Fix 3rd

.gitignore

@@ -13,4 +13,3 @@ dist/
 Makefile
 !docs/Makefile
 version.py
-setup.py

.gitmodules

@@ -6,4 +6,4 @@
 	url = https://github.com/gabime/spdlog.git
 [submodule "3rd/eigen3"]
 	path = 3rd/eigen3
-	url = https://github.com/eigenteam/eigen-git-mirror.git
+	url = https://gitlab.com/libeigen/eigen.git

MANIFEST.in

@@ -0,0 +1,9 @@
+include install/cuda_setup.py
+include 3rd/json11/json11.hpp
+include 3rd/json11/json11.cpp
+recursive-include 3rd/eigen3/Eigen/ *
+recursive-include 3rd/eigen3/unsupported/Eigen/ *
+recursive-include 3rd/spdlog/include/spdlog/ *
+recursive-include buffalo/ *.py *.pyx *.hpp
+recursive-include include/ *.hpp
+recursive-include lib/ *.cc *.cu

buffalo/algo/_als.pyx

@@ -1,21 +1,18 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool
 from libcpp.pair cimport pair
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/algo_impl/als/als.hpp" namespace "als":
     cdef cppclass CALS:
         void release() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void initialize_model(float*, int,
                               float*, int) nogil except +
         void precompute(int) nogil except +
@@ -51,8 +48,6 @@ cdef class CyALS:
     def precompute(self, axis):
         self.obj.precompute(axis)
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update(self,
                        int start_x,
                        int next_x,

buffalo/algo/_bpr.pyx

@@ -1,20 +1,17 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/algo_impl/bpr/bpr.hpp" namespace "bpr":
     cdef cppclass CBPRMF:
         void release() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void initialize_model(float*, int32_t,
                               float*, int32_t,
                               float*,
@@ -66,8 +63,6 @@ cdef class CyBPRMF:
                              int cum_table_size):
         self.obj.set_cumulative_table(&cum_table[0], cum_table_size)
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def add_jobs(self,
                  int start_x,
                  int next_x,

buffalo/algo/_cfr.pyx

@@ -1,20 +1,16 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool as bool_t
 from libcpp.string cimport string
-from libcpp.vector cimport vector
-
-import numpy as np
 
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/algo_impl/cfr/cfr.hpp" namespace "cfr":
     cdef cppclass CCFR:
-        bool_t init(string) nogil except +
+        bint init(string) nogil except +
         void set_embedding(float*, int, string) nogil except +
         void precompute(string) nogil except +
         double partial_update_user(int, int,
@@ -46,17 +42,13 @@ cdef class CyCFR:
                       np.ndarray[np.float32_t, ndim=2] F, obj_type):
         self.obj.set_embedding(&F[0, 0], F.shape[0], obj_type)
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update_user(self, int start_x, int next_x,
                             np.ndarray[np.int64_t, ndim=1] indptrs,
                             np.ndarray[np.int32_t, ndim=1] keys,
                             np.ndarray[np.float32_t, ndim=1] vals):
         return self.obj.partial_update_user(start_x, next_x,
                                             &indptrs[0], &keys[0], &vals[0])
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update_item(self, int start_x, int next_x,
                             np.ndarray[np.int64_t, ndim=1] indptrs_u,
                             np.ndarray[np.int32_t, ndim=1] keys_u,
@@ -68,8 +60,6 @@ cdef class CyCFR:
                                             &indptrs_u[0], &keys_u[0], &vals_u[0],
                                             &indptrs_c[0], &keys_c[0], &vals_c[0])
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update_context(self, int start_x, int next_x,
                                np.ndarray[np.int64_t, ndim=1] indptrs,
                                np.ndarray[np.int32_t, ndim=1] keys,

buffalo/algo/_plsi.pyx

@@ -1,17 +1,16 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
-import numpy as np
 
 cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool as bool_t
 from libcpp.string cimport string
 
