cumo 0.4.3 → 0.5.1

data/ext/cumo/narray/narray.c

@@ -324,19 +324,19 @@ cumo_na_setup(VALUE self, int ndim, size_t *shape)
 
   @example
     i = Cumo::Int64.new([2,4,3])
-    #=> Cumo::Int64#shape=[2,4,3](empty)
+    # => Cumo::Int64#shape=[2,4,3](empty)
 
     f = Cumo::DFloat.new(3,4)
-    #=> Cumo::DFloat#shape=[3,4](empty)
+    # => Cumo::DFloat#shape=[3,4](empty)
 
     f.fill(2)
-    #=> Cumo::DFloat#shape=[3,4]
+    # => Cumo::DFloat#shape=[3,4]
     # [[2, 2, 2, 2],
     #  [2, 2, 2, 2],
     #  [2, 2, 2, 2]]
 
     x = Cumo::NArray.new(5)
-    #=> in `new': allocator undefined for Cumo::NArray (TypeError)
+    # => in `new': allocator undefined for Cumo::NArray (TypeError)
     #   	from t.rb:9:in `<main>'
 
 */
@@ -420,10 +420,10 @@ cumo_na_initialize_copy(VALUE self, VALUE orig)
  *  but for typed NArray subclasses, e.g., DFloat, Int64.
  *  @example
  *    a = Cumo::DFloat.zeros(3,5)
- *    => Cumo::DFloat#shape=[3,5]
- *    [[0, 0, 0, 0, 0],
- *     [0, 0, 0, 0, 0],
- *     [0, 0, 0, 0, 0]]
+ *    # => Cumo::DFloat#shape=[3,5]
+ *    # [[0, 0, 0, 0, 0],
+ *    #  [0, 0, 0, 0, 0],
+ *    #  [0, 0, 0, 0, 0]]
  */
 static VALUE
 cumo_na_s_zeros(int argc, VALUE *argv, VALUE klass)
@@ -471,8 +471,8 @@ cumo_na_s_ones(int argc, VALUE *argv, VALUE klass)
 
   @example
     a = Cumo::DFloat.linspace(-5,5,7)
-    => Cumo::DFloat#shape=[7]
-    [-5, -3.33333, -1.66667, 0, 1.66667, 3.33333, 5]
+    # => Cumo::DFloat#shape=[7]
+    # [-5, -3.33333, -1.66667, 0, 1.66667, 3.33333, 5]
  */
 static VALUE
 cumo_na_s_linspace(int argc, VALUE *argv, VALUE klass)
@@ -510,11 +510,12 @@ cumo_na_s_linspace(int argc, VALUE *argv, VALUE klass)
 
   @example
     Cumo::DFloat.logspace(4,0,5,2)
-    => Cumo::DFloat#shape=[5]
-       [16, 8, 4, 2, 1]
+    # => Cumo::DFloat#shape=[5]
+    #    [16, 8, 4, 2, 1]
+
     Cumo::DComplex.logspace(0,1i*Math::PI,5,Math::E)
-    => Cumo::DComplex#shape=[5]
-       [1+4.44659e-323i, 0.707107+0.707107i, 6.12323e-17+1i, -0.707107+0.707107i, ...]
+    # => Cumo::DComplex#shape=[5]
+    #    [1+4.44659e-323i, 0.707107+0.707107i, 6.12323e-17+1i, -0.707107+0.707107i, ...]
  */
 static VALUE
 cumo_na_s_logspace(int argc, VALUE *argv, VALUE klass)
@@ -548,10 +549,10 @@ cumo_na_s_logspace(int argc, VALUE *argv, VALUE klass)
   @return [Cumo::NArray]  created NArray.
   @example
     a = Cumo::DFloat.eye(3)
-    => Cumo::DFloat#shape=[3,3]
-    [[1, 0, 0],
-     [0, 1, 0],
-     [0, 0, 1]]
+    # => Cumo::DFloat#shape=[3,3]
+    # [[1, 0, 0],
+    #  [0, 1, 0],
+    #  [0, 0, 1]]
 */
 static VALUE
 cumo_na_s_eye(int argc, VALUE *argv, VALUE klass)
@@ -889,6 +890,39 @@ cumo_na_check_contiguous(VALUE self)
     return Qfalse;
 }
 
