cumo 0.4.3 → 0.5.0

data/bench/cumo_bench.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'cumo/narray'
 require 'benchmark'
 
@@ -5,7 +7,7 @@ NUM = (ARGV.first || 100).to_i
 
 # warm up
 a = Cumo::Float32.new(10).seq(1)
-b = Cumo::Float32.new(10).seq(10,10)
+b = Cumo::Float32.new(10).seq(10, 10)
 c = a + b
 c.free
 
@@ -14,7 +16,7 @@ def elementwise(num = nil)
   puts "elementwise(#{num})"
   Benchmark.bm do |r|
     a = Cumo::Float32.new(10000).seq(1)
-    b = Cumo::Float32.new(10000).seq(10,10)
+    b = Cumo::Float32.new(10000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**4') do
@@ -25,7 +27,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(100000).seq(1)
-    b = Cumo::Float32.new(100000).seq(10,10)
+    b = Cumo::Float32.new(100000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**5') do
@@ -36,7 +38,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(1000000).seq(1)
-    b = Cumo::Float32.new(1000000).seq(10,10)
+    b = Cumo::Float32.new(1000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**6') do
@@ -47,7 +49,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(10000000).seq(1)
-    b = Cumo::Float32.new(10000000).seq(10,10)
+    b = Cumo::Float32.new(10000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**7') do
@@ -58,7 +60,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(100000000).seq(1)
-    b = Cumo::Float32.new(100000000).seq(10,10)
+    b = Cumo::Float32.new(100000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**8') do
@@ -130,8 +132,8 @@ def dot(num = nil)
   num ||= 1
   puts "dot(#{num})"
   Benchmark.bm do |r|
-    a = Cumo::Float32.new(100,100).seq(1)
-    b = Cumo::Float32.new(100,100).seq(10,10)
+    a = Cumo::Float32.new(100, 100).seq(1)
+    b = Cumo::Float32.new(100, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**4') do
@@ -141,8 +143,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,1000).seq(1)
-    b = Cumo::Float32.new(1000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 1000).seq(1)
+    b = Cumo::Float32.new(1000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**5') do
@@ -152,8 +154,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,10000).seq(1)
-    b = Cumo::Float32.new(10000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 10000).seq(1)
+    b = Cumo::Float32.new(10000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**6') do
@@ -163,8 +165,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,100000).seq(1)
-    b = Cumo::Float32.new(100000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 100000).seq(1)
+    b = Cumo::Float32.new(100000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**7') do
@@ -174,8 +176,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,1000000).seq(1)
-    b = Cumo::Float32.new(1000000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 1000000).seq(1)
+    b = Cumo::Float32.new(1000000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**8') do

data/bench/numo_bench.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'numo/narray'
 require 'benchmark'
 
@@ -5,7 +7,7 @@ NUM = (ARGV.first || 100).to_i
 
 # warm up
 a = Numo::Float32.new(10).seq(1)
-b = Numo::Float32.new(10).seq(10,10)
+b = Numo::Float32.new(10).seq(10, 10)
 c = a + b
 
 def elementwise(num = nil)
@@ -13,31 +15,31 @@ def elementwise(num = nil)
   puts "elementwise(#{num})"
   Benchmark.bm do |r|
     a = Numo::Float32.new(10000).seq(1)
-    b = Numo::Float32.new(10000).seq(10,10)
+    b = Numo::Float32.new(10000).seq(10, 10)
     r.report('10**4') do
       NUM.times { (a + b) }
     end
 
     a = Numo::Float32.new(100000).seq(1)
-    b = Numo::Float32.new(100000).seq(10,10)
+    b = Numo::Float32.new(100000).seq(10, 10)
     r.report('10**5') do
       NUM.times { (a + b) }
     end
 
     a = Numo::Float32.new(1000000).seq(1)
-    b = Numo::Float32.new(1000000).seq(10,10)
+    b = Numo::Float32.new(1000000).seq(10, 10)
     r.report('10**6') do
       NUM.times { (a + b) }
     end
 
