size
+            temp_expected_size = fmin(temp_expected_size, max_expected)
             # Store result as int
             self.expected_size = <Py_ssize_t>temp_expected_size
         else:

diff --git a/MANIFEST.in b/MANIFEST.in
@@ -1,6 +1,2 @@
-include CyRK/array/*.pyx
-include CyRK/array/*.pxd
-include CyRK/rk/*.pyx
-include CyRK/rk/*.pxd
-include CyRK/cy/*.pyx
-include CyRK/cy/*.pxd
+global-include *.pxd
+global-include *.pyx
diff --git a/Performance/cyrk_performance-DOP853.csv b/Performance/cyrk_performance-DOP853.csv
@@ -16,3 +16,4 @@ CyRK Version, Run-on Date, cython (avg), cython (std),CySolver (avg),CySolver (s
 0.7.0a6, 28/08/2023 15:54:30, 1.0249, 0.0023, 0.1009, 0.0001, 0.1745, 0.0003, 10.0953, 0.0187, 0.5086, 0.0004, 1.6063, 0.0157, 0.4882, 0.0005, 0.0831, 0.0001, 0.1019, 0.0039, 4.6144, 0.0984, 0.3383, 0.0006, 0.7983, 0.0067, 1.4988, 0.0178, 0.1394, 0.0006, 0.2681, 0.0004, 20.1680, 0.0990, 1.2133, 0.0049, 3.4577, 0.0201, 1.7582, 0.0131, 0.1490, 0.0031, 0.5073, 0.0059, 24.8074, 0.4213, 1.2755, 0.0208, 6.5607, 0.0177
 0.7.1, 30/08/2023 13:06:08, 1.0108, 0.0108, 0.0930, 0.0004, 0.1709, 0.0004, 10.0508, 0.0336, 0.4942, 0.0005, 1.5619, 0.0081, 0.4779, 0.0045, 0.0741, 0.0001, 0.0993, 0.0009, 4.5178, 0.0105, 0.3211, 0.0018, 0.7910, 0.0070, 1.4697, 0.0163, 0.1282, 0.0002, 0.2644, 0.0022, 20.0282, 0.2074, 1.1535, 0.0006, 3.3897, 0.0012, 1.7311, 0.0088, 0.1364, 0.0027, 0.4774, 0.0019, 23.5503, 0.0176, 1.1954, 0.0037, 6.4767, 0.0802
 0.8.0, 06/09/2023 15:03:15, 0.9615, 0.0011, 0.0561, 0.0001, 0.1602, 0.0004, 9.8846, 0.1049, 0.4796, 0.0010, 1.4545, 0.0154, 0.4413, 0.0005, 0.0376, 0.0001, 0.0953, 0.0001, 4.3959, 0.0170, 0.2890, 0.0007, 0.7862, 0.0168, 1.4213, 0.0046, 0.0937, 0.0005, 0.2464, 0.0012, 19.5317, 0.0414, 1.1684, 0.0054, 3.1531, 0.0143, 1.6563, 0.0055, 0.0989, 0.0018, 0.4797, 0.0064, 22.9818, 0.1217, 1.2122, 0.0250, 6.4178, 0.0221
+0.8.3, 05/10/2023 16:59:21, 0.9504, 0.0037, 0.0551, 0.0002, 0.1624, 0.0001, 9.9624, 0.2841, 0.4721, 0.0009, 1.4638, 0.0075, 0.4286, 0.0025, 0.0381, 0.0012, 0.0953, 0.0002, 4.2439, 0.0103, 0.2896, 0.0004, 0.7662, 0.0078, 1.3670, 0.0023, 0.0912, 0.0015, 0.2437, 0.0042, 19.0453, 0.0910, 1.1357, 0.0015, 3.0867, 0.0039, 1.6214, 0.0179, 0.0931, 0.0003, 0.4685, 0.0014, 22.4041, 0.1208, 1.1700, 0.0008, 6.2697, 0.0249
diff --git a/Performance/cyrk_performance-RK23.csv b/Performance/cyrk_performance-RK23.csv
@@ -16,3 +16,4 @@ CyRK Version, Run-on Date, cython (avg), cython (std),CySolver (avg),CySolver (s
 0.7.0a6, 28/08/2023 15:52:13, 4.5831, 0.0535, 0.3040, 0.0014, 0.8701, 0.0066, 44.4575, 0.1312, 2.5780, 0.2150, 10.0452, 0.5722, 2.8942, 0.0355, 0.2581, 0.0066, 0.6178, 0.0021, 28.1092, 0.4533, 2.0150, 0.0244, 6.2654, 0.0942, 4.8829, 0.0073, 0.3678, 0.0006, 1.0106, 0.0144, 83.3562, 0.5923, 5.4591, 0.0302, 22.3938, 0.7174, 5.8152, 0.0915, 0.3946, 0.0008, 1267.8939, 2192.9856, 102.3280, 1.0445, 7.0396, 0.2830, 36.1755, 0.3725