+VALUE
+cumo_na_check_fortran_contiguous(VALUE self)
+{
+    int i;
+    ssize_t st0;
+    cumo_narray_t *na;
+
+    switch(CUMO_RNARRAY_TYPE(self)) {
+    case CUMO_NARRAY_DATA_T:
+    case CUMO_NARRAY_FILEMAP_T:
+        return Qfalse;
+    case CUMO_NARRAY_VIEW_T:
+        CumoGetNArray(self,na);
+
+        // not contiguous if it has index
+        for (i=0; i < CUMO_NA_NDIM(na); i++) {
+            if (CUMO_NA_IS_INDEX_AT(na,i))
+                return Qfalse;
+        }
+
+        // check f-contiguous
+        st0 = cumo_na_element_stride(self); // elmsz
+        for (i=0; i < CUMO_NA_NDIM(na); i++) {
+            if (CUMO_NA_SHAPE(na)[i] == 1)
+                continue;
+            if (CUMO_NA_STRIDE_AT(na, i) != st0)
+                return Qfalse;
+            st0 *= CUMO_NA_SHAPE(na)[i];
+        }
+    }
+    return Qtrue;
+}
+
 VALUE
 cumo_na_as_contiguous_array(VALUE a)
 {
@@ -1388,7 +1422,7 @@ static VALUE cumo_na_inplace( VALUE self );
 /*
   Load marshal data.
   @overload marshal_load(data)
-  @params [Array] Array containing marshal data.
+  @param [Array] Array containing marshal data.
   @return [nil]
  */
 static VALUE
@@ -1833,6 +1867,9 @@ cumo_na_equal(VALUE self, volatile VALUE other)
             return Qfalse;
         }
     }
+    if (na1->size == 0) {
+      return Qtrue;
+    }
     vbool = rb_funcall(self, cumo_id_eq, 1, other);
     return (rb_funcall(vbool, cumo_id_count_false_cpu, 0)==INT2FIX(0)) ? Qtrue : Qfalse;
 }
@@ -1929,6 +1966,7 @@ Init_cumo_narray()
     rb_define_method(cNArray, "debug_info", cumo_na_debug_info, 0);
 
     rb_define_method(cNArray, "contiguous?", cumo_na_check_contiguous, 0);
+    rb_define_method(cNArray, "fortran_contiguous?", cumo_na_check_fortran_contiguous, 0);
 
     rb_define_method(cNArray, "view", cumo_na_make_view, 0);
     rb_define_method(cNArray, "expand_dims", cumo_na_expand_dims, 1);

data/ext/cumo/narray/ndloop.c

@@ -56,7 +56,7 @@ typedef struct CUMO_NA_MD_LOOP {
                               // [2,3,4], 111b for sum(), 010b for sum(axis: 1), 110b for sum(axis: [1,2])
     VALUE  loop_opt;
     cumo_ndfunc_t  *ndfunc;
-    void (*loop_func)();
+    void (*loop_func)(cumo_ndfunc_t *, struct CUMO_NA_MD_LOOP *);
 } cumo_na_md_loop_t;
 
 #define LARG(lp,iarg) ((lp)->user.args[iarg])

data/ext/cumo/narray/struct.c

@@ -812,7 +812,7 @@ nst_s_add_type(int argc, VALUE *argv, VALUE mod)
 
 #define NST_TYPEDEF(tpname,tpclass)                 \
 static VALUE                                        \
-nst_s_##tpname(VALUE argc, VALUE *argv, VALUE mod)  \
+nst_s_##tpname(int argc, VALUE *argv, VALUE mod)  \
 {   nstruct_add_type(tpclass,argc,argv,mod);        \
     return Qnil;                                    \
 }

data/lib/cumo/cuda/compile_error.rb

@@ -1,4 +1,4 @@
-require_relative '../cuda'
+# frozen_string_literal: true
 
 module Cumo::CUDA
   class CompileError < StandardError

data/lib/cumo/cuda/compiler.rb

@@ -1,32 +1,33 @@
+# frozen_string_literal: true
+
 require 'tmpdir'
 require 'tempfile'
 require 'fileutils'
 require 'digest/md5'
-require_relative '../cuda'
 
 module Cumo::CUDA
   class Compiler
     VALID_KERNEL_NAME = /\A[a-zA-Z_][a-zA-Z_0-9]*\z/
     DEFAULT_CACHE_DIR = File.expand_path('~/.cumo/kernel_cache')
-  
+
     @@empty_file_preprocess_cache ||= {}
-    
+
     def self.valid_kernel_name?(name)
       VALID_KERNEL_NAME.match?(name)
     end
-  
+
     def compile_using_nvrtc(source, options: [], arch: nil)
       arch ||= get_arch
       options += ["-arch=#{arch}"]
-    
+
       Dir.mktmpdir do |root_dir|
         path = File.join(root_dir, 'kern')
         cu_path = "#{path}.cu"
-    
+
         File.open(cu_path, 'w') do |cu_file|
           cu_file.write(source)
         end
-   
+
         prog = NVRTCProgram.new(source, name: cu_path)
         begin
           ptx = prog.compile(options: options)
@@ -41,12 +42,12 @@ module Cumo::CUDA
         return ptx
       end
     end
-    
+
     def compile_with_cache(source, options: [], arch: nil, cache_dir: nil, extra_source: nil)
       # NVRTC does not use extra_source. extra_source is used for cache key.
       cache_dir ||= get_cache_dir
       arch ||= get_arch
-    
+
       options += ['-ftz=true']
 