     a = Numo::Float32.new(10000000).seq(1)
-    b = Numo::Float32.new(10000000).seq(10,10)
+    b = Numo::Float32.new(10000000).seq(10, 10)
     r.report('10**7') do
       NUM.times { (a + b) }
     end
 
     a = Numo::Float32.new(100000000).seq(1)
-    b = Numo::Float32.new(100000000).seq(10,10)
+    b = Numo::Float32.new(100000000).seq(10, 10)
     r.report('10**8') do
       NUM.times { (a + b) }
     end
@@ -79,32 +81,32 @@ def dot(num = nil)
   num ||= 1
   puts "dot(#{num})"
   Benchmark.bm do |r|
-    a = Numo::Float32.new(100,100).seq(1)
-    b = Numo::Float32.new(100,100).seq(10,10)
+    a = Numo::Float32.new(100, 100).seq(1)
+    b = Numo::Float32.new(100, 100).seq(10, 10)
     r.report('10**4') do
       num.times { a.dot(b) }
     end
 
-    a = Numo::Float32.new(100,1000).seq(1)
-    b = Numo::Float32.new(1000,100).seq(10,10)
+    a = Numo::Float32.new(100, 1000).seq(1)
+    b = Numo::Float32.new(1000, 100).seq(10, 10)
     r.report('10**5') do
       num.times { a.dot(b) }
     end
 
-    a = Numo::Float32.new(100,10000).seq(1)
-    b = Numo::Float32.new(10000,100).seq(10,10)
+    a = Numo::Float32.new(100, 10000).seq(1)
+    b = Numo::Float32.new(10000, 100).seq(10, 10)
     r.report('10**6') do
       num.times { a.dot(b) }
     end
 
-    a = Numo::Float32.new(100,100000).seq(1)
-    b = Numo::Float32.new(100000,100).seq(10,10)
+    a = Numo::Float32.new(100, 100000).seq(1)
+    b = Numo::Float32.new(100000, 100).seq(10, 10)
     r.report('10**7') do
       num.times { a.dot(b) }
     end
 
-    a = Numo::Float32.new(100,1000000).seq(1)
-    b = Numo::Float32.new(1000000,100).seq(10,10)
+    a = Numo::Float32.new(100, 1000000).seq(1)
+    b = Numo::Float32.new(1000000, 100).seq(10, 10)
     r.report('10**8') do
       num.times { a.dot(b) }
     end

data/bench/reduction_fp32.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'benchmark'
 require 'cumo/narray'
 
@@ -5,7 +7,7 @@ num_iteration = 100
 Cumo::CUDA::Runtime.cudaDeviceSynchronize
 
 Benchmark.bm 30 do |r|
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum" do
     num_iteration.times do
       x.sum
@@ -13,7 +15,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 0)" do
     num_iteration.times do
       x.sum(axis: 0)
@@ -21,7 +23,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 1)" do
     num_iteration.times do
       x.sum(axis: 1)
@@ -29,7 +31,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(keepdims: true)" do
     num_iteration.times do
       x.sum(keepdims: true)
@@ -37,7 +39,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 0, keepdims: true)" do
     num_iteration.times do
       x.sum(axis: 0, keepdims: true)
@@ -45,7 +47,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 1, keepdims: true)" do
     num_iteration.times do
       x.sum(axis: 1, keepdims: true)
@@ -53,7 +55,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max" do
     num_iteration.times do
       x.max
@@ -61,7 +63,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 0)" do
     num_iteration.times do
       x.max(axis: 0)
@@ -69,7 +71,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 1)" do
     num_iteration.times do
       x.max(axis: 1)
@@ -77,7 +79,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(keepdims: true)" do
     num_iteration.times do
       x.max(keepdims: true)
@@ -85,7 +87,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 0, keepdims: true)" do
     num_iteration.times do
       x.max(axis: 0, keepdims: true)
@@ -93,7 +95,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 1, keepdims: true)" do
     num_iteration.times do
       x.max(axis: 1, keepdims: true)