+np.import_array()
+
 
 cdef extern from "buffalo/algo_impl/plsi/plsi.hpp" namespace "plsi":
     cdef cppclass CPLSI:
-        bool_t init(string) nogil except +
+        bint init(string) nogil except +
         void release() nogil except +
         void swap() nogil except +
         void reset() nogil except +
@@ -51,8 +50,6 @@ cdef class CyPLSI:
     def normalize(self, alpha1, alpha2):
         self.obj.normalize(alpha1, alpha2)
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update(self, int start_x, int next_x,
                        np.ndarray[np.int64_t, ndim=1] indptr,
                        np.ndarray[np.int32_t, ndim=1] keys,

buffalo/algo/_w2v.pyx

@@ -1,20 +1,17 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t, uint32_t
-from libcpp cimport bool
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/algo_impl/w2v/w2v.hpp" namespace "w2v":
     cdef cppclass CW2V:
         void release() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void initialize_model(float*, int32_t,
                               int32_t*,
                               uint32_t*,
@@ -58,8 +55,6 @@ cdef class CyW2V:
     def launch_workers(self):
         self.obj.launch_workers()
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def add_jobs(self,
                  int start_x,
                  int next_x,

buffalo/algo/_warp.pyx

@@ -1,20 +1,17 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/algo_impl/warp/warp.hpp" namespace "warp":
     cdef cppclass CWARP:
         void release() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void initialize_model(float*, int32_t,
                               float*, int32_t,
                               float*,
@@ -66,8 +63,6 @@ cdef class CyWARP:
     def launch_workers(self):
         self.obj.launch_workers()
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def add_jobs(self,
                  int start_x,
                  int next_x,

buffalo/algo/cuda/_als.pyx

@@ -1,21 +1,18 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool
 from libcpp.pair cimport pair
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/cuda/als/als.hpp" namespace "cuda_als":
     cdef cppclass CuALS:
         CuALS() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void set_placeholder(int64_t* lindptr, int64_t* rindptr, size_t batch_size)
         void initialize_model(float*, int,
                               float*, int) nogil except +
@@ -46,7 +43,7 @@ cdef class CyALS:
 
     def set_placeholder(self, np.ndarray[np.int64_t, ndim=1] lindptr,
                         np.ndarray[np.int64_t, ndim=1] rindptr,
-                        batch_size):
+                        size_t batch_size):
         self.obj.set_placeholder(&lindptr[0], &rindptr[0], batch_size)
 
     def precompute(self, axis):
@@ -55,8 +52,6 @@ cdef class CyALS:
     def get_vdim(self):
         return self.obj.get_vdim()
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def partial_update(self,
                        int start_x,
                        int next_x,

buffalo/algo/cuda/_bpr.pyx

@@ -1,30 +1,27 @@
-# cython: experimental_cpp_class_def=True, language_level=3
+# cython: language_level=3, boundscheck=False, wraparound=False
 # distutils: language=c++
-import cython
 
+cimport numpy as np
 from libc.stdint cimport int32_t, int64_t
-from libcpp cimport bool
 from libcpp.pair cimport pair
 from libcpp.string cimport string
 
-import numpy as np
-
-cimport numpy as np
+np.import_array()
 
 
 cdef extern from "buffalo/cuda/bpr/bpr.hpp" namespace "cuda_bpr":
     cdef cppclass CuBPR:
         CuBPR() nogil except +
-        bool init(string) nogil except +
+        bint init(string) nogil except +
         void set_placeholder(int64_t* indptr, size_t batch_size) nogil except +
         void set_cumulative_table(int64_t*) nogil except +
         void initialize_model(float*, int,
-                              float*, float*, int, int64_t, bool) nogil except +
+                              float*, float*, int, int64_t, bint) nogil except +
         pair[double, double] partial_update(int, int,
                                             int64_t*, int32_t*) nogil except +
         double compute_loss(int, int32_t*, int32_t*, int32_t*) nogil except +
         int get_vdim() nogil except +
-        void synchronize(bool)
+        void synchronize(bint)
 
 
 cdef class CyBPR:
@@ -66,8 +63,6 @@ cdef class CyBPR:
     def wait_until_done(self):
         return
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def add_jobs(self,
                  int start_x,
                  int next_x,
@@ -78,8 +73,6 @@ cdef class CyBPR:
                                        &indptr[0],
                                        &keys[0])
 
-    @cython.boundscheck(False)
-    @cython.wraparound(False)
     def compute_loss(self,
                      np.ndarray[np.int32_t, ndim=1] user,
                      np.ndarray[np.int32_t, ndim=1] pos,