       env = [arch, options, get_nvrtc_version]
@@ -57,15 +58,15 @@ module Cumo::CUDA
         @@empty_file_preprocess_cache[env] = base
       end
       key_src = "#{env} #{base} #{source} #{extra_source}"
-    
+
       key_src.encode!('utf-8')
       digest = Digest::MD5.hexdigest(key_src)
       name = "#{digest}_2.cubin"
-    
+
       unless Dir.exist?(cache_dir)
         FileUtils.mkdir_p(cache_dir)
       end
-   
+
       # TODO(sonots): thread-safe?
       path = File.join(cache_dir, name)
       cubin = load_cache(path)
@@ -74,7 +75,7 @@ module Cumo::CUDA
         mod.load(cubin)
         return mod
       end
-    
+
       ptx = compile_using_nvrtc(source, options: options, arch: arch)
       cubin = nil
       cubin_hash = nil
@@ -85,7 +86,7 @@ module Cumo::CUDA
       end
 
       save_cache(path, cubin_hash, cubin)
-    
+
       # Save .cu source file along with .cubin
       if get_bool_env_variable('CUMO_CACHE_SAVE_CUDA_SOURCE', false)
         File.open("#{path}.cu", 'w') do |f|
@@ -97,7 +98,7 @@ module Cumo::CUDA
       mod.load(cubin)
       return mod
     end
-  
+
     private
 
     def save_cache(path, cubin_hash, cubin)
@@ -105,7 +106,7 @@ module Cumo::CUDA
       tf.write(cubin_hash)
       tf.write(cubin)
       temp_path = tf.path
-      File.rename(temp_path, path)
+      FileUtils.mv(temp_path, path)
     end
 
     def load_cache(path)
@@ -121,29 +122,29 @@ module Cumo::CUDA
       end
       nil
     end
-  
+
     def get_cache_dir
       ENV.fetch('CUMO_CACHE_DIR', DEFAULT_CACHE_DIR)
     end
-  
+
     def get_nvrtc_version
       @@nvrtc_version ||= NVRTC.nvrtcVersion
     end
-    
+
     def get_arch
       cc = Device.new.compute_capability
       "compute_#{cc}"
     end
-    
+
     def get_bool_env_variable(name, default)
       val = ENV[name]
       return default if val.nil? or val.size == 0
       Integer(val) == 1 rescue false
     end
-  
+
     def preprocess(source, options, arch)
       options += ["-arch=#{arch}"]
-    
+
       prog = NVRTCProgram.new(source, name: '')
       begin
         result = prog.compile(options: options)

data/lib/cumo/cuda/cudnn.rb

@@ -1,4 +1,4 @@
-require 'cumo'
+# frozen_string_literal: true
 
 module Cumo
   [SFloat, DFloat].each do |klass|

data/lib/cumo/cuda/device.rb

@@ -1,4 +1,4 @@
-require_relative '../cuda'
+# frozen_string_literal: true
 
 module Cumo::CUDA
   class Device

data/lib/cumo/cuda/link_state.rb

@@ -1,4 +1,4 @@
-require_relative '../cuda'
+# frozen_string_literal: true
 
 module Cumo::CUDA
   # CUDA link state.
@@ -25,7 +25,7 @@ module Cumo::CUDA
     end
 
     def complete
-      cubin = Driver.cuLinkComplete(@ptr)
+      Driver.cuLinkComplete(@ptr)
     end
   end
 end

data/lib/cumo/cuda/module.rb

@@ -1,4 +1,4 @@
-require_relative '../cuda'
+# frozen_string_literal: true
 
 module Cumo::CUDA
   # CUDA kernel module.
@@ -37,4 +37,3 @@ module Cumo::CUDA
     end
   end
 end
-

data/lib/cumo/cuda/nvrtc_program.rb

@@ -1,4 +1,5 @@
-require_relative '../cuda'
+# frozen_string_literal: true
+
 require_relative 'compile_error'
 
 module Cumo::CUDA
@@ -18,7 +19,7 @@ module Cumo::CUDA
       begin
         NVRTC.nvrtcCompileProgram(@ptr, options)
         return NVRTC.nvrtcGetPTX(@ptr)
-      rescue NVRTCError => e
+      rescue NVRTCError
         log = NVRTC.nvrtcGetProgramLog(@ptr)
         raise CompileError.new(log, @src, @name, options)
       end

data/lib/cumo/cuda.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 module Cumo
   module CUDA
   end

data/lib/cumo/linalg.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'cumo'
 
 # Provide compatibility layers with numo/linalg