data/bin/console

@@ -1,4 +1,5 @@
 #!/usr/bin/env ruby
+# frozen_string_literal: true
 
 require "bundler/setup"
 require "cumo"

data/cumo.gemspec

@@ -1,9 +1,9 @@
-# coding: utf-8
+# frozen_string_literal: true
+
 lib = File.expand_path("../lib", __FILE__)
 $LOAD_PATH.unshift(lib) unless $LOAD_PATH.include?(lib)
 
 cumo_version = File.read(File.join(__dir__, "ext/cumo/include/cumo.h")).match(/CUMO_VERSION "([^"]+)"/)[1]
-numo_narray_version = File.read(File.join(__dir__, "numo-narray-version")).strip
 
 Gem::Specification.new do |spec|
   spec.name          = "cumo"
@@ -16,7 +16,9 @@ Gem::Specification.new do |spec|
   spec.homepage      = "https://github.com/sonots/cumo"
   spec.license       = "BSD-3-Clause"
 
-  spec.files         = `git ls-files -z`.split("\x0").reject do |f|
+  spec.required_ruby_version = ">= 3.0.0"
+
+  spec.files = `git ls-files -z`.split("\x0").reject do |f|
     f.match(%r{^(test|spec|features)/})
   end
   spec.test_files    = `git ls-files -- {test,spec,features}/*`.split("\n")
@@ -24,9 +26,4 @@ Gem::Specification.new do |spec|
   spec.executables   = spec.files.grep(%r{^exe/}) { |f| File.basename(f) }
   spec.require_paths = ["lib"]
   spec.extensions    = ["ext/cumo/extconf.rb"]
-
-  spec.add_runtime_dependency "numo-narray", numo_narray_version
-
-  spec.add_development_dependency "bundler", "~> 1.15"
-  spec.add_development_dependency "rake", "~> 10.0"
 end

data/ext/cumo/cuda/cudnn.c

@@ -50,7 +50,7 @@ cumo_cuda_cudnn_handle()
   @return [Boolean] Returns true if cuDNN is available
  */
 static VALUE
-rb_cudnn_available_p()
+rb_cudnn_available_p(VALUE self)
 {
 #if CUDNN_FOUND
     return Qtrue;
@@ -72,7 +72,7 @@ Init_cumo_cuda_cudnn(void)
     rb_define_const(mCUDA, "Cudnn", mCUDNN); // alias
     eCUDNNError = rb_define_class_under(mCUDA, "CUDNNError", rb_eStandardError);
 
-    rb_define_singleton_method(mCUDNN, "available?", RUBY_METHOD_FUNC(rb_cudnn_available_p), 0);
+    rb_define_singleton_method(mCUDNN, "available?", rb_cudnn_available_p, 0);
 #ifdef CUDNN_FOUND
     rb_define_const(mCUDNN, "CUDNN_POOLING_MAX", INT2NUM(CUDNN_POOLING_MAX));
     rb_define_const(mCUDNN, "CUDNN_POOLING_MAX_DETERMINISTIC", INT2NUM(CUDNN_POOLING_MAX_DETERMINISTIC));

data/ext/cumo/cumo.c

@@ -114,13 +114,17 @@ Init_cumo()
     const char* env;
     VALUE mCumo;
 
+#ifdef HAVE_RB_EXT_RACTOR_SAFE
+    rb_ext_ractor_safe(true);
+#endif
+
     mCumo = rb_define_module("Cumo");
 
     rb_define_const(mCumo, "VERSION", rb_str_new2(CUMO_VERSION));
 