 0.7.1, 30/08/2023 13:03:46, 4.5783, 0.0691, 0.2913, 0.0003, 0.8521, 0.0037, 44.6829, 0.8883, 2.4223, 0.0259, 9.0549, 0.2553, 2.8797, 0.0312, 0.2429, 0.0002, 0.6060, 0.0057, 27.9517, 0.2481, 1.9483, 0.0099, 5.9059, 0.0472, 4.7863, 0.0076, 0.3516, 0.0008, 0.9815, 0.0151, 81.9632, 0.6779, 5.3389, 0.0114, 20.6745, 0.4038, 5.6843, 0.0111, 0.3861, 0.0015, 1229.8100, 2127.0902, 99.4341, 0.2647, 6.5017, 0.2603, 35.6680, 0.7413
 0.8.0, 06/09/2023 15:00:58, 4.4794, 0.0786, 0.2262, 0.0003, 0.8199, 0.0021, 43.1476, 0.0854, 2.1466, 0.0018, 10.1946, 0.1933, 2.7552, 0.0105, 0.1866, 0.0004, 0.5903, 0.0024, 26.9694, 0.1524, 1.7797, 0.0257, 5.9678, 0.0812, 4.7003, 0.0324, 0.2855, 0.0005, 0.9508, 0.0055, 80.1011, 0.3480, 4.8504, 0.0219, 20.8530, 0.2461, 5.5239, 0.0051, 0.3178, 0.0014, 1270.6912, 2197.9259, 95.6354, 0.1723, 5.4006, 0.0797, 35.0990, 0.8415
+0.8.3, 05/10/2023 16:57:04, 4.4364, 0.0974, 0.2495, 0.0005, 0.8070, 0.0028, 42.8127, 0.0596, 2.4683, 0.1174, 9.0078, 0.0796, 2.6867, 0.0297, 0.1936, 0.0001, 0.5844, 0.0005, 27.1970, 0.8808, 1.8609, 0.0142, 6.0141, 0.1670, 4.6906, 0.1724, 0.3072, 0.0008, 0.9267, 0.0016, 78.5589, 0.3792, 5.1886, 0.0043, 20.6397, 0.8234, 5.4013, 0.0234, 0.3169, 0.0029, 1258.9088, 2177.5555, 94.7052, 0.3245, 5.3609, 0.0110, 34.2569, 0.6742
diff --git a/Performance/cyrk_performance-RK45.csv b/Performance/cyrk_performance-RK45.csv
@@ -16,3 +16,4 @@ CyRK Version, Run-on Date, cython (avg), cython (std),CySolver (avg),CySolver (s
 0.7.0a6, 28/08/2023 15:53:37, 1.2422, 0.0017, 0.1071, 0.0001, 0.2161, 0.0008, 12.0481, 0.0702, 0.5705, 0.0113, 2.0538, 0.0396, 0.8788, 0.0042, 0.1046, 0.0002, 0.1716, 0.0003, 8.3498, 0.0205, 0.5435, 0.0008, 1.4974, 0.0019, 1.6924, 0.0167, 0.1459, 0.0001, 0.3222, 0.0085, 24.6399, 0.0578, 1.3955, 0.0234, 4.2863, 0.0472, 1.9800, 0.0062, 0.1529, 0.0005, 0.5698, 0.0008, 29.2582, 0.4323, 1.4362, 0.0020, 8.3727, 0.1087
 0.7.1, 30/08/2023 13:05:11, 1.2473, 0.0174, 0.0985, 0.0005, 0.2110, 0.0020, 12.1557, 0.1808, 0.5338, 0.0012, 1.9231, 0.0144, 0.8757, 0.0034, 0.0959, 0.0000, 0.1693, 0.0015, 8.4225, 0.0678, 0.5282, 0.0034, 1.4849, 0.0082, 1.6579, 0.0112, 0.1361, 0.0009, 0.3019, 0.0024, 24.1373, 0.0868, 1.3549, 0.0143, 4.1917, 0.0442, 1.9731, 0.0101, 0.1435, 0.0012, 0.5592, 0.0011, 28.7442, 0.0958, 1.3859, 0.0005, 8.1105, 0.0455
 0.8.0, 06/09/2023 15:02:20, 1.2080, 0.0171, 0.0582, 0.0001, 0.2023, 0.0008, 11.8032, 0.0256, 0.4807, 0.0012, 1.8475, 0.0119, 0.8318, 0.0034, 0.0569, 0.0002, 0.1659, 0.0003, 8.1177, 0.0578, 0.4684, 0.0018, 1.4517, 0.0058, 1.6145, 0.0023, 0.0921, 0.0005, 0.2928, 0.0011, 23.6490, 0.0748, 1.2090, 0.0118, 4.0150, 0.0179, 1.8902, 0.0034, 0.0969, 0.0009, 0.5533, 0.0009, 27.6976, 0.0946, 1.2947, 0.0416, 7.9394, 0.0140
+0.8.3, 05/10/2023 16:58:27, 1.1821, 0.0195, 0.0597, 0.0001, 0.2044, 0.0013, 11.6865, 0.0342, 0.5119, 0.0012, 1.8802, 0.0457, 0.8004, 0.0018, 0.0574, 0.0001, 0.1661, 0.0008, 7.8608, 0.0145, 0.4825, 0.0004, 1.4451, 0.0078, 1.5689, 0.0075, 0.0986, 0.0009, 0.2891, 0.0002, 23.1983, 0.0209, 1.3082, 0.0034, 4.0478, 0.0406, 1.8522, 0.0163, 0.1005, 0.0002, 0.5495, 0.0016, 27.1767, 0.0465, 1.3709, 0.0245, 7.9300, 0.1021
diff --git a/README.md b/README.md
@@ -11,12 +11,12 @@
 
