Issue 179 compilation speedup PIMPL

brian2cuda/binomial.py

@@ -4,11 +4,17 @@
 import numpy as np
 
 from brian2.core.functions import DEFAULT_FUNCTIONS
-from brian2.units.fundamentalunits import check_units
 from brian2.utils.stringtools import replace
 from brian2.input.binomial import (BinomialFunction, _pre_calc_constants,
                                    _pre_calc_constants_approximated)
 from brian2 import prefs
+#ths can be removed when brian2 is updated to forward compiler_kwds from target generators
+_CUDA_BINOMIAL_COMPILER_KWDS = {
+    "headers": [
+        '"rand.h"',
+        "<curand_kernel.h>",
+    ],
+}
 
 
 def _generate_cuda_code(n, p, use_normal, name):
@@ -103,3 +109,15 @@ def _generate_cuda_code(n, p, use_normal, name):
 
 
 BinomialFunction.implementations['cuda'] = _generate_cuda_code
+
+_original_binomial_init = BinomialFunction.__init__
+
+
+def _binomial_init_with_cuda_headers(self, n, p, approximate=True, name="_binomial*"):
+    _original_binomial_init(self, n, p, approximate=approximate, name=name)
+    cuda_impl = self.implementations._implementations.get("cuda")
+    if cuda_impl is not None:
+        cuda_impl.compiler_kwds = dict(_CUDA_BINOMIAL_COMPILER_KWDS)
+
+
+BinomialFunction.__init__ = _binomial_init_with_cuda_headers

brian2cuda/brianlib/clocks.h

@@ -1,7 +1,6 @@
 #ifndef _BRIAN_CLOCKS_H
 #define _BRIAN_CLOCKS_H
 #include<stdlib.h>
-#include<iostream>
 #include<algorithm>
 #include<brianlib/stdint_compat.h>
 #include<math.h>
@@ -75,4 +74,4 @@ class EventClock : public BaseClock
 	}
 };
 
-#endif
+#endif

brian2cuda/brianlib/cuda_utils.h

@@ -1,9 +1,11 @@
 #ifndef BRIAN2CUDA_ERROR_CHECK_H
 #define BRIAN2CUDA_ERROR_CHECK_H
 #include <stdio.h>
-#include <thrust/system_error.h>
-#include "objects.h"
-#include "curand.h"
+#include <cuda_runtime.h>
+
+namespace brian{
+  extern size_t used_device_memory;
+}
 
 // Define this to turn on error checking
 #define BRIAN2CUDA_ERROR_CHECK
@@ -23,9 +25,9 @@
 #define THRUST_CHECK_ERROR(code)  { try {code;} \
     catch(...) {_thrustCheckError(__FILE__, __LINE__, #code);} }
 
-
+#ifdef BRIAN2CUDA_CURAND_HOST
+#include <curand.h>
 // adapted from NVIDIA cuda samples, shipped with cuda 10.1 (common/inc/helper_cuda.h)
-#ifdef CURAND_H_
 // cuRAND API errors
 static const char *_curandGetErrorEnum(curandStatus_t error) {
   switch (error) {
@@ -71,39 +73,37 @@ static const char *_curandGetErrorEnum(curandStatus_t error) {
 
   return "<unknown>";
 }
-#endif
-
 
-inline void _cudaSafeCall(cudaError err, const char *file, const int line, const char *call = "")
+inline void _cudaSafeCall(curandStatus_t err, const char *file, const int line, const char *call = "")
 {
 #ifdef BRIAN2CUDA_ERROR_CHECK
-    if (cudaSuccess != err)
+    if (CURAND_STATUS_SUCCESS != err)
     {
         fprintf(stderr, "ERROR: %s failed at %s:%i : %s\n",
-                call, file, line, cudaGetErrorString(err));
+                call, file, line, _curandGetErrorEnum(err));
         exit(-1);
     }
 #endif
 
     return;
 }
+#endif // BRIAN2CUDA_CURAND_HOST
 
 
-inline void _cudaSafeCall(curandStatus_t err, const char *file, const int line, const char *call = "")
+inline void _cudaSafeCall(cudaError err, const char *file, const int line, const char *call = "")
 {
 #ifdef BRIAN2CUDA_ERROR_CHECK
-    if (CURAND_STATUS_SUCCESS != err)
+    if (cudaSuccess != err)
     {
         fprintf(stderr, "ERROR: %s failed at %s:%i : %s\n",
-                call, file, line, _curandGetErrorEnum(err));
+                call, file, line, cudaGetErrorString(err));
         exit(-1);
     }
 #endif
 
     return;
 }
 
-
 inline void _cudaCheckError(const char *file, const int line, const char *msg)
 {
 #ifdef BRIAN2CUDA_ERROR_CHECK_BLOCKING
@@ -176,5 +176,4 @@ inline void _thrustCheckError(const char *file, const int line,
             code, file, line);
     throw;
 }
-
 #endif  // BRIAN2CUDA_ERROR_CHECK_H

brian2cuda/brianlib/spikequeue.h

@@ -1,4 +1,3 @@
-#include<iostream>
 #include<vector>
 #include<algorithm>
 #include<inttypes.h>

brian2cuda/codeobject.py

@@ -95,15 +95,19 @@ class CUDAStandaloneAtomicsCodeObject(CUDAStandaloneCodeObject):
 rand_code = '''
     #define _rand(vectorisation_idx) (_ptr_array_%CODEOBJ_NAME%_rand[vectorisation_idx])
     '''
+
+_RAND_HEADERS = {"headers": ['"rand.h"']}
 rand_impls = DEFAULT_FUNCTIONS['rand'].implementations
 rand_impls.add_implementation(CUDAStandaloneCodeObject,
                               code=rand_code,
-                              name='_rand')
+                              name='_rand',
+                              compiler_kwds=_RAND_HEADERS)
 
 randn_code = '''
     #define _randn(vectorisation_idx) (_ptr_array_%CODEOBJ_NAME%_randn[vectorisation_idx])
         '''
 randn_impls = DEFAULT_FUNCTIONS['randn'].implementations
 randn_impls.add_implementation(CUDAStandaloneCodeObject,
                                code=randn_code,
-                               name='_randn')
+                               name='_randn',
+                               compiler_kwds=_RAND_HEADERS)

brian2cuda/cuda_generator.py

@@ -989,7 +989,13 @@ class CUDAAtomicsCodeGenerator(CUDACodeGenerator):
     CUDACodeGenerator,
     code=poisson_code,
     name='_poisson',
-    compiler_kwds={"headers": ["<curand.h>"]}
+    compiler_kwds={
+        "headers": [
+            "<curand.h>",
+            "<curand_kernel.h>",
+            '"rand.h"',
+        ],
+    }
 )