-    rb_define_singleton_method(mCumo, "enable_compatible_mode", RUBY_METHOD_FUNC(rb_enable_compatible_mode), 0);
-    rb_define_singleton_method(mCumo, "disable_compatible_mode", RUBY_METHOD_FUNC(rb_disable_compatible_mode), 0);
-    rb_define_singleton_method(mCumo, "compatible_mode_enabled?", RUBY_METHOD_FUNC(rb_compatible_mode_enabled_p), 0);
+    rb_define_singleton_method(mCumo, "enable_compatible_mode", rb_enable_compatible_mode, 0);
+    rb_define_singleton_method(mCumo, "disable_compatible_mode", rb_disable_compatible_mode, 0);
+    rb_define_singleton_method(mCumo, "compatible_mode_enabled?", rb_compatible_mode_enabled_p, 0);
 
     // default is false
     env = getenv("CUMO_COMPATIBLE_MODE");

data/ext/cumo/depend.erb

@@ -1,3 +1,5 @@
+MAKEFLAGS = <%= ENV.fetch('MAKEFLAGS', "-j#{Etc.nprocessors}") %>
+
 TAGSRC = \
  ../../ruby/include/ruby/*.h \
  ../../ruby/*.c \
@@ -11,17 +13,17 @@ tags : TAGS
 TAGS : $(TAGSRC)
 	etags $(TAGSRC)
 
-C_TMPL = <%=Dir.glob("narray/gen/tmpl*/*.c").join(" ")%>
-CU_TMPL = <%=Dir.glob("narray/gen/tmpl*/*.cu").join(" ")%>
+C_TMPL = <%=Dir.glob("#{__dir__}/narray/gen/tmpl*/*.c").join(" ")%>
+CU_TMPL = <%=Dir.glob("#{__dir__}/narray/gen/tmpl*/*.cu").join(" ")%>
 
-C_COGEN = narray/gen/cogen.rb
-CU_COGEN = narray/gen/cogen_kernel.rb
-C_DEPENDS = $(C_TMPL) narray/gen/*.rb
-CU_DEPENDS = $(CU_TMPL) narray/gen/*.rb
+C_COGEN = <%= __dir__ %>/narray/gen/cogen.rb
+CU_COGEN = <%= __dir__ %>/narray/gen/cogen_kernel.rb
+C_DEPENDS = $(C_TMPL) <%= __dir__ %>/narray/gen/*.rb
+CU_DEPENDS = $(CU_TMPL) <%= __dir__ %>/narray/gen/*.rb
 
 <%
 list_type_c = []
-list_type_rb = Dir.glob("narray/gen/def/*.rb")
+list_type_rb = Dir.glob("#{__dir__}/narray/gen/def/*.rb")
 list_type_rb.each do |type_rb|
   type_name = File.basename(type_rb, ".rb")
   next if ENV['DTYPE'] and !type_name.downcase.include?(ENV['DTYPE'].downcase)
@@ -34,7 +36,7 @@ list_type_rb.each do |type_rb|
 
 <%
 list_type_cu = []
-list_type_rb = Dir.glob("narray/gen/def/*.rb")
+list_type_rb = Dir.glob("#{__dir__}/narray/gen/def/*.rb")
 list_type_rb.each do |type_rb|
   type_name = File.basename(type_rb, ".rb")
   next if ENV['DTYPE'] and !type_name.downcase.include?(ENV['DTYPE'].downcase)
@@ -47,12 +49,12 @@ list_type_rb.each do |type_rb|
 
 src : <%= list_type_cu.join(" ") %> <%= list_type_c.join(" ") %>
 
-build-ctest : cuda/memory_pool_impl_test.exe
+build-ctest : <%= __dir__ %>/cuda/memory_pool_impl_test.exe
 
-run-ctest : cuda/memory_pool_impl_test.exe
+run-ctest : <%= __dir__ %>/cuda/memory_pool_impl_test.exe
 	./$<
 