 ---
 
-<a href="https://github.com/jrenaud90/CyRK/releases"><img src="https://img.shields.io/badge/CyRK-0.8.1 Alpha-orange" alt="CyRK Version 0.8.1 Alpha" /></a>
+<a href="https://github.com/jrenaud90/CyRK/releases"><img src="https://img.shields.io/badge/CyRK-0.8.3 Alpha-orange" alt="CyRK Version 0.8.1 Alpha" /></a>
 
 
 **Runge-Kutta ODE Integrator Implemented in Cython and Numba**
 
-CyRK provides fast integration tools to solve systems of ODEs using an adaptive time stepping scheme. CyRK can accept differential equations that are written in pure Python, njited numba, or cython-based cdef classes. These kinds of functions are generally easier to implement than pure c functions. Using CyRK can speed up development time while not making a huge sacrifice when it comes to performance. 
+CyRK provides fast integration tools to solve systems of ODEs using an adaptive time stepping scheme. CyRK can accept differential equations that are written in pure Python, njited numba, or cython-based cdef class methods. These kinds of functions are generally easier to implement than pure c functions and can be used in existing Python software. Using CyRK can speed up development time while avoiding the slow performance that comes with using pure Python-based solvers like SciPy.
 
 The purpose of this package is to provide some 
 functionality of [scipy's solve_ivp](https://docs.scipy.org/doc/scipy/reference/generated/scipy.integrate.solve_ivp.html) with greatly improved performance.
@@ -28,7 +28,7 @@ The [cython](https://cython.org/) `CySolver` class that works with cython-based
 An additional benefit of the two cython implementations is that they are pre-compiled. This avoids most of the start-up performance hit experienced by just-in-time compilers like numba.
 
 
-<img style="text-align: center" src="https://github.com/jrenaud90/CyRK/blob/main/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-0.png" alt="CyRK Performance Graphic" />
+<img style="text-align: center" src="https://github.com/jrenaud90/CyRK/blob/main/Benchmarks/CyRK_SciPy_Compare_predprey_v0-8-3.png" alt="CyRK Performance Graphic" />
 
 ## Installation
 

diff --git a/pyproject.toml b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name='CyRK'
-version = '0.8.3a0.dev6'
+version = '0.8.3'
 description='Runge-Kutta ODE Integrator Implemented in Cython and Numba.'
 authors= [
     {name = 'Joe P. Renaud', email = '[email protected]'}