-cuda/memory_pool_impl_test.exe: cuda/memory_pool_impl_test.cpp cuda/memory_pool_impl.cpp cuda/memory_pool_impl.hpp
-	nvcc -std=c++14 <%= ENV['DEBUG'] ? '-g -O0 --compiler-options -Wall' : '' %> -L. -L$(libdir) -I. $(INCFLAGS) -o $@ $< cuda/memory_pool_impl.cpp
+<%= __dir__ %>/cuda/memory_pool_impl_test.exe: <%= __dir__ %>/cuda/memory_pool_impl_test.cpp <%= __dir__ %>/cuda/memory_pool_impl.cpp <%= __dir__ %>/cuda/memory_pool_impl.hpp
+	nvcc -std=c++14 <%= ENV['DEBUG'] ? '-g -O0 --compiler-options -Wall' : '' %> -L. -L$(libdir) -I. $(INCFLAGS) -o $@ $< <%= __dir__ %>/cuda/memory_pool_impl.cpp
 
-CLEANOBJS = *.o */*.o */*/*.o *.bak narray/types/*.c narray/types/*_kernel.cu *.exe */*.exe
+CLEANOBJS = <%= __dir__ %>/*.o <%= __dir__ %>/*/*.o <%= __dir__ %>/*/*/*.o <%= __dir__ %>/*.bak <%= __dir__ %>/narray/types/*.c <%= __dir__ %>/narray/types/*_kernel.cu <%= __dir__ %>/*.exe <%= __dir__ %>/*/*.exe

data/ext/cumo/extconf.rb

@@ -1,39 +1,19 @@
+# frozen_string_literal: true
+
 require 'rbconfig.rb'
+require 'fileutils'
 require "erb"
+require 'etc'
 require_relative '../../3rd_party/mkmf-cu/lib/mkmf-cu'
 
-if RUBY_VERSION < "2.0.0"
-  puts "Cumo::NArray requires Ruby version 2.0 or later."
-  exit(1)
-end
-
-def have_numo_narray!
-  version_path = File.join(__dir__, "..", "..", "numo-narray-version")
-  version = File.read(version_path).strip
-  gem_spec = Gem::Specification.find_by_name("numo-narray", version)
-
-  $INCFLAGS += " -I#{gem_spec.gem_dir}/ext/numo/narray"
-  if !have_header("numo/narray.h")
-    puts "
-    Header numo/narray.h was not found. Give pathname as follows:
-    % ruby extconf.rb --with-narray-include=narray_h_dir"
-    exit(1)
-  end
-
-  if RUBY_PLATFORM =~ /cygwin|mingw/
-    $LDFLAGS += " -L#{gem_spec.gem_dir}/ext/numo"
-    unless have_library("narray","nary_new")
-      puts "libnarray.a not found"
-      exit(1)
-    end
-  end
+def d(file)
+  File.join(__dir__, file)
 end
 
 def create_depend
   message "creating depend\n"
-  depend_path = File.join(__dir__, "depend")
-  File.open(depend_path, "w") do |depend|
-    depend_erb_path = File.join(__dir__, "depend.erb")
+  File.open(d("depend"), "w") do |depend|
+    depend_erb_path = d("depend.erb")
     File.open(depend_erb_path, "r") do |depend_erb|
       erb = ERB.new(depend_erb.read)
       erb.filename = depend_erb_path
@@ -42,7 +22,7 @@ def create_depend
   end
 end
 
-rm_f 'include/cumo/extconf.h'
+rm_f d('include/cumo/extconf.h')
 
 MakeMakefileCuda.install!(cxx: true)
 
@@ -52,10 +32,10 @@ end
 $CXXFLAGS << " -std=c++14"
 #$CFLAGS=" $(cflags) -O3 -m64 -msse2 -funroll-loops"
 #$CFLAGS=" $(cflags) -O3"
-$INCFLAGS = "-Iinclude -Inarray -Icuda #{$INCFLAGS}"
+$INCFLAGS = "-I$(srcdir)/include -I$(srcdir)/narray -I$(srcdir)/cuda #{$INCFLAGS}"
 
-$INSTALLFILES = Dir.glob(%w[include/cumo/*.h include/cumo/types/*.h include/cumo/cuda/*.h]).map{|x| [x,'$(archdir)'] }
-$INSTALLFILES << ['include/cumo/extconf.h','$(archdir)']
+$INSTALLFILES = Dir.glob(%w[include/cumo/*.h include/cumo/types/*.h include/cumo/cuda/*.h]).map { |x| [x, '$(archdir)'] }
+$INSTALLFILES << ['include/cumo/extconf.h', '$(archdir)']
 if /cygwin|mingw/ =~ RUBY_PLATFORM
   $INSTALLFILES << ['libcumo.a', '$(archdir)']
 end
@@ -113,17 +93,10 @@ cuda/cudnn
 cuda/cudnn_impl
 )
 
-if RUBY_VERSION[0..3] == "2.1."
-  puts "add kwargs"
-  srcs << "kwargs"
-end
-
-$objs = srcs.map {|src| "#{src}.o" }
+$objs = srcs.map { |src| "#{src}.o" }
 
 dir_config("narray")
 
-have_numo_narray!
-
 if have_header("dlfcn.h")
   exit(1) unless have_library("dl")
   exit(1) unless have_func("dlopen")
@@ -147,14 +120,14 @@ end
 
 have_type("bool", stdbool)
 unless have_type("u_int8_t", stdint)
-  have_type("uint8_t",stdint)
+  have_type("uint8_t", stdint)
 end
 unless have_type("u_int16_t", stdint)
-  have_type("uint16_t",stdint)
+  have_type("uint16_t", stdint)
 end
 have_type("int32_t", stdint)
 unless have_type("u_int32_t", stdint)
-  have_type("uint32_t",stdint)
+  have_type("uint32_t", stdint)
 end
 have_type("int64_t", stdint)
 unless have_type("u_int64_t", stdint)
@@ -162,17 +135,22 @@ unless have_type("u_int64_t", stdint)
 end
 have_func("exp10")
 have_func("rb_arithmetic_sequence_extract")
+have_func("RTYPEDDATA_GET_DATA")
 
 have_var("rb_cComplex")
 have_func("rb_thread_call_without_gvl")
 
-create_header('include/cumo/extconf.h')
+create_header d('include/cumo/extconf.h')
 $extconf_h = nil # nvcc does not support #include RUBY_EXTCONF_H
 
+# Create *.o directories
+FileUtils.mkdir_p('narray')
+FileUtils.mkdir_p('cuda')
+
 create_depend
 
-HEADER_DIRS = (ENV['CPATH'] || '').split(':')
-LIB_DIRS = (ENV['LIBRARY_PATH'] || '').split(':')
+HEADER_DIRS = (ENV['CPATH'] || '').split(File::PATH_SEPARATOR)
+LIB_DIRS = (ENV['LIBRARY_PATH'] || '').split(File::PATH_SEPARATOR)
 dir_config('cumo', HEADER_DIRS, LIB_DIRS)
 
 have_library('cuda')
@@ -186,4 +164,12 @@ if have_library('cudnn') # TODO(sonots): cuDNN version check
   $CXXFLAGS << " -DCUDNN_FOUND"
 end
 
+have_library('stdc++')
+
 create_makefile('cumo')
+
+begin
+  require 'extconf_compile_commands_json'
+  ExtconfCompileCommandsJson.generate!
+rescue LoadError
+end

data/ext/cumo/include/cumo/cuda/cudnn.h

@@ -14,9 +14,11 @@ extern "C" {
 #endif
 #endif
 
+extern VALUE cumo_cuda_eCUDNNError;
+
 #ifdef CUDNN_FOUND
 
-VALUE cumo_na_eShapeError;
+extern VALUE cumo_na_eShapeError;
 
 #define CUMO_CUDA_CUDNN_DEFAULT_MAX_WORKSPACE_SIZE 8 * 1024 * 1024
 

data/ext/cumo/include/cumo/intern.h

@@ -79,6 +79,7 @@ void cumo_na_parse_enumerator_step(VALUE enum_obj, VALUE *pstep);
 // used in aref, aset
 int cumo_na_get_result_dimension(VALUE self, int argc, VALUE *argv, ssize_t stride, size_t *pos_idx);
 VALUE cumo_na_aref_main(int nidx, VALUE *idx, VALUE self, int keep_dim, int result_nd, size_t pos);
+VALUE cumo_na_at_main(int nidx, VALUE *idx, VALUE self, int keep_dim, int result_nd, size_t pos);
 
 // defined in array, used in math
 VALUE cumo_na_ary_composition_dtype(VALUE ary);

data/ext/cumo/include/cumo/narray.h

@@ -141,7 +141,7 @@ extern "C" {
 # endif
 #endif
 
-#if SIZEOF_VALUE > 4
+#if SIZEOF_LONG > 4
 # undef INT322NUM
 # undef UINT322NUM
 # define INT322NUM(x) INT2FIX(x)
@@ -329,6 +329,12 @@ _cumo_na_get_narray_t(VALUE obj, unsigned char cumo_na_type)
 #define CUMO_RNARRAY_VIEW(val)       ((cumo_narray_view_t*)DATA_PTR(val))
 #define CUMO_RNARRAY_FILEMAP(val)    ((cumo_narray_filemap_t*)DATA_PTR(val))
 
+#ifdef HAVE_RTYPEDDATA_GET_DATA
+#define CUMO_RENUMERATOR_PTR(ptr)    ((cumo_enumerator_t *)RTYPEDDATA_GET_DATA(ptr))
+#else
+#define CUMO_RENUMERATOR_PTR(ptr)    ((cumo_enumerator_t *)DATA_PTR(ptr))
+#endif
+
 #define CUMO_RNARRAY_NDIM(val)       (CUMO_RNARRAY(val)->ndim)
 #define CUMO_RNARRAY_TYPE(val)       (CUMO_RNARRAY(val)->type)
 #define CUMO_RNARRAY_FLAG(val)       (CUMO_RNARRAY(val)->flag)
@@ -483,6 +489,12 @@ typedef unsigned int CUMO_BIT_DIGIT;
 #include "cumo/ndloop.h"
 #include "cumo/intern.h"
 
+// for Ractor support code
+#ifndef HAVE_RB_EXT_RACTOR_SAFE
+#   undef RUBY_TYPED_FROZEN_SHAREABLE
+#   define RUBY_TYPED_FROZEN_SHAREABLE 0
+#endif
+
 #if defined(__cplusplus)
 #if 0
 { /* satisfy cc-mode */

data/ext/cumo/include/cumo/template.h

@@ -112,9 +112,8 @@
         size_t dig = (pos) / CUMO_NB;                \
         int    bit = (pos) % CUMO_NB;                \
         ((CUMO_BIT_DIGIT*)(adr))[dig] =              \
-            (((CUMO_BIT_DIGIT*)(adr))[dig] & ~(1u<<(bit))) | ((val)<<(bit)); \
+            (((CUMO_BIT_DIGIT*)(adr))[dig] & ~(1u<<(bit))) | (((val)&1u)<<(bit)); \
     }
-// val -> val&1 ??
 
 #define CUMO_STORE_BIT_STEP( adr, pos, step, idx, val )\
     {                                           \
@@ -129,9 +128,8 @@
             pos += step;                        \
         }                                       \
         ((CUMO_BIT_DIGIT*)(adr))[dig] =              \
-            (((CUMO_BIT_DIGIT*)(adr))[dig] & ~(1u<<(bit))) | ((val)<<(bit)); \
+            (((CUMO_BIT_DIGIT*)(adr))[dig] & ~(1u<<(bit))) | (((val)&1u)<<(bit)); \
     }
-// val -> val&1 ??
 
 static inline int
 cumo_is_aligned(const void *ptr, const size_t